Oral Presentations

Cell & Tissue Engineering

Elucidating the Role of Nicotinamide Adenine Dinucleotide (NAD) in the Impairment of First Phase Insulin Release (FPIR) during Early-stage Type 2 Diabetes.

Nitya Gulati^1,2, Mahnoor Memon^1,2, Yufeng Wang^1,2, Romario Regeenes¹, Jonathan Rocheleau^1,2

¹University of Toronto, Toronto, Canada. ²University Health Network, Toronto, Canada

Rationale & Objectives: Pancreatic islet β-cells exhibit biphasic response to glucose stimulation, characterised by a rapid spike of first-phase insulin release (FPIR) followed by secondary oscillations, a process crucial for glucose homeostasis. Understanding the metabolic triggers behind FPIR is crucial, especially considering its early loss in type 2 diabetes (T2D). Nicotinamide adenine dinucleotide (NAD), an essential co-factor in insulin secretion regulation, is often overlooked in this context. Pro-inflammatory cytokines associated with T2D induce nitric oxide (NO) production via inducible nitric oxide synthase (iNOS), leading to DNA damage and activation of Poly [ADP-ribose] polymerase 1 (PARP-1). PARP-1 consumes NAD+ as a substrate thus depleting NAD+ pools.  

We aim to investigate how decreased NAD pools contribute to FPIR loss in early-stage T2D, to identify potential therapeutic interventions. 

Methods: We utilize islet-on-a-chip devices and live cell fluorescence imaging to study islet glucose metabolism, insulin secretion, and NADP pools in pancreatic islets with high temporal resolution. Pro-inflammatory cytokines (TNFα, IL1-β, IFNγ) were used to mimic T2D conditions in vitro. 

Results: Our data demonstrate a clear reduction in NAD pools within pancreatic β-cells exposed to pro-inflammatory cytokines, paralleled by decreased FPIR. Furthermore, we observe that NAD depletion inhibits lower glycolysis, corroborating its role in blocking FPIR. Importantly, we successfully rescued the loss of FPIR by blocking iNOS activity, blocking NAD consumption, and replenishing NAD pools using precursor. 

Conclusions & Significance: These findings underscore the critical role of NAD metabolism in FPIR and highlight potential therapeutic strategies for T2D. By elucidating the mechanisms linking NAD depletion, inflammatory pathways, and FPIR impairment, our study contributes to advancing precision medicine approaches in diabetes treatment. 

Modeling diastolic dysfunction in a human heart-on-a-chip platform

Omar Mourad¹, Stéphane Masse², Fatemeh Mirshafiei³, Julia Plakhotnik³, Xuetao Sun², Kumaraswamy Nanthakumar², Jason T Maynes³, Sara S Nunes²

¹University of Toronto, Toronto, Canada. ²University Health Network, Toronto, Canada. ³Hospital for Sick Children, Toronto, Canada

Rationale: Diastolic dysfunction is a manifestation of several cardiovascular diseases and is a hallmark of heart failure with preserved ejection fraction (HFpEF), which is a tremendous burden on healthcare systems worldwide due to high mortality rates. Leveraging established heart-on-a-chip (HOC) technology, our goal is to generate a human-relevant model of diastolic dysfunction that recapitulates key features of HFpEF, such as increased passive tension, cardiomyocyte hypertrophy, impaired calcium handling, slowed relaxation, and interstitial fibrosis. 

Methods: Using a HOC platform that enables the real-time assessment of passive tension and active force, we engineered human cardiac microtissues from stem cell-derived cardiomyocytes and cardiac fibroblasts. Co-treatment of microtissues with endothelin-1 (ET-1) and transforming growth factor beta-1 (TGF-β) – factors upregulated in HFpEF samples that are known to induce cardiac hypertrophy and fibrosis – led to an increase in passive tension while active force of contraction was unaffected.

Results: HOC microtissues treated with ET-1 and TGF-β (disease) displayed a 3-fold increase in passive tension with unchanged active force, and pathological CM hypertrophy compared to untreated (healthy) tissues. Calcium handling of disease tissues was impaired due to slowed reuptake kinetics. In addition, high-resolution pixel tracking analysis of disease tissues revealed that the velocity of tissue relaxation was reduced.  Disease tissues were mildly fibrotic, although not to a degree severe enough to impair systolic function. Co-treatment of disease HOCs with the anti-senescence drugs dasatinib and quercetin partially reversed the disease phenotype by reducing passive tension by 45% compared to vehicle.

Conclusion: Inspired by the clinical features of HFpEF, this novel human HOC model of diastolic dysfunction recapitulates several key hallmarks of the disease including impairment of both passive and active relaxation. This human-relevant model will improve our understanding of the biological pathways that contribute to diastolic dysfunction and can be used as a drug testing platform.

Effects of Low-Magnitude High-Frequency Vibration on Prostate Cancer Progression and Bone Metastasis

Amel Sassi¹, Kimberly Seaman¹, Xin Song¹, Lidan You^1,2

¹University of Toronto, Toronto, Canada. ²Queen’s University, Kingston, Canada

Prostate cancer preferentially metastasizes to bone, whereby 80% of men who die from prostate cancer exhibit signs of bone metastases. To mitigate these effects, exercise is often recommended to cancer patients due to the beneficial effects on bone remodelling. However, physical activity may be challenging for elderly or bedridden patients. As such, vibration has emerged as a safe, effective, and easy to perform alternative therapy. Specifically, low magnitude high frequency (LMHF) has been shown to activate osteocytes and reduce breast cancer cell migration. Nevertheless, the effects of vibration on prostate cancer cell extravasation remains to be elucidated. We hypothesize that LMHF vibration (0.3 g, 60 Hz, 1h/day) will decrease prostate cancer cell extravasation through the activation of osteocytes and LMHF will directly influence prostate cancer cell growth. MLO-Y4s (osteocytes), HUVECs (endothelial cells) and PC3s (prostate cancer cells) were seeded into the appropriate channels of the bone metastasis-on-a-chip platform. Devices were then placed on a custom-made vibration platform for 1 hour every day for 3 days and extravasation distance was monitored. To examine potential mechanisms, we examined PC3 adhesion onto HUVECs treated with either static or vibration-stimulated MLO-Y4 conditioned media. Additionally, to examine the effects on prostate cancer cells, an apoptosis, viability, and colony formation assay were carried out. LMHF vibration significantly reduced extravasation distance by approximately 42.8%. We also observed that 29% more PC3 cells remained adhered to HUVECs in static MLO-Y4 CM. Furthermore, a significant reduction in PC3 colony formation was observed, but not in apoptosis or viability. Results indicate that LMHF vibration is effective at reducing the incidence of prostate cancer bone metastases through an overall reduction in extravasation and colony formation. These findings are critical for understanding how mechanical stimulus can be utilized to employ alternative or adjunctive prostate cancer therapies in the clinical setting.

Molecular Engineering

ROS-Responsive Hydrogel Biosensor for Improved Monitoring of Inflammatory Bowel Diseases

Lucia Huang, Abigail Clapperton, Helen Tran, Caitlin Maikawa

University of Toronto, Toronto, Canada

Inflammatory bowel diseases (IBD) are characterized by chronic unregulated inflammation of tissue in the digestive tract. Patients present with symptoms such as abdominal pain, diarrhea, bloody stool, and have increased risk of infection and cancer. Current treatment strategies aim to completely heal damaged intestinal tissue using therapeutics and require regular assessments of intestinal inflammation to determine progress towards this target. Accessible, point-of-care tools to measure intestinal inflammation based on objective biomarkers would enable informed adjustments to therapeutic regimens and improve treatment effectiveness. Intestinal levels of reactive oxygen species (ROS) have been demonstrated to significantly increase during inflammation. In this work, we synthesized an orally administered ROS-responsive hydrogel biosensor for monitoring IBD. The hydrogel is cross-linked using ROS-responsive functional groups and disassembles in the presence of elevated ROS levels to release a fluorescent reporter. The release kinetics of the ROS-responsive hydrogel are concentration-dependent and can be tuned by modulating the mechanical stiffness through cross-linking density. The fluorescent signal in stool is easily visualized under UV light to help direct patient treatments to halt IBD progression.

Multi-Axes Electromagnetic Sample Handler (ESH) For Multi-View Imaging Of Organoids

Thaisa Luup Carvalho Kannen^1,2, Christopher Yip¹, Laurence Pelletier², Aaron Au¹

¹University of Toronto, Toronto, Canada. ²Lunenfeld-Tanenbaum Research Institute, Toronto, Canada

Rationale & Objectives: Current microscope stages position a given sample along linear XYZ-axes. In order to capture further three-dimensional (3D) views of biological specimens, multi-view microscopy can be used. It images samples from multiple angles and reconstructs these views in a 3D image. However, it is costly and has a limited number of views due to the physical constraints of the microscope. Live organoid-based studies also face light penetration limitations in deeper microscopic planes due to light scattering. Furthermore, the imaging systems introduce an unavoidable image distortion, the point spread function (PSF). These challenges limit the image quality.

Methods: Our proposed alternative for multi-view microscopy comprises an electromagnetic sample handler (ESH), which effectively addresses the limitations of traditional multi-view microscopy. The ESH significantly improves the overall image quality of organoid structures by applying a multi-axis rotational system for imaging from multiple angles and reconstructing it into a 3D image. The acquired images are processed and aligned by matching fiducial marks in the different views and using mathematical algorithms to compute their spatial relationships.

Results: The ESH provides optimized angular positioning in the XYZ-theta space, which is useful for studying dynamic processes within living specimens, such as organoids. ESH provides improved resolution and helps mitigate the limitation of light depth penetration in organoid imaging. Multi-view systems also improve the PSF deconvolution. 

Conclusions & Significance: Examining samples with stability and precision from multiple angles provides scientists valuable insights into organoids’ 3D structure and dynamics. This advanced imaging approach provides higher resolution in 3D reconstructions than traditional methods. It is a valuable tool for biological and biomedical research applications, enabling the investigation and precise quantification of intricate 3D structures and dynamics of biological specimens in unprecedented detail.

Towards DNA Origami Therapeutics: How PEG Length, Density, and Oligolysine Length Controls Hybrid DNA-Origami-Polymer Material Function

Mohammadamir G. Moghadam, Leo Chou

University of Toronto, Toronto, Canada

Rationale & Objectives: DNA nanostructures (DNs) are nanoparticles made from DNA. They hold promise in biomedical applications like cancer vaccines and drug delivery, thanks to their precise control over synthesis, shape, size, and surface chemistry. Yet, their clinical application is limited by challenges such as nuclease degradation and the need for unphysiological salt levels. Our study aims to overcome these obstacles by introducing DNA-nanostructure-polymer hybrid materials, focusing on oligolysine-PEG (polyethylene glycol) coatings, to enhance stability and functionality in physiological conditions.

Methods: We synthesized the hybrid materials through electrostatic coating of DNs with oligolysine-PEG. Starting with 40 coatings, we explored a broad design space of oligolysine and PEG lengths, and PEG density. After confirming hybrid formation (coating), we evaluated peptide-polymer coating affinity to DNs, eliminating 63% of the coatings based on their performance compared to the current widely used coating in the field. The remaining candidates underwent stability assays in physiological salt levels and DNase I conditions. The successful candidates will be examined for their cell toxicity and impact on antibody-decorated DN binding to cells.

Results: Successful synthesis of the hybrid materials was achieved. Notably, increasing PEG length and density reduced the coating’s affinity for DNs. We hypothesize that the absolute number of PEG molecules on each peptide chain affects coating binding affinity more significantly than the density of these PEG molecules on the peptide, within the range of peptide lengths studied.

Conclusions & Significance: This study advances the application of DNA nanostructures in biomedicine, particularly for creating more effective, targeted therapeutic delivery systems. By overcoming major obstacles and proving the feasibility of DN-polymer hybrids, we establish a novel platform for targeted delivery, significantly impacting cancer vaccines and immunomodulatory therapeutics.

Clinical Engineering

Development of an Artificial Intelligence Model to Assess the Risk of Exercise in Persons with Asthma

Shaghayegh Chavoshian^1,2, Xiaoshu Cao^1,2, Yan Fossat³, Azadeh Yadollahi^1,2

¹University of Toronto, Toronto, Canada. ²University Health Network, Toronto, Canada. ³Klick Health Inc., Toronto, Canada

Rationale & Objectives: Asthma is a chronic inflammatory disease of the small airways, affecting over 200 million people globally. Regular physical exercise can improve lung function and asthma control. Nevertheless, being active or exercising could be a risk factor for asthma exacerbations. We hypothesized that monitoring physiological signals during exercise may help to detect potential worsening in asthma and could be used to help persons with asthma adjust their exercise. 

Methods: In this study, adults with asthma were asked to cycle for 10 minutes. During exercise, Electrocardiogram (ECG) and thoraco-abdominal motion/respiration belt signals were measured continuously using electrodes and calibrated respiratory inductance plethysmographs, respectively. Before and after exercise, small airway narrowing was assessed based on respiratory impedance measured by forced oscillation technique. Therefore, two class labels, including airway narrowing and not airway narrowing, were determined from respiratory impedance changes. The signals were segmented into 60-second windows with no overlap between successive segments. After signal processing, we extracted time and frequency domain features from ECG and respiration signals. After feature selection, to detect airway narrowing, different machine learning classifiers were fine-tuned and evaluated using a leave-one-subject-out cross-validation approach. 

Results: A total of 23 individuals (11 females, age: 56.3 ± 10.9 years, BMI: 27.4 ± 5.7 kg/m²) with asthma were enrolled in the study. Out of 204 windows of signals, 45% were from individuals with airway narrowing and 55% were from individuals without airway narrowing. The random forest model performed the best compared to other models with an accuracy of 77.01%, a recall of 82.69%, a specificity of 68.57%, a precision of 79.63%, and an F1 score of 81.13%. 

Conclusions & Significance: These results provided proof of concept that technologies with embedded respiratory and cardiac signal monitoring may be able to predict airway narrowing during exercise in individuals with asthma.

Detecting Slips Using Audio Signals: A Deep Learning Approach

Davood Dadkhah^1,2, Hamed Ghomashchi², Tilak Dutta^1,2,3

¹BME, University of Toronto, Toronto, Canada. ²KITE – Toronto Rehabilitation Institute – UHN, Toronto, Canada. ³Rehabilitation Sciences Institute, University of Toronto, Toronto, Canada

Rationale & Objectives: Falls on icy surfaces are a significant public health concern. This study aims to enhance pedestrian safety by developing a more accurate method for detecting slips using wearable sensor data, ultimately improving winter footwear performance tests by making them more objective.

Methods: We recruited 27 participants to walk on icy surfaces in the WinterLab, capturing data with a range of devices including wearable sensors on footwear (Android phones, wireless microphones, and embedded microphones), GoPro cameras, Vicon motion capture system, and participant’s self-report. This setup allowed for the collection of a comprehensive dataset, including kinematic, audio, and video data, to analyze slip events in detail. We employed deep learning models, particularly convolutional neural networks, trained on a dataset comprising over 80,000 steps from participants wearing various winter footwear on icy inclines. The models were tasked with distinguishing steps with slip events from steps without slips.

Results: Preliminary results showed promising accuracy (~80%), indicating the model’s potential to reliably identify slip events. This suggests an innovative approach to assessing the slip resistance of winter footwear.

Conclusions & Significance: The study’s findings highlight the feasibility of using wearable sensors and deep learning for slip detection on icy surfaces. This approach could significantly contribute to reducing fall-related injuries, providing a data-driven basis for footwear assessment and public safety recommendations. Further research will aim to refine these models for broader application in real-world settings.

Efficacy Evaluation of the Remote Surgical Skill Training with the Marshall III Thoracic Surgery Simulation Training Kit

Kate Kazlovich^1,2, Nicholas Bernards², Fumi Yokote², Hiroyuki Ogawa², Takahiro Yanagihara², Kenta Nakahashi², Takamasa Koga², Yoshihisa Hiraishi², Andrew Effat^1,2, Nadia Mohammed², Blair Marshall³, Laura Donahoe², Kazuhiro Yasufuku²

¹U of T Biomedical Engineering, Toronto, Canada. ²Division of Thoracic Surgery, Toronto General Hospital, University Health Network, Toronto, Canada. ³Division of Thoracic Surgery, First Physicians Group, Florida, USA

Rationale & Objectives: The Marshall III Thoracic Surgery Simulation Training Kit has been designed and manufactured to help hone skill development and retention among senior thoracic surgery trainees. A study focused on remote training for pulmonary artery anastomosis techniques used in transplant surgery was conducted to validate the efficacy of the kit. 

Methods: Training sessions for 7 participants were conducted with an onboarding demonstration from a senior thoracic surgeon at the Toronto General Hospital (TGH) followed by a subsequent at-home practice with the return for final in-person skill and performance assessment by two senior surgeons. Participants completed a total of 10 independent practices at home. Parameters such as Estimated Time to Completion (ETC), anastomosis water leak test, Thoracic Anastomosis Skill Assessment Checklist (TASAC), and User Experience (UX) feedback were collected.

Results: Overall ETC (n = 7) improved by 18.8% between the first and last recorded training session. The TASAC showed an increase from 62% to 97% completion of the 7-parameter performance scale. The overall impression of ability rated on a scale of 1 to 5 showed an average rating increase from 3.2 to 4.8 points, indicating that most participants had excellent results towards the end of the study. The evaluation of the participant’s ability to independently complete anastomosis increased from 2.8 to 4.9 points on a scale of 1 to 5, with 5 representing one’s ability to perform anastomosis without supervision. 

Conclusions & Significance: A portable anastomosis kit is a viable tool that can be used for training and skill retention in senior thoracic surgery trainees. Its relatively low manufacturing cost of CAD 300 reduces the barrier of access to on-demand training. Device portability provides the user with access to on-demand training. All study participants agreed that they would use or recommend the training kit for practice and retention of surgical skills. 

Poster Presentations

Cell & Tissue Engineering

Abstract 1: Development of a Comprehensive Computational Model for Simulating the Proliferation Dynamics of Human Pluripotent Stem Cells

Ferdinand Avikpe¹, Faisal Alibhai², David Romero Torres¹, Michael LaFlamme^1,2, Cristina Amon¹

¹University of Toronto, Toronto, Canada. ²University Health Network, Toronto, Canada

Rationale & Objectives: In silico or computational models have emerged as vital tools for understanding complex biological processes in the field of tissue engineering, and in particular for human pluripotent stem cells (hPSCs). Traditional in vitro and in vivo methods have been foundational in advancing our understanding of hPSCs, providing essential data that are instrumental in the development of in silico models. While these methods are invaluable, they can be resource-intensive and time-consuming. In silico methods complement these approaches by using computational modeling to offer a cost-effective and efficient alternative for accelerating research processes and enhancing productivity. This study aims to bridge a crucial gap in stem cell research by developing a comprehensive in silico model that accurately simulates the proliferation dynamics of hPSCs in relation to their biochemical environments.

Methods: Through the integration of experimental design protocols, mathematical statistics, and optimization concepts,  we developed a system of ordinary differential equations (ODEs), to predict the proliferation dynamics of hPSC populations using eight parameters that have been carefully inferred from optimally designed in vitro experiments. Variables such as the initial cell density and the concentrations of nutrients (glucose) and waste products (lactate) were incorporated into the model through ODE source/sink terms and modulation terms which follow constant, linear and Michaelis-Menten mathematical forms.

