Weak outcome predictors of multimodal rehabilitation at one-year follow-up in patients with chronic pain—a practice based evidence study from two SQRP centres

Subjects

This study collected data using the SQRP on 464 patients between 18 and 65 years old who had chronic mainly musculoskeletal pain. These patients were referred to the Pain and Rehabilitation Centre at University Hospital of Linköping or the Pain Rehabilitation Clinic at University Hospital of Umeå between 2008 and 2012. In total, 464 patients with chronic pain were referred to these two facilities to participate in a MMRP. The present study is primarily based on patients who answered the SQRP both at baseline (before the MMRP) and at the 12-month follow-up (n = 227).

MMRP at the two sites

At the Umeå site, the referred patients were assessed by teams consisting of a specialist physician in rehabilitation medicine or by specialists in training under supervision of a senior colleague, a physiotherapist, a social worker, and a psychologist and occupational therapist if needed. If the team considered that the patient needed MMRP and fulfilled the inclusion criteria (disabling chronic pain, on sick leave or experiencing major interference in daily life, no further investigations needed) they were accepted to participate in the program. The MMRP was conducted in groups of six to nine participants and was based on CBT principles and included physiotherapy, ergonomics, training in coping strategies, and education in pain management. The patients were encouraged to take an active part in goal setting. At the Linköping site, medical assessments were performed by senior physicians, primarily from specialists in rehabilitation medicine or similar specialities, or by specialists in training under the supervision of a senior colleague. The majority of the participants were also assessed by a psychologist, an occupational therapist, and a physiotherapist. The Linköping program followed the same main principles as the Umeå program—group treatment from a CBT perspective, group physiotherapy, interventions targeting improved ergonomics, and occupational therapy. At both sites the MMRPs lasted between 6 and 8 weeks for at least 20 h per week of group-based activities. In addition, lectures in basic pain science and pain management were offered for both patients (both sites) as well as for relatives, friends, and colleagues (only Linköping). Both programs included work related advice and support, and individually tailored sessions with team members were available. Further individual sessions might also be required for a few weeks following the program.

The following inclusion criteria for the MMRP were used: (i) disabling chronic pain (on sick leave or experiencing major interference in daily life due to chronic pain); (ii) age between 18 and 65 years; (iii) no further medical investigations needed; (iv) written consent to participate and attend to the MMRP;(v) and agreement not to participate in other parallel treatments. The following exclusion criteria were used: (i) ongoing major somatic or psychiatric disease; (ii) a history of significant substance abuse; and (iii) state of acute crisis. General exclusion criteria included severe psychiatric morbidity, abuse of alcohol and/or drugs, diseases that did not allow physical exercise, and specific pain conditions with other treatments options available (red flags).

Methods

The SQRP is recognized as a national registry by the Swedish Association of Local Authorities and Regions. The SQRP is mainly based on questionnaires. The patients completed the SQRP questionnaires on three occasions: 1) at the assessment during the first visit to the clinical department (pre-MMRP or baseline); 2) immediately after MMRP (post-MMRP); and 3) at the 12-month follow-up (FU-12). Demographic data are only collected before MMRP. In the present study, the predictions concerned the FU-12.

The main purpose of the SQRP is to present the results of MMPR at a group level to the participating clinical departments. Based on these data, health care providers and researchers can develop a process that will encourage continued improvement of rehabilitation programs. In a review of the overall quality of approximately 60 nationwide registries, the Boston Consulting Group considered the SQRP as one of the top ten national registries in Sweden [35].

The SQRP includes descriptive variables of the patients’ background, pain characteristics, other symptoms such as depression and anxiety, function, activity/participation, and quality of life. Generally validated Swedish language versions of the instruments were used.

Data from the SQRP

Listed below are the variables and instruments of the SQRP that this study used. These are also summarized in Additional file 1: Table S1.

Statistics

All statistics were performed using the statistical package IBM SPSS Statistics (version 20.0) and SIMCA-P+ (version 13.0; Umetrics Inc., Umeå, Sweden); in all tests, a probability of <0.05 (two-tailed) was accepted as the criteria for significance. In the tables and text, the mean value ± one standard deviation (±1SD) of the investigated variables are given.

Effect sizes of the rehabilitation program were calculated using the template developed by Lakens [52]. In this paper, Cohen’s d_av was seen as the most appropriate way to present the effects as this uses the average standard deviation of the two measurement points [53]. Roughly, one can interpret the Cohen’s effect sizes using the following criteria: d = 0.2 is small, d = 0.5 is medium, and d = 0.8 is large [54]. Bonferroni correction was used in the detailed analyses over time for the different outcomes. Results were deemed significant at the p = 0.00083 level.

