A Hybrid Machine Learning Approach for Enhanced Diabetes Prediction: Integrating Image and Numerical Data

Abstract

Diabetes mellitus (DM) continues to escalate as a worldwide health emergency issue, with approximately 537 million adults currently diagnosed and forecasts estimating a further increase to 643 million by 2030. Early and precise foretelling of DM remains a decisive factor for timely intervention, thereby mitigating severe downstream sequelae such as cardiovascular disease, peripheral neuropathy, and diabetic retinopathy (DR). Conventional prognostic frameworks typically depend on exclusively either structured tabular measurements or visual medical imagery, which constrains comprehensive diagnostic capacity. This contribution confronts such limitation by advancing a hybrid machine learning (ML) methodology that synergistically combines deep learning—specifically, convolutional neural networks (CNNs) dedicated to retinal photograph scrutiny—with gradient-boosting machines (GBMs) that ingest structured demographic and clinical variables. Two publicly accessible repositories supplied training material: the Pima Indians Diabetes Database for tabular covariates and the Asia Pacific Tele-Ophthalmology Society (APTOS 2019) Blindness Detection corpus for fundus imagery. Retinal studies underwent standardised pre-processing re-scaling, pixel normalisation, Gaussian denoising, and multiplicative augmentation while tabular patient records underwent rigorous feature ranking. Outcome representations from both data strata were concatenated into a consolidated tensor, thereby rendering simultaneous latent-space learning achievable. The experimental results demonstrate that the hybrid model outperforms single-modality models, achieving an accuracy of 96%, a macro average F1 score of 0.96, and an area under the receiver operating characteristic curve (AUC-ROC) of 0.994. The proposed approach offers a comprehensive diagnostic framework by combining systemic and localized disease indicators, thereby enhancing robustness, reducing variance, and supporting more informed clinical decision-making. This work highlights the potential of multimodal ML integration for complex disease prediction and sets the stage for future extensions to other chronic conditions.