Artificial Intelligence-Derived Fractional Flow Reserve in Routine Clinical Practice: An International Multicenter Retrospective Study

Abstract

Background

Artificial intelligence-based fractional flow reserve (AI-FFR) incorporates a machine learning-based algorithm to derive FFR directly from angiography. Its accuracy compared with invasive FFR has not been assessed.

Methods

AI-FFR was compared with wire-based FFR in patients with a single intermediate lesion (diameter stenosis ≥40% to <70%) at 5 centers in the United States and Israel. AI-FFR assessments were performed by core laboratory analysts blinded to invasive FFR results. Diagnostic performance metrics were calculated using an FFR threshold of ≤0.80.

Results

A total of 504 vessels from 496 patients were analyzed. AI-FFR computation time was 36.1 ± 7.7 seconds. Lesion detection was fully automatic in 371 of 504 (73.6%) vessels, whereas semiautomated analysis with manual marking was performed in 133 of 504 (26.4%) vessels. Mean wire-based FFR was 0.85 ± 0.07 and mean AI-FFR was 0.85 ± 0.08 (mean difference, 0.00 ± 0.08; 95% CI, −0.15 to 0.15; P = .41). AI-FFR showed a sensitivity of 90.2%, specificity of 94.9%, positive predictive value of 83.5%, negative predictive value of 97.1%, and overall diagnostic accuracy of 93.8%. The area under the receiver operating characteristic curve (AUC) was 0.93 (95% CI, 0.89-0.96). Among 151 lesions with wire-based FFR values in the borderline “gray zone” (0.75-0.85), AI-FFR demonstrated a diagnostic accuracy of 91.4% and an AUC of 0.91 (95% CI, 0.86-0.96). AI-FFR demonstrated high diagnostic accuracy across vessel types and lesion locations in men and women and between the US and Israeli cohorts.

Conclusions

AI-FFR, an automated, machine learning-based tool, demonstrated high diagnostic accuracy compared with wire-based FFR. Its speed, simplicity, and independence from complex procedural steps may facilitate broader adoption during coronary angiography.