A machine learning approach with SHAP interpretability for classifying drug craving levels

Background: Drug addiction is a chronic relapsing brain disease. Drug craving is the strongest independent predictor of relapse. However, traditional linear models often fail to capture complex non-linear addiction patterns. Although machine learning (ML) demonstrates superior performance, its "black-box" nature limits clinical credibility. Objective: This study aimed to construct a classification model for drug craving scores based on ML and explore the impact of multifactorial features. Methods: A total of 692 abstainers were recruited from Compulsory Isolation Drug Rehabilitation Centers. Craving was assessed using a 34-item Drug Craving Scale. After preprocessing (imputing missing values, removing outliers via IQR), 629 valid samples remained. Eighteen demographic and behavioral features were analyzed. Samples were split 7:3 into training and test sets. SMOTE was employed for class imbalance. Seven algorithms-including Logistic Regression, XGBoost, and LightGBM-were compared. The optimal model was selected using 10-fold cross-validation and grid search, then evaluated on the independent test set using multi-dimensional metrics. SHAP was introduced for interpretability. Results: Logistic Regression was the optimal model. On the independent test set, it achieved 66.13% accuracy and 0.85 micro-average AUC, demonstrating encouraging potential in identifying the high-craving group (AUC = 0.84). SHAP quantified feature contributions: frequency of drug use, duration of use, and heroin use were core factors. Behavioral features positively correlated with high craving, whereas sociodemographic features exhibited a protective effect that diminished as addiction severity increased. Conclusion: The Logistic Regression model combines predictive performance and interpretability. By applying SHAP, this study visually elucidated specific feature contributions, enhancing model transparency for preliminary clinical evaluations, pending rigorous validation in independent and diverse populations.