Evaluation of a risk-stratification strategy to improve primary care for low back pain: the MATCH cluster randomized trial protocol

We developed the QI strategy between March 2013 and April 2014 with the support of key GH primary care and PT leaders, several of whom actively served on our project team (ML, DP, PR). To fully benefit from the expertise of the UK group that had developed, tested and implemented the STarT Back approach, we included three key UK members (NF, JH, GS) on our team. We also invited four local individuals with extensive personal experience with chronic pain to serve on our project team. Three are actively involved in helping others with chronic pain through classes and support groups. These “Patient Partners” provided valuable perspectives on the implementation strategy, including identification of ways clinicians could more effectively communicate with patients and outcome measures to use in our evaluation.

To ensure we understood the perspectives of GH patients and clinicians on key issues related to the design and implementation of the STarT Back approach we convened 3 focus groups with a total of 28 patients and 3 focus groups with a total of 15 clinicians (10 PCPs and 5 PTs). To maximize the potential for efficiently implementing our intervention strategy in other U.S. healthcare settings, we recruited a diverse group of national advisors representing patients, large employers, major governmental and independent healthcare systems, primary care practice networks, major government payers and insurers, and institutions that train complementary and alternative medical care (CAM) providers. Most advisers were in positions to influence policies within their organizations.

We used the STarT Back tool without modification, but selected treatment options matched to each risk stratum that were deemed appropriate for a US, and specifically GH, setting. We used GH’s new guidelines for assessing and treating patients with back pain (Additional file 1) to identify the evidence-based treatment options that could be matched to each patient subgroup. These guidelines adapted leading back pain guidelines in the US [26] and UK [27], for use at GH. For example, the US guidelines recommend a multi-disciplinary intensive rehabilitation program as an option for persistent back pain [26], but this was not readily available at GH. However, GH patients do have varying degrees of access to the remaining six treatment options recommended by the US guidelines [26]. CBT (available from GH’s behavioral health service), exercise programs (available from GH PTs), spinal manipulation therapy (chiropractic)/acupuncture/massage (available to most GH members through a contracted network of external providers), and yoga (available in the community, but not covered by insurance). Given the similar clinical effectiveness observed in previous systematic reviews [28, 29], we encouraged clinicians to try to interest patients in active treatments (exercise, yoga) before recommending passive options (chiropractic, massage, acupuncture).

Clinicians in the intervention clinics were trained to routinely administer the STarT Back tool to determine each patient’s risk of persistent disabling pain and to tailor their treatment recommendations to the patient’s risk level [12]. Patients could be offered treatment options appropriate for patients at lower risk levels. The treatments recommended for each risk level were:

Prior to implementation, we met with the primary care and PT teams at the intervention clinics to orient them to the study and explain how it fit into GH’s broader clinical improvement initiative for back pain. Over the next six months (May 2014 – October 2014), we worked with primary care and PT leaders to implement training.

We used the EHR to identify all adult patients (18+ years old) who received “primary” (i.e., first-listed) diagnoses consistent with non-specific back pain (e.g., lumbago, back pain not otherwise specified). To maintain broad applicability of the study population we only excluded patients with specific causes of their pain (such as pregnancy, disc herniation, vertebral fracture, or spinal stenosis) or whose pain resulted from a job injury since these patients are directed to the GH Occupational Medicine clinic. Otherwise, all patients visiting GH PCPs for non-specific back pain were eligible for the study.

Within one week of their visits for back pain, patients were mailed letters informing them that GH was surveying patients in an effort to improve its care for back pain. Patients receiving care from the intervention clinics were not aware that a QI activity was occurring. The letter stated that a research specialist would call them within a few days to invite their participation. A phone number was provided for those not wishing to be contacted. Research specialists called patients between 1 and 3 weeks after their back pain visits to explain the study, answer questions, confirm eligibility and obtain verbal informed consent for participation in a baseline and two follow-up interviews. To maximize participation and follow-up rates, we offered patients $20 for completing each questionnaire. This recruitment process continued until we reached our required sample size for the pre-implementation and the post-implementation periods.

Data were collected by trained interviewers using a computer-assisted telephone interview (CATI) version of the questionnaires to minimize errors and missing data. We tracked the disposition of the consecutive patients seen in the primary care clinics, noting how many were successfully contacted by phone, agreed to participate, and completed each questionnaire.

During both the pre- and post-intervention periods we measured outcomes between about 2 weeks (range: 1 to 3 weeks) after the index clinician visit (baseline) and again two and six months later. We included two primary patient outcome measures. Back-related physical function in the previous week was measured with the modified Roland-Morris Disability Questionnaire (RMDQ) [31]. This instrument asks 23 yes/no questions selected for their relevance for patients with back problems. The RMDQ has been found reliable, valid, and sensitive to clinical changes [32, 33]. We also asked patients to rate their back pain severity during the previous week on a 0-to-10 scale where 0 represents “no pain” and 10 “pain as bad as it could be.” Such numerical rating scales have been found to be valid and reliable measures of pain [34].

We measured the following secondary outcomes:

The primary purpose of this trial is to estimate the effectiveness of the risk stratification strategy for reducing patient dysfunction and pain severity related to back pain. We also are conducting pre-specified subgroup analyses to assess risk-strata-specific differences in physical function and pain severity in the intervention and control group since we hypothesize there will be more benefit in the medium and high risk strata.