Advanced multivariate analyses (MVDA) are essential in PBE studies. Classical statistical methods such as multiple linear regression (MLR) and logistic regression (LR) can quantify the level of relations of individual factors but disregard interrelationships among different factors and thereby ignore system-wide aspects (e.g., when a group of variables correlates with the investigated dependent outcome) [55]. Classical methods assume variable independence when interpreting results [56]. Eriksson and co-workers discussed the risks with considering one separate variable at a time (COST)[57]. They point out that the COST approach has serious drawbacks; the information in multivariate data often remains hidden and there is a risk for spurious results [57]. To handle these drawbacks, we used advanced MVDA (i.e., Principal Component Analysis, PCA) for the multivariate correlation analyses of all investigated variables and Partial Least Square Regression (PLS) for the multivariate regressions using SIMCA-P+. These techniques do not require normal distribution [58].

PCA was used to extract and display systematic variation in a data matrix. All variables were log transformed before the statistical analyses if necessary. A cross validation technique was used to identify nontrivial components (p). Variables loading on the same component p were correlated, and variables with high loadings but with opposing signs were negatively correlated. Variables with high absolute loadings—i.e., 95% jack-knife uncertainty confidence interval non equal to zero—were considered significant. Hence, the most important variables were those with high absolute loadings. The obtained components are per definition not correlated and are arranged in decreasing order with respect to explained variation. R² describes the goodness of fit—the fraction of sum of squares of all the variables explained by a principal component [59]. Q² describes the goodness of prediction—the fraction of the total variation of the variables that can be predicted by a principal component using cross validation methods [59]. Outliers were identified using two methods: 1) score plots in combination with Hotelling’s T² and 2) distance to model in X-space. No extreme outliers were detected.

PLS was used for the multivariate regression analyses [59]. The VIP variable (variable influence on projection) indicates the relevance of each X-variable pooled over all dimensions and Y-variables—the group of variables that best explain Y. VIP ≥ 1.0 was considered significant if VIP had 95% jack-knife uncertainty confidence interval non equal to zero. The PLS analyses were made in two steps. First, all potential regressor variables were entered. If this regression was significant, regressors with VIP > 0.80 were selected and used in the final PLS. Coefficients (PLS scaled and centred regression coefficients, Coeff) were used to note the direction of the relationship (positive or negative). SQRP uses predetermined rules when handling single missing items of a scale or a subscale; details about this is reported in Additional file 1: Table S6. SIMCA-P+, in contrast to traditional statistical packages e.g. SPSS, uses the NIPALS algorithm when compensating for missing data (i.e. in the present study the scales and subscales of a certain subject).

MLR or LR could have been an alternative, but they assume that the regressor (X) variables are independent. If multicollinearity (i.e., high correlations) occurs among the X-variables, the regression coefficients become unstable and their interpretability breaks down. MLR and LR also assumes that a high subject-to-variables ratio is present (e.g., > 5), an assumption not required for PLS. In fact, PLS can handle subject-to-variables ratios < 1. Moreover, PLS, in contrast to MLR and LR, can handle several Y-variables simultaneously.

Drop out analysis

When comparing those who did not and who did complete the questionnaire at FU-12, only one just barely significant difference was found—the pain intensity variable NRS-7days. Those who did not complete the questionnaire had a mean score of 7.02 (SD 1.06) in comparison to those who did (M = 6.61, SD = 1.89) (t(437) = 2.09, p = 0.04). Hence, no statistically significant differences were found for demographic variables, psychological well-being, quality of life, acceptance, kinesiophobia, general health, or psychological functioning as measured by the SF-36.

Background data and characteristics of pain

The number of anatomical regions with pain (PRI) was 14.7 ± 7.9 out of 36 anatomical regions. Mean pain duration was 2550 ± 2609 days; corresponding figures for persistent pain were 1856 ± 2218 days. Health care seeking was prevalent; 58.9% had visited a physician four or more times during the previous year.

Effects of MMRP

Predictions of changes in outcomes at 12-month follow-up

The difference (i.e., changes) between pre and FU-12 values for the selected outcomes were calculated and transformed if necessary so that an improvement had a positive value.

