Between December 2008 and January 2020, we conducted a single-center longitudinal study in an EACVI-accredited imaging laboratory, with a retrospective screening of all consecutive symptomatic patients with newly diagnosed obstructive CAD on CCTA, defined by the presence of at least one ≥ 50% stenosis (5, 14). Patients with previously known CAD before index CCTA were excluded. All patients with moderate renal failure defined by glomerular rate <60 ml/min/1.73 m² were excluded to avoid any confounding factor in the analysis of coronary plaque composition. Those symptomatic patients with ≥1 coronary stenosis on CCTA and further referred for stress CMR (within 3 months after index CCTA) to evaluate the significance of that stenosis were included. The flowchart of the study is depicted in Figure 1. Symptomatic patients were defined by the presence of angina or dyspnea on exertion. In the first place, patients with a high-grade > 90% stenosis on CCTA were directly referred for invasive coronary angiography without stress CMR exam. The main exclusion criteria were known CAD, contraindication to CCTA, CMR, or dipyridamole (detailed list in Supplementary Material 1). Clinical data including symptoms were collected according to the medical history and clinical examination on the day of CCTA. All patients gave informed written consent for CCTA, CMR, and enrolment in the clinical research study. The study was approved by the local Ethics Committee of our Institutions and conducted in accordance with the Declaration of Helsinki. This study followed the STROBE reporting guidelines for cohort studies. Clinical, CCTA, and CMR data were prospectively recorded into a dedicated database (Clinigrid software, Hemolia, France).

The follow-up consisted of a clinical visit as part of usual care (67%) or by direct contact with the patient or the referring cardiologist (33%). Data collection was ended on January 2021. Cardiovascular (CV) events were checked by medical reports collected from the affiliated hospitals. The primary composite endpoint was the occurrence of at least one of the combined major adverse clinical events (MACE), defined as CV mortality or non-fatal MI. The secondary endpoints were CV mortality and all-cause mortality. All these clinical events were defined according to standardized definitions (15, 16), and are detailed in Supplementary Material 2. Three patients who experienced peri-procedural events after percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG) <90 days after the index CCTA examination were excluded.

Between December 2008 and January 2020, all CCTA studies were performed on multidetector CT scanners with ≥64 detector rows (Brilliance 64, Philips Healthcare, Eindhoven, Netherlands, between 2008 and 2011; Philips iCT 128, between 2012 and 2018; Aquilion One Genesis, Canon Medical Systems, Otawara, Japan from 2019 to 2020), and the imaging protocol adhered to the Society of CV Computed Tomography guidelines on appropriateness and performance of CCTA available at the time of scanning (4, 17–19). Each CCTA protocol is detailed in Supplementary Material 3 and radiation exposure assessment is described in Supplementary Material 4.

Coronary computed tomography angiography data were interpreted using multi-planar reconstruction and maximum intensity projections. The data were analyzed based on the 16-segment coronary artery model (20). Coronary segments were scored visually for the presence and composition of coronary plaque and degree of luminal stenosis. In each coronary segment, coronary atherosclerosis was defined as any tissue structure > 1 mm either within the coronary artery lumen or adjacent to the coronary artery lumen that could be discriminated from surrounding pericardial tissue, epicardial fat, and the vessel lumen itself, as previously described (21). In each coronary artery segment, plaques were classified as non-calcified, mixed, or calcified, as previously defined (5). Following the CAD-RADS classification (22), the severity of CAD was categorized according to the highest value of stenosis of the diameter among segments: normal (0% luminal stenosis), non-obstructive CAD (1–49% luminal stenosis), obstructive CAD (50–69% luminal stenosis), or severe obstructive CAD (≥70% luminal stenosis). Within the category of obstructive CAD, we further divided CCTA findings as 1-, 2-, or 3-vessel disease/left main according to the number of major epicardial vessels with the presence of ≥50% stenosis. The presence of ≥50% stenosis in the left main coronary artery was considered a 3-vessel disease equivalent. The number of segments with any plaque or stenosis, a specific plaque composition, specific luminal stenosis, or a specific topography of the plaque (proximal vs. no proximal) were assessed for each patient, following the same method analysis of previous studies (5, 14).

