Wearable Missingness as a Behavioral Signal: Detecting Wear Pattern Anomalies Around Cardiovascular Visits

Missing data in wearable device streams is typically treated as noise and is imputed or discarded. We argue that when and how a patient stops wearing their device carries clinically meaningful behavioral information, and propose a personalized anomaly detection framework that models each individual's day-type-specific wear patterns and flags deviations. Applying this framework to Fitbit data from 390 participants (756 cardiovascular emergency visits) in the NIH All of Us Research Program linked with electronic health record (EHR) cardiovascular emergency visits, we find that wear pattern anomalies rise to $1.45\times$ baseline in the week before a visit and to $2.00\times$ baseline in the week after. Critically, aggregate missingness rates show no statistically significant change between the baseline and around visits, indicating that the anomaly detector captures temporal pattern shifts that summary statistics miss.