AlcoWear: Detecting blood alcohol levels from wearables

Alcohol abuse causes 88,000 deaths annually. Alcohol affects neuromotor functions such as walk patterns, making them a reliable bio-measure of intoxication. In this paper, we present AlcoWear, a machine learning based system that passively senses a drinker's Blood Alcohol Content (BAC) by classifying accelerometer and gyroscope data gathered from their smartphone and smartwatch. While gait sensor readings taken from a device attached to the user's trunk (smartphone) are the most accurate, users often do not carry their phones (e.g. leave them on a table) while walking around during their day. Smartwatches are worn continuously but are less accurate due to noisier sensor readings (e.g. confounding hand gestures). AlcoWear extracts and classifies features such as sway area (gyroscope) and cadence (accelerometer) from smartphone sensor data, and features such as Total Harmonic Distortion (accelerometer) and angular velocity (gyroscope) from the same users' smartwatches. On the smartphone, the J48 classifier was the most accurate, classifying user gait patterns into BAC ranges of [0.00-0.08), [0.08-0.15), [0.15-0.25), [0.25+) with an accuracy of 89.45%. For the smartwatch, AlcoWear classifies users as being in BAC ranges [<; 0.08 vs >= 0.08] (2 bins) with a 79.8% accuracy using a Random Forest classifier. AlcoWear classifies the smartphone's data when the user carries it, and uses data from the user's smartwatch when the user is not carrying their phone. AlcoWear is the first to combine the smartphone and smartwatch in a collaborative system to sense BAC from gait.