Towards Accessible, Equitable, Generalizable and Useful Camera Health Sensing

Abstract

The COVID-19 pandemic has prompted a shift in the delivery of healthcare globally, with a growing emphasis on scalable health sensing. Currently, biomedical contact sensors are considered the gold standard for measuring vital signals, but they are not widely accessible, particularly in under-resourced areas. Camera-based health sensing offers the potential to reach a wider population by using regular RGB cameras to detect changes in electromagnetic radiation (light) reflected from the body that result from physiological processes. However, existing camera-based health sensing methods are inaccessible due to their high computational costs, inequitable due to poor generalizability across skin tones, lighting, and movements, and not fully validated for use in clinical settings. To address these challenges, this thesis explores the development of on-device neural networks, few-shot adaptation, federated learning, and data augmentation systems and algorithms for camera-based health sensing. A transnational clinical study is also conducted to evaluate the usefulness of these methods in real-world clinical settings and to advance the field of camera-based health sensing beyond well-studied physiological signals. Finally, this research introduces an open-source toolbox to promote reproducibility and fair benchmarking comparisons.