Use of machine learning models to predict mortality in dialysis patients

Area of Science:

• Nephrology
• Artificial Intelligence
• Biostatistics

Background:

• High mortality rates persist in maintenance hemodialysis patients.
• Traditional statistical models struggle with complex clinical data.
• Need for advanced predictive tools in hemodialysis care.

Purpose of the Study:

• To systematically develop and compare 19 machine learning algorithms.
• To validate models for predicting all-cause mortality in hemodialysis patients.
• To identify optimal algorithms for clinical risk stratification.

Main Methods:

• Retrospective study of 538 maintenance hemodialysis patients.
• Data split into 70% for training and 30% for testing.
• Hyperparameter optimization using accuracy, F1-score, and ROC AUC.

Main Results:

• Gradient boosting models (XGBoost, AdaBoost) showed superior performance.
• XGBoost (accuracy-optimized) achieved F1=0.683, ROC AUC=0.899.
• AdaBoost (F1-optimized) achieved highest ROC AUC=0.903, F1=0.682.
• AdaBoost demonstrated consistent performance across strategies.

Conclusions:

• Machine learning offers tailored, high-performing mortality risk models.
• Potential to enhance identification and management of high-risk hemodialysis patients.
• AdaBoost is suitable for clinical implementation requiring balanced risk prediction.