AN EXPLAINABLE MACHINE LEARNING FRAMEWORK FOR EARLY DRUG-ABUSE RISK ASSESSMENT AND STRATIFICATION
: https://doi.org/10.65725/JCISE/2/4/006
JOURNAL OF COMPUTATIONAL INTELLIGENCE SCIENCE AND ENGINEERING (JCISE) 
ISSN: 3107-8168
Volume 2 Issue 4, Oct – Dec 2026

Abstract: 
Drug abuse and substance use disorders are major public-health concerns that require effective approaches for the early identification of elevated risk and for preventive intervention. Existing risk-assessment approaches often rely on single risk factors, specific populations, clinical evaluations, or isolated psychological and behavioral indicators and they frequently lack interpretable, multidimensional risk stratification. This study addresses this gap by proposing an explainable machine-learning framework that integrates psychological, behavioural, social, familial, lifestyle, academic/work, and exposure-related predictors for early drug-abuse risk assessment. The dataset comprises approximately 5,000 observations; following data-quality review, leakage assessment, and feature selection, 35 multidimensional predictors were retained. Six supervised machine-learning algorithms, namely Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, XGBoost, and LightGBM, were compared for three-level risk classification into Low, Moderate, and High Risk, using leakage-aware feature selection, pipeline-based preprocessing, and an 80:20 train-test split. SHAP-based explainability was applied to identify which variables contributed to model predictions. In the held-out evaluation, Logistic Regression achieved the highest accuracy (70.5%), Macro F1-score (70.0%), and ROC-AUC (0.863) and was selected as the final model. The framework is intended to support early identification of elevated risk among individuals assessed at a given time, including those who may already have a history of substance exposure, rather than to predict first exposure or to provide a clinical diagnosis. SHAP findings are presented as model-based feature attribution, not as causal evidence. The results provide a basis for external validation, fairness and calibration analysis, longitudinal evaluation, and responsible deployment in drug-abuse prevention and public-health settings.

Authors:  Dr. A. Maria Vinitha., P. M. Jesu Melwin

Keywords: Drug Abuse, Risk Assessment, Early Warning System, Machine Learning, Explainable AI, Risk Stratification.