ntegrating Mammographic Breast Arterial Calcifications and Clinical Data with Multi-Modal Transformers and GATs for Cardiovascular Risk Prediction

Cardiovascular disease (CVD) remains the leading global cause of mortality, necessitating innovative, non-invasive screening methods. Breast arterial calcifications (BAC), visible inmammograms, serve as a robust biomarker for CVD risk.Thisstudyproposesamultimodaldeeplearning framework integrating mammogram analysis with clinical data, using Swin Transformer V2 for image processingandGraphAttentionNetworks(GATs)for clinical data modeling, with late fusion for preciserisk stratification. The dataset combines INbreast, CBIS-DDSM, and a private hospital dataset, totaling 15,000 mammograms with clinical records. Themodel achieves an accuracy of 96.2%, sensitivity of 0.96, and AUC-ROC of 0.99, outperforming prior approaches. Advanced visualizations, including t- SNE embeddings, saliency maps, and feature importance plots, enhance interpretability, supporting clinical adoption. This framework leverages routine mammography for dual-purpose screening, offering a scalable, high-accuracy solution for early CVD detection.