Determination of four homogeneous subgroups of patients with antiphospholipid syndrome: a cluster analysis based on 509 cases

Abstract

Introduction

APS is defined by the occurrence of arterial and/or venous thromboembolism (VTE) and/or pregnancy morbidity, associated with persistent aPL, including lupus anticoagulant (LA), anticardiolipin antibodies (aCL) and/or anti-β2 glycoprotein 1 (β2GP1) antibodies [1]. APS can either occur as a primary disorder (primary APS) or be associated with a CTD, most frequently SLE. Several clinical phenotypes have traditionally been distinguished in APS, based on the classification criteria: obstetric, defined as isolated obstetric manifestations; venous APS as VTE (± obstetric manifestations); and arterial APS as arterial thrombosis (± obstetric and/or venous manifestations). In addition, APS severity varies widely from VTE or obstetric manifestations to catastrophic antiphospholipid syndrome (CAPS), a life-threatening condition resulting from rapidly progressive widespread thromboses mainly affecting the microvasculature in the presence of aPL [2]. Finally, APS is now considered a multisystem disease as other ‘non-criteria’ manifestations of APS may occur, including cardiac, neurological, cutaneous, renal or haematological features [3].

Thus, it is not clear if the traditional classification criteria of APS—venous, arterial and obstetric—accurately represent all phenotypic profiles of APS and consider the potential associations between the broad spectrum of its clinical and laboratory manifestations. Identifying different subgroups that include different manifestations and comorbidities might help us to better understand the pathophysiology and thus the treatment of APS.

Hierarchical cluster analysis is an unsupervised statistical method of data partitioning, whereby individuals are clustered, that is, categorized, into homogeneous groups on the basis of similarity. This method has been previously used for autoimmune diseases [4–7] including APS [8–10] to identify phenotypic groups. Thus, we aimed to use this method to determine distinct homogeneous phenotypes among APS patients.

Patients and methods

Patients

This study used data from the French multicentre ‘APS and SLE’ registry, a database including APS and/or SLE patients from different centres in France (clinicaltrials.gov NCT02782039, with Institutional Review Board approval, previously described by Morel et al. [11]). Two centres (Cochin Hospital in Paris and Lille University Hospital) are the main contributors, but other centres (n = 8) have contributed by including their patients with CAPS.

This study included patients aged 18 or older, meeting the APS criteria of the Sydney classification, and with sufficient data about their clinical and laboratory manifestations [1]. Patients with aPL on laboratory tests but with no clinical criteria of APS were not included.

Clinical variables

The authors (Y.N., V.L.G., N.C.C.) selected 27 characteristics of clinical and laboratory manifestations to cover the broad APS spectrum, based on their clinical relevance and their availability in the registry (Supplementary Table S1, available at Rheumatology online). Those manifestations could occur at APS diagnosis or during follow-up. We included all manifestations included in the Sydney classification criteria with their definitions: vascular thrombosis (arterial thrombosis, deep VTE, or biopsy-proven small vessel thrombosis); ≥3 consecutive fetal losses before 10 weeks of gestation, unexplained death of a fetus at or after 10 weeks of gestation; premature birth before 34 weeks of gestation because of eclampsia or placenta insufficiency; one or more unexplained deaths of a morphologically normal fetus at or beyond the 10th week of gestation; and LA, aCL antibodies of the IgG and/or IgM isotypes, anti-β2GP1 antibodies of the IgG and/or IgM isotypes (present on two or more occasions, at least 12 weeks apart). Additionally, we included sex, CAPS diagnosis, other ‘non-criteria’ manifestations of APS (livedo reticularis, seizures, migraine, chorea, cardiac valvulopathy, aPL-associated nephropathy, haemolytic anaemia, and thrombocytopenia), SLE diagnosis, other autoimmune diseases, cardiovascular comorbidities (diabetes mellitus or dyslipidaemia), and other laboratory features (such as antinuclear antibodies [ANA] and low complement C3 levels). Lymphopenia was also included as it has been described in primary APS populations [12].

