Data‐Driven Clustering Approach to Identify Different Phenotypes of Primary Central Nervous System Vasculitis

ABSTRACT

Background

To determine whether hierarchical unsupervised cluster analysis identifies a phenotypic distinction in adult patients with primary CNS vasculitis (PCNSV).

Methods

An agglomerative hierarchical cluster analysis based on the Ward method was conducted, including 153 patients with complete baseline phenotypic characterization in the COVAC' registry.

Results

The hierarchical analysis identified two main clusters. In Cluster 1 (n = 109 patients, 71%), patients more frequently had a motor deficit (p = 0.039), ≥ 1 acute brain infarct (p < 0.001), and ≥ 1 intracranial stenosis on CT or MR angiogram (p < 0.001) than patients in Cluster 2 (n = 44 patients, 29%). Conversely, patients in Cluster 2 more frequently had seizures (p < 0.001), cognitive impairment (p = 0.002), gadolinium‐enhanced parenchymal lesions (p < 0.001), leptomeningeal enhancement (p < 0.001), ≥ 1 cerebral microbleed (p < 0.001), and intracranial hemorrhage(s) (p < 0.001). In multivariable logistic regression, gadolinium‐enhanced parenchymal lesions were significantly associated with Cluster 2 lesions (OR = 35.53 [95% CI: 3.91–322.81], p = 0.002). Conversely, ≥ 1 acute brain infarct was significantly associated with Cluster 1 (OR = 0.003 [95% CI: 0.01–0.03], p < 0.001). A CNS biopsy was positive in 11/40 (28%) patients from Cluster 1 and 35/37 (95%) patients from Cluster 2 (p < 0.001). At 12 months, functional independence (modified Rankin scale score ≤ 2) did not differ between the two groups (p = 0.17). Relapse and mortality rates did not differ between the clusters (p = 0.17 and p = 0.23, respectively).

Conclusion

This unsupervised analysis of a large PCNSV cohort identified two different clinical and radiological phenotypes with different diagnostic work‐ups, which confirms the relevance of distinguishing PCNSV phenotypes according to the sizes of affected vessels.

1. Introduction

Primary central nervous system vasculitis (PCNSV) is a rare inflammatory disorder affecting CNS vessels that is responsible for high morbidity. In the absence of a CNS biopsy showing transmural inflammation of the vessel wall, diagnosis is challenging and often relies on indirect evidence of vascular involvement combined with inflammatory findings on CSF analysis or contrast imaging [1, 2, 3, 4]. In the last two decades, few cohorts with > 60 patients have been reported [5, 6, 7, 8, 9]. Initially, patients were classified according to the methodology used for diagnosis. Patients with a positive biopsy result were compared to those with an angiogram‐based diagnosis, with a negative or no biopsy [10, 11]. Some differences regarding their clinical and radiological presentations were observed between these groups. Given that patients with a positive biopsy often have a negative angiogram, they are considered to present a pattern of small‐vessel (sv) PCNSV. Conversely, those with positive angiograms are considered to present large/medium‐vessel (l/mv) PCNSV. A comparison of patients with sv‐ and l/mv‐PCNSV produced relatively similar results to those of biopsy‐proven and angiogram‐diagnosed patients. Patients with biopsy‐proven or sv‐PCNSV often present with seizures, cognitive impairment, gadolinium‐enhanced parenchymal or leptomeningeal lesions, cerebral microbleeds, and elevated CSF white blood cells. Conversely, angiogram‐diagnosed or l/mv patients often present with focal deficits in relation to brain infarcts on MRI and arterial stenoses on angiograms [8, 12, 13, 14].

Although practical, this artificial phenotypic distinction includes some important biases. First, we cannot exclude the possibility that patients with negative angiograms were more likely to be referred for a biopsy than patients with positive angiograms. Second, some patients had overlapping presentations and may have been misclassified (e.g., positive biopsy and abnormal angiogram). Finally, important data were lacking for some patients (gadolinium‐injected sequences, biopsy performance), and their results may have changed the phenotypic distinction.

