A Contrastive Siamese 1D-CNN with Hybrid Feature Optimization for Accurate Diabetes Prediction from EHR Data

Abstract

Diabetes mellitus is a rapidly increasing global health concern, necessitating accurate and early risk prediction to support timely clinical intervention. This study proposes a compact and efficient diabetes prediction framework that integrates a hybrid IGF–DMO–RFO feature-selection strategy with a Siamese one-dimensional convolutional neural network enhanced by Global Spatial–Channel Attention (GSCA). The model is evaluated on a large-scale public diabetes dataset comprising 100,000 patient records with demographic, lifestyle, comorbidity, and metabolic features. The hybrid feature-selection pipeline identifies a concise and highly informative subset of predictors, while the Siamese architecture employs contrastive learning to generate discriminative embeddings from structured clinical data. Experimental results demonstrate that the proposed framework achieves an overall classification accuracy of 97%, with a ROC-AUC of 0.9726 and reliable probability calibration, outperforming conventional machine-learning and single-branch deep-learning baselines. The lightweight architecture enables fast inference and robustness to class imbalance, making it suitable for large-scale diabetes screening and clinical decision-support applications.