Future cardiovascular events prediction from invasive coronary angiography: A graph representation learning...

Improving risk stratification for coronary artery disease (CAD), the leading global cause of death, remains a daily challenge in clinical practice. This highlights the urgent need for innovative approaches to early prediction of future cardiovascular events from invasive coronary angiography (ICA), the gold standard imaging modality for CAD diagnosis. However, current methods, including clinical indices and data-driven deep learning models, fail to tackle this challenge. In this work, we propose AngioGraphCAD, a graph neural network-based framework for patient-level future cardiovascular events prediction from ICA. By design, the framework flexibly accommodates the variable number of coronary lesions per patient, enabling robust patient-level prediction. For each lesion, AngioGraphCAD constructs a geometric graph representation and learns a lesion-level embedding that leverages coronary artery geometry alongside clinical data, together with a lesion-level prediction objective. These lesion-level embeddings are then fused using a novel masked-attention mechanism to form a unified and interpretable patient-level representation. Across two clinical cohorts (FAME2: 563 patients, 1551 stenoses; FCL: 83 patients, 382 stenoses), AngioGraphCAD achieves strong lesion-level performance, significantly outperforming clinical measures and CNN-based models, with AUCs of 0.71 (FAME2) and 0.73 (FCL). Building on these lesion-level representations, AngioGraphCAD further achieves an AUC of 0.70 for patient-level prediction on FAME2. Overall, this study underscores the predictive value of coronary artery geometry from ICA and highlights geometry-aware representations as a promising step toward more personalized CAD management.