Automatic Anatomical and Functional Classification of Coronary Arteries With Artificial Intelligence.

Coronary disease is the leading cause of death in the world with 18 million deaths according to the WHO The exploration of chest pain suggesting coronary artery disease now gives preference to coronary CT angiography (CCTA) for its high sensitivity (95%) in the non-invasive detection of patients with coronary artery disease.

This is a class I recommendation with a high level of evidence according to the AHA/ACC recommendations of November 2021. The interpretation of the images aims to define the degree of stenosis of the vessels, a stenosis >= 50% being likely to limit coronary flow. However, the degree of stenosis, especially between 40%-90%, is not directly correlated with its functional impact. This must therefore be assessed by an invasive intra-arterial examination during coronary angiography. The measurement of the post-stenosis flow (FFR) may indicate a stent or bypass operation when the FFR is less than 0.8.

260,000 coronary angiographies are performed each year in France, of which two-thirds do not lead to an interventional procedure. CCTA coupled with an intelligent predictive analysis system could reduce this rate of invasive examinations that do not require an interventional procedure. Various decision support systems have been developed recently using AI methods, either for stenosis assessment or for FFR estimation. Their overall sensitivity is insufficient, mainly because their training base is small (<100 cases) and they have not been validated in a multicenter setting. On the contrary, the radiology department of the IMM has built a large base of images (n=5000) of CCTA from various machines, qualified by an expert and associated with FFR values. This learning base feeding a deep learning system has very good predictive performances of stenosis and FFR on a new test base of CCTA images alone. Obtaining these 2 parameters in a single acquisition would enhance the radiologist's accuracy and avoid the need for invasive FFR. It therefore seems appropriate to reinforce this system with a multicenter feed and to perform an external validation on an independent sample.

OBJECTIVES Main : Predictive performance, at the coronary vessel level, of an intelligent Coronary CT AI based image analysis system on the detection of a stenosis requiring intervention, versus invasive coronary angiography with reference measurement (FFR).

Secondary:- predictive performance regarding the indication for intervention at the patient level (i.e., the synthesis of all the assessments of his or her vessels) - medico-economic analysis of the cost-effectiveness type comparing two diagnostic strategies (CCTA+AI, vs. usual care = CCTA + invasive FFR) in terms of effectiveness (shortening of the time to obtain treatment, unnecessary invasive coronary angiography avoided, complications avoided), cost and incremental cost-effectiveness ratio

JUDGEMENT CRITERIA Primary: This criterion is calculated on the validation sample. Sensitivity at the vessel level will be calculated as the ratio of the number of stenotic vessels classified as interventional by the AI system to the total number of stenotic vessels classified as interventional by the reference method. The other metrics (specificity, likelihood ratios, prevalence and predictive errors) will be calculated, all with their 95% confidence intervals.

Secondary:- Predictive performance regarding the indication of intervention at the patient level: sensitivity, specificity and other metrics- Cost per complication avoided. It will be calculated from the Differential Cost Outcome Ratio (DCOR), which is the difference in costs (from a Medicare perspective) divided by the difference in the number of coronary complications between the two strategies studied, and will be supplemented by sensitivity analyses.

METHODOLOGY A multicenter study that will collect CCTA images and invasive FFR measurements from consecutive patients under standardized conditions. The total sample will be randomly partitioned into a learning sample representing approximately 60% of the population and a validation sample (40%). The reference results will be obtained:- for the % of stenosis, at the vessel level: by consensus of independent experts (Delphi method on dedicated WWW site) on an a posteriori examination of the images (vessel, anonymized, blinded to the local interpretation and FFR) - for FFR at the vessel level: by a standardized invasive procedure performed as soon as stenosis is ≥40% according to the local radiologist's estimation; the result is first quantitative and then dichotomized at the FFR threshold ≤0.8- for patient classification (indication for stenting/bypassing or not) by the conjunction of reference results obtained vessel by vessel: indication for stenting/trimming if at least one vessel has an FFR≤0.8; no indication if no vessel has stenosis ≥ CAD-RADS 3 (≥50%) or otherwise has an FFR value ≤0.8

ELIGIBILITY CRITERIA- Adult patient,- with no diagnosed coronary status or history of stenting or bypass surgery- whose CCTA evaluation by the local radiologist results in at least an intermediate stenosis ≥40% on at least one vessel with indication for FFR measurement.- Who has not expressed opposition to the use of their data.

RESEARCH SCHEME 11 participating centers

NUMBER OF SUBJECTS A minimum is a sensitivity of 95% on the test basis. To estimate this parameter with an accuracy of +-2.5% (92.5%-97.5%), with a 5% two-sided risk of error, 300 observations (at the vessel level) are required, that is, at a rate of approximately 1.5 vessels with lesions per patient, 200 patients with an indication for intervention. Knowing that 2/3 of the FFR will be negative, the number of new patients with an FFR is 600 and taking into account 10% of uninterpretable images, 670 new patients for the test base (40% of the total). The number of patients for the learning base (60%) must be 1000. This means a total recruitment of 1670 patients. 340 patients for the learning base will be included retrospectively and the remaining 1330 patients to complete the learning base and the validation base will be included prospectively.

RESEARCH QUALIFICATION

Category 3 research involving the human person:

For patients who will be included in the retrospective part: only their data from their medical records will be used.

Patients who will be included in the prospective part will additionally complete the EQ5D5L questionnaire before coronary angiography and at the end of the patient's participation (4 months after the CCTA).

STUDY TIME FRAME Patient participation time: 4 months Entry time: 2 years