Machine learning models identify key predictors of driving under the influence of alcohol or cannabis

The frequency of substance use, early age of initiation, and cannabis-related memory impairments are among the primary factors contributing to driving under the influence, according to a new analysis using machine learning. Impaired driving is known to be influenced by a range of behavioral, demographic, and contextual factors.

Most research has explored several of these at a time—alcohol use patterns, socioeconomic status, peer influences, and so on—in analyses that cannot fully capture the multidimensional and interconnected array of influences.

For the study in Alcohol: Clinical & Experimental Research, investigators used two machine learning algorithms to identify the most salient predictive factors.

Machine learning approaches have advantages over traditional analytic methods. In this study, regularized regression allowed for a more nuanced understanding of the variables influencing DUI, while random forests accounted for complex interactions.

Researchers worked with data from young adults in Washington state who filled out comprehensive health surveys between 2015 and 2022. The investigators included 79 candidate variables potentially relating to DUIA (alcohol) and 88 for DUIC (cannabis), including demographics, social roles, living situations, beliefs, and substance use behaviors.

Among almost 10,000 participants with recent alcohol use, 18% reported DUIA in the past month, and 15% reported DUIC. For alcohol-impaired driving, both machine learning models found drinking frequency the strongest predictor. Other top factors included age, age at drinking initiation, and maximum number of drinks on a recent occasion.

In the same group, cannabis-use frequency was linked to a lower likelihood of DUIA, perhaps pointing to participants substituting cannabis for alcohol and perceiving DUIC as less dangerous than DUIA.

Among 5,000 participants with recent cannabis use, 37% reported past-month DUIC and 16% reported DUIA. Frequency of cannabis use was the strongest predictor of DUIC. Other influential factors were cannabis-related memory problems, using alcohol and cannabis on the same occasion, and the age of initiating cannabis use. The findings suggest that people reporting DUIC may be at elevated risk of cannabis use disorder.

The study demonstrated the ability of machine learning to analyze multiple risk domains, facilitating a more comprehensive understanding of DUI. Using two machine learning approaches helped researchers leverage the strengths of both while mitigating the limitations of each. The findings were consistent and helped flag risk factors to prioritize in substance use interventions and treatments.