Adaptive Multi-Model Fusion Framework for Heart Attack and Diabetes Risk Prediction

Abstract

The increase in the prevalence of diabetes and cardiovascular diseases in the world requires accurate and easy to access diagnostics. These chronic diseases are asymptomatic and need early screening in an attempt to prevent complications and reduce healthcare expenditure. The contemporary systems yield tremendous data about patients, their demographics, lifestyle, medical history, lab results, and ECGs, and the traditional tools are processing these data separately. Such a fragmented approach lacks the ability to capture complex risk interactions and predictively forecast the likelihood, slowing down intervention, and putting strain on resources in high-volume or constrained environments. Our system solves this by providing a full preprocessing pipeline: data cleaning, feature scaling, temporal slicing, and missing value imputation of consistent and reliable inputs. It uses a hybrid modelling model based on XGBoost and TabNet with a rule-based logic module with clinical heuristics. The Multimodal Transformer Fusion layer is central, as it dynamically weighted and combined predictions across all modalities so that cross-model interactions could be used to provide context-related outputs. This provides subtle patient particular predictions, enhancing accuracy of diagnosis in profiles. One of the best is the mechanism of clinician-in-the-loop feedback. It provides SAP and SHAP explainability visuals, post-prediction, which clearly explain the contribution of features. Clinicians perfect the model based on practical knowledge, which includes AI and human intelligence that increases reliability and minimizes human supervision. Focused on clinicians, diagnostic laboratories, health professionals, and health-tech innovators, this explainable and easy-to-use platform allows detecting the disease at an early stage and designing a unique treatment regimen.