Predicting intradialytic exercise intolerance in maintenance hemodialysis patients: an interpretable machine learning approach integrating functional assessments.

Exercise intolerance, primarily manifesting as intradialytic hypotension (IDH), is common in maintenance hemodialysis patients, interrupting interventions and contributing to adverse outcomes. This prospective observational study enrolled 128 adults on thrice-weekly hemodialysis for ≥3 months, generating 896 dialysis records. Four classifiers-logistic regression with elastic net penalty, Random Forest, histogram-based gradient boosting decision tree (HGBDT), and eXtreme Gradient Boosting (XGBoost)-were evaluated using a 10-fold GroupKFold cross-validation strategy on the entire dataset. Performance was assessed via receiver operating characteristic area under the curve (ROC AUC), precision-recall area under the curve (PR AUC), accuracy, harmonic mean of precision and recall (F1 score), and Brier score. SHapley Additive exPlanations (SHAP) analysis enhanced interpretability, with calibration via reliability curves and clinical utility via decision curve analysis (DCA). Exercise intolerance occurred in 29.7% of patients (30.2% of records). The Random Forest model outperformed other models, achieving a mean ROC AUC of 0.914 ± 0.024, accuracy of 0.845 ± 0.025, and F1 score of 0.750 ± 0.047. Model stability was confirmed across validation folds. Calibration showed good agreement between predicted and observed probabilities, and DCA confirmed superior net benefit across threshold probabilities of 0.1-0.6. SHAP analysis identified Timed Up and Go (TUG) duration, predialysis systolic blood pressure, and ultrafiltration parameters (rate and total volume) as the top predictors of exercise intolerance. The Random Forest model, enhanced by functional assessments and SHAP interpretability, offers a robust, transparent tool for predicting intradialytic exercise intolerance, supporting precision nephrology through tailored risk management.