To examine whether spondylarthropathy (SpA) disease manifestations would combine in any ordered pattern among patients from SpA multiplex families.
To examine whether spondylarthropathy (SpA) disease manifestations would combine in any ordered pattern among patients from SpA multiplex families.
SpA patients (n = 540) belonging to 190 multiplex families were thoroughly investigated. Clinical data was collected, systematic pelvic radiographs were taken, and HLA–B27 status was determined. The patterns of SpA manifestations were examined by several methods, including multiple correspondence analysis, nonhierarchical and hierarchical clustering, and discriminant analysis.
The nonhierarchical cluster analysis allowed us to classify patients, independent of disease duration, into 2 major groups of comparable size. Group A contained a majority of the women, whereas group B predominantly consisted of men. The 2 groups were very similar regarding axial symptoms, radiographic sacroiliitis, and uveitis. Group B was characterized by a younger age at onset and a higher frequency of clinical enthesitis, peripheral arthritis, dactylitis, psoriasis, and inflammatory bowel disease than group A. Patients belonging to those groups exhibited some degree of familial aggregation, thereby supporting their intrinsic validity.
Pattern analysis of SpA manifestations among familial SpA allowed us to recognize 2 main clusters independent of disease duration. Furthermore, there was a trend toward aggregation by cluster among families, suggesting that they are determined by specific genetic factors. These clusters may indeed correspond to different severity patterns.
The term spondylarthropathy (SpA) refers to a group of inflammatory rheumatic disorders that currently is di- vided into distinct diagnostic entities: ankylosing spondylitis (AS), which is characterized by predominant axial skeletal involvement and advanced radiographic sacroiliitis; a subset of psoriatic arthritis; arthritis associated with idiopathic inflammatory bowel diseases; reactive arthritis (ReA), which is characterized by an antecedent triggering infection; and undifferentiated SpA, which is defined by the lack of any of the former distinguishing features. A close relationship between the different SpA subtypes is supported by several observations, including anatomic studies of the skeletal lesions that have isolated enthesitis as a characteristic feature of SpA (1–3) and the role of HLA–B27, a genetic predisposing factor shared by the different varieties of SpA (4). There is, however, some difficulty in establishing a formal distinction between subtypes as defined above because of frequent overlapping presentations in individual patients. Hence, the SpA spectrum refers to the variety of skeletal and extraarticular inflammatory manifestations that may combine differently among patients (5, 6). This set includes axial manifestations (spinal and buttock pain or stiffness, radiographic sacroiliitis), peripheral arthritis, dactylitis, peripheral enthesitis (such as subtalar pain), acute anterior uveitis, psoriatic lesions, and inflammatory bowel disease (IBD; i.e., Crohn's disease, or ulcerative colitis).
A classic interpretation of the interrelationship between SpA phenotypes is that they result from different combinations of factors that contribute independent of each other to produce overlapping diseases (7). Such interpretation was supported by studies reporting that conditions such as AS and ReA breed true within families (8, 9), but not by our recent analysis of a large collection of French SpA multiplex families (10, 11). In the latter study, all possible SpA subtypes were represented among the patients. Nevertheless, axial skeletal involvement appeared almost constant. Furthermore, all the manifestations belonging to the spectrum of SpA appeared linked together and to HLA–B27 within families (11). Hence, neither articular or extraarticular SpA features segregated independent of each other, as predicted if independent factors were determining those manifestations. Our results supported a novel model in which a critical core of predisposing factors, including HLA–B27, were shared by different forms of SpA, whereas secondary factors determined phenotypic variations. Thus, SpA appeared more homogenous than previously thought, and we concluded that SpA subtypes should be considered as phenotypic variations of a unique disease rather than truly different conditions (11). One major consequence of this conclusion is that SpA should be studied as a whole for the purpose of genetic studies. This prediction allowed us to successfully map a non-major histocompatibility complex locus linked to SpA predisposition in multiplex families (12).