Regression of changes in pain intensity aspects

Regression of changes in emotional functioning

In the regression of changes in three psychological variables (MPI-Distress, HAD-D, and HAD-A) (R² = 0.14, Q² = 0.10), the strongest regressors of the background and baseline variables were HAD-A, HAD-D, MPI-Distress, and SF36-MCS (Table 3, model 1). Hence, improvements in psychological status were associated with high psychological and mental distress at baseline. When the three psychological variables at baseline were omitted as regressors, the explained variation decreased (Table 3, model 2). In addition, low life control (MPI-LifeCon) and low satisfaction with life in general (LISAT-life) to some extent was associated with improvements in the three psychological variables.

Regression of changes in physical functioning

In the regression of changes in physical functioning (i.e., MPI-PainInterfer and SF36-PF; R² = 0.19, Q² = 0.13), the three significant regressors were SF36-PF, RTW-when, and Own prognosis-RTW (Table 3, model 1). Hence, improvements in physical functioning were associated with the following circumstances at baseline: low physical functioning (SF36-PF), perceptions of relatively soon return to work (i.e., RTW-when), and relatively easy to return to work (i.e., Own prognoisis—RTW). When excluding the two physical functioning variables at baseline as regressors, the explained variation decreased (Table 3, model 2). To some extent low physical health (SF36-PSC) at baseline was associated with improvements in physical functioning.

Regression of changes in coping aspects

It was not possible to significantly regress the changes in the four coping aspects (i.e., the 4 Y-variables: MPI-LifeCon, CPAQ-AE, CPAQ-PW, and Tampa) using the background and baseline variables as regressors.

Regression of changes in health aspects

In the significant regression of changes in the health variables (i.e., the two EQ5D variables and LISAT-life; R² = 0.19, Q² = 0.09), the two significant regressors were EQ5D-VAS and EQ5D-index (Table 3, model 1). Hence, improvements in health aspects were associated with low reported health at baseline.

When omitting the health variables at baseline, the explained variation decreased (Table 3, model 2). Short pain durations and low age were then associated with improvements in health aspects.

Retrospective evaluations of pain and in ability to handle life situation in general

At FU-12, 53.6% retrospectively reported improvements for pain and 80.1% for general life situation (Table 2). For Retro-pain, the following background and baseline variables were multivariately important (R² = 0.11, Q² = 0.04): Education (VIP = 1.84 +), MPI-Pain severity (VIP = 1.42-), EQ5D-index (VIP = 1.07+), and Work-importance (VIP = 1.07-). Hence, improvements were associated with relatively high education, low pain intensity (MPI-pain-severity), high health (EQ-5D-index), and perceiving work as important for reasons other than economic (i.e. Work- importance). For Retro-life situation, the important variables at baseline (R² = 0.09, Q² = 0.05) were MPI-Pain severity (VIP = 1.77-), EQ5D-index (VIP = 1.72+), NRS-7days (VIP = 1.25-), and education (VIP = 1.20+). Hence, improvements in life situation were associated with low pain intensities (NRS-7days and MPI-Pain-severity), high health (EQ5D-index), and relatively high education at baseline.

PCA of outcome variables

To understand to what extent the changes in outcomes of the different domains together with the two retrospective items above were independent, a PCA was made for the situation at FU-12. Only one significant component (R² = 0.34; Q² = 0.24) was achieved according to the cross-validation rules in SIMCA-P+. The four most important variables – i.e., highest absolute loadings—were changes in MPI-PainInterfer (loading = 0.33), MPI-LifeCon (loading = 0.31), MPI-Pain severity (loading = 0.30), and MPI-Distress (loading = 0.29). Hence, these four variables showed the most prominent variability across subjects.

Regression of changes in the four MPI variables

When regressing the changes in the four most important MPI variables according to the PCA, the most important background/baseline variables were MPI-LifeCon (VIP = 2.05-), MPI-Distress (VIP = 1.46+), and Work–importance (VIP = 1.08-)(R² = 0.16, Q² = 0.09, n = 227). Hence, improvements in these four MPI variables were associated at baseline with low life control, high psychological distress, and perceiving work as important for reasons other than economic.

Effectiveness of MMRP

This PBE study confirms the generally positive results for MMRPs according to SRs [16–18], health economics studies, and 10-year follow-ups [19–21] (Table 2). Both the repeated measurements (i.e., baseline vs. FU-12) and the retrospective items (Table 2) showed improvements for this cohort of patients.

Retrospective evaluations at FU-12

The majority of patients retrospectively reported improvements according to both pain and in their ability to handle their life situation in general (Table 2). Fortunately and in agreement with other studies [61, 62], relatively small proportions (3–7%) of the patients reported worse situations at FU-12 (Table 2). Qualitative explorations have been suggested as one way to develop interventions for patients who do not benefit from MMRP [63].

