Cluster analysis in early axial spondyloarthritis predicts poor outcome in the presence of peripheral articular manifestations

Abstract

Objectives

To assess whether two cluster analysis-based axial SpA (axSpA) endotypes (A for purely axial; B for both axial and peripheral) are stable over time and are associated with different long-term disease outcomes.

Methods

K-means cluster analysis was performed at each visit (until 5 years) on 584 patients from the DESIR cohort, who completed all planned visits, and validated in 232 consecutive axSpA patients from the BeGiant cohort. Cluster stability overtime was assessed by kappa statistics. A generalized linear mixed-effect analysis was applied to compare outcomes between clusters. Classification and regression tree (CART) analysis was performed to determine a decision rule able to assign a given patient to a definite cluster at onset.

Results

Both endotypes remained remarkably stable over time. In the DESIR cohort, patients in cluster B showed higher disease activity, worse functional outcome and higher need for anti-rheumatic drugs than patients in cluster A. CART analysis yielded three main clinical features (arthritis, enthesitis and dactylitis) that accurately determined cluster assignment. These results could be replicated in the Be-GIANT cohort.

Conclusion

Cluster-based axSpA endotypes were reproducible in two different cohorts, stable over time and associated with different long-term outcome. The axSpA endotype with additional peripheral disease manifestations is associated with more severe disease and requires more intensive drug therapy

Clinical trial registration

clinicaltrials.gov, https://clinicaltrials.gov, NCT01648907.

Introduction

Axial spondyloarthritis (axSpA) is a chronic inflammatory disease, with elective involvement of the spine, triggered by genetic and environmental factors [1]. Prevalence of axSpA is estimated between 0.3 and 0.7% [2, 3] and seems to increase over time thanks to a better recognition of the disease. As in many other chronic diseases, the natural course of axSpA is highly heterogeneous. Hence, the degree and duration of inflammation in axSpA vary substantially among patients as well as the distribution of tissue involvement [1]. These factors are likely to impact disease course suggesting that early axSpA can develop into severe progressive disease or stay as rather mild and limited disease over time [4]. Early identification of patients with poor outcome would therefore help to develop a more tailored treatment approach for axSpA and reduce unjustified costs in terms of health assessment and drugs use [5].

Inflammatory back pain (IBP) is the leading and most frequently reported symptom in axSpA [1]. However, peripheral articular manifestations, including arthritis, enthesitis and dactylitis and extra-articular manifestations, including anterior uveitis, psoriasis and IBD seem to critically impact the burden of disease as well [6]. Few studies previously attempted to link the initial clinical presentation of SpA to disease outcome: Amor and colleagues identified seven variables (hip joint involvement, high erythrocyte sedimentation rate, poor efficacy of NSAIDs, limited lumbar spine motion, dactylitis, oligoarthritis and disease onset before age of 16 years) as being associated with poor outcome [7]. Brophy and colleagues showed that extra-articular manifestations such as psoriasis, uveitis and IBD were associated with the severity of ankylosing spondylitis [8]. An additional study suggested that female sex, older age, lower educational level, active smoking status and high inflammatory disease activity at baseline were independently associated with a poor functional outcome at 2 years [9]. Finally, peripheral arthritis and enthesitis have been advocated to adversely influence disease burden [10–12]. While all those studies underlined the value of individual disease manifestations as potential predictors of the progression and disease severity of the whole group of SpA, an integrative approach (such as cluster analysis) to define high-risk patients of poor outcome in axSpA has not been done and an ordered distribution of the manifestations in such patients has not been defined yet.