Taken together, these observations require validation using an unsupervised analysis of easily accessible baseline variables. In this study, we aimed to determine whether a hierarchical cluster analysis conducted in the COVAC' registry may identify PCNSV subgroups that share similarities with those suggested and reported in cohorts that did not use clustering methods.

2. Patients and Methods

2.1. The COVAC' PCNSV Registry

The COVAC' (cohort of primary CNS vasculitis—cohorte de vascularite cérébrale primitive) registry of patients with PCNSV was created in 2010 and first described in 2014 [10]. Initially, the included patients were diagnosed with PCNSV using the criteria of Calabrese and Malek, which were the standard criteria used at the time [15]. With increasing knowledge, several important differential diagnoses, such as intracranial atherosclerotic disease or reversible cerebral vasoconstriction syndrome (RCVS), have been identified as challenging mimickers in patients with PCNSV who are diagnosed exclusively based on a positive angiogram [16, 17]. In 2021, the registry was thus updated, and each included patient was individually reanalyzed. The inclusion criteria were designed to reduce the risk of including PCNSV mimickers. For inclusion in the registry, patients had to satisfy the following criteria: (1a) vessel wall inflammation on CNS biopsy or (1b) ≥ 1 otherwise unexplained intracranial stenosis with ≥ 1/4 additional findings supportive of vasculitis (inflammatory CSF, concentric vessel wall enhancement, nonischemic parenchymal or leptomeningeal gadolinium enhancement, or rapidly progressive and persistent stenoses) and (2) no secondary etiology of CNS vasculitis (e.g., CNS infection, connective tissue disorder, systemic vasculitis, sarcoidosis, inflammatory bowel disease, or active cancer). Owing to clear phenotypic distinctions between biopsy‐positive PCNSV and amyloid beta‐related angiitis, patients with the latter diagnosis were excluded from this analysis [18]. Similarly, one patient with vasculitis isolated in the spinal cord was excluded. At the time of this study, 193 patients satisfied the inclusion criteria.

The COVAC' registry complies with the Declaration and Helsinki, and the study protocol was approved by the CHU Caen‐Normandie institutional review board (#3786), without the requirement for written informed consent. In France, patients were informed of their inclusion and had the right to request the exclusion of their data. Additional local review board approval was obtained when required by individual sites.

2.2. Model Variables

After discussion (AN, HdB), the following set of variables at baseline was selected for inclusion in the hierarchical cluster analysis: age, sex, motor or sensory deficit, aphasia, visual impairment, seizure, vertigo, ataxia, cognitive impairment, altered level of consciousness, headache, leukocyte level, and protein concentration on CSF analysis, ≥ 1 acute brain infarct, nonischemic parenchymal gadolinium‐enhanced lesion, leptomeningeal gadolinium‐enhanced lesion, ≥ 1 cerebral microbleed, intracranial hemorrhage, subarachnoid hemorrhage, tumor‐like lesion, and ≥ 1 arterial stenosis on CT or MR angiogram.

All the data were completely available for 153 patients. At least one variable was missing in the 40 other patients, including gadolinium‐injected sequences on MRI in 29 patients, T2*‐based MRI sequences in 11 patients, and CSF analysis in two patients.

Data regarding biopsy performance and results were analyzed and compared across the different clusters. We did not include these variables in the unsupervised analysis since the procedure was not performed in all patients and because the indication for a CNS biopsy is probably biased by imaging results.

2.3. Outcomes

The outcomes included relapse, death, and 12‐month modified Rankin scale (mRS) score. Relapse was defined as the occurrence of a new neurological event along with new findings on neuroimaging attributable to PCNSV in a patient who had previously achieved disease remission. Patients with an mRS score ≤ 2 (functional independence) at 12 months were considered to have a favorable prognosis. The outcome variables were not included in the cluster analysis.

