This updated Cochrane Systematic Review of duloxetine for the treatment of chronic pain and fibromyalgia identified 12 more studies than the original review that fitted the predefined quality criteria for inclusion, bringing the number of included studies to 18. These studies covered painful peripheral neuropathy (diabetic), central neuropathic pain, fibromyalgia, and painful physical symptoms in people with major depressive disorder and no underlying explanation for their pain. We excluded 25 studies identified in our searches for various reasons; many concerned pain from other conditions (for example, pelvic pain or osteoarthritis of the knee) that are the topics of other reviews. Some did not fulfil our criteria of treating participants for at least eight weeks (e.g. Brannan 2005), some were open label (e.g. Raskin 2006a and Raskin 2006b) and some not randomised or controlled (Canovas 2007). One trial comparing a novel agent ADL‐5859 to duloxetine and placebo has been completed but remained unpublished at the time of writing; no data were available from the company. Five studies are ongoing (NCT00457730; NCT00619983; NCT01179672; NCT01237587; NCT01552057) and the results of these are likely to become available in due course. Following extensive searches of ClinicalTrials.gov and reference databases, handsearching reference lists, cross‐correlating NCT codes, titles and abstracts, the authors did not identify any further trials. The Lilly Trials Database is freely available and contains extensive details of all trials. Unfortunately, neither the ClinicalTrials.gov NCT number nor the final published title of the research are published in the Lilly database, which make it very difficult to check whether all trials registered on ClinicalTrials.gov have been published. We have no reason to suspect extensive publication bias, certainly in the last few years, but we could not find publications corresponding to five ClinicalTrials.gov entries (Appendix 6).

We analysed the effects of duloxetine on painful diabetic neuropathy, central pain, fibromyalgia, and painful physical symptoms in depression separately. Each has a different pathogenesis and hence analysis of all conditions together would be meaningless. Furthermore, patients and caregivers are more likely to glean benefit from data presented for individual diseases. There is then the potential, should the need arise, to extrapolate such data as a guide to similar diseases where no evidence exists. In meta‐analyses, the magnitude of the benefit in terms of pain relief was similar in all the individual conditions.

For painful diabetic neuropathy, the RR of ≥ 50% reduction in pain at eight to 12 weeks at all doses of duloxetine versus placebo was 1.53 (95% CI 1.21 to 1.92) (NNTB 7 (5 to 10)). For the standard dosage of 60 mg duloxetine daily, this corresponds to an NNTB of 5 (95% CI 4 to 7) (summary of findings Table for the main comparison). For ≥ 30% improvement of pain, the RR at 60 mg was 1.53 (95% CI 1.33 to 1.75), corresponding to an NNTB of 5 (95% CI 4 to 8) (summary of findings Table for the main comparison). These NNTBs are of similar magnitude to the NNTB for tricyclic antidepressants for the achievement of moderate pain relief (NNTB 3.6, 95% CI 3 to 4.5) (Saarto 2007; Kajdasz 2007), and for amitriptyline for 'improvement' (NNTB) of 4.6 (95% CI 3.6 to 6.6) (Moore 2012). The mean difference in pain at a dose of 60 mg duloxetine daily was ‐0.96 points (95% CI ‐1.26 to ‐0.65 points) compared with placebo, on an 11‐point scale. A small dose‐effect may be present, but in most outcome measures, 20 mg to 30 mg is not clearly effective, and 40 mg to 120 mg is effective, with a very small efficacy increment as the dose is increased. There were improvements in the other prespecified secondary outcome measures at 40 mg, 60 mg, and 120 mg daily doses of duloxetine, again with slightly but not significantly greater improvements with the 120 mg dose. The PGIC secondary outcomes were statistically significant, but only border on magnitudes of change that are currently considered to be clinically significant (Dworkin 2008). We discovered no RCTs of the effect of duloxetine on painful diabetic neuropathy for periods longer than 12 weeks.

A number of studies exist presenting the minimal clinically important difference (MCID) for pain in various painful conditions on a 100 mm VAS, which can be roughly translated into an 11‐point VAS. Studies vary in design and inclusion criteria, and the MCID or minimal clinically important change varies with the position on the scale from which participants start. However, for people with pain measured in the middle of the VAS (moderate pain) at baseline, MCID is in the region of two to three points on an 11‐point scale (Mease 2011; Salaffi 2004). Changes of one point or less are unlikely to be clinically relevant. However, as with all pain studies, it is recognised that there is a U‐shaped response curve, with some people responding maximally and some not at all. This makes globalised linear measures of improvement moderately meaningless as the lack of response in some dilutes the effect that occurs in others. The proportion of people responding becomes a more useful measure.

