Chronic pain patients can be classified into four groups: Clustering-based discriminant analysis of psychometric data from 4665 patients referred to a multidisciplinary pain centre (a SQRP study)

Objective To subgroup chronic pain patients using psychometric data and regress the variables most responsible for subgroup discrimination. Design Cross-sectional, registry-based study. Setting and subjects Chronic pain patients assessed at a multidisciplinary pain centre between 2008 and 2015. Methods Data from the Swedish quality registry for pain rehabilitation (SQRP) were retrieved and analysed by principal component analysis, hierarchical clustering analysis, and partial least squares–discriminant analysis. Results Four subgroups were identified. Group 1 was characterized by low “psychological strain”, the best relative situation concerning pain characteristics (intensity and spreading), the lowest frequency of fibromyalgia, as well as by a slightly older age. Group 2 was characterized by high “psychological strain” and by the most negative situation with respect to pain characteristics (intensity and spreading). Group 3 was characterized by high “social distress”, the longest pain durations, and a statistically higher frequency of females. The frequency of three neuropathic pain conditions was generally lower in this group. Group 4 was characterized by high psychological strain, low “social distress”, and high pain intensity. Conclusions The identification of these four clusters of chronic pain patients could be useful for the development of personalized rehabilitation programs. For example, the identification of a subgroup characterized mainly by high perceived “social distress” raises the question of how to best design interventions for such patients. Differentiating between clinically important subgroups and comparing how these subgroups respond to interventions is arguably an important area for further research.


Introduction
Chronic pain with moderate to severe intensity affects about 20% of the general population [1] and available treatment strategies rarely provide adequate analgesia [2]. A visionary goal for pain medicine would be the ability of basing analgesic treatment and other interventions-e.g., multimodal rehabilitation programs (MMRP)-on a precise understanding of the mechanisms of different pain conditions [3,4]. It seems probable that prevalent pain syndromes such as chronic widespread pain (CWP) or unspecific chronic low back pain (CLBP) are heterogeneous categories encompassing different (albeit probably inter-related) mechanisms. However, in randomized controlled trials, systematic reviews, and meta-analyses, these broad conditions are often regarded as homogenous categories [5,6]. Hence, true treatment effects may be "diluted" because only a subgroup of patients respond.
Subgrouping chronic pain patients into different "phenotypes" is therefore an important endeavour. For instance, using quantitative sensory testing, neuropathic pain patients and fibromyalgia patients have been classified into broad phenotypes irrespective of aetiology [7,8]. Another psychophysical phenotyping method is conditioned pain modulation (CPM) experiments [9]. Subgrouping of pain patients can also be done by combining interview questions with a simple neurological examination [8] or by using self-reported psychometric data from validated questionnaires (e.g., about pain intensity and pain-related interference, quality of life, depression, anxiety, fear-avoidance, acceptance, or catastrophizing) [8,10,11]. Hence, it is thought that subgrouping pain patients is an important gateway into a more personalized practice of pain medicine [12][13][14][15] and may provide support for health care systems to optimize resources and costs [16]. Identifying subgroups is particularly of interest in conditions such as chronic pain conditions associated with high prevalence, burden, and costs as well as diagnostic and therapeutic uncertainty [17,18].
Pain is a subjective experience modulated by psychosocial and contextual factors. Engel's biopsychosocial (BPS) model of disease and [19,20] has been very influential for pain medicine [21]. Per the BPS model, chronic pain is influenced by and interacts with physical, psychological, and social factors, and pain is often said to have three facets (cognitive-evaluative, sensory-discriminative, and affective-motivational aspect) [22]. It is commonly held that affective factors such as fear and depression, and cognitive factors such as catastrophizing, are important to take into consideration in chronic pain patients. However, the question of causality is not easy to answer [23][24][25][26]. The BPS model is also in keeping with the ability of the brain to modulate nociception via top-down pathways [22,[27][28][29].
Using hierarchical cluster analysis (HCA), this study subgroups patients assessed at a multidisciplinary pain centre using all psychometric data registered in the Swedish quality registry for pain rehabilitation (SQRP) [30]. Based on this cluster analysis, we regressed the variables most responsible for subgroup discrimination to uncover a deeper understanding of what The sponsors of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit for publication. The authors had full access to all the data in the study and had final responsibility for the decision to submit for publication.
Competing interests: One of the authors (AIL) is a statistician who owns a small company (Quality Stat AB) that sells consultancy services mainly to different quality registers and research groups in Sweden. For several years, SQRP (with funding from SALAR) has purchased her services to handle practical tasks around the registry including various statistical summaries to the participating clinics. This was also done for this study, requiring more demanding statistical analyses. However, in this specific study, EB performed the statistical analyses and the role of AIL was of a supervisory nature (hence the contributions statement in the manuscript and in the online submission form). In conclusion, her company did not sponsor the project and her commercial affiliation did not play a role in the project. AFA Insurance and research-ALF (Linköping University hospital) provided support in the form of salaries for authors EB and BG. SALAR provided support in the form of salaries for authors ALI. This does not alter our adherence to PLOS ONE policies on sharing data and materials. There are no patents, products in development or marketed products to declare. The specific roles of these authors are articulated in the 'author contributions' section.
correspond to high levels of well-being. On that basis, two summary components are calculated, i.e., a physical part (SF36-Phys) and a mental (psychological) part (SF36-Ment) [35]. Chronic pain acceptance questionnaire (CPAQ). The 20-item CPAQ was used [36]. The patient rates each item on a scale from 0 (never true) to 6 (always true), and results are summarized in two subscales: activity engagement (CPAQ-E) with scores ranging from 0-66 and pain willingness (CPAQ-W) with scores ranging from 0-54. High values mirror high activity engagement and high pain willingness, respectively.
Tampa scale for kinesiophobia (Tampa). In this instrument, individuals report fear of movement and (re)injury, i.e., movement is (wrongly) assumed to cause a new injury [37]. A 4-point Likert scale, ranging from ''strongly disagree" to ''strongly agree", is used on 17 items. Hence, the total score ranges from 17 to 68. The cut-off for kinesiophobia is 36 for women and 38 for men.