In our study, the significance of diversity among familial SpA was previously addressed by examining the clustering of SpA manifestations among families, and we found that extraarticular manifestations, such as psoriasis, appeared evenly distributed, which in turn suggested that factors influencing the occurrence of this manifestation were ubiquitous. In contrast, we found some degree of familial clustering both for peripheral arthritis (10) and for arthritis associated with IBD (4), as if specific predisposing factors were influencing those manifestations. Finally, uveitis could not be analyzed with confidence because of its dependence on disease duration. Nevertheless, it was clear that SpA tended not to aggregate among families as distinct entities. Hence, the classic subtypes that have been defined until now may not represent adequate subdivision. This assumption is more likely to be true that radiographic sacroiliitis, which has been taken as an essential distinction criteria, is highly dependent on disease duration.
The aim of the present study was to determine if any of the elementary manifestations that are part of the SpA spectrum combine in an orderly fashion, and whether this would lead us to propose a model of discrimination between SpA phenotypes different from the classic one. To achieve this goal, the phenotypic presentation of the disease in the context of familial SpA was thoroughly analyzed by means of clustering methods.
One hundred ninety French families with multiple cases of SpA according to the criteria of Amor et al (5) or the European Spondylarthropathy Study Group criteria (6), collected by the Groupe Français d'Etude Génétique des Spondylarthropathies as previously described (10, 11), were included in the present study. This updated cohort consisted of 540 SpA patients. Demographic and clinical characteristics of the patients are shown in Table 1. Proband refers to the SpA patient from each family who was examined first. Standard anteroposterior radiographs of the pelvis were routinely obtained and scored for sacroiliitis using an established grading system (13). A grade of II bilaterally or of III unilaterally was required for a definite diagnosis of radiolographic sacroiliitis. The diagnosis of psoriasis required the presence of typical lesions or a definite diagnosis by a dermatologist. A diagnosis of anterior uveitis was retained if established by an ophthalmologist. The diagnosis of IBD required typical gut endoscopic or histologic findings of Crohn's disease or ulcerative colitis.
|Characteristic||Probands (n = 190)||All patients (n = 540)||Men (n = 301)||Women (n = 239)||P†|
|Men, no. (%)||111 (58)||301 (56)||301 (100)||0 (0)||ND|
|Age, median (range) years||42 (16–78)||40 (9–82)||40 (9–82)||40 (10–77)||0.4|
|Age at onset, median (range) years||22 (6–58)||22 (5–59)||22 (6–59)||23 (5–59)||0.1|
|Disease duration, median (range) years||17 (0–59)||16 (0–63)||17 (0–59)||15 (0–63)||0.025|
|HLA–B27, no. positive/no. tested (% positive)||184/187 (98)||506/524 (97)||285/293 (97)||221/231 (96)||0.34|
|Inflammatory back pain, no. (%)||174 (92)||484 (90)||274 (91)||210 (88)||0.26|
|Buttock pain, no. (%)||169 (89)||460 (85)||251 (83)||209 (87)||0.22|
|Sacroiliitis, no. positive/no. tested (% positive)‡||127/189 (67)||321/524 (61)||210/290 (72)||111/234 (47)||< 0.0001|
|Peripheral enthesitis, no. positive/no. tested (% positive)||112/190 (59)||282/537 (52)||152/299 (50)||130/238 (54)||0.19|
|Peripheral arthritis, no. positive/no. tested (% positive)||82/190 (43)||198/537 (37)||119/299 (40)||79/238 (33)||0.13|
|Dactylitis, no. positive/no. tested (% positive)||43/190 (23)||110/539 (20)||52/300 (17)||58/239 (24)||0.053|
|Uveitis, no. positive/no. tested (% positive)||61/190 (32)||144/539 (27)||90/301 (30)||54/238 (23)||0.063|
|Psoriasis, no. positive/no. tested (% positive)||33/188 (17)||115/535 (21)||69/296 (23)||46/239 (19)||0.29|
|IBD, no. positive/no. tested (% positive)||9/190 (5)||25/539 (5)||8/300 (3)||17/239 (7)||0.022|
Standard serologic methods were used for HLA–B typing (14). In several cases, HLA–B typing was performed on DNA extracted from peripheral venous blood leukocytes using a polymerase chain reaction–based sequence-specific method (15). For individuals who had already been typed as positive for HLA–B27, retyping was not systematically performed.