For participants with fibromyalgia, the magnitude of improvement in pain at 12 weeks was similar to that seen in participants with diabetic peripheral neuropathy. There was no clear difference between the effects of duloxetine in fibromyalgia or diabetic peripheral neuropathy when assessed on the primary outcome measures, and so even if fibromyalgia is a "different sort of pain" (Dadabhoy 2006), it seems to respond to duloxetine in a similar way. However, the absolute risk reduction was marginally less than that for diabetic peripheral neuropathic pain and the corresponding NNTB for ≥ 50% or more pain relief at a duloxetine dose of 60 mg daily was 8 (95% CI 5 to 17) (summary of findings Table 2). When we combined results for all doses, the NNTB was 9 (95% CI 7 to 13). Although there were fewer data for fibromyalgia, there again seemed to be a floor effect of dosage, where 30 mg was not effective but higher doses appeared to be. The ceiling effect (or lack of any additional therapeutic effect at 120 mg) was not evident as there were too few data.

It is notable that in fibromyalgia the magnitude of improvement in the SF‐36 mental subscore (MD, 120 mg dose, 4.22, 95% CI 2.43 to 6.02) was double that of peripheral neuropathic pain at the same dose (MD 2.23, 95% CI 0.69 to 3.77) or central neuropathic pain; whereas the magnitude of improvement in bodily pain scores in the neuropathic pain studies (MD 8.19, 95% CI 4.33 to 12.05) was 40% more than that in fibromyalgia (MD 5.96, 95% CI 3.76 to 8.16). As is noted above, the authors of Brecht 2007 comment on the difference in tempo of improvement of depression and pain scores, suggesting that different, but similar, mechanisms are responsible for the two phenomena. In Gaynor 2011a "..patients who met the >30% or >50% BPI response criteria at the 8 week LOCF endpoint had rates of [depression] remission (MADRS total scores…) that were higher compared to duloxetine treated patients who did not meet the BPI response criterion", and in Gaynor 2011b "..remission rates [of depression] were three times greater for patients taking duloxetine who reported at least 30% reduction in pain versus those who did not." However, in Raskin 2005, depression scales did not change despite similar pain scale improvements but the trial specifically excluded people with pre‐existing depression from entry. Russell 2008 used two regression models to separate the direct effect of duloxetine on pain and the indirect effect on the treatment of depression in the treatment of pain. The trial authors calculated that between 21% and 38% of the pain treatment effect was due to treatment of depressive symptoms. Fava 2004 estimated a 50% influence. This remains circumstantial evidence that the mechanisms of pain relief involve both mood and direct pain components and where mood is more involved there may be a greater influence on pain relief.

This new version of the review contains two large, new studies including people with major depressive symptoms and pain of unclear origin. Although these studies were generally well performed, limitations of allocation concealment, blinding and randomisation, along with the potential heterogeneous mix of causes of pain (including various pains of unknown origin), potentially leads to extensive imprecision and a lack of certainty about the results achieved. The mean improvement in pain was only ‐0.55 points on an 11‐point VAS (95% CI ‐0.75 to ‐0.35). The RR of people achieving 50% or more relief of pain at under 12 weeks was 1.37 (95% CI 1.19 to 1.59; NNTB 8, 95% CI 5 to 14) (summary of findings Table 3), once again of a similar magnitude to the pain in fibromyalgia. Unfortunately, there were no data on the SF‐36 to make further comparisons.

This updated review also contains a single study independent from the makers of duloxetine exploring the effect of duloxetine versus placebo in the treatment of central neuropathic pain from spinal cord disorders or stroke. This single, high quality, independent and well‐performed study suffers from its small size and so effect estimates are associated with large CIs. The only result of statistical significance was the PGIC, with a RR favouring duloxetine that was just statistically significant at 2.75 (95% CI 1.02 to 7.44). The magnitude of improvement in pain was also similar to other conditions, although it was nonsignificant, with wide CIs.

