HeartFormer++ an Explainable Multimodal Deep Learning Framework for Early Detection of Heart Disease

According to the World Health Organization (WHO) report cardiovascular diseases (CVDs) are the leading causes of human deaths, mostly because of late diagnosis and the lack of interpretability of standard diagnosing systems. This work proposes a new multimodal deep learning technique to detect heart disease at an early stage with high accuracy using ECG, Echo, and clinical data following the traditional Transformer architecture. The model uses cross-modal attention to combine spatial, temporal and textual features. It also incorporates XAI for clinical interpretability. Using multi-institutional datasets (UK Biobank, PhysioNet, and MIMIC-IV) for an experimental evaluation, we achieve superior predictive accuracy 96% and improved clinical trustworthiness, compared to existing CNN- and RNN-based methods. This work offers a scalable, interpretable and privacy-aware diagnostic scheme potentially useful for real-time monitoring and telemedicine applications for cardiac health.