2.4. Statistical Analysis

We conducted hierarchical cluster analysis for the 153 patients with complete data (primary analysis) and the 193 patients after imputation of missing data (sensitivity analysis). Agglomerative hierarchical cluster analysis based on Ward's method was used to construct homogeneous clusters of patients (see Data S1). The optimal number of clusters within the studied population was estimated using the “NbClust” R package, with Euclidean distance and the Ward method (Ward D). Twenty‐four indices for determining the number of clusters were computed. According to the majority rule, the best number of clusters was determined. To visualize these clusters, a dendrogram and a cluster plot were created. An ellipse was drawn around each cluster.

The baseline characteristics of patients were described between clusters as counts (percentages) and means (standard deviations) or medians (interquartile ranges) for qualitative and quantitative variables, respectively. These characteristics were compared between clusters using the χ ² or Fisher's exact test for categorical variables and the Student's t test or the Mann–Whitney U test for continuous variables, as appropriate.

Variables that were statistically significant (p < 0.05) in the univariate analysis were considered for inclusion in the multivariable logistic regression analysis. We explored collinearity among variables potentially selected in our multivariable model by calculating the variance inflation factor, with a value above 5 indicating collinearity. We used backward stepwise selection to identify independent variables associated with Cluster 2, with a threshold of p < 0.10 to remain in the model. The goodness‐of‐fit of the logistic regression model was tested with the Hosmer–Lemeshow test.

Favorable 12‐month prognosis, the number of relapses and the number of deaths were compared between clusters using the χ ² test. Cramer's V was estimated as a measure of association between the clusters and the outcomes, ranging from 0 (no association) to 1 (complete association). Overall survival (in all patients) and relapse‐free survival (in patients who achieved remission) were compared between clusters using the Kaplan–Meier method and the log‐rank test.

Sensitivity analyses were performed to evaluate the stability of the findings. For sensitivity analyses, missing data were imputed using the mean or median for quantitative variables depending on the distribution and by creating a third class for categorical variables. The same analyses described above were subsequently performed.

An agreement analysis was performed to compare the clustering of the main and sensitivity analyses, using McNemar's test as well as the kappa coefficient and its 95% confidence interval (CI). A kappa coefficient > 0.81 indicated almost perfect agreement, and a coefficient of 1 indicated complete agreement of the clustering.

A p < 0.05 was considered statistically significant; all p values were two‐tailed. We used SAS statistical software, version 9.4 (SAS Institute Inc., Cary, NC, USA); R software, version 4.3.3 (The R Foundation for Statistical Computing); and RStudio version 2023.12.1 Build 402 (2009–2024 Posit Software, PBC).

3. Results

3.1. PCNSV Patient Subgroups

Among the 153 patients, the male/female ratio was 89/64, and the median age was 49 [38–59] years. The diagnosis was based on a positive CNS biopsy in 46 (30%) patients, whereas the 107 others had an imaging‐based diagnosis.

Hierarchical analysis of the 153 patients with complete data revealed two main clusters (agglomerative coefficient: 0.88; dendrogram and cluster plot in Figure 1A,B). The baseline characteristics of each cluster are described and compared in Table 1.

FIGURE 1.

Cluster analysis. (A) Cluster dendrogram. Hierarchical clustering revealed two clusters. (B) Cluster plot showing patient distribution in the clusters.

TABLE 1.

Baseline characteristics of patients with primary CNS vasculitis according to their assigned cluster.

