Efficacy, acceptability, and safety of antidepressants for low back pain: a systematic review and meta-analysis

The primary outcomes were pain intensity and acceptability. Pain intensity was measured at the follow-up assessment closest to the end of treatment. Acceptability, defined as overall acceptability of the medicine, was measured using all-cause discontinuation during treatment [15, 26].

The secondary outcomes included low back-specific function, symptoms of depression, safety, harm, and tolerability. Low back-specific function and symptoms of depression were measured at the follow-up assessment closest to the end of treatment. Safety and harm, defined as the incidence of adverse effects and serious adverse effects [27], were measured by reports of adverse effects and serious adverse effects during treatment. Tolerability was defined as the tolerability of adverse effects sustained during treatment, measured by reports of discontinued treatment due to adverse effects.

We used comprehensive search strategies to search electronic databases and clinical trial registries for records of randomized clinical trials of antidepressant medicines in LBP (Appendix S1 in Additional file 2) [28, 29]. We piloted the strategies using records of trials included in a previous systematic review [23]. We searched the Cochrane Back and Neck Group’s Trials Register and the Cochrane Central Register of Controlled Trials (CENTRAL) (Cochrane Library), MEDLINE, Embase (Ovid), and CINAHL (EBSCO) databases from inception to May 15, 2020. We searched ClinicalTrials.​gov (ClinicalTrials.​gov), the EU Clinical Trials Register (www.​clinicaltrialsre​gister.​eu), and the WHO International Clinical Trial Registry Platform (apps.who.int/trialsearch/Default.aspx) from inception to May 15, 2020. We included records written in English, Italian, Spanish, Portuguese, German, and French.

We included published and trial registry reports of randomized controlled trials (RCTs) that allocated adult participants with LBP to receive (i) a systemically administered dose of an antidepressant medicine or (ii) a sham (placebo) medicine, (iii) continuation of usual care, (iv) a waiting list, or (v) no-treatment. LBP was defined as pain of any duration between the 12th rib and buttock crease, with or without associated leg pain [30]. Trials that only included participants with symptoms of nerve root compromise (sciatica) [31] or LBP due to specific medical conditions (e.g., spinal fracture, inflammatory disease, aortic dissection, malignancy, or infection) were excluded. We included trials of mixed samples (e.g., non-specific LBP and LBP with sciatica, or non-specific LBP and large joint osteoarthritis) if separate data for the non-specific LBP sample were available. We included trials that tested the efficacy of selective serotonin reuptake inhibitors (SSRIs), serotonin and norepinephrine reuptake inhibitors (SNRIs), tricyclic antidepressants (TCAs), tetracyclic antidepressants (TeCA), heterocyclic antidepressants (HCAs), monoamine oxidase inhibitors (MAOIs), or atypical antidepressants, provided they were listed on the WHO ATC [32] and licensed in at least one of the following jurisdictions: USA (FDA) [33], Australia (TGA) [34], UK (MHRA) [35], or Europe (EMA) [36].

We screened records for inclusion in two stages. Pairs of authors from a team of six (MCF, MAW, AGC, MDJ, HBL, RRNR) independently screened record titles and abstracts in duplicate. The full texts of potentially eligible records were retrieved and independently screened again (MCF, MAW) to confirm inclusion. Disagreements were resolved through discussion or recourse to a third author (MKB or JHM).

We linked records to identify unique studies using a hierarchy. Records that were published and reported the results of a trial were classified as primary records, followed by other published records of a trial (e.g., secondary analyses), conference abstracts, and lastly, trial registry records. We classified the trial registry record as the primary record if there was no evidence of publication.

Pairs of authors (MCF, MAW, AGC, HBL, RRNR, and MDJ) independently extracted data using standardized, piloted, data extraction forms and assessed study-level risk of bias using the Cochrane “Risk of bias” tool (version 5.1.0) [37] and published recommendations [38, 39]. Outcomes were rated as low overall risk when three or fewer domains are rated “unclear” risk, and no domains were rated “high”; moderate risk if a single domain was rated as “high” risk, but four or more were rated as “unclear” and high overall risk in all other instances. We resolved conflicts by consensus or, where necessary, through arbitration with a third author (MKB, JHM). We extracted, for each trial, the following: participant age, sex, duration of symptoms, and sample size; outcome value and measure of variance for pain intensity, function, and symptoms of depression; number of adverse and serious adverse effects; and the number of participants that discontinued treatment for any reason or due to adverse effects. We used an established hierarchy to preference data from continuous measures of pain, function, and symptoms of depression and converted all outcome data to a 0–100-point scale [24]. We used recommended methods [40, 41] to calculate standard deviations when these were not available.

