SHAP-Integrated Deep Neural Network for Interpretable Early Detection of Chronic Kidney Disease

Abstract

: Chronic kidney disease (CKD) remains a global health burden, often progressing silently until advanced stages, thereby underscoring the urgent need for early and interpretable diagnostic tools. We propose a SHAP-integrated deep neural network framework for the early detection of CKD from routine clinical indicators. The methodology begins with rigorous preprocessing, including median imputation for missing values and z-score normalization of continuous features to ensure stable convergence. A fully connected deep neural network, composed of multiple hidden layers with ReLU activations and dropout regularization, is then trained to map the normalized feature vector to a binary probability of CKD presence. The model is optimized using the Adam algorithm to minimize binary cross-entropy loss. To address the critical requirement of clinical interpretability, we incorporate SHAP (SHapley Additive exPlanations) as a post-hoc explanation module. This module decomposes each patient’s prediction into additive feature contributions, thereby quantifying how variables such as serum creatinine or hemoglobin influence the risk score. The Shapley values are computed efficiently via approximation, yielding both global feature importance rankings and local, patient-specific force plots. Our framework thus delivers two key contributions: first, a high-performance deep learning classifier tailored for early CKD prediction; second, a transparent explanation mechanism that aligns with clinical decision-making needs. The significance of this work lies in bridging the gap between predictive accuracy and model interpretability in nephrology, potentially enabling earlier interventions and improved patient outcomes. Experimental evaluations on benchmark clinical datasets demonstrate that the proposed method achieves competitive predictive performance while providing actionable insights into the driving factors behind each diagnosis