Baseline characteristics	Cluster 1 (n = 109)	Cluster 2 (n = 44)	p
Demographics
Age, mean ± SD	48 ± 13	46 ± 16	0.43
Female, n (%)	44 (40)	20 (45)	0.57
Clinical manifestations, n (%)
Headaches	74 (68)	24 (55)	0.12
Motor deficit	74 (68)	22 (50)	0.039
Sensory deficit	34 (31)	10 (23)	0.30
Aphasia	38 (35)	10 (23)	0.15
Visual disturbance	38 (35)	8 (18)	0.042
Seizure	7 (6)	24 (55)	< 0.001
Vertigo	18 (17)	3 (7)	0.13
Ataxia	27 (25)	3 (7)	0.013
Cognitive impairment	49 (45)	32 (73)	0.002
Altered level of consciousness	9 (8)	2 (4)	0.52
Cerebrospinal fluid analysis
White blood cell count, median [IQR]	14 [3–58]	9 [2–30]	0.29
Protein concentration (g/L), median [IQR]	0.70 [0.44–1.00]	0.59 [0.39–0.95]	0.34
Imaging findings
≥ 1 Acute brain infarct	104 (95)	3 (7)	< 0.001
Gadolinium‐enhanced parenchymal lesions	17 (16)	38 (86)	< 0.001
Leptomeningeal gadolinium‐enhancement	6 (5)	17 (39)	< 0.001
≥ 1 Cerebral microbleed	18 (17)	26 (59)	< 0.001
Intracerebral hemorrhage	0 (0)	10 (23)	< 0.001
Subarachnoid hemorrhage	9 (8)	8 (11)	0.55
Tumor‐like lesion	2 (2)	17 (39)	< 0.001
≥ 1 Intracranial stenosis on CT‐ or MR‐angiography	87 (80)	4 (9)	< 0.001

Cluster 1 included 109 (71%) patients. Compared with those in Cluster 2 (44 patients, 29%), patients in Cluster 1 more frequently presented with motor deficits (68% vs. 50%, p = 0.039), visual disturbances (35% vs. 18%, p = 0.042), ataxia (25% vs. 7%, p = 0.013), ≥ 1 acute brain infarct (95% vs. 7%, p < 0.001), and ≥ 1 intracranial stenosis on CT or MR angiogram (80% vs. 9%, p < 0.001).

Compared with those in Cluster 1, patients in Cluster 2 were more likely to present with seizures (55% vs. 6%, p < 0.001), cognitive impairment (73% vs. 45%, p = 0.002), nonischemic gadolinium‐enhanced parenchymal (86% vs. 16%, p < 0.001) or leptomeningeal (39% vs. 5%, p < 0.001) lesions, ≥ 1 cerebral microbleed (59% vs. 17%, p < 0.001), intracranial hemorrhage (23% vs. 0%, p < 0.001), and tumor‐like lesions (39% vs. 2%, p < 0.001).

In the multivariable logistic regression analysis (Table 2), gadolinium‐enhanced parenchymal lesions were identified as an independent factor associated with Cluster 2 (OR = 35.53 [3.91–322.81], p = 0.002). Conversely, ≥ 1 acute brain infarct (OR = 0.003 [0.01–0.03], p < 0.001) and ataxia (OR = 0.03 [0.01–0.58], p = 0.02) were independently associated with Cluster 1.

TABLE 2.

Independent baseline factors associated with Cluster 2 after multivariate logistic regression.

Baseline characteristics	OR	95% CI	p
Ataxia	0.03	[0.01–0.58]	0.02
≥ 1 Acute brain infarct	0.003	[0.01–0.03]	< 0.001
Gadolinium‐enhanced parenchymal lesions	35.53	[3.91–322.81]	0.002

Abbreviations: 95% CI, 95% confidence interval; OR, odds ratio.

A CNS biopsy was obtained from more patients in Cluster 2 (37/44, 84%) compared with Cluster 1 (40/109, 37%) (p < 0.0001). The biopsy was positive in 11/40 (28%) patients from Cluster 1 and in 35/37 (95%) patients from Cluster 2 (p < 0.0001).

Patients in cluster 1 with a positive biopsy were compared to patients with a negative biopsy (Table S1). We observed that patients in cluster 1 with a positive biopsy more frequently presented at baseline with cognitive impairment, CSF abnormalities, gadolinium‐enhanced parenchymal lesions, and less frequently showed motor deficits, acute brain infarct(s) and intracranial stenosis on CT‐ or MR‐angiography.

