Data-Driven Predictive Modeling for Early Diagnosis and Risk Stratification of Chronic Diseases

Abstract

Chronic kidney disease (CKD) often remains undetected until advanced stages, limiting opportunities for timely intervention and increasing the burden of complications and treatment. This study proposes a data-driven hybrid ResNet50-SVM framework for the early diagnosis and risk stratification of CKD. Clinical data were prepared through missing-value treatment, data cleaning, Min-Max normalization, reduction of redundant information, and optimized feature selection. The processed dataset was divided into training and testing subsets using an 80:20 ratio. ResNet50 was employed to extract high-level representations, while a support vector machine performed the final classification. Model performance was evaluated using accuracy, precision, recall, F1-score, confusion-matrix analysis, and training and validation behavior. The proposed framework achieved an accuracy of 98.76%, precision of 97.65%, recall of 99.21%, and F1-score of 98.34%. It also outperformed logistic regression, a convolutional neural network, and a decision tree across the reported evaluation metrics. The high recall indicates a strong capacity to identify patients with disease while reducing missed positive cases. These findings demonstrate the potential of hybrid deep-learning and machine-learning methods to support accurate, scalable, and timely CKD screening. Further validation using larger, more diverse clinical datasets and explainable AI methods is needed before deployment in real-world clinical decision-support systems.