Project Leads: Soshi Samejima, UW Rehabilitation Medicine. Joanna Zhou, Arya Kuttiyan, and Ellie Bieler
Data Scientist: Valentina Staneva
Cardiovascular disease is a leading health concern for individuals living with a spinal cord injury (SCI), making it critical to monitor cardiorespiratory fitness during rehabilitation and exercise. Clinically, the gold standard for measuring aerobic fitness is tracking oxygen consumption (VO2) during a cardiopulmonary exercise test. However, this testing requires expensive metabolic equipment and wearing a breathing mask during strenuous activity, which creates a significant burden for patients and limits accessibility in everyday clinical and home settings.
Our project explores whether machine learning can provide a low-burden, non-invasive alternative by estimating oxygen uptake in real time using heart rate variability (HRV) derived from standard electrocardiogram (ECG) recordings, respiratory rate, and basic demographic information. We evaluated several regression models—including Linear Regression, Random Forest, Support Vector Regression, and XGBoost—across both an able-bodied cohort and adults with chronic SCI performing graded arm-bike exercise. While nonlinear models like XGBoost accurately tracked metabolic demand in able-bodied individuals, predicting oxygen consumption in people with SCI proved more complex due to injury-related changes in autonomic nervous system control. Incorporating SCI-specific clinical factors, such as time since injury and neurological level of injury, significantly improved model performance, highlighting the importance of personalized, injury-informed algorithms for accessible fitness monitoring in rehabilitation populations.