Finally, the treatments administered (low-dose aspirin, anticoagulation by therapeutic doses of heparin and/or vitamin K antagonists [VKA], hydroxychloroquine, and immunosuppressive therapies, including azathioprine, rituximab, mycophenolate mofetil and cyclophosphamide) and death during follow-up were also added in the descriptive analyses (i.e. not included in the clustering process, as the choice of therapy could be investigator-dependent).

CAPS was defined according to the criteria of Asherson et al., revised in 2010 [2, 13, 14], and SLE by the SLICC criteria [15]. Thrombocytopenia was defined as platelet counts equal to or <150 × 10⁹ platelets/l, and lymphopenia below 1.4 × 10⁹ cells/l.

Statistical analysis

We used the 27 selected variables to perform multiple correspondence analysis (MCA) and then considered the coordinates of the observations of the factorial axes retained as new variables for the cluster analysis.

The first k-axes, which explained at least 90% of the total variability, were considered. We then performed hierarchical clustering based on the Ward method, followed by consolidation (k-means algorithms) to build homogeneous clusters of APS patients. This technique starts with every case considered to be its own cluster and successive two-by-two mergers of clusters until the final merger with all subjects falling into a single category. The metric used to assess the proximity between two classes is the Euclidian distance, the most commonly used measure of (dis)similarity. The clustering process can then be plotted as a dendrogram, with horizontal branches representing the combination of two clusters and vertical branches the degree of dissimilarity between combined clusters; long distances of the vertical segments indicate large differences between the clusters that have been formed.

Two distinct approaches were used to estimate the optimal number of clusters within the population. First, we used a visual distance criterion by cutting the dendrogram horizontally at the level of highest dissimilarity. Then, we checked the gain in within-cluster inertia achieved at each clustering. We then tested the robustness of the clusters itself by determining the optimal number of clusters in our data set with the NbClust function [16]. Finally, the clusters resulting from the grouping process were described and named by their most prominent summary characteristics.

Categorical variables are presented as counts and percentages, and continuous variables as means (standard deviation [SD]) or medians (interquartile range [IQR]). Different subgroups were compared with logistic regression and analysis of variance. Obstetric events were compared only among women. A two-tailed P ≤ 0.05 was considered statistically significant.

Missing values of variables included in the cluster analysis were handled by using multiple imputations with the missMDA package in R [17]. missMDA imputes missing values in such a way that the imputed values have no weight (i.e. have no effect and the method is performed with only the observed values) on the MCA results. We assumed that the missing data were missing at random. Variables with the highest rate of missing data were low C3 levels (44%), ANA positivity (15%) and low lymphocyte counts (6%).

Two sensitivity analyses were conducted. First, we excluded patients with SLE and/or auto-immune diseases, to conduct a cluster analysis on only primary APS patients. Second, we performed an analysis that excluded patients with a high rate of missing values (>3 missing clinical values) and did not incorporate ‘low C3’ and ‘ANA’ in the clustering process.

All analyses were performed with R version 4.3.1 (R Foundation for statistical computing, Vienna, Austria) and the FactoMineR library for cluster analysis [18].

Results

Study population

In July 2019, among the 1239 patients included in the SLE and APS registry, 604 were classified with APS according to the Sydney criteria. No data were available for the laboratory characteristics of 95, who were excluded from the analyses. Thus, this study included 509 patients, mainly women (77.8%). Most received care at two tertiary centres: Cochin Hospital in Paris (n = 220; 43.2%) and Lille University Hospital (n = 223; 43.8%).

The patients' main clinical characteristics at APS diagnosis or during following-up are presented in Table 1. Mean (s.d.) age at APS diagnosis was 36.2 (14.6) years, and mean length of follow-up since this diagnosis was 10.3 (8.5) years. In all, 55.4% of patients had VTE, 41.7% arterial thrombosis, and 124 of the 396 women (31.3%) had pregnancy morbidity. CAPS occurred in 90 patients (17.7%), and APS was associated with SLE in 123 (24.2%) cases and with other autoimmune diseases (including Sjögren’s syndrome, systemic sclerosis and rheumatoid arthritis) in 55 more (10.8%).

