Evaluating machine learning accuracy in detecting significant coronary stenosis using CCTA-derived fractional flow reserve: Meta-analysis and systematic review

Background

The use of machine learning (ML) based coronary computed tomography angiography (CCTA) derived fractional flow reserve (ML-FFR_CT), shortens the time of diagnosis of ischemia considerably and eliminates unnecessary invasive procedures, when compared to invasive coronary angiography with invasive FFR (iFFR). This systematic review aims to summarize the current evidence on the diagnostic accuracy of (ML-FFR_CT) compared with iFFR for diagnosis of patient- and vessel-level coronary ischemia.

Methods

To identify suitable studies, comprehensive literature search was performed in PubMed, the Cochrane Library, Embase, up to August 2023. The index test was ML derived FFR and studies with diagnostic test accuracy data of ML-FFR_CT at a threshold of 0.8 were included for the review and meta-analysis. Quality of evidence was assessed using QUADAS-2 checklist.

Results

After full text review of 230 identified studies, 17 were included for analysis, which encompassed 3255 participants (age 62.0 ± 3.7). 8 studies reported patient-level data; and 12, vessel-level data. With iFFR as the reference standard, the pooled patient-level sensitivity, specificity, and area-under-curve (AUC) of ML-FFR_CT were 0.86 [95 % CI: 0.79, 0.91], 0.87 [95 % CI: 0.76, 0.94], and 0.92 [95 % CI: 0.89–0.94], respectively; and pooled vessel-level sensitivity, specificity, and AUC, 0.80 [95 % CI: 0.74–0.84], 0.84 [95 % CI: 0.77–0.89), and 0.88 [95 % CI: 0.85–0.91], respectively.

Conclusions

This systemic review demonstrated the favourable diagnostic performance of ML-FFR_CT against standard iFFR, although heterogeneity exists, providing support for the use of ML-FFR_CT as a triage tool for non-invasive screening of coronary ischemia in the clinical setting.