Comparison of PTSD Risk Prediction Models for Rescue Workers Based on High-dimensional Features.

Abstract

Objective: Rescue workers are frequently exposed to traumatic events and are at a high risk of developing post-traumatic stress disorder (PTSD). We developed and compared machine-learning models for PTSD risk and assessed their generalizability.

Methods: This multicenter cross-sectional study included 13,462 rescue workers from 13 units for model development/internal validation and 8,757 workers from two additional units for external validation. Probable PTSD was defined as a PTSD Checklist for DSM-5 score ≥ 33. Sixteen variables were reduced using principal component analysis (PCA) or exploratory factor analysis (FA). Nine machine-learning algorithms were trained using a stratified 70:30 split, 5-fold cross-validation, and synthetic minority oversampling.

Results: Among PCA-based models, extreme gradient boosting (XGBoost) achieved the highest external area under the curve (AUC, 0.8931), whereas random forest showed the highest sensitivity. Among the factor-analysis-based models, XGBoost achieved the highest external AUC (0.8832), whereas random forest provided a sensitivity-oriented alternative. Original-variable SHapley Additive exPlanations (SHAP) analysis identified acute stress disorder (ASD), self-rating depression scale (SDS), self-rating anxiety scale (SAS), passive smoking, monthly family income, smoking status, body mass index (BMI), and age as the main contributors.

Conclusion: Dimensionality-reduced machine-learning models showed measurable discriminatory ability for PTSD risk stratification, and model choice should reflect screening priorities.