Enhancing Interpretable, Transparent, and Unobtrusive Detection of Acute Marijuana Intoxication in Natural Environments: Harnessing Smart Devices and Explainable AI to Empower Just-In-Time Adaptive Interventions: Longitudinal Observational Study.

Abstract

Background: Acute marijuana intoxication can impair motor skills and cognitive functions such as attention and information processing. However, traditional tests, like blood, urine, and saliva, fail to accurately detect acute marijuana intoxication in real time.

Objective: This study aims to explore whether integrating smartphone-based sensors with readily accessible wearable activity trackers, like Fitbit, can enhance the detection of acute marijuana intoxication in naturalistic settings. No previous research has investigated the effectiveness of passive sensing technologies for enhancing algorithm accuracy or enhancing the interpretability of digital phenotyping through explainable artificial intelligence in real-life scenarios. This approach aims to provide insights into how individuals interact with digital devices during algorithmic decision-making, particularly for detecting moderate to intensive marijuana intoxication in real-world contexts.

Methods: Sensor data from smartphones and Fitbits, along with self-reported marijuana use, were collected from 33 young adults over a 30-day period using the experience sampling method. Participants rated their level of intoxication on a scale from 1 to 10 within 15 minutes of consuming marijuana and during 3 daily semirandom prompts. The ratings were categorized as not intoxicated (0), low (1-3), and moderate to intense intoxication (4-10). The study analyzed the performance of models using mobile phone data only, Fitbit data only, and a combination of both (MobiFit) in detecting acute marijuana intoxication.

Results: The eXtreme Gradient Boosting Machine classifier showed that the MobiFit model, which combines mobile phone and wearable device data, achieved 99% accuracy (area under the curve=0.99; F₁-score=0.85) in detecting acute marijuana intoxication in natural environments. The F₁-score indicated significant improvements in sensitivity and specificity for the combined MobiFit model compared to using mobile or Fitbit data alone. Explainable artificial intelligence revealed that moderate to intense self-reported marijuana intoxication was associated with specific smartphone and Fitbit metrics, including elevated minimum heart rate, reduced macromovement, and increased noise energy around participants.

Conclusions: This study demonstrates the potential of using smartphone sensors and wearable devices for interpretable, transparent, and unobtrusive monitoring of acute marijuana intoxication in daily life. Advanced algorithmic decision-making provides valuable insight into behavioral, physiological, and environmental factors that could support timely interventions to reduce marijuana-related harm. Future real-world applications of these algorithms should be evaluated in collaboration with clinical experts to enhance their practicality and effectiveness.