Multiple correspondence analysis was used to graphically assess the positive or negative associations of the studied manifestations. Basically, multiple correspondence analysis corresponds to a weighted principal components analysis for multiway tables, and provides a 2-dimensional graphic representation of the association between the levels of categorical variables. On the graphs, a point in the Euclidean space is drawn for each level of the variables, and the distance between any 2 points is a measure of the strength of association between the 2 corresponding characters. As in principal components analysis, the axes are selected as best explaining the variability of the observed data.
Cluster analysis was used to identify subgroups of patients with similar characteristics. For clustering, a nonhierarchical method (16) based on a modified K-means algorithm was used, called nearest centroid sorting (17). The analysis was repeated for numbers of clusters varying from 1 to 15, and the optimal number of clusters was determined using the pseudo-F and the cubic clustering criterion statistics (18).
Preliminary analyses showed that clusters constructed on the whole sample of patients were highly dependent on disease duration and age at onset, both of which were linked together (Table 2). Disease duration may indeed affect clinical presentation, as previously described (10). Hence, the frequency of such manifestations as inflammatory back pain, radiographic sacroiliitis, uveitis, and IBD increased in relation to disease duration, whereas the frequency of other manifestations, such as clinical sacroiliitis (i.e., buttock pain), peripheral enthesitis, peripheral arthritis, dactylitis, and psoriasis, appeared stable whatever the disease duration. We therefore decided to perform separate analyses within the 3 following subgroups of patients, as defined by their disease duration: <10 years, 10–19 years, and ≥20 years. We also excluded from the clustering algorithm variables highly dependent on disease duration, i.e., uveitis, IBD, and radiographic sacroiliitis, to allow the identification of similar clusters in the 3 formerly defined subgroups. Variables considered for cluster construction were sex, age at onset, inflammatory back pain, inflammatory buttock pain, peripheral enthesitis, dactylitis, peripheral arthritis, and psoriasis. To achieve a standardized contribution of all variables, independent of their measurement scale, the variables were converted into z-scores. Otherwise, the continuous variable age of onset would have had more influence on the classification than other variables.
|<10 years (n = 146)||10–19 years (n = 161)||≥20 years (n = 197)|
|Age at onset, median (range) years||25 (6–59)||22 (6–55)||20 (5–50)||< 0.0001§|
|Inflammatory back pain, %||81||93||95||< 0.0001|
|Buttock pain, %||84||87||89||0.16|
|Sacroiliitis, %¶||37||72||76||< 0.0001|
|Peripheral enthesitis, %||51||53||56||0.41|
|Peripheral arthritis, %||34||35||41||0.15|
|Uveitis, %||11||29||40||< 0.0001|
Reproducibility of the classifications obtained within each subgroup was assessed using a nonparametric bootstrap procedure (19) based on 100 independent samples randomly constituted from the original observations, with the same balance of patients between clusters. The clustering method was then applied to each of these samples, the optimal number of clusters was selected as described above, and misclassification rates as compared with the original classification were then derived.
Additionally, the results of the nonhierarchical clustering method were contrasted to that of an agglomerative hierarchical algorithm, using Ward's cluster method (16, 20). Consistency between both methods was then assessed by the misclassification rates and the kappa statistic (21).
Most discriminating variables were identified by means of discriminant analysis with variable selection, and misclassification rates for such analyses were computed. The model for prediction of cluster membership obtained in the first 2 subgroups of patients (disease duration <20 years) was then applied for predicting classification in the last subgroup of patients.
To test if family members had increased probability to be in the same cluster as their proband, a generalized linear model with generalized estimating equations, which accounts for correlation between observations from the same family, was used (22). Results were given as odds ratios and 95% confidence intervals.
Differences between groups were tested by Wilcoxon's rank-sum test (2 groups; numerical data), by Kruskal-Wallis test (>2 groups; numerical data), or by Fisher's exact test (categorical data). All tests were 2-sided and a P value < 0.05 was considered statistically significant. All analyses were performed on SAS version 8.1 software (SAS Institute Inc., Cary, NC).
