Development of an alcohol biosensor non-wear algorithm: laboratory-based machine learning and field-based deployment.

Wrist-worn alcohol biosensors can continuously track alcohol consumption, but their measurements are disrupted when the device is removed. Left unaddressed, non-wear data compromises observations of alcohol use and subsequent predictions of intoxication. To advance beyond commonly used temperature cutoffs and enable more precise detection of non-wear, we trained a random forest algorithm using laboratory ground truth data. Participants in Study One (N = 36) wore a wrist-worn alcohol biosensor (BACtrack Skyn) across 61 five-hour laboratory sessions, generating ground truth non-wear by removing and re-applying the device at specified times. Algorithm features included temperature, motion, and their time-series quadratic coefficients. According to device-based cross-validation, the algorithm performed with excellent sensitivity to detect non-wear (0.96) and specificity to confirm wear (0.99), out-performing all univariable temperature cutoffs from 25 to 30 °C. The algorithm was then used to evaluate biosensor adherence in Study Two, a four-week field study where participants (N = 114) wore the Skyn and self-reported non-wear intervals each day. The algorithm detected 1.6 h of daily non-wear per participant and had more agreement with self-report compared with the temperature cutoff method. This non-wear algorithm can assess biosensor adherence in field studies and may also facilitate precise data imputation, resulting in more objective models of alcohol-related outcomes.