Predicting burnout in radiology nurses: an interpretable machine learning model developed and externally validated in...

To develop and externally validate an interpretable machine learning model for predicting individual burnout risk among radiology nurses, and to identify modifiable, non-linear risk thresholds to guide precision prevention.

A two-center, two-stage design was used. A development cohort (n = 219) from six tertiary hospitals in China trained an XGBoost model. An independent external validation cohort (n = 234) was collected 1.5 years later from two tertiary centers (one partially overlapping with the development sites but with entirely independent participants) to rigorously test generalizability. Feature selection used bootstrapped LASSO with permutation testing. Model performance was assessed by discrimination, calibration, and decision curve analysis. SHAP (SHapley Additive exPlanations) provided global and individual-level interpretability.

The XGBoost model achieved strong discrimination in external validation (AUC = 0.859). In the internal test set, XGBoost outperformed logistic regression and other machine learning algorithms (AUC: 0.963 vs. 0.910 for logistic regression). SHAP analysis identified Effort-Reward Imbalance and Overcommitment as dominant predictors, and revealed a potential inflection point at approximately 65 weekly work hours, beyond which burnout risk escalated disproportionately in the SHAP analysis. Decision curve analysis confirmed net clinical benefit across threshold probabilities of 5 -65%. Individualized risk profiles were generated for actionable nurse management.

This externally validated, interpretable XGBoost-SHAP framework moves beyond population-averaged associations. By uncovering specific non-linear thresholds (e.g., 65 h/week workload) and providing personalized risk profiles, it enables nurse managers to shift from reactive crisis management to data-driven, precision prevention of burnout in radiology nurses.