automated reduced order CT-FFR model for diagnosis of hemodynamically significant coronary artery disease- a prospective validation study

Abstract Background/Introduction CT derived fractional flow reserve (CT-FFR) is introduced in current guidelines as a gatekeeper for invasive coronary angiography (ICA) in symptomatic patients with intermediate stenoses on coronary CT angiography (CCTA). Traditionally, CT-FFR is validated with knowledge of the precise location of invasive FFR measurements to ensure a direct comparison. However, in a clinical setting, this information will not be available and may affect the diagnostic performance of CT-FFR. Purpose To validate the diagnostic performance of a novel physics augmented machine learning algorithm based on reduced order modelling for CT-FFR calculations, with an automated method for defining the site of CT-FFR evaluations. Methods In this prospective validation study, 108 consecutive patients with stable coronary artery disease (CAD) and at least one suspected obstructive stenosis of questionable hemodynamic significance on CCTA were included. All patients underwent a clinically indicated invasive coronary angiography. The diagnostic performance of the CT-FFR algorithm was assessed with invasive FFR at a threshold of ≤ 0,80 as reference standard. CT-FFR at the location of invasive FFR measurements (CT-FFRatloc) was compared with CCTA, CT-FFR at automatically selected locations (CT-FFRauto) and CT-FFR predictions at the distal end of the vessel (CT-FFRdistal). Results CT-FFRatloc showed good correlation with invasive FFR measurements (r=0,67) and demonstrated improved diagnostic performance compared to CCTA diameter stenosis ≥50%. alone both at lesion level (AUC= 0,85 vs 0,63, p<0,01) and at patient level (AUC=0,87 vs 0,74, p<0,01). Both specificity (84% vs 40%, p=0,01) and accuracy (81% vs 63%, p<0,01) of CT-FFRatloc was significantly increased compared to CCTA. CT-FFRauto performed equally well as CT-FFRatloc, both at lesion level (AUC= 0,82 vs AUC= 0,85, p=0,27) and patient level (AUC= 0,86 vs AUC= 0,87(p=0,67). When comparing CT-FFRdistal to CT-FFRatloc the overall diagnostic performance was poorer (AUC=0,79 vs 0,85, p<0,01). Conclusions Compared to CCTA, CT-FFRatloc showed improved diagnostic performance, including accuracy and specificity for the detection of hemodynamically significant coronary artery disease in a symptomatic population. Predictions using an automated location algorithm, CT-FFRauto, was comparable to evaluations at the location of invasive FFR measurements, CT-FFRatloc.Figure 1Figure 2