We recently performed a cross-sectional cluster analysis focussing on baseline characteristics of patients enrolled in a large prospective axSpA cohort DESIR (DEvenir des Spondyloarthrites Indifférenciées Récentes) encompassing subjects with recent onset of IBP highly suggestive of SpA [13]. We identified two clinical clusters within this population: the first, characterized by an isolated axial disorder (A for axial), and the second, characterized by the additional presence of peripheral manifestations (B for both) [14]. Because these two distinct endotypes of axSpA presented at baseline a different disease burden in term of function and disease activity, we hypothesised that they would probably be different in disease progression and outcome. Therefore, we aimed to assess whether these two endotypes were stable over time and might predict different disease outcomes along 5 years of follow-up. We also questioned their generalizability and applicability to a different patient group by studying SpA patients with axial manifestations from the BelGian Inflammatory Arthritis and spoNdylitis cohorT (Be-Giant) cohort.

Patients and methods

Patients

DESIR cohort is a French longitudinal prospective multicentre cohort of early axSpA. Between October 2007 and April 2010, consecutive patients aged from 18 to 49 years were included if they had IBP for at least 3 months but <3 years duration, and symptoms suggestive of SpA according to the opinion of the rheumatologist investigator (>5/10 on a visual analogue scale). Patients should be biologic-naïve at inclusion. Patients underwent a standardized clinical and biological evaluation every 6 months during the first two years and then on a yearly basis up to 5 years. Full description of the 708 patients at inclusion is provided elsewhere [13]. We only included in the present study the patients who completed all the seven scheduled follow-up visits until 5 years (n = 584). We used the dataset as locked on 20 June 2016.

Be-Giant is a multicentre, observational cohort that includes 367 consecutive patients, newly diagnosed with SpA by their treating rheumatologist. Biologic-naïve patients that fulfil the ASAS classification criteria of axial or peripheral SpA are prospectively followed in the cohort every 6 months with a planned total duration of 20 years [15, 16]. In this study, we only included the patients with past or current IBP at inclusion (n = 232, with 84 of them completing the follow-up visits until 2 years). We used the dataset as locked on 1 March 2020.

Collected data

Collected information in DESIR comprised demographic, clinical, biological and imaging data. At each visit, clinical manifestations of the disease (articular and extra-articular) and treatments were collected. Disease activity was evaluated using the BASDAI [17] and the Ankylosing Spondylitis Disease Activity Score using CRP (ASDAS-CRP) [18]. Functional outcomes were assessed with BASFI [19] and HAQ Disability Index for the Spondyloarthropathies (HAQ-S) [20]. Quality of life was assessed by patient-reported outcomes: Bath Ankylosing Spondylitis Global Score (BASG) [21], 36-Item Short Form Survey (SF36, with a higher score defining a more favourable health state) [22] and Ankylosing Spondylitis Quality of Life Questionnaire (AsQOL) [23]. Radiographs and MRI of the whole spine and pelvis were performed at baseline, 2 years and 5 years of follow-up in DESIR (follow-up MRIs were only performed in a subset of centres including 203 patients at 2 years and 153 at 5 years).

The same clinical and imaging information were collected at each visit in Be-Giant except for BASG, SF36 and AsQOL. Conventional radiographs of the pelvis and the whole spine were performed at baseline and repeated every two and four years, respectively. MRI of sacroiliac joints was only performed at baseline with follow-up imaging upon clinical indication.

Ethics statement

Both studies were conducted according to the current regulations of the International Conference on Harmonization guidelines and the principles of the Declaration of Helsinki. DESIR cohort (clinicaltrials.gov ID: NCT01648907) obtained the approval of the appropriate local medical ethics committee (CPP Ile-de-France, P070302). Be-Giant cohort was authorized by the local medical ethics committee of the University Hospital of Ghent and every participating centre (EC/2012/497). Written informed consent was obtained from all patients before inclusion.

Statistical analysis

The statistical analysis was primarily performed in DESIR cohort, in which the clusters were initially described [14], then replicated in the Be-Giant cohort by including only patients with axial manifestations to independently validate the results obtained in DESIR. Differences between patients with or without missing follow-up visits were assessed using Student’s t test or Wilcoxon test for continuous variables and χ² test or Fisher’s exact test for categorical variables.

