Machine Learning Prediction of Opioid Involvement in Connecticut Drug Overdose Deaths: Comparative Performance of Penalized Regression, Random Forest, and Support Vector Models

The opioid epidemic remains a critical public health challenge in the United States, with Connecticut experiencing particularly high rates of opioid involvement in accidental drug overdose deaths. This study evaluates the predictive capacity of machine learning models—Ridge regression, LASSO regression, Random Forest, and Support Vector Machine (SVM)—to classify opioid involvement among 11,979 de-identified overdose cases recorded between 2012 and 2021. After rigorous preprocessing, stratified train–test splits, and hyperparameter optimization, model performance was assessed using accuracy, precision, recall, F1, and calibration metrics. All models demonstrated strong predictive utility, with Random Forest and SVM offering slightly better discrimination compared to penalized regressions, though at the cost of interpretability. LASSO identified key demographic and geographic predictors, including place of death and race/ethnicity, providing actionable insights for public health interventions. Findings highlight the potential of machine learning to enhance surveillance, risk stratification, and targeted resource allocation in combating the opioid crisis in Connecticut.