Artificial intelligence–driven three-vessel OCT quantification identifies the vulnerable patient and predicts 1-year outcomes in acute coronary syndromes

Background Acute coronary syndromes (ACS) result from plaque instability; however, conventional imaging focuses primarily on culprit lesions rather than systemic vulnerability. Optical coherence tomography (OCT) enables high-resolution plaque characterization but remains time-intensive and operator-dependent. Artificial intelligence (AI) may allow automated pancoronary vulnerability quantification and improve risk stratification beyond traditional clinical scores. Purpose To evaluate whether an AI-based three-vessel OCT model can automatically quantify pancoronary plaque vulnerability and improve prediction of 1-year clinical outcomes in ACS. Methods This was a retrospective analysis of a prospectively maintained three-vessel OCT registry including consecutive ACS patients between January and December 2024. A convolutional neural network performed automated lipid-rich plaque segmentation, thin-cap fibroatheroma (TCFA) classification, fibrous cap thickness (FCT) quantification, and macrophage detection. Model performance was validated against independent core laboratory assessment. Patients were followed for 12 months (through December 2025) for major adverse cardiovascular events (MACE: cardiovascular death, non-fatal myocardial infarction, or unplanned revascularization). Multivariable Cox regression adjusted for age, sex, diabetes, LDL-C, and GRACE score. Results Among 312 patients (mean age 59±11 years; 27% female; 38% diabetes; median GRACE 128 [IQR 112–146]), complete follow-up was available in 300 (96%). For TCFA classification, AI achieved AUC 0.92 (95% CI: 0.89–0.95), sensitivity 90% (95% CI: 86–93%), specificity 87% (95% CI: 83–91%), precision 0.88, recall 0.90, and F1-score 0.89. Lipid segmentation showed Dice coefficient 0.88 (95% CI: 0.86–0.90). FCT measurements strongly correlated with manual assessment (r=0.91, 95% CI: 0.88–0.93; mean absolute error 12 μm). Macrophage detection yielded AUC 0.86 (95% CI: 0.82–0.90), sensitivity 82%, specificity 79%, and F1-score 0.81. Pancoronary vulnerability was common: TCFA was present in 72% (95% CI: 66–78%) of culprit and 24% (95% CI: 20–29%) of non-culprit lesions; 42% (95% CI: 36–48%) had ≥1 non-culprit TCFA. Non-culprit TCFA burden independently predicted MACE. At 12 months, MACE occurred in 60 patients (19.2%, 95% CI: 15.0–23.4%). Patients in the highest AI-derived lipid tertile had HR 3.52 (95% CI: 2.10–5.80; p<0.001), with adjusted HR 2.94 (95% CI: 1.76–4.90). Addition of AI-derived OCT metrics improved C-statistic from 0.71 (95% CI: 0.66–0.76) to 0.82 (95% CI: 0.78–0.87; p=0.003), with net reclassification improvement of 0.27 (95% CI: 0.14–0.40). Conclusion Automated pancoronary AI-OCT quantification provides objective assessment of systemic plaque vulnerability and independently predicts 1-year adverse outcomes in ACS, significantly improving risk discrimination beyond traditional clinical scoring.