In both cohorts, cluster analyses were performed as previously described [14]. First, we performed a multiple correspondence analysis (MCA) from the package FactoMineR [24] in the R version 3.5.1 (http://www.r-project.org/) to obtain standard coordinates of the principal components of the individual observations. Then we applied an iterative partitioning K-means method on these standard coordinates, using the Hartigan–Wong algorithm based on Euclidean distances [25]. The following variables were used as variables of construction: sex, age at disease onset (<40 years or ≥40 years), duration of disease symptoms (<2 years or ≥2 years), HLA-B27 positivity, buttock pain, peripheral arthralgia or arthritis, enthesitis, dactylitis, uveitis, psoriasis, IBD, triggering infection and response to NSAIDs. These data were collected at each time point in a cumulative manner (once positive, always positive). Optimal number of clusters was determined and confirmed at baseline using a combination of several statistical criteria (further information is available upon request from the corresponding author). The stability of our solution was tested with a resampling method [26].

Student’s t test or Wilcoxon test for continuous variables and χ² test or Fisher’s exact test for categorical variables were used to identify differences between clusters at inclusion. All tests were 2-sided and P-values <0.05 after Holm–Bonferroni correction were considered significant.

To assess cluster stability at each follow-up timepoint, we used Light’s kappa to estimate the global agreement of cluster allocation over the eight repeated classifications, and to evaluate patients changing cluster allocation during follow-up, we estimated differences in disease activity variables and in drugs initiation before and after the switch, by performing permutation tests that estimate the same parameters in random samples of patients who never switched. P-values were calculated as the fraction of permutation values that are at least as extreme as the observed parameter.

To compare outcomes between the two clusters while taking into account the repeated measures within each patient, we performed generalized linear mixed-effect regressions using identity as link junction for quantitative responses and logit link function for binary responses, both using functions of the package lme4 [27] in the R version 3.5.1, and using only incident, non-cumulative, variables. Clusters at baseline were considered as fixed effects and subjects as random effect. P-values were obtained from likelihood ratio tests (LRT) comparing the full model including cluster with the null model without cluster. Coefficient estimates of the variable ‘cluster’ were also given together with 95% CIs with cluster A as reference. P-values <0.05 after Holm–Bonferroni correction for the number of LRT performed were considered significant.

To identify the best predictors of cluster allocation of each patient at baseline into a given cluster, we first evaluated the accuracy (e.g. sensitivity, specificity, positive and negative predictive value and positive likelihood ratio) of each variable of construction separately; then, we tested combinations of variables by using a recursive partition tree (CART) analysis within the package caret in R [28]. A 10-fold cross validation was used for estimating misclassification rates. A positive LR of the resulted prediction was calculated, with a LR ≥3 being considered as suitable. The same variables of cluster construction were used in both cohorts, except for alternate buttock pain and arthritis only, instead of buttock pain and arthralgia/arthritis, in Be-GIANT and DESIR, respectively.

Results

Longitudinal cluster determination in ‘DESIR completers’

Because we used a subset of the DESIR cohort (i.e., patients who completed all follow up visits), we first compared the 584 completers to the 95 patients with missing follow-up visits. Patterns of missing follow-up visits are summarized in Supplementary Fig. S1, available at Rheumatology online. Baseline clinical, laboratory and imaging characteristics were similar in both groups (Supplementary Table S1, available at Rheumatology online) except for a greater proportion of good responders to NSAIDs treatment among completers (92.9% vs 64.5%, P = 0.003).