Multiple correspondence analysis was used to graphically assess patterns of patients according to the combination of clinical characteristics. Results of such an analysis performed on the whole cohort of patients are displayed in Figure 1A. Because HLA–B27-negative patients seemed to differ markedly from other patients, all subsequent analyses were performed after excluding this subgroup of patients. The new analysis showed a positive association between a lack of back pain and a lack of radiographic sacroiliitis (Figure 1B). A positive association between enthesitis, arthritis, dactylitis, and psoriasis could also be drawn from the figure. Remarkably, patients with IBD seemed to fall apart from others, although this manifestation appeared to be most associated with uveitis on one side and with peripheral arthritis and dactylitis on the other.
The construction of clusters was performed on the basis of all available variables that were not influenced by disease duration, except for back pain, which was included because it allowed improved clustering. Both for patients with disease duration <10 years (group 1) and for patients with disease duration of 10–19 years (group 2), the analysis resulted in an optimal division of the sample into 2 clusters, herein referred to as clusters A and B, each of comparable size (Table 3). Characteristics of the resulting clusters were very similar between groups 1 and 2. Indeed, clusters 1A and 2A included more women (52% and 80%, respectively) with older than average age at onset and with slightly more buttock pain than the B clusters. Clusters 1B and 2B, on the contrary, included more male patients (52% and 74%, respectively) with earlier onset and with more frequent peripheral skeletal manifestations, such as arthritis and dactylitis. In group 1, enthesitis was very discriminant for cluster separation, but not in group 2. The frequency of advanced radiographic sacroiliitis was remarkably similar between both clusters in either group. Finally, regarding extraarticular manifestations, patients in the B clusters had more frequent psoriasis, in both groups 1 and 2, whereas uveitis and IBD were equally frequent between clusters A and B (Table 3).
|<10 years (group 1)||10–19 years (group 2)||≥20 years (group 3)|
|Cluster 1A (n = 73)||Cluster 1B (n = 73)||P†||Cluster 2A (n = 78)||Cluster 2B (n = 83)||P†||Cluster 3A (n = 83)||Cluster 3B (n = 42)||Cluster 3C (n = 72)||P‡|
|Age at onset, median (range) years||27 (13–59)||23 (6–45)||0.0009||26.5 (14–55)||20 (6–40)||<0.0001||22 (6–50)||22.5 (11–43)||20 (5–38)||0.0003|
|Inflammatory back pain, %||83.6||78.1||0.53||92.3||92.8||1||96.4||92.9||95.8||0.62|
|Buttock pain, %||89||78||0.12||91||83.1||0.16||85.5||88.1||93.1||0.33|
|Peripheral enthesitis, %||8.2||94.5||<0.0001||52.6||54.2||0.88||38.6||78.6||62.5||<0.0001|
|Peripheral arthritis, %||13.9||54.2||<0.0001||11.5||56.6||<0.0001||22||55.8||55.6||<0.0001|
In contrast, patients with disease duration ≥20 years (group 3) were best divided into 3 clusters (Table 3). Clusters 3A and 3C were of comparable size, but differed from each other by sex distribution (a majority of men in 3A versus a majority of women in 3C), age at onset (slightly older in 3A), a higher frequency of radiographic sacroiliitis in 3A, but a much higher frequency of peripheral arthritis, dactylitis, enthesitis, and IBD in 3C. The remaining cluster (3B) was of smaller size, contained a similar proportion of the male and female populations (23% and 19%, respectively), and shared many characteristics with cluster 3C (i.e., frequency of peripheral arthritis, dactylitis, enthesitis, IBD, and radiographic sacroiliitis), whereas age at onset was similar to cluster 3A. Finally, cluster 3B differed mostly from 3C by the frequency of psoriasis (all patients in 3B and none in 3C had psoriasis) (Table 3).
