Early Prediction of Chronic Kidney Disease Using Fine-Tuned VGG19 and densenet121 with SHAP-Based Interpretability

Abstract

Chronic Kidney Disease (CKD) is a progressive and potentially fatal condition that often goes undetected until its advanced stages. Timely prediction is essential to reduce the burden on healthcare systems and improve patient outcomes. In this study, we propose a robust CKD prediction model based on the VGG19 deep convolutional neural network, enhanced through fine-tuning, class weighting, techniques. The model is learned on a formatted clinical dataset of principal renal biomarkers and patient characteristics. For handling data imbalance, class weight adjustment was done during training and model checkpoints were saved for resume. Our trained VGG19 model recorded 99.13% accuracy while Densenet121 recorded 99% accuracy, surpassing traditional classifiers and yielding improved generalization. Furthermore, SHAP (SHapley Additive exPlanations) values were used to provide feature-level interpretability, confirming the biological significance of top-performing predictors such as serum creatinine, albumin, and blood pressure. Comparative comparisons with baseline models and earlier work highlight the reliability and clinical usefulness of our approach. The proposed model not only delivers precise predictions but also offers transparent decision-making, thus constituting a high-value resource for early intervention in CKD management.