A Risk-Oriented and Explainable Hierarchical AI Framework for Chronic Kidney Disease Classification.

Area of Science:

• Nephrology
• Machine Learning
• Health Informatics

Background:

• Chronic kidney disease (CKD) is a silent global health issue often detected late.
• Current machine learning CKD detection methods lack routine clinical workflow integration.
• Existing risk stratification models rarely combine with hierarchical decision-making.

Purpose of the Study:

• To develop a risk-oriented, explainable, hierarchical machine learning framework for CKD classification.
• To utilize real-world laboratory data for CKD detection, staging, and risk assessment.
• To create a system aligned with clinical practice for early CKD identification.

Main Methods:

• A hierarchical machine learning pipeline using XGBoost for CKD detection and MLP for staging.
• Recursive Feature Elimination (RFE) and SelectKBest for feature selection.
• Development using routine laboratory data from 746 Saudi patients.

Main Results:

• Achieved 97% accuracy and F1-score for binary CKD classification.
• Reached 85% accuracy and 86% F1-score for CKD stage classification.
• Enabled early CKD pattern identification via interpretable risk scoring and SHAP explanations.

Conclusions:

• The proposed framework supports preventive nephrology through transparent, deployable CKD risk assessment.
• The system facilitates clinically meaningful decision-making for early CKD intervention.
• Highlights the potential of routine lab data and explainable AI in managing CKD.