Management of Acute Pain From Non–Low Back, Musculoskeletal Injuries: A Systematic Review and Network Meta-analysis of Randomized TrialsFREE

An academic librarian developed search strategies (10) for MEDLINE, EMBASE, CINAHL, PEDro (Physiotherapy Evidence Database), and Cochrane Central Register of Controlled Trials from inception to 2 January 2020. To identify additional eligible studies, we reviewed reference lists from eligible trials and relevant reviews and guidelines.

Eligible studies were parallel-design trials that enrolled adult patients (aged ≥18 years) with acute pain from non–low back–related musculoskeletal injuries (pain duration ≤4 weeks or defined by authors as “acute”); randomly assigned at least 10 patients per study group; and compared currently available pain relief interventions, provided in an outpatient setting, with one another or versus a placebo or sham. Ten pairs of reviewers independently screened titles, abstracts, and full-text articles of potentially eligible studies and resolved disagreement through discussion. We used online systematic review software (DistillerSR [Evidence Partners]) for literature screening.

Seven pairs of reviewers independently extracted data from eligible studies and assessed risk of bias by using a modified Cochrane Risk of Bias Tool (11, 12). Abstracted data included participant and trial characteristics, details of interventions and comparators, and patient-important outcomes (namely pain, physical function, health-related quality of life, patient satisfaction, return to work, proportion of patients with relief, reinjury, and adverse events). For trials with different follow-up lengths, we abstracted data from the longest follow-up for all outcomes except pain, which we abstracted at the 3 most common posttreatment time points reported among eligible trials: 15 minutes to 2 hours, 1 to 7 days, and 3 weeks to 6 months. We contacted study authors for missing or unclear information.

Pooling different instruments that report on a common domain typically is done by converting each instrument to SD units and combining effects across studies as the standardized mean difference; however, this approach has limitations, including difficulties in interpretation and vulnerability to baseline heterogeneity of enrolled patients (13, 14). Therefore, by using linear transformation and assuming that instruments reporting on shared domains have similar measurement properties, we converted all measures of pain intensity and physical functioning to 10-cm visual analogue scales (VASs) (15). We pooled each direct paired comparison of pain or physical function reported by more than 1 study as the weighted mean difference (WMD) and associated 95% CI by using change scores from baseline to the end of follow-up to address interpatient variability. If authors did not report change scores, we estimated them by using the baseline and end-of-study scores and the associated SDs and the median correlation coefficient reported by the trials at lowest risk of bias. To optimize interpretability of our findings for continuous outcomes, we used the network estimate of treatment effects to model the risk difference (RD) for achieving the minimally important difference (MID) (15), the smallest change in a patient-reported outcome that patients perceive as important (16). We used an MID of 1 cm for both the 10-cm VAS for pain (17) and the 10-cm VAS for physical function (18). In our presentation of the results, RDs refer to the modeled estimates in relation to the MID (that is, the proportion of patients who achieve an MID gain in the intervention vs. the control group).

For dichotomous outcomes, we calculated the pooled odds ratio (OR) and modeled RD with corresponding 95% CIs. Initially, we performed a conventional pairwise meta-analysis by using a DerSimonian–Laird random-effects model and then performed a frequentist NMA using the methodology of multivariate meta-analysis assuming a common heterogeneity parameter (19, 20), using the mvmeta command and network suite in Stata (StataCorp) (21, 22). For direct comparisons with 3 or more studies, we performed a sensitivity analysis using a Hartung–Knapp random-effects model for conventional pairwise meta-analysis.

We assessed heterogeneity between RCTs for each direct comparison with visual inspection of forest plots and the I ² statistic (23). Heterogeneity of 0% to 40% was considered as “might not be important,” 30% to 60% as “moderate heterogeneity,” 50% to 90% as “substantial heterogeneity,” and 75% to 100% as “considerable heterogeneity.” The Cochrane Collaboration has proposed overlapping categories to convey that there are no strict cutoffs for interpreting heterogeneity, and categorization depends on the magnitude and direction of effects, as well as the strength of evidence for heterogeneity. When possible, we explored the following a priori hypotheses, determined before analysis through discussions between our study team and a technical expert panel, to explain heterogeneity between trials: First, different clinical conditions will show different treatment effects. Second, more severe injuries will show smaller treatment effects than less severe injuries. Third, older patients will show smaller treatment effects than younger patients. Fourth, longer follow-up will show smaller treatment effects than shorter follow-up. Fifth, higher-dose or higher-intensity treatment will show larger treatment effects. For all direct comparisons, if at least 10 RCTs contributed to a meta-analysis, we assessed small-study effects by using the Harbord test for binary outcomes and Egger test for continuous outcomes (24).

