Background: This study evaluated the diagnostic capability of on-site coronary computed tomography-derived computational fractional flow reserve (CT-FFR) determinations for detecting coronary artery disease (CAD), as assessed by invasive fractional flow reserve (FFR).Methods and Results:Seventy-four patients with coronary artery calcium scores <1,500 who underwent coronary CT angiography (CTA) and invasive FFR measurements within 90 days were retrospectively reviewed. CT-FFR was computed using a prototype machine-learning (ML) algorithm in 91 vessels; 47 vessels of 42 patients were determined to have significant CAD (FFR ≤0.8). Correlation between CT-FFR and FFR was good (r=0.786, P<0.001). Per-vessel area under the curve was significantly larger for CT-FFR (0.907, 95% confidence interval: 0.828-0.958) than for CTA stenosis ≥50% (0.595, 0.487-0.697) or ≥70% (0.603, 0.495-0.705) (both P<0.001). Standard coronary CTA classifications recommended further functional tests in 57 patients with moderate or worse stenosis on CTA. CT-FFR analysis (mean analysis time: 16.4±7.5 min) corrected the standard coronary CTA classification in 18 of 74 patients and confirmed it in 45 of 74 patients. Thus, the per-patient diagnostic accuracy of the classifications was improved from 66% (54-77%) to 85% (75-92%).
Conclusions: On-site CT-FFR based on a ML algorithm can provide good diagnostic performance for detecting hemodynamically significant CAD, suggesting the high value of coronary CTA for selected patients in clinical practice.
Keywords: Artificial intelligence; Computed tomography; Coronary artery disease; Fractional flow reserve; Machine learning.