Machine learning and computational fluid dynamics derived FFRCT demonstrate comparable diagnostic performance in patients with coronary artery disease; A Systematic Review and Meta-Analysis

Abstract

Background: As a new noninvasive diagnostic technique, computed tomography-derived fraction flow reserve (FFRCT) has been used to identify hemodynamically significant coronary artery stenosis. FFRCT can be calculated using computational fluid dynamics (CFD) or machine learning (ML) approaches. It was hypothesized that ML-based FFRCT (FFRCT_ML) has comparable diagnostic performance with CFD-based FFRCT (FFRCT_CFD). We used invasive FFR as the reference test to evaluate the diagnostic performance of FFRCT_ML vs. FFRCT_CFD.

Methods: We searched PubMed, Cochrane Library, EMBASE, WOS, and Scopus for articles published until March 2024. We analyzed the synthesized sensitivity, specificity, and diagnostic odds ratio (DOR) of FFRCT_ML vs FFRCT_CFD at both the patient and vessel levels. We generated summary receiver operating characteristic curves (SROC) and then calculated the area under the curve (AUC).

Results: This meta-analysis included 23 studies reporting FFRCT_CFD diagnostic performance and 18 studies reporting FFRCT_ML diagnostic performance. In the FFRCT_CFD group, 2501 patients and 3764 vessels or lesions were analyzed. In the FFRCT_ML group, 1323 patients and 4194 vessels or lesions were analyzed. Our results showed that at the per-patient level, FFRCT_CFD and FFRCT_ML had comparable pooled specificity (Z ​= ​-0.59, P ​= ​0.55) and AUC (P ​= ​0.5). At the per-vessel level, FFRCTCFD and FFRCTML also showed comparable specificity (Z ​= ​0.94, P ​= ​0.34), DOR (Z ​= ​0.7, P ​= ​0.48), and AUC (P ​= ​0.74). However, the sensitivity of FFRCT_ML was significantly lower compared to FFRCT_CFD at both patient (Z ​= ​-3.85, P ​= ​0.0001) and vessel (Z ​= ​-2.05, P ​= ​0.04) levels.

Conclusion: The FFRCT_ML technique was comparable to standard CFD approaches in terms of AUC and specificity. However, it did not achieve the same level of sensitivity as FFRCT_CFD.

Keywords: Angiography; Computational fluid dynamics; Computed tomography; Coronary artery disease; Fractional flow reserve; Machine learning.