Low back pain (LBP) ranks the highest of all conditions in terms of years lived with disability, both in Western Europe and Worldwide [1]. This has enormous economic consequences on health care expenditure and loss of work productivity [2, 3]. Complex biopsychosocial mechanisms are known to be involved in the development of chronic disabling LBP and as a result the variation in treatment prognosis for individual patients is considerable [4]. This means, that although a number of effective treatments are available, selecting appropriate treatment for individual patients remains a substantial clinical challenge for primary care where most patients are managed [5–7].

An internationally agreed research priority is to test whether the effectiveness of treatment results might be improved by early identification of patient-subgroups that are most at risk for developing chronic disabling pain [8]. Several different approaches for subgrouping patients with low back pain have been developed [9]. One with feasibility and demonstrated clinical and cost effectiveness is the STarT Back stratified care approach involving the use of the Keele STarT Back Tool (SBT), which has specifically been designed for primary care to fast-track patients to appropriate treatment based on prognostic information. The tool assists clinicians in determining an individual’s likely prognosis and consists of nine items, with the first four relating to biomedical factors and the last five identifying modifiable psychosocial risk factors [10]. Patients are then allocated to one of three prognostic groups (low, medium and high risk) using established scoring cut-offs which each have different matched treatment recommendations [11, 12]. Patients at low risk receive support to self-manage and are deliberately not over treated or investigated. Patients at medium risk are provided with evidence-based physiotherapy treatment such as exercise and manual therapy, and patients at high risk receive psychologically informed physiotherapy, which integrates physical and psychological approaches, aiming at reducing obstacles to recovery such as unhelpful beliefs and illness behaviors. Following promising research findings the approach is gradually being implemented into routine practice in the UK [13] and some initial research has reported promising findings that implementing this approach in Germany might be possible [14].

The SBT was originally developed in the UK and has been translated into over 25 languages with 15 articles validating the psychometric properties of different versions [15]. The German version (STarT-G) was developed following a structured and widely accepted cross-cultural translation process [16]. In addition, some initial psychometric testing on a small cohort in Switzerland has been reported [17]. However, as the sample size was small and only a limited number of measurement properties were tested, we aimed to conduct a larger scale German validation study.

The specific objectives of our study were to test the STarT-G’s construct validity, discriminative ability, internal consistency and test-retest-reliability and to compare findings with the values given for the original English version.

A cohort study with a two-week retest has been conducted. Patients with LBP were recruited from primary care (general practices and orthopaedic practices both with free access to care). The first questionnaire (t0) took place within the practice before consultation, and the second questionnaire was sent by post (t1) 10 days later. Patients who did not respond to the postal questionnaire were contacted by telephone.

Inclusion criteria were patients with non-specific LBP, aged 18 to 60 years. The diagnosis of low back pain was defined as being specific, if a patient had a cauda equina syndrome, an inflammatory disorder such as ankylosing spondylitis, or had a suspected serious pathology such as a tumor or vertebral fracture. No restrictions were placed on the duration of a patient’s back pain symptoms. Patients were excluded, if they had consulted the physician within the last twelve weeks, had undergone spinal surgery within the last six months, or if they were unable to complete the study questionnaires due to poor German language skills. Anonymized information on eligible patients’ age and gender was obtained regardless of study participation (“consent list”).

The retest-material was sent to patients 10 days after the baseline assessment from the study-center. This duration was set to counter memory effects. Since it was likely that the health status would change at least for a part of the patients, an additional question on the subjective estimation of whether their complaints had changed over this period, was added [18]. Patients who did not respond to the postal questionnaire within two weeks were telephoned and reminded to send the questionnaire, or alternatively asked to answer a limited set of questions. The retest-process was managed by a specifically constructed database to ensure the maintenance of the predefined time intervals.

Ethical approval was granted by the Ethics Committee of the University of Heidelberg (registration ID: S-414/2013). All patients gave their written informed consent for participation before entering the study in the participating practice.

Due to missing answers (9 questionnaires) or invalid answers (12 questionnaires, both (missing and invalid answers) 2 questionnaires), it was not possible to calculate the STarT-G total score for 23 patients and subscale score for 19 patients.

