Clinically-interpretable and large-scale machine learning to monitor mood disorders with wearables

Abstract

Mood Disorders (MDs) are common and severe psychiatric conditions with a relapsing-remitting course. Timely intervention during impending episodes improves outcomes, but pre-emptive measures are limited due to infrequent patient reviews and limited symptom reporting from patients. MDs involve changes in energy levels, circadian rhythms, and neurovegetative functions, correlating with changes in physiological data, like acceleration and galvanic skin conductance. Personal sensing, leveraging data from wearables, offers a way to monitor MDs remotely with objective biomarkers, which psychiatry currently lacks, relying mainly on clinical observation and patient self-reports. AI can harness wearable data to realise remote monitoring. I outline a unifying perspective on personal sensing for MDs and make original contributions to the field, using a prospective, observational cohort (TIMEBASE/INTREPIBD), recorded with an Emaptica E4 device.

Our first contribution focuses on Heart Rate Variability (HRV), an indicator of the autonomic nervous system functionality. I find HRV increases as symptoms subside after acute episodes, suggesting it as a potential symptom improvement biomarker. Due to limited HRV study samples, I use Bayesian statistics to propose an interpretable probabilistic model, explaining the HRV data generating process. Longitudinal HRV data collection is indeed labour-intensive, often resulting in small samples that undermine frequentist statistics reliability.

Personal sensing research typically attempted to detect the mere presence of acute episodes or the total score on a psychometric scale, missing actionable clinical information. I propose inferring all symptoms from two popular scales assessing the full MD symptom spectrum, akin to a concept bottleneck. This approach ensures AI output is interpretable, recognizing that different symptom combinations require varied therapeutic strategies. I developed a model for this task and investigate key AI challenges.

Lastly, to address labelled data scarcity in AI systems for personal sensing, I gather open-access datasets using the E4 wearable, regardless of the task they are concerned with, and make such collection publicly available. I propose a Transformer model tailored to the E4 and show that self-supervised learning, repurposing unlabelled data to learn useful representations through surrogate tasks, is viable in personal sensing. This method outperforms fully supervised models, whether using deep learning or classical machine learning with hand-crafted features.