Machine learning to predict postdialysis fatigue in patients undergoing hemodialysis

Area of Science:

• Nephrology
• Artificial Intelligence
• Medical Informatics

Background:

• Postdialysis fatigue (PDF) is a prevalent complication in hemodialysis (HD) patients.
• Accurate prediction models for PDF are crucial for improving patient care.
• Machine learning (ML) offers potential for developing such predictive models.

Purpose of the Study:

• To explore the efficacy of various ML models in predicting PDF among Chinese HD patients.
• To identify key clinical factors associated with PDF using ML.

Main Methods:

• A cross-sectional study involving 1,281 Chinese HD patients from six tertiary hospitals.
• Seven ML models (LR, DT, RF, LGBM, CatBoost, XGB, GBT) were evaluated for PDF prediction.
• The TRIPOD+AI guidelines were followed for reporting study findings.

Main Results:

• The Random Forest (RF) model demonstrated optimal performance (AUC=0.855, Accuracy=0.773).
• Significant predictors of PDF included resilience, appetite, potassium levels, sleep quality, constipation, fistula surgery history, diastolic blood pressure, and co-existing diseases.
• The SHapley Additive exPlanations (SHAP) approach provided model interpretability.

Conclusions:

• ML models offer a practical tool for screening and assessing PDF risk in HD patients.
• Interpretable ML models enhance clinical understanding of PDF determinants.
• This framework supports better clinical decision-making for managing PDF.