Identifying compassion fatigue indicators in Chinese psychiatric nurses: a machine learning approach with SHAP interpretation

To develop and internally evaluate interpretable machine-learning models for classifying current questionnaire-threshold-defined compassion-fatigue (CF) risk among Chinese psychiatric nurses, while quantifying feature-selection stability and outcome-definition dependence. Compassion-fatigue-related symptoms are important occupational-health concerns among psychiatric nurses. Cross-sectional classification may support additional assessment, but it cannot establish future risk or causal determinants. This multicentre cross-sectional study recruited 607 psychiatric nurses from four psychiatric hospitals in Northeast China (valid response rate: 94.10%). Participants were allocated to a training set (n = 424) and a held-out test set (n = 183) using a centre-preserving 7:3 split. Categorical variables were appropriately encoded during feature selection and model fitting. Feature stability was assessed using repeated stratified five-fold cross-validation (10 repeats; 50 resamples) with Boruta, CV-LASSO, and grouped backward regression; variables selected by ≥ 2 methods in ≥70% of resamples were retained. Eleven algorithms were tuned and compared using conditional five-fold OOF predictions, with model selection considered exploratory. The held-out test set was evaluated using frozen predictions, with uncertainty estimated from 2,000 bootstrap resamples. The primary questionnaire-threshold-defined outcome was present in 495/607 participants (81.55%). For the primary outcome, five variables met the predefined stability criterion: perceived pressure, age, professional identity, depression, and physical exercise during the past year. SVM was selected through an exploratory model-comparison process. In the held-out test set, SVM achieved an AUC of 0.830 (95% CI 0.741–0.907), balanced accuracy of 0.703 (95% CI 0.605–0.790), sensitivity of 0.706, specificity of 0.700, PPV of 0.923, NPV of 0.318, and Brier score of 0.134 (reference Brier 0.138; Brier skill score 0.028). Calibration assessment showed an intercept of 1.491 and a slope of 0.784. Alternative outcome definitions resulted in different selected models and test AUCs of 0.817 and 0.927, indicating that model performance was dependent on outcome operationalization. Machine-learning models identified stable indicators associated with questionnaire-based threshold-defined compassion fatigue outcome among Chinese psychiatric nurses, including perceived pressure, age, professional identity, depression, and physical exercise. These findings may inform future occupational-health assessment and risk-stratification research. However, further hospital-level and prospective validation is required before practical implementation.