Results: The validated in silico model can predict hPSC proliferation patterns, capturing both the observed experimental magnitudes and trends. Furthermore, the model provides valuable insights into the threshold levels required for optimal hPSC proliferation, which will allow the simulation of various experimental scenarios.

Conclusions & Significance: This model will be applicable for optimizing cell culture conditions and experimental protocols, thereby improving cell yield and reducing cost. Thus, the proposed model has significant potential to advance the development of regenerative stem cell therapies.

Abstract 2: Biomaterial fabrication of angle-ply and cross-ply structures

Chantel Campbell, Axel Guenther, Jiayu Li, Jacqueline Zhu

University of Toronto, Toronto, Canada

Angle-ply and cross-ply structures, characterized by stacks of ultrathin aligned collagen composites, are the hallmark of numerous load-bearing tissues. They impart tensile strength to blood vessel walls and underlie the biomechanics of corneal and musculoskeletal tissues. However, their regeneration is limited, and no biomimetic replacements exist. Available multilamellar fabrication techniques are limited by the preparation of ultrathin (<5 micron) aligned collagen filaments, and often produce structures with biomechanically poor and inconsistent interlamellar adhesion. Here, we present multilamellar biomaterials, prepared from aligned collagen composites, that recapitulate native tissue biomechanics while exhibiting consistent interlamellar adhesion. 90-degree and 0-degree bilamellar composites were engineered, with 3.5-3.6 micron thickness, 32-35 megapascal vs. 14-50 megapascal anisotropic elastic modulus, and 470-560 pascals interlamellar adhesive strength respectively. The collagen ultrastructure, of an engineered penta-lamellar, with 0-degree , 45-degree and 90-degree interlamellar angles, revealed consistent interlamellar contact. We anticipate future applications of these biomaterials as angle-ply and cross-ply tissue substitutes.

Abstract 3: Do Culture Enhanced Mesenchymal Stromal Cells Secrete Extracellular Vesicles with Improved Immunomodulatory Properties?

Griffin Copp^1,2, Sowmya Viswanathan^1,2

¹University Health Network, Toronto, Canada. ²University of Toronto, Toronto, Canada

Extracellular vesicles (EVs) secreted by mesenchymal stromal cells (MSCs) have emerged as primary therapeutic agents due to their potential cargo of nucleic acids, proteins, lipids, and signalling molecules. Despite the mixed clinical success of MSC therapies, recent recognition of EVs as primary therapeutic agents warrants investigation into strategies enhancing their efficacy. This study explores the impact of various non-genetic cell-enhancing techniques on MSC-secreted EVs, explicitly focusing on their ability to polarize macrophages toward an M2 (anti-inflammatory) phenotype. Understanding these immunomodulatory properties is crucial for optimizing cell-based therapeutic strategies. 

Adipose tissue-derived MSCs are enhanced through inflammatory priming, transient hypoxic conditioning, or transient 3D culturing . Small EVs are isolated using differential ultracentrifugation and characterized through nanoparticle tracking analysis (NTA), transmission electron microscopy (TEM), Bradford assay (BCA), and Western blot. A macrophage polarization assay, assessed via real-time qPCR, is employed to determine the immunomodulatory capacity of EVs. 

Preliminary data reveals a distinct clustering of enhanced MSC-EVs in a principal component analysis of macrophage polarization read-outs, indicating altered immunomodulatory effects compared to naive MSC-EVs. TEM images verify the morphology of small EVs, while NTA demonstrates particle quantity and size distribution (peaking < 200nm). BCA quantifies protein content, and Western blot validates the presence of exosome-specific markers (CD-9, -63, -81). 

This study sheds light on the increased immunomodulatory potential of MSC-derived EVs following various cell-enhancing techniques. The observed differences in macrophage polarization suggest a nuanced interplay between MSC modifications and EV functionality. These findings contribute valuable insights for developing consistently therapeutic cell products, emphasizing the pivotal role of EVs in cellular strategies for clinical applications. 

Abstract 4: Investigating culture-expanded adipose tissue mesenchymal stromal cell fitness in autologous fat transfer for Trapeziometacarpal Osteoarthritis

Kevin Fan^1,2, Kevin Robb^1,2, Sowmya Viswanathan^1,2,3

¹Osteoarthritis Research Program, Division of Orthopedic Surgery, Schroeder Arthritis Institute, Toronto, Canada.²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³Department of Medicine, Division of Hematology, University of Toronto, Toronto, Canada

Rationale & Objectives: Adipose tissue-derived mesenchymal stromal cells (MSC(AT)) produce paracrine factors involved in immunomodulation and angiogenesis, which are thought to play an important role in autologous fat transfer (AFT)—a minimally invasive treatment for Trapeziometacarpal Osteoarthritis (TMOA). Given its rich source in adipose tissue, we hypothesized that correlations between patient-reported outcomes and MSC(AT) fitness (the biological ability of cells to exert immunomodulatory, angiogenic, or other functions) could reveal important cellular characteristics involved in successful AFT treatments.

Methods: MSC(AT) (N = 9 human donors) obtained from patient fat grafts were cultured with or without addition of pro-inflammatory licensing cytokines. Clinical scores (qDASH, VAS, TASD, MPQ, pinch and grip) were collected from patient-reported questionnaires at baseline, 3, 6, and 12 months. We used a matrix of assays to measure 24 genes and cell proliferation and morphology characteristics. PCA and Spearman correlation analysis were used as statistical tools to identify significant correlations between clinical scores and MSC(AT) characteristics. 

Results: Delta qDASH (>20) and TASD (>15) scores between timepoints identified 3 responders and 6 non-responders to AFT treatment. PCA analysis suggested that responder MSC(AT) donors displayed distinct immunomodulatory gene expression profiles compared to non-responder MSC(AT) donors. Spearman correlation analysis revealed significant correlations between higher angiogenic/immunomodulatory gene expression, lower cell proliferation and greater clinical improvement. Growth factors correlated negatively with clinical improvement but positively with cell proliferation. 

Conclusions & Signficance: Correlations to clinical improvement in TMOA revealed that donor heterogeneity in MSC(AT) within fat grafts may drive this biology; MSC(AT) with augmented angiogenic and immunomodulatory fitness and reduced progenitor properties appear more likely to result in TMOA clinical improvements, suggestive of potential mechanisms of action that will inform future clinical studies.

Abstract 5: Application of Vacuum Thermoforming for Manufacturing of GLAnCE Platform With Improved Optical Properties

Aleksandra Fomina¹, Adam Tam², Jennifer Lam², Alison McGuigan^2,1

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ²Department of Chemical Engineering and Applied Chemistry, University of Toronto, Toronto, Canada

GLAnCE (Gels for Live Analysis of Compartmentalized Environments) is a 3D in vitro platform of an array of micro-molded cell-containing hydrogels that allow longitudinal image-based monitoring of tumor cell dynamics and therapy response. Currently, image acquisition in the GLAnCE platform can be completed at low magnifications only (up to 10x) due to the great thickness of the bottom component produced by the hot embossing microfabrication technique (0.7mm). This platform feature limits its use with objectives of high magnification and numerical aperture and thus hinders the investigation of tumor cell phenotypes at single-cell and subcellular levels.

Here, we employ the vacuum thermoforming microfabrication technique to enable the manufacturing of the GLAnCE bottom component from thin polystyrene films (0.192mm). Using a desktop thermoformer, positive aluminum molds, and methods for polydimethylsiloxane casting and brightfield imaging, we first optimize vacuum thermoforming parameters for a 3-by-3-channel mold and then scale up the process to a 12-by-8-channel version. Lastly, we validate the thermoformed GLAnCE bottom component in the current platform assembly and cell seeding workflow and acquire images using confocal microscopy and high-magnification objectives. 

The optimized vacuum thermoforming process reduces the platform’s thickness 14-fold and allows the manufacturing of GLAnCE bottom components as thin as 0.050mm. The thermoformed GLAnCE bottom component is compatible with the current workflow for platform assembly and cell seeding, as evidenced by the negative leak test and cell proliferation in three different matrices over time. Importantly, the reduced platform thickness allows imaging of fluorescent protein-expressing cells and immunofluorescently stained cells and intracellular structures in 3D at up to 63x magnification.

The thermoformed GLAnCE platform offers a unique opportunity to study tumor cell phenotypes in 3D with single-cell and subcellular resolution. We envision applying this platform for studies on single-cell dynamics of modeled tumor microenvironments to provide insights into the drivers of disease progression.

Abstract 6: Unveiling muscle stem cell quiescence re-entry: Combining in silico predictions with an engineered tissue assay identifies a responsible skeletal muscle niche cue

Erik Jacques^1,2, Pauline Garcia³, Yechen Hu¹, Sidy Fall³, Cyril Degletagne³, Maira Pedroso De Almeida², Stephane Angers², Fabien Le Grand³, Penney M. Gilbert^1,2,4

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ²Donnelly Center, University of Toronto, Toronto, Canada. ³Institue NeuroMyoGène, Université Claude Bernard Lyon, Lyon, France. ⁴Department of Cell and Systems Biology, University of Toronto, Toronto, Canada

Upon skeletal muscle injury, muscle stem cells (MuSCs) launch a regenerative response to repair damaged, and create new, myofibers. Orchestration of this process requires communication between MuSCs and the niche; the primary component being the myofibers. MuSCs leverage the reversible state of cell-cycle arrest, termed ‘quiescence’, which allows for the preservation of stemness and persistence of this population over a lifetime. Thus, a critical juncture during regeneration is MuSCs returning to quiescence. Niche influences over this decision remain incompletely understood, but are poised to serve as a target to restore MuSC function in cases of dysregulated quiescence such as aging. We recently developed mini-IDLE (Inactivation and Dormancy LEveraged in vitro); a 3D biomimetic niche that induces a state of quiescence unto MuSCs. Characterization of the system revealed that the myotube (immature myofiber) component of the biomimetic niche was essential to driving the phenotype, suggesting a possible culprit in vivo. To rapidly expand knowledge of possible quiescence-inducing cues, we established a discovery pipeline whereby in silico predictions were tested in mini-IDLE. First, we performed snRNAseq of regenerating muscle to capture the full breadth of myogenic progression. Using cell communication inference tools, we highlighted receptor-ligand interactions between MuSCs and myofibers. Next, we conducted a functional genomic screen of the predicted interactions using Cas9-based technologies to perturb ligand genes in mini-IDLE and evaluate MuSC quiescence re-entry. Results and subsequent in vivo validation yielded perlecan, a proteoglycan that is secreted into the niche during myofiber formation that serves as a pro-quiescence ligand. Our group then demonstrated that the temporal profile and overall levels of perlecan are altered with age, impacting MuSC function. However, this was amenable to rescue with exogenous treatment. Thus, coupling in silico predictions with our in vitro assay, has made possible the study of a previously inaccessible window of MuSC biology. 

Abstract 7: Optimization of a HAMVEC-ASC seeded gelatin-polyurethane scaffold to support stem cell derived cardiomyocyte maturation

Alexandra Jucan, Yizhou Chen, C.W. Brian Webb, Kate D. MacQuarrie, J. Paul Santerre

University of Toronto, Toronto, Canada

Cardiac tissue regeneration using transplanted cardiomyocytes (CMs) is limited by the rapid death of most cells after their transplant, due to immature cells and the lack of oxygen in the transplant area. Support cells such as adipose derived stem cells (ASCs) and endothelial cells (ECs) have been shown to provide paracrine signals that can inhibit cardiomyocyte (CM) cell death and increase the expression of key functional proteins (e.g. Connexin-43), as well as assemble into vessels and drive angiogenesis that can be used to rapidly perfuse the transplanted cells. The delivery of HAMVECs and ASCs alongside CMs has the potential to improve CM cell therapies. This work seeks to establish an injectable system to deliver ECs and ASCs to support CM maturation and survival. 

Gelatin-polyurethane scaffolds are electrospun and laser cut into 500um by 500um micro-scaffolds, that can be injected through standard needle gauges. Currently, 7 days of culture of ECs and ASCs on the micro-scaffolds grow over the micro-scaffolds and form a cell sheet that cannot be injected in the form of micro-scaffolds. In this work, we have sought to vary the cell culture conditions on the micro-scaffolds, by seeding ECs and ASCs in a 1:2 ratio at either a high (400,000 cells/cm²) or low (100,000 cells/cm²) cell density, with or without FBS, and cultured for 24 vs 72 hours. Using the ideal culture condition, the effect of paracrine signals from HAMVECs and ASCs on CMs will be evaluated using a Transwell coculture model and staining for cardiac functional proteins. Co-culture-support of  CM tissue engineered constructs will enhance cardiac regenerative therapies. 

Abstract 8: Polarized trafficking of cell-cell adhesion proteins facilitates scarless wound healing

Sofia Mendez-Lopez^1,2, Kate MacQuarrie^1,2, Katheryn Rothenberg^1,2, Rodrigo Fernandez-Gonzalez^1,2,3,4

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ²Ted Rogers Centre for Heart Research, University of Toronto, Toronto, Canada. ³Department of Cell and Systems Biology, University of Toronto, Toronto, Canada. ⁴Developmental and Stem Cell Biology Program, The Hospital for Sick Children, Toronto, Canada

Collective cell movements drive the formation and repair of tissues in development and the spread of metastatic disease. To understand how cells coordinate their migration, we investigate wound healing in the epidermis of Drosophila embryos. Upon wounding, a supracellular cable composed of the cytoskeletal protein actin and the molecular motor non-muscle myosin II assembles around the wound, creating tension to coordinate cell movements and drive wound closure. The actomyosin cable forms through the polarization of actin and myosin in the cells adjacent to the wound. In parallel, adherens junction proteins, including E-cadherin, are depleted from former bicellular junctions at the wound edge and accumulate at former tricellular junctions (TCJs) around the wound. The reorganization of cell-cell adhesions is necessary for rapid wound healing. E-cadherin is removed from the former bicellular junctions via endocytosis, but it is unclear how E-cadherin is delivered to TCJs, and whether the E-cadherin accumulating at TCJs is recycled protein that was previously endocytosed or protein originating elsewhere in the cells. To examine the potential role of protein recycling in E-cadherin accumulation at TCJs, we manipulated the activity of the small GTPase Rab11, which marks vesicles for slow recycling. Reducing Rab11 activity by overexpressing a dominant-negative form (Rab11DN), slowed down wound closure by 29%. While myosin polarization occurred normally in Rab11DN embryos, the accumulation of E-cadherin at TCJs decreased by 33%. Together, our results suggest that Rab11 is partially responsible for E-cadherin remodeling during wound repair, and that cell-cell adhesion rearrangements control the rate of wound healing independent of cytoskeletal polarity. Future work will investigate the role that other endocytic and recycling compartments, as well as diffusion of junctional material along cell membranes play in the redistribution of E-cadherin and rapid wound repair.

Abstract 9: Facile Fabrication of Anisotropic Porous Collagen Scaffolds

Kenneth Kimmins¹, Qin Wang², Christopher McCulloch², Eli Sone^1,3,2

¹Institute of Biomedical Engineering, Toronto, Canada. ²Faculty of Dentistry, Toronto, Canada. ³Faculty of Materials Science & Engineering, Toronto, Canada

Collagen scaffolds are promising candidates for enhancing tissue regeneration as they share certain properties of native tissues. We considered that scaffold anisotropy and porosity are properties that collectively guide cell function to promote tissue regeneration by enhancing the assembly of oriented collagen fibrils.But fabricating scaffolds with these properties is not easily achieved. Here we report a novel but simple gas diffusion fabrication method for creating anisotropic, macroporous collagen scaffolds to obtain insights into the behavior of cultured cells.

Type I rat tail collagen in acetic acid was neutralized with ammonia vapors to create collagen scaffolds with aligned macropores. Optical and scanning electron microscopy were used to evaluate scaffold morphology. Immortalized human gingival fibroblasts were cultured on scaffolds treated with fibronectin to improve cell attachment. Cell nuclei were stained with DAPI and actin filaments with phalloidin. Scaffolds were immunostained for collagen to study the spatial relationship between cells and the scaffolds. Cells were imaged with confocal microscopy.

Consistent with our hypothesis that pore formation in collagen is a phase separation phenomenon, we found that the size and numbers of anisotropic macropores are affected by the pH used for collagen assembly. Acetic acid and ammonia concentrations exerted marked effects on porosity; fine-tuning these concentrations enabled the fabrication of scaffolds of defined porosity. Cells formed long, actin-enriched extensions, similar to human gingiva. 

We hypothesize that cells will align in response to anisotropic cues presented by the scaffolds. This behavior will lead to the production of oriented collagen, verified by picrosirius red staining of the nascent collagen. We conclude that collagen scaffold fabrication via ammonia gas diffusion is a novel and promising approach for studying the effects of anisotropy and macroporosity without the need for complex and expensive fabrication methods. We anticipate that these methods will create new opportunities for tissue regeneration.

Abstract 10: Computational analysis of donor heterogeneity and critical processing parameters in mesenchymal stromal cells identifies suitable donors and CPP conditions for improved MSC expansion and potency.

Oreoluwa Kolade¹, Julie Audet², Sowmya Viswanathan¹

¹Krembil Research Institute, Toronto, Canada. ²University of Toronto, Toronto, Canada

Rationale & Objectives: Mesenchymal Stromal Cells (MSCs), known for their inflammation-regulating properties, face variability challenges hindering clinical and commercial success. Employing Design of Experiments (DoE) and desirability analysis, this study investigates donor heterogeneity and Critical Processing Parameters (CPPs) used to manufacture MSCs. It assesses their dual impact on potency and expansion. 

Methods: We use DoE to assess input parameters—donor heterogeneity, CPP parameters including plating MSC density, medium composition, and oxygen concentrations—across 13 bone marrow donors, including osteoarthritis (OA) patients from a previous clinical trial (NCT02351011). An 8-gene curated panel of anti-inflammatory/angiogenic genes under licensed conditions serves as MSC critical quality attributes (CQAs). Using the DoE, we evaluate novel effects of donor heterogeneity and CPP conditions on MSC expansion and CQAs. 

Results: Principal component analysis (PCA) unveils heterogeneity among 12 CPP combinations and 8 genes across 13 donors. Significant PC1 scores (p<0.01 for responders vs. healthy, p<0.0001 for responders vs. non-responders) underscore donor’s role in MSC potency heterogeneity. The 8-gene panel, ranked via sensitivity analysis, identifies TSG-6 (TNF-stimulated gene 6 protein), VEGF (Vascular endothelial growth factor), and PD-L2 (Programmed cell death 1 ligand 2) as most sensitive to variations in donor heterogeneity (CV ≥ 10%). Using individual gene CV as weights, MSC donors are ranked based on calculated desirability scores across 12 CPP combinations with an experimentally determined threshold. Our rankings correlate with clinical data, stratifying MSC donors based on responder and non-responder status in patient-reported outcome measures (PROMs).

Conclusion: Our computational approach correlates MSC fitness, measured via curated gene expression, with clinical efficacy. It identifies MSCs with high desirability scores, aligning with stringent OARSI responder status, validating our approach. Furthermore, it facilitates the evaluation of CPP combinations yielding potent, expanded MSCs, aiding in personalized process optimization.

