Feasibility of an artificial intelligence based fractional flow reserve assessment for coronary artery disease

BACKGROUND: The implementation of artificial intelligence has been investigated in many aspects of cardiovascular disease. OBJECTIVES: To develop deep learning models based on coronary angiograms to detect functionally significant coronary stenoses. METHODS: A total of 610 frames from 122 coronary arteries that received pressure wire-based fractional flow reserve (FFR) assessment were analyzed. Deep learning models were developed for the segmentation and classification of coronary stenoses. Both internal and external validation of the deep learning models were performed. RESULTS: The mean FFR value was 0.84 ± 0.08. The artificial intelligence-based FFR was significantly correlated with wire-based FFR with an average correlation coefficient of 0.68 and a mean absolute error of 0.05. The diagnostic performance of artificial intelligence-based FFR versus wire-based FFR was accuracy 87.6%, F1 score = 83.6%, and recall = 81.1%. The artificial intelligence-based FFR showed good discriminative performance with an area under the receiver operating characteristic curve of 86.5% (95% CI: 79.3-93.6). CONCLUSION: The artificial intelligence-based FFR showed moderate agreement with pressure wire-based FFR and showed promising diagnostic performance in the internal cohort, although reduced performance was observed in external validation, warranting further refinement and multicenter validation.