Predicting blood pressure variability in hemodialysis using an explainable boosting machine model.

Area of Science:

• Nephrology
• Cardiovascular Medicine
• Artificial Intelligence in Healthcare

Background:

• Intradialytic hypotension and hypertension increase cardiovascular risk in hemodialysis patients.
• Current systolic blood pressure (SBP) prediction models have limitations, including consistent error ranges and lack of transparency.
• Improved SBP monitoring can enhance patient care beyond traditional event classification.

Purpose of the Study:

• To develop a transparent SBP prediction model using Explainable Boosting Machine (EBM).
• To evaluate the generalizability of the developed EBM model across different hospital branches.
• To compare the performance of the EBM model against other common machine learning methods.

Main Methods:

• Retrospective analysis of data from 524 hemodialysis patients (2016-19).
• Utilized Explainable Boosting Machine (EBM) for SBP prediction, incorporating hemodialysis parameters, vital signs, and predialytic measurements.
• Employed cross-branch validation for generalizability assessment and feature selection, comparing EBM with five other machine learning models.

Main Results:

• The EBM model achieved competitive performance (MAE: 10.57-11.33 mmHg, r: 0.80-0.83) in cross-validation.
• A waterfall plot visualized feature contributions to SBP prediction, offering transparency.
• High correlations (r: 0.81-0.87) were found between blood pressure variability and prediction errors.

Conclusions:

• EBM models demonstrated excellent cross-branch generalizability and feature-level transparency, aligning with clinical understanding.
• The transparent nature of EBM contrasts with black-box models, reducing the need for post-hoc explanation methods.
• Findings support the development of patient-specific solutions for managing highly variable blood pressure patterns in hemodialysis.