We used the “design-by-treatment” model (global test) to assess the coherence assumption for each network (25). We used the node-splitting method to evaluate local (loop-specific) incoherence (26, 27) in each closed loop of the network separately as the difference between direct and indirect evidence.

After reviewing the literature (28) and consulting with our technical expert panel, we elected to combine all acute musculoskeletal conditions in our analyses. Upon consultation with a clinical pharmacologist, a pharmacist, and the technical expert panel, we included pharmacologic treatments with similar properties and clinical effects in single nodes, which was our primary network of interventions. We explored the appropriateness of these groupings by deriving the statistical consistency of each network and local loops of evidence for each outcome, neither of which showed evidence of incoherence. We conducted a sensitivity analysis considering all interventions as separate nodes, as a secondary network of interventions.

We did not perform NMA if 10 or fewer studies reported an outcome, but we did report a conventional pairwise meta-analysis if 2 or more studies were available. We estimated the ranking probabilities by using the surface under the cumulative ranking curve, mean ranks, and rankograms. We used Stata 15.1 for all analyses.

We used the GRADE (Grading of Recommendations Assessment, Development and Evaluation) approach for rating the certainty of evidence for each network estimate (29–31). Network estimate certainty was based on the source of evidence, direct or indirect, that most contributed to the network estimate. We rated down the certainty for incoherence between the indirect and direct estimates (31, 32), or we used the direct or indirect estimate of effect instead of the network estimate if supported by higher certainty of evidence. We did not rate down the certainty rating of the network estimate twice if both intransitivity and incoherence were present. We evaluated imprecision by using the network estimate; if the 95% CI excluded the null effect, we did not rate down for imprecision unless the comparison was informed by fewer than 300 observations for continuous outcomes or 300 events for binary outcomes.

We categorized interventions from most to least effective on the basis of the treatment effect estimates for benefits and harms obtained from the NMA, as well as their associated certainty of evidence. For each effectiveness outcome, we created groups of interventions as follows: the reference intervention (placebo) and interventions no different from placebo, which we refer to as “among the least effective”; interventions superior to placebo but not superior to other interventions, which we describe as “inferior to the most effective but superior to the least effective” (category 2 interventions); and interventions that proved superior to at least 1 category 2 intervention (which we defined as “among the most effective”). We used the same approach for harm outcomes but designated groups as follows: no more harmful than placebo, less harmful than some alternatives but more harmful than placebo, and among the most harmful. We then categorized interventions as those with moderate- or high-certainty evidence and those with low- or very-low-certainty evidence relative to placebo (33).

The funder had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.

Most trials (76% [158 of 207]) were at high risk of bias for at least 1 domain; 115 (56%) adequately generated their randomization sequence, 150 (73%) concealed allocation, 121 (59%) blinded patients, 116 (56%) blinded health care providers, 128 (62%) blinded data collectors, and 129 (62%) blinded outcome assessors. Forty-three trials (21%) reported 20% or more of the outcome data as missing (Supplement Tables 9 to 11).

Twenty-eight RCTs involving 4464 patients reported pain relief at 2 hours or less (Supplement Figure 2). In 10 of the 25 direct comparisons, 2 or more studies were available for conventional pairwise meta-analysis, in which heterogeneity was substantial (I ² ≥ 54%) in 7 comparisons (Supplement Table 12). We found no evidence of global or loop-specific incoherence (Supplement Figure 3).

Compared with placebo, moderate-certainty evidence showed that topical NSAIDs reduced pain within 2 hours of treatment, with a mean effect approximating the MID (WMD, −1.02 cm [95% CI, −1.64 to −0.39 cm] on a 10-cm VAS for pain; MID of 1 cm; RD for achieving the MID, 23%), as did topical NSAIDs plus menthol gel (WMD, −1.68 cm [CI, −3.09 to −0.27 cm]; RD, 36%), oral NSAIDs (WMD, −0.93 cm [CI, −1.49 to −0.37 cm]; RD, 21%), acetaminophen plus diclofenac (WMD, −1.11 cm [CI, −2.00 to −0.21 cm]; RD, 25%), and acetaminophen alone (WMD, −1.03 cm [CI, −1.82 to −0.24 cm]; RD, 23%) (Figure 1, Appendix Figure 1, and Supplement Table 13).

Low-certainty evidence suggested that compared with placebo, transbuccal fentanyl (WMD, −3.52 cm [CI, −4.99 to −2.04 cm]; RD, 57%) was the most effective treatment in reducing pain within 2 hours. Low-certainty evidence suggested that acetaminophen plus ibuprofen plus codeine (WMD, −1.36 cm [CI, −2.49 to −0.23 cm]; RD, 30%), transcutaneous electrical nerve stimulation (WMD, −1.94 cm [CI, −2.90 to −0.98 cm]; RD, 40%), specific acupressure (WMD, −1.59 cm [CI, −2.52 to −0.66 cm]; RD, 34%), and joint manipulation (WMD, −1.75 cm [CI, −2.68 to −0.81 cm]; RD, 37%) were more effective than placebo but less effective than fentanyl (Figure 1, Appendix Figure 1, and Supplement Table 13).