Statistics
Multivariate data analysis. We used SIMCA-P+ version 13.0 (Umetrics AB, Umeå, Sweden) for multivariate data analysis by projection (MVDA). More precisely, we performed principal component analysis (PCA), hierarchical clustering analysis (HCA) and, based on the groups defined by HCA, partial least squares-discriminant analysis (PLS-DA). We have previously in detail described the principles of PCA and PLS-DA [10, 39,40], and this will not be repeated here. Briefly, PCA is a technique that models the correlation structure of a dataset, and thereby enables the identification of multivariate outliers [41,42]. After outlier detection with PCA, we applied a bottom-up HCA to the principal component score vectors using the default Ward linkage criterion to identify relevant subgroups of patients. HCA complements PCA in the sense that while PCA identifies distinct clusters in multivariate space, HCA can find subtle clusters. In the resulting dendrogram, clusters were identified and, based on these groups, PLS-DA was performed using group belonging as Y-variables and psychometric data as predictors (X-variables). The PLS-DA model was computed to identify associations between the X-variables and the subgroups. This was visualized on a corresponding loading plot.
Traditional statistics. Based on the four groups defined by HCA, traditional inferential statistics (Kruskal Wallis Test, Pearson Chi-Square, Mann-Whitney U Test) were computed using IBM1 SPSS1 Statistics version 23. Effect sizes by Cohen's d were computed according to the formula where n is the number of individuals in group a or b, and SD is the standard deviation of group a or b [43][44][45]

Data overview and outlier detection
Between 2008 and 2015, data were available for 5111 patients. However, 394 patients with >50% missing values were excluded from further analysis. Hence, data from 4717 patients were overviewed by PCA. Because 49 strong outliers and three moderate outliers were excluded, 4665 patients were retained for subsequent analyses. The resulting PCA model (n = 4665, 3 PC, R 2 = 0.41, Q 2 = 0.25) had a well-centred score plot (Fig 1), in accordance with the removal of strong outliers.

Hierarchical clustering analysis
Based on this PCA model, a HCA was performed. In the resulting dendrogram, a level of four clusters/groups was chosen for subsequent analyses (Fig 2): group 1 (n = 1305, 28%); group 2 (n = 778, 17%); group 3 (n = 726, 16%); and group 4 (n = 1856, 40%). The interpretation of groups 1, 2, and 3 was simplified by the fact that the "group dot" was almost on either one of the axes of the plot. For group 4, however, this was not the case, and to  interpret group 4, an additional interpretative axis was added in the loading plot, in accordance with the standards of how to interpret PLS-DA models (Fig 5). The findings of the two loading plots (Figs 4 and 5) can be summarized as follows (a selection of variables is mentioned). More detail is provided below. Group 1 was characterized by the following:
• Relative good coping and acceptance skills, i.e., positive associations with CPAQ-E and CPAQ-W.
• Good psychological status, i.e., positive correlations with SF36-Ment and LiSat-Ment and negative correlations with HAD-A, HAD-D, and MPi-Distre.
• Low pain intensity (i.e., negative correlations with NRS7d and MPI-severe).  Group 2 was characterized as the opposite of group 1. Group 3 was characterized by the following: • High levels of "social distress", i.e., group 3 correlates positively with MPI-Pun and negatively with MPI-Soli, MPI-Supp, and MPI-Distra.
Group 4 was in many ways like group 2 in the sense that it shared much of group 2's characteristics concerning, e.g., HAD-A and the EQ5D-index (Fig 5). Moreover, in contrast to group 3, group 4 was characterized by low levels of "social distress", i.e., group 4 correlated positively with MPI-Soli, MPI-Supp, and MPI-Distra.