Analysis of bootstrap resampling results consisted of computing the agreement between the classification obtained on the bootstrap samples and the original one. For group 1, the cluster algorithm divided the sample into 2 clusters for 80 of the 100 artificial data sets, in 3 clusters for 10 of 100, in 4 clusters for 6 of 100, and in 5 clusters for 1 of 100. The different criteria (pseudo-F and cubic clustering criterion) failed to select the same numbers of clusters for 3 artificial samples. In the case of the 80 artificial samples in which 2 clusters were selected, a correspondence was assigned between each novel cluster generated on artificial samples and the original cluster that had the largest proportion of its element in the new group, then misclassification rates were computed. The mean misclassification rates were 17% for cluster A and 29% for cluster B, corresponding to an overall misclassification rate of 23%. For group 2, two clusters were obtained in 56% of cases, 3 in 16%, 4 in 10%, and 5 in 8%. The method failed to conclude in 10% of the samples. For the 56 samples in which 2 clusters were selected, misclassification rates were 21% for cluster A and 25% for cluster B, which yielded an overall rate of 23%. For group 3, three clusters were selected as optimal in 35% of the samples, 2 in 35%, 4 in 12%, and 5 in 17%. The method failed in 1% of bootstrap samples. Correspondence between new and original clusters was more difficult to establish than in both former cases, because assignment of artificial samples into clusters 3B and 3C often seemed to result in a split of each original cluster into 2 new clusters. Global misclassification rates were therefore more elevated than for groups 1 and 2, namely 50%, 56%, and 37% for clusters A, B, and C, respectively. However, if clusters 3B and 3C were grouped into 1 single cluster, the misclassification rate for that composite cluster was only 16%.
Results obtained with another clustering algorithm (a hierarchical algorithm with Ward's method) are displayed in Table 4. For patients in group 1, Ward's method also yielded 2 clusters as best dividing the sample. Seventy (97%) of 72 patients originally classified in cluster 1A were also found in a same new cluster (A′), and 44 (61%) of 72 patients from the original cluster 1B were classified into the second new cluster (B′). Two patients could not be classified because of missing data. The global misclassification rate was 21% and the kappa statistic evaluating chance-corrected agreement was 0.58. We thus concluded that the classification was reasonably reproducible. For patients from group 2, two clusters were also found optimal. The first new cluster of patients (A′) comprised a majority of patients from the original cluster 2A (79.2%), and 49 patients from the original cluster 2B (60.5%) were in the alternative new cluster (B′). Three patients were excluded from the new classification because of missing data, and the kappa statistic agreement between both classifications was 0.40. In group 3, Ward's method divided the sample into 3 clusters, as the nonhierarchical algorithm. Each new cluster was formed with a majority of patients from the same original cluster. Respectively, 74.1%, 65.9%, and 48.6% of the patients from cluster 3A, 3B, and 3C were found in the corresponding new cluster (A′, B′, and C′). The global misclassification rate was 37% and agreement between both classifications was evaluated by a kappa statistic of 0.41.
|Ward's method||Disease duration (nonhierarchical classification)*|
|<10 years (group 1)||10–19 years (group 2)||≥20 years (group 3)|
|Cluster 1A (n = 72)||Cluster 1B (n = 72)||Cluster 2A (n = 77)||Cluster 2B (n = 81)||Cluster 3A (n = 81)||Cluster 3B (n = 41)||Cluster 3C (n = 72)|
|Cluster A′, no.||70||28||61||32||60||1||36|
|Cluster B′, no.||2||44||16||49||3||27||1|
|Cluster C′, no. (if selected)||-||-||-||-||18||13||35|
In group 1, discriminant analysis associated with crossvalidation showed that age at onset, arthritis, enthesitis, and uveitis were sufficient to classify 94.6% of the patients in the right cluster. In group 2, age at onset, sex, arthritis, back pain, psoriasis, and uveitis yielded 7% misclassification; and in group 3, age at onset, sex, arthritis, dactylitis, and psoriasis were kept to obtain a 6.1% misclassification rate.