The detailed stress CMR protocol has been previously published (23, 24) and is described in Supplementary Material 5. Briefly, CMR was performed on a 1.5 T scanner (MAGNETOM Espree and Aera, Siemens Healthcare, Erlangen, Germany). Vasodilation was induced with dipyridamole injected at 0.84 mg/kg over 3 min. Then, a bolus of gadolinium-based contrast agent (Dotarem^®, Guerbet, France, 0.1 mmol/kg) was injected at a rate of 5 ml/s. Stress perfusion imaging was performed using an ECG-triggered saturation-prepared balanced steady-state free-precession sequence (Siemens, Erlangen, Germany). A series of six slices (four short-axis views, a two-chamber, and a four-chamber view) were acquired every other heartbeat. Then, 10 min after contrast injection, a breath-hold contrast-enhanced 3D T1-weighted inversion-recovery gradient-echo sequence was acquired to detect late gadolinium enhancement (LGE). CMR sequence parameters are detailed in Supplementary Material 6.

Left ventricular volumes and functions were quantified on the short-axis cine stack. Stress perfusion and LGE images were evaluated according to the 17-segment model of the American Heart Association (25). The analysis of perfusion images was done visually by two experienced physicians blinded to clinical and follow-up data. Inducible ischemia was defined as a subendocardial perfusion defect that (1) occurred in at least one myocardial segment, (2) persisted for at least three phases beyond peak contrast enhancement, (3) followed a coronary distribution, and (4) in the absence of co-location with LGE in the same segment (8). An unrecognized MI was defined by LGE with ischemic patterns defined by subendocardial or transmural LGE (26). LGE with non-ischemic patterns was defined by any location that did not involve the subendocardium and was not transmural. For LGE with ischemic patterns, a myocardial segment was considered viable if LGE thickness was <50% and non-viable when LGE thickness was ≥ 50% of the myocardial wall (27). The total number of ischemic and LGE segments was assessed for each patient. Mild, moderate, and severe ischemia were defined as the involvement of 1–2, 3–5, and ≥6 myocardial segments, respectively (8).

A total of 44 clinical characteristics and CCTA/CMR imaging data were determined (Supplementary Material 7). The absence of collinearity between those 44 baseline variables was verified by principal component analysis (PCA). After the exclusion of collinear variables, 17 categorical variables were selected for the clustering model and the definition of phenogroups which were as follows: age, gender, obesity (body mass index ≥ 30 kg/m²), dyslipidemia, diabetes mellitus, hypertension, current or former smoker, family history of CAD, presence of typical angina, dyspnea on exertion, presence of atrial fibrillation (AF) on 12-lead ECG, history of peripheral arterial disease (PAD) defined by revascularization procedures involving the peripheral arterial circulation (15), presence of LV dilatation defined by LV end-diastolic volume indexed (LVEDVi > 100 ml/m²) (28), presence of LV systolic dysfunction defined by LV ejection fraction (LVEF) value <50% (28), ≥1 proximal segment with stenosis > 50%, rate of segments with non-calcified plaques ≥ 50%, the presence of LGE, and the presence of inducible ischemia. An unsupervised hierarchical clustering of principal components (HCPC function, from FactoMineR package, Vienna, Austria) algorithm was conducted using two steps: a multiple correspondence analysis (MCA) by which the principal components were obtained, followed by a hierarchical clustering analysis using Euclidean distance measures. Notably, the use of MCA was justified by the fact that all continuous variables included for clustering analysis were further classified into categorical variables. In addition, four dimensions were retained in the MCA output. The optimal number of clusters was determined based on the gain in within-inertia (inside group variance) and using the Nbclust package (Vienna, Austria). A detailed description of the used cluster analysis methods is provided in Supplementary Material 8.

Descriptive results were presented as percentages for categorical data and M ± SD or median [interquartile range (IQR)] for continuous variables, depending on the normality of their distribution. Comparisons between clusters were analyzed by ANOVA for numeric and chi-square test or Fisher exact test, as appropriate. The over- or under-representation of variables in each phenogroup were assessed by v-test, based on the hypergeometric distribution (HCPC function, from FactoMineR package, Vienna, Austria). Cox proportional hazards methods were used to assess the prognostic significance of each phenogroup and the prognostic value of inducible ischemia in each cluster. The assumption of the proportional HR was verified. The additional predictive value of phenogrouping for predicting MACE was calculated using Harrell's C-statistic increment, continuous net reclassification improvement (NRI), and the integrative discrimination index (IDI). A two-tailed p-value <0.05 was considered statistically significant. Statistical analysis and clustering were performed using R software, version 4.0.3 (R Project for Statistical Computing, Vienna, Austria).

