Comparison Between Statistical Model and Machine Learning Methods for Predicting the Risk of Renal Function Decline...

Area of Science:

• Nephrology
• Medical Informatics
• Data Science

Background:

• Predicting renal function decline (RFD) is crucial for managing chronic kidney disease.
• Machine learning (ML) offers advanced methods for analyzing complex risk factor interactions.
• Routine clinical data holds potential for developing accurate predictive models.

Purpose of the Study:

• To evaluate and compare the performance of various ML algorithms against logistic regression for predicting RFD risk.
• To assess the utility of ML in identifying individuals at risk of renal function decline using standard clinical data.

Main Methods:

• A retrospective cohort study of 2166 participants (aged 35-74) from an adult health screening program.
• Seven ML models (random forest, gradient boosting, etc.) and logistic regression were employed.
• Baseline estimated glomerular filtration rate (eGFR) was the primary predictive variable, with 24 independent variables.

Main Results:

• All models demonstrated strong predictive performance, with Area Under the Receiver Operating Characteristic Curve (AUROC) above 0.85.
• Gradient boosting achieved the highest prediction accuracy (AUROC: 0.914).
• ML models generally improved RFD prediction over logistic regression (AUROC: 0.882), though differences were small and not always significant.

Conclusions:

• ML models are applicable and validated for RFD risk prediction using health screening data.
• Logistic regression provides comparable performance to ML models for RFD prediction with simple clinical predictors.
• The incremental benefit of complex ML models over logistic regression may be limited with readily available clinical data.