The distribution of clusters according to the diagnosis group (biopsy‐proven diagnosis versus imaging‐based diagnosis) is shown in Figure 2. Among patients with a biopsy‐proven diagnosis, 77% belonged to Cluster 2. In contrast, 92% of patients with an imaging‐based diagnosis belonged to Cluster 1.

FIGURE 2.

Diagram showing the cluster distributions in patients with biopsy‐proven and angiogram‐based diagnoses.

3.2. Outcomes

Patients from Clusters 1 and 2 were followed up for a median of 35 (24–105) months and 43 (23–88) months, respectively (p = 0.92).

The proportion of patients who achieved 12‐month functional independence did not differ between Clusters 1 and 2 (63% vs. 75%, p = 0.17).

The number of relapsing patients did not differ between the two clusters (23 [21%] in Cluster 1 vs. 14 [32%] in Cluster 2, p = 0.17). Relapse‐free survival is shown in Figure 3 (log‐rank p = 0.16).

FIGURE 3.

Kaplan–Meier relapse‐free survival of the 153 patients with complete datasets in the registry from Cluster 1 (blue curve) and Cluster 2 (red curve).

Death occurred in 16% of patients from Cluster 1 and 11% of patients from Cluster 2 (p = 0.50). Survival curves are shown in Figure S1 (log‐rank p = 0.48).

3.3. Sensitivity Analyses

Sensitivity analyses were performed on the entire cohort (n = 193). After imputation, the hierarchical cluster analysis distinguished 142 (74%) patients in Cluster 1 and 51 (26%) in Cluster 2 (Table S2). The agreement between the clustering obtained with the complete case and imputation analyses was almost perfect (kappa coefficient = 0.92 [95% CI: 0.85–0.99], Table S3).

The significant differences in clinical and neuroimaging findings between Clusters 1 and 2 were consistent in the imputation analysis. In addition, females were more likely to belong to Cluster 2 (53% vs. 36% in Cluster 1, p = 0.034), and aphasia was more common in Cluster 1 (37% vs. 20% in Cluster 2, p = 0.026). In the multivariable logistic regression analysis (Table S4), factors independently associated with Clusters 1 or 2 remained the same.

A positive CNS biopsy was obtained from 13/49 (27%) patients in Cluster 1 and 41/44 (93%) patients in Cluster 2 (p < 0.0001).

At 12 months, patients in Cluster 2 were more likely to achieve functional independence (61% in Cluster 1 versus 80% in Cluster 2, p = 0.014). More relapses were observed in patients in Cluster 2 (37% vs. 20% in Cluster 1, p = 0.018). Figure S2 shows the relapse‐free survival (log‐rank p = 0.006).

The death rate did not differ between the two groups (17% in Cluster 1 vs. 10% in Cluster 2, p = 0.23). Overall survival is shown in Figure S3 (log‐rank p = 0.21).

4. Discussion

Using unsupervised agglomerative hierarchical cluster analysis, the present study identified two distinct phenotypes of patients with PCNSV. Patients in Cluster 1 preferentially presented with stroke, with ≥ 1 intracranial stenosis on 80% of the CT or MR angiograms performed, along with ≥ 1 acute brain infarct and subsequent focal neurological deficits. In contrast, patients in Cluster 2 presented mainly with encephalopathic findings, that is, seizures or cognitive impairment associated with gadolinium‐enhanced parenchymal lesions in more than 85% of the patients. Intracranial stenoses and acute brain infarcts were observed in less than 10% of these patients. Gadolinium‐enhanced parenchymal lesions were the strongest predictive factor associated with Cluster 2, whereas ≥ 1 acute brain infarct was the strongest predictive factor for belonging to Cluster 1. The outcomes did not differ in the complete case analysis (153 patients). However, after missing data imputation (193 patients), patients from Cluster 2 had a better prognosis, indicated by an mRS ≤ 2 at 12 months, although they relapsed more frequently than patients from Cluster 1.