The AUCs for STarT-G’s ability to discriminate reference standard cases ranged from 0.70 to 0.77, indicating acceptable discrimination (Table 2 and Fig. 1a–d). Since the lower CIs all fell short of 0.7, the STarT-G’s discriminative abilities were not equivalent to the original SBT version according to our pre-defined criteria.

	AUC	SE	95 % CI
STarT-G Total - RMDQ	0.76	0.04	0.68	0.83
STarT-G Sub - PCS	0.70	0.04	0.63	0.78
STarT-G Sub - HADS-D	0.71	0.05	0.61	0.81
STarT-G Sub - CRS	0.77	0.09	0.60	0.94

HADS-D: Hospital Anxiety and Depression Scale (Depression), PCS: Pain Catastrophizing Scale, RMDQ: Roland Morris Disability Questionnaire, CRS: composite reference standard

With α = 0.55 for the subscale score, and α = 0.52 for the total score, internal consistency was poor. Since Cronbach’s alpha fell short of the predefined threshold for equivalence, equivalence to the original version of the SBT was not reached.

Data from 30 patients, who self-rated their health problems as unchanged, was included in the analysis for the test-retest-reliability. With Kappa = 0.67/0.68 (95 % CI 0.46-0.81/0.48-0.84) for the total score and the subscale score respectively, agreement was good and equivalent to the original version of the SBT. An additional calculation for the risk-groups resulted in Kappa = 0.28 (95 % CI 0.23-0.55). The planned sensitivity analysis excluding retest data gathered via telephone was not possible as none of the four patients responding via telephone rated their health problems as unchanged.

With equivalence to the original SBT version, no floor or ceiling effects were found for the STarT-G (3.9 % (n = 8) patients with 0 points; no patients with 9 points).

The STarT-G did not show identical psychometric properties to the original English version although our risk-group distribution was comparable to the original SBT validation study. The measurement properties of the STarT-G showed good test-retest reliability for the total score and the subscale score, acceptable discriminative ability and no floor or ceiling effects. Correlations with the total score reference standards for disability, and subscale score for catastrophising were acceptable. The correlations with the subscale score reference standards for kinesiophobia and depression and the test-retest reliability for risk-group allocation among patients who reported stable symptoms were weak. The psychosocial subscale of the STarT-G (items five to nine) should not be regarded as one distress factor, as the internal consistency of the corresponding items was poor.

According to pre-defined criteria the properties of the STarT-G differ in a number of respects from those of the original SBT. It is common that measurement tools perform less well in further validation studies in comparison to the developmental study [35]. In addition, we were not able to identify commonly accepted methods for the comparison of the measurement properties of the two instrument versions. A method we found to compare the tools’ discriminative abilities has been subject to criticism [27]. Therefore we used a pragmatic approach, which may have led to an over-estimation of the differences between both instruments. Nevertheless, international comparisons between different cohorts using the STarT-G and the SBT should be undertaken cautiously.

Our decision to exclude patients who had consulted their physician within the previous twelve weeks meant that in comparison to the original study sample and other SBT related studies we had greater numbers of patients with a short duration back pain symptoms [36–41]. As a result, some of the differences in the STarT-G’s psychometric properties may be due to the differences between the populations studied. The results of Beneciuk et al. support this perspective, as they specifically analyzed the relevance of the timing of the SBT and found that especially in acute high-risk patients, tool scores often change within a few weeks. They therefore suggest further research examining the potential of repeated measures in helping to improve the accuracy of prognostic assessment in this subgroup [42].

In the original SBT validation study the psychosocial subscale was confirmed to be one overall ‘distress’ factor. However, this finding was not replicated in our STarT-G validation study and in fact is consistent with a Danish validation study which described the SBT as ‘a multidimensional questionnaire consisting of one or two screening questions for each of eight underlying constructs’ and therefore recommended a stronger item-based approach for validation [36]. Whilst this finding is of interest, it might not have a large impact on the STarT-G’s ability of stratify patients in primary care.