The flowchart of the study is presented in Figure 1. Overall, 2,015 symptomatic patients with newly diagnosed obstructive CAD on CCTA and further referred for stress CMR completed the clinical follow-up and constituted our study cohort. The median (IQR) delay between CCTA and stress CMR was 9 (4-12) days. No patient had non-fatal MI, CV mortality, or death between CCTA and stress CMR. Among the 2,015 patients who were symptomatic at the time of CCTA and further referred to stress CMR, 2,001 (99.3%) had similar symptoms at the time of stress CMR. No severe adverse events occurred and detailed safety results are presented in Supplementary Material 9. Baseline patient characteristics and imaging data are shown in Table 1.

Among the 2,015 patients (46.3% men, mean age 70 ± 12.2 years), 64.1% had hypertension, 49.0% had dyslipidemia, 31.6% had diabetes mellitus, 30.2% had a family history of CAD, 25.6% had obesity, and 20.9% were smokers. On CCTA, 94% of patients had 1-vessel obstructive CAD. Among the seven patients with LM stenosis > 50%, none had stenosis evaluated at > 70% in CCTA. By CMR, the study cohort had a mean LVEF of 55 ± 10%.

The presence of stress CMR inducible ischemia was detected in 302 (15%) patients with a mean extent of 2.6 ± 1.6 segments, and LGE was identified in 228 (11.3%) patients (Table 1).

Among 302 patients with ischemia, 287 (95%) underwent coronary angiography. Among those, 250 (82.7%) had obstructive CAD confirmed by invasive angiography and 239 (83.3%) underwent CMR-related coronary revascularization [238 (99.5%) PCI and 1 (0.5%) CABG]. Baseline patient characteristics and outcomes of patients with inducible ischemia are shown in Supplementary Material 10 according to the presence of obstructive CAD as defined by ICA.

The results further showed that 20 clinical and CMR input variables had different contributions in defining the phenogroups (Figure 2). Hierarchical clustering established that three phenogroups yielded the highest gain in inertia (inside group variance) and was suggested by the HCPC function for k-means clustering (Supplementary Material 11). This led to the identification of three phenogroups (Figure 3) with significant differences in their clinical characteristics (Supplementary Material 12). Supplementary Video 1presented the three phenogroups according to the first three principal components in hierarchical clustering assessed on the three axes.

Phenogroup 1 (PG1; n = 854; 42.4%) included the oldest patients (75 ± 10 years), predominantly women (64%), with the lowest proportion of traditional risk factors (57% of hypertension, 38% of dyslipidemia, 14% of diabetes, 15% of family history of CAD, 11% of smoking, and 9% of obesity) compared with other patients. On CCTA, PG1 had the lowest proximal stenosis > 50% rate (6%) and the lowest proportion of non-calcified plaques (5% of patients with ≥ 50% of non-calcified plaques). On CMR, PG1 had the lowest LGE rate (9%) with a preserved mean LVEF value (58 ± 9%) and lower LV dilatation rate (5%).

Phenogroup 2 (PG2; n = 681; 33.8%) included younger patients (69 ± 10 years), predominantly women (61%), with the highest proportion of hypertension (86%), dyslipidemia (70%), diabetes (63%), obesity (56%), and family history of CAD (49%), but a lower rate of smokers (12%) compared with other patients. PG2 had the highest rate of calcified plaques (88%) on CCTA, with a preserved mean LVEF value (58 ± 8%) and the lowest LV dilatation rate (3%) on CMR.

Phenogroup 3 (PG3; n = 480; 23.8%) included the youngest patients (65 ± 11 years), predominantly men (75%), with the highest proportion of smokers (51%). PG3 had the highest proximal stenosis > 50% rate (40%) and the highest proportion of non-calcified plaques (18% of patients with ≥ 50% of non-calcified plaques). On CMR, PG3 had the highest LGE rate (15%) with the lowest mean LVEF value (45 ± 10%) and the highest LV dilatation rate (73%) (Supplementary Material 13).

