Efficacy and tolerability of atypical antipsychotics for acute bipolar depression: a network meta-analysis

A systematic literature review was conducted to identify randomized controlled trials of atypical antipsychotic monotherapy in bipolar depression. A protocol was developed for this review and NMA and can be found in Additional file 1, however it was not registered. The amendments to the protocol are listed in Additional file 1 section of protocol. The PRISMA 2020 expanded statement for reporting of systematic reviews incorporating network meta-analyses can be found in the Additional file 1: Table 1 [17]. The most recent bipolar depression NMA identified trials that were completed prior to May 2015 [15]. This update included searches in Embase, MEDLINE, Cochrane Library, and PsycINFO for studies published between May 2015 and 04 May 2020. Eligibility criteria for trial inclusion remained consistent with the previous NMA [15], which was developed using the Patient, Intervention, Comparator, Outcome, and Study type (PICOS) paradigm to minimize bias and identify as many relevant studies as possible [18]. To be included in this NMA, studies must have been double-blinded randomized controlled trials (RCTs) of adults (≥18 years old), with either bipolar I disorder or bipolar II disorder (at least 50% with bipolar I disorder), treated with an atypical antipsychotic as monotherapy, and reported at least one outcome of interest at study endpoint of week 8 or less (Table 1). The exact search terms and resulting number of records returned are reported in Additional file 1: Table 2a. In addition, conference abstracts were reviewed for the 2019–2020 meetings of 11 psychiatry professional organizations to identify secondary publications to supplement already included studies that were published in peer reviewed journals (Additional file 1: Table 2b). A clinical trial registry (https://​www.​clinicaltrials.​gov/​) was searched for additional studies on 04 May 2020 and during the data extraction process. The registry was searched using keywords related to the interventions listed in the PICOS and disease terms related to bipolar I disorder. Study investigators, authors and companies were not contacted for additional data and only the published data was used for this review.

All references/publications were screened based on title and abstract against the inclusion and exclusion criteria by two independent reviewers and discrepancies were resolved by a third reviewer. The full text of publications retained during abstract review were again reviewed by two independent reviewers and similarly discrepancies resolved by a third reviewer. Following best practices, data extraction from selected studies was conducted by two independent researchers with results cross-checked to ensure accuracy.

The primary efficacy outcome was change from baseline in the MADRS total score reported week 8 or before. In addition, change in CGI-BP-S-overall and CGI-BP-S-depression scores, response rate (≥ 50% improvement from baseline in MADRS), and remission rate (MADRS ≤12 and ≤ 10 at endpoint) were also examined. Discontinuation outcomes included all-cause discontinuation, discontinuation due to adverse events, and discontinuations due to lack of efficacy. Metabolic outcomes included change in weight, rate of ≥7% weight gain, changes in triglycerides, total cholesterol, low-density lipoprotein (LDL) cholesterol, glucose, and prolactin. Additional tolerability outcomes included rates of investigator reported somnolence, extrapyramidal symptoms (EPS), akathisia, and switch to mania.

Missing standard errors (SEs) for continuous outcomes were estimated from reported standard deviations, 95% confidence intervals, p values or standard errors (reported for baseline and endpoint values) [18].

The NMA was conducted according to guidance published by National Institute for Health and Care Excellence’s Decision Support Unit [19] and the International Society of Pharmacoeconomics and Outcomes Research Task Force on Indirect Treatment Comparisons [20]. When available, results from mixed models for repeated measures (MMRM) were favored over those using the last observational carried forward method to handle missing data. For trials with multiple fixed dose arms, the results were pooled across dose, which is consistent with methods used in several past meta-analyses [15, 16]. Network diagrams were drawn for each outcome and can be found in Additional file 1: Figure 1a-d.

