Wearable health devices are widely adopted tools that continuously collect physiological data, and thus
make possible passive, continuous, data-driven assessments of people’s health. Algorithms can detect
acute illnesses, like COVID-19, by distinguishing physiological time series data that contains anomalous
patterns, often during sleep, from presumably healthy, stable baseline data. However, some individuals’
baseline states contain shifts and fluctuations that can look like anomalous patterns but are actually
dynamic characteristics of that individual’s baseline. Precision dropped substantially (-23.4% in AUC) for
individuals with dynamic baselines in COVID-19 detection experiments because anomaly detection
algorithms are designed to rely on the stability of baseline states. We used 5 million nights of sleep data
to investigate new approaches to modeling dynamic baselines and show our temporal model improves
separability by 4-10x across acute health conditions (COVID-19, flu, and fever). With this model, we
drastically recovered performance (+19.4% in AUC) with large reduction in false positive errors.
Modeling how people are dynamic over time is essential not only to identifying anomalous health states
but also to building robust health monitoring systems in the real world, where people are inherently
dynamic, and empowering individuals to take data-informed actions that meaningfully preserve their
health.