A number of studies were not included in the meta‐analysis and are included here for completeness. Kaur 2011 performed a randomised cross‐over trial of amitriptyline and duloxetine. It was not included in a formal meta‐analysis because it was the only trial comparing those interventions. It did not meet our predefined inclusion criteria for length but we included it because it was the only one of its type. It had significant carryover between the cross‐over arms, which greatly affected quality. Duloxetine has a superior response to pregabalin in Tesfaye 2013, but the pregabalin response was at the level of a placebo response in other trials; other studies (for example, Moore 2009) suggest that pregabalin is effective and hence the comparative benefit of duloxetine over placebo remains unclear. Brannan 2005 performed a double‐blind RCT of duloxetine in major depressive disorders, in which they also measured the effect on pain. The outcomes were measured at seven weeks and hence it did not meet our eligibility criteria. Furthermore, the placebo and active arms of the trial were not balanced for depression as measured by the Hamilton Rating Scale for Depression at entry. However, depression scores did not change significantly and similar improvements were seen in pain scores as in the fibromyalgia and painful diabetic neuropathy trials. This compares favourably with the longer studies of Gaynor (Gaynor 2011a; Gaynor 2011b). The Raskin 2006a trial was a 52‐week extension phase of the prior 13‐week study (Raskin 2005) but although this trial was randomised, it was open, with a non‐standardised control group ("standard care"). Hence we excluded it. The only subscale outcomes of the SF‐36 eligible for this review from this trial did not change over one year, but duloxetine was safe and well‐tolerated over that period. Wernicke 2006b was a similar randomised but open extension phase study of Goldstein 2005. Again, safety and tolerability were the main focus. The SF‐36 bodily pain subscore improved significantly but by a much smaller magnitude than in the studies included in the meta‐analysis. Finally, the Goldstein 2004 study reported, with a dose of duloxetine of 80 mg, a median improvement in pain severity on a VAS for pain of ‐7.5% (interquartile range ‐25% to 1%), assessed at week eight. This was the only pain measure to show statistically significant improvement. Since pain was not an inclusion criterion for the trial (designed primarily to look at depression), the trial did not match our pre‐specified inclusion criteria. Hence it was not included in the analysis above but adds support to the therapeutic efficacy of duloxetine in the treatment of pain.

We included TSA in this review for the first time. These analyses are useful in a number of ways. There was not enough information to create a TSA for central pain. For the primary outcome in painful diabetic neuropathy, treatment with the standard dose of duloxetine had convincing evidence of non‐futility even though the optimal information size had not been reached. Within the predefined parameters, even after the first two studies the Z‐score had crossed the boundary line of efficacy. It remains that none of these trials was performed by an independent investigator, but the trials were generally of high to moderate quality and there was no reason to suspect that further independent trials would significantly skew the result to futility. In fibromyalgia, the optimal information size was some way off and further independent trials are recommended. For painful physical symptoms associated with depression, the optimal information size had been exceeded but there were insufficient data to calculate an area of futility. However, the efficacy in this indication is clear. As indicated above, the magnitude of improvement in terms of MD should be considered.

Adverse events were very common in these trials but were, in general, mild. The rates of any adverse event and adverse events leading to cessation of treatment were significantly greater with duloxetine than with placebo at the 60 mg and 120 mg doses, which are the doses used in clinical practice. However, withdrawals because of adverse events were relatively few (duloxetine, all doses combined, 12.6% versus placebo 5.8% (RR 1.99, 95% CI 1.67 to 2.37)). The NNTH for cessation of treatment was 17 (95% CI 13 to 26). These figures are in line with the retrospective analysis of Gahimer et al. in 23,983 patients in the duloxetine integrated exposures database (Gahimer 2007), where approximately 20% of patients withdrew because of adverse events. Adverse event rates were also dose related, being more common with the 120 mg than the 60 mg dose and with the 60 mg than the 20 mg or 30 mg doses, whether any adverse event or events leading to cessation are considered. The rates of serious adverse events were not greater with duloxetine than with placebo (RR 0.81, 95% CI 0.53 to 1.25). The adverse event profile was broadly in line with adverse events in the Cochrane Systematic Review of duloxetine in stress incontinence (Mariappan 2009). From observational studies, the most common side effects of duloxetine are quoted as nausea (37%), dry mouth (32%), dizziness (22%), somnolence (20%), insomnia (20%) and diarrhoea (14%) (Aronson 2007). In the studies included here, sweating (60 mg daily 6.5%, 120 mg daily 9.3%, versus control 0.88%) was also common, occurring at about the same frequency as reported by Gahimer et al. (6.2%) (Gahimer 2007). Tremor was also commonly reported, especially at the 120 mg dose in one trial (60 mg daily 3.3%, 120 mg daily 10.2%) (Russell 2008). There were small changes in blood pressure and heart rate in some but not all of the included studies, although we have not formally analysed these. However, in a review of the trials in depression, duloxetine was reported to produce a small, statistically significant rise in heart rate (2 beats per minute) and a sustained increase in systolic blood pressure versus placebo (1% rise versus 0.4%) (Aronson 2008). A company‐performed retrospective database review of 8504 participants in 42 placebo‐controlled studies of duloxetine covering five indications identified no significant cardiovascular risk (Wernicke 2007). There were no significant increases in suicide or suicidal thoughts in studies where these event were reported.

