Abstract:

Wearables provide a valuable new lens on health by enabling longitudinal monitoring of physiology. Core technologies include photoplethysmography (PPG), electrocardiography (ECG) and electrodermal activity (EDA) sensing, which measure cardiovascular and autonomic nervous system function. These data can facilitate new longitudinal insights about health and disease to complement existing cross-sectional evidence in medicine. They can also enable various medical applications, such as remote patient monitoring for disease progression and acute event detection.

Given its strengths in pattern recognition, machine learning is a powerful framework for processing physiological data. However, these data also present two substantial challenges that limit the utility of standard machine learning techniques. First, while the proliferation of measurement devices has created an abundance of physiological records, it is expensive to obtain structured annotations given that this requires human effort. As such, most records are unlabelled, prohibiting the use of supervised learning at the full scale of the data. Second, physiological signals exhibit strong between-individual variation, not all of which is relevant clinically. Without careful model design, this variation can dominate the learning process, obscuring the subtler within-individual changes that often carry critical health-related information.

I present three projects to address these challenges. I first consider depression monitoring in a well-characterised cohort of patients with major depressive disorder. Because both depression and its physiological correlates vary strongly between individuals, I develop a mixed-effects machine learning framework that adjusts for this variation when using wearable features to assess severity. Second, to confront the sparsity of labels for ECG records, I propose a multimodal contrastive learning method where ECG waveforms are concurrently aligned to multiple types of unstructured clinical notes that occur sparsely and asynchronously in a patient's medical record. This produces an ECG foundation model that learns strong ECG representations that support the accurate detection of many cardiovascular, pulmonary and metabolic conditions. It outperforms prior foundation models on over 100 disease and irregularity detection tasks, and transfers well to reduced-lead configurations common in wearable devices. Finally, I present a PPG foundation model that represents both the within-individual and between-individual variance in the PPG signal. It performs strongly across a spectrum of downstream tasks, ranging from the detection of chronic or slowly-varying conditions such as hypertension, to rapidly-changing states such as sleep stages.

This thesis makes several novel contributions. First, I demonstrate that actively managing between-individual differences in self-supervised and supervised learning improves representation quality and detection accuracy. Second, I show that pre-training physiological signal models by aligning to several modalities in parallel results in stronger and more universal representations that support a wider range of downstream tasks than aligning to any of these modalities in isolation. Finally, I contribute two state-of-the-art foundation models for ECG and PPG that are validated using broader benchmarks than prior work, matched to representative use cases. These models can serve as a bedrock for future research and applications in the field of wearable health.

Rosalind Picard
Professor of Media Arts and Sciences
MIT Media Lab

John Guttag
Professor
MIT CSAIL

Nicholas Cummins
Senior Lecturer
Department of Biostatistics and Health Informatics at King’s College London