Predictive ability of the start back tool: an ancillary analysis of a low back pain trial from Danish general practice

STarT integrates biological, psychological, and social factors and includes nine questions that are used to subgroup patients into a low-, medium- or high-risk subgroup according to the risk of persistent disabling pain [13]. For each subgroup, the STarT follows a set of recommendations for treatment. Patients in the low-risk group are recommended to receive information on LBP and advice to stay as physically active as possible and to continue daily activities. Supplementary to information and advice, GPs are expected to recommend standardized treatment focusing on addressing physical symptoms and function to patients in the medium-risk group. In addition, healthcare professionals are expected to pay special attention to cognitive behaviour to address psychosocial obstacles to recovery for patients in the high-risk group [14]. The STarT has been found to be effective in predicting functional outcomes and has also been found to be effective when applied in two large studies in UK settings [15, 16]. Currently, stratification by the STarT is recommended in the newly published NICE guidelines [10].

Numerous studies performed in different healthcare settings have tested the predictive ability of the STarT. The findings from these studies are inconclusive, hampering widespread use across different healthcare settings [17‐24]. In a recent guideline implementation trial for patients with LBP, a subgroup of patients completed a range of questionnaires, including the STarT at baseline and the Roland-Morris Disability Questionnaire (RMDQ), after 4, 8, and 52 weeks [25]. These data provide the opportunity to perform an ancillary analysis of the guideline implementation trial and study the STarT’s predictive ability in a population consisting solely of patients consulting general practice. In a UK primary care setting, a 30% improvement between baseline and follow-up has been estimated as guidance for defining clinically relevant improvement in function when applying the RMDQ [26].

The aim was to study whether the STarT score for patients consulting general practice with LBP was predictive of a functional improvement of 30% in the RMDQ score after 8 weeks.

This was an ancillary analysis of a cluster randomised controlled trial on guideline implementation for LBP in Danish general practice. Reporting of the present study follows the STROBE Statement [27].

From January 2013 to July 2014, 60 general practices in the North Denmark Region participated in a guideline implementation trial. The cluster randomised controlled trial compared two strategies for supporting the implementation of LBP guidelines with the primary aim of reducing the referral of patients from primary care to secondary care. General practices in the intervention group had an outreach visit from a guideline facilitator, were offered access to feedback on their treatment of low back pain, and had the opportunity to score their patients with STarT (which was embedded in their electronic medical record). The GPs’ STarT scoring results are not reported in this study. Practices participating in the guideline implementation study had a project module installed in their electronic medical record system, and GPs were encouraged to perform diagnostic coding during consultations with LBP patients [25]. The International Coding for Primary Care (ICPC-2) diagnostic codes L02, L03, L84, and L86 [28] triggered a pop-up in the medical record system. If a patient met the inclusion criteria, the GP invited the patient to participate in the guideline implementation study. The inclusion criteria were consulting general practice with LBP of any duration for the first time within 3 months, age 18 to 65 years, with or without associated radiculopathy, and a complete STarT questionnaire at baseline. The exclusion criteria were insufficient language skills to fill out questionnaires in Danish, pregnancy, and serious underlying disease (e.g., signs of fracture, osteoporosis, cauda equina syndrome, malignancy, or spinal inflammatory arthritis) [29]. Patients consenting to participate in the guideline implementation study were informed that participation with questionnaires was not a requirement for study participation, but they were encouraged to do so. For this ancillary analysis, we included patients from both the intervention group and the control group who filled in the RMDQ and had a complete STarT questionnaire at baseline. Patients with perfect function (RMDQ = 0) at baseline were excluded.

Patients filled in a questionnaire at home after the initial consultation and were sent follow-up questionnaires after four, eight, and 52 weeks. Patients could choose to complete the questionnaires on the internet or to fill out and return paper versions. Paper versions of the questionnaires were sent to the research unit in a prepaid envelope and the responses were typed into the database by two of the researchers (AR and CEJ). When completing the questionnaires on the internet, the data were directly stored in the project’s database. Every nine STarT items were programmed with a limiter, prompting the patient to respond to all nine items before access to page two of the questionnaire was possible. The 23 RDMQ items were, however, not provided with a limiter. The use of limiters to avoid missing values was not possible in the paper version of the questionnaires, but text was inserted encouraging a reply to all questions. If patients did not respond to a questionnaire, reminders (emails or postal letters) were sent following one- and two-week delays [25]. The database was hosted by an external data manager at the North Denmark Region Department of Information Technology. The project database was provided with access login, written recording and daily backup copying.

The predictor variable was the patient reported STarT risk group (low, medium, or high) at baseline. The primary outcome was assessed by a relative risk combining the low-risk group and the medium-risk group and comparing these to the high-risk group in terms of good outcomes. A good outcome was defined as receiving a minimal clinically relevant improvement in the RMDQ score (0–23 points) after 8 weeks [30]. The outcome was dichotomised using a standard cut-off at 30% improvement [21, 22, 26, 31]. As previous studies also included a secondary cut-off point between the low-risk group and the medium-risk group, this was applied as a secondary analysis. Furthermore, clinically relevant improvements after four and 52 weeks were included as secondary analyses.

For each STarT risk group, baseline characteristics were presented with numbers (%) for categorical variables and mean (sd) or median [iqr] for continuous variables. Baseline characteristics were patients’ age, gender, college education (y/n), employment (y/n), sick leave within 14 days (hours), RMDQ score (0–23 points) [30], numerical pain rating (0–10 points) [32], and self-reported health (EQ VAS, 0–100 points) [33]. Differences in baseline characteristics were tested by Fischer’s Exact test for categorical outcomes (gender, education level, and employment status), with the Student’s t-test (age, numerical pain rating, EQ VAS, and RMDQ), or by the Mann-Whitney test (sick leave) for continuous outcomes. For continuous outcomes, the tests were only comparing the low-risk group with the high-risk group.

For estimating the predictive ability of the STarT, a combination of the low-risk group and the medium-risk group was compared with the high-risk group and the low-risk group was compared with the medium-risk group + the high-risk group by relative risks. A regression analysis was performed to study whether the allocation group in the guideline implementation study was likely to have introduced bias into the estimates. The regression analysis includes baseline RMDQ score, allocation group in the cluster randomised controlled trial together with all the following baseline variables: age (continuous), gender (male/female), college level education (yes/no), employment status (employed, yes/no), sick leave (any LBP-related sick leave 14 days prior to baseline), numerical pain rating (continuous), and EQ VAS (continuous).

The study size was 441 patients by including all patients with a complete STarT from the guideline implementation trial [25]. Single responses from the 23-item RMDQ were coded 0 (no) if they were missing, allowing the inclusion of these observations in the analysis. Patients with a RMDQ score of 0 (optimal function) were excluded from the analysis as they could not achieve a 30% improvement. Throughout the analyses, a P value of <0.05 was considered statistically significant. Analyses were performed using Stata, IC version 14.0 (College Station, Texas, USA).