An Explainable Lightweight Deep Learning Framework for Early Diabetes Risk Prediction Using Clinical Dataset and Hybrid Optimization Techniques

Abstract

Early diabetes risk identification is vital for preventing progression to disease and managing long-term complications of health care. However, many existing AI risk prediction models suffer from poor generalization due to several issues; class imbalance, irrelevant clinical features, lack of explainability, high computational complexity, and inadequate validation on heterogeneous patient data sets. This research examines these issues utilizing the Early Stage Diabetes Risk Prediction Dataset 2025, which contains demographic, clinical, and symptom attributes that can be used for early diabetes risk screening purposes. The framework includes missing value imputation, z-score normalization, data augmentation using Conditional Tabular Generative Adversarial Network (CTGANs), mutual information and chi-square feature selection, hybrid optimization methods, and Integrated Gradient-based methods for providing explanation of predictions. Classification is performed using a low computational complexity convolutional model, Lightweight Convolutional Integrated Gradients-Griffon Vultures Optimization Algorithm and Revolution Optimization Algorithm (LwCIG-GVROA). Experimental evaluation resulted in 95.20% accuracy, 96.22% precision, 97.12% recall and 99.65% F1 score; indicating that this approach to early diabetes risk prediction has good reliability, computational efficiency, and clinical interpretability.