Cluster analysis performed using the baseline characteristics of the 584 completers resulted in two clusters, with 288 patients assigned to cluster A and 296 to cluster B (Supplementary Fig. S2A and B, available at Rheumatology online). Cluster characteristics were very similar to the previously published: cluster A was characterized by more frequent axial manifestations, while cluster B was mainly characterized by a higher frequency of arthritis, enthesitis and dactylitis and of psoriasis (Table 1, Fig. 1). Moreover, patients from cluster B presented higher disease activity scores (as measured by BASDAI and ASDAS-CRP), higher functional impairment (as measured by BASFI and HAQ-AS) and lower quality of life (as measured by SF-36, ASQoL and BASG).

To address cluster stability over time, we repeated cluster analyses at each follow-up visit in the DESIR cohort. Most of the patients stayed in the same cluster during the entire follow-up (Supplementary Fig. S3, available at Rheumatology online) with consistent differences in the distribution of clinical parameters between the two clusters over time (Supplementary Table S2, available at Rheumatology online). Global agreement of cluster allocation over the eight repeated classifications was high (Light’s kappa of 0.80; 95% CI: 0.77, 0.84). Few patients changed cluster allocation during follow up: 42 out of 288 (14.5%) moved from cluster A to B, while 27 out of 236 (11.4%) patients moved from cluster B to A. Transitions from cluster A to cluster B were mostly due to new onset of enthesitis (n = 31), peripheral arthralgia/arthritis (n = 18) or dactylitis (n = 12). Transitions from cluster B to cluster A were not associated with any modification of variables of construction and were therefore considered as cluster fluctuations.

We then looked at the characteristics of the 42 patients who moved from cluster A to cluster B and estimated the differences in BASDAI and ASDAS-CRP scores before and after the switch. We found that patients who never switched from cluster A had a higher improvement of BASDAI overtime as compared with patients who moved from cluster A to B (P-value <0.001; Supplementary Fig. S4A, available at Rheumatology online), even though there was no significant difference in the ASDAS-CRP evolution between the two groups (P = 0.1; Supplementary Fig. S4B, available at Rheumatology online). Thirty-three of the 42 patients were not treated by bDMARDs at the time they moved from cluster A to B. Eight of them (24.2%) initiated bDMARDs after the switch, which was significantly higher than patients who never switched (P-value = 0.04; Supplementary Fig. S4C, available at Rheumatology online).

Five years outcome prediction of baseline clusters in DESIR

Regression analyses showed that both clusters maintained their baseline clinical and imaging differences during follow-up (Table 2): i.e., higher frequency of peripheral involvement (arthralgia, arthritis and enthesitis) in cluster B (Supplementary Fig. S5A–C, available at Rheumatology online), and higher prevalence of isolated axial symptoms and imaging abnormalities of sacroiliac joints in cluster A (Supplementary Fig. S5D–F, available at Rheumatology online). Over time, worse disease outcome was observed in cluster B than in cluster A, including higher disease activity as measured by BASDAI and ASDAS-CRP, higher functional impairment as measured by BASFI and HAQ, and a lower quality of life as measured by ASQoL and by both physical and mental components of SF-36. Furthermore, axSpA patients in cluster B received more frequently conventional DMARDs or anti-TNF therapy (Supplementary Fig. S6, available at Rheumatology online). Table 2 shows the comparison of disease parameters between both clusters during the entire follow-up.

Prediction of cluster assignment

Despite the high specification, none of the considered variables, taken individually, was able to discriminate at baseline future or progressive cluster allocation (A or B), even though peripheral manifestations, and in particular dactylitis, presented the highest positive LR due to its nearly perfect specificity to predict allocation to cluster B (Supplementary Table S3, available at Rheumatology online). CART confirmed that at baseline, cluster allocation was determined by the presence of peripheral arthralgia and/or arthritis, followed by, or associated to, enthesitis, or dactylitis (Fig. 2A). Given this tree, the accuracy to classify patients into cluster B was very good (90%) with a sensitivity of 83%, specificity of 96% and positive LR of 20.0 (95% CI: 11.5, 34.9). The same partitioning method gave similar results when applied to follow-up data (5 years) (Fig. 2B).

