Enhancing Interpretable, Transparent, and Unobtrusive Detection of Acute Marijuana Intoxication in Natural...

Area of Science:

• Digital phenotyping
• Wearable technology
• Artificial intelligence in health

Background:

• Acute marijuana intoxication impairs cognitive and motor functions.
• Traditional detection methods (blood, urine, saliva) lack real-time accuracy.
• Need for unobtrusive, real-time methods to detect marijuana intoxication.

Purpose of the Study:

• To explore smartphone and wearable sensor integration for detecting marijuana intoxication.
• To assess passive sensing technologies for enhancing algorithm accuracy and interpretability.
• To understand digital device interaction during algorithmic decision-making for intoxication detection.

Main Methods:

• Collected smartphone and Fitbit data from 33 young adults over 30 days.
• Used experience sampling method with self-reported intoxication levels (0-10).
• Analyzed models using phone data, Fitbit data, and combined (MobiFit) data.

Main Results:

• The MobiFit model achieved 99% accuracy in detecting marijuana intoxication.
• Combined data significantly improved sensitivity and specificity compared to individual sources.
• Explainable AI identified elevated heart rate, reduced movement, and increased noise energy with intoxication.

Conclusions:

• Smartphone and wearable devices enable interpretable, unobtrusive monitoring of marijuana intoxication.
• Algorithmic decision-making offers insights for timely interventions to reduce harm.
• Future applications require clinical expert collaboration for enhanced practicality.