To evaluate the overall effectiveness of the proposed risk stratification intervention we estimate the difference in change scores between the control and intervention groups, while accounting for possible effects due solely to the difference in calendar time between the pre- and post-implementation periods (Table 4). Potential effects such as seasonal trends and health care system policy changes will be accounted for in our analyses by using information from concurrent control clinics during the same period of time. The primary analysis time point for the study is 6-months following baseline, though we will also evaluate 2-month changes.

Differences in change scores by intervention group and time period will be estimated using linear mixed models (LMM) [46] with random effects to control for correlation within provider, clinic, or both. Including random effects in the model will yield valid statistical inference while increasing the overall statistical efficiency of the pre- and post-treatment design. The general mean model framework will be the following:

where Y is the change score for the specified follow-up time (t = 2 or 6 months), CompStrat indicates randomization to either a risk stratification or a control clinic, Post is 1 if the patient entered the study during the post-intervention period and 0 if they entered in the pre-intervention period, and \( \overrightarrow{Z} \) is a vector of potential baseline confounders to be controlled for in the analysis. Baseline variables that have previously been shown to be related to dysfunction and pain intensity outcomes include age, gender, RMDQ and pain severity scores, duration of pain, medication use, work-related exertion, overall health status, and psychological issues related to health [47‐49]. To control for any additional confounding, we will also adjust for baseline variables that are associated with the changes in RMDQ and pain intensity and are imbalanced between the intervention groups, or between pre and post time periods.

This model framework allows for the calculation of the confounder-adjusted difference between change scores in the risk stratification and control groups (β ₃ ) taking into account changes due solely to time period (β ₂ ). Table 4 summarizes quantities of interest and their respective parameter estimates from the model. The model presented here is a simplified version of what will actually be used for the final analysis. In the complete model we will include both follow-up times in the same model to more efficiently estimate the effects at each time point (efficiency gained from estimating confounder model parameters just once). We will use the same model framework to calculate the risk-stratum specific estimates and secondary outcomes. For binary secondary outcomes we use generalized linear mixed models (GLMM) [50] with logistic or log link functions to estimate odds ratios or relative risks instead of mean change scores.

We gave respondents $20 per completed follow-up questionnaire to minimize missing data. Further, we are using a model framework of LMM and GLMM models that assume data are missing at random given baseline confounders. We will also conduct sensitivity analyses using imputation to assess the missing data assumptions.

We based our power on detecting clinically meaningful differences of 2.5 points on the RMDQ score and 1.5 points on the pain severity score reported in the literature [31]. We assumed that within the medium- and high- risk groups clinically meaningful differences will be observed in the change scores between intervention and control groups, but within the low- risk group we assumed that there will not be significant differences. We assumed that the baseline distribution of the risk stratification groups will be 40 % low- risk, 40 % medium- risk and 20 % high- risk [51]. Therefore the overall effect difference between the intervention and control groups would be 1.5 points (0*0.4 + 2.5*0.4 + 2.5*0.2) for RMDQ and 0.9 points for pain intensity. These differences in mean change scores correspond to standardized effect sizes of 0.30 and 0.36 for RMDQ and pain severity, respectively. We assume a standard deviation of 5 points for the RMDQ score and 2.5 for pain intensity based on previous studies [52‐54]. For simplicity we have determined the sample size assuming no correlation of outcomes within provider or clinic, yielding conservative estimates of sample size (larger sample size than may be required). Most cluster randomized designs require an increase in sample size relative to independent randomization, but because we have included a cluster with cross-over design (pre versus post) it is a more efficient design than independent randomization. The same model specified in the previous statistical analysis section was used for the sample size and power calculations. For simplification we have set β₀, β₁, and β₂ all to 0 since their values do not affect the power of β₃, the estimate of interest. When estimating β₃ we use the full model including estimates for β₀, β₁, and β₂ to arrive at the correct power estimates. Power calculations assumed two-sided tests at the 0.05 α-level and were performed using R version 2.14.0 for Windows XP Professional.

A total sample size of 1,410 participants provides 80 % power to detect a between-group, overall, standardized effect size of 0.30 on the RMDQ. This sample size provides approximately 92 % power to detect a standardized effect of 0.36 for pain intensity. We assumed that 20 % of participants would drop out of the study and thus inflated the estimated sample size of 1,410 to 1,763. The inflated sample size of 1,763 was divided into four equal groups corresponding to intervention and control clinics in both the pre- and post-intervention. The result was rounded down to a final sample size of 1,760 participants (4 groups of 440). Within risk strata groups we estimate >90 % power to detect an effect size of 0.50 (2.5 points on the RMDQ score) for the medium- and high- risk groups combined, and ~80 % power to detect an effect size of 0.68 (3.4 points on RMDQ) for just the high- risk group. Similarly, for pain intensity, the study will have >95 % power to detect an effect size of 0.60 (1.5 points) for the pooled medium- and high- risk groups, and ~80 % power to detect an effect size of 0.68 (1.7 points) in the high- risk group. Therefore we have adequate power for all comparisons of interest. Additionally, though there may not be sufficient power to detect the noted effect sizes, we will also estimate the effect of the intervention separately in each of the risk strata.