Clustering replication in the Be-GIANT cohort

Baseline cluster analysis in the Be-Giant cohort also delineated two clusters (A and B), with differences similar to those obtained in DESIR: cluster A (201 patients) was characterized by higher prevalence of isolated axial manifestations and cluster B (31 patients) by higher frequency of peripheral manifestations and psoriasis, higher disease activity scores, higher functional impairment and a lower quality of life (Table 3 and Supplementary Fig. S2C and D, available at Rheumatology online).

Eighty-four patients completed all the follow-up visits up to 2 years (77 in cluster A and 7 in cluster B). Regression analyses showed the same trends in disease outcomes as in DESIR cohort with a higher disease activity, a higher functional impairment and a lower quality of life in cluster B as compared with cluster A. Use of cDMARDs and bDMARDs was also more frequent in patients in cluster B (Supplementary Table S4 and Fig. S7, available at Rheumatology online). CART built from DESIR baseline data was finally applied to Be-Giant data at baseline and after 24 months of follow-up. Performances of the tree to classify patients into cluster B at baseline were very good with a positive LR of 11.7 (95% CI: 6.0, 22.9) at baseline and of 3.9 (95% CI: 3.2, 4.9) at 24 months (Fig. 2C and D).

Discussion

AxSpA is a complex disease, presentation of which is built on an orchestra of individual disease domains. Single disease characteristics over time were previously used to define axSpA trajectories [29–31]. However, integrating several disease domains rather than considering them separately may help to better characterize patients endotypes. Hence, cluster analysis, previously performed in cross-sectional manner in two independent cohorts, including the present DESIR cohort, demonstrated to be very useful in defining two clinically relevant subgroups of axSpA: one characterized by an isolated axial disorder, the other by the additional presence of peripheral manifestations [14, 32].

Here, longitudinal analysis over 5 years of those two clusters allowed us to assess their stability and whether they differ in key disease outcomes. We showed that both clusters were stable over time, maintaining the initial clinical pattern distribution (axial presentation for cluster A, axial and peripheral presentation for cluster B), and mainly characterized by different long-term outcomes. Cluster B showed a worse disease outcome with higher disease activity and functional impairment over time as well as lower health-related quality of life, despite a higher use of conventional and biological anti-rheumatic treatments. Moreover, an optimal partition in two baseline clusters with the aforementioned differences in the clinical pattern of disease was also observed in the Be-GIANT cohort. Worse disease outcome was found in the cluster B as compared with cluster A in the Be-GIANT cohort, reinforcing the validity of the results obtained in DESIR.

As the disease course of axial SpA is variable and influences treatment strategies over time, it is important to define predictors of severe disease to reduce costs. Severity may include multiple facets such as pain, inflammatory activity, radiographic structural damage, reduced physical function or social impairment. Here we show that peripheral manifestations rather than axial ones are predictive of a poor health-related quality of life and of a sustained high disease activity. These findings are consistent with the lack of correlation observed between the modified Stoke Ankylosing Spondylitis Spinal Score (mSASSS) and quality of life found in two recent studies [33, 34]. This accumulation of evidence challenges the role of axial imaging changes being relevant for disease severity. Indeed, in our analysis, mSASSS was not different between the two clusters.

A link between peripheral manifestations and poor disease outcome has been suggested in other studies, but none showed that it can be predicted at baseline [10]. Traditionally, differences in the clinical presentation of SpA were based on retrospective description of single disease domains. We, instead, integrated all disease domains and then prospectively analysed the outcome of these disease endotypes. While cluster analysis is helpful to identify different clinical patterns, which may end in different disease outcome, it is not tailored to estimate cluster affiliation in daily practice. Therefore, we elaborated a simple algorithm to accurately adjudicate a patient in one of the two clusters, based on only three clinical variables: peripheral arthralgia/arthritis, enthesitis and dactylitis. The generalizability of our clustering and the applicability of our decision tree algorithm has now been demonstrated in two independent cohorts, DESIR and Be-GIANT, suggesting that this algorithm may be useful in all patients with axSpA.