We used the difference in means and accompanying 95% confidence intervals for analyses of effects of antidepressant medicines on continuous outcomes (pain, function, symptoms of depression). We followed recommended guidance for trials with multiple arms by dividing the control group sample size by the number of arms in the study (Cochrane Handbook, Version 6) [42]. For cross-over trials where we were unable to obtain the first phase outcome data from the study authors, we included the overall effect (reflecting both phases) adjusted to correct for the correlation between the two phases [41]. The minimal clinically important difference in means is established as 10 points on a common 0–100-point scale for both pain and function [42]. We used the odds ratio and accompanying 95% confidence intervals for analyses of effects of antidepressant medicines on binary outcomes (acceptability, safety, harm, tolerability).

We constructed extended funnel plots using Stata (version 14.2) [46] to simulate the influence of hypothetical parameters of a future RCT on the pooled effect estimate for pain intensity [47, 48]. The extended funnel plot augments a funnel plot with overlays to provide an illustration of the impact of a new trial on a given meta-analysis [48]. We used 10 points on a 0–100 pain intensity scale as the threshold for the smallest worthwhile effect. We did not perform this analysis for acceptability as there is no known smallest worthwhile effect for this outcome.

Two authors (MCF, MAW) used the Grading of Recommendations Assessment Development and Evaluation (GRADE) [49] framework to develop judgements of high, moderate, low, or very low confidence in the evidence for each outcome. We assessed the domains of study limitations, inconsistency, imprecision, and publication bias, using planned criteria [24]. Publication bias was evaluated using visual assessment of funnel plot symmetry, and Egger’s tests where 10 or more studies were available for an outcome [50].

The search identified 2598 records. We removed 371 duplicates and screened the titles and abstracts of 2227 records for inclusion. We excluded 2104 records and retrieved the full-texts of 123 potentially eligible records (Fig. 1). We excluded 63 records and included 60 records that comprised 23 unique trials [51‐69] (Table 1).

Eighteen trials used a parallel design, and five trials used crossover designs. Four trials were reported in trial registries. We identified a single ongoing trial, a single withdrawn trial, and a single terminated trial. Seventeen trials provided data for inclusion in the meta-analysis. These 17 trials randomized a total 2517 participants to one or more of 11 different antidepressant medicines or placebo. We did not identify any trials of antidepressant medicines compared to waiting list, usual care or no-treatment. The analyses presented below are for the effect of antidepressant medicines compared to placebo.

We assessed completed trials (n = 20) for overall risk of bias (Table S1 in Additional file 2); 15 were assessed as high risk, four at moderate risk, and a single trial at low risk of bias. All twenty trials reported an appropriate method of blinding. Fourteen trials reported either high dropout rates or differences in dropouts between arms. Seven trials reported that they maintained complete control over the publication of results or had no funding-related conflicts of interests.

Visual inspection of funnel plots for each outcome suggested that the effects were evenly distributed around the mean (Figures S1-14 in Additional file 2). For all outcomes, visual inspection of contour-enhanced funnel plots provided no evidence of effects clustered around the threshold for statistical significance. Egger’s tests were conducted for outcomes with 10 studies; only a single study indicated statistically significant asymmetry. A single completed trial report from a trial registry (NCT01225068) was included in our analyses.

The GRADE assessment of confidence in the evidence for each main analysis is presented in Appendix S2 in Additional file 2 and referred to below.

Sixteen of the 23 included trials reported data for pain. We downgraded confidence in the evidence by two levels due to trial limitations. There is low confidence that the pooled effect of antidepressant medicines compared to placebo is − 4.33 [95% CI − 6.15 to − 2.50; Tau² = 2.20] on a 0–100 point scale (Fig. 2).