The studies included in this meta‐analysis contribute more than 6400 participants to meta‐analyses of the use of duloxetine in neuropathic pain, fibromyalgia and painful symptoms in depressive illness, and this is the most comprehensive assessment of any individual antidepressant drug for these indications of which we are aware. This provides a greater level of certainty to the conclusions of this review. Other tricyclic antidepressants have greater efficacy in terms of lower NNTBs than duloxetine, but the trials are small, many have methodological deficiencies and so efficacy may be overestimated. The European Federation of Neurological Societies (EFNS) guidelines for the pharmacological treatment of neuropathic pain recommend duloxetine as second line to tricyclic antidepressants and gabapentin or pregabalin, except where cardiovascular risk factors are present, when duloxetine is favoured (Attal 2006). The National Institue for Health and Clinical Excellence (NICE) recommends duloxetine as first line therapy in diabetic neuropathic pain in non‐specialist care settings but fails to mention it in the treatment of other painful neuropathies (http://publications.nice.org.uk/neuropathic‐pain‐cg96/guidance).

The cost effectiveness of duloxetine has not been formally investigated in RCTs. Four studies have addressed cost effectiveness (Wu 2006; Beard 2008; O'Connor 2008, Bellows 2012); the company manufacturing the drug performed two of them (Wu 2006; Beard 2008). The quality of the earlier studies was questionable but the latest independent study estimates similar levels of economic cost associated with duloxetine use. Wu 2006 performed an analysis of the cost effectiveness of duloxetine in participants completing the Goldstein 2005 trial in the USA. Using trial outcomes, the study authors compared the costs to healthcare, society and employers with those of the 'standard therapies' in the control group. Given that duloxetine was shown to be more effective than standard therapies, duloxetine was also shown to significantly reduce societal and employer costs with a trend towards cost effectiveness in medical costs when compared to standard treatment with other pain management therapies. Beard 2008 used a decision analytic model to represent the sequential management of people with diabetic peripheral neuropathic pain based upon current prescribing practice in the UK. Calculations in the model were based upon data for clinical efficacy that are good for duloxetine and pregabalin, but less robust for tricyclic antidepressants and gabapentin ('first line' drugs). However, recognising the inherent limitations, the authors calculated that duloxetine added as a second line therapy resulted in a predicted cost saving of GBP 77 per patient, on the basis that an additional 29 patients per 1000 achieved a full pain response compared to standard treatment. When duloxetine is added to the standard prescribing hierarchy as a second line therapy, the clinical benefit can also be expressed as adding an additional 0.0019 quality‐adjusted life years (QALYs) per patient. O'Connor 2008 also used a decision analytic model incorporating published and unpublished data from RCTs and cross‐sectional studies of duloxetine, gabapentin, pregabalin and desipramine. The QALY cost of duloxetine using this model varied from USD 47,700 to USD 867,000 per QALY depending upon the assumptions made in the analysis of the trial data. In the latest study, Bellows 2012 used a decision tree model in both clinical trial and real world studies comparing duloxetine to pregabalin. Duloxetine demonstrated an incremental cost of minus USD 187 per patient or incremental effectiveness of 0.011 QALYs. At a cost per QALY threshold of 50,000 USD (almost equivalent to the UK's GBP 30,000), duloxetine is more cost effective than pregabalin. This latter analysis presents the best case scenario, since although the cost effectiveness in people with good pain relief is high, the cost effectiveness can be overwhelmed by those patients in whom a drug does not work, and so cost effectiveness studies of this sort are questionable.