A mix of background and baseline factors were important regressors. The two retrospective items had several regressors in common. That is, low pain intensity, high education, and high health perception were related to improvements. Similar results have been reported in other studies using retrospective evaluations [33, 61]. In contrast, a SR of MMRP for fibromyalgia patients summarized that the global treatment effect was predicted by the MPI dysfunctional profile (i.e., high pain, disability, and high solicitous responses from significant others) and/or worse baseline status [28]. Hence, it is presently unclear if a relatively good or bad situation with respect to, for example, pain intensity is associated with improvements when evaluated retrospectively.

Changes in outcomes according to the repeated measurements

In the repeated measurements (i.e., changes from baseline to FU-12), the most prominent effects according to Cohen’s d were for emotional and pain aspects (Table 2). Effect sizes of 0.51–0.61 for the two pain intensity variables at FU-12 were found (Table 2) and in fact both variables showed increasing effect sizes over time (Additional file 1: Tables S2 and S4). These results were in contrast to some SRs, which reported no evidence for efficacy with respect to pain intensity [16, 17]. Many pain patients consider pain intensity as the most important aspect to improve in treatments. Several of the included RCTs in SRs of MMRP do not include pain intensity due to the fact that the interventions are not focussed on the pain itself but rather on its consequences [16, 17]. Some authors have suggested that it might be detrimental for the patient to focus on attempting to control, reduce, or cure pain, since this might shift emphasis away from the aspects that are important for the patient’s health and quality of life, such as daily functioning and emotional well-being [64, 65]. Although these lines of arguments appear reasonable, it may ethically problematic if both clinical practice and research ignore the reports of the patients regarding pain intensity. Moreover, patients with high pain intensities require greater reductions in pain intensity than patients with lower pain intensity levels in order to obtain clinically important improvements [66]. In addition to focusing on pain intensity, it might be wise to focus on the other outcomes emphasised by IMMPACT when communicating with patients. As an outcome of MMRP, pain intensity might be associated with problems when interpreting the results of MMRPs (e.g., a successful intervention leads to less fear of movement) as the patient may be more prone to participate in strenuous activities, which may lead to increased pain intensity in a more active patient. However, changes in pain intensity, psychological distress, and coping aspects were positively intercorrelated in this study, findings that do not support such an argument. It is presently unclear if a more prominent focus on pain intensity within the bio-psycho-social context of MMRP can increase the effect sizes or will actually diminish them.

A general pattern was observed where changes in outcomes were significantly predicted by their respective baseline scores (Table 3, model 1). For example, high pain intensity at baseline was associated with more prominent changes in pain intensity aspects. Such results have been observed in other studies and could be due to the fact that a more severe initial status gives more room for improvement [61]. Another related explanation could be regression to the mean [33]. Recently, Bonstra et al. briefly summarized the literature concerning regressors of repeated measures of outcomes and better outcomes were reported for patients with high scores on depression, high fear avoidance beliefs, high perceived stress, optimistic attitudes, low need to socialize, younger age, or high educational level [61]. Hence, the present results (Table 3, model 1) and a study of chronic widespread pain generally agree with that review [33]. In contrast to this literature, a SR concerning fibromyalgia reported that poor outcomes were associated with depression at baseline [28]. Differences across studies with respect to important predictors may be due to factors such as number of potential prognostic factors, cohort characteristics, chosen outcomes, and/or content of MMRP [18, 29, 67, 68].

For two of the domains of outcomes—pain intensity and physical functioning (Table 3, model 1)—additional background and baseline variables were significant. Hence, the regression of the pain intensity changes was also important for satisfaction with psychological health, working/studying, and relatively high social support at baseline. For changes in physical functioning, additional significant regressors were optimistic attitudes towards RTW (i.e., RTW-when and Own-prognosis–RTW).

Based on the regressions of the changes in the repeated measurements (Table 3, model 1), it seems important to choose patients with a relatively severe clinical picture in order to be able to achieve substantial improvements. It is reasonable to assume that this is relevant only up to a certain level of severity. However, even though it was possible to significantly regress most of the domains of outcomes, the explained variations (R²) were low (15–25%). Similar results have been reported by other researchers [61]. Hence, other—not investigated—factors are important and the present results concerning significant background and baseline variables cannot be applied directly in clinical practice with any precision.

Is high or low pain intensity indicative of positive outcomes?

