Privacy-Preserving Federated Learning for Early Prediction of Diabetes and Cardiovascular Disease using Multimodal Data from Rural Primary Health Centers

Abstract

Diabetes mellitus and cardiovascular disease (CVD) are significant non-communicable diseases for which early detection of risk may help with screening, referral, prevention and ongoing management. The information generated by the primary health centers (PHCs) in rural areas has fragmented information of clinical value through the electronic health records (EHRs), vital signs, laboratory tests, electrocardiograms, lifestyle information and, if available, low cost imaging. This is quite a big deal when it comes to privacy, governance and infrastructure issues, as they have to bring all that information together for AI development. This paper suggests a multimodal federated learning solution to early prediction of diabetes and CVD without transferring raw patient records from the PHCs to other parts of the healthcare system. The proposed architecture consists of modality-specific local encoders, attention-based multimodal fusion, federated optimization, gradient clipping, differential privacy and secure aggregation. If the rural PHC is heterogeneous, the experimental protocol for the site is defined and an evaluation is performed using the discrimination, calibration, sensitivity, specificity, F1-score, communication cost and privacy budget. No actual patient-level data was provided for this manuscript and the numerical results are a methodological demonstration only to show how the evaluation procedure is to be applied and do not reflect clinical evidence. In the simulated experiment, the proposed privacy-preserving multimodal federated model achieved an AUROC of 0.927, accuracy of 89.4%, sensitivity of 87.8%, specificity of 90.6%, F1-score of 0.884 and calibration error of 0.031, compared with 0.895 AUROC for conventional FedAvg and 0.914 for a centralized reference model. The multimodal configuration resulted in higher level of simulated discrimination than the individual EHR, laboratory or ECG configuration. The findings underscore the potential of multimodal federated learning as an effective technical solution for collaborative prediction of disease risk in rural communities with a high degree of control over local data. External validation: perspective of a perspective on perspective external validation. Subgroup analysis: perspective about perspective subgroup analysis.