Building risk prediction models for daily use of marijuana using machine learning techniques

Area of Science:

• Public Health
• Data Science
• Behavioral Science

Background:

• Standard statistical methods struggle to identify characteristics of adults with recent marijuana use.
• A unique approach is needed to understand patterns of daily marijuana consumption.

Purpose of the Study:

• To evaluate machine learning models for predicting daily marijuana use.
• To identify key factors associated with daily marijuana use among adults.

Main Methods:

• Utilized pooled data from the 2016-2019 Behavioral Risk Factor Surveillance System (BRFSS) Survey.
• Developed and compared prediction models using Logistic Regression, Decision Tree, Random Forest, and Naïve Bayes algorithms.
• Assessed model performance using accuracy, AUC, precision, and recall on training and testing samples.

Main Results:

• The study included 253,569 respondents, with 5.9% reporting daily marijuana use.
• Random Forest (AUC 0.97) and Decision Tree (AUC 0.95) demonstrated superior predictive performance.
• Significant factors for daily use included e-cigarette/combustible cigarette use, male gender, unmarried status, poor mental health, depression, cognitive decline, abnormal sleep, and high-risk behaviors.

Conclusions:

• Machine learning models are efficient for predicting daily marijuana use and identifying associated risk factors.
• Data mining techniques reveal crucial insights into behavioral health risks from complex surveys.
• Findings can inform targeted public health interventions for daily marijuana users.