The classification of patients with non-specific low back pain (NSLBP) into subgroups has a number of advantages. First, this approach is thought to be superior in guiding treatment compared with other approaches (1–4). Secondly, classifying patients based on their signs and symptoms has demonstrated diagnostic success in studies investigating specific low back pain (LBP) (5) and is a feasible, practical, cost-effective and clinically applicable process. Finally, clinical reasoning is facilitated by the utilization of diagnostic/prognostic classification systems, especially in cases where a specific cause of disease is indefinable (e.g. in NSLBP) (6).

However, a primary shortcoming of classification systems is their subjective nature. Most classification systems have been developed using a judgmental rather than a statistical approach (basing their development on small samples and relying on clinicians’ personal experience and expertise); therefore this is likely to bias the process, and reduce the generalizability of the results (7, 8). Furthermore, little work has been undertaken to address issues of reliability and to ensure agreement on the selection of discriminatory criteria for the classifying criteria. As such, many classification systems lack validity and reliability, and do not lead to confidence in the derived subgroup profiling. Thus, although there is emerging evidence that classification systems are the best method for diagnosing and/or guiding treatment for NSLBP, basic steps towards their development require further elaboration.

In view of the high incidence of LBP in the general population and the lack of an established NSLBP taxonomy adopted by clinicians, it would seem appropriate to implement basic steps in subgroup development by trying to reach consensus on reliable clinical items that are believed to be discriminatory for identifying different NSLBP subsets. Consensus amongst experienced clinicians has important benefits. First, if a high level of consensus is reached, it is more likely that these agreed items will be useful and worth further consideration and study. It is also possible that these items will suggest a degree of discriminative validity in identifying LBP subsets. Secondly, the outcomes may facilitate cross-cultural comparisons, since cultural trends in the diagnosis of LBP are suggested across different settings (7, 9).

Thus, the aims of this study, which forms the second part of a two-part series of studies, were to explore consensus on potentially discriminatory clinical items in a NSLBP assessment list, and to identify whether homogenous sub-groups are being developed. The initial study (10) reported the inter-tester reliability of this discriminatory list, enabling the identification of reliable items, which then formed the examination procedure for investigating sub-group developments.

METHODS

Sample

A convenience sample of Greek patients with NSLBP was invited to participate in the study. Patients were diagnosed and referred for physiotherapy by specialist doctors (mostly orthopaedic specialists) working predominantly within the private sector. The evaluations took place in 3 different testing sites (physiotherapy clinics) within Greece (Athens, Patras and Lamia) and volunteers were recruited consecutively as they visited each clinic over a 10-month period. Patients were excluded if: (i) their native language was not Greek, (ii) their LBP was due to specific spinal pathology, (iii) they had undergone lumbar surgery, (iv) they were pregnant, or (v) they had severe neurological problems (that influenced their cognitive or motor performance). A total of 106 patients with LBP consented to participate. In addition, 7 physiotherapists (5 men, 2 women), clinically experienced in treating LBP, agreed to perform the assessments (2 based in Athens, 2 in Patras and 3 in Lamia). Ethical approval was obtained from the ethics committees of the Technological Educational Institute (TEI) of Lamia, Greece and the University of Manchester, UK.

The first study accompanying this publication (10) reports the inter-tester reliability of a suggested discriminatory list, derived from the consensus of a large number of health professionals (11, 12). This reliability study enabled the identification of items that presented with a kappa value of 0.41 or more, corresponding to moderate and substantial/excellent agreement (13). These items were included in a final examination list. Items with kappa value of less than 0.41 were excluded from the examination, as they were considered to be of low reliability. In addition, items with almost perfect percentage agreement (for which kappas could not be computed) were not included; the reason for their exclusion was the non-existence of positive findings on these items, which indicated that they would not have any discriminative ability (14), and were thus discarded from the examination. Based on these criteria, the final examination list comprised 82 items (51 oral questions and 31 clinical tests). Thus, each physiotherapist (PT) conducted a full clinical examination (Appendix I), addressing the items that were deemed reliable, as reported in the accompanying study (10).

