Deep Neural Network Approach for Early-Stage Diabetes Risk Prediction using Hybrid SMOTE-ENN and GAN with SHAP-Based Feature Explanations

Abstract

Type 2 diabetes mellitus (T2DM) is prevalent in India and remains a major global health concern. Timely recognition of type 2 diabetes is essential for successful management, especially in resource-constrained environments where access to standard laboratory testing might be restricted. This research introduces a deep learning framework aimed at predicting early-stage diabetes, emphasizing clarity and practical application in clinical environments through the use of non-invasive, symptom-based inputs. The Southern India Diabetes Dataset (SIDD) is a notable regional collection, comprising 1,680 ethically sourced patient records along with 17 clinical and demographic variables. A hybrid augmentation strategy was employed to address the class imbalance, incorporating SMOTE-ENN in aggregation with Generative Adversarial Networks (GANs). Furthermore, SHAP (SHapley Additive exPlanations) values were utilized to identify essential predictive features, enhancing the model interpretability. We employed and assessed two neural architectures: the Radial Basis Function Neural Network (RBFNN) and the Deep Neural Network (DNN). The proposed DNN model achieved a test accuracy of 98%, surpassing the performance of models trained on standard datasets such as PIMA. The proposed framework shows significant potential for application in essential healthcare settings, due to its incorporation of interpretable artificial intelligence, strong augmentation, and pertinent clinical data. This will facilitate prompt intervention and improve patient outcomes.