A wearable alcohol biosensor: Exploring the accuracy of transdermal drinking detection.

Abstract

Background: Trace amounts of consumed alcohol are detectable within sweat and insensible perspiration. However, the relationship between ingested and transdermally emitted alcohol is complex, varying across environmental conditions and involving a degree of lag. As such, the feasibility of real-time drinking detection across diverse environments has been unclear. In the current research we revisit sensor performance using new tools, exploring the accuracy of a new generation of rapid-sampling transdermal biosensor for contemporaneous drinking detection across diverse environments via machine learning.

Methods: Regular drinkers (N = 100) attended three laboratory sessions involving the experimental manipulation of alcohol dose, rate of consumption, and environmental dosing conditions. Participants further supplied breath alcohol concentration (BAC) readings in the field over 14 days. Participants wore compact wrist sensors capable of rapid sampling (20sec intervals). Transdermal sensor data was translated into alcohol use estimates using machine learning, integrating only transdermal data collected prior to the point of BAC assessment.

Results: A total of 5.39 million transdermal readings (28,615hours) and 12,699 BAC readings were collected for this research. Models indicated strong transdermal sensor accuracy for real-time drinking detection across both laboratory and field contexts (AUROC, 0.966, 95 % CI, 0.956-0.972; Sensitivity, 89.8 %; Specificity, 90.6 %). Models aimed at differentiating high-risk (≥0.08 %) drinking levels yielded intermediate (AUROC, 0.738; 95 % CI, 0.698-0.777; only drinking episodes) to strong (AUROC, 0.941, 95 % CI, 0.929-0.954; all data) accuracy levels.

Conclusions: Results indicate a range of useful future applications for transdermal alcohol sensors including long-term health tracking, medical monitoring, and just-in-time relapse prevention.