Machine learning prediction of blood alcohol content: A digital signature of behavior
Some people use digital devices to track their blood alcohol content (BAC). A BAC-tracking app that could anticipate when a person is likely to have a high BAC could offer coaching in a time of need. Kirstin Aschbacher shares a machine learning approach that predicts user BAC levels with good precision based on minimal information, thereby enabling targeted interventions.
Talk Title | Machine learning prediction of blood alcohol content: A digital signature of behavior |
Speakers | Kirstin Aschbacher (UCSF Cardiology) |
Conference | Strata Data Conference |
Conf Tag | Big Data Expo |
Location | San Francisco, California |
Date | March 26-28, 2019 |
URL | Talk Page |
Slides | Talk Slides |
Video | |
Individuals can track their blood alcohol content (BAC) using commercially available devices, providing an opportunity for behavioral health interventions such as personalized messaging to target specific users at high-risk times or locations. To do so, a machine learning (ML) model would need to predict user BAC levels with high precision based on minimal information, such as timestamps, geolocation, and device/app engagement. Kirstin Aschbacher shares a machine learning approach to identify a digital signature of self-monitored BAC levels that predicts the times, locations, and circumstances under which a user is likely to exceed the legal BAC driving limit of 0.08%. Kristin and her teammates analyzed over 1 million data points from 33,452 distinct users of the BACtrack device (with established accuracy comparable to police-grade devices) collected between 2013 and 2017. Extensive feature generation was performed on BAC levels, app engagement, timestamps, and geolocation. They used census data to quantify zip codes by % rural/urban and integrated state-level motor vehicle death rates from the Center for Disease Control. Feature selection was performed using a gradient-boosted classification tree model (XGBoost; learning rate=0.1, max depth=5; boosting rounds=30). They optimized around precision specifically because commercial devices that employ ML-driven strategies to reach out to users should consider that recommendations made on the basis of low precision (predicting a user has a high BAC when they do not) could harm product trust or engagement. In a separate test set, they predicted whether BAC≥0.08% for a given user at a given time and location, with an average precision (positive predictive value) of 79%. The most predictive features in rough order of importance were users’ prior behavior (average BAC, subjective estimation of their BAC, tracking frequency, engagement quantity), temporal features (time of day/day of week), and geolocation (elevation, distances traveled between subsequent measurements, country, rural/urban percentage). Join in to learn how BAC levels exceeding the safe legal driving limit of 0.08% can be predicted with good precision using machine learning to quantify a digital phenotype. BAC prediction from minimal information establishes the foundation to conduct precision medicine behavioral interventions using a digital app and BAC tracking device. Kirstin wants to acknowledge and thank her UCSF coauthors on this work: R. Avram, G. Tison, K. Rutledge, M. Pletcher, J. Olgin, and G. Marcus.