Predicting Chronic Kidney Disease from Biomarkers: An Explainable Machine Learning Approach.

Area of Science:

• Nephrology
• Artificial Intelligence
• Machine Learning

Background:

• Chronic kidney disease (CKD) is often diagnosed late, necessitating improved early detection methods.
• Current screening tools lack the required discrimination and transparent operating thresholds for clinical deployment.

Purpose of the Study:

• To develop and validate a robust, clinically deployable machine learning framework for early CKD detection.
• To ensure the model provides high discrimination, an explicit operating threshold, and transparent explanations.

Main Methods:

• A CKD detection framework integrating structured preprocessing, class imbalance handling, and 10-fold cross-validation with out-of-fold (OOF) prediction.
• Clinically oriented threshold selection using the Youden index and explainability via SHAP and LIME.
• Evaluation across two datasets using ten machine learning models, with a focus on gradient boosting methods.

Main Results:

• LightGBM demonstrated superior clinical composite performance on both datasets, achieving near-ceiling OOF discrimination (ROC-AUC up to 99.98).
• The model provided excellent sensitivity (e.g., 99.20%) and specificity (e.g., 99.60%) at optimal Youden thresholds, with robust cross-validation stability and strong calibration.
• SHAP and LIME explanations confirmed alignment with clinically meaningful renal function and biomarker patterns.

Conclusions:

• The proposed machine learning framework offers a reliable and explainable solution for CKD screening.
• The model's performance and transparent decision-making support its translation into clinical workflows for improved patient outcomes.
• Gradient boosting methods, specifically LightGBM, are highly effective for developing high-performance CKD detection models.