A Multimodal Deep Learning Approach for Cardiac Risk Assessment

Cardiovascular diseases cause large numbers of deaths worldwide. This creates a need for precise and individualized diagnostics. This study presents a multimodal predictive framework that combines image-based deep learning and structured clinical data analysis. The image branch uses a convolutional neural network trained on cardiac MRI datasets to classify patients as normal or diseased. The second branch processes patient data, including clinical, biological and genetic features, using an XGBoost model. The system fuses outputs from both branches in a joint layer to produce a final patient-level prediction. Experiments show high predictive accuracy in both branches. Grad-CAM and SHAP provide visual and quantitative insight into model behavior. A prototype front-end shows practical usage. Clinicians upload cardiac images and patient parameters and receive prediction scores and suggested treatments.