The main strength of this study was the ability to distinguish disease phenotypes using an unsupervised analysis with complete data from 153 patients. Sensitivity analysis confirmed the observed results and enhanced the analytical power with the addition of 40 patients, allowing us to demonstrate differences in outcomes. Importantly, after missing data imputation, the concordance of the clustering remained almost perfect.

Our results are in accordance with the published literature. Our two clusters are relatively similar to the main subgroups that were distinguished in the PCNSV cohorts. Cluster 1 mainly includes patients who were previously classified into the angiogram‐diagnosed or l/mv‐PCNSV groups, whereas Cluster 2 mainly corresponds to patients with a biopsy‐proven or sv‐PCNSV [8, 10, 11, 12, 13, 14]. The present results confirm and strengthen the diagnostic yield of CNS biopsy in patients in Cluster 2 and the risk of negative results (> 2/3 of cases) in patients in Cluster 1 [6]. In addition, the 28% of patients in cluster 1 with a positive CNS biopsy presented at baseline with characteristics that were more often observed in Cluster 2, except for acute brain infarct(s). Overall, the present study helps to better define the diagnostic strategy to adopt in patients with suspected PCNSV and provides practical messages to specifically identify patients in whom a CNS biopsy may have a greater yield.

According to the sensitivity analyses, patients in Cluster 2 had a better prognosis at 12 months; however, they relapsed more frequently. This observation is in accordance with previous results. Our group and others previously suggested that patients with sv‐PCNSV have a more relapsing course than patients with l/mv‐PCNSV [8, 12, 19]. In addition, these patients might have a better response to treatment, which may explain why they ultimately have a better prognosis [14]. To our knowledge, there are no pathophysiological explanations to support these observations, and we cannot exclude the possibility that different therapeutic managements potentially influenced the disease trajectory. Notably, in patients without a positive biopsy, especially those with l/mv‐PCNSV, the absence of response to treatment may be suggestive of another diagnosis, especially intracranial atherosclerotic disease. Finally, little is known about whether these two phenotypes depict two different faces of the same disease or two different diseases.

Many important points should be discussed to interpret these results. The variables used to perform the hierarchical cluster analysis were available for all 153 patients. However, some current tools used at baseline, such as digital subtraction angiography or vessel wall imaging, were not included in the model because the procedures were not performed in all patients [20]. Indeed, DSA is not systematically performed in patients with abnormal MRA considering the additional value of the procedure is limited. In opposition, vessel wall imaging is performed in patients with abnormal MRA. Since we aimed initially to include only procedures systematically included in the working diagnosis of PCNSV, we did not consider these two variables, even in sensitivity analyses. Further work is required to implement these tools in supervised analysis and determine whether this would consolidate the distinction between the two clusters. The retrospective and multicenter nature of this study may have resulted in some heterogeneity in the data, especially with respect to imaging. MRI devices and protocols varied between centers, and we did not conduct a centralized blinded analysis of images by a neuroradiologist. In addition, we pooled together large‐ and medium‐sized vessels, although the distinction of both involvements might have allowed identification of patients with overlapping involvement between medium‐ and small‐sized vessels.

Although we hardened our inclusion criteria in the registry, we cannot exclude the possibility that patients without a positive biopsy, especially those in Cluster 1, may have another diagnosis. In these patients, the diagnosis of PCNSV was retained based on the association of vascular abnormalities and findings potentially suggestive of an inflammatory process. We acknowledge that, alone, CSF abnormalities or contrast patterns on MRI might not be specific and sensitive enough to distinguish PCNSV from other mimickers. Of note, with a median 3‐year follow‐up in our patients, no alternative diagnoses were retained.

Tumor‐like lesions were included in the analyses since we aimed to depict the full spectrum of PCNSV presentations. Most patients with this presentation were classified into cluster 2. Obviously, the diagnostic work‐up in these patients is less complicated since the need for a biopsy is well recognized. However, we believe that the inclusion of these patients in the unsupervised analysis is of interest since it confirms that this disease presentation often corresponds to small‐vessel PCNSV.

