Smartphone Inference of Alcohol Consumption Levels from Gait

Excessive alcohol use is the third leading lifestyle-related cause of death in the United States. Smartphone sensing offers an opportunity to passively track alcohol usage and record associated drinking contexts. Drinkers can reflect on their drinking logs, detect patterns of abuse and self-correct or seek treatment. In this paper, we investigate whether a smartphone user's alcohol intoxication level ( how many drinks) can be inferred from their gait. Accelerometer data was gathered from the smartphones of a group of drinkers. Time and frequency domain features were then extracted and used for classification in a machine learning framework. Various classifiers were compared for a task of classifying the number of drinks consumed by a user into ranges of 0-2 drinks ( sober), 3-6 drinks ( tipsy) or > 6 drinks ( drunk). Random Forest proved to be the most accurate classifier, yielding 56% accuracy on the training set, and 70% accuracy on the validation set. Using these results, AlcoGait, an Android smartphone application was developed and evaluated by real users. The results of user studies were encouraging.