Abstract 11: A Scalable Approach for the Fabrication of Meter-Long Aligned Collagen Sheets for Load-Bearing Scaffolds

Samuel Lasinski, Lihua Wei, Yuming Zhang, Chantel Campbell, Wuyang Gao, Axel Guenther

University of Toronto, Toronto, Canada

Collagen is integral to the function of load-bearing tissues due to the structural alignment of collagen fibrils, which typically run parallel to the applied load. Despite efforts to replicate native morphology and biomechanics in collagen-based materials, achieving a tensile strength comparable to native tissues remains elusive; this is predominantly attributable to the absence of native-like ultrastructure. Previous work has demonstrated the formation of ultra-thin, aligned collagen sheets capable of recapitulating native-like collagen fiber alignment and tensile properties. However, challenges associated with material handling limit the length of sheet that can be produced, hindering continuous production of materials for large-scale tissue fabrication. This study devises a roll-to-toll (R2R) method for the scalable processing of aligned collagen sheets.

We devised a R2R approach to continuously extrude and collect meter-long aligned collagen sheets. Collagen solution and wet-spinning buffer were co-extruded through a microfluidic printhead into a buffer bath. At a distance from the printhead, the emergent collagen sheet was deposited onto a moving nylon carrier film. The carrier film’s speed was set to 9X the average velocity of collagen. This relative velocity difference strains the collagen sheet before gelation, thereby inducing fiber alignment. The laminated collagen sheet and film were guided over a conveyor belt and finally collected on a rotating drum. 

~3m of continuous sheet were produced – an order of magnitude greater than published methods. The sheets possess ultralow thickness (2.2±0.4 mm) and an ultimate tensile strength (4.1±1.2 MPa) exceeding previous approaches (2.7 MPa). To illustrate the aligned collagen sheet’s capacity as a substrate for composite biomaterials, we deposited a soft collagen layer via temperature-induced gelation on the extruded collagen sheet. 

This technique facilitates the formation of a meter-long collagen sheets with increased tensile properties, underlining the potential of this material as a building block for human-scale, composite load-bearing tissues. 

Abstract 12: Vascularized Heart-on-a-chip for Recapitulating Physiological Complexity

Mengyuan Li, Kayla Soon, Sara Vasconcelos

University of Toronto, Toronto, Canada

Rationale/Objectives: Human pluripotent stem cells (hPSCs) and microfabrication have enabled accurate modeling of cardiac environments using organ-on-a-chip techniques. However, creating a functional cardiac construct with vascular interactions remains challenging. In the native heart, cardiomyocytes (CMs) and endothelial cells (ECs) interact, regulating cardiac functions and vessel growth. These interactions are crucial for cardiac development and fibrosis progression. The goal of my project is to create an in vitro cardiac-vasculature-on-a-chip model for real-time force measurement during cardiac development and fibrotic responses.

Methods: We designed a thermoplastic-based platform where hPSC-derived CMs self-organize into a 3D structure between flexible rods near a central microwell for force measurements. ECs and stromal cells were introduced to self-assembly into a microvascular network, induced by gravity-driven media flow. The co-culture was maintained for 3 weeks. Contractile force, sarcomeric organization, and gap junction localization were assessed. Microvascular network properties, including permeability, vessel density, and diameter, were characterized. For pro-inflammatory studies, ECs were activated by TGF-β and co-cultured with CMs to investigate fibrotic responses.

Results: A lumenized microvasculature was self-assembled by ECs and stromal cells. Selective permeability of the vasculature was tested using fluorescent dextran. Cardiac tissue compaction was achieved. Contractile force and sarcomeric organization are expected to improve with microvasculature co-culture. Activation of ECs is expected to lead to alterations in cardiac contractility and maturation, as well as secrete pro-fibrotic factors and modulate the cardiac microenvironment, promoting the activation of cardiac fibroblasts and the deposition of extracellular matrix components. 

Conclusion/Significance: The vascularized heart-on-a-chip model will capture key hallmarks of cardiac development and microvascular dysfunction, promoting advanced tissue maturity and stable function for up to a month. This study provides a valuable tool for high-fidelity therapeutic screening, aiding in the management of cardiovascular pathology and the development of new functional myocardium.

Abstract 13: Differentiated and Non-Differentiated Adipose-Derived Stromal Cells Accelerate Autologous, Non-Thrombogenic Endothelialization of an Electrospun Scaffold

Kate MacQuarrie¹, Katya D’Costa¹, Xi Lei², Guangheng Zhu², Stefan Hofer³, Heyu Ni², Paul Santerre^1,3

¹Ted Rogers Centre for Heart Research, Toronto, Canada. ²St. Michael’s Hospital, Toronto, Canada. ³University Health Network, Toronto, Canada

Synthetic vascular grafts lose patency at internal diameters <6 mm, often due to thrombosis, as they lack autologous endothelia. Although endothelialized grafts maintain patency longer, the time required to achieve endothelialization limits its clinical use. Previous work from our group has shown that contralateral co-culture of adipose-derived stromal cells (ASCs) and human adipose-derived microvascular endothelial cells (HAMVECs) facilitates complete endothelialization within 48 hours. These ASCs can be differentiated towards a vascular smooth muscle cell (VSMC)-like phenotype (ASC-VSMCs), which allows them to behave like native VSMCs. When ASC-VSMCs are co-seeded with monocytes, more collagen and elastin are produced to mechanically stabilize the graft. Here, we compare the abilities of ASCs, ASC-VSMCs, and ASC-VSMCs with monocytes to support rapid endothelialization, then assess the thrombogenicity of these endothelia.

ASCs and HAMVECs were isolated from human adipose tissue, while monocytes were isolated from human peripheral blood. ASCs were seeded as-is or differentiated to produce ASC-VSMCs. Support cells were cultured contralaterally with HAMVECs across an electrospun polyurethane scaffold. Cell-seeded scaffolds were immunostained to measure endothelialization and proliferation. To assess thrombogenicity, whole blood was applied to endothelialized or non-endothelialized scaffolds, and platelet activation was measured with flow cytometry.

Here, we show that all support cell types produce similar effects: ASCs, ASC-VSMCs, and ASC-VSMCs with monocytes support 70-80% endothelialization within 24 hours. Preliminary results suggest that these endothelialized surfaces induce less platelet activation, evidenced by lower P-selectin expression and PAC-1 binding, than the acellular and Dacron controls. Additional preliminary work suggests that the endothelialized scaffolds have acceptably low (<5%) hemolytic ratios.

We demonstrated that all three support cell types facilitate rapid endothelialization of polyurethane scaffolds. Preliminary results suggest that the endothelialized surfaces may be less thrombogenic than acellular polyurethane or Dacron. This work demonstrates the potential of using autologous cells to produce biomimetic, small-diameter vascular grafts.

Abstract 14: Biomimetic Scaffold Using Graphene Quantum Dots-Hybrid Hydrogel for Diabetic Wound Healing

Siddhartha Pahari¹, Dr Monalisa Mukherjee²

¹Chemical Engineering & Applied Chemistry (ChemE), Toronto, Canada. ²AICCRS Amity Universitya, Noida, India

Introduction: To  address  diabetic  wound  healing  challenges,  we  developed  a  biomimetic  scaffold  for wound healing with minimal inflammation using graphene quantum dot (GQD)-polyacrylic acid (PAA) hybrid hydrogel, leveraging unique properties of GQDs and the versatility of PAA.

Methods: The GQD-PAA hybrid hydrogel was synthesized via aqueous homopolymerization of acrylic acid using APS and TEMED as accelerator and initiator. Comprehensive characterization techniques, cytotoxicity assays, wound healing rates in diabetic rats along with an assessment of inflammatory cytokines were performed to analyse the structure, biocompatibility and wound-healing process.

Results: The  average  pore  size  of  the  hybrid  hydrogel  after  swelling  was  about  120  µm.  The incorporation of GQDs led to a strong interaction with the inherent functional groups of the hydrogel.  It  facilitated  the  preferential  stacking  and  rolling  of  GQD  sheets  during polymerization. Cytotoxic studies demonstrated cell viabilities of over 95%, indicating high biocompatibility and complete wound healing on the 13 day in diabetic wounds treated with the 0.05–0.1% GQD-PAA hybrid hydrogel, indicative of expedited rate of reepithelialization and accelerated wound healing process. Moreover, the expression of IL-10 (around 210 pg/ml) was highest in 0.05% and 0.1% GQD-PAA hydrogel composites, further supporting their role in promoting a favorable wound healing environment.

Conclusion: Incorporating 0-D GQD into PAA hydrogel for a biomimetic scaffold enhances diabetic wound healing  in  rat  models  resulting  in  complete  diabetic  wound  closure  by  day  13, accompanied by favorable pro- and anti-inflammatory  responses. The GQD-PAA hybrid hydrogel  emerges  as  a promising  band-aid  by  promoting  expedited  angiogenesis  and maintaining a moist environment.

Abstract 15: Mechanical loading of osteocytes via oscillatory fluid flow regulates prostate cancer cell extravasation to bone in vitro

Kimberly Seaman¹, Chun-Yu Lin¹, Xin Song¹, Amel Sassi¹, William W. Du², Burton Yang², Yu Sun¹, Lidan You^1,3

¹University of Toronto, Toronto, Canada. ²Sunnybrook Research Institute, Toronto, Canada. ³Queen’s University, Kingston, Canada

Bone metastasis occurs in 80% of advanced-stage prostate cancer patients. Recently, exercise has been shown to attenuate tumour progression in vivo. As the major mechanosensors and regulators of bone, the role of osteocytes under mechanical loading warrants further investigation. Previous in vitro studies have indicated that direct prostate cancer-osteocyte interactions in loading conditions promote cancer cell growth and migration. However, these findings do not correlate with in vivo results, and are more reflective of late-stage colonization. Therefore, the aim of this study is to elucidate the role of osteocytes during the initial stages of bone metastasis, specifically when cancer cells extravasate through the endothelial barrier before colonizing bone. 

Conditioned media and microfluidic approaches were used for this study. Oscillatory fluid flow (OFF) was applied to osteocytes at a frequency of 1Hz and shear stress of 1Pa for two hours to simulate physiologically relevant flow in vivo. Conditioned media collected from static or OFF-stimulated MLO-Y4 osteocyte-like cells were used to assess PC-3 prostate cancer cell adhesion and trans-endothelial migration. A well-established microfluidic tissue model was used to assess PC-3 extravasation towards MLO-Y4 cells and primary osteocytes extracted from 2-month-old male mice through a lumen lined with human umbilical vein endothelial cells (HUVECs). 

Conditioned media assays indicated that mechanical loading of osteocytes reduced PC-3 adhesion to HUVEC monolayers and trans-endothelial migration. Moreover, mechanical loading of MLO-Y4s and primary osteocytes reduced both the extravasation distance and rate of PC-3s in the microfluidic device. Application of a neutralizing vascular cellular adhesion molecule 1 antibody abolished the difference in adhesion and extravasation rate between groups. 

Taken together, these findings will provide more information on the role of mechanical loading of osteocytes during the initial stages of prostate cancer bone metastasis, and aid in the development of osteocyte-targeted cancer mechanotherapies to improve patient outcomes.

Abstract 16: Vibration-Driven Protection Against Radiation-Induced Osteocyte Damage

Xin Song¹, Kimberly Seaman¹, Amel Sassi¹, Chun-Yu Lin¹, Lidan You^1,2

¹University of Toronto, Toronto, Canada. ²Queen’s University, Kingston, Canada

Radiotherapy remains a cornerstone in breast cancer treatment, yet its efficacy is hindered by collateral damage to healthy tissues, particularly bones, due to their heightened radiation absorption. Addressing this challenge, we investigate the therapeutic potential of low-magnitude high-frequency (LMHF) vibration as a non-invasive intervention. Leveraging the bone’s inherent anabolic response to mechanical stimuli, our study focuses on osteocytes, primary mechanosensors in bones whose regulatory functions extend to cancer metastasis modulation. Previous studies have suggested that mechanical loading, may shield osteocytes, fostering recovery from chemical or physical damage. Despite the prevailing focus on chemically-mediated radioprotection strategies such as P7C3 and PTH1-34, the potential of mechanical stimulation in alleviating radiation-induced bone damage remains unexplored. Therefore, our research is to study the effects of radiation on osteocyte viability and regulatory functions, as well as to explore the potential of LMHF vibration in alleviating radiation-induced osteocyte damage.

Our findings reveal that radiation exposure triggers osteocyte apoptosis, whereas LMHF vibration reduces this effect. Irradiated osteocytes retain their mechanosensitivity, as demonstrated by the increased expression of COX-2 following vibration treatment. Furthermore, irradiated osteocytes express more RANKL, thereby promoting the formation of bone-resorbing osteoclasts. This process can be mitigated by vibration, which suppresses the formation of multinucleated giant osteoclasts. By utilizing a microfluidic platform to mimic the bone-cancer microenvironment, we demonstrated that irradiated osteocytes attract breast cancer cells to extravasate more, thus facilitating metastatic spread. Remarkably, LMHF vibration could reduce the cancer extravasation by 18%.

In summary, our study unveils the therapeutic potential of LMHF vibration in mitigating radiation-induced osteocyte damage and its consequential effects on osteoclast formation and cancer metastasis. These findings underscore the promise of non-invasive mechanical interventions in preserving bone health and optimizing cancer treatment outcomes.

Abstract 17: Identification of congenital aortic valve malformations in juvenile natriuretic peptide receptor 2 deficient (Npr2+/-) mice using high frequency ultrasound

Vrushali Guruji^1,2, Yu-Qing Zhou^1,2, Craig Simmons^1,2

¹University of Toronto, Toronto, Canada. ²Ted Rogers Centre for Heart Research, Toronto, Canada

Bicuspid aortic valve (BAV) disease is characterized by two functioning leaflets instead of the normal three in a tricuspid aortic valve (TAV). Most BAV patients develop ascending aortic dilation (AAD), increasing their risk of fatal aortic dissection or rupture. We hypothesize that the elevated risk of developing AAD in BAV disease is a consequence of aberrant hemodynamics inducing transcriptional and proteomic perturbations in the aorta. To facilitate this investigation, a mouse model with a mutation in the natriuretic peptide 2 receptor (Npr2) was employed, enabling the isolation and evaluation of consequences accompanying aberrant blood flow. Notably, though genetically identical, 10% of Npr2+/-mice develop BAVs, whereas the remaining 90% have TAVs. To expedite the progression of BAV-AAD, mice are fed a high-fat western diet starting at 6 weeks of age. A novel echocardiogram protocol was developed to diagnose mice with BAVs or TAVs at 4 weeks of age, prior to diet introduction. Color Doppler identified BAV-associated regurgitation and recirculation, pulsed-wave Doppler detected peak systolic velocities ≥1250 mm/sec, and electrocardiogram-gated kilohertz visualization recognized closing patterns along commissures that identified valvular abnormalities. Gross anatomy images of echo-assessed mice indicated hemodynamic dysfunction correlated to thickened valves, partially fused BAVs, or Sievers type 0 BAVs. Movat’s pentachrome staining on coronal sections of the aortic valves confirmed that abnormal valves had increased collagen content, cellularity, and proteoglycan content relative to the normal TAVs. Collectively, this study establishes a robust protocol to classify and characterize valve morphologies in 4-week old mice. This is an essential first step to studying associations between congenital valve abnormalities and acquired valvulopathies and aortopathies, such as the role of hemodynamics in BAV-associated AAD. The ultrasound protocol can also be adapted by other investigators for investigation of other congenital cardiovascular diseases in small animal models. 

Abstract 18: Investigating the interplay between gut microbiome, monocytes/macrophages, and osteoarthritis

Atoosa Ziyaeyan^1,2,3, Mozhgan Rasti^1,2, Aida Feiz Barazandeh^1,2, Sowmya Viswanathan^1,2,3

¹Shroeder Arthritis Institute, Toronto, Canada. ²Krembil Research Institute, Toronto, Canada. ³University of Toronto, Toronto, Canada

Osteoarthritis (OA) is a complex degenerative joint disease influenced by various factors. This research aims to unravel the complex connection between metabolites produced by intestinal commensal bacteria and their impact on local immune effectors within the joint. The synovial joint is not a sterile environment, as shown by elevated levels of bacterial metabolites, such as lipopolysaccharide (LPS), correlating with OA severity. We hypothesize that gut metabolites influence local and systemic immune effectors, specifically monocytes/macrophages (MΦ), contributing to OA pathogenesis. To test this hypothesis, we are measuring the correlation of levels of short-chain fatty acids (SCFAs) which are bacterial metabolites, LPS, LPS-binding protein (LBP) with i) patient-reported outcomes (PROMs) and ii) with frequencies of classical (CD14+CD16low), intermediate (CD14+CD16+) and non-classical (CD14lowCD16+) MΦ subsets.

LBP levels are measured by an ELISA assay, and a Kinetic Chromogenic LAL Assay measures LPS levels. The level of SCFAs is measured using liquid chromatography with tandem mass spectrometry (LC-MS/MS). The MΦ populations in the synovial fluid are investigated by flow cytometry. 

With an N=78, there was a significant negative correlation between the SF LBP and the frequency of intermediate MΦ in SF, indicating that patients with higher levels of LBP in their SF had lower frequencies of intermediate MΦ in their SF. Of the 78 patients in our study, 43 were females, and 35 were males. There was a significant negative correlation between LBP levels in SF and SF intermediate MΦ frequency in males and females.

In conclusion, the observed correlations in the patients suggest that gut metabolites, specifically LBP, have systemic effects on OA severity. Overall, our study helps us understand the complex relationship between gut microbiota and OA, allowing us to improve patient outcomes and develop innovative interventions for OA.

Molecular Engineering

Abstract 19: Using Digital Microfluidic Isolation of Single Cells to Understand Microglia Heterogeneity

Savina Cammalleri^1,2, Erica Scott^2,3, Aaron Wheeler^1,2,3

¹Institute of Biomaterials and Biomedical Engineering, Toronto, Canada. ²Donnelly Centre for Cellular and Biomolecular Research, Toronto, Canada. ³Department of Chemistry, Toronto, Canada

Microglia are yolk-sac derived immune cells of the brain meaning they are the only innate immune cell permanently residing in the central nervous system (CNS). The role of microglia are broad encompassing phagocytosis, production of cytokines, synaptic plasticity, neuronal monitoring, and overall maintaining homeostasis. To accomplish this wide array of tasks, microglia must be capable to both, sense diverse stimuli and respond to these stimuli by altering morphology thus changing their cell state. While some microglial states have been characterised to be involved in specific function such as the role of ramified microglial as cell surveyors to work in screening the CNS other microglial morphologies are not understood.

Recent work uses fluorescence activated cell sorting (FACS) to obtain microglia for single cell -omics resulting in loss of spatial information. While other techniques can obtain single cell spatial information, the data is limited to transcriptomics. Using digital microfluidic isolation of single cells will allow for single cell spatial information and multimodal -omics data resulting in a rich layered catalog per cell, which will be especially important to fully understand the highly heterogenous population of microglia within the brain. 

