Unveiling the relative efficacy, safety and tolerability of prophylactic medications for migraine: pairwise and network-meta analysis

The medical literature for relevant studies in PubMed and EMBASE were systematically searched using electronic search strategies, and 1315 records were identified through searching the following key words, for example “migraine”, “topiramate”, “propranolol”, “gabapentin”, “amitriptyline”, “divalproex” and “valproate”. Two additional references were obtained from reviewers. As flow chart Fig. 1 illustrates, 556 duplicated records were identified and removed. Another 486 irrelevant studies were excluded from the remaining 761 records and a final 32 studies were subject to full-text review (Table 1).

Articles were excluded in our study according to the following criteria: (1) the diagnose of migraine was not firmly confirmed in the study; (2) contain treatments that cannot form a closed network; (3) have no comparisons between different treatments; (4) contain outcomes without proper information; (5) does not have any relevant clinical outcomes or treatments; (6) studies without blinding procedures or studies with sample size less than 30; (7) non-randomized clinical trials such as reviews. A study was not considered to be eligible if any of the above criteria was fulfilled.

We selected several clinical outcomes in order to measure the relative efficacy, safety and tolerability of prophylactic migraine medications: monthly migraine headache days, headache frequency, the percentages of patients with at least 50% reductions in migraine attacks (efficacy), the number of patients with all adverse events such as nausea, somnolence or dizziness (safety) and the number of patients who withdrew from studies (tolerability). The following data were extracted from eligible studies and shown in Table 1, including country of study, sample size, histology and clinical outcomes. The corresponding data were extracted into a database after two independent investigators reviewed the manuscripts of all the studies. A Jaded scale table was produced for the purpose of study quality assessment (Additional file 1: Table S1). After data extraction was performed for each study, a network plot with respect to each clinical outcome was produced for demonstrating direct and indirect comparisons (Figs. 2, 3 and 4).

We implemented a two-step approach in our system review and evidence synthesis. Firstly, a conventional pairwise meta-analysis was carried out in order to pool all direct evidence in the current literature. For continuous outcomes such as monthly migraine headache days and headache frequency, raw mean differences (MD) between two groups were synthesized and a summary effect was produced based on raw mean differences, sample size and sample standard deviation. We selected raw mean differences for evidence synthesis since all eligible studies reported the above continuous outcomes in the same scale. On the other hand, the statistic of odds ratio (OR) was pooled from each eligible study and a summary OR was produced for each binary outcome. The pairwise meta-analysis was implemented based on the random-effects model because we did not have full knowledge of study participants or implementation for each eligible study [11].

The second step in our study involves conducting a NMA by synthesize both direct and indirect evidence. Similar to the pairwise meta-analysis, summary mean differences and summary ORs were produced using the Bayesian framework and the Markov Chain Monte Carlo (MCMC) sampling technique. The corresponding ranking of each intervention was obtained by using the surface under the cumulative ranking curve (SUCRA). If an intervention exhibited a higher SUCRA value compared to other interventions, then it is potentially more preferable than others with respect to an endpoint. After that, the assumption of consistency between direct and indirect evidence was assessed by using the node-splitting method whereas publication bias was visually inspected by using funnel plots [12, 13].

The prescribed searching strategy and exclusion criteria enabled us to identify and include 32 studies with a total number of 6052 subjects (Table 1) [14‐45]. All included studies were carried out by using single or double blinding procedures. These studies were carried out between 1986 and 2015 with a maximum following-up duration of 1 year. The majority of the included studies belonged to typical randomized controlled trials (RCTs), however, we identified and included five crossover RCTs in which participants were randomized to receive a sequence of interventions.

A total of ten direct comparisons with respect to each endpoint were produced by using pairwise meta-analysis (Table 2). Patients with topiramate exhibited significantly less average headache days, less headache frequency, a higher likelihood of at least 50% reduction compared to those with placebo (migraine headache days: −0.28, 95% CI = −0.53 to −0.03; headache frequency: −0.31, 95% CI = −0.45 to −0.17; ≥ 50% reduction: OR = 2.33, 95% CI = 1.58–3.42). However, patients with topiramate appeared to have significantly higher risk of all-adverse events and withdrawal due to adverse events compared to those with placebo (all-adverse events: OR = 1.35, 95% CI = 1.06–1.73, withdrawal due to adverse events: OR = 2.08, 95% CI = 1.56–2.78). Patients with propranolol exhibited a significantly less average headache days but higher risk of all-adverse events, somnolence and withdrawal due to adverse events compared to those with placebo (migraine headache days: −0.29, 95% CI = −0.49 to −0.09; all-adverse events: OR = 2.02, 95% CI = 1.05–4.08, somnolence: OR = 4.33, 95% CI = 1.21 to 15.53, withdrawal due to adverse events: OR = 1.87, 95% CI = 1.09 to 3.09). Although there is no significant differences in the average migraine days, headache frequency or the likelihood of at least 50% reduction in headache attacks between patients with gabapentin and those with placebo, gabapentin appeared to be associated with an increased risk of somnolence and dizziness (somnolence: OR = 2.23, 95% CI = 1.11 to 4.46; dizziness: OR = 3.13, 95% CI = 1.73 to 5.56). Patients treated with amitriptyline or divalproex exhibited a reduced headache days or headache frequency as well as a better performance in at least 50% reduction in headache attacks compared to those with placebo (amitriptyline: headache frequency: −0.36, 95% CI = −0.62 to −0.10; ≥ 50% reduction: OR = 1.81, 95% CI = 1.03–3.20; divalproex: migraine headache days: −0.40, 95% CI = −0.61 to −0.18; ≥ 50% reduction: OR = 4.27, 95% CI = 1.30–13.99), however, this was offset by an increased risk of all-adverse events or nausea (amitriptyline: all-adverse events: OR = 2.20, 95% CI = 1.04–4.66; divalproex: nausea: OR = 2.23, 95% CI = 1.21–4.10). Besides that, we were not able to identify any significant results between direct comparisons produced by conventional meta-analysis. Besides, propranolol was safer comparing to topiramate (all-adverse events: OR = 0.57, 95% CI = 0.36–0.90; withdrawal: OR = 0.66, 95% CI = 0.44–0.99; withdrawal due to adverse events: OR = 0.58, 95% CI = 0.37–0.91).