The NMA was conducted with a Bayesian framework using OpenBUGS v3.2.3 (OpenBUGS Foundation) and R 3.6.1 (R Foundation for Statistical Computing, Vienna, Austria), following codes provided in the NICE Decision Support Unit (DSU) Technical Support Document 2 (TSD2) [21]. The methodology also followed guidance from the ISPOR Task Force on Indirect Treatment Comparisons [20, 22]. In addition, correlations induced by multi-arm trials were taken into account using the methods and codes recommended in the DSU TSD2 [21]. Continuous variables were modeled assuming an identity link and a normal distribution. Dichotomous variables were modeled using a logit link and binomial distribution. Results for continuous variables were reported as the difference in change from baseline and dichotomous variables were reported as odds ratios (OR). All results were reported along with the 95% credible intervals, the Bayesian analogue to confidence intervals. Treatments were ranked using the surface under the curve cumulative ranking (SUCRA) probabilities [23]. This ranking hierarchy was obtained by ordering the effects from the most to least effective (or tolerable) treatment in comparison to placebo. The base case models were fit with random effects models, which estimate additional variance parameters associated with study heterogeneity and generally have larger credible intervals than fixed effects models.

The number needed to treat, and number needed to harm were computed using risk differences derived from the network meta-analysis results following methods described in the Cochrane Handbook for Systematic Reviews of Interventions Version 6.1.0 [18, 24]. NNT and NNH values were calculated by taking the reciprocal of the absolute values of the risk difference [25, 26]. NNT and NNH were rounded up to the next whole number. NNT is a measure of effect size that indicates how many patients would need to be treated with the medication of interest instead of a comparator (i.e., placebo in the case of all trials included in the present study) for a single patient to benefit. Lower NNT values represent superior performance of the treatment of interest on a given outcome. NNH is similar to NNT but measures the number of patients who would likely need to be administered a treatment in order for a single patient to encounter the adverse event. Higher values for NNH represent better performance (i.e., a greater number of patients are likely to be treated before a single patient experiences an adverse event).

We examined the impact of pooled vs. stratified doses of each atypical antipsychotics as well as restricted the data at the 6-week timepoint (i.e., data reported at other timepoints was removed.

The Grading of Recommendation, Assessment, Development and Evaluation (GRADE) approach for NMA was used to evaluate the confidence of evidence [27, 28]. The evidence was evaluated for each direct comparison within each network separately, and since there were no closed loops, an evaluation of indirect comparison was not applicable. Ratings were based on the following domains: study design, risk of bias, inconsistency, indirectness, imprecision, and publication bias [28]. Risk of bias was assessed using the Cochrane Risk of Bias version 2 (RoB2) [24, 29]. Publication bias was assessed by comparison-adjusted funnel plots, with tests for asymmetry applied to cases with > = 10 studies [30].

The transitivity assumption of NMA was evaluated by comparing the distribution of potential effect modifies across clinical trials that were included in the NMA, to ensure that their populations were suitably comparable. The following baseline patient characteristics were examined: age, gender, body weight, BMI, percent with bipolar I disorder, baseline MADRS, baseline CGI-BP-S depression score, and age of bipolar onset. In the NMA random effects models, a common heterogeneity parameter across the various treatment comparisons was assumed, and heterogeneity was assessed by the between-study variance τ² (tau squared) for each outcome, and further characterized by comparing with its predictive distribution [31]. Since there were no closed loops in the NMA, assessment of incoherence was not applicable to this analysis.

The update to the systematic literature review identified 1791 records in EMBASE, MEDLINE, the Cochrane Library, and PsycINFO for screening (Fig. 1). A total of 111 full-text articles and secondary conference abstracts were reviewed with 17 records meeting all the inclusion criteria. In addition, 6 more records were identified from searches of the conference abstracts. These 17 records reflected 10 unique RCTs, 4 of which [32‐35] had not been included in the earlier NMA [15]. Additionally, the results from one RCT that had been included in the earlier NMA based on a conference presentation has since been published [36]. Study results which were published only in conference abstracts and not in a peer-reviewed manuscript were excluded. A total of 18 trials from 25 references were included in the NMA [32‐47]. Figure 1 gives the PRISMA flow diagram showing the reasons for record exclusion.

All studies were multi-site, randomized, double-blind, and placebo-controlled clinical trials. There were no head-to-head trials of atypical antipsychotics. Most trials were multi-national, but eight recruited only from sites in the United States (US) [35, 37, 39, 41‐43] and two recruited only from sites in the People’s Republic of China [36, 46]. Most trials lasted 8 weeks in duration, but six studies [33, 34, 38, 39, 44] lasted only 6 weeks in duration.

