Interpretable machine learning for predicting chronic kidney disease progression risk.

Area of Science:

• Nephrology
• Medical Informatics
• Biostatistics

Background:

• Chronic kidney disease (CKD) presents a significant global health challenge.
• Accurate early prediction of CKD progression is crucial for timely intervention.
• Conventional prediction models often lack the necessary accuracy and interpretability for clinical use.

Purpose of the Study:

• To develop and compare machine learning (ML) models for predicting CKD progression.
• To enhance the interpretability of ML models for clinical decision-making.
• To identify key risk factors and their associations with CKD progression.

Main Methods:

• Utilized a cohort of 491 patients with clinical data, randomly split into training and testing sets.
• Developed and evaluated four ML algorithms (logistic regression, random forests, neural networks, XGBoost) for classification.
• Employed Cox proportional hazards regression (COX) and random survival forests (RSFs) for survival analysis.
• Assessed model performance using AUC-ROC, C-index, and integrated Brier score.
• Interpreted models using variable importance, partial dependence plots, and restricted cubic splines.

Main Results:

• XGBoost achieved the highest predictive performance for CKD classification (AUC-ROC: 0.867).
• Random survival forests (RSF) demonstrated superior discrimination and calibration in survival analysis compared to COX.
• Key predictors for CKD progression included estimated glomerular filtration rate, age, and creatinine.
• Identified non-linear associations between age and cholesterol levels with CKD progression.

Conclusions:

• Interpretable ML models are effective for predicting CKD progression.
• ML, particularly RSF, offers advantages over traditional methods in survival analysis for CKD.
• This study advances CKD risk prediction by elucidating non-linear risk factor relationships and comparing predictive techniques.