Results produced by NMA are displayed in Table 3 which determined the relative efficacy, safety and tolerability of prophylactic migraine interventions by using both direct and indirect evidence. Patients with three interventions exhibited significantly less average migraine headache days compared with those treated by placebo (topiramate: −1.20, 95% CrI = −1.83 to −0.70; propranolol: −0.98, 95% CrI = −1.86 to −0.07; divalproex: −1.28, 95% CrI = −2.44 to −0.27; Table 3, Fig. 5). Moreover, patients with topiramate and valproate exhibited a significantly increased likelihood of at least 50% reduction in migraine headache attacks compared to those with placebo (topiramate: OR = 4.28, 95% CrI = 1.35 to 14.70; valproate: 11.38, 95% CrI = 1.31 to 111.11; Table 3, Fig. 5). Patients with topiramate or propranolol also exhibited significantly reduced headache frequency compared to those with placebo (topiramate: −1.17, 95% CrI = −1.98 to −0.35; propranolol: −1.37, 95% CrI = −2.49 to −0.29; Table 3, Fig. 5).

Our NMA also provides results for the relative safety of migraine interventions. Patients with Divalproex exhibited significantly higher risk of nausea compared to those with placebo, topiramate, propranolol, gabapentin and amitriptyline (all ORs > 1). Furthermore, patients with amitriptyline exhibited a significantly elevated risk of all-adverse events compared to those with propranolol or placebo (all ORs > 1). However, patients with amitriptyline exhibited a significantly reduced risk of all-adverse events compared to those with divalproex (OR = 0.24, 95% CrI = 0.07 to 0.85). Patients with topiramate also exhibited significantly higher risk of all adverse events compared to those with placebo (OR = 2.44, 95% CrI = 1.55 to 3.88; Table 3, Fig. 6). Our NMA also discovered that patients with GABAPENTIN were associated with a significantly increase in the risk of dizziness in comparison to those received placebo (OR = 3.69, 95% CrI = 1.17 to 9.63; Table 3, Fig. 6). The relative tolerability of various migraine interventions were assessed by using the endpoints of all-cause withdrawal and withdrawal due to adverse events. As suggested by Table 3 and Fig. 7, patients with divalproex exhibited a significantly increased risk of all-case withdrawal compared to those with propranolol or placebo (propranolol: OR = 2.09, 95% CrI = 1.11 to 4.52; placebo: OR = 1.68, 95% CrI = 1.14 to 3.67). On the other hand, patients treated with topiramate or divalproex were associated with an increased risk of withdrawals due to adverse events compared to those with placebo (topiramate: OR = 2.33, 95% CrI = 1.55 to 3.45; divalproex: OR = 2.25, 95% CrI = 1.01 to 5.49).

A ranking plot with respect to each endpoint was produced and a SUCRA table was created in order to differentiate the above migraine interventions (Figs. 8, 9 and 10, Additional file 2: Table S2). Divalproex appeared to have the largest SUCRA value with respect to migraine headache days. Propranolol, topiramate and gabapentin exhibited the largest three SUCRA values with respect to headache frequency. Moreover, valproate, topiramate and divalproex were more preferable than other interventions with respect to the endpoint of at least 50% reduction in migraine attacks. A similar ranking scheme was produced for the above interventions with respect to their safety and tolerability (Additional file 2: Table S2). We also conducted a cluster analysis for grouping the above prophylactic migraine interventions based on the SUCRA values of two endpoints. Overall, propranolol seemed to be the most desirable intervention when several endpoints were simultaneously considered (Fig. 11).

Since we implemented a consistency model when conducting the NMA, the node-splitting method was used to assess the validity of this assumption (Figs. 12, 13 and 14). Net heat plots were also produced by software in order to visualize the consistency pattern existed in each comparison (Figs. 15, 16 and 17). As suggested by both the node splitting method (P-value > 0.05) and net heat plots, there is no significant inconsistency between direct and indirect evidence for the majority of comparisons. Therefore, we concluded that the consistency model is valid in our NMA.

Potential publication bias contained in our NMA was evaluated by using the funnel plots produced by software (Additional file 3: Figures S1, Additional file 4: Figure S2 and Additional file 5: Figure S3). As suggested by the funnel plots, there is no significant asymmetry pattern and most studies were evenly distributed in the funnel plot. Therefore, we concluded that there is no significant publication bias in our study.