Post-Infarction Risk Prediction with Mesh Classification Networks

Post-myocardial infarction (MI) patients are at risk of major adverse cardiac events (MACE), with risk stratification primarily based on global image-based biomarkers, such as ejection fraction, in current clinical practice. However, these metrics neglect more subtle and localized shape differences in 3D cardiac anatomy and function, which limit predictive accuracy. In this work, we propose a novel geometric deep learning approach to directly predict MACE outcomes within 1 year after the infarction event from high-resolution 3D cardiac anatomy meshes. Its architecture is specifically designed for direct and efficient processing of surface mesh data with a hierarchical, multi-scale structure to enable both local and global feature learning. We evaluate the binary MACE prediction capabilities of the proposed mesh classification network on a multi-center dataset of post-MI patients. Our results show that the proposed method outperforms corresponding clinical benchmarks by ∼ 16% and ∼ 6% in terms of area under the receiver operating characteristic (AUROC) curve for 3D shape and 3D contraction inputs, respectively. Furthermore, we visually analyze both 3D cardiac shapes and 3D contraction patterns with regards to their MACE predictability and demonstrate how task-specific information learned by the network on a balanced dataset successfully generalizes to increasing levels of class imbalance. Finally, we compare our approach to both clinical and machine learning benchmarks on our original highly-imbalanced dataset of post-MI patients and find average improvements in AUROC scores of ∼ 9% and ∼ 3%, respectively.