Our work to date has developed methods to allow for cell culturing and immunohistochemistry staining on digital microfluidic devices thus allowing for a streamlined approach to automate fluid handling and delivery of reagents for culture and staining. The technique aims to minimize human interaction with the sample to reduce variability, improve assay reproducibility, and requires less time than its manual counterpart. Using these stained samples, we perform DISCO to isolate single microglia cells with diverse morphologies to perform single cell transcriptomics and proteomics. The DISCO isolation and -omics data will provide a highly rich dataset between morphologically distinct microglia. 

Abstract 20: A novel lateral flow assay (LFA) to detect a urinary acute kidney injury biomarker for pediatric patient monitoring

Kevin Da^1,2,3, Ryan Li^1,4, Craig Simmons^1,2,3,4, Xinyu Liu^1,2,4

¹University of Toronto, Toronto, Canada. ²Institute of Biomedical Engineering, Toronto, Canada. ³Translational Biology and Engineering Program, Toronto, Canada. ⁴Department of Mechanical Engineering, Toronto, Canada

Rationale & Objectives: Acute kidney injuries (AKIs) are a clinical syndrome where an abrupt loss in renal function leads to fibrosis and increased mortality if not properly monitored. Current monitoring standards are delayed and imprecise, leading to AKI underrecognition particularly in pediatric individuals. Monitoring urinary biomarkers may help improve AKI management by providing higher disease specificity. Lateral flow assays (LFAs) provide an inexpensive point of care (POC) platform satisfying clinical needs in AKI monitoring. The objective of this project is to develop a novel LFA for the clinically relevant detection of a urinary AKI biomarker to enable improved patient monitoring and outcomes. 

Methods: 40 nm carboxylic activated gold nanospheres were conjugated with capture antibodies using 1-ethyl-3-(3-dimethylaminopropyl)carbodiimide hydrochloride/N-Hydroxysuccinimide (EDC/NHS) chemistry, validated with UV-VIS spectroscopy and dynamic light scattering (DLS). LFAs were then manually assembled and tested across a clinically relevant concentration spiked in artificial urine. 

Results: Successful synthesis of antibody conjugated gold nanospheres was confirmed using UV-VIS spectroscopy by a right plasmonic absorbance shift. DLS demonstrated an increase in hydrodynamic diameter (p < 0.0001) following antibody conjugation as a monodisperse population across all synthesized probes validating conjugation synthesis. Conjugate formulations can be stable in solution up to 2 months with the lowest antibody loadings showing minimal aggregation. Hand-fabricated preliminary LFAs demonstrate a captured detection range down to 25 ng/mL which can be detected within a POC setting (15 minutes). 

Conclusions & Significance: A preliminary LFA developed for urinary AKI biomarker detection in artificial urine is demonstrated across a clinically relevant range that helps inform patient decisions with performance like commercial POC tests. This project demonstrates the potential to supplement AKI management through assessing urinary biomarkers using a cost-effective and easily translatable LFA platform to improve patient outcomes and quality of life.

Abstract 21: Spatially defined and decorated DNA origami to investigate immune cell Fc-gamma receptor biology

Travis Douglas, Dr. Leo Chou

BME, Toronto, Canada

The immune system uses various mechanisms to fight off foreign pathogens. Antibodies play a key role in neutralizing the activities of these intruders. When antibodies bind antigenic epitopes of pathogens, an immune complex (IC) is formed. IC formation leaves the fragment crystallizable (Fc) region of antigen-bound antibodies available to interact with cell-surface Fc-gamma receptors (FcγRs) to activate immune cells. These receptors are expressed on a variety of innate immune cells like monocytes, macrophages, dendritic cells, and neutrophils to facilitate the internalization of ICs. FcγR-mediated internalization of the ICs does not solely serve to eliminate the pathogen, as ingeseted contents engage FcγRs to modulate signaling pathways that control immune cell polarization and effector functions including antigen presentation for priming adaptive immunity.

However, the extent to which the physical properties of an IC contribute to governing immune cell signaling remains poorly understood in some contexts. This is due to the complexity of FcγRs and antibody subtypes, as well as challenges in generating ICs with defined strucutral characteristics. Here, we leverage DNA nanotechnology to engineer synthetic immune complexes with precisely controlled parameters to tune the physical characteristics of the ICs.

Our findings demonstrate the ability to manipulate IC structure using previously engineered DNA nanostructures, providing a platform to investigate FcγR-mediated responses to synthetic ICs. This research holds promise for elucidating the structural properties that influence FcγR signaling and their contribution to downstream immune activation across a variety of antibody subtypes and cell types. Moreover, it may have implications for the development of immunotherapies and vaccines by offering insights into the design of more effective and targeted IC therapeutics to overcome limitations associated with current approaches reliant on laborious engineering methods and uncontrollable IC structures generated from admixing antibodies and antigens.

Abstract 22: Antibody Synthesis in Cell-Free for High-Throughput Production and Screening

Sabina Panfilov, Masoud Norouzi, Keith Pardee

University of Toronto, Toronto, Canada

Rationale: Antibody production in cell-based platforms is known for its time-consuming and costly processes. Cell-free protein synthesis (CFPS) systems offer an alternative for expressing immunoglobulins and antibody fragments. However, successfully reproducing E.Coil’s periplasmic environment is crucial for antibody production, requiring the optimization of additives for each antibody and making the process time-consuming. This project aims to develop a platform for expedited prototyping and screening of antibodies utilizing an E.coli lysate based CFPS system. This kind of platform can be further used to expedite the screening process of antibody libraries.

Methods: The proposed platform integrates our in-house CFPS reactions with the replication-termination Tus-Ter system for linear DNA protection and disulfide-bond supporting additives like glutathione, DsbC, PDI, and Erv1p. CFPS reactions, arranged in a multi-well format, are followed by immediate Western blot and ELISA assays to identify properly folded antibodies with target binding affinity and high yield. Conjugation experiments employ the S.Pyogenes SpyTag/SpyCatcher system, demonstrating immunofluorescence of HER2-positive cells.

Results: Preliminary experiments identified necessary additives for disulfide bond formation, and selected conditions were applied to express 38 antibodies across various modalities and targets. The choice of lysate was found to significantly impact successful antibody expression. Thirty initial conditions were identified as a promising starting point for every antibody within the high-throughput platform. Testing focused on three anti-HER2 antibodies: scFv (small), scFab (medium), and scFv-Fc (large), with further anti-TNFα antibody library screen.
Additionally, the platform exhibited versatility by successfully conjugating a fluorescent payload (YFP) to CFPS-produced antibodies using the SpyTag/SpyCatcher system, validated through an immunostaining assay with HER2-positive cells.

Conclusion: While cell-free systems have proven successful for antibody production, the lack of a universal expression system necessitates a high-throughput platform. This platform streamlines the identification of optimal conditions for each antibody, facilitating scalable production and future expansion for payload conjugation and library screening.

Abstract 23: Engineering a Model to Simulate Intestinal Inflammation for Biomaterials Testing

Lu (Kelly) Yin, Caitlin L. Maikawa

University of Toronto, Toronto, Canada

Inflammatory Bowel Disease (IBD) affects over 1 in 130 Canadians and has been experiencing a rising global incidence. Biomaterials are a promising strategy to enable targeted delivery of IBD therapeutics to inflamed intestinal tissue for improved efficacy and reduced off-target effects. There is potential to develop new materials that exploit the differences in electrostatic charge between healthy and inflamed tissue for localized drug delivery to inflamed sites. However, successful screening of inflammation-targeting materials is heavily dependent on in vitro testing that mimics the intestinal environment. Current in vitro testing methods fall short of mimicking the dynamic and complex conditions of the intestinal flow in the intestines. Thus, our project goal is to develop a flow-based model that mimics intestinal conditions. This model will enable effective screening of potential inflammation-targeting materials by offering an understanding of (1) the proportion of delivered material that adheres to simulated inflamed tissue and (2) the residence time of materials following delivery while under shear stress. Our design uses a mucin-gelatin hydrogel (anionic) as a healthy tissue mimic and a transferrin-gelatin hydrogel (cationic) to simulate inflamed tissue. These hydrogels are patterned onto the inner wall of a tube and then charged materials (ex. microgels) are flowed through to observe adhesion. Preliminary results demonstrated that cationic chitosan microgels adhere to anionic regions. Future work aims to quantify microgels effectively adhered to the anionic and cationic coating regions using fluorescent microscopy and image analysis, as well as the residence time of microgels following shear stress as fluids with different viscosities are flowed through the model. Our in vitro screening tool has the potential to facilitate the development of improved materials for targeted intestinal drug delivery with enhanced retention for more effective IBD treatments.

Clinical Engineering

Abstract 24: A Multi-Modal Physiological Monitoring System Utilizing Multi-Wavelength Photoplethysmography and Bioimpedance for Advanced Hemodynamic Classification and Regression

Rawad Alkallas¹, Matthew Lee¹, Samantha Unger¹, Farida Abdelmalek^1,2, Daniel Franklin¹

¹University of Toronto, Toronto, Canada. ²University of Waterloo, Waterloo, Canada

Cardiovascular disease remains a leading cause of mortality worldwide, with blood pressure regulation being a critical aspect of its management. Continuous, non-invasive blood pressure (cNiBP) devices allow for patients and clinicians to monitor the disease remotely and in between clinical visits. However, blood pressure readings alone do not provide a comprehensive picture of cardiovascular status, as in cases of compensated heart failure where blood pressure remains within healthy ranges. This research project presents a multimodal system that monitors physiological metrics correlated with blood pressure and its regulatory mechanisms, such as vascular resistance and cardiac output. The system employs bio-impedance spectroscopy to quantify cardiac output. A multi-wavelength photoplethysmography system estimates systemic vascular resistance. Combined with ECG, the system can also estimate blood pressure through pulse arrival time measurements. Monitoring this specific set of parameters, the system provides real-time quantification of the cardiovascular status, improving the classification of hemodynamic in various pressor scenarios. Our system provides a practical and more comprehensive tool for monitoring the cardiovascular system, which will improve prevention, early detection, and management of cardiovascular disease. Once deployed remotely, the system will provide clinically actionable insights into the progression of cardiovascular diseases and enable personalized treatment plans.

Abstract 25: Gamified Assessment: Extracting and characterizing reaching performance in children with cerebral palsy using data collected during a movement tracking video game, Bootle Blast

Sorsha Asady^1,2, Elaine Biddiss^1,2,3, Soowan Choi^1,2, Ajmal Khan¹

¹Bloorview Research Institute, Toronto, Canada. ²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³Institute of Rehabilitation Sciences, University of Toronto, Toronto, Canada

Cerebral palsy (CP) is the leading cause of childhood motor disability. Hemiplegic CP, which affects movement on one side of the body, can impact upper limb (UL) force production, coordinated movement and the precision of hand transportation toward a target. This may compromise reaching performance and how individuals with CP engage with their environment. Bootle Blast (BB) is a therapy game in which children can practice motor skills associated with reaching. BB uses a low cost 3D camera, the Orbbec Persee+ (OP+) to track 19 body points during game play. UL tracking with the OP+ has shown good concurrent validity relative to gold standard motion capture.

Our overall goal is to understand if data collected by the OP+ during BB game play can be used to characterize reaching performance in children with CP. Specific objectives include: 1) to compare movement metrics associated with reaching performance between the OP+ and gold standard motion capture and 2) to compare movement metrics associated with reach performance between typically developing children and those with CP.

Two previously collected datasets will be used: an in lab dataset (20 adults/8 children with and without CP) and an at-home dataset (13 children with CP). In-game reaching movements will be segmented and the following movement metrics calculated: movement time, peak hand velocity (PHV), percentage time to PHV, smoothness, maximum extent of hand reach and symmetry.

Preliminary results show that a child with CP required longer to complete a reach task and displayed a greater number of movement units (19 vs. 8) than a typically developing child, suggesting less smooth movements. This may imply that differences in reaching performance are discernible from data collected during BB playtime.

Future analyses will use larger datasets and look at potential changes in reaching performance over an 8 week intervention.

Abstract 26: The impact of chronic opioids therapy on sleep efficiency: A sex-specific comparative analysis

Atousa Assadi^1,2, Muammar Kabir², Frances Chung^1,3, Clodagh Ryan^2,1, Azadeh Yadollahi^2,1

¹University of Toronto, Toronto, Canada. ²KITE-Toronto Rehabilitation Institute, Toronto, Canada. ³University Health Network, Toronto, Canada

Rationale: Chronic opioids use, for managing moderate-to-sever pain, is associated with decreased sleep efficiency by increasing time in bed and reducing time spent sleep. However, the sex-differences in sleep efficiency is poorly understood.

Methods: We performed retrospective analysis on sleep efficiencies of 174 individuals (18+ years, 103 women) on chronic opioids therapy for >3 months with a stable daily dosages >4 weeks. Sleep efficiency was defined as total sleep time/total time spent in bed. Age, body mass index (BMI), and daily opioids dosages were categorized using cut-offs of 50 years, 30 kg/m², 50 milligram morphine equivalent (MME), respectively. To examine the effect of sex and its interaction with various factors, we employed multivariable ANOVA analysis, including main effects (age, sex, BMI, opioids dosages) and 2-factor to 4-factor interactions terms involving sex. Posthoc analysis included Wilcoxon test, given non-normality of data, to further investigate the sex-differences of sleep efficiency.

Results: Participants’ age was 51.9±13.0 years with corresponding BMI:29.2±6.5kg/m², opioids dosage:412.1±3278.7MME, and sleep efficiency:79.6±15.7. Age, BMI, and opioids dosages were not significantly different between sexes. In total, women had significantly higher sleep efficiency than men (81.9±14.8 vs. 76.3±16.5, p=0.010). The ANOVA model revealed significant main effects for sex (p=0.023) and opioids dosages (p=0.047) with notable interaction effect for sex:BMI:opioids dosages (p=0.022). Sleep efficiency was significantly higher in women than men in individuals with BMI≥30kg/m² and opioids dosages<50MME (84.5±12.7 vs. 76.0±12.7, p=0.048) and those with BMI<30kg/m² and opioids dosages≥50MME (86.6±9.8 vs. 77.2±14.7, p=0.008). In women with BMI<30kg/m², sleep efficiency was significantly higher with opioids dosage≥50MME than <50MME (86.6±9.8 vs. 73.3±20.4, p=0.012).

Conclusion: Women on chronic opioids therapy had significantly higher sleep efficiency than men, particularly influenced by BMI and opioid dosage.

Abstract 27: Patient Reported Outcome Measures In The Elderly-Do These Reflect Healing Post FFP?

Samantha Bartman^1,2, Diane Nam^1,2, Cari Whyne^1,2

¹Sunnybrook Research Institute, Toronto, Canada. ²University of Toronto, Toronto, Canada

Purpose: With an aging population and increasing rates of osteoporosis, fragility fractures of the pelvis (FFPs) have become increasingly common, accounting for ~7% of all fragility fractures. This study aimed to document functional outcomes and quality of life (QoL) measures in individuals with non-operatively treated FFPs and determine if a relationship exists between these parameters and healing status.

Methods: This was a case series of a limited cohort prospectively collected at a single institution. Of the 53 patients (age ≥ 65 years) who sustained a non-operative FFP as diagnosed on x-ray, follow up x-ray imaging was only available for a subset (N=35). Musculoskeletal Function Assessment (MFA) and 36-Item Short Form Health Survey (SF-36) scores were collected over a 2-year period to assess function and QoL. Healing status was assessed on available follow up imaging by 4 orthopaedic surgeons and categorized as either less than 50% healed, 50-75% healed, or >75% healed.

Results: Patient MFA scores increased from baseline for all categories except ‘Fine Motor’ and ‘Employment’, indicating worsening function. Additionally, SF-36 scores decreased from baseline with some variability in the ‘Role Limitation due to Physical Problems’ category, indicating deteriorating QoL. Confirmation of healing on x-ray (>75% healed) was found for only 7 individuals. No relationship was found between any MFA or SF-36 category and healing status for patients with coincident reporting (N=16). Despite the limited number of follow up x-rays, this cohort highlights the limited healing that may occur in non-operatively treated FFPs. Lack of follow up imaging even without documented healing may occur as no treatment exists to address the lack of boney fusion. 

Conclusion: Function and QoL are not improving in elderly patients with non-operatively treated FFPs from baseline, with many categories demonstrating a steady decline out to 24 months. Furthermore, no clear relationship exists between these measures and healing status.

Abstract 28: Improving Tele-Rehabilitation with Automated Exercise Assessment

Naomi Opia-Evans¹, Nikkole Chow¹, Raymond Hawkins¹, Pedram Karimi¹, Ali Barzegar Khanghah^1,2, Atena Roshan Fekr^1,3,2

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ²KITE Research Institute, Toronto Rehabilitation Institute, University Health Network, Toronto, Canada. ³Rehabilitation Science Institute, University of Toronto, Toronto, Canada

Rationale & Objectives: By eliminating the need for patients to travel to rehabilitation facilities or pay outrageous costs for in-home services, tele-rehabilitation (tele-rehab) has the potential to fundamentally change the remote patient care and rehabilitation practices. Despite the benefits of video conferencing consultations with a doctor, they can also be costly and have limited-service availability. An automated tele-rehab platform that can provide accurate feedback to the patients is required. Creating such a platform would increase patient adherence, facilitate access to rehabilitation services, and eventually result in improved patient outcomes. The objective of this study is to develop and validate an automated exercise assessment algorithm that can provide accurate feedback to the patient.

Methods: This study used an available dataset consisting of 16 patients and 14 healthy individuals performing 5 physical exercises considering integrating data from left and right sides. We used a SpatioTemporal Graph Convolutional Network (STGCN) to distinguish between correct and incorrect exercises based on skeleton joint position data. The model merges STGCN and Long layers for binary classification. STGCN extracts spatiotemporal features from joint positions, while Long Short Term Memory (LSTM) captures temporal patterns. Hyperparameter tuning is conducted, and models are trained using class weights to address the imbalance of the dataset. Five binary models are trained and validated using 10-fold and Leave-One-Subject-Out (LOSO) techniques. 

Results: Incorporating class weights significantly enhanced performance in identifying incorrectly executed exercises. Our 5 classifiers achieved an average accuracy of 96.60% ± 0.90%, a micro F1-Score of 88.34% with 10-Fold, and an average accuracy of 75.09 %± 28.56% with LOSO validation. We also demonstrated that the network could operate in real-time without compromising classification performance. 

Conclusions & Significance: The proposed system has the potential to integrate automated feedback into tele-rehab platforms, helping patients consistently perform exercises and facilitate the recovery process.

Abstract 29: Generative Platform for Diabetic Foot Ulcer Electronic Health Reporting and Image Synthesis

Reza Basiri^1,2, Shlok Desai¹, Milos R. Popovic^1,2, Shehroz S. Khan^1,2

¹University of Toronto, Toronto, Canada. ²KITE Research Institute, Toronto Rehabilitation Institute, Toronto, Canada

Rationale & Objectives: Diabetic Foot Ulcer (DFU) patients’ clinical reports often lack key wound descriptors, hindering communication between specialist groups and at-home nurses. Moreover, the scarcity of diverse DFU imagery limits research and healthcare training. WoundVista2.0 is an in-house platform developed by a DFU-based Large Language Model (LLM). WoundVista2.0 aims to enable swift and consistent image-to-text transcription of wound conditions and text-to-image generation for DFU images.

