Multi-Level Risk of Polysubstance use Trajectory and Leveraging Machine Learning Approaches to Detect Overdose Risk

Columbia University Health Sciences, New York NY

Investigators

Abstract

PROJECT SUMMARY/ABSTRACT Significance. Recent national drug-involved overdose death data has shown that the current opioid epidemic is characterized by four salient waves that claimed over 106,000 lives in 2021 alone. The current fourth wave of the epidemic is characterized as an opioid-stimulant polysubstance use. However, it is still unclear how individuals transition through different substance use profiles, the individual and structural determinants that are associated with these transitions, and polysubstance use profiles that are at increased risk of overdose. Additionally, no study to date has used a multilevel approach to develop a machine learning risk prediction model to predict overdose risk. Career Development Plan. Dr. Cadet’s training program will include seminars, workshops, coursework, and conferences to develop her skills and expertise in Multilevel Latent Markov Modeling, Bayesian Inference, longitudinal causal modeling, and in machine learning risk prediction modeling, which are necessary for conducting her proposed research plan and achieving her career goals of becoming an independent substance use epidemiology scientist who conducts large observational studies that will inform the targeted public health intervention and polysubstance use related overdose prevention. Mentorship. A highly accomplished team of mentors (Drs. Martins, Musci, Stingone, Tabb, Aiello) who are experts in substance use epidemiology, Bayesian Inference, longitudinal structural equation modeling, biostatistics, and machine learning, will support Dr. Cadet’s research and training goals. Research Plan. Dr. Cadet will conduct a multilevel epidemiological study informed by the Risk Environment Model that leverages the size and scope of the National Longitudinal Study of Adolescent to Adult Health (Add Health) from 1994-2026 (N> 20,000 participants) to answer the following aims: 1) Examine the dynamic transitions across substance use patterns using Multilevel Latent Markov Models (Latent Transition Analysis) among people who used drugs enrolled in the Add Health study; 2) Examine whether polysubstance use typologies have poorer survival risk of fatal overdose by using latent class modeling with a time-to-event distal outcome joint-model approach to explore latent polysubstance use classes with higher risk of fatal overdose and all-cause mortality in the Add Health study; 3) Develop a multilevel machine learning algorithms using a train-test split procedure to predict people who use drugs (PWUD) who are at increased risk of overdose episode by using micro-level (e.g., behavioral, psychosocial) and macro-level (e.g., area deprivation) factors in the Add Health study. Findings from this study will inform an R01 to lead a mixed-methods approach to better understand the nuances between intentional polysubstance use and treatment implications. Public Health Impact. The cross-disciplinary methodologies such as combining structural equation modeling, Bayesian modeling, and data science is crucial for addressing the multifaceted challenges of polysubstance use effectively.