Federated Multimodal Contrastive-Learning Transformer with Calibrated Risk Scoring for CHD

Coronary Heart Disease (CHD) remains a leading cause of morbidity and mortality worldwide, necessitating the earlier, precision and privacy preserving prediction systems in assisting the clinical decision making process. The recent advancement of technology and the evolutions of Deep Learning (DL) algorithms have demonstrated promising results in the diagnosis process, leveraging the multimodal clinical data such as Electrocardiogram (ECG), medical imaging and Electronic Health Records (ECR). Despite of advancements in state of art diagnosing methodologies, the precision experiences a setback in data privacy concern, institutional data silos, heterogeneous data distribution, poor model calibration and cross modal representational alignment due to the regulatory constraints, domain drifts and inconsistent data quality. To overcome these challenges, this manuscript proposes a Federated Multimodal Contrastive Learning Transformer with Calibrated Risk Scoring (FMCL-CRS) for a robust and privacy preserving CHD risk prediction. This framework integrates the transformer based multimodal encoder with contrastive learning method to align with the input heterogeneous feature representations across multi modalities. The framework is trained and tested using UCI heart disease dataset, PhysioNet ECG dataset and MIMIC-IV clinical records to analyze the performance in terms of accuracy, precision, recall, F1 score, ROC curve, brier score, Expected calibration Error (ECE) and Communication efficiency. The experimental analysis is anticipated to achieve superior performance and enhanced robustness for heterogeneous data and is suitable for the real world CHD risk assessment applications.