Four clusters of chronic pain patients
To illustrate important group differences, nine selected variables (three "directly painrelated", three "psychological", and three "social") are displayed as boxplots in Fig 6 (for significance levels and effect sizes, see Tables 2 and 3, respectively).
Demographics. Although the model was not influenced very much by sex or age in relation to other X-variables (Fig 4), both sex and age were nonetheless statistically significant ( Table 1).  Sex distribution differed statistically between group 2 and group 3 (P = 0.005) and between group 3 and group 4 (P = 0.016) ( Table 2). Hence, group 3 was characterized by a higher frequency of females. Although the difference was small ( Table 1), it is nonetheless notable that females were overrepresented in group 3 and characterized by high "social distress". Group 1 was characterized by a statistically higher age than the other groups ( Table 2). However, the differences were small, i.e., in median 3-4 years. Directly pain-related variables. Fig 4 suggests that pain intensity (NRS7d and MPI-sev) would be higher in groups 2 and 4, and this was confirmed by traditional statistics (Tables 1  and 2).
As seen in Fig 4, pain durations (i.e., PainDur and PainDurPer) were positively associated with group 3, which had indeed by far the highest median value of all groups for these variables ( Table 1).
NbPainReg differed statistically between all groups except between group 3 and 4 (near significance, Table 2). Notably, the greatest difference between groups concerning NbPainReg was between groups 1 and 2, the median value being almost twice as high in group 2 ( Table 1). Hence, spreading of pain was associated with high psychological strain (Fig 4).
Moreover, neuropathic pain conditions (i.e., neuralgia, lumbago with sciatica, and polyneuropathy) had low associations with group 3 compared to the other groups (Tables 4 and 5).

