Prediction of pain outcomes in a randomized controlled trial of dose–response of spinal manipulation for the care of chronic low back pain

Data were obtained through a randomized clinical trial looking for the dose effect of spinal manipulative therapy for cLBP [ 3, 6]. Four hundred participants were randomized to 4 dose groups, 100 patients per group. Participants were treated 3 times per week for 6 weeks. They received 0, 1, 2, or 3 SMT sessions per week (0, 6, 12, or 18 total SMT visits) with control visits, consisting of light massage, on non-manipulation visits. All care was provided in the Portland metro area. Outcomes were collected at baseline and 6, 12, 18, 24, 39, and 52 weeks after randomization on. Randomization coincided with the first treatment visit.

All patients provided informed consent for participation in this study which was approved of by University of Western States Institutional Review Board. Patients were recruited through craigslist, mailers, and newspaper advertisements in the Portland Metro area to participate in our study. Details on subject enrollment, inclusion/exclusion criteria, design, and analysis are reported elsewhere [ 3]. Details on cost effectiveness and the doctor/patient interaction have also been reported [ 6, 19].

Patients were considered chronic if their low back pain was at least 3 months in duration and they had at least 30 days of low back pain in the last 6 weeks. A minimum pain score of 25 (0–100 scale) was required in order to prevent floor effects. Participants were excluded for contraindications to SMT, such as active cancer, spine pathology, inflammatory arthropathies, autoimmune disorders, and anti-coagulant conditions. Also excluded were those with potential confounders to pain and disability improvement; including neurodegenerative diseases, pain radiating below the knee, organic referred pain, and disability compensation [ 3].

The intervention was SMT consisting of manual high velocity, low amplitude, thrust spinal manipulation [ 20]. The control appointments consisted of a brief light massage, shorter and lighter than what would be considered appropriate in a therapeutic massage practice [ 21, 22]. In addition to SMT or light massage, at each visit, all participants received 5 min of hot pack treatment to relax the back musculature and 5 min of very low intensity pulsed ultrasound (0.5 watts/cm ²) to enhance credibility of care. This was deemed necessary for fidelity of care in patients receiving only or mostly a brief light massage.

We assessed cLBP with the modified Von Korff pain and disability scales, validated by Underwood et al. [ 23]. The pain scale consisted of asking patients to rate their pain today, average pain over the last four weeks, and worst pain over the past 4 weeks each on a 0–10 pain scale with no pain and as bad as pain could be as the anchors. The disability scale consisted of asking patients to rate how much their low back pain interfered with their daily activities; changed their ability to take part in recreational, social, and family activities; and changed their ability to work (including house and yard work) with no interference/change and extreme interference/change as the anchors. For both pain and disability, the 3 0–10 scales were averaged and then multiplied by 10 to create the two modified Von Korff scores on 0 to 100 scales.

Pain intensity was measured by the Modified Von Korff pain scale described above [ 23]. Pain outcomes were analyzed either as a continuous score (future pain intensity) or as a binary indicator for a responder. Choosing a dichotomous responder indicator is recommended by the NIH Task Force for Research Standards on Chronic Low Back Pain [ 24]. For a given follow-up visit, a responder was defined as having at least 50 % improvement relative to the baseline pain intensity [ 24, 25]. Fifty percent improvement in back pain is considered substantial [ 26]. Note that the same participant could be classified as a responder during one visit and not at another. The pretreatment predictor models were used to predict the outcomes at the 6, 12, 18, 24, 39 and 52-week follow-ups and the post-treatment predictor models were used to predict the outcomes at the 12, 18, 24, 39 and 52-week follow-ups.

A more extensive discussion of the outcomes used in this study is provided in our paper on main results from our dose response study [ 3]. Potential pretreatment predictors were collected at baseline by questionnaire (Table  1). Demographics included age, gender, education, race, and ethnicity. Validated low back pain characteristics were the primary outcomes, pain intensity and disability [ 23], as well as pain unpleasantness [ 27]. Other common condition-specific variables we have used in the past were days with pain, days with disability, duration, and previous treatment [ 28, 29]. General Health status measures were comorbidity, smoking, and the validated EuroQol-5D [ 30– 32] and SF-12 [ 33, 34]. Psychosocial variables were the Fear Avoidance Beliefs Questionnaire [ 35] and previously used confidence in the success of SMT and light massage [ 28, 29]. Dose was determined during randomization at baseline and coded as 0, 1, 2, or 3 representing the multiple of 6 SMT sessions each study group received. Follow-up time point was used as a continuous variable measured in weeks from randomization.