MCA and selection of classification models

MCA retained 11 axes, which explained 95% of the total variability. Hierarchical clustering analysis suggested a four-class solution both visually and with the gain in inertia (Figs 1 and 2). In addition, most indexes tested in the NbClust function suggested an optimal number of clusters of four (Supplementary Data S1, available at Rheumatology online).

Figure 1.

Dendrogram for cluster model for antiphospholipid syndrome. Horizontal branches represent the combination of two clusters, and vertical branches the degree of dissimilarity between combined clusters; the grey zone represents the level of truncation, thus resulting in four groups

Open in new tabDownload slide

Figure 2.

Factor map showing the raw individual data used to generate the dendrogram. The first two dimensions cumulatively explained 20.7% of the variance. We obtained a hierarchical tree positioned on the factorial map. The colours indicate individuals according to the cluster to which they belong. There were some overlaps between the four clusters. APL: antiphospholipid; CAPS: catastrophic antiphospholipid syndrome; LA: lupus anticoagulant; VTE: venous thromboembolism

Open in new tabDownload slide

Cluster description

Cluster 1 (n = 181) included mostly patients with VTE (78%) and premature births due to placental insufficiency (14.1%). Only 2.2% had an associated autoimmune disease.

Cluster 2 (n = 130) included older patients (mean age 45.8 years), the lowest proportion of women (64.6%) and patients with a history of arterial events (88.5%). Cardiac valvulopathy (21.1%), migraines (20.8%), livedo (35.2%), arterial hypertension (48.8%) and other cardiovascular risk factors (diabetes mellitus and/or dyslipidaemia) were relatively frequent, while obstetric manifestations were relatively rare. Patients were most often treated with anticoagulation (84.3%), while hydroxychloroquine treatment was fairly infrequent.

Cluster 3 (n = 102) included younger patients (mean age 30.7 years), the highest proportion of women (92.2%), and those with associated SLE (76.5%) or other autoimmune diseases (30.4%). They frequently had a history of VTE (73.5%) and of pregnancy morbidity (36.2%). Lymphopenia, thrombocytopenia (48.0%), haemolytic anaemia (13.5%), ANA, low C3, and LA (84.3%) were relatively frequent. This group had the highest percentages of patients treated with hydroxychloroquine (79.6%) and with immunosuppressive therapies (25.3%).

Cluster 4 (n = 96) included mainly patients with a history of CAPS (77.1%), of arterial thrombosis (68.8%), and/or APL-associated nephropathy (92.6%), and pregnancy morbidity (36.2%), with more premature births because of eclampsia or placental insufficiency and more unexplained fetal deaths at or beyond the 10th week of gestation. Livedo (35.9%), seizures (14.3%), chorea (4.4%), cardiac valvulopathy (43.6%), thrombocytopenia, LA (87.5%) and triple positivity (65.6%) were relatively frequent, as was associated SLE, although less so than in cluster 3. Most patients were treated with anticoagulation (therapeutic heparin and/or VKA) (95.3%). Other treatments included low-dose aspirin (48.8%), hydroxychloroquine (52.3%) and immunosuppressive therapies (16.5%). The proportion of deaths was highest in this cluster (16.7%).

The four clusters were thus named ‘venous’, ‘arterial’, ‘associated with autoimmune disease’, and ‘arterial microthrombotic’. Their clinical and laboratory characteristics are described in Table 1. Obstetric manifestations were found in all four clusters, mostly in clusters 3 and 4 (36.2% in each). Non-criteria manifestations of APS, including skin, neurological and cardiovascular manifestations, were mainly found in clusters 2 (arterial) and 4 (arterial microthrombotic). Among the laboratory features, aCL and anti-β2-glycoprotein I antibodies were most likely to be found in cluster 4 (arterial microthrombotic), as was triple positivity. LA positivity was most frequent in clusters 3 (autoimmune) and 4 (arterial microthrombotic). Most patients received anticoagulation (therapeutic heparin and/or VKA), especially in cluster 4, and/or low-dose aspirin, but hydroxychloroquine and/or immunosuppressive therapies was prescribed most often in cluster 3 (autoimmune), but also in cluster 4 (arterial microthrombotic).

