Trained on 2,595 carotid artery plaques stained with nine histological markers, the PHENOMICL framework — an attention-based multiple instance learning model — detected intraplaque hemorrhage (IPH) with an AUROC of 0.86 using routine H&E staining alone, rising to 0.92 when combining H&E with CD68 or Verhoeff-Van Gieson stains. Critically, the AI-derived IPH quantification outperformed conventional manual histological scoring when predicting preoperative symptoms and major adverse cardiovascular events (MACE). Transcriptomic integration implicated TNF-alpha signaling, macrophage foam-cell activity, and the CCL-ACKR1 communication axis in IPH-driven angiogenesis and plaque instability.
IPH is among the most dangerous features of vulnerable atherosclerotic plaques, accelerating rupture risk that leads to stroke and myocardial infarction — yet its quantification has remained cumbersome, subjective, and poorly standardized across clinical centers. If validated, PHENOMICL could transform routine carotid endarterectomy pathology into a scalable risk-stratification tool without requiring specialized staining protocols. The transcriptomic findings add mechanistic depth, particularly the CCL-ACKR1 axis, which may represent a therapeutic target for plaque stabilization. However, several limitations deserve attention: the dataset is carotid-specific, so generalizability to coronary or other vascular beds remains untested; the causal direction of identified molecular pathways cannot be established from this design; and external prospective validation is absent. As a preprint posted to medRxiv and not yet peer-reviewed, these performance metrics and biological associations should be considered preliminary until independent scrutiny confirms the findings. Still, the combination of interpretable spatial probability maps with multi-omic integration marks this as a meaningfully ambitious advance.