Objective physical exam measures included lumbar range of motion, segmental hypomobility, and pain pressure threshold. Lumbar range of motion was measured in flexion , extension, and lateral bending [ 36– 38]; reliability has been established by Keeley et al. [ 39]. The accompanying self-reported pain was evaluated using the validated 0 to 10 numeric pain rating scale [ 40]. The sum of these pain scores, the difference between right and left lateral bending pain scores, and maximum pain score were also of interest for prediction model development. A modified Schober test [ 41, 42] was conducted on each patient; test reliability has been established [ 43– 45]. Segmental hypomobility from L1 through L5 was determined by manual motion palpation [ 46, 47] and defined for this study as restriction of motion in any plane of motion. The total number of hypomobile joints were also recorded [ 46, 47]. Stochkendahl et al. [ 47] recently observed that global assessment of hypomobility has clinically acceptable reproducibility. Pain pressure thresholds (PPT) were assessed from L1 through S1 using a validated pressure algometer [ 48, 49].

Potential post-treatment predictors were collected several days after the end of treatment (6 weeks after randomization) and included pain intensity, disability, LBP unpleasantness, days with LBP, days with disability due to LBP, satisfaction with care [ 8, 19, 50, 51] and objective physical exam measures, as well as dose and time (follow-up time points listed under outcome measures).

The data set was randomly split into two sets. The training set included all data for 75 % of the participants and the test set included all data for the remaining 25 % of the participants. The multivariate prediction models were developed on the training-set and the ability of the developed models to predict the outcomes was evaluated on the test set. This random split enabled internal validation of the developed model in the same study and addressed the potential issue of over-optimism (over-fitting) of the model in the training set.

For each of the 2 outcomes, we developed a pretreatment and a post-treatment predictor model using the variables identified above. The outcomes were modeled by logistic regression (responder outcome) and linear regression (future pain intensity outcome) using generalized estimating equations (with the AR1 correlation structure) to account for the repeated measures for the same subject [ 52– 54]. Analysis included all time points in aggregate (as repeated measures) to improve the stability and power of the estimated model and to avoid the complexity of reporting and interpreting results from separate predictor models for each time point. A sensitivity analysis (not shown) of models for individual time points showed results that were consistent with and warranting the repeated-measures analysis.

Table  1 shows the descriptive statistics for the baseline variables that were considered for entry into pretreatment predictor models. On average, these patients were 41 years old, equally split in gender, and white non-Hispanic; they had moderate pain (52/100) and disability (45/100), chronic back pain of about 12 years in duration, moderate baseline health, previous care for their back problems, and confidence in success of study care [ 3]. Mean global lumbar range of motion was 43° flexion, 15° extension, and 19° lateral bending. Overall, reported pain with lumbar motion was mild to moderate (3/10). Pain pressure thresholds were reported to be about 6 kg/cm ² from L1 through L5, on average. Table  1 also notes the statistically significant variables in the univariate analyses ( p < .05) that were considered for inclusion into the multivariate models.

There were no missing values in the variables considered in the responder multivariate model and <2 % missing values for variables considered in the multivariate model for future pain intensity. Less than <3 % of values were missing for the remaining variables. In aggregate, only 2.6 % observations had to be dropped due to missing values in variables for building the multivariate model for future pain intensity.

The predictive ability of the pretreatment responder model in the training and test sets were quantified by the AUC statistic and illustrated visually by the receiver operating characteristics (ROC) curves (Table  2 and Fig.  1). While the model appeared to predict some risk of being a responder in the training set (AUC = 0.624), in the test set the model performed similarly to chance only (AUC = 0.479 vs. 0.500 for chance). Among the variables selected into the model, greater odds of 50 % improvement in pain were associated with greater dose of SMT, while poorer odds were associated with comorbidity, less tendency towards self-care, and greater pain with lumbar motion.

The ability to predict future pain intensity was also poor (Table  2 and Fig.  2). The model R ² was 0.268 for the training set and 0.065 for the test set. Hence, the model could account for only a trivial percentage of the variability in future pain. The prediction error (RMSE) was large for both the training and test set data (RMSE = 17.4 and 19.4 points, respectively). Predictors in the model associated with greater future pain were greater baseline pain intensity and unpleasantness, as well as greater pain with lumbar motion; less future pain was related to greater dose of SMT and greater general health evaluated with the EuroQol visual analogue scale.

