Artificial intelligence-based analysis of novel and established lipid plaque measures in relation to clinical outcome

Background Intracoronary optical coherence tomography (OCT) enables high-resolution visualization of lipid-rich plaque structures. Established OCT-based prognostic markers include thin-cap fibroatheroma (TCFA) and lipid arc. Previous studies have demonstrated the capability of artificial intelligence (AI)-based models to quantify plaque characteristics using the outer boundary of lipid pools, enabling metrics such as plaque burden. However, the prognostic value of these boundary-based OCT-derived metrics remains unknown. Purpose To investigate the association of AI-derived plaque characteristics on OCT with major adverse cardiovascular events (MACE) in patients after myocardial infarction (MI). Methods Using our validated AI segmentation model (OCT-AID), 488 non-flow-limiting, non-culprit lesions were analyzed from 415 MI-patients enrolled in the PECTUS-obs study. OCT-pullbacks were automatically segmented by OCT-AID, focusing on frames within the predefined, core-lab selected lesions. We introduce the focal lipid-to-cap ratio (LCRf), a novel version of the lipid-to-cap ratio (LCR), defined as the AI-estimated lipid plaque thickness divided by the AI-estimated FCT at each of 360 circumferential degrees, measured from the centroid of the lumen. LCRf was designed to capture lipid-intima ratios at a more focal level, using automated AI-based calculation of the outer boundary of the lipid pool in every angle. The prognostic value of LCRf was compared to other metrics. The following metrics were calculated: TCFA, total number of frames with TCFA, and the maximum lipid arc, plaque burden, LCR and LCRf. Continuous metrics were split into tertiles, with patients with at least one lesion in the highest tertile being assigned to the “high” group, and remaining patients to the “low” group. Primary outcome measure was MACE at 2 year follow-up, defined as the composite of death from any cause, non-fatal myocardial infarction, and unplanned revascularization. Cox proportional hazard models were used to assess prognostic value. Results At two-year follow-up, MACE occurred in 34/415 cases (8.2%). Both binary TCFA and the number of TCFA frames were a significant predictor of MACE (HR 2.10, 95% CI 1.07-4.11, p=0.031 and HR 2.04, 95% CI 1.04-4.01, p=0.037, respectively). The new metric LCRf was associated with MACE (HR 2.39, 95% CI 1.21-4.70, p=0.012). No statistical association with MACE was found for lipid arc (HR 1.18, 95% CI 0.59-2.33, p=0.640), OCT-derived plaque burden (HR 1.22, 95% CI 0.61-2.41, p=0.572 ) and LCR (HR 1.55, 95% CI 0.79-3.05, p=0.204). Conclusions We found that established markers such as TCFA and the number of TCFA frames were associated with cardiovascular events. Also, the novel LCRf metric was significantly associated with MACE, suggesting that AI-enabled lipid plaque quantification using OCT-estimations of lipid plaque thickness may be worth exploring further.