Mobile phone sensors and supervised machine learning to identify alcohol use events in young adults: Implications for just-in-time adaptive interventions

Abstract

Background: Real-time detection of drinking could improve timely delivery of interventions aimed at reducing alcohol consumption and alcohol-related injury, but existing detection methods are burdensome or impractical.

Objective: To evaluate whether phone sensor data and machine learning models are useful to detect alcohol use events, and to discuss implications of these results for just-in-time mobile interventions.

Methods: 38 non-treatment seeking young adult heavy drinkers downloaded AWARE app (which continuously collected mobile phone sensor data), and reported alcohol consumption (number of drinks, start/end time of prior day's drinking) for 28days. We tested various machine learning models using the 20 most informative sensor features to classify time periods as non-drinking, low-risk (1 to 3/4 drinks per occasion for women/men), and high-risk drinking (>4/5 drinks per occasion for women/men).

Results: Among 30 participants in the analyses, 207 non-drinking, 41 low-risk, and 45 high-risk drinking episodes were reported. A Random Forest model using 30-min windows with 1day of historical data performed best for detecting high-risk drinking, correctly classifying high-risk drinking windows 90.9% of the time. The most informative sensor features were related to time (i.e., day of week, time of day), movement (e.g., change in activities), device usage (e.g., screen duration), and communication (e.g., call duration, typing speed).

Conclusions: Preliminary evidence suggests that sensor data captured from mobile phones of young adults is useful in building accurate models to detect periods of high-risk drinking. Interventions using mobile phone sensor features could trigger delivery of a range of interventions to potentially improve effectiveness.

Keywords: AWARE app; Alcohol; Machine learning; Smartphone sensors.

Figure 1

Daily Query Response Rates Over 28 Study Days

Figure 2

Drinking Episodes by Day of Week

Figure 3. Cumulative Accuracy of Random Forest Models (30-minute window, 3-day historical data) Over 28 Days

X-axis represents number of days (maximum 28 days), Y-axis represents accuracy of classifying non-drinking, low-risk, and high-risk drinking. The Red and Blue lines represent results from Random Forest (30-minute window, 3-days of historical data) model, since it had the best overall performance in classifying non-drinking, low-risk, and high-risk drinking. Red line: only top-20 features were used for classification. Blue line: only 2 features, time of day and day of week, were used for classification. ZeroR model (dashed line) is a naïve model that just predicts the most frequent ‘N’ class. The graph depicts cumulative accuracy up to a given day, and not accuracy per day. Cumulative accuracy was determined by incrementally training models on successively larger sets of data. The figure shows higher classification accuracy (i.e., non-drinking, low-risk drinking, high-risk drinking) when using the top-20 features (red line) compared to the model using only time of day and day of week (blue line).