Deep-learning-driven optical coherence tomography analysis for cardiovascular outcome prediction in patients with acute coronary syndrome

Abstract

Aims: Optical coherence tomography (OCT) can identify high-risk plaques indicative of worsening prognosis in patients with acute coronary syndrome (ACS). However, manual OCT analysis has several limitations. In this study, we aim to construct a deep-learning model capable of automatically predicting ACS prognosis from patient OCT images following percutaneous coronary intervention (PCI).

Methods and results: Post-PCI OCT images from 418 patients with ACS were input into a deep-learning model comprising a convolutional neural network (CNN) and transformer. The primary endpoint was target vessel failure (TVF). Model performances were evaluated using Harrell's C-index and compared against conventional models based on human observation of quantitative (minimum lumen area, minimum stent area, average reference lumen area, stent expansion ratio, and lesion length) and qualitative (irregular protrusion, stent thrombus, malapposition, major stent edge dissection, and thin-cap fibroatheroma) factors. GradCAM activation maps were created after extracting attention layers by using the transformer architecture. A total of 60 patients experienced TVF during follow-up (median 961 days). The C-index for predicting TVF was 0.796 in the deep-learning model, which was significantly higher than that of the conventional model comprising only quantitative factors (C-index: 0.640) and comparable to that of the conventional model, including both quantitative and qualitative factors (C-index: 0.789). GradCAM heat maps revealed high activation corresponding to well-known high-risk OCT features.

Conclusion: The CNN and transformer-based deep-learning model enabled fully automatic prognostic prediction in patients with ACS, with a predictive ability comparable to a conventional survival model using manual human analysis.

Clinical trial registration: The study was registered in the University Hospital Medical Information Network Clinical Trial Registry (UMIN000049237).

Keywords: Acute coronary syndrome; Deep learning; Optical coherence tomography; Survival model.

Graphical Abstract

Deep-learning-driven optical coherence tomography (OCT) analysis for cardiovascular outcome prediction in patients with acute coronary syndrome (ACS). A deep-learning model utilizing a convolutional neural network (CNN) and a transformer enables automatic prediction of cardiovascular outcomes from post-percutaneous coronary intervention (PCI) OCT images in patients with ACS. Harrell’s C-index is significantly higher than that of the conventional model including only quantitative factors, and comparable to that of the conventional model including both quantitative and qualitative factors.

Figure 1

A schematic overview of the deep-learning model. Post-percutaneous coronary intervention optical coherence tomography images consisting of 50 random slices (250 × 250 pixels) are input into the convolutional neural network (Resnet 50), generating 256 element vectors. These vectors are fed into a transformer encoder model, along with a class token. Following layer normalization, a linear layer outputs two patient-specific parameters—µ and σ—that define the patient-specific hazard curve. The weights are optimized using the maximum likelihood process to match the actual survival data. Finally, the optimized hazard function is used to predict patient-specific survival curves.

Figure 2

A patient flow chart.

Figure 3

A comparison of model performance between conventional and deep-learning models. Harrell’s C-index values for conventional Model 1 (including only quantitative factors), conventional Model 2 (including quantitative and qualitative factors), and the ensemble deep-learning model are shown. The error bars represent a 95% confidence interval calculated using the bootstrap method.

Figure 4

Representative hazard and survival curves in the target vessel failure and non-target vessel failure groups. Hazard and survival curves estimated using µ and σ outputs from the deep-learning model for individual patients are shown: (A) target vessel failure group and (B) non–target vessel failure group. The upper figure shows the survival curve, while the lower figure shows the hazard curve. The dotted lines represent the curves estimated from each cross-validation and the solid lines represent the curves resulting from the ensemble model averaged over each cross-validation. The black arrows indicate the time of the target vessel failure event and the green arrows indicate the time of censoring.

Figure 5

Visualization of the attention of deep-learning models. The original optical coherence tomography images (left panels) and heat map images highlighting the regions with the highest attention within each optical coherence tomography cross-sectional image (right panels). In (A) and (B), the highlighted areas correspond to thin-cap fibroatheroma and lipid-rich plaques. In (C), almost the entire lumen with a small area is highlighted. (D) and (E) may indicate calcified nodule and thrombus, respectively, whereas (F), (G), and (H) suggest stent malapposition, stent edge dissection, and irregular protrusion.