Fourteen of the 23 included trials reported data for acceptability (all-cause discontinuation). We downgraded confidence in the evidence by two levels due to trial limitations. There is low confidence that the odds of all-cause discontinuation are higher for antidepressants than for placebo: odds ratio 1.27 [95% CI 1.03 to 1.56; Tau² = 0] (Fig. 3).

Six of the 23 included trials reported data for function. We downgraded confidence in the evidence by two levels due to trial limitations. There is low confidence that the pooled effect of antidepressants compared to placebo is − 3.22 [95% CI − 4.96 to − 1.48, Tau² = 0] on a 0–100 point scale (Figure S15 in Additional file 2).

Four of the 23 included trials reported data for symptoms of depression. We downgraded confidence in the evidence by two levels for trial limitations and an additional level for imprecision. There is very low confidence that the pooled effect of antidepressants compared to placebo is − 1.72 [95% CI − 3.88 to 0.44; Tau² = 0] (Figure S16 in Additional file 2) on a 0–100 point scale.

Nine of the 23 included trials reported data for safety (adverse effects). We downgraded confidence in the evidence by two levels for trial limitations. There is low confidence that the odds of experiencing an adverse effect are higher for antidepressants than for placebo: odds ratio 1.58 [95% CI 1.28 to 1.93; Tau² = 0] (Figure S17 in Additional file 2).

Six of the 23 included trials reported data for harm (serious adverse effects). We downgraded confidence in the evidence by two levels for trial limitations and an additional level for imprecision. There is very low confidence that the odds of experiencing a serious adverse effect are higher for antidepressants than for placebo: odds ratio 1.29 [95% CI 0.56 to 2.94; Tau² = 0] (Figure S18 in Additional file 2).

Ten of the 23 included trials reported data for tolerability (discontinuation due to adverse effects). We downgraded confidence in the evidence by two levels for trial limitations. There is low confidence that the odds of discontinuing treatment due to an adverse effect are higher for antidepressants than for placebo: odds ratio 2.39 [95% CI 1.71 to 3.34; Tau² = 0] (Figure S19 in Additional file 2).

We conducted subgroup analyses for pain by antidepressant type and dose to provide additional clinical information (Fig. 4). There were no trials that evaluated the efficacy of HCA or MAOI antidepressants on LBP symptoms. The results for additional subgroup and sensitivity analyses are presented in Supplementary results with corresponding forest plots in Figures S20-23 in Additional file 2.

The extended funnel plots (Figures S24, S25 in Additional file 2) suggest the upper bound of the confidence interval for the pooled effect would cross the threshold for clinical meaningfulness if the meta-analysis included an additional hypothetical trial with approximately 400 participants per arm and an effect for pain of approximately − 30 on a 0–100 scale (antidepressants more favorable than placebo).

Duloxetine is noted in the 2017 American College of Physicians guideline to have small effects on pain and function compared to placebo, for chronic LBP [11]. We repeated the main analyses on five trials that evaluated duloxetine compared to placebo. The effect of duloxetine on pain intensity post-treatment was − 5.87 [95% CI − 7.88 to − 3.86; Tau² = 0] (Figure S26 in Additional file 2). The odds ratio for all-cause discontinuation of duloxetine compared to placebo was 1.17 [95% CI 0.90 to 1.52; Tau² = 0] (Figure S27 in Additional file 2). The odds ratio for experiencing adverse effects of duloxetine compared to placebo was 1.50 [95% CI 1.21 to 1.85; Tau² = 0] (Figure S28 in Additional file 2). The odds ratio for experiencing serious adverse effects of duloxetine compared to placebo was 1.35 [95% CI 0.56 to 3.27; Tau² = 0] (Figure S29 in Additional file 2). The odds ratio for discontinuing treatment due to adverse effects of duloxetine compared to placebo was 2.53 [95% CI 1.70 to 3.77; Tau² = 0] (Figure S30 in Additional file 2).

The REML estimator may underestimate between-study variance for binary outcomes when events are rare [70]. We repeated the analyses for acceptability, safety, harm, and tolerability using DerSimonian-Laird, Paule and Mandel and Mantel-Haenszel methods of estimation (Table S2 in Additional file 2). A single additional post hoc sensitivity analysis is reported in Supplementary Results and Figure S31 in Additional file 2.