No information was available regarding whether biopsies were targeted to CNS lesions.

The unsupervised analysis yielded a dichotomous clustering approach. However, some overlaps might exist since we observed that patients in Cluster 1 with a positive CNS biopsy presented overlapping features with Cluster 2.

Finally, a treatment analysis was not conducted, and outcomes were not analyzed while taking into consideration the treatments received. We focused mainly on diagnosis in this work, but further analysis of treatments is planned.

In conclusion, this unsupervised analysis of a large PCNSV cohort confirmed the existence of two distinctive clinical and radiological patient phenotypes. The yields of angiogram and CNS biopsy were opposite in the two observed clusters. This distinction is particularly relevant for helping clinicians obtain a CNS biopsy in patients with a presentation suggestive of Cluster 2, that is, seizure or cognitive impairment associated with gadolinium‐enhanced parenchymal lesions but no intracranial stenosis or acute brain infarct.

Author Contributions

Hubert de Boysson: conceptualization, writing – original draft, investigation, methodology, formal analysis, supervision, writing – review and editing. Ahmad Nehme: conceptualization, investigation, writing – original draft, methodology, formal analysis, writing – review and editing. Anais R. Briant: methodology, formal analysis, writing – review and editing. Sonia Alamowitch: validation, investigation, writing – review and editing. Achille Aouba: investigation, validation, writing – review and editing. Caroline Arquizan: investigation, validation, writing – review and editing. Grégoire Boulouis: investigation, validation, writing – review and editing. Jean Capron: investigation, validation, writing – review and editing. Barbara Casolla: investigation, validation, writing – review and editing. Christian Denier: investigation, validation, writing – review and editing. Nelly Dequatre: investigation, validation, writing – review and editing. Olivier Detante: validation, writing – review and editing, investigation. Laurent Derex: investigation, validation, writing – review and editing. Sophie Godard: investigation, validation, writing – review and editing. Cédric Gollion: investigation, validation, visualization. Benoit Guillon: investigation, validation, visualization. Lisa Humbertjean: investigation, validation, visualization. Clothilde Isabel: investigation, validation, visualization. Philippe Kerschen: investigation, validation, visualization. Laurent Kremer: investigation, validation, visualization. Nicolas Lambert: validation, visualization, investigation. Sylvain Lanthier: validation, visualization, investigation. Adil Maarouf: investigation, validation, visualization. Antoine Néel: investigation, visualization, validation. Thomas Papo: investigation, visualization, validation. Alexandre Y. Poppe: investigation, validation, visualization. Alexis Régent: investigation, validation, visualization. Amina Sellimi: investigation, validation, visualization. Igor Sibon: investigation, validation, visualization. Benjamin Terrier: validation, investigation, visualization. Emmanuel Touzé: validation, visualization, investigation. Stéphane Vannier: validation, visualization, investigation. David Weisenburger‐Lile: investigation, validation, visualization. Mathieu Zuber: investigation, validation, visualization. Jean‐Jacques Parienti: methodology, formal analysis, writing – review and editing. Christian Pagnoux: conceptualization, investigation, writing – original draft, methodology, validation, writing – review and editing, supervision.

Conflicts of Interest

Hubert de Boysson reports receiving fees for serving on advisory boards from Roche‐Chugai and Novartis and lecture fees from Roche‐Chugai, Novartis, Fresenius Kabi, GlaxoSmithKline, Amicus therapeutics, and Sanofi. Christian Pagnoux reports receiving fees for serving on advisory boards from Chemocentryx, Otsuka, GlaxoSmithKline, AstraZeneca, Sanofi, and Hoffman‐La Roche; lecture fees from Hoffman‐La Roche and GlaxoSmithKline; and educational grant support from Hoffman‐La Roche, Otsuka, Pfizer, TEVA, Amgen, and GlaxoSmithKline. The other authors declare no conflicts of interest.

Supporting information

Data S1.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.