Interpretable machine learning for chronic kidney disease prediction: Insights from SHAP and LIME analyses

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Renal Medicine

Background:

• Chronic kidney disease (CKD) necessitates early detection for improved patient outcomes.
• Machine learning (ML) adoption in healthcare faces transparency challenges.
• Interpretable AI is crucial for clinical integration and trust.

Purpose of the Study:

• To develop and evaluate an interpretable machine learning framework for transparent CKD prediction.
• To assess the clinical relevance of identified predictors using explainability techniques.
• To demonstrate the reliability of the framework across diverse datasets.

Main Methods:

• Developed an interpretable ML framework using XGBoost, optimized with SMOTE.
• Integrated SHapley Additive exPlanations (SHAP) for global interpretability.
• Utilized Local Interpretable Model-agnostic Explanations (LIME) for local, patient-level insights.
• Validated the model on two distinct datasets: UAE Tawam Hospital and UCI CKD data.

Main Results:

• Achieved high accuracy: 88.4% (AUC=0.904) on hospital data and 94.6% (AUC=0.948) on UCI data.
• Identified key predictors: eGFRBaseline, HbA1c, CholesterolBaseline (hospital); specific gravity, hemoglobin, serum creatinine (UCI).
• SHAP and LIME analyses showed convergence, confirming model reliability and clinical relevance.

Conclusions:

• The interpretable ML framework enhances transparency in CKD prediction.
• The approach ensures clinically realistic performance and addresses barriers to AI adoption in healthcare.
• Provides a foundation for integrating interpretable AI into CKD screening and management workflows.