Development of an accelerometer-based wearable sensor approach for alcohol consumption detection.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Public Health

Background:

• Alcohol use is a significant public health concern with limited, stigmatized treatment options.
• Wearable accelerometer sensors offer a potential for accessible and affordable alcohol use detection.
• Innovations in alcohol treatment are needed due to access and affordability barriers.

Purpose of the Study:

• To compare distributional and random forest classification algorithms for detecting alcohol consumption using smartwatch data.
• To evaluate the accuracy and sensitivity of these algorithms in identifying drinking behavior.
• To assess the feasibility of using consumer-grade smartwatches for objective alcohol use monitoring.

Main Methods:

• Participants (n=194) wore an Android smartwatch at a state fair, consuming water at intervals.
• Confounding behaviors (e.g., touching face) were interspersed with drinking to test algorithm specificity.
• Data analysis involved comparing distributional and random forest classification models.

Main Results:

• The distributional algorithm achieved 95% accuracy and higher sensitivity (76%) compared to the random forest model (93% accuracy, 32% sensitivity).
• The distributional algorithm demonstrated significant equivalency to ground truth for sip duration and between-sip intervals.
• The random forest model did not show significant equivalency for between-sip intervals.

Conclusions:

• Consumer-grade smartwatches, combined with machine learning and distributional algorithms, can effectively detect and measure alcohol use.
• This research provides a methodological basis for studying alcohol's behavioral pharmacology.
• The findings support the development of accessible, just-in-time clinical interventions for alcohol use.