Construction and application of machine learning models for predicting intradialytic hypotension

Area of Science:

• Nephrology
• Artificial Intelligence
• Biomedical Informatics

Background:

• Intradialytic hypotension (IDH) is a frequent and serious complication during hemodialysis (HD).
• Predicting and preventing IDH is crucial for improving patient outcomes in chronic kidney disease management.

Purpose of the Study:

• To develop and validate machine learning models for predicting IDH using multiple definitions.
• To identify key clinical features contributing to IDH prediction.
• To create simplified, accessible models for clinical application.

Main Methods:

• Utilized a large dataset of 26,690 HD sessions for training/testing and 12,293 for temporal validation.
• Developed predictive models for five distinct IDH definitions using ten machine learning algorithms.
• Interpreted models, simplified features, and validated both complex and streamlined models temporally.

Main Results:

• The CatBoost algorithm achieved superior predictive performance across all five IDH definitions (ROC-AUC up to 0.880).
• Left ventricular mass index (LVMI) was consistently identified as a top predictive feature.
• Both complex and simplified machine learning models demonstrated robust performance on the temporal validation cohort.

Conclusions:

• Machine learning provides a reliable approach for predicting IDH in hemodialysis patients.
• LVMI is a critical factor in predicting intradialytic hypotension.
• Simplified predictive models are accessible for potential clinical integration.