UC San Diego

Most people today occasionally visit a clinic or hospital, leading to gaps in time between assessments of a person’s health. These gaps could be filled by wearable devices measuring physiological data that use anomaly detection algorithms to identify health conditions like COVID-19. However, anomaly detection algorithms raise too many false alarms in real-world data, preventing wearables from being precise tools for continuous health monitoring. I hypothesized that the high false alarm rate of anomaly detection algorithms stems from these algorithms’ assumption that physiological data before illness is static compared to data during illness. If these algorithms could treat baseline data as dynamic, they may be able to identify illness-related deviations more precisely. In this work, I test this hypothesis and show that to improve the precision of anomaly detection algorithms, I must represent individuals’ physiological data as non-static by encoding patterns of an individual that vary with time. First, I represent patterns of sleep quality with 5 million nights of wearable data, showing these temporal representations improve 2-10x separability of various health conditions and substantially reduce the false alarm rate of COVID-19 detection algorithms. Then, I represent skin temperature patterns using wearable data from 11,000 people and combine these with the temporal representations of sleep to enable noninvasive detection of people with diabetes. This work demonstrates that we can improve the precision of anomaly detection algorithms by representing peoples’ physiological data as non-static. My work is one of the first to use insights from large-scale, longitudinal data to design novel wearable health algorithms. The longitude of this data made it possible for us to represent the temporal structure of a person’s baseline physiology in a data-driven manner, and the breadth showed how widely the structure of baseline physiology varies across people. With the broad adoption of commercially available wearable health devices and the implementation of detection algorithms on these devices, temporal representations of people’s health will not only improve precision when detecting health conditions but also enable people with highly variable lifestyles, those who may need continuous health monitoring most, to benefit from advancements in health detection technology equitably.