As all discriminant analyses selected quite similar variables, we used the model derived for pooled groups 1 and 2 for classification of patients in the third group. Such a model was based on age at onset, arthritis, enthesitis, and psoriasis. The overall misclassification rate for group 3 patients was then 27.7%. More precisely, the classification obtained for patients with the shortest disease duration (groups 1 and 2) would have placed 57% of the patients from cluster 3A into cluster A and 78% of patients from clusters 3B and 3C into cluster B. Results are detailed in Table 5.
|New classification (as in group 1 + 2)‡||Original cluster (group 3)†|
|Cluster 3A (n = 81)||Cluster 3B (n = 41)||Cluster 3C (n = 72)|
A statistically significant trend toward familial aggregation by clusters was observed. The calculated odds ratio for being in the same cluster as the proband was 1.64 (95% confidence interval 1.04–2.58, P = 0.033).
The early description of entities such as AS, psoriatic arthritis, and ReA reflected the most striking presenting symptoms of patients. However, these categories were not expected to necessarily best reflect the underlying pathogenesis. Most intriguing was the early recognition of frequent clinical overlap between subsets, which eventually led to develop the SpA classification criteria (5, 6).
By studying familial SpA, we observed the relative inadequacy of the current subclassification to describe meaningful differences between patients (4). Here, we tried to generate a more appropriate description of phenotypes by performing pattern analysis. The principle of this study was to examine the combination of elementary manifestations that are recognized as part of the SpA spectrum, rather than the classic subtypes. From multiple correspondence analysis, we inferred that the few B27-negative individuals (about 3% of the whole set) were quite different from the rest of the population, and we decided to exclude them from further analysis to improve internal consistency. We could also observe some trends toward grouping of axial manifestations on one side and of peripheral skeletal manifestations and psoriasis on the other, whereas IBD separated from the rest of manifestations.
As a limitation, this first method could not take in account continuous variables, such as age at onset or disease duration. However, age at onset could influence disease presentation (23). Therefore, it was included among variables studied in the cluster analysis. Such variables as radiographic sacroiliitis, uveitis, and IBD appeared to increase in parallel with disease duration and were then excluded from the algorithm used to elaborate clusters. Moreover, in this type of cross-sectional retrospective study, disease duration is linked to some bias: 1) long disease duration correlates with young age at onset, as a consequence of left censoring; 2) long disease duration associates with increase in the risk of missing remitting manifestations, such as peripheral arthritis, peripheral enthesitis, dactylitis, and uveitis, which were routinely retrieved from medical history; and 3) disease duration would seem to affect presentation of the disease if the natural history of disease was changing over time. Therefore cluster analysis was limited to the 504 HLA–B27 patients with known disease duration, which were split into 3 groups of 150–200 patients each, spanning distinct disease duration intervals. Advantages in performing separate analyses in these subgroups were that analysis performed in the groups with the shortest disease duration were less sensitive to left sensoring effect and that consistency of the classification could be tested by comparing results obtained in several samples from the same original population.
In both groups with the shortest disease duration (1 and 2), the nonhierarchical cluster method classified patients into 2 major clusters (A and B), with quite similar characteristics. A majority of the women fell into the A clusters (65% altogether), whereas a majority of the men were in the B clusters (65%). Clusters A and B were very similar regarding axial symptoms, radiographic sacroiliitis, uveitis, and IBD. However, both B clusters were characterized by a younger age at onset and a higher frequency of peripheral arthritis, dactylitis, and psoriasis than the A clusters. The high similarity of results obtained in 2 independent sets of patients, with distinct diseases durations, was quite remarkable and unexpected to happen by chance. Reproducibility of the classification was also assessed by bootstrap resampling of the nonhierarchical method and by applying a hierarchical cluster Ward's method, both of which showed a good reproducibility of the clusters selected in groups 1 and 2. Interestingly, the best reproducibility was achieved in group 1, which was the least sensitive to the bias inherent to long disease duration.
Few variables segregated differently between clusters in group 2, as compared with group 1. The proportion of men and women in each cluster was more balanced in group 1 than in group 2, in which a large majority of men fell into cluster B and a majority of women into cluster A. This should be interpreted along with the increased proportion (+25%) of men in group 2 as compared with group 1. This observation could possibly be explained by a left censoring effect, according to which group 2 was enriched in patients with early onset (as shown by a younger average age at onset in this group compared with group 1), and thereby in men from cluster B. Peripheral enthesitis differed strikingly between clusters A and B in group 1 but not in group 2. However, this manifestation was also one of the most difficult to assess with confidence, especially in the group with the longest disease duration.