Pain relief at 1 to 7 days was reported in 69 studies involving 10 829 patients (Supplement Figure 4) and including 33 direct comparisons, among which 11 had 2 or more studies available for conventional pairwise meta-analysis; heterogeneity was substantial in 8 comparisons (I ² ≥ 56%) (Supplement Table 14). The design-by-treatment interaction model showed no evidence of incoherence; however, we observed incoherence in 3 loops of the evidence (Supplement Figure 5), in which the difference between direct and indirect comparison was statistically significant for acetaminophen plus opioids versus oral NSAIDs, acetaminophen plus opioids versus placebo, and oral NSAIDs versus transcutaneous electrical nerve stimulation (Supplement Table 14). We therefore used the direct estimate of effect from conventional meta-analysis, over the network estimate, for acetaminophen plus opioid versus placebo, which showed high-certainty evidence of improved pain relief (WMD, −1.71 cm [CI, −2.97 to −0.46 cm] on a 10-cm VAS for pain; MID of 1 cm; RD for achieving the MID, 19%).

Compared with placebo, moderate-certainty evidence showed that acetaminophen alone (WMD, −1.07 cm [CI, −1.89 to −0.24 cm]; RD, 15%), topical NSAIDs (WMD, −1.08 cm [CI, −1.40 to −0.75 cm]; RD, 15%), and oral NSAIDs (WMD, −0.99 cm [CI, −1.46 to −0.52 cm]; RD, 14%) were among the most effective treatments in reducing pain at 1 to 7 days. Low-certainty evidence showed no difference in pain relief at 1 to 7 days between acetaminophen plus opioids and NSAIDs (Figure 1, Appendix Figure 1, and Supplement Table 15).

Low-certainty evidence suggested that acetaminophen plus chlorzoxazone (WMD, −2.92 cm [CI, −5.41 to −0.43 cm]; RD, 24%), specific acupressure (WMD, −2.09 cm [CI, −3.86 to −0.32 cm]; RD, 21%), and transcutaneous electrical nerve stimulation (WMD, −1.18 cm [CI, −2.09 to −0.28 cm]; RD, 16%) were among the most effective treatments (Figure 1, Appendix Figure 1, and Supplement Table 15).

Thirty studies involving 3549 patients reported physical function (Supplement Figure 6) in 15 direct comparisons, among which 8 had 2 or more studies available for conventional pairwise meta-analysis; heterogeneity was substantial in 3 comparisons (I ² ≥ 70%) (Supplement Table 16). We found no evidence of global or loop-specific incoherence (Supplement Figure 7).

Moderate-certainty evidence showed that compared with placebo, topical NSAIDs (WMD, 1.66 cm [CI, 1.16 to 2.16 cm] on a 10-cm VAS for physical function; MID of 1 cm; RD for achieving the MID, 15%) was the most effective treatment in improving physical function. Oral NSAIDs (WMD, 0.73 cm [CI, 0.17 to 1.30 cm]; RD, 9%) were inferior to topical NSAIDs but superior to placebo. Among the interventions supported by low- or very-low-certainty evidence, only specific acupressure (WMD, 1.51 cm [CI, 0.42 to 2.61 cm]; RD, 14%) improved physical function compared with placebo (Figure 1, Appendix Figure 1, and Supplement Table 17).

Seventeen studies involving 10 390 patients and addressing 15 direct comparisons reported treatment satisfaction (Supplement Figure 8 and Supplement Table 18). In 3 comparisons, 2 or more studies were available for conventional pairwise meta-analysis; the heterogeneity was substantial in 1 comparison (I ² = 89%) (Supplement Table 19). We found no evidence of global or loop-specific incoherence (Supplement Figure 9).

High-certainty evidence showed that compared with placebo, only topical NSAIDs (OR, 5.20 [CI, 2.03 to 13.33]; RD, 34%) demonstrated a statistically significant increase in the likelihood of treatment satisfaction (Figure 1 and Supplement Tables 20 and 21).

