Deep learning approach for early-stage diabetes risk prediction using Southern Indian Diabetes Dataset

Author : POOJA

Abstract :This proposed work investigates the efficiency of a hybrid data augmentation approach (SMOTE-ENN + GAN), SHAP-based feature selection, and deep learning models for classifying the 2001 dataset, which exhibits significant class imbalance. Five deep learning models Radial Basis Function Neural Networks (RBF-NN), Long Short-Term Memory (LSTM), Simple Recurrent Neural Networks (RNN), Convolutional Neural Networks (CNN), and Deep Neural Networks (DNN) were evaluated for their performance. The proposed hybrid augmentation technique significantly improved the model's ability to simplify by balancing the class distribution, resulting in a notable increase in model accuracy. The DNN and CNN models achieved a test accuracy of 98%, outperforming other approaches in the literature. SHAP-based feature selection further enhanced model interpretability, identifying critical features such as age, obesity, and visual blurring. The comparative results confirm that the proposed methodology, combining hybrid data augmentation and SHAP feature selection, provides substantial improvements in classification tasks, showcasing its potential for real-world applications where high accuracy is crucial.

Keywords :Deep Learning, Feature Importance, Class Imbalance, Generative Adversarial Networks.

Conference Name :National Conference on Advanced Computer Science and Information Technology (NCACSI-25)

Conference Place Guwahati India

Conference Date 21st Jun 2025