Patient characteristics evaluated within several days after completion of care (Table  3) show that, on average, patients were in mild to moderate pain (31/100) and disability (28/100). Patients generally agreed or strongly agreed, on average, that they were satisfied with care. Patient perception of provider confidence in treatment was more neutral because the study doctors were trained to show equipoise with regards to study interventions. Lumbar global range of motion increased only about 2°, on average, in each direction. Overall, patients had greater pain pressure thresholds than at baseline.

There were no missing values in most variables considered in the post-treatment multivariate models and for those with missing values 8 % or fewer were missing. In aggregate a total of 13.8 and 9.2 % of observations had to be dropped due to missing values for building the multivariate model for responders and for future pain intensity, respectively.

For responder prediction, AUC was 0.750 in the training set and 0.665 for the test set (Table  4, Fig.  3). Greater odds of 50 % improvement in pain intensity were associated with greater dose of SMT and the passage of time. Poorer odds were associated with greater pain intensity, days with pain, and difference in pain with left and right lateral bending.

The ability to predict future pain in the study population can be considered moderate with an R ² of 0.366 for the training set and 0.261 for the test set. (Table  4, Fig.  4). However, the RMSE was fairly large, 16.3 points in the training set and 17.5 points in the test set. Greater future pain intensity was related to greater 6-week pain intensity and pain in right lateral bending.

In addition, we would like to point out that the 6-week pain intensity predictor alone had better ability to predict future pain score (R ² = 0.350 in the test set) compared to the full multivariate predictor model (R ² = 0.261 in the test set). The RMSE for the 6-week pain intensity only model was 16.5 points for the training set and 17.6 points for the test set, almost identical to the RMSE for the full multivariate model.

The goal in developing the multivariate predictor models was to predict pain outcome for individuals. This is a different objective from the development of a model that seeks to identify specific predictors that impact the outcomes. While our models do identify potential predictors of pain (in an exploratory sense), they cannot reliably claim that variables selected into the models are determinants of pain and variables not selected into the models are not. For example, functional disability might in general be expected to predict pain outcomes. We found that among single predictors other than pain, this variable had the strongest correlation with future pain intensity but it was not included in the multivariate model because of its correlation with the pain predictors that were included. In general, the effort of building a good predictive model does not necessarily lead to reliable identification of specific risk factors that truly impact the outcome. The ability to identify these variables is compromised by the large number of the predictors and by the correlation among them.

A related issue is the use of p-values for selection of the specific variables in our results. The univariate p-values should be interpreted as indicators of variable importance to guide variable selection, not as traditional hypothesis tests of statistical significance. The confidence intervals and p-values for the coefficients and odds ratios in the multivariate regression models must also be viewed as exploratory because both tend to be over-optimistic in forward variable selection. We considered only variables with univariate p < 0.05 in the multivariate analysis (a univariate filter) to reduce this issue at least to some extent. This choice may have theoretically eliminated some true predictors from the multivariate model. However, reducing the potential for over-fitting by considering fewer variables in the multivariate selection process was more desirable in our opinion.

We also would like to emphasize that we did not endeavor to create a clinical prediction rule. We did not have a meaningful clinical comparator such as another efficacious treatment or no treatment at all. We also did not have the power to evaluate potential effect modifiers of intervention including the interaction effect between treatment alternatives and the final prediction-rule recommendation [ 61]. However, the results from our study data do suggest that it may be difficult to develop a prediction rule to identify patients who might benefit from SMT for cLBP before they receive treatment.

The principal limitation was that the study was a randomized controlled trial with treatment mostly limited to SMT of the low back. The baseline models included a rigorously controlled treatment environment for the first six weeks. Even though the post-treatment models are based on the clinical course of cLBP, the influence of other practice characteristics on patient outcomes not included in the prediction models cannot be ruled out, such as full-spine manipulation, physical modalities, supervised exercise, supplements, and advice on posture, diet, stress management, and other lifestyle and self-care measures [ 62]. The observation that decreased future pain is associated with decreased pain with right lateral bending, and not left, is questionable and may simply be an artifact of assessing so many predictors. Pain with some form of lumbar motion does, however, appear in all models. External validation in common clinical practice applied to a variety of patient population is required to evaluate the robustness of applicability of our models.