Explainable illicit drug abuse prediction using hematological differences.

This study aimed to develop a reliable and explainable predictive model for illicit drug use (IDU). The model uses a machine learning (ML) algorithm to predict IDU using hematological differences between illicit drug users (IDUr) and non-users (n-IDUr). A total of 286 IDUr and 302 n-IDUr were included in this study. By comparing the IDU prediction performance of eight ML models, an explainable LGB model with 13 features was established, which could accurately predict IDU in both internal (area under the curve [AUC] = 0.925) and external (AUC = 0.915) validation sets. Using SHapley Additive exPlanations (SHAP) to explain our prediction model, we found that chloride (Cl), β-hydroxybutyrate (BHB), and anion gap (AG) had a strong influence on the results predicted. Many of the features used for model training are interrelated, serving as important indicators of kidney, liver, and thyroid function in hematological examinations. Combining these manifestations can enable doctors to perform preliminary screening for IDU while conducting corresponding organ examinations in ordinary patients, which has a profound importance in clinical practice.