Influence of diabetes mellitus on the diagnostic performance of machine learning-based coronary CT angiography-derived fractional flow reserve: a multicenter study.

Objectives: To examine the diagnostic accuracy of machine learning-based coronary CT angiography-derived fractional flow reserve (FFR) in diabetes mellitus (DM) patients.

Methods: In total, 484 patients with suspected or known coronary artery disease from 11 Chinese medical centers were retrospectively analyzed. All patients underwent CCTA, FFR and invasive FFR. The patients were further grouped into mild (25~49 %), moderate (50~69 %), and severe (≥ 70 %) according to CCTA stenosis degree and Agatston score < 400 and Agatston score ≥ 400 groups according to coronary artery calcium severity. Propensity score matching (PSM) was used to match DM (n  = 112) and non-DM (n  = 214) groups. Sensitivity, specificity, accuracy, and area under the curve (AUC) with 95 % confidence interval (CI) were calculated and compared.

Results: Sensitivity, specificity, accuracy, and AUC of FFR were 0.79, 0.96, 0.87, and 0.91 in DM patients and 0.82, 0.93, 0.89, and 0.89 in non-DM patients without significant difference (all p > 0.05) on a per-patient level. The accuracies of FFR had no significant difference among different coronary stenosis subgroups and between two coronary calcium subgroups (all p > 0.05) in the DM and non-DM groups. After PSM grouping, the accuracies of FFR were 0.88 in the DM group and 0.87 in the non-DM group without a statistical difference (p > 0.05).

Conclusions: DM has no negative impact on the diagnostic accuracy of machine learning-based FFR.

Key Points: • ML-based FFR has a high discriminative accuracy of hemodynamic ischemia, which is not affected by DM. • FFR was superior to the CCTA alone for the detection of ischemia relevance of coronary artery stenosis in both DM and non-DM patients. • Coronary calcification had no significant effect on the diagnostic accuracy of FFR to detect ischemia in DM patients.