Machine learning CT-FFR for physiologic decision-making in diffuse coronary artery disease
Z An, Z Ma, M Zhang, L Feng, S Yang, M Zhang, C Tu, X Song, H ZhangAbstract
Background
Diffuse coronary artery disease (CAD) is characterized by gradual, long-segment pressure loss, which may not be fully captured by anatomy-based stenosis severity alone. Computed tomography-derived fractional flow reserve (CT-FFR) provides noninvasive physiological assessment, but its performance in diffuse disease and its relationship with treatment strategy and outcomes remain incompletely defined.
Purpose
To evaluate the agreement and diagnostic performance of machine learning (ML)-based CT-FFR versus invasive pressure wire-based FFR in diffuse versus non-diffuse CAD, and to explore its associations with revascularization decisions and prognosis.
Methods
This multicenter retrospective study included consecutive patients from 3 hospitals in China. CCTA DICOM datasets and invasive FFR data acquired within 30 days were retrospectively collected. ML-based CT-FFR was centrally computed by offline post-processing in the Keya Medical core laboratory. Diffuse disease was defined as lesion length ≥20 mm (patient-level diffuse: ≥1 diffuse vessel). Vessel-level and patient-level analyses used invasive FFR ≤0.80 as the reference standard. Treatment-decision matching (0.80 threshold) was assessed against actual management, and major adverse cardiovascular events (MACE) were analyzed using Kaplan–Meier and multivariable Cox models.
Results
Among 328 patients, 136 (41.5%) had diffuse CAD; 374 vessels were included for vessel-level analysis. CT-FFR showed strong correlation with invasive FFR at both vessel and patient levels (vessel: Spearman ρ=0.834, n=374, p<0.001; patient: Spearman ρ=0.828, n=328, p<0.001), with stronger correlations in diffuse than non-diffuse disease. Diagnostic discrimination was excellent overall (AUC 0.941 at both vessel and patient levels) and remained robust in diffuse CAD (patient-level AUC 0.971). Concordance between physiology-guided recommendations and actual revascularization was similar across diffuse and non-diffuse strata, with discordance predominantly driven by nonconcordant revascularization rather than nonconcordant deferral. In 208 patients with follow-up (48 MACE), Kaplan–Meier curves stratified by CT-FFR threshold did not separate significantly, and CT-FFR was not independently associated with MACE in adjusted Cox analysis (HR 1.33 per 0.10-unit change, 95% CI 0.90–1.94; p=0.15), likely reflecting attenuation of baseline physiological risk gradients by downstream revascularization.
Conclusions
ML-based CT-FFR closely agrees with invasive FFR and maintains strong diagnostic performance in both diffuse and non-diffuse CAD, with particularly favorable performance in diffuse disease. These findings support CT-FFR as a practical noninvasive physiological tool to inform revascularization planning, especially when anatomy alone may underestimate ischemic burden in diffuse lesions.Correlation analysis and AUC curve