In summary, we show that axSpA is characterized by stable clinical endotypes, which differ in their disease course and outcome, and that can be easily identified by a simple algorithm. Based on these data it will be possible to define a subset of axSpA that are prone to have poor disease outcome. A tighter monitoring of disease activity may be proposed to these patients with the possibility of early therapeutic intervention. However, additional studies are needed to confirm the interest of this cluster-based therapeutic approach.

Acknowledgements

The DESIR cohort was sponsored by the Département de la Recherche Clinique et du Développement de l'Assistance Publique–Hôpitaux de Paris. This study is conducted under the umbrella of the French Society of Rheumatology and INSERM (Institut National de la Santé et de la Recherche Médicale). The database management is performed within the department of epidemiology and biostatistics (Professor Paul Landais, D.I.M., Nîmes, France). An unrestricted grant from Pfizer was allocated for the 10 years of the follow-up of the recruited patients. The authors thank the different regional participating centres: Prof. Maxime Dougados (Paris– Cochin B), Prof. André Kahan (Paris—Cochin A), Prof. Philippe Dieudé (Paris—Bichat), Prof. Pierre Bourgeois (Paris—La Pitié Salpetrière), Prof. Francis Berenbaum (Paris—Saint Antoine), Prof. Pascal Claudepierre (Créteil), Prof. Maxime Breban (Boulogne Billancourt), Dr Bernadette Saint-Marcoux (Aulnay-sous-Bois), Prof. Philippe Goupille (Tours), Prof. Jean-Francis Maillefert (Dijon), Dr Xavier Puéchal, Dr Emmanuel Dernis (Le Mans), Prof. Daniel Wendling (Besançon), Prof. Bernard Combe (Montpellier), Prof. Liana Euller-Ziegler (Nice), Prof. Philippe Orcel, Dr Pascal Richette (Paris—Lariboisière), Prof. Pierre Lafforgue (Marseille), Dr Patrick Boumier (Amiens), Prof. Jean-Michel Ristori, Prof. Martin Soubrier (Clermont-Ferrand), Dr Nadia Mehsen (Bordeaux), Prof. Damien Loeuille (Nancy), Prof. René-Marc Flipo (Lille), Prof. Alain Saraux (Brest), Prof. Corinne Miceli (Le Kremlin Bicêtre), Prof. Alain Cantagrel (Toulouse), Prof. Olivier Vittecoq (Rouen). The authors also thank URC-CIC Paris Centre for the coordination and monitoring of the study.

Funding: The sponsors (Abbvie, AP-HP, Pfizer) had no role in the study design or in the collection, analysis, or interpretation of the data, the writing of the manuscript, or the decision to submit the manuscript for publication. Publication of this article was not contingent upon approval by Abbvie, AP-HP or Pfizer.

Disclosure statement: F.C.: Speakers fees from Lilly. Consultancy fees from Novartis, UCB. F.VdB.: Grants from AbbVie; personal fees from AbbVie, Eli Lilly, Galapagos, Janssen, Novartis. D.E.: Grants from AbbVie; personal fees from AbbVie, Eli Lilly, Galapagos, Janssen, Novartis, Pfizer, UCB. M.-A.D’A.: Speakers fees from Abbvie, BMS, Novartis, Celgene. Consultancy fees from Novartis, Janssen, Abbvie. Financial support for Academic Sponsored International Studies from Pfizer. The other authors have declared no conflicts of interest.

Data availability statement

All the material collected in the DESIR cohort are available for any researcher. The overall description of the cohort and the procedures of the application of a research project can be obtained at http://www.lacohortedesir.fr/desir-in-english/. For any further information, please contact us at the following [email protected].

Supplementary data

Supplementary data are available at Rheumatology online.

References