Lesion-Specific Prediction of Segmental Fractional Flow Reserve Using Deep Learning and Optical Coherence Tomography-Derived Features

Abstract

Background: Intracoronary imaging-derived physiologic indices enable vessel-level assessment of coronary flow impairment by integrating obstructive plaque burden. However, accurately evaluating hyperemic flow remains challenging due to the difficulty of simultaneously incorporating both vascular geometry and plaque composition.

Methods and Results: We aimed to develop and validate a deep learning algorithm using optical coherence tomography (OCT) to predict continuous, lesion-specific changes in fractional flow reserve (∆FFR) along coronary segments. Model performance was evaluated using Pearson’s correlation, a confusion matrix, and a receiver operating characteristic (ROC) curve, with invasive wire-based FFR as the reference. The model was trained and tested on 157 OCT pullbacks from 86 patients, demonstrating enhanced predictive performance by incorporating anatomical and compositional OCT features compared to conventional feature-based machine learning approaches. Predicted ∆FFR_OCTvalues correlated strongly with wire-based ∆FFR (R=0.932, P<0.001), achieving an accuracy of 0.952 for detecting physiologically significant lesions and an area under the curve of 0.953 in the ROC analysis. The model accurately localized segmental pressure gradients, effectively distinguishing focal from diffuse disease patterns.

Conclusions: This novel deep learning-based algorithm enabled lesion-specific, pressure wire-free physiologic assessment from a single OCT pullback. By integrating anatomical and compositional data, it supported improved diagnostic accuracy and personalized treatment for coronary artery disease.