AutoML-driven ensemble learning for intradialytic hypotension prediction.

OBJECTIVE

Intradialytic hypotension (IDH) is a common and serious complication during hemodialysis (HD), associated with increased morbidity and mortality. This study aims to develop an artificial intelligence (AI)-driven early warning system to predict IDH events and recommend real-time dialysis parameter adjustments to enhance patient safety.

MATERIALS AND

METHODS

A total of 9,634 HD sessions from 178 patients were collected. Advanced feature engineering techniques were applied, including domain-specific variable generation, Synthetic Minority Oversampling Technique (SMOTE), and Recursive Feature Elimination with Cross-Validation (RFECV). An AutoML pipeline tested 100 hyperparameter combinations per run to tailor hospital-specific models. This approach enables rapid development of customized models optimized for institutional variability.

RESULTS

Among the evaluated models, CatBoost and LightGBM with Extra Trees achieved the highest performance. CatBoost yielded the best results with a ROC-AUC of 0.934 and an F1-macro score of 0.508. Feature selection and SMOTEEEN resampling further enhanced performance, improving the ROC-AUC to 0.949.

CONCLUSION

The implementation of AutoML for large-scale hyperparameter optimization enables the efficient development of hospital-specific AI models with minimal manual input. This approach supports scalable customization across diverse clinical settings. Our findings underscore the value of AI-powered prediction systems in facilitating early identification and management of intradialytic hypotension (IDH), thereby contributing to improved patient safety and hemodialysis outcomes.