In the first section of the examination, which involved history-taking, the PT read each question from the examination booklet and recorded the answers. The second section involved a detailed physical examination. All PTs were familiarized with each testing procedure, following a training session directed by the principal investigator. Training was accompanied by a booklet summarizing the key examination points. Each PT assessed 15 patients, except for one PT who assessed 16. In addition, self-administered outcome measures for pain, disability and psychosocial status, which had undergone a cross-cultural validated adaptation procedure into the Greek language, and were highly recommended for use in LBP studies (15, 16) were also administered to the sample. For each patient the whole evaluation procedure lasted approximately 40 min (15 min for the questionnaires and 25 min for the examination).

Data analysis

To investigate whether there were distinct clinical subgroups within the sample, a cluster analysis approach was utilized. Cluster analysis is an exploratory procedure that aims to classify data into subgroups (17). In this study, a K-means non-hierarchical cluster analysis approach was utilized; this method is considered appropriate for producing exactly K different cluster solutions of maximum possible distinction. However, as this procedure cannot indicate how many clusters exist in the given data, in order to obtain an indication of the possible clustered options, hierarchical clustered modelling (14) was applied initially, and the dendrogram (formed by the hierarchical clustering) suggested further investigation of 2–6 cluster solutions. In addition, from the cluster analysis of 4 memberships, two of the groups contained two patients each, whose characteristics did not appear to be clinically different from the other two subgroup options. Thus, as this 4 membership option was not able to provide clinically meaningful results, a decision was made to stop the statistical cluster analysis at the 4 membership option. Thus, a K-means cluster analysis was carried out from 2- up to 4-group membership. K-means clustering generates an analysis of variance (ANOVA) table, whereby the mean on each item (for each cluster option) was explored to assess how “distinct” the within-cluster differences were in relation to the p-value and the magnitude of the F values. Thus, both F and p-values were used for descriptive purposes. Statistically significant p-values set at 95% level (p < 0.05) were also used to distinguish between items. Cross-tabulations on each item within each cluster solution were also utilized as a means to determine whether observed differences in magnitude across the clusters were clinically “important” in size; and from the resulting χ2 values the strength of association of each variable assigned to each group was tested. Finally, the distribution of each item across the subgroups was used. All data were analysed in SPSS (version 15.0).

RESULTS

A total of 106 patients (43 males, 63 females) participated in the study, mean age 36 years (standard deviation (SD) 15.9 years, age range 18–70 years). Fifty-eight patients (54.7%) had LBP for more than 12 weeks and, for most of them (87.7%), the pain was of a recurrent nature. The sample’s profile and main examination findings are shown in Tables I and II, respectively. The examinations were conducted by 7 PTs with mean clinical experience with LBP of 11.8 years (range 7–19 years).

K-means cluster analysis generated an ANOVA table, whereby each item’s mean explored how “distinct” the clusters were.

For each cluster option 104 patients were finally computed (2 were missed due to missing data).

Two-cluster option

The 2-subgroup option consisted of a small group of 24 patients with NSLBP and a larger group of 80 patients with NSLBP. A total of 29 items (21 from the patient’s history and 8 from physical examination) yielded statistically significant values. The small group (Group 1) had greater “severity” in their presentation and outcome measures’ scores compared with Group 2. More than half of this subgroup’s patients had pins and needles (58.3%), neck pain (75%), restricted lumbar movements (62.5–66.7%) and had their lumbar spines radiographed (83.3%), as opposed to less than half from the larger group (23.7–42.5%) with these symptoms. In addition, the smaller group comprised more patients with deep pain, predominant leg pain, aggravation in lying and night pain. Two pain location items (left foot and anterior leg pain) were present in larger percentages (20.8% and 12.5%) in the small group compared with the larger one (3.7% and 0%), indicating greater peripheralization of symptoms. A larger proportion of the small group presented with stiffness (66.7%) and dragging feet (25%) and had undertaken MRI diagnostic tests (40%) compared with the larger group (36.2%, 2.5% and 18.7%, respectively). In addition, straight leg raise (SLR) presented with pain and positive responses in 33.3% and 37.5%, respectively, in the small group, compared with the large one (8.7% and 13.8%, respectively). Table III summarizes the ANOVA outputs, χ2 scores and clinical item distributions for the two membership option (due to the large volume of data, 82 items being included, only statistically significant items are presented).