The importance of the level of pain intensity for positive outcome differed between the repeated measurements and the retrospective evaluations. Improvements in pain intensity aspects according to the repeated measures were associated with high pain intensity at baseline (Table 3, model 1), while in the retrospective evaluations the improvements were associated with low pain intensity at baseline. There are several problems with retrospective items such as recall time, “telescoping”, desirability, and memory aspects [69–71]. When using retrospective items to evaluate changes in a treatment context, additional problems may exist (e.g., overly optimistic assessments) [72]. Hence, the results concerning predictors for the retrospective items may have been associated with bias. On the other hand, repeated evaluations using self-report questionnaires in intervention studies may be problematic [73]. The change that the patient undergoes because of the intervention may affect the interpretations of the questions when presented at follow-ups. MMRPs aim to influence the perception and coping of pain, psychological distress, and various behaviours. Hence, there may be a risk that the attitude towards the items of the questionnaire changed as a consequence of MMRP. This could mean that the follow-up tests cannot simply be compared with the test before MMRP. In order to deepen the knowledge concerning efficacy of MMRPs, objective registrations (e.g., sick-leave registrations, actigraphic recordings, and qualitative interviews [63]) can be important additional sources.

Evaluations of outcomes of MMRP

There is a need to develop clinically applicable and standardized ways to evaluate multiple outcomes of MMPRs in individual patients in trials and in PBE studies. This involves both the statistical methods (cf. Statistics) and strategies for handling multiple outcomes. This study used different approaches for evaluation of outcomes. The number of outcomes in MMRPs are generally high; one SR including 46 RCTs found in median nine outcomes and that these generally were not divided into primary and secondary outcomes [16]. In a recent SR, the outcomes were evaluated separately [74], which may be problematic if the outcomes are intercorrelated and in fact they often are as found in the present study. The present study used 14 outcomes divided into six domains that were mainly recommendations from IMMPACT [12, 28, 34]. Different definitions of a positive outcome of a MMRP trial have been presented such as the majority of outcomes had to be significantly better than for the control intervention [16, 17]. The authors of another SR predetermined primary and secondary outcomes and what was necessary to classify an intervention as positive before reviewing the RCTs [75].

The analysis of several outcomes may raise an issue of multiple comparisons and Bonferroni corrections are frequently applied [76] and recommended in pain trials [77]. However, this is a conservative approach when the number of tests increases [76, 78, 79], and this approach reduces the chances to detect real treatment effects. Furthermore, such corrections were designed for corrections of independent comparisons [78]. The latter is generally not present when evaluating the outcome of MMRPs. IMMPACT has presented hierarchical or “gatekeeping” procedures that do not require adjustment for multiplicity [77] but a natural hierarchy of the outcomes. Outcomes may be combined into a single composite outcome [80], but this may be problematic with respect to missing cases and when the components of the composite endpoint are measured on different scales such as non-commensurate outcomes [80].

Multivariate methods that are able to handle non-commensurate outcomes in one analysis have been presented [80]. The presently applied PLS regression can also handle non-commensurate outcomes in one analysis. We used a more comprehensive approach when regressing four outcomes simultaneously. These four outcomes were chosen based on the advanced multivariate correlation analysis (PCA). In agreement with the regressions of the repeated measures for different domains, we found that a relatively severe clinical presentation at baseline was associated with improvements.

Limitations

There are several limitations to consider. No control group was included and outcomes can be due both to the effects of MMRP and to other factors such as natural course. On the other hand, these patients on average had their pain for a considerable period (on average more than 2500 days), a period that reduces but does not eliminate the risk that effects are related to the natural course. Patients participating in MMRPs have often tried different unimodal treatments with little or no effects and thus represent a selection of patients. The present cohort of patients also represents a selection of patients with the most complicated clinical pictures as they were referred to a university hospital. Hence, the generalizability of the present results is limited and the results of MMRP in primary care may be different. Not all the MMRP participants answered the questionnaire at FU-12. The only significant difference at baseline was a small difference in pain intensity. But bias cannot be ruled out since the non-responders may have had worse effects of MMRP than the responders. It was not possible to analyse our results in relation to the contents and it was not possible to control for small changes in content over time. There are no internationally accepted definitions of MMRP [61, 81] and no taxonomies for the contents of these programs exist [17, 61], both issues that hinder further improvements. The general goals of MMRP may not be relevant to—or seen as relevant by—every patient. In the present study, the importance of the general goals was not investigated and handled in the regressions. Another limitation is that we have no records regarding the level of patient attendance on the programs.