Absolute risk prediction for substance use disorders using statistical and machine learning

Abstract  Substance use disorders (SUDs), particularly alcohol use disorder (AUD) and cannabis use disorder (CUD), are major public health concerns in the United States (US). Early substance use during adolescence is a strong precursor of SUDs in adulthood. Identifying individuals at high risk before the onset of problematic use provides a critical window of opportunity for timely prevention and intervention strategies. To address this need, we developed statistical and machine learning based absolute risk prediction models that estimate an individual’s personalized probability of developing AUD or CUD within a given time frame. Such models are widely used for diseases like cancer and heart disease, but are largely missing for SUDs. We propose three models: • A Bayesian machine learning model to predict CUD risk among cannabis users. • A joint Bayesian model that predicts both AUD and CUD while accounting for their correlation through a shared frailty structure. • A deep learning model to predict AUD risk among alcohol users. These models are trained and validated using data from the National Longitudinal Study of Adolescent to Adult Health (Add Health), which followed over 20,000 US adolescents for more than two decades. The Bayesian models are further validated using an external dataset from the Christchurch Health and Development Study in New Zealand. Key predictors include demographics, delinquent behavior, peer influence, and personality traits such as conscientiousness and neuroticism. Overall, these models provide flexible and interpretable tools for personalized, time-specific risk prediction of SUDs, supporting early identification of high-risk individuals and enabling targeted prevention strategies. 

I will also share my experiences and lessons learned from the job search process.

Slides

Photos

Participants: 20