FFR vs iFR for clinical outcomes: insights from machine-learning risk stratification
M Vilela, D F Ferreira, D C Cazeiro, J C Cravo, C O Oliveira, T R Rodrigues, A R F Francisco, J M Marques, P C F Ferreira, C J Jorge, P C Cardoso, F P Pinto, M M MenezesAbstract
Introduction
Coronary revascularization guided by physiological assessment improves outcomes versus angiography alone, and trials have shown iFR to be noninferior to FFR at 1 year. Yet 5-year results from DEFINE-FLAIR (higher mortality with iFR) and iFR SWEDEHEART (neutral) have created uncertainty about long-term differences. We analyzed a large clinical cohort to compare FFR and iFR, evaluate prognostic performance, and apply machine learning to identify high-risk phenotypes among patients with negative physiological measurements.
Methods
All patients undergoing FFR or iFR assessment from 2012–2022 were included. Clinical outcomes comprised 5-year all-cause mortality, acute coronary syndrome (ACS), and major adverse cardiovascular events (MACE: death, nonfatal myocardial infarction, or unplanned revascularization). Outcomes were evaluated with Kaplan–Meier analysis and Cox proportional hazards models. In patients with negative iFR values, unsupervised machine-learning clustering identified subgroups, which were analyzed for 5-year outcomes.
Results
We analyzed 1,314 physiological measurements from 915 patients (349 FFR; 566 iFR) with a mean 5.9-year follow-up. Baseline characteristics were similar and comparable to DEFINE-FLAIR, though iFR patients more often presented with ACS and had less prior PCI. All-cause mortality was higher with iFR (HR 1.52; CI 1.17–1.98), while rates of ACS, MACE, and unplanned revascularization were similar. Survival curves separated between 12–24 months, mirroring DEFINE-FLAIR. Physiologically positive lesions were numerically lower with iFR. To identify high-risk individuals within the iFR-negative population, we applied machine-learning clustering based on features distinguishing true from false negatives, identifying three clusters. Since clusters 2 and 3 showed no outcome differences, they were merged as low-risk, while cluster 1 was classified as high-risk. In five-year survival analyses, excess adverse events occurred in the high-risk group, which had higher rates of death (HR 2.6; p=0.006) and MACE (HR 1.83; p=0.03). This group was older with worse renal function, more anatomical and clinical complexity, and heavier comorbidity burden. Using the strongest prognostic variables, we built software that automatically identifies high-risk patients through an integrated risk-classification interface.
Conclusion
In this large real-world cohort, iFR was associated with higher 5-year mortality despite similar rates of ACS, MACE, and revascularization, echoing divergence seen in DEFINE-FLAIR. Machine-learning identified a high-risk phenotype within the iFR-negative population that accounted for the excess events. These findings challenge the assumption that a negative iFR uniformly denotes low risk and highlight the need for closer follow-up and potentially intensified management in selected patients.