Projects

Our deep learning fusion model integrates preoperative ECG waveforms clinical data to outperform traditional risk scores in predicting major postoperative outcomes. By using counterfactual ECGs, it offers interpretable insights, linking waveform features to risk. This innovation paves the way for personalized, actionable interventions in surgical care.

Have you ever noticed that when we grasp objects, one finger almost always makes contact before the others? Perhaps not, because this millisecond-scale gap is negligible to human perception. However, prosthetics and robotics can certainly take advantage of it. We used vibrations from this initial contact window to estimate object stiffness, enabling potential grip modulation by the other fingers.

This work builds on another project, listed below, "Optimizing cross-modal matching for multimodal motor rehabilitation."

Cross-modal matching (CMM) is widely used to calibrate sensory feedback across modalities, but it is typically assumed to be bidirectional and invertible. We test this assumption using visual–vibrotactile and co-located haptic (pressure–vibration) datasets and show that it does not hold. At the individual level, mappings are systematically non-invertible, with directional asymmetries arising from structured, intensity-dependent distortions rather than noise. As a result, CMM compresses the perceptual range of one modality relative to another, making the choice of mapping direction consequential. These findings show that CMM is not a neutral calibration step but a directional design decision that shapes perception and learning in human–machine systems.

This work builds on another project, listed below, "Optimizing cross-modal matching for multimodal motor rehabilitation."

After a stroke or neurological injury, many people struggle not just with movement, but also with sensing and controlling their hands. Rehab tools often mix visual, touch, and sound cues to help retrain the brain – but these cues need to be carefully balanced so one sensation doesn't overpower the others. That's where cross-modal matching comes in: a way to "calibrate" perception across senses. We built SenseMatch, a smartphone app that makes this calibration process easier and more accessible. In our study, we compared SenseMatch to a clinical rehab tool device. The long-term vision is big: with tools like SenseMatch, we can move toward objective, personalized measures of sensory deficits, helping to guide rehabilitation and track recovery beyond the clinic.

After a stroke or injury, many people struggle not just with movement, but also with sensing and controlling their hands. Rehab tools often mix visual, touch, and sound cues to help retrain the brain – but these cues need to be carefully balanced so one sensation doesn't overpower the others. That's where cross-modal matching comes in: a way to "calibrate" perception across senses. This project calibrates visual and haptic feedback stimuli through cross-modal matching to establish perceptual equity across the feedback channels. By aligning subjective intensities, we isolate true performance differences from perceptual confounds. Using statistical modeling and Monte Carlo optimization, we develop a streamlined calibration protocol that cuts session time by over 5×, enabling practical clinical use.

We are designing an experiment to explore how different forms of feedback — visual, haptic, and combined — affect motor performance in a precision grip task. Participants used isometric force to “pinch” virtual objects while receiving real-time feedback through either vision, vibration, or both. The study showed that while visual feedback alone often led to high success rates, combining it haptic feedback offered perceptual and learning benefits in specific contexts. These findings underscore the importance of tailoring multimodal feedback to user needs and task demands, especially in the design of next-generation rehabilitation tools for stroke recovery and sensorimotor training.

In neuroimaging, skull stripping is vital for removing non-brain tissue from scans, improving accuracy in analysis, segmentation, and 3D modeling. While traditional methods are dependable, they struggle large, multi-scanner datasets. Deep learning, powered by U-Net architectures, provides a faster, more efficient solution, robust against multi-scanner variability.

The feeling of being in control of our actions, known as the Sense of Agency, is a fascinating aspect of human experience. This sensation, rooted in the brain's response to the consequences of our actions, underpins our perception of autonomy and active engagement in the world. Neuroscience seeks to unravel the complexities behind this phenomenon, offering insights that span philosophy, psychology, robotics, and AI. In this project, we delve into how the brain generates this sense of control, examining the interplay of cognitive processes and sensory feedback.

Wild blueberries, loved for their bold flavors, grow close to the ground in rocky terrains, posing challenges for automated harvesting setups. To address this, I aided in the development of automated harvester-head height adjusting, and a camera-vision-based system to automate berry volume measurement and bin switching to reduce operator workload.

This project was funded by Mitacs Globalink, a program supporting top undergraduates from 17 countries for fully funded summer research in Canada.