Three-cluster option

The 3-cluster option comprised a large group of patients (n = 73) and 2 smaller ones (n = 9 and n = 22 patients, respectively). Twenty-two history and 8 physical examination items achieved statistical significance. From the distribution of outcomes and clinical features, one of the groups (n = 22) appeared more “severe” compared with the other two groups in terms of referred and widespread (i.e. neck) pain, other symptoms, recurrent episodes, investigations and restricted movement. However, the other groups did not show any consistent differences (indicative of a particular pattern) in their characteristics.

The 4-subgroup option entailed a large group (n = 79), a smaller one (n = 21) and two 2-patient groups. Eighteen out of 36 items achieved statistical significance; however, in view of the limited number of patients in the two subgroups, consistent distribution patterns or distinctive characteristics across groups were not obtained.

DISCUSSION

This study aimed to explore a list of reliable and potentially discriminatory items for NSLBP, on their ability to distinguish amongst different patient subsets. The sample utilized was predominantly recruited by medical referrals from the private sector, which is well developed in Greece (18), and consisted of a mix of patients with acute and chronic LBP, who were moderately disabled by LBP. The sample’s profile had comparable features to that of some previous classification studies (19, 20), and a marginally less severe clinical profile compared with some others (21).

Following cluster analysis within the 2-cluster option, the small group had a more severe clinical profile compared with the larger group. Based on the distribution of the clinical items on the 3- and 4-cluster options, it could be argued that they both included a small and a larger group that possessed similarities to the 2-subgroup option. However, in the 3-cluster option one of the smaller groups was not distinctively different to the other two, whereas for the 4-group option, two of the groups were extremely small and their distribution patterns did not subsequently indicate any clinically meaningful solutions. Thus, given the above information, the 2-subgroup option provided the most comprehensible and clinically meaningful presentation profiles (compared with the other 2-cluster options). Furthermore, despite outcome measures not being included in the cluster model, the scores for the small group on the outcomes of pain intensity, disability, anxiety, etc. were much higher, further verifying the presence of more severe cases in this group.

Although direct comparisons between this study’s two subgroups and previous classification reports are difficult to perform with accuracy due to methodological differences, some similarities are evident. Pain location, aggravating factors, 24-h pain patterns, pain response to movement, and symptom duration are criteria included in several European (19, 22, 23), Canadian (24), New Zealand (25) and US (26) studies. Similarly, a wide range of physical features have been utilized, most common of which are SLR (22–24, 27) and lumbar mobility pain provocation tests (28, 29). These similarities give weight to the inclusion of these factors in such classification approaches.

From the studies limited to 2–3 subgroups, comparisons are still difficult due to variability in their designs and lack of reported detail on item distribution across the generated groups. Whilst Langworthy & Green (30) identified 3 criteria with similarity to this study (pain getting worse, night pain and exacerbation in lying), the subgroups’ distribution on these items was not detailed. Similarly, despite the interesting results in the study by Hill et al. (21) comparing 2 validated questionnaire instruments in their ability to identify subgroups requiring early intervention, limited information was given on their subgroups’ distribution characteristics. McCarthy et al. (19) were the only group that adopted a similar approach to this study and generated similar results. Their cluster analysis pointed towards two subgroups; a smaller more dysfunctional one (called the “hypervigilant” group), and a larger (less dysfunctional) one. Distribution of age, gender and outcome measures across their subgroups was comparable with this study. In addition, distribution on several clinical items (i.e. SLR test, upper lumbar palpation, pain provocation tests) pointed towards more positive (severe) responses for the hypervigilant group. The presence of a UK study (19) with similarities to this Greek one provides some confidence in our results, and their possible generalizability in wider population samples.

