Abstract We0067: Intraplaque hemorrhage prediction using attention based multiple instance learning

Background: Atherosclerosis is a leading cause of cardiovascular morbidity and mortality, driven by the development of vulnerable plaques characterized by phenotypes such as intraplaque hemorrhage (IPH). IPH is a hallmark of plaque instability and is associated with major adverse cardiovascular events (MACE). Traditional histological methods for plaque assessment are labor-intensive and subjective, highlighting the need for scalable, objective approaches. Recent advancements in machine learning, particularly attention-based multiple instance learning (MIL), offer a promising avenue for automated, precise plaque phenotyping and risk stratification. Methods: We applied an attention-based additive MIL model to detect and quantify IPH using whole slide images (WSIs) from carotid endarterectomy plaques. Plaques were stained with H&E and eight immunohistochemical markers, including EVG, Picrosirius Red, CD34, and CD68, to capture diverse cellular and extracellular plaque features, consisting of 13,345 WSIs from 2595 patients. Patch-level MIL scores were used to localize and quantify IPH, with manual binary labels serving as ground truth. Additionally, bulk RNA-seq data from 1,045 plaques were analyzed for differentially expressed genes (DEG), linking histological features to molecular signatures. Results: The MIL model achieved an AUROC of 0.86 for IPH detection, with H&E demonstrating the strongest staining performance. IPH levels quantified through MIL correlated with erythrocyte presence, as verified by glycophorin C staining. DEG analysis revealed transcriptional signatures associated with IPH, including upregulation of macrophage-related genes (e.g. gene CXCL1, SIGLEC1) and downregulation of smooth muscle cell markers (e.g. gene ACTA2). Cox regression analysis showed a clear impact of IPH on survival (P = 0.04). Finally, using XGBoost, integration of clinical data identified IPH, SMA, and glomerular filtration rate as critical predictors of MACE and patient symptoms. Conclusions: This study demonstrates the potential of attention-based MIL to automate IPH detection and quantify plaque vulnerability features at scale. The integration of histological, molecular, and clinical data provides a comprehensive framework for understanding plaque instability and improving risk prediction. Future work will expand this approach to investigate genetic determinants of plaque vulnerability and further validate these findings in large, multi-cohort datasets.