Methods: We used the Zivot DFU dataset, collected in collaboration with our partners in Calgary, to train the deep learning vision model and LLM. The Zivot dataset consisted of 3,700 RGB, thermal, and depth images from 270 unique patients and wound text descriptors. A pre-trained stable diffusion baseline model with progressive distillation was employed for image generation. The Segment Anything vision transformer model, pre-trained on 1 billion generic images and fine-tuned using DFU wound bounding boxes, was utilized for image segmentation. Vision language models (e.g., CLIP, Kosmos-2) will be experimented with for text integration.

Results: The unconditional diffusion model generated low-resolution DFU images. The CLIP and diffusion models generated higher-resolution DFU images from the prompt “DFU foot condition,”. The Segment Anything model successfully segmented and tokenized areas important for creating DFU reports, such as “DFU,” and “calluses.”

Conclusions & Significance: WoundVista2.0 demonstrates the potential for generating consistent DFU image descriptions and diverse synthetic DFU images. This platform can be integrated into existing Electronic Medical Record systems and serve as an educational tool for training clinicians. By improving communication and expanding the available DFU imagery, WoundVista2.0 has the potential to enhance DFU care and treatment, as well as facilitate research in wound analytics and deep learning.

Abstract 30: Does Human Pose Estimation benefit from Depth Data?

Gloria-Edith Boudreault-Morales^1,2, José Zariffa^1,2,3,4

¹KITE, Toronto Rehabilitation Institute, University Health Network, Toronto, Canada. ²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, Canada. ⁴Rehabilitation Sciences Institute, University of Toronto, Toronto, Canada

Rationale: The rise of new neuromodulation therapies has created a need to track motor performance to evaluate and personalize new approaches. Human Pose Estimation (HPE) neural networks can be used for this purpose. Most current published approaches use color (RGB) data as an input. Advancements in technology have made depth (RGB-D) data more accessible. Incorporating depth data into HPE models should be advantageous, but this has not yet been fully demonstrated.

We wish to develop a self-contained system that can collect data and perform HPE in a clinical rehabilitation scenario. The main aim is to determine how depth data affects a lightweight monocular RGB HPE model’s performance (accuracy and speed).

Methods: 1) Select a RGB model and a RGB-D dataset from the literature. 2) Train the model with the dataset’s RGB data. 3) Modify the model’s architecture to include depth as an input. 4) Train the modified model using the dataset’s RGB-D data. 5) Compare model performance.

Results: The Dite-HRNet model was chosen because it is the most accurate lightweight RGB option. The CMU Panoptic dataset was selected due to its camera viewpoints variety. The RGB model achieved an average precision of 0.921 and a speed of 27.6 ± 0.1 milliseconds/image. The RGB-D model achieved an average precision of 0.931 and a speed of 28.3 ± 0.3 milliseconds/image.

Conclusion: Adding depth as an extra channel improved the accuracy without significantly affecting the model’s speed (p = 0.1). Next steps include investigating the effect of data fusion techniques/model architectures to better demonstrate the effect of depth on HPE. Implementing HPE models into rehabilitation processes could lead to more precise ongoing motor assessment and support personalized interventions. It could also provide a scalable approach to capturing data about recovery trajectories and play a key role in improving the evidence base for interventions.

Abstract 31: Optimizing spatial frequency domain imaging for defocus-invariance towards robust remote patient monitoring

Dylan Dao, Jie Jiao, Lindsay Kuramoto, Ofer Levi

University of Toronto, Toronto, Canada

Rationale & Objectives: Real-time monitoring technologies for tissue oxygenation are desired to track the progression of cardiovascular disease. Spatial Frequency Domain Imaging (SFDI) is an emerging technique for wide-field, non-invasive measurement of tissue optical properties, from which tissue oxygenation images can be estimated based on how much the tissue blurs incident projection patterns due to optical absorption/scattering processes. Since SFDI is a projector-camera technique, it is particularly sensitive to defocus errors due to contributions from both projection and imaging paths. We quantify and model SFDI error due to mid-range defocus and propose defocus-invariant computational imaging techniques to correct these errors.

Methods: To characterize optical property measurement performance, a mid-range SFDI system is used to measure the optical properties of reference tissue-mimicking optical phantoms at best focus. Then, to directly evaluate the optical system, image sharpness is evaluated in various defocus conditions using test targets on 1) the imaging path, 2) the illumination path, and 3) the full SFDI system. Full imaging system defocus blur is simulated in lens design raytracing software, compared to experimental results, and computationally removed to restore image sharpness.

Results: The baseline SFDI system was found to measure tissue optical properties with 5-10% error under best focus conditions. At 20cm off best focus, optical system imaging sharpness was found to degrade up to 40% with typical projection patterns, resulting in significant optical property measurement error. Defocus simulations agreed with experimental observations on each path independently as well as combined.  

Conclusion: We demonstrate that the optical pathways in SFDI can be modelled independently, and that optical property measurement in defocus can be computationally recovered provided a priori knowledge of imaging distance. Towards a defocus-invariant SFDI system that does not require knowledge of imaging distance, further investigations will attempt applying defocus-invariant optical elements on each optical pathway.

Abstract 32: Correlation Of Sleep And Mobility Sensor Data With Social Isolation And Functional Decline In Older Adults After Lower Limb Fractures

Faranak Dayyani^1,2, Charlene Chu^2,3, Ali Abedi², Shehroz Khan^2,1

¹Institute of Biomaterials and Biomedical Engineering, University of Toronto, Toronto, Canada. ²KITE – Toronto Rehabilitation Institute – University Health Network, Toronto, Canada. ³Lawrence S. Bloomberg Faculty of Nursing, Toronto, Canada

Older adults (OAs) post lower limb fractures experience social isolation (SI) and functional decline (FD) once discharged from inpatient rehabilitation. This population experience SI and FD due to reduced physical mobility, changes in sleeping patterns and lack of motivation. Our team has developed MAISON (Multimodal AI-based Sensor platform for Older iNdividuals), a multimodal sensor system supporting the collection of data from various smart devices. This study aims to assess correlations of SI and FD with sensor data in OAs after lower limb fractures.

Data were collected from seven OAs with lower limb fractures after their inpatient rehabilitation, monitored for an 8-week period at home. During this time, five clinical metrics were obtained via video chat biweekly which included three questionnaires (Social Isolation Scale (SIS), Oxford Hip and Knee Score), and two physical tests (Timed Up and Go, 30 second chair stand). Moreover, MAISON collects various modalities of data such as raw acceleration, heartrate, step count, frequency of indoor motion, GPS and sleep metrics.

From the sensor data collected, 107 statistical and domain-specific features were extracted. Using Spearman coefficient, correlation analysis was performed between the extracted features and the clinical data. SIS and its two subcategories of Connectedness and Belonging, had moderate correlations (r=0.5,p<0.05) with sleep metrics. The physical tests and the Oxford Knee Score had correlations with GPS data (i.e., distance travelled from home (r=-0.4, p<0.05), time spent outside (r=0.5, p<0.05)).

The correlations found between SIS and sleep metrics were consistent with previous studies, indicating a direct positive correlation between SI and sleep quality in OAs. Oxford Knee score and the physical tests represent the degree of FD among this population. Preliminary results also show the usefulness of data from other sensors, and this information will be used to develop AI models to predict the recovery trajectory of these patients.

Abstract 33: Exploring Pain Phenotypes in Osteoarthritis using Ultrasound: Preoperative Assessment of Total Joint Arthroplasty Patients

Joana Dilipkumar^1,2, David Koivisto^1,2, Josh Downer^1,2, Daniel West^2,1, Dinesh Kumbhare^2,1

¹University of Toronto, Toronto, Canada. ²University Health Network, Toronto, Canada

Osteoarthritis (OA) is a disease in which the joint breaks down, causing pain. Total joint arthroplasty (TJA) is a first-line therapy for OA with 130,000 hip and knee joint replacement procedures performed each year in Canada. One of the strongest predictors of patient dissatisfaction is persistent pain; however, the factors underpinning persistent pain are unclear, and it is difficult to predict who will experience a poor versus good outcome after TJA. Identifying the pain phenotype would allow for targeted treatment of these subgroups postoperatively and suggests the possibility for improved patient outcomes. Thus, this study aimed to explore differentiating between nociceptive and mixed pain phenotypes that are heavily represented in the OA population. Twenty patients scheduled for TJA were recruited (n = 6 classified as mixed, n = 14 classified as nociceptive) and assessed using various clinical surveys (CSI, PainDETECT, & WOMAC). Survey scores were associated with different types of pain: nociceptive, neuropathic and nociplastic. Patients exhibiting minimal neuropathic and nociplastic pain were considered the nociceptive pain phenotype, while a greater distribution of pain types indicated the mixed pain phenotype. Ultrasound images of the vastus medialis were acquired and texture features were extracted from individual frames along the region of the muscle. A support vector machine was used to classify the images into the two phenotypic groups. Performance was evaluated using F1-score due to the imbalanced dataset. The F1-score was 0.507 ± 0.084 and suggests no discriminability was possible between the two groups using ultrasound texture features alone. These results can be attributed to similarities within the phenotype groups. Patients categorized as a mixed pain phenotype may still exhibit dominating nociceptive pain scores. To improve the separation accuracy of the model, additional biomarkers or imaging modalities should be investigated.

Abstract 34: Wearable, smart textile sensors for centre of pressure estimation using machine learning for a standing neuroprosthesis

Stephanie DiNunzio^1,2, Hani E. Naguib^1,3, Milos R. Popovic^1,2

¹KITE – Toronto Rehabilitation Institute, University Health Network, Toronto, Canada. ²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³Department of Mechanical & Industrial Engineering, University of Toronto, Toronto, Canada

Functional Electrical Stimulation (FES) therapy has proven to be an effective rehabilitation technique for retraining standing and balancing in individuals with Spinal Cord Injury (SCI). Several FES systems with closed-loop control have been developed for balance control, where the Centre of Pressure (COP) is often used as a quantifiable measurement for balance as feedback to the system. The equipment typically used to measure these quantities, such as force plates or motion capture systems, are expensive and cumbersome. This confines the usage of these systems to lab settings, which requires frequent travel to receive therapy, which is not ideal for SCI patients with limited mobility. Recently, Inertial Measurement Units (IMUs) have been demonstrated to provide a wearable, low-cost alternative to estimate the COP. This study presents smart textile sensors that can be used as an input to a closed-loop FES system for standing and balancing therapy that is portable, reliable, and accessible so that individuals with SCI can benefit from FES therapy in the comfort of their own home. A system of IMU sensors embedded into a textile that can be worn as a garment for the lower limbs and trunk has been designed. Using the IMU measurements from one able-bodied participant with IMUs placed on the thigh and sacrum, an artificial neural network was trained and implemented to estimate the COP and compared against measurements from a force plate.  Good Pearson’s correlation in both the medial-lateral (ML) and anterior-posterior (AP) directions in quiet standing (ML: 0.54, AP: 0.56) and dynamic standing (ML: 0.97, AP: 0.92) is reported. The result is a basis for which a neuroprosthesis for standing can be developed.

Abstract 35: Active Tunable Integrated Silicon Photonic Refractive Index Sensors for Biosensing Applications

Matthew Downing, Sarvath Sharma, Ofer Levi

University of Toronto, Toronto, Canada

Rationale & Objectives: Many in-vivo imaging and sensing applications are ultrasound-based due to the modality’s high spatiotemporal resolution, general non-invasiveness, and lack of ionizing radiation. Our previous work demonstrated silicon photonic (SiP) ultrasound sensors designed around passive integrated resonator-based refractive index sensors such as 2D photonic crystal slabs. The resonance spectral locations of passive sensors often shift due to fabrication errors and ambient temperature variations, necessitating the use of large expensive tunable lasers for readout. We explore the first of two approaches towards the realization of a more compact and robust sensing system: 1) Active tunable integrated sensors that can adjust resonance locations post-fabrication. 2) Miniature integrated tunable lasers.

Methods: Lumerical finite-difference time domain and frequency domain simulation tools were used to design and optimize the geometry of microscale active tunable Silicon-on-Insulator (SOI) racetrack resonators for compatibility with commercial SiP fabrication processes. Several sensor configurations were fabricated through an external foundry then characterized in-house using a custom grating-coupler-based chip interrogation setup. A tunable near-infrared light source was used to determine sensor resonance locations and loaded quality factor (Q_L) values. Bulk sensitivity (S_bulk) values were determined by measuring the relative wavelength shift induced with different refractive-index-matching gels.

Results: Fabricated sensors demonstrated Q_L and S_bulk values within around ~10⁴ — ~10⁵ and ~10¹ — ~10² [nm/RIU], respectively. Initial thermal tuning results for air cladding yielded thermal efficiency values of around ~250[mW/nm] (~4[pm/mW]); further thermal characterization with more application-optimized cladding materials is in progress.

Conclusions & Significance: We progress towards a more cost-effective, fully integrated, and compact ultrasound sensing system for applications such as imaging tissue physiology, monitoring laser ablation during surgery, and to increase access to diagnostic imaging within under-served regions worldwide. Future work will focus on increasing device sensitivity and evaluating more application-specific device performance.

Abstract 36: Personalization of a myoelectric classification system to optimize information transfer after spinal cord injury.

Jonathan Eby^1,2, Jose Zariffa^1,2,3,4

¹Institute of Biomaterials and Biomedical Engineering, University of Toronto, Toronto, Canada. ²KITE – Toronto Rehabilitation Institute – University Health Network, Toronto, Canada. ³Rehabilitation Sciences Institute, University of Toronto, Toronto, Canada. ⁴Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, Canada

Rational and Objectives: Spinal cord injury (SCI) affects over 85,500 Canadians, with over 4500 new cases annually. Recovery of arm and hand function is a top priority for SCI individuals. Sophisticated assistive technologies exist but require high-bandwidth information transfer between the user and device. Myoelectric (sEMG) control algorithms can be used but do not account for varying muscle impairment patterns after SCI. By modifying the number of gestures recognized (GN) and the rate of change of gestures (GR), we can optimize information transfer for an individual’s level of impairment. The objectives are: 1) demonstrate that optimization of system parameters (GN and GR) for a user with SCI can increase the information transfer rate (ITR), a common brain-computer interface metric of efficiency; and 2) characterize the relationship between optimal design parameters and remaining neural activity in users with SCI.

Methods: 20 participants, 10 uninjured and 10 SCI, were recruited. Each participant was instrumented with 8 sEMG channels located on forearm anatomical landmarks. Tracing tasks assessed the level of volitional control of participants’ sEMG signals. Residual strength was assessed using manual muscle testing (MMT). Participants performed a series of trials varying GN and GR for a gesture classification system.

Results: Preliminary results show participants with SCI have peak ITR at non-maximal conditions of GN and GR, unlike uninjured participants. A weak correlation was observed between impairment metrics and the optimal GN. No significant correlation was observed for GR. 

Conclusions and Significance: With additional participants, we expect the relationships between impairment and model parameters will be clarified. From this, a model predicting optimal GN and GR for a given impairment pattern can be created. By optimizing the myoelectric classification system based on the individual’s pattern of impairment, we can improve assistive technologies such as exoskeletons or functional electrical stimulation for users with SCI.

Abstract 37: Using Pressure Mat Technology for Body Position Monitoring to Prevent Pressure Injuries

Najat ElFarra¹, Zahra Niazi¹, Georgia Bardaklis¹, Aleah Corbeth¹, Lindsay Stern^1,2, Atena Roshan Fekr^1,2

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ²KITE Reseach Institute, Toronto Rehabilitation Institute, UHN, Toronto, Canada

Purpose: Monitoring sleep positions has emerged as a pivotal strategy for preventing Pressure Injuries (PI). Patients with limited mobility have a higher risk of PI, as their inability to adjust positions frequently can result in damage to the skin and/or surrounding tissue. This abstract proposes an algorithm to detect the body positions automatically during sleep, using pressure distribution images. The aim of this algorithm is to ensure timely reminders for repositioning, thereby preventing the onset of PI.

Methods: We used an open-access dataset provided by Pouyan et al. [1] from ten healthy individuals who were instructed to lie on a mat, Vista Medical FSA SoftFlex with 2,048 pressure sensors (32×64 matrix). This experiment captured their movements across 17 unique body positions, maintaining each position for 2 minutes. The pressure data was then reconfigured into pressure distribution images to train and validate a pre-trained 2D Convolutional Neural Network (CNN), ResNet18, to classify these 17 postures. A Leave-One-Subject-Out (LOSO) cross validation technique was used to evaluate the algorithm.

Results: The proposed model achieved an accuracy of 96.56% ± 1.29% to classify all 17 body positions using LOSO. The algorithm demonstrated exceptional accuracy, with only 10 instances of misclassification against 1,188 correct classifications. This high accuracy indicates that the ResNet18 model can accurately distinguish between 17 positions, including postures captured on an inclined mattress. 

Conclusions: The proposed model can provide a basis to determine how long a patient is lying in a singular position and prompt family members or caregivers to reposition them to prevent the PI development.

Abstract 38: Feasibility of Depth Camera-Based Systems in Estimating the Joint Angle and Range of Motion

Alireza Ettefagh^1,2, Atena Roshan Fekr^1,2

¹KITE – University Health Network, Toronto, Canada. ²Institute of Biomedical Engineering, Toronto, Canada

Rationale and Objectives: Tele-rehabilitation is the delivery of physiotherapy services to individuals in their own homes, yet current tele-rehabilitation platforms rely on clinician’s presence throughout virtual sessions. Joint angle estimation using contactless technology plays a pivotal role in tele-rehabilitation as they enable monitoring and assessment of patient movements and Range of Motions (RoM).

This study aims to assess the feasibility of using depth cameras with a skeleton tracking module for estimating joint angles during lower limb rehabilitation exercises.

Methods: RGB and Depth data were collected using two depth cameras, alongside Motion Capture (Mocap) data from a single healthy participant performing Hip Flexion (HF) and Hip External Rotation (HER) exercises. The MediaPipe (MP) skeleton tracking model was used to find the lower body joint locations in RGB data. The joint coordinates were fused with depth data to enhance accuracy.

Results: To assess the accuracy of the proposed system, estimated knee joint angles were compared to Mocap measurements for each repetition. The estimated angles showed an average Mean Absolute Error (MAE) of 7.92° ± 1.59° and a Mean Squared Error (MSE) of 9.92° ± 2.26° compared to the ground truth, with an error of 1.71° ± 1.35° at the extremum point (MinError) across 7 repetitions of HF exercise. For HER exercise, the average MAE, MSE, and MinError over 5 repetitions were 7.76° ± 1.09°, 8.86° ± 1.27°, and 2.18° ± 0.46°, respectively.

Conclusions and Significance: The integration of depth information with skeleton data shows potential in tele-rehabilitation with errors of less than 8° in joint angle estimation, and with an error margin of less than 2° in estimating the RoM of the knee joint. Therefore, the proposed system holds promise for enhancing tele-rehabilitation practices, offering clinicians a tool for assessing patient movements with reasonable accuracy.