Clusters generated in group 3 were more difficult to interpret, and correspondence with those generated in both former groups was not straightforward. Hence, 3 clusters were generated, both by nonhierarchical and by hierarchical methods. However, bootstrap resampling produced with equal frequencies 2 and 3 clusters. One major variance with groups 1 and 2 concerned the early age at onset in all 3 clusters, which was similar to clusters 1B and 2B. This was expected as a consequence of left censoring, but questions the true significance of clusters formed in group 3. Nevertheless, most variables opposed cluster A to clusters B and C, except for clinical axial manifestations and uveitis (which were similar between the 3 clusters) and psoriasis (which was the main distinguishing feature between clusters 3B and 3C).There were several arguments to indicate that clusters 3B and 3C indeed corresponded to clusters 1B and 2B. The misclassification rate was improved during bootstrap resampling by combining the 2 former clusters. Furthermore, discriminant analysis was used to derive a classification model based on the 2 groups of shorter disease duration. When this model was applied to the third group, a large majority of clusters 3B and 3C fell into new cluster B. In contrast, cluster 3A was almost equally divided between new clusters A and B (suggesting the possibility that left censoring might have affected the formation of cluster 3A).
From these analyses, it appears that 2 major clusters were consistently described. The first one would correspond to a rather pure axial disorder with a relatively late age at onset. The second would correspond to a more diffuse pattern of disease with a high proportion of peripheral enthesitis, arthritis, dactylitis, and psoriasis. Both clusters appeared rather similar with respect to the prevalence of uveitis and radiographic sacroiliitis. There remains some uncertainty about sex distribution between clusters because women were more prevalent in the A clusters, in group 1, and above all in group 2, but not in group 3. However, due to the left censoring, it is possible that data in group 3 were less interpretable than in both former groups. The place of IBD also remains unclear because this manifestation was selective for cluster B according to data from group 3 but not from group 2. Only a prospective study could reliably help clarify these points. Interestingly, patients in cluster B might have been classified as having ReA if low stringency criteria were used (i.e., not requiring a recent proven triggering infection) (24). Interestingly also, characteristic features of cluster B, i.e., a young age at onset, peripheral arthritis, dactylitis, and psoriasis, have been identified as severity markers of SpA (23, 25). Then, the 2 clusters of SpA that were described in the present study could correspond to severity subsets. Their mild degree of familial aggregation, which was statistically significant, strongly suggests that these phenotypes are likely determined by widespread genetic factors. Consistently, heritability of age of onset and severity have previously been demonstrated in AS (26, 27). If this hypothesis is correct, it could also explain the previously reported true breeding of AS and ReA within families (8, 9). Furthermore, cluster assignment could be useful in genomescan analysis to map genes responsible for the diversity of SpA expression, and also for its severity.
Possible limitation of this study relates to the familial form of the disease, which may not adequately represent the general SpA population. Nevertheless, it should be mentioned that a recent survey found the familial form may account for as many as 50% of all SpA patients (E. Dernis-Labous et al, personal communication), and that no major phenotypic difference was identified between familial and sporadic cases in previous studies (4). Also noteworthy, the present study focused on the HLA–B27-positive patients because too few HLA–B27-negative patients were available. Then, additional studies will be necessary to determine if the model of classification proposed herein applies to other populations of SpA, including sporadic and HLA–B27-negative cases.
We gratefully acknowledge the contribution of Drs. Bernard Amor, Francis Berenbaum, Jean-Marie Berthelot, Jean-Paul Blanquet, Bernard Combe, Pascal Claudepierre, Emmanuelle Dernis-Labous, Agnès Duché, Jacques Fechtembaum, Sandrine Guis, Pierre Miossec, Aleth Perdriger, Anne-Marie Prieur, Xavier Puéchal, and Alain Saraux for patient selection.