Non-invasive derivation of instantaneous free-wave ratio from Invasive Coronary Angiography using a new Deep Learning Artificial Intelligence model and comparison with Human operators’ performance

Abstract Background Invasive coronary physiology is largely underused and not without risk of complications despite its advantages over angiography alone. Artificial intelligence (AI) use in coronary physiology is still rudimentary. Methods We developed fully automated AI models capable of automatic coronary angiography segmentation and binary instantaneous free-wave ratio (iFR) lesion classification, based on a single coronary angiography (CAG) image. Three Interventional Cardiologists were asked to classify the target lesions binarily (cut-off ≤ 0,89) based on “raw” fluoroscopic frames AI-segmented frames. Their performance was then compared to AI. Results 250 measurements from 223 patients were included. Regarding the CAG analysis of all the arteries, the AI model presented an accuracy of 72%, a positive predictive value (PPV) of 48%, a negative predictive value (NPV) of 90%, a sensitivity of 77% and a specificity of 71%. The NPV was particularly high in the circumflex (Cx) and in the right coronary artery (CD) – 96% and 98%, respectively. Regarding human performance, accuracy ranged from 54–74%, PPV from 32–50%, sensitivity from 43–66%, specificity from 50–85% and NPV was 81%. The Operators’ NPV was also high regarding the Cx and RCA (95–97% and 94–97%, respectively), but significantly lower than AI in the left anterior descending artery (60–64% versus 78%). Conclusions We developed an AI model capable of binary iFR classification of lesions, slightly outperforming experienced Interventional Cardiologists overall. While not mature enough for clinical use, these results highlight the potential of AI in CAG-based coronary lesions assessment.