In relation to relevant reference standards our results identified that the psychosocial subscale of the STarT-G and the TSK (for fear avoidance beliefs) had the lowest correlation. The reason for this issue is unclear although it is noted that in an updated version of the TSK, the instrument includes six items less than the version we used. This reduction was triggered by problems described for the TSK factor-structure in the international literature [20]. Therefore, in future studies it might be useful to include other instruments for fear avoidance or other versions of the TSK. For the study described in this manuscript we slightly modified the wording of two previous STarT-G items, number five about fear of movement (kinesiophobia) and number eight about low mood (depression). It is possible that the lower correlation between the subscale score and reference standards for kinesiophobia and depression was due to this re-wording, however this would seem unlikely, especially since the new formulation of item five resulted in a higher conformity to the original TSK-item which was used to develop the STarT-Tool [8].

Missing answers leading to invalid STarT-G scores - although anticipated - were only a small proportion (<4 %), with more than half of these due to patients giving ambiguous responses, e.g. ticks in between the ‘Agree’/’Disagree’ boxes. For the Chinese version of the SBT missing data was also an issue although to a lesser extent [41]. Our results suggest that an estimated tenth of all patients return incomplete STarT-scores making it impossible to calculate their risk subgroup without further enquiry. This occurred despite specific attention from the study team to ensure appropriate instructions were in bold, such as ‘answer each question if possible and try not to spend too long over your answers but pick the answer that first comes to mind’. In clinical practice therapists clearly have the opportunity to ask patients if this occurs. However, we recommend that a clear definition and evaluation method of missing-data for the SBT is considered in order that more responses with missing data are able to be used.

In Germany, it is likely that physiotherapists and primary care physicians will be the primary users of the STarT-G [13]. This study, provides them with clear information on the instrument’s measurement properties. The stratification tool is designed to assist clinicians in their decision-making process and not to replace their decision-making. In addition, the STarT Back approach upskills physiotherapists, through a training course, to address the complex needs of high-risk patients through the delivery of ‘psychologically informed physiotherapy’. Evidence suggests that trained physiotherapists are effective in managing around 85 % of this high-risk complex patient subgroup [43–46]. Nevertheless, the systematic review by Kenny et al. concludes that, from an international perspective, the SBT should still be considered as being in a developmental stage and should therefore be used with caution in practice [15]. Our results indicate that this also applies for the German version.

Strengths of this study were that the design was clearly planned a priori and rigorously applied. Experiences from a first small validation pilot in Switzerland reinforced the selection of primary care practices as the appropriate setting for this study since the clinical context has a strong influence on the answers patients provide. Moreover, in contrast to the Swiss study, pre first-contact consultation as the first time-point for administrating the STarT-G was chosen to include patients who had not yet received any treatment [17, 47]. In addition, a half-year mailing is currently being carried out, to determine the predictive ability of the STarT-G. Other strengths include achieving an appropriate sample size and ensuring non-responders were contacted by telephone to ensure a high follow-up rate. In respect to interpreting the test-retest-reliability findings some caution is appropriate, since the Kappa-value calculations were limited to a relatively small subsample of patients who self-reported their health problems as being unchanged for the observed study period.

The potential for cost savings and better cost-effectiveness of the STarT-approach has been demonstrated by Hill et al. in a large randomized controlled trial [11] and is also supported by the results from a prospective population-based sequential comparison observing the implementation of the STarT-Back-Approach in routine health care [13]. Nevertheless, results have yet to be replicated in other countries. Foster et al. propose that randomized trials are the preferred optimal design for this [9]. Correspondingly, the next step should be a feasibility pilot to examine the approach in Germany and provide information to help develop a clear implementation strategy. We have therefore conducted a related qualitative study with GPs and physiotherapists to compliment this research [14].

The STarT-G shows overall acceptable psychometric properties, although some differences with the original English version were identified. These included the items previously regarded as a psychosocial subscale being found to be more than one construct, and so we recommend this subscale is better understood to be a collection of individual psychosocial items. Further research utilizing the STarT-G should consider our findings and pay attention to establishing methods to deal with missing values.