During follow-up, 23 (4.5%) patients died, mainly in cluster 4 (16 patients, 16.7%). Only two patients died ≤1 year after their APS diagnosis, and both were in cluster 4. None of the deaths in clusters 1, 2 or 3 was related to APS. In contrast, of the 16 deaths in the CAPS group, seven (43.8%) were directly attributable to CAPS or thrombotic events.

Sensitivity analyses

The first sensitivity analysis, excluding patients with SLE and/or other autoimmune diseases yielded three clusters, corresponding to clusters 1, 2 and 4, and did not modify our previous findings, as 95.5% of the patients were classified in the same cluster as in the main analysis (Supplementary Table S2, available at Rheumatology online).

The second sensitivity analysis, in which patients with a high rate of missing values (>3 clinical missing values) were excluded and in which ‘low C3’ and ‘ANA’ were not incorporated in the clustering process, was performed on 486 patients (Supplementary Fig. S1, available at Rheumatology online). It also yielded four clusters, with 95.7% of the patients classified in the same cluster as in the main analysis (Supplementary Table S3, available at Rheumatology online).

Discussion

Using an unsupervised clustering method for this study, we distinguished four homogeneous subgroups of APS patients: predominantly venous; arterial; associated with SLE or other autoimmune disease; and arterial microthrombotic. Compared with traditional phenotypes based on APS classification criteria, we did not find an individualized obstetric phenotype of patients. Instead, obstetric comorbidities were relatively frequent in all clusters, except cluster 2.

Cluster 3 included most of the patients with APS associated with SLE, but SLE was also present in cluster 4. Interestingly, our group previously demonstrated that a significant proportion of patients with primary APS may meet SLICC classification criteria for SLE even though they have no SLE because in addition to having antiphospholipid antibodies, other manifestations/criteria related to APS may be mistakenly attributed to SLE (e.g. thrombocytopenia, haemolytic anaemia that can be due to microangiopathy, and seizures) [12]. Since we are particularly aware of this, we believe that all patients classified as SLE indeed have SLE. In particular, the haemolytic anaemia present in both clusters corresponded mainly to autoimmune anaemia or mechanical anaemia with the presence of schistocytes during thrombotic microangiopathy.

The most frequent systemic non-criteria manifestations of APS, namely skin, cardiac and neurological manifestations, were mainly included in the arterial and the arterial microthrombotic clusters (2 and 4). This association has already been found by our group in a cohort study of 200 patients with APS, in which livedo reticularis was significantly associated with cerebral or ocular ischaemic arterial events, all arterial events, hypertension and heart valve abnormalities [19].

We also found an arterial microthrombotic cluster, which includes mostly patients with CAPS, of varying clinical manifestations, severity and outcomes, and increased mortality. SLE was not rare in this cluster.

Four previous studies have attempted to identify different APS phenotypes by using unsupervised clustering methods (Supplementary Table S4, available at Rheumatology online). In a series of 246 patients with APS, Krause et al. [8] sought to define clusters of disease manifestations by factor analysis with 14 variables. Their analyses yielded five patterns: (i) association between cardiac valve abnormalities, livedo reticularis, autoimmune haemolytic anaemia and neurological manifestations; (ii) arthritis, thrombocytopenia, and leucopenia; (iii) recurrent fetal loss and IUGR; (iv) inverse correlation between arterial and VTE; and (v) epilepsy and migraine. Their clusters 1 and 5 are close to our cluster 2, which includes patients with many non-criteria neurological, skin and cardiovascular manifestations, while their cluster 2 could represent patients with SLE (a variable that they did not specifically study).