Discussion
The major findings of the study were that the four groups/clusters were identified, which had the following characteristics: Four clusters of chronic pain patients • Group 1 was characterized by low psychological strain and the best relative situation with respect to pain characteristics (intensity and spreading) and lowest frequency of fibromyalgia diagnosis as well as a slightly higher age.
• Group 2 was characterized by high psychological strain as well as the most negative situation with respect to pain characteristics.
• Group 3 was characterized by high "social distress", the longest pain durations, as well as a statistically higher frequency of females. The frequency of three neuropathic pain conditions was generally lower in this group.
• Group 4 was characterized by high psychological strain and low "social distress", as well as high pain intensity (NRS7d and MPI-sev).
Chronic pain patients are often treated as a homogenous group. Such an indiscriminate approach is problematic [46]. The Swedish Council on Health Technology Assessment (SBU) suggested that RCTs need to describe the included chronic pain patients in a more comprehensive and systematic way [47]. In the present study, four groups were identified. Although significant differences in distribution of certain diagnoses were found, the differences between the groups were small. The importance of the clinical diagnosis seems to be limited when planning for rehabilitation interventions.
Although cognitive behavioural therapy and MMRP are evidence-based treatments [31, 48,49], only low to medium effect sizes are usually reported. Effect sizes may increase if patients can be classified into meaningful groups [50][51][52]. Identifying subgroups has been identified as research priority to facilitate development of "tailor-made" interventions [52][53][54]. Subgrouping of patients with chronic pain has been done in various ways, and there is no consensus concerning the most suitable method and/or data set optimally to be used for subgrouping/ clustering [7, 10, [55][56][57][58]. From a clinical perspective, it appears important that subgrouping is built on easily assessed and clinically useful data.
Chronic pain is often co-morbid with psychological conditions [59]; 35% of the chronic pain population has co-morbid depression [60] and anxiety is reported in 17 to 35% of chronic pain cohorts [61,62]. In agreement with earlier studies, these comorbidities were of importance when the four groups were identified especially for groups 1 and 2 [12,13,16].
Groups 1 and 2 were the two extremes (together 45% of the cohort) and similar results have been presented earlier [63,64]. Group 1 presented a relatively good situation concerning pain characteristics, common comorbidities, and health aspects. However, this group still had on average a moderate pain intensity. This group also reported relatively positive situations for psychological status (including satisfaction), kinesiophobia (TAMPA), and acceptance (CPAQ). All subjects included in this study were complex patients in one or several aspects as judged by the refereeing physician (generally in primary health care). Thus, there is a substantial proportion of patents with complex pain conditions who do not report psychological comorbidities such as anxiety and depression [11]. Group 2 was characterized by high psychological strain and high pain intensity. This constellation of psychological symptoms and pain was associated with low health and suffering as demonstrated earlier [14,[64][65][66][67][68][69][70]. Patients in group 2 perceived their situation as being "worse than death" (i.e., negative figures) at group level per the EQ5D-Index. The identification of patients belonging to this group seems very important especially since there are reports that such a constellation of symptoms may negatively affect treatment outcomes, feed treatment resistance, and sick leave as well as be associated with high health costs [23,[71][72][73][74][75][76][77][78][79][80][81].
Group 3 was one of two intermediate groups and one of the pregnant characteristics of this group was the perceived lack of social support and support from significant others. Hence, the results of group 3 pinpoint that not only psychological but also social characteristics seem important to consider when assessing chronic pain patients. In a sense, for a subgroup of patients, we have confirmed the importance of the "social" component in the bio-psychosocial model. Further studies on the social dimensions of the chronic pain experience are warranted. Group 3 had a significantly higher proportion of women, and it has recently been shown that women patients (together with patients high in depressive symptoms) seem more vulnerable to spouse criticism/hostility [82]. Social support has been found to be associated with long-term functioning in post neck trauma patients [83]. Future research should evaluate to what extent and how gender aspects contribute to the rating of social support and evaluate whether certain interventions can be implemented to address the low social support, e.g., education for relatives. There is a lack of social support-based interventional studies [84]. As this group was also characterized by long pain durations, it appeared these patients to a certain extent handled their situation without markedly increasing their depressive and anxiety symptoms; this was not the case for group 2, where the highest values were found.
Group 4 -the second intermediary group-was the largest group (40%) of all the four groups and generally (except for two variables) (Tables 1, 2 and 3) had a somewhat worse (significant) situation than group 3, although these differences were small. Group 4 also differed significantly from group 2 (except for age and gender). Variables reflecting the perceived health situation (e.g., EQ5D-Index and EQ5D-VAS) were markedly lower in group 4 than in groups 1 and 3 (Table 1). Interestingly, group 4 was also characterized by relatively short pain durations compared to groups 2 and 3 (no difference compared to group 1). It seems that patients in group 4 are at risk for developing a clinical picture similar to group 2; however, properly addressing this speculation would require longitudinal studies.
In the future, subgrouping patients into meaningful clusters might have profound clinical implications. Should for instance group 2 focus (at least initially) on pharmacological treatment of depression and/or anxiety, and should group 1 focus on e.g. increased physical activity and not psychological treatment? Is perhaps group 3 in dire need of involving relatives in the treatment programme, for instance through education sessions? And, given that cognitive behavioural therapy (CBT) is less effective in patients with high levels of emotional problems [85], is perhaps group 4 better suited for CBT-based rehabilitation programmes than group 2? All of this is highly speculative but illustrates the potential of using complex patient-related outcomes measures (PROM) in order to hopefully increase the effect size of our interventions. A visionary goal for a truly personalized practice of pain medicine would be to combine PROM with clinical judgement (expressed in e.g. a diagnosis according to the upcoming 11 th version of ICD, where chronic pain diagnoses will have their own section) or even perhaps with new biomarkers and/or psychophysical tests mirroring chronic pain pathophysiology [86][87][88][89], in order to choose the best treatment for each individual patient.

Strengths and limitations
One of this study's strengths is the use of data from a steady influx of patients to a multidisciplinary pain centre responsible for the care of patients with severe chronic pain in a county in Sweden. The large sample size was associated with good power for identifying the four groups. On the other hand, these patients were mainly refereed from the primary care and hence represent a selection of the most complex chronic pain cases. Another strength is that this study was based on a wide array of aspects per the BPS model and in agreement with important facets suggested by ICF and the IMMPACT group. Using MVDA was also a strength since these methods are especially designed to handle and take advantage of the complex intercorrelation pattern of the investigated variables, and using this methodology we found substantial effect sizes ( Table 3)-confirming the validity of the MVDA methodology and suggesting that the subgroups might be clinically relevant. The cross-sectional study design was unable to identify directions of causality. Another limitation was that we used questionnaires instead of a faceto-face clinical examination of anxiety and depression, which has been shown to be associated with more robust assessments; however, systematic clinical assessments of depression and anxiety were not within the economic resources of the present project. The only systematic clinical assessment was the medical diagnosis.

Conclusion
Using MVDA four subgroups/clusters were identified. One group (group 2) was characterized by high psychological strain as well as the most negative situation with respect to pain characteristics and health aspects. The other extreme (group 1) had relatively low intensity of symptoms and a relatively good health situation. Two intermediary groups were also found; one of these (group 3) was characterized by high "social distress", the longest pain durations, as well as a statistically higher frequency of females. The fourth group (group 4) had a somewhat worse situation than group 3 but was also associated with relatively short pain durations. One can speculate that patients in group 4 could be at risk for developing a clinical picture similar to group 2. The four groups showed marked differences in clinical pictures. This may indicate the need to design "tailor-made" interventions.