Some issues were of clinical importance. First, a large set of signs, symptoms and clinical measures have been utilized. Secondly, most previous classification studies lack the standards required to be deemed methodologically “rigorous” (i.e. lack of statistically developing clusters, utilization of empirical methods, lack of item reliability, etc.). Thirdly, apart from utilizing a statistical procedure for generating subgroups, which is considered more sophisticated than utilizing other methods (i.e. observational or judgmental approaches) (8), the present study selected items suggested to be “discriminatory” by a consensus procedure involving a large and representative sample of health professionals dealing with NSLBP within Greece (11). In addition, only the most reliable of these selected items were utilized (10) in order to improve confidence and objectivity in outcomes. In addition, this study aimed to provide a meaningful, practical and useful taxonomy within the Greek healthcare setting. In view of existing cultural variations (7), it has not been assumed that classification systems developed in a given cultural setting can be adopted and utilized in the Greek setting. Therefore, this exploratory study tried to develop clinical subgroups based on practical, clinically applicable, reliable and generalizable classifying criteria. However, whether this approach is clinically useful as an assessment-based process for targeting treatment or as a prognostic guiding path is currently untested and needs to be further investigated.

In terms of the methodology utilized, the K-means cluster approach was considered more appropriate than a hierarchical one because it is easy to use, reliable (no “multiplicity” effects, repeatable cluster generation, etc.), has the ability to produce distinct non-overlapping clusters (14, 17) and has been used in LBP exploratory studies (23, 30, 31). However, in view of its limitation in indicating a priori the exact number of existing clusters, a second clustering approach, the hierarchical method, a commonly used adjunct approach (17, 32), was utilized to verify the generated clusters and increase the validity of the findings (14). In addition, interpretation of the resulting partitions by descriptive means (as shown in Table III) was conducted, as recommended, for evaluating their clusters’ clinical utility (14). However, whilst cluster analysis has merit as a data-driven analysis procedure (as indicated above), it has pointed towards a relatively simple model of analysis, which can provide only basic discrimination of data. Thus, given the complexity of the NSLBP problem, as well as the fact that there is currently no optimal classification scheme, it appears that further testing is required in order to ensure reproducibility and enhance credibility of the study’s subgroups. Perhaps utilizing more complex analyses (such as neural networks or data-mining) or even combining approaches, to allow subgroups’ comparison and observation of similarities and differences between approaches would be desirable and beneficial. Interestingly, in Kent et al.’s review (33) on the research methods utilized for sub-grouping LBP, they proposed a method framework comprising 6 stages, including hypothesis setting and testing studies, validation studies, as well as impact analyses studies. They highlighted the need for sub-grouping research to proceed through all phases of study, in order for the developed subgroups to gain credibility, generalizability and applicability within clinical practice.

A limitation of the present study is that most discriminatory items utilized were biomedically based despite the profound role of psychosocial factors in NSLBP (34, 35). Psychosocial factors were excluded on the basis of their poor reliability. Nevertheless, certain social factors (work, hobbies, physical activities, etc.) and the psychosocial measures’ scores (Hospital Anxiety and Depression Scale, Fear-Avoidance Beliefs Questionnaire, Pain Catastrophizing Scale, presented in Table I) differed across the two groups, indicating greater psychosocial overlay for the small group compared with the large one. Interestingly, only 4 studies incorporated biopsychosocial elements in their classification (36–39); yet, they were utilized in a very different and, consequently, non-comparable way. Another limitation concerns the sample utilized, which was limited in terms of representing the more “disabled” patients; the presenting sample consisted of LBP patients with, in general, low levels of disability. This could have precluded the potential development of another subgroup with more “dysfunctional” features. However, this exploratory study constitutes only the first step towards developing a classification system within a particular cultural context.

In conclusion, this cross-sectional exploratory study identified the existence of two distinct subgroups by utilizing a cluster analysis approach; a small group with more “severe” and widespread clinical signs and symptoms, and a large group with low severity, dysfunction and symptom presentation. This preliminary study forms the first step in developing a classification system within Greece based on discriminatory and reliable clinical criteria. However, despite its advantages, cluster analysis provides a simplistic method of subgroup research, and further work should thus explore each subgroup’s clinical and diagnostic utility, in larger samples.

ACKNOWLEDGEMENTS

The authors would like to thank physiotherapists Ioannis Stathopoulos, Panagiotis Trigkas, Savvas Spanos, George Krekoukias and Nikolaos Strimpakos, as well as all the patients who participated in this study. This research was funded by the Manipulation Association of Chartered Physiotherapist’s Elsevier Science and Doctoral Awards for Research in Manipulative Physiotherapy.

REFERENCES