Towards Automated, Interpretable and Unobtrusive Detection of Acute Marijuana Intoxication in the Natural Environment: Harnessing Smartphones, Wearables, Machine Learning and Explainable AI to Empower Clinical Decision Support for Just-In-Time Adaptive Interventions (Preprint)

Abstract

Acute marijuana intoxication can impair motor skills and cognitive functions (e.g., attention, information processing). However, existing tools (e.g., blood, urine, saliva tests) do not accurately reflect ‘real-time’ acute marijuana intoxication.

Considering the absence of screening tools to detect acute marijuana intoxication and impairment-related harms, our objective is to examine whether integration of smartphone-based sensors with a wearable activity tracker (Fitbit), as more accessible devices using passive sensing, can enhance detection of episodes of acute marijuana intoxication in real-world settings. No prior work has determined the potential of utilizing data from both phone sensors and a wearable device to improve the accuracy of algorithms in detecting acute marijuana intoxication in real-life scenarios (‘outside of lab settings’), nor focused on developing explainable AI (XAI) to provide insights into the algorithmic decision-making process, specifically in detecting episodes of moderate-intensive marijuana intoxication, leveraging passive sensing technologies captured in real-world contexts.

To address these aims, we collected daily data using the Experience Sampling Method (ESM) for up to 30 days from 33 young adults using personal smartphone sensors and a Fitbit, and self-reported marijuana use. Participants provided subjective ratings of marijuana intoxication within 15 min of starting to use marijuana and during semi-random prompts 3 times per day: “low-intoxication” (rating = 1–3) vs “moderate-intensive intoxication” (rating = 4–10) vs. “not-intoxicated” (rating = 0).

Using the EXtreme Gradient Boosting Machine classifier (XGBoost) to model this data, our results indicated that the best model (MobiFit-model), which combined data from off-the-shelf mobile phone and wearable technologies, achieved accuracy of 99% (AUC=0.99, F1-score =0.85) in detecting acute marijuana intoxication (i.e., subjective sense of intoxication) in the natural environment. F1-score, which balances sensitivity and specificity, showed a significant improvement of 13% and 11% for the combined model (MobiFit) compared to using Mobile and Fitbit individually, respectively. Explainable AI (XAI) presented algorithmic decisions which revealed that self-reported moderate-intensive marijuana intoxication was associated with smartphone sensors and Fitbit features, specifically: elevated minimum heart rate, increased micro-movements, but reduced macro-movement (i.e., a smaller radius of gyration via GPS), and increased noise energy level around the participants.

This study demonstrates the promise that mobile phone sensors and off-the-shelf wearable devices hold for automated and continuous detection of acute marijuana intoxication in daily life. Advanced algorithmic decision-making processes could provide insight into behavioral, physiological and environmental features’ contributions that may be most useful, for example, in triggering the delivery of just-in-time interventions to prevent marijuana-related harm; however, in order to make the algorithm applicable in real-world settings, the usefulness and effectiveness of such algorithms-driven decisions need to undergo robust evaluation in collaboration with clinical experts.