Symptom relief was reported in 26 studies involving 4067 patients (Supplement Figure 10 and Supplement Table 22) addressing 21 direct comparisons, among which 5 had 2 or more studies available for conventional pairwise meta-analysis. Heterogeneity was substantial in 1 comparison (I ² = 82%) (Supplement Table 23). We found no evidence of global or loop-specific incoherence (Supplement Figure 11). Compared with placebo, moderate-certainty evidence showed that topical NSAIDs alone (OR, 6.39 [CI, 3.48 to 11.75]; RD, 40%), oral NSAIDs (OR, 3.10 [CI, 1.39 to 6.91]; RD, 17%), and acetaminophen plus diclofenac (OR, 3.72 [CI, 1.02 to 13.52]; RD, 21%) were the most effective treatments for symptom relief (Figure 1 and Supplement Tables 21 and 24).

Low- to very-low-certainty evidence suggested that joint manipulation (OR, 167.71 [CI, 6.67 to 4217.10]; RD, 81%) was the most effective treatment for symptom relief, and that topical NSAIDs combined with menthol gel (OR, 13.34 [CI, 3.30 to 53.92]; RD, 54%), low-level laser therapy (OR, 32.08 [CI, 4.93 to 208.60]; RD, 59%), and mobilization (OR, 7.99 [CI, 1.29 to 49.41]; RD, 21%) were inferior to joint manipulation but superior to placebo (Figure 1 and Supplement Tables 21 and 24).

Gastrointestinal adverse events were reported in 45 studies involving 7070 patients (Supplement Figure 12 and Supplement Table 25) addressing 24 direct comparisons, among which 5 had 2 or more studies available for conventional pairwise meta-analysis; there was no statistical heterogeneity in any of the comparisons (I ² = 0%) (Supplement Table 26). We found no evidence of global or loop-specific incoherence (Supplement Figure 13).

Compared with placebo, moderate-certainty evidence showed that fentanyl (OR, 59.38 [CI, 6.21 to 567.71]; RD, 37%) and acetaminophen plus opioids (OR, 5.63 [CI, 2.84 to 11.16]; RD, 13%) were among the most harmful treatments, increasing the likelihood of gastrointestinal adverse events. Moderate-certainty evidence existed that oral NSAIDs significantly increased the likelihood of gastrointestinal adverse events (OR, 1.77 [CI, 1.33 to 2.35]; RD, 4%) compared with placebo but were less harmful than fentanyl or acetaminophen plus opioids (Figure 1, Appendix Figure 2, and Supplement Table 27).

Neurologic adverse events were reported in 37 studies involving 6245 patients (Supplement Figure 14 and Supplement Table 25). Of the 22 available direct comparisons, 6 had 2 or more studies available for conventional pairwise meta-analysis, in which heterogeneity was low (I ² < 40%) (Supplement Table 28). We found no evidence of global or loop-specific incoherence (Supplement Figure 15).

Relative to placebo, high- to moderate-certainty evidence showed that acetaminophen plus opioids (OR, 3.53 [CI, 1.92 to 6.49]; RD, 16%), fentanyl (OR, 5.73 [CI, 1.20 to 27.47] [moderate certainty]; RD, 22%), and tramadol (OR, 6.72 [CI, 1.24 to 36.39]; RD, 22%) caused greater harm due to neurologic events. Low-certainty evidence suggested that ibuprofen and cyclobenzaprine combination therapy increases the likelihood of neurologic adverse events (OR, 4.91 [CI, 1.45 to 16.61]; RD, 20%) (Figure 1, Appendix Figure 2, and Supplement Table 29).

Thirty-eight studies involving 7235 patients reported dermatologic adverse events (Supplement Figure 16 and Supplement Table 25). Of the 13 available direct comparisons, 4 had 2 or more studies available for conventional pairwise meta-analysis, in which heterogeneity was low (I ² < 40%) (Supplement Table 30). We found no evidence of global or loop-specific incoherence (Supplement Figure 17). Our analysis showed that no intervention caused any statistically significant dermatologic-related harm compared with placebo (Supplement Table 31).

Supplement Tables 32 to 34 and Supplement Figures 18 to 25 provide details of rankings and values for surface under the cumulative ranking curve for all outcomes. We were unable to explore between-study heterogeneity based on clinical condition or injury severity, because most eligible trials enrolled mixed populations. We performed network metaregression to explore the impact of allocation concealment, blinding, missing participant data, and length of follow-up, and found no statistically significant coefficients for any of our outcomes; however, we rated down certainty of evidence for risk of bias (when present) because of concerns that our analyses may have been underpowered to detect associations. We were unable to explore associations between dose or intensity of treatment and age with treatment effects because of limited variability. Supplement Tables 35 to 37 present summaries of findings for comparisons that were not included in any network. Sensitivity analyses using the Hartung–Knapp–Sidik–Jonkman method for pooling showed point estimates of effect consistent with the DerSimonian–Laird method; some associated measures of precision were wider; however, they would not alter our network estimates, because we assumed a common heterogeneity parameter (Supplement Tables 38 and 39).