It is worth noting that the association between the proportion of non-calcified plaques and the presence of LGE with ischemic pattern within each phenogroup is depicted in Supplementary Material 14.

During a median follow-up of 6.8 (IQR 5.9–9.2) years, there was 277 (13.7%) MACE, including 149 (7.4%) CV mortality and 128 (6.4%) non-fatal MI. Furthermore, 184 all-cause mortality (9.1%) were recorded. Annualized event rates were 3.3% for MACE, 1.8% for CV mortality, and 1.9% for all-cause mortality.

Kaplan-Meier analysis showed significant differences for the occurrence of MACE across phenogroups, CV mortality, and all-cause mortality (all p < 0.001, Figure 4). PG1 has the lowest while PG3 has the highest risk of MACE (0.9 vs. 3.9 events per 100 patient-years; p < 0.001), CV mortality (0.4 vs. 2.1 events per 100 patient-years; p < 0.001), and all-cause mortality (0.5 vs. 2.6 events per 100 patient-years; p < 0.001) (Table 2). In multivariate Cox regression analysis, PG2 and PG3 were independently associated with a higher incidence of MACE compared with PG1 (HR: 1.98, 95% CI: 1.32–2.97 and HR: 2.15, 95% CI: 1.40–3.31, respectively, both p < 0.001) (Table 2; Supplementary Material 15). The addition of phenogroup characterization improved the prediction of MACE beyond clinical, CCTA and CMR predictors (C-statistic improvement: 0.04, p = 0.028; IDI = 0.033, p = 0.019; NRI = 0.21, p = 0.024) (Table 3). To note, the distribution of CCTA findings according to the occurrence of MACE in patients with ischemia is presented in Supplementary Material 16.

In each PG, the presence of inducible ischemia by CMR was associated with the occurrence of MACE (PG1, HR = 3.09, 95% CI, 1.7–5.62; PG2, HR = 3.62, 95% CI, 2.31–5.7; PG3, HR = 3.55, 95% CI, 2.3–5.49; all p < 0.001; Figure 5). Using CCTA data, the number of proximal segments with non-calcified plaques was associated with the occurrence of MACE within each PG (PG1, HR = 1.57, 95% CI, 1.31–1.89; PG2, HR = 1.73, 95% CI, 1.22–2.45; PG3, HR = 2.06, 95% CI, 1.63–2.53; all p < 0.001; Supplementary Material 17).

Although patients were included prospectively, the study design was retrospective with 7.8% of patients lost to follow-up. Baseline data for medications were not collected. In the absence of contraindications, all patients with obstructive CAD on CCTA received optimal medical treatment, including statins, as recommended by current guidelines (2). Although a recent study has emphasized the effects of statins on plaque composition (35), the detailed medical regimen was not prospectively collected and this question was beyond the scope of the study. Symptoms were assessed by the sole presence of symptomatic angina or dyspnea on exertion without standardized classification. Patients with a high-grade > 90% stenosis on CCTA were referred for invasive coronary angiography and excluded from analysis, limiting the extrapolation of the findings in this group of patients. The exclusion of patients with moderate to severe renal failure limited the extrapolation of results to a general patient population. The analysis of CMR scans was visual, which represented the most widely accepted clinical method with optimal diagnostic accuracy. The coronary artery calcium score was not systematically performed before CCTA in symptomatic patients. Although CCTA protocol followed current guidelines, specific and quantitative plaque analysis might be affected. In addition, the quantitative assessment of the low-attenuation plaques by CCTA was not systematically performed in this study. Dipyridamole was used as a stress agent mainly because of medico-economic reasons and a very close efficacy/safety profile compared to adenosine. The current analysis did not address the question of whether unsupervised learning based on CCTA alone could predict outcomes or functionally significant CAD as assessed by stress CMR. The current study was not designed to assess which stress CMR or CCTA parameter was most powerful in predicting MACE. Although recent studies have highlighted the role of biomarkers such as Troponin T in risk stratification (36), these data were not available in this study. Although the current approach might represent a major shift from traditional studies, it was a novel attempt toward a more personalized approach to patients with high CV risk. Whether automated unsupervised phenogrouping of CAD patients could improve clinical decision-making should be further explored in prospective studies.