Abstract 39: Development of a Wearable Biofeedback System to Elicit Temporal Gait Abstract 39: Asymmetry using Rhythmic Auditory Stimulation and an Assessment of Immediate Effects

Aliaa Gouda^1,2, Jan Andrysek^1,2

¹University of Toronto, Toronto, Canada. ²Bloorview Research Institute, Toronto, Canada

Temporal gait asymmetry (TGA) is commonly observed in individuals facing mobility challenges. Rhythmic auditory stimulation (RAS) can improve temporal gait parameters by promoting synchronization with external cues. While biofeedback for gait training, providing real-time feedback based on specific gait parameters measured, has been proven to successfully elicit changes in gait patterns, RAS-based biofeedback as a treatment for TGA has not been explored. In this study, a wearable RAS-based biofeedback gait training system was developed to measure temporal gait symmetry in real-time and deliver RAS accordingly. Three different RAS-based biofeedback strategies were compared: open- and closed-loop RAS at constant and variable target levels. The main objective was to assess the ability of the system to induce TGA with able-bodied participants and evaluate and compare each strategy. With all three strategies, temporal symmetry was significantly altered (ranging between 4% and 10% TGA) compared to the baseline. The closed-loop strategy yielding the most significant changes when comparing at different target levels. Speed and cadence remained largely unchanged during RAS-based biofeedback gait training. Setting the metronome to a target beyond the intended target may potentially bring the individual closer to their symmetry target. These findings hold promise for developing personalized and effective gait training interventions to address TGA in patient populations with mobility limitations (e.g., lower-limb prosthesis users) using RAS.

Abstract 40: Respiratory Signal Extraction from ECG using a Phase-Amplitude Cross-Frequency Coupling Index

Adam Gravitis, Cesar Augusto Suarez Garcia, Berj Bardakjian

University of Toronto, Toronto, Canada

Rationale & Objectives: Direct measurement of respiration is comparatively rarer than recordings of ECG and EEG for patients under clinical study, especially those in epilepsy monitoring units. Methods: We propose a new method for extracting respiration from ECG using phase-amplitude cross-frequency coupling (PAC), which is more resilient to artifacts than existing extraction methods. Results: When analysed with a standard ECG and respiration dataset, PAC performs comparably to ECG envelope extraction, the most commonly employed approach. There was no significant difference between the wavelet phase coherence of this method versus ECG envelope, calculated against actual respiration, nor any significant difference between the distribution of predicted respiration rates. Conclusion: The PAC method’s decreased sensitivity to high-amplitude noise suggests it as a candidate for extracting respiration during high-amplitude seizure events when only ECG recordings are available.

Abstract 41: Real-time Location System to assess motor agitation in people living with dementia

Zain Hasan

University of Toronto, Toronto, Canada

Real-time location systems (RTLS) allow tracking of people over time and space with minimal obstruction to their daily lives. In long-term care settings, this data can then be used to analyze people’s motion, walking trajectory and social dynamics, residents’ interaction with their environment, and with each other. When used in conjunction with clinical metrics, including medication history (e.g., falls, records of motor agitation events), this data can provide insights and potentially predictive capabilities for various clinically relevant outcomes.  In the case of people living with dementia (PwD), an RTLS platform could help describe correlations between the movement patterns of residents when agitated, where they spend their time, how often they socialize with others, and what medications or dosages preceded motor agitation events.

In this study, we collected RTLS data from a specialized dementia unit caring for PwD’s behavioural and psychological symptoms of dementia and explored patterns in their motion to help train machine learning algorithms to classify motor agitation events. This analysis was conducted using solely RTLS-based motion dynamics features, such as motion speed, trajectory, non-linearity, to establish a baseline for predicting motor agitation events. Next, the same predictive modeling was carried out using purely clinical features (medication history, fall events, clinical agitation scores, mobility aids etc. to establish a second prediction baseline. Lastly, predictive models were built using both clinical and RTLS data. Our results showed that combining both trajectory metrics and clinical measures improve the predictability of motor agitation above both individual baselines. We also present comments on the efficacy, strengths, and weaknesses of predicting acute clinical events versus chronic events using RTLS features, to inform future avenues of research.

Abstract 42: Long-term Renal Fibrosis Evaluation Through Implantable Spectrometry-based Device

Hugo Higueros, Anat Usatisnky, Daniel Franklin

University of Toronto, Toronto, Canada

Renal fibrosis is a condition defined by the chronic buildup of scar tissue inside the kidneys. This pathology is most seen in aging adults and in people with chronic kidney disease. It is estimated that renal fibrosis affects 10% of the global population. To this day, there is no widely accepted paradigm for slowing the progression of renal fibrosis, and no consistent technique to reliably assess the disease progression stage precisely. The primary goal of this work is the development and validation of a long-term implantable optical device that is able to assess kidney function in a comprehensive and precise manner.  To achieve this a series of preliminary animal studies will be executed to determine the optimal wavelengths and device geometries for accurately measuring renal tissue oxygenation, water content and collagen deposition due to scar tissue. Afterwards, a state-of-the-art implantable device capable of spectroscopy analysis, Bluetooth communication, and battery management will be developed. Said device will be utilized in a series of in-vivo studies to precisely characterize kidney health in murine models under various subtypes and stages of renal fibrosis. Initial animal work has shown the potential for multiwavelength spectroscopy analysis through prototype devices that are able to characterize renal tissue oxygenation, which is direct indicator of renal composition and function. To further understand how the kidney precisely interacts with light, a series Monte Carlo simulations have also been developed. The development and completion of this projects opens new opportunities for preclinical models, as well as translational research. By offering a quantitative evaluation of kidney fibrosis, this project has a substantial potential to better the diagnosis and prognosis of a currently irreversible pathology that affects 10% of the general population.

Abstract 43: Development of a Gamified Modular Robotic Rehabilitation System to Enhance Motivation and Engagement in Post-Stroke Upper Limb Recovery Exercises

Deniz Jafari

University of Toronto, Toronto, Canada

Stroke is a major global health issue, impacting about 15 million people yearly with long-term disabilities. Extensive research emphasizes the importance of repetitive, high-intensity training for effective rehabilitation. Active patient engagement and personalized therapy are critical for optimizing outcomes.

We are currently in the process of developing a gamified exercise platform tailored for post-stroke rehabilitation. The platform incorporates a webcam and a vision-based pose-tracking deep neural network to monitor exercise performance. Real-time audio and visual feedback are provided to users to ensure proper posture during the exercises. To achieve this, a set of gamified upper limb exercises, aligned with the independent exercises prescribed by clinicians, has been designed. These exercises are intended for patients to perform autonomously between their therapy sessions.

The platform offers detailed instructions to patients and provides feedback on their performance, enabling the tracking of progress over time. To assess the platform’s effectiveness and user engagement, we are planning to conduct a usability study with both healthy older adults and post-stroke patients. The study aims to collect qualitative and quantitative data to evaluate the level of engagement during rehabilitation exercises.

Currently the study is in the development phase. In testing phase, participants from both groups will be asked to report on their level of engagement with the rehabilitation platform, as well as their perception of its effectiveness in addressing their arm rehabilitation needs. Additionally, feedback will be collected regarding their willingness to continue training with the system even after the study concludes.

Our research addresses challenges in adopting rehab platforms. We developed a customizable, engaging exercise system empowering patients to perform high repetition exercises independently. The platform provides optimal therapy with performance tracking, promoting patient engagement and autonomy without constant expert supervision.

Abstract 44: Simultaneous blood oxygenation and flow imaging with coherent optics

Jie Jiao, Dylan Dao, Sidy Ndiongue, Ofer Levi

UofT, Toronto, Canada

Rationale and objectives: In the field of biomedical imaging for clinical applications, accurate measurements of blood oxygenation and flow rate are desired for monitoring cardiovascular conditions of patients. An emerging technique for tracing blood oxygenation is SFDI (Spatial Frequency Domain Imaging), that provides a larger field of view than traditional pulse-oximetry with non-invasive features. In SFDI, a series of spatially modulated light patterns are projected onto the region of interest, with optical properties (absorption and scattering coefficients) of tissue recovered from the reflected light. Oxygen concentration in blood could therefore be evaluated. On the other hand, LSCI (Laser Speckle Contrast Imaging) applies lasers to create speckles as a coherent effect. The map of speckle contrast gives information on the blood flow velocity in capillaries. I am presenting an imaging system that is capable of measuring both factors simultaneously, with coherent optics.

Methods: For SFDI, laser speckles are usually not desired due to their spatial non-uniformities. Current sweeping is implemented on the laser source to reduce its coherence over a longer time period, with controlled framerate on camera to switch between SFDI and LSCI. Based on laser speckle statistics, real-time control of aperture size is another alternative of changing speckle intensity, to be explored in my study.

Results: By operating the laser in a sweeping mode, I have achieved a significant speckle contrast reduction of 90%, to create better uniformity for SFDI. In the meanwhile, a shorter capture period allows LSCI measurement in parallel. Expanding the aperture size of the camera has also produced expected speckle reduction.

Conclusions & Significance: An integrated blood oxygenation and flow imaging system could be achieved with laser current sweeping, to allow real-time and remote monitoring with a single coherent source. This technique could lead to clinical applications that reduce the risk of acute heart failure.

Abstract 45: Providing Hand Use Context for Outpatient Neurorehabilitation by Detecting Activities of Daily Living

Adesh Kadambi^1,2, José Zariffa^1,2

¹KITE Research Institute, Toronto, Canada. ²University of Toronto, Toronto, Canada

Rationale & Objectives: Assessing activities of daily living (ADLs) is crucial for guiding outpatient neurorehabilitation. However, traditional methods of gathering information about ADLs, such as self-report and direct observation, often fall short due to biases and environmental limitations. To address this gap, we previously designed a framework for reporting hand performance metrics to therapists via a dashboard based on egocentric video data, but therapists identified the need for context on the ADLs performed and object interactions. Our objective is to develop a machine learning model that can capture this contextual information by classifying the type of ADL being performed, thereby enabling therapists to make informed decisions and tailor therapy plans effectively.

Methods: We processed 2261 minutes of egocentric video recordings of individuals with spinal cord injury (SCI) performing unconstrained ADLs in a naturalistic environment by splitting the original recordings into 1-minute video snippets and extracting frames at 1 FPS, taking the sum of frame-level object detections across frames of a video snippet, and identifying active objects to create our “bag of objects” feature vector. We trained classifiers (logistic regression, random forest, gradient boosting, extreme gradient boosting, and a multi-layer perceptron) to classify between 7 ADL classes. We used leave-one-subject-out cross-validation for evaluation, and mean weighted F1-score and percentage of participants >0.5 weighted F1-score as evaluation metrics.

Results: The logistic regression model had the highest performance, with a mean weighted F1-score of 0.78. However, identifying grooming & health management and leisure activities was challenging due to data imbalances and the diversity of activities within these categories.

Conclusions & Significance: This study provides crucial insights into the effectiveness of machine learning in identifying ADLs in egocentric videos. This approach offers therapists with more objective contextual information about hand use at home, thereby enabling more personalised and effective therapeutic strategies.

Abstract 46: Real-Time Location Systems to Discern Clusters of Rest-Activity in People with Dementia

Yasser Karam^1,2, Shehroz S. Khan^1,2, Andrea Iaboni^1,2

¹University of Toronto, Toronto, Canada. ²KITE UHN, Toronto, Canada

Disruption to circadian rhythm is a hallmark of dementia, which manifests in the rest-activity rhythms of people with dementia. Their sleep-wake cycles are fragmented, with reduced daily activity and sometimes day-night reversal. These disruptions accelerate the need for residential care, where the rhythm is further disturbed by medications and environmental factors, such as light and noise. Sleep disturbances are linked to increased agitation, which can compromise their safety and quality of care. Subjective clinical assessments related to rest-activity rhythms are prone to mistakes related to caregivers’ recall bias and manual entries underscoring the necessity for incorporating objective measures. Reliable identification and classification of rest-activity rhythms can allow for more personalised and timely therapies for people with dementia. 
Real-time location systems (RTLS) technology utilizes radio frequency communication between a transmitter and receiver, and has applications in healthcare for patient and asset tracking. RTLS data offers a distinct advantage over conventional wrist-actigraphy for rest-activity measurement by allowing the incorporation of spatiotemporal information in the extraction of rest-activity features. This project aims to measure the rest-activity of people with dementia and explore longitudinal clinical correlations using RTLS data.

The sixteen rest-activity features extracted from the RTLS data primarily encompass the magnitude of activity, activity rhythmicity and time in bed. These features are investigated for identifying groupings of people using a hierarchical clustering algorithm. Results show that there are six different rest-activity clusters identifiable using these extracted features. The resulting clusters are investigated for clinical correlations using a multi-level mixed effects model. Preliminary results indicate that clusters differ based upon Neuropsychiatric Inventory sleep score that measures nighttime behaviour disturbances, as well as the age of participants (p<0.05). Further investigation is underway with clinical assessments measuring mobility, cognitive impairment, depression and agitation.

Abstract 47: The Assessment of Motor Unit Excitability using Electromyography-Based Techniques: A Scoping Review

David Koivisto^1,2, Joana Dilipkumar^1,2, Ryan Koh¹, Dinesh Kumbhare^2,1

¹University of Toronto, Toronto, Canada. ²University Health Network, Toronto, Canada

Historically, assessment of motor unit (MU) excitability has been concentrated on the MU pool (the group of MUs that innervate one muscle). Recent developments utilize decomposition algorithms to assess excitability at the neuronal level, allowing for novel insights. The heterogeneity of MU excitability evaluation poses a problem to researchers and clinicians. Therefore, we set out to evaluate: what electromyography-based technique in literature best quantifies changes in lower MU pool excitability? A clear definition of MU excitability is absent in current literature and has been defined in this review as the change in the facilitation or inhibition of synaptic input to the motor neuron characterized by MU firing. PubMed, MEDLINE, Embase, Web of Science, and Scopus were searched from inception to November 29th, 2023 for peer-reviewed articles that met the eligibility criteria. In total, 44 instances of various techniques were employed across 42 studies, which included MU decomposition (24), global surface electromyography (EMG) analysis (6), mathematical models (2), compound muscle action potentials (7), H-reflex (1), F-wave (3) and recurrence quantification analysis (1). From the excitability measures extracted, a considerable number were repeated across techniques. The most prevalent excitability measures were the number of identified MUs (16/44) and MU firing rate (19/44). The heterogeneity of the field is further expressed through the variation in how each of the excitability measures are reported. In total, 8 variations of firing rate were reported across studies that depend on the proposed research question such as maximum, mean, median firing etc. An optimal set of excitability measures cannot be recommended due to the differences in data acquisition between studies (muscles evaluated, strength of contraction and stimulation parameters). By extension, no single technique can be concluded to best describe MU excitability. To appropriately compare results between techniques, cohesive experiments are needed to limit variation.

Abstract 48: Depressive-like phenotype induced by AAV-mediated overexpression of human α-synuclein in midbrain dopaminergic neurons

Laura Kondrataviciute^1,2, Minesh Kapadia³, Jimmy George³, Taufik Valiante^1,3,4, Luka Milosevic^1,4,3, Lorraine Kalia^1,3,4, Suneil Kalia^1,3,4

¹University of Toronto, Toronto, Canada. ²Max Planck Institute for Intelligent Systems, Stuttgart, Germany. ³Krembil Research Institute, Toronto, Canada. ⁴University Health Network, Toronto, Canada

Parkinson’s disease (PD), as the most prevalent neurodegenerative motor disorder, presents a complex spectrum of symptoms encompassing both motor impairments (bradykinesia, resting tremor, freezing of gait) and non-motor manifestations (depression, anosmia, anxiety). Robust animal models that faithfully replicate this multifaceted pathology are indispensable for comprehensive PD research. This study aims to investigate if depressive-like behaviour, prominent in up to 40% of PD patients (Laux, 2022), emerges in the bilateral human mutated alpha-synuclein (A53T) rat model of PD. 

40 adult Sprague-Dawley rats were bilaterally injected into substantia nigra with either empty AAV1/2 vector or AAV1/2-expressing human mutated A53T-alpha synuclein. (Paxinos & Watson, 2013). Sucrose preference (Liu et al., 2018) and novelty suppressed feeding (Blasco-Serra et al., 2017) tests were performed on week 3 and week 6 post virus injection. Immunofluorescence staining for TH+, alpha-synuclein, DAPI was done to confirm neurodegeneration on week 6. 

Both groups of rats showed a preference for the 20% sucrose solution with no statistically significant difference between groups on week 3 (p=0.313, t test). Statistically significantly lower preference for sucrose was exhibited on week 6 by alpha-synuclein expressing animals compared to control group (p=0.0099). Reduced responsiveness to highly palatable food consumption has also been observed in novelty suppressed feeding test, but statistically significant differences between A53T and EV have not been reached at any timepoint.

Our findings suggest that the A53T-alpha synuclein rat model manifests depressive-like behavior, evidenced by diminished responsiveness to palatable stimuli. This model holds promise for investigating non-motor pathologies associated with PD.

Abstract 49: Development of Deep Learning Models for Motion Artifact Mitigation in Wearable Photoplethysmography Devices

Matthew Lee, Will Gao, Daniel Franklin, Chris McIntosh

University of Toronto, Toronto, Canada

Wearable devices are assuming a larger role in remote healthcare, fitness tracking, and athletics. Predominantly based on photoplethysmography, these devices use light to non-invasively detect changes in blood flow and oxygenation within peripheral circulation – leading to estimates of heart rate, pulse oximetry, and the identification of arrhythmias. However, the presence of motion greatly reduces the quality and interpretability of data from wearable devices and limits the development of ML/AI models. Currently, accelerometers are used to detect excessive motion and dispose of ‘contaminated’ data. However, accelerometers capture global motion, not the relative motion at the sensor-skin interface – the predominant source of motion artifacts. Here, we propose the reconstruction of physiological waveforms in the presence of motion through by developing 1) a multimodal sensor that captures multiwavelength photoplethysmography and force data and 2) a deep learning model that reconstructs denoised waveforms. The sensor combines force and multiwavelength optical measurements to capture relative motion at the sensor interface. Deep learning models are then trained to reconstruct the photoplethysmography waveform. Initial testing and training consists of lab-controlled motion artifact induction through a linear actuator and is then translated into real-world examples of motion. Utilizing a novel dataset (n=10), the deep learning algorithms have displayed output denoised photoplethysmography signals that have an average Pearson Correlation Coefficient of 0.78 with the reference signals. The initial results show indicators that the use of force and multiwavelength optical data can be used in conjunction with deep learning algorithms to reconstruct these important physiological signals in the presence of motion with strong generalization capabilities. This work will enable real-time motion artifact cancellation in wearable optical devices and lead to more advanced remote healthcare devices, athletic performance trackers, and algorithms.                                                       

Abstract 50: Investigating Muscle Electrophysiological Profiles in Cervical SCI Through Surface EMG Clustering Analysis for Tailored Rehabilitation

Guijin Li^1,2, Gustavo Balbinot^1,3, Julio C Furlan^1,4,5,6,7, Sukhvinder Kalsi-Ryan^1,4,8, José Zariffa^1,2,4,9

¹KITE Research Institute, University Health Network, Toronto, Canada. ²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³Krembil Research Institute, University Health Network, Toronto, Canada. ⁴Rehabilitation Sciences Institute, University of Toronto, Toronto, Canada. ⁵Department of Medicine, Division of Physical Medicine and Rehabilitation, University of Toronto, Toronto, Canada. ⁶Division of Physical Medicine and Rehabilitation, Toronto Rehabilitation Institute, University Health Network, Toronto, Canada. ⁷Institute of Medical Sciences, University of Toronto, Toronto, Canada. ⁸Department of Physical Therapy, University of Toronto, Toronto, Canada. ⁹Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, Canada

Rationale & Objectives: Spinal cord injury (SCI) can cause significant impairment and disability, impacting individuals’ quality of life and independence. Informative assessments are essential for efficient treatment planning. Surface electromyography (sEMG) is a sensitive and non-invasive technique to measure muscle activity and has demonstrated great potential in capturing the impact from SCI. One of the major barriers for a more successful clinical translation is the reliable extraction and interpretation of sEMG signal features. To further explore the clinical relevance of sEMG signals and help promote their clinical translation for precision rehabilitation, we seek to explore whether distinct muscle electrophysiological profiles can be identified from sEMG signal features after cervical SCI. We hypothesize that distinct clusters can be identified in the sEMG feature space. 

