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

Area of Science:

• Neuroscience and Addiction Research
• Machine Learning in Healthcare
• Public Health and Drug Policy

Background:

• Drug addiction is a chronic, relapsing brain disease where craving strongly predicts relapse.
• Traditional linear models struggle with complex addiction patterns.
• Machine learning (ML) shows promise but often lacks clinical transparency due to its 'black-box' nature.

Purpose of the Study:

• To develop a machine learning classification model for predicting drug craving scores.
• To explore the influence of multifactorial features on drug craving.
• To enhance the interpretability of ML models in addiction research.

Main Methods:

• Utilized data from 629 abstainers in drug rehabilitation centers, assessing craving with a 34-item scale.
• Applied data preprocessing, SMOTE for class imbalance, and compared seven ML algorithms (e.g., Logistic Regression, XGBoost, LightGBM).
• Employed 10-fold cross-validation, grid search for model selection, and SHAP for feature interpretability.

Main Results:

• Logistic Regression emerged as the optimal model, achieving 66.13% accuracy and 0.85 micro-average AUC on the test set.
• SHAP analysis identified drug use frequency, duration, and heroin use as core predictors of high craving.
• Behavioral features correlated with high craving, while sociodemographic factors showed a protective effect that decreased with addiction severity.

Conclusions:

• The Logistic Regression model offers a balance of predictive accuracy and interpretability for identifying high-craving individuals.
• SHAP visualization clarifies feature contributions, improving model transparency for potential clinical application.
• Further validation in diverse populations is necessary before widespread clinical adoption.