Truace
Subscribe

Performance of artificial intelligence-based coronary CT angiography-derived fractional flow reserve across cohorts with different prevalence of obstructive coronary artery disease

Background Coronary computed tomography angiography (CCTA) provides accurate anatomical assessment of coronary stenosis severity but has limited ability to determine lesion-specific ischemia as assessed by invasive fractional flow reserve (FFR). We aimed to evaluate the diagnostic performance of an artificial intelligence-based deep learning CCTA-derived FFR model (CT-FFR AI ) in cohorts undergoing CCTA and invasive FFR as part of a study protocol and a real-world cohort undergoing invasive FFR based on suspecte…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2078· cite & verify
Performance of artificial intelligence-based coronary CT angiography-derived fractional flow reserve across cohorts with different prevalence of obstructive coronary artery disease
"Cardiac CT and Coronary Angiography Imaging in Navi Mumbai" by Goleisureintl is licensed under CC BY 4.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/.

In brief

Researchers tested a deep learning tool called CT-FFR AI (CorEx, SPIMED-AI) that estimates fractional flow reserve from coronary CT angiography images. They checked it against invasive FFR measurements in two groups: 208 patients in the PACIFIC-1 study and 202 patients from routine care who had a CCTA lesion of at least 70% stenosis.

In the study cohort, the AI model performed about as well as stenosis-based CCTA reads, while in the real-world cohort it was clearly more accurate. This suggests that the tool's value depends on how common obstructive disease is and on the spread of stenosis severity in the vessels being assessed, so results from one setting may not carry over to another.

Main points

  1. In PACIFIC-1 (612 vessels), per-vessel accuracy was 77% for CT-FFR AI versus 76% for a 50% stenosis threshold and 74% for a 70% threshold, with no significant difference.
  2. In the real-world cohort (325 vessels), CT-FFR AI reached 63% accuracy versus 52% and 57% for the 50% and 70% CCTA thresholds.
  3. The study's added-value measure beyond CCTA was 7.9% (P = 0.256) and 17.7% (P = 0.012) in PACIFIC-1, versus 23.2% in the real-world cohort.
  4. The authors conclude CT-FFR AI may outperform anatomic assessment alone, depending on disease prevalence.

The gain

In a real-world cohort of 202 patients, the AI CT-FFR model identified lesion-specific ischemia more accurately per vessel (63%) than CCTA stenosis thresholds of 50% (52%) or 70% (57%).

The rundown

The two cohorts differed in design. PACIFIC-1 patients had invasive FFR in all major coronary vessels regardless of CCTA findings. The real-world cohort included only patients with at least one CCTA-defined lesion of 70% or more stenosis who went on to invasive angiography with FFR.

The evaluation used an intention-to-diagnose approach, comparing the AI model with CCTA at 50% and 70% stenosis cutoffs, with invasive FFR as the reference standard.

Background context: CCTA gives accurate anatomical assessment of stenosis but has limited ability to determine lesion-specific ischemia, which is the gap the AI-derived FFR is meant to address.

Sources

  1. Peer-reviewedJournal of Cardiovascular Computed Tomography2026-10-19

The debate