Methods: We recruited 22 participants with cervical SCI. Prior to their functional electrical stimulation therapy, baseline sEMG signals from 184 muscle groups were recorded during submaximal (50%) voluntary contractions with resistance provided according to the manual muscle testing protocol. Features from the SEMG signals in both time and frequency domains were extracted. Multiple feature sets were then curated with dimensionality reduction and feature selection methods. We performed clustering analysis using k-means, k-medoids, and agglomerative hierarchical clustering algorithms. After the initial parameter tuning evaluated by silhouette score, resampling using bootstrapping with replacement was done to ensure stability. 

Results: The algorithms were able to identify two to three clusters from the curated sEMG feature sets, with the silhouette score ranging from 0.60 to 0.73, suggesting that distinct clusters do exist. Next, these clusters will be related to other neurological and functional indicators of muscle function after SCI to investigate their clinical relevance.

Conclusions & Significance: Through the identification of electrophysiological profiles, our study could contribute to sEMG interpretation after SCI as well as guiding clinical planning.

Abstract 51: Using deep learning fusion architectures to classify impaired hands after stroke using multiple egocentric videos of activities of daily living

Anne Mei^1,2, Meng-Fen Tsai^1,2, José Zariffa^1,2,3,4

¹KITE – Toronto Rehabilitation Institute – University Health Network, Toronto, Canada. ²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³Rehabilitation Sciences Institute, University of Toronto, Toronto, Canada. ⁴Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, Canada

Rationale & Objectives: After stroke, structured clinical assessments are conducted to determine the effectiveness of new rehabilitation therapies in upper-extremity motor control, but do not reflect true behaviour in activities of daily living (ADLs) at home. Wearable (egocentric) cameras provide a way to capture hand function information in natural environments. Automated analysis of egocentric video has previously been explored to analyze impaired hand function due to neurological disorders, but using single-video inputs. In this study, we investigated deep learning approaches to categorize hand impairment severity of individuals after stroke, using multiple ADL video inputs, to reflect clinical hand function assessments that rely on observing hands in several functional tasks.

Methods: An egocentric video dataset of 38 ADLs performed by 6 stroke survivors in a home simulation lab was used to train a SlowFast baseline model to classify hands into impaired and unimpaired categories. Using SlowFast as a feature extractor, four fusion methods were investigated to develop multi-video architectures to accept sets of three videos as inputs.

Results: Through evaluation with Leave-One-Subject-Out-Cross-Validation, the baseline model (F1-score of 0.61 ± 0.13) has been found to perform similarly to the two late fusion methods explored, majority voting and a fully-connected network (F1-scores of 0.60 ± 0.19 and 0.61 ± 0.18, respectively). In comparison, the two intermediate fusion methods explored, feature concatenation and sequential chain, have a lower performance (F1-scores of 0.48 ± 0.21 and 0.5 ± 0.07, respectively).

Conclusions & Significance: The fusion methods’ lack of improvement in performance may be attributed to the challenging dataset, in which mildly impaired and unimpaired hands are difficult to visually distinguish. This study investigates, for the first time, the impact of using multiple ADLs for the automated assessment of hand impairment at home and evaluation of new rehabilitative interventions on the daily lives of stroke survivors.

Abstract 52: Detection of Agitation in People with Dementia using Different Video Camera Views

Pratik K. Mishra^1,2, Andrea Iaboni^1,2, Bing Ye², Kristine Newman³, Alex Mihailidis^1,2, Shehroz S. Khan^1,2

¹University of Toronto, Toronto, Canada. ²KITE Research Institute, Toronto, Canada. ³Toronto Metropolitan University, Toronto, Canada

Rationale & Objectives: Managing behavioural symptoms of dementia within Long Term Care (LTC) homes poses significant challenges in ensuring resident safety with limited staff resources. Although many LTC facilities have installed video surveillance systems in common areas, these streams typically go unmonitored. In this study, we introduce an approach for utilizing these video streams to detect clinically important episodes of agitation in individuals with dementia.

Methods: Given the rarity of agitation episodes compared to normal behaviours, we formulate this as an anomaly detection problem, focusing on monitoring the scene rather than tracking individuals. Our method involves training a custom spatio-temporal convolutional autoencoder on normal behaviours and identifying agitation as anomalous behaviour during testing. We validate our approach using video data from a specialized dementia unit annotated for agitation events. This work is an extension of our previous work with one participant and one camera to now ten participants and three cameras. 

Results: By training the autoencoder on an average of 1224.4 minutes of normal activities and testing on 199.2 minutes of videos containing both normal activities and agitation events, we achieved the area under the receiver operating characteristic curve of 0.844, 0.818 and 0.765, respectively, on three cameras. 

Conclusions & Significance: This research opens avenues for leveraging existing surveillance infrastructure in LTC and other mental health settings to detect agitation, potentially enhancing health and safety outcomes.

Abstract 53: Automatic Recognition of Lower Limb Rehabilitation Exercises Using a Pressure-Sensitive Mat

Pourya Moghadam^1,2, Ruiwen Xian², Shirui Zhuang², Yat Ching Kwong², Alireza Ettefagh^1,2, Atena Roshan Fekr^1,2,3

¹The KITE Research Institute, Toronto Rehabilitation Institute, Toronto, Canada. ²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³Rehabilitation Research Institute, University of Toronto, Toronto, Canada

Rationale & Objectives: Tele-rehabilitation (Tele-rehab) is a crucial aspect of recovery, but adherence to prescribed exercise routines remains a significant challenge. This study investigates a novel approach using a pressure-sensitive mat for remote monitoring of lower limb exercises in patients undergoing at-home rehabilitation. The vision-based approaches often struggle with recognizing exercises due to occlusion. Pressure mats offer an alternative data source focusing on ground reaction forces, potentially leading to more reliable exercise recognition, especially for lower limb exercises.

Methods: The research involves analyzing data from 30 healthy participants (P1-P30) performing 7 physiotherapy exercises (Knee-rolling, Bridging, Pelvic Tilt, The Clam, Repeated Extension in Lying, Prone punches, and Superman) from an open-source dataset proposed by Wijekoon et. al (Multi-modal Exercises Dataset – MEx). A Convolutional Neural Network (CNN) model, known as SlowFast, was trained on pressure map frames sampled at a frequency of 15Hz to classify various exercises performed by each subject. This model processes video data through both its slow and fast pathways, enabling the model to capture spatial details and temporal dynamics effectively. For validation purposes, participants were divided into 6 groups, each consisting of 5 individuals.

Result: The accuracy values range from 78.31% (for group #5: P21-P25) to 93.02% (group #1: P1-P5) with an average accuracy of 86.78% ± 5.80%.

Conclusion and Significance: The creation of automated exercise monitoring platforms offers a promising approach to enhancing both the adherence to and accessibility of rehabilitation programs. By leveraging data from a pressure-sensitive mat, this study introduces a novel approach to address the limitations of existing systems in recognizing lower limb exercises.

Abstract 54: Protocol for Development of AVA: A Platform for Avatar-Assisted Virtual Cardiac Rehabilitation

Fateme Pourghasem^1,2, Ali Abedi², Tracey JF. Colella^3,2, Shehroz Khan^2,1

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ²KITE, Toronto Rehab, UHN, Toronto, Canada. ³Lawrence S Bloomberg Faculty of Nursing, Rehabilitation Sciences Institute, University of Toronto, Toronto, Canada

Rationale and Objectives: Virtual cardiac rehabilitation is proven to improve mortality and reduce future cardiac events. However, its effectiveness is often compromised by issues such as low accessibility, low adherence, and staff shortages leading to patient dropout and reduced health outcomes. To address these challenges, the Adaptive Virtual Assistant (AVA), an AI-driven virtual avatar, has been developed to provide at-home exercise guidance and exercise quality monitoring, aiming to improve cardiac health and reduce dropout rates. The study will be conducted using a participatory design framework to ensure user-centered development of the virtual assistants.

Method: We will develop a digital platform featuring an on-screen 3D avatar, customizable appearance, initially validated through testing with five healthy volunteers. A co-design strategy will be employed, engaging patients, caregivers, clinicians, and researchers to select the most suitable avatars. We will then assess the feasibility of the AVA’s platform in real-world scenarios, focusing on its impact on patient engagement, exercise quality, and rehabilitation completion rates. Following platform refinement based on patient feedback, a two-phase study at the Rumsey Cardiac Center, Toronto Rehabilitation Institute will recruit 25 cardiac patients. The first phase will involve developing personalized AI algorithms to evaluate and provide feedback on patients’ exercise techniques and quality. The second phase will validate the performance of these algorithms in a real-world setting.

Results: AVA will enhance at-home cardiac rehabilitation delivery, especially to remote communities, and expected to lower patient dropout rates, thereby improving long-term health outcomes and quality of life for cardiovascular patients.

Conclusions: AVA system can improve patient engagement and rehab completion rates in VRehab programs. It expands equitable rehabilitation services and has the potential to improve health outcomes while reducing healthcare costs. The study’s findings will have implications for the development of personalized telerehabilitation interventions and the delivery of rehabilitation services in remote settings.

Abstract 55: Multi-Sensor Wearable Technology to Safely Navigate Exercise in Hot and Cold Environments

Jahir Ibna Rafiq, Dr Shehroz Khan

University of Toronto, Toronto, Canada

Cardiovascular Disease (CVD) remains a leading cause of mortality worldwide, claiming 20.5 million lives in 2021 and 17.9 million in 2019. While exercise is known to mitigate cardiovascular risks, engaging in physical activity in extreme temperatures poses unique dangers for CVD patients. Existing wearable devices lack the capacity to alert users to these risks. To address this gap, our study aims to develop a multi-sensor wearable device with an algorithm tailored to detect vulnerable situations.

Forty overweight or obese patients with diabetes mellitus and a history of hypertension will participate. They will undergo 30-minute exercise sessions in optimal, hot, and cold climates at the Climate Labs, Toronto Rehabilitation Institute. Physiological parameters, including heart rate, electrocardiogram data, skin and core temperature, and galvanic skin response, will be continuously monitored. Participants will also provide subjective feedback on fatigue, exertion, and comfort levels. Environmental factors such as temperature, humidity, and CO2 levels will be measured for impact analysis.

We will explore deep learning regression model for checking the variability of physiological parameters and classification model for identifying potential risk for this collected time series data where RNNs model, such as LSTM and GRU, as well as transformer network can be applied for both regression task and classification task. Advanced machine learning techniques for analysing and visualizing data including t-SNE and Grad-CAM will be exploited.

Some pilot data has already been obtained and encouraging results were seen although data collection will take more and thus the final work can only be done after that.

This research offers insight into physiological responses during exercise in extreme temperatures, crucial for CVD patients. It will advance wearable tech for personalized risk monitoring. This research promises to revolutionize personalized risk monitoring for cardiovascular patients, enabling millions of Canadians to exercise with greater confidence and understanding of their vulnerability.

Abstract 56: Development of a Novel Approach for Assessing Range of Motion in Upper Extremity Post-Stroke or Spinal Cord Injury

Aisha Raji^1,2, Cesar Marquez-Chin^1,2, Urvashy Gopaul², Milos Popovic^1,2,3

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ²KITE Research Institute, Toronto Rehabilitation Institute – University Health Network, Toronto, Canada. ³Rehabilitation Sciences Institute, University of Toronto, Toronto, Canada

Rationale & Objectives: Over half of individuals with stroke and spinal cord injury (SCI) experience upper extremity impairment, highlighting the need for effective rehabilitation. Robotic rehabilitation offers promise in improving upper extremity motor function by encouraging intensive, repetitive training and providing objective measure of performance. However, current evidence suggests its outcome is not superior to most conventional rehabilitation methods. Hence, our objective is to develop a new robotic approach that promotes functional reaching and grasping, using relevant objects tailored, to individual user needs and abilities. 

Methods: This study involves: 

a. creating standardized objects for the robot to train patients to perform functional tasks such as grasping a cup, handling a credit card, and picking up a phone.

b. constructing a dedicated shelf for consistent object positioning and storage. 

c. configuring the robotic system and its safety components to ensure patient safety during object presentation, retrieval, and training. 

d. creating a program for the robot to assess users’ range of motion

Results: Objects from the Toronto Rehabilitation Institute – Hand Function Test (TRI-HFT) were selected because they encourage the practice of various reaching, grasping, and object manipulation functions. These objects have been modified and 3D-printed with handles to facilitate proper gripping by the robot. A prototype of an object storage shelf has been constructed, and the robotic system has been configured with its safety components. Range of motion assessment will be conducted by the robot through spatial exploration of users’ hand capacity using one of the TRI-HFT objects.

Conclusions & Significance: This work aims to improve the outcome of robotic rehabilitation by integrating functional tasks.  By assessing users’ extent of reach and adapting object presentation accordingly, gains can be easily translated into daily activities. This innovative approach has the potential to advance upper extremity motor rehabilitation post-stroke and SCI.

Abstract 57: Development of multilayer pigmented biomedical optical phantoms with vasculature microchannels for testing photoplethysmography devices

Megh Rathod^1,2, Farida Abdelmalek¹, Heather Ross^2,3, Daniel Franklin¹

¹Institute of Biomedical Engineering, University of Toronto, Toronto, ON, Canada. ²Ted Rogers Center for Heart Research, Toronto, ON, Canada. ³Peter Munk Cardiac Center- University Health Network, Toronto, ON, Canada

Advances in wearables have led to the rise of non-invasive optical devices that perform pulse oximetry. The impact of skin pigmentation on pulse oximetry and photoplethysmography (PPG) is critical amid numerous evidence that pulse oximetry is less accurate on pigmented skin. This study combined the fabrication of tunable pigmented skin phantoms, with a method to generate mechanically flexible phantoms with vasculature channels. 

We have developed dynamic pulsatile phantoms for PPG signal evaluation to create non-invasive, wearable devices for monitoring cardiovascular health. Skin phantoms were developed to quantify the impact of melanin concentrations on reflectance values. Phantoms were custom-made with PDMS (polydimethylsiloxane) as the base because of the similarity of the refractive index to epidermis tissue. TiO2 was used for scattering, and granulated melanin was used as the absorbing agent. The phantom manufacturing process allowed for the control of the optical properties, interfaces and geometrical form factors known to influence the PPG signals. 

The phantoms were characterized via spectroscopy with a high-resolution VIS-IR spectrometer with a dual reflectance probe. Generated pulse waves were created by pumping fluid through a microchannel, and a reflectance-mode Maxim 86180 optical data acquisition system evaluated the impact of pigmentation on optical signals. 

This work demonstrated a methodology to rapidly fabricate dynamic phantoms that can provide a platform for testing pulse oximeter performance across the spectrum of human pigmentation and can be adapted for other forms of variability. This study furthers our understanding of how melanin impacts wearable optical sensing and provides an in-vitro testing platform for device calibration. Skin-tone invariant wearables are essential for the equitable future of cardiac and respiratory monitoring devices. Models developed in this work have wide relevance to existing and future clinical and commercial monitors and allow further exploration into advanced hemodynamic monitoring and its equitable development.

Abstract 58: From Caregivers’ Activity Recognition to Subsequent PatientInfection Risk Prediction

Koorosh Roohi^1,2, Atena Roshan Fekr^1,2

¹KITE – Toronto Rehabilitation Institute – University Health Network, Toronto, Canada. ²Institute of Biomaterials and Biomedical Engineering – University of Toronto, Toronto, Canada

Rationale & Objectives: Healthcare-acquired infections (HAIs) pose a significant challenge to patient safety, leading to extended hospital stays and increased healthcare costs. While factors such as caregivers’ compliance with Hand Hygiene (HH) play an important role, the actions of caregivers during patient interactions are critical in the chain of HAI transmission. The objective of this study is to address this issue by developing a predictive system to identify potential infection risks based on the actions of caregivers during patient care.

Methods: The study used a unique dataset known as the CARECOM Nurse Care Activity Dataset. This dataset includes six different nursing activities performed by eight caregivers. The collected motion capture data was preprocessed and resampled to train two models: Spatiotemporal Graph Convolutional Neural Networks (ST-GCN) and an enhanced version, ST-GCN++. The pre-trained version of these models was fine-tuned using skeleton data from 23 joints, providing a comprehensive representation of caregiver actions.

Results: The ST-GCN++ model demonstrated superior performance, achieving an accuracy of 82.41% on unseen test data. This model outperformed other skeleton-based and multimodal approaches in recognizing caregiver activities while keeping the performance at a real-time level. However, it’s important to note that there were some misclassifications in activities such as vital sign measurements and blood collection, which can be attributed to imbalances in the dataset.

Conclusions & Significance: The lightweight ST-GCN++ model’s high accuracy in recognizing caregiver activities indicates its potential for integration into monitoring systems to assess infection risk. This could lead to significant improvements in patient safety and a reduction in HAIs in healthcare settings. The findings of this study underscore the importance of accurate activity recognition in infection control strategies and pave the way for future research in this area.

Abstract 59: Influence of Body Posture and Electrode Placement on ECG Signal Measurement Using Textile Electrodes

Delaram Sadatamin^1,2, Ivana Culjak², Saba Sadatamin^1,3, Bryan Piper^1,2, Baptiste Garnier², Azadeh Yadollahi^2,1

¹BME UofT, Toronto, Canada. ²KITE UHN, Toronto, Canada. ³Wilfred and Joyce Posluns Centre for Image-Guided Innovation and Therapeutic Intervention,Hospital for Sick Children, Toronto, Canada

Rationale & Significance: Wearable technology has become increasingly popular because it easily fits into our daily lives. These devices can gather a lot of information to help us monitor and control different health aspects more effectively. For example, smart clothes can track our heart activity through the electrocardiogram (ECG), which can be done conveniently, without being in a hospital, and continuously. However, the accuracy of the ECG readings can be affected by the body posture and electrode placement.

Objective: The main goal of this research is to explore how these factors—body posture and electrode placement—affect the quality of the ECG readings.