In an Italian multicentre cohort including 486 patients with aPL (not all with APS), Sciascia et al. also attempted to identify phenotypes of patients by a cluster analysis [9]. It too produced five clusters: (i) VTE and high frequency of triple aPL positivity (34.7%) (matching our cluster 1); (ii) SLE and high rates of anti-dsDNA positivity (matching our cluster 3); (iii) pregnancy morbidity; (iv) a bridging condition between pure primary APS and defined SLE with lower thrombotic risk but higher general features such as ANA and cytopenia; and (v) asymptomatic aPL carriers. Their results might not be comparable to ours, as their selected populations included a high rate of asymptomatic aPL carriers (27.4%), distributed among different clusters (mainly clusters 4 and 5).

More recently, using the international APS Alliance for Clinical Trials and International Networking (APS ACTION) Registry, Zuily et al. used a similar method of hierarchical cluster analysis with 497 patients and 30 clinical data points and found three different clusters [10]: (i) female patients with no other autoimmune diseases but with VTE and triple-aPL positivity; (ii) female patients with SLE, VTE, aPL-associated nephropathy, thrombocytopenia, haemolytic anaemia and positive LA; and (iii) older men with arterial thrombosis, heart valve disease, livedo, skin ulcers, neurological manifestations and cardiovascular disease risk factors. We found several similarities with our cluster analysis, and their classification could match our ‘venous’, ‘autoimmune’, and ‘cardiovascular’ clusters. Like us, they did not find an ‘obstetric cluster’: patients with obstetric morbidity were evenly distributed among the three groups. As in [9], their population might differ from ours, as they also included asymptomatic aPL carriers (10.4%), and their prevalence of CAPS patients was probably low (not reported).

Finally, in a Japanese series of 168 patients, Ogata et al. found three clusters, very similar to those found by Zuily et al.: (i) APS associated with other autoimmune diseases; (ii) cardiovascular risk and arterial thrombosis; and (iii) VTE and triple positivity [20]. Although they did not study systemic non-criteria manifestations of APS, such as livedo or valve involvement, they analysed outcomes according to the clusters and found cluster 2 (arterial) was associated with a higher rate of thrombosis, severe bleeding or death. Our study, which is the largest thus far, confirmed these previous findings and demonstrated for the first time the existence of an arterial microthrombotic cluster.

We acknowledge some limitations to our study. First, our results have not been validated in an independent validation cohort by the same methodology. Second, patients received care mainly in two tertiary centres, and because the register favours the inclusion of patients with CAPS, the proportion of such patients is high (17%). The relative proportions of patients in each cluster may thus not be representative of real life. This is nonetheless also a strength of our study since it allowed us to identify a specific cluster of patients with arterial microthrombotic APS. In addition, as patients’ characteristics included in the cluster analysis were collected throughout the follow-up, we could not compare different outcomes (such as a new arterial event or a new VTE) between the clusters. Other strengths of our study were the large number of patients and the relative low quantity of missing data, handled by multiple imputations. In addition, sensitivity analysis excluding missing data yielded the same results.

To conclude, our study distinguished four distinct homogeneous subgroups of APS patients: venous; arterial; associated with SLE or other autoimmune disease; and arterial microthrombotic. Heterogeneous pathophysiological mechanisms may explain these findings.

Supplementary data

Supplementary data are available at Rheumatology online.

Data availability

Data are available upon reasonable request.

Contribution statement

All authors contributed to the manuscript. Y.N., N.M., V.L.G. and N.C.C. were responsible for conception and design. Y.N., C.Y., N.M., V.L.G. and N.C.C. were responsible for data collection and analysis. All authors were responsible for the interpretation of data. Y.N. wrote the first version of the manuscript. All authors critically revised and approved the final version of the manuscript.

Funding

No specific funding was received from any bodies in the public, commercial or not-for-profit sectors to carry out the work described in this article.

Disclosure statement: The authors have declared no conflicts of interest.

Acknowledgements

The authors are indebted to all patients for their participation, and to all physicians who included patients in the APS and SLE registry. The authors would like to thank Ada Clarke for her help in data collection.

References