Methods: In this study, three healthy adults participated. We collected ECG data using gel-based and textile-based electrodes at the same time. We then compared the heart rate readings from the textile electrodes to those from the gold standard gel electrodes in different body postures and electrode placements. We performed statistical analyses to determine the textile electrodes’ accuracy in capturing ECG compared to the gold standard method, under different conditions. We also calculated the difference in heart rate readings between the two types of sensors across various conditions to identify which combinations provide the most accurate heart rate readings compared to the standard method.

Results & Conclusion: Preliminary findings suggest that body posture and electrode placement can change the heart rate measurement accuracy. We plan to collect more data to provide stronger evidence on how these factors influence the accuracy of ECG readings.

Abstract 60: Optimizing MRgLITT Patient Monitoring through Time-series based Deep Learning Methods: A Comparative Study of ConvLSTM and U-Net

Saba Sadatamin^1,2, Paola Driza¹, Gemma Postill¹, Steven Robbins³, Richard Tyc³, Rahul Krishnan¹, Lueder Kahrs¹, Adam Waspe¹, James Drake¹

¹University of Toronto, Toronto, Canada. ²Hospital for Sick Children, Toronto, Canada. ³Monteris Medical, Winnipeg, Canada

Magnetic resonance-guided laser interstitial thermal therapy (MRgLITT) is a novel, minimally invasive therapeutic approach that leverages thermal ablation to treat drug-resistant focal epilepsy. Patient-specific heat sinks, such as blood vessels, complicate planning of MRgLITT as it creates patient-level variability in how heat from the laser propagates, thus potentially undermining treatment efficacy. To simulate the MRgLITT outcomes, we developed a deep learning framework which can predict the next steps of the monitoring system helps the surgeons to figure out, how long they want to keep ablating. We evaluate the outcome using both spatial and temporal metrics. We demonstrate strong performance of our best framework, ConvLSTM, with a structural similarity index metric of 88\%, Dice score of 0.85\% and sensitivity of 0.77\% which shows that heat propagation predicted highly similar to the ground truth. Our findings can be used by neurosurgeons to improve the delivery of MRgLITT.

Abstract 61: Modifying Ankle Muscle Stiffness using Neuromuscular Electrical Stimulation

Fatemeh Shomal Zadeh^1,2, Jae W. Lee^1,2, Derrick Lim^1,2, Kai Lon Fok^1,2, Jonguk Lee^1,2, Dinesh Kumbhare^1,2, Kei Masani^1,2

¹Institute of Biomedical Engineering, University of Toronto, Ontario, Canada. ²KITE – Toronto Rehabilitation Institute – University Health Network, Ontario, Canada

Individuals with incomplete spinal cord injury (iSCI) often suffer from impaired balance ability. Controlling plantar flexors (i.e., soleus (SOL)) and dorsiflexors (tibialis-anterior (TA)) are crucial for maintaining balance, which are often deteriorated after iSCI. Regulating ankle joint stiffness plays a key role in maintaining standing balance, which is modified through regulating the ankle muscle stiffness of SOL and TA. We have developed a therapy for standing balance using neuromuscular electrical stimulation (NMES) on SOL and TA, which showed a promising performance in a pilot study. Further details such as the ability of NMES on modifying the ankle muscle stiffness need to be further studied to enhance the therapeutic effect of the proposed system. Therefore, we are currently investigating the ability of NMES in modifying the ankle muscle stiffness using shear wave elastography (SWE). Two female able-bodied participants were positioned on an isometric electro-dynamometer. An ultrasonic probe was used to acquire SWE-images. During plantar- and dorsi-flexion, SWE of SOL and TA as well as the ankle joint torque were recorded at 1) relax, 2) voluntary contraction, 3) contraction induced by NMES, and 4) combination of 2) and 3). From the SWE-images, the Young’s elastic modulus was calculated as a surrogate of stiffness for each image. SWE can measure increase in muscle stiffness resulting from voluntary vs. artificially induced contractions. There was an approximately linear relationship between ankle torque and muscle stiffness. This can provide information about the required stimulation intensity in therapy to maintain balance. While the data collection is ongoing, these preliminary results suggest that SWE is a potential tool to quantify muscle stiffness, which can contribute to the ankle stiffness and hence the postural stability. In the next step, the effect of NMES on improving the postural stability will be assessed using this tool, specifically in individuals with iSCI.

Abstract 62: Effect of Mattress Stiffness and Sleeping Aids on Posture Detection via Pressure Imaging

Lindsay Stern^1,2, Atena Roshan Fekr^2,1

¹University of Toronto, Toronto, Canada. ²KITE Research Institute, Toronto, Canada

Rationale & Objectives: The prevalence of pressure injuries (PIs) is highest among the aging population due to a gradual decline in overall health, including aspects such as nutrition and mobility. This decline can result in injuries to the skin and nearby tissues due to extended periods in one position. To prevent this, clinical practice recommends regular repositioning of patients. This study aims to (1) identify in-bed postures using pressure distribution images captured while participants were lying with a blanket and pillow and (2) to evaluate the effects of mattress stiffness on in-bed posture detection.

Methods: Ten healthy participants were recruited to lie on 3 mattresses using sleeping aids, such as pillows and blankets. The 3 mattress environments include a hospital bed (high stiffness), a home bed (medium stiffness), and a foam mattress topper placed on top of a bed (low stiffness). Data was collected from a pressure-sensitive mat, composed of 1,056 sensors. Participants were asked to lie in 17 postures related to supine, prone, right side, and left side classes. Additionally, data was collected from seated positions and an empty bed. A 2D pre-trained convolutional neural network (CNN), the ResNet-18, was trained and validated using Leave-One-Subject-Out (LOSO) cross validation. 

Results: The F1-scores obtained on the hospital bed, home mattress, and foam topper were 90.97% ± 9.56%, 90.53% ± 10.06%, and 81.86% ± 5.17%, respectively. These results indicate that the algorithm performed well even when sleeping aids were considered, creating a more realistic analysis. Additionally, these results show that as mattress stiffness decreases, the performance of the algorithm decreases as well.

Conclusions & Significance: The proposed method has great potential to be used in hospitals and at home for PI prevention. It can notify the user or caregiver when it is time to reposition if this has not occurred naturally.

Abstract 63: Improving selective peripheral neural recordings through transfer learning

Yinghe Sun^1,2, José Zariffa^1,2,3,4

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ²KITE, Toronto Rehabilitation Institute, University Health Network, Toronto, Canada. ³Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, Canada. ⁴Rehabilitation Sciences Institute, University of Toronto, Toronto, Canada

Rationale & Objectives: Peripheral nerve interfaces can be used to interface advanced assistive technologies with the human nervous system. Current prototypes use neural networks to classify activities from different neural pathways based on spatiotemporal neural features that vary across subjects, making the model subject-specific. It would be desirable to use data from other subjects to improve the generalizability of the model. Therefore, the objective was to leverage data from different subjects through transfer learning and investigate the impact it brought to the model’s performance in classification tasks.

Methods: The study applied the Extraneural Spatiotemporal Compound Action Potentials Extraction Network to classify afferent naturally evoked compound action potentials corresponding to three different sensory stimuli—ankle dorsiflexion, plantarflexion and heel pricking. The datasets used were electroneurographic recordings sampled from the sciatic nerves of 9 Long-Evans Rats through 7×8 multi-channel nerve cuff electrodes. 9 models were developed on the dataset from each rat as baseline. Then, the models were trained starting from weights trained on one other rat. Every combination of two rats was included. The mean accuracy and mean macro F1 score of these models were evaluated.

Results: No statistically significant differences (p>0.05) were found regarding the changes in the mean macro F1 score and the mean accuracy of the best performance of transfer learning models on each rat compared to the baseline models, with the baseline models and best performing transfer learning models having 0.733±0.121 compared to 0.751±0.123 mean macro F1 scores, and 0.794±0.114 compared to 0.811±0.107 mean accuracies.

Conclusions & Significance: The study found that applying transfer learning across different datasets with the same size did not result in significant improvements towards the models’ performance in classifying neural signatures. The study provides new perspectives on improving the robustness of neural interfaces with deep learning.

Abstract 64: Progression of Osteosarcopenia in Prostate Cancer Patients using AI-Enabled Musculoskeletal Imaging Biomarkers

Saleh Tabatabaei^1,2, Tayler Declan Ross^1,3,4, Geoff Klein^1,5, Joel Finkelstein^1,4, Cari M. Whyne^1,2,3,4, Urban Emmenegger^1,3, Michael Hardisty^1,4

¹Sunnybrook Research Institute, Toronto, Canada. ²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³Institute of Medical Science, University of Toronto, Toronto, Canada. ⁴Department of Surgery, University of Toronto, Toronto, Canada. ⁵Department of Medical Biophysics, University of Toronto, Toronto, Canada

Rationale: Osteosarcopenia (combined osteopenia [decreased bone mineral density-BMD] and sarcopenia [low muscle mass and strength]) is prevalent in prostate cancer patients. Osteosarcopenia decreases musculoskeletal (MSK) health, leading to poor patient outcomes. Traditionally, these conditions were studied separately at individual time points, limiting understanding of interactions and progression. Objective: To retrospectively quantify osteosarcopenia progression in advanced prostate cancer patients using imaging biomarkers. Hypothesis: Sarcopenia will progress over time, but BMD may rise due to sclerotic metastatic lesions.

Methods: MSK biomarkers were calculated for patients treated with systemic therapy at Sunnybrook Odette Cancer Centre (2009-2021) using opportunistic prostate cancer surveillance images (REB#1862). Lumbar Spine 3D CT scans were reconstructed to a higher resolution of 1mm. Using a U-Net based convolutional neural network, the lumbar vertebral bodies (L1-L5) and psoas muscles were segmented. The psoas muscles’ superior and inferior limits were defined from the L2/L3 to L4/L5 disc mid-points. BMD of the lumbar vertebral bodies and psoas muscle volume and density were extracted.

Results: 142 male patients (age=71years,range=44-90years) with 475 CT scans were analyzed with average follow up of 1034days. Throughout this period BMD increased (17.1%,STD=48%), psoas volume decreased (5.7%,STD=34.9%), and psoas density decreased (17.9%,STD=46.7%). BMD was stable or decreased in 51.8% of patients, with 25.9% experiencing a >30% increase. Psoas metrics decreased in 2⁄3subjects. Moderate to weak correlations were observed between the temporal rate of change of MSK biomarkers (psoas volume vs. BMD: R2=-0.46,p<0.0001).

Conclusion & Significance: The majority of patients experienced sarcopenia progression. This study showed larger magnitude changes in osteopenia measures than previous reports, likely attributable to longer follow-up. BMD increases were measured secondary to sclerotic metastatic lesions. The osteosarcopenia quantitative 3D imaging biomarkers are more sensitive to changes than 2D biomarkers for longitudinal assessment. These methods can track osteosarcopenia progression and monitor metastatic progression and other MSK health-impacting conditions.

Abstract 65: Feasibility of Employing Speech Representations in Screening Obstructive Sleep Apnea

Behrad Taghibeyglou^1,2, Alexander Chow², Parker McLaurin³, Oviga Yasokaran², Rene Adams², Majida Mohammed², Mandeep Singh⁴, Najib Ayas⁵, Sachin R. Pendharkar⁶, Fernanda Almeida⁵, Valeria Rac³, Azadeh Yadollahi^1,2

¹University of Toronto, Toronto, Canada. ²Toronto Rehabilitation Insitute- University Health Network, Toronto, Canada. ³Toronto General Hospital- University Health Network, Toronto, Canada. ⁴Toronto Western Hospital-University Health Network, Toronto, Canada. ⁵University of British Columbia, Vancouver, Canada. ⁶University of Calgary, Calgary, Canada

Rationale: Obstructive Sleep apnea (OSA) is highly underdiagnosed in adults. In particular, underserved population, such as people who live in shelters have even more barriers to sleepcare, e.g. difficulty visiting medical facilities. Our preliminary results show a high prevalence of sleep apnea in shelter residents. However, the clinical diagnosis using polysomnography is not accessible to shelter residents. Speech, as a non-invasive biomarker for monitoring physiological changes in the pharyngeal airway, can be a potential tool for objective assessment of the risk of sleep apnea.

Objective: This study proposes a novel pipeline based on classical acoustic features to estimate the risk of OSA using five vowels and two phonemes recorded in standing and sitting postures.

Methods: Adult shelter residents were included. Two research assistants, a person with lived experience of homelessness and sleep apnea and an expert in speech and sleep physiology collected the data. We asked the participants to repeat five vowels (/see/, /sahh/, /set/, /so/, /soo/), and two nasal phonemes (/n/, /m/) twice, while seated and holding an audio recorder in front of them. Afterwards, participants were set up with a level-II portable polysomnography device for the overnight sleep study. An obstructive apnea-hypopnea-index threshold of 10 events/hour was used to differentiate OSA vs. non-OSA groups. A transfer learning approach incorporating magnitude and phase-based representations alongside a pre-trained SEResNet-50 was used to compare our model. Our framework includes extracting spectrotemporal speech features.

Results: We recruited 35 residents (18 women, 45.4±12.4 years). Eighteen participants were categorized as OSA (one woman) and 17 were non-OSA. Our proposed framework achieved a remarkable F1 score of 0.93 in classifying individuals living in homeless shelters.

Conclusion: Our findings reveal the feasibility of employing acoustic models in accessible speech-based OSA screening.

Abstract 66: Unveiling Dataset Bias: Impeding Generalization of Remote Photoplethysmography for Heart Rate Monitoring

Anshul Verma, Shehroz S. Khan

University of Toronto, Toronto, Canada

About one in 12 Canadian adults aged 20 and over have diagnosed heart disease, and it is the second leading cause of death in the country. Despite the frequent use of wearable devices, their issues with clinical validation and compliance are prevalent. Hence, there is a need for clinically useful, non-invasive, and user-friendly alternatives. Remote photoplethysmography (rPPG) and heart rate (HR) monitoring through videos are promising alternatives. However, there is a notable gap in extending these algorithms to clinical settings for patient use. Our objective is to investigate the reasons behind this gap. To achieve this, we conducted a narrative literature review and identified the limitations of the current rPPG algorithms. We found 19 publicly available rPPG datasets, mostly consisting of younger and healthy individuals with a mean age of 32.06 ± 9.2 years. Remarkably, only one dataset contains data from six patients with atrial fibrillation (A-fib or AF). Our analysis revealed that different datasets used for training rPPG algorithms are significantly different and easily distinguishable. Classification experiment to identify which dataset a video sample (or its spatial-temporal maps) belongs to could achieve 97% test accuracy. We also noticed this behavior in out-of-distribution accuracies, where training the rPPG algorithm on one dataset (VIPL-HR) leads to inferior performance on other datasets (r=0.64 MANHOB-HCI, r=0.69 MMSE-HR) compared to when it is trained on the same dataset (r=0.87 MANHOB-HCI, r=0.86 MMSE-HR). The representation bias of fewer older people and those with cardiac problems further affects the generalization of existing remote HR algorithms trained on younger and healthy individuals. Therefore, to make rPPG a viable solution in clinical settings, we recommend collecting equitable and representative data, and consider algorithmically mitigating potential model biases. 

Abstract 67: Semi-Automated Sensory Assessment in Spinal Cord Injury

Siti Nurfaezah Binti Zahari^1,2, Julio Furlan^2,3, Stephanie Iwasa², Shehroz Khan^1,2, Milos R Popovic^1,2

¹Institute of Biomedical Engineering, University of Toronto, Toronto, Toronto, Canada. ²The KITE Research Institute, Toronto, Canada. ³Department of Medicine, Division of Physical Medicine and Rehabilitation, University of Toronto, Toronto, Canada

There is a pressing need for precise metrics of outcomes in sensory deficit assessments for spinal cord injury (SCI) patients. Current clinical sensory evaluation methods, such as light touch and pinprick tests, are limited by accuracy, variability and low sensitivity to subtle sensory deficit changes in post-recovery patients. These limitations underscore the necessity for more precise and consistent assessments of SCI individuals. The electrical perceptual threshold (EPT) test addresses this limitation by utilizing non-painful electrical stimulation increased in intensity in small increments to quantify sensory perception. 

However, manual stimulus intensity adjustment in the EPT test can introduce variability depending on the test administrator. Our study addresses this bias by developing a computer-controlled EPT test (CCEPT) device. The CCEPT device ensures uniform stimulus delivery, objective recording of patient responses, and quick and easy administration.

We tested our CCEPT device on one healthy individual using a monophasic square waveform (0.5ms pulse width, 3Hz frequency). Unlike the manual EPT test, we employed a computer-controlled protocol to increase the current intensity by 0.1mA per step, from 0mA to 10mA, with a 1-second stimulation time per step. The participant pressed a push-button when they felt the sensation, and stimuli were applied three times over four dermatomes bilaterally (C3, T1, S2, and L3). The EPT (sensory) score was the mean intensity for which the participant perceived sensation, described as a light tapping or gentle pulsing without pain.

Initial findings suggest that the CCEPT device can replicate EPT assessments. Future studies will validate its applicability in clinical settings by testing on a larger cohort of healthy participants and individuals with SCI. This innovative approach addresses the limitations of current sensory assessment methods, aiming to provide more accurate outcome measures in SCI individuals. Additionally, it allows us to study how different test administrators might affect sensory assessments.

Abstract 68: A Novel Noise Reduction Framework for Electrocardiogram Signals Captured by Textile Electrodes

Luka Zigomanis¹, Barry Bytensky¹, Dhana Abdo¹, Holly Liu¹, Ryan Zhang¹, Delaram Sadatamin^2,3, Azadeh Yadollahi^2,3

¹Engineering Science, University of Toronto, Toronto, Canada. ²Institute of Biomedical Engineering, University of Toronto, Toronto, Canada. ³KITE, University Health Network Toronto Rehabilitation Institute, Toronto, Canada

Electrocardiogram (ECG) signals measure the electrical activity of the heart to diagnose a range of conditions. Textile-based wearable alternatives aim to more comfortably match the diagnostic capabilities of gold-standard gel electrodes, but are commonly degraded by noise. Therefore, we aimed to develop a signal processing framework to reduce noise in textile ECG measurements and reliably extract heart rates (HR) across various breathing types. Our solution framework was tested on a novel textile-based wearable, developed to continuously monitor ECG signals.

To evaluate our solution, textile and gel electrode ECGs were obtained during normal breathing, deep breathing, and breath holding. Textile ECGs were processed using our solution framework, as well as 4 baseline algorithms. As we were unable to capture gel and textile data simultaneously, we compared gel and textile ECG distributions using Z-score, standard deviation, Kullback-Leibler divergence, and percentage of physiologically viable heart rates.  

In our solution framework, ECG signals were split into 4.5-second segments and processed by an infinite impulse response notch filter. Subsequent bandpass and envelope filtering smoothened our final processed ECG signals. The Neurokit2 library was then used to extract HR. 

Our solution consistently produced ECG heartbeats and derived HR that most closely matched those obtained using gel electrodes. Our solution extracted HR that was 10x more similar to the reference data over Fourier preprocessing, the next best alternative. The framework also produced heartbeats that were at least 20% more uniform than any alternative. 

While our framework performed best when applied to data from normal breathing periods, further development is needed to ensure consistent performance during deep breathing, and breath holding. With further refinement through a larger sample size and simultaneous capturing of gel and textile data, a framework can be achieved for effective noise reduction and accurate biosignal extraction across a variety of individuals.