Table 2 gives the baseline characteristics used in the assessment of heterogeneity for each of the included RCTs. The study populations were similar in terms of mean age (29.2–43.6 years), gender (34.3–48.1% male) distribution, mean baseline MADRS score (26.9–32.0), and mean baseline CGI-BP-S-overall (4.2–4.5) and depression (4.3–4.9) scores. In the quetiapine trials [36, 37, 40, 41, 43, 47], as few as 50.9% of patients were diagnosed with bipolar I disorder (the remainder were diagnosed with bipolar II disorder), whereas in the other trials 100% of patients were diagnosed with bipolar I disorder. Mean baseline body weight was not reported across all studies, but in 11 studies [32‐34, 36, 38, 40‐42, 46, 47] where it was reported, baseline body weight ranged from 63.9 to 88.8 kg. Mean age of onset was only reported in three trials [32, 38, 44], and ranged from 25.4 to 28.4 years.

Lurasidone, olanzapine, quetiapine, and cariprazine (but not aripiprazole or ziprasidone) were all significantly more efficacious than placebo in change from baseline in MADRS but there was a larger magnitude of change for lurasidone, quetiapine and olanzapine versus placebo (Table 3). According to SUCRA rankings, lurasidone, olanzapine and quetiapine ranked first for change in MADRS score versus placebo followed by cariprazine, ziprasidone and aripiprazole (Table 14). In pairwise comparison for change from baseline in MADRS, lurasidone was similar to cariprazine, olanzapine, and quetiapine, and was significantly better than aripiprazole and ziprasidone (Table 3).

For response (≥50% improvement in MADRS score), lurasidone, quetiapine, olanzapine and cariprazine (but not aripiprazole and ziprasidone) were associated with significantly greater odds of response compared to placebo (Table 3). According to SUCRA rankings lurasidone ranked the first in terms of response followed by quetiapine, olanzapine, cariprazine, aripiprazole and ziprasidone as compared to placebo (Table 14). In pairwise comparison for response lurasidone had significantly greater odds of response than cariprazine, aripiprazole, and ziprasidone (Table 3).

For change in CGI-BP-S-overall score, lurasidone, cariprazine and quetiapine were significantly better than placebo, but olanzapine and ziprasidone were not. According to SUCRA rankings lurasidone ranked first in improving the overall CGI-BP-S score followed by quetiapine, olanzapine, cariprazine and ziprasidone (Table 14). In pairwise comparison for change in CGI-BP-S-overall score, lurasidone was associated with a significantly larger improvement than cariprazine and ziprasidone but showed similar improvements to quetiapine and olanzapine. Quetiapine was significantly better than cariprazine and ziprasidone in improving the overall severity assessed by CGI-BP-S (Table 4).

For studies reporting on the change in CGI-BP-S-depression score, lurasidone was significantly better than placebo, but olanzapine and aripiprazole were not (Table 4). No studies reporting on CGI-BP-S-depression score were identified for cariprazine, quetiapine, and ziprasidone. Lurasidone ranked first in improving the CGI-BP-S-depression score followed by olanzapine and aripiprazole (Table 14).

The odds of remission (MADRS ≤12) were significantly greater for lurasidone, quetiapine, and olanzapine compared to placebo, but not ziprasidone (Table 5). The most effective treatment in improving remission rates as per the SUCRA rankings was lurasidone, followed by quetiapine, olanzapine ad ziprasidone (Table 14). In pairwise comparison for remission, lurasidone had greater odds of remission compared to ziprasidone and quetiapine had greater odds of remission compared to olanzapine as well as ziprasidone. Clinical trials for cariprazine defined remission as MADRS ≤10 at endpoint, and data for this definition was available only for lurasidone. Both lurasidone and cariprazine had significantly greater odds of remission compared to placebo (Table 5); and lurasidone ranked higher than cariprazine according to SUCRA rankings (Table 14).

All-cause discontinuation for lurasidone, olanzapine, cariprazine and, ziprasidone, were comparable to placebo. However, the odds of all-cause discontinuation for aripiprazole were significantly higher than placebo (Table 6). Based on SUCRA ranking, olanzapine and quetiapine ranked the best tolerated treatments in terms of all-cause discontinuation followed by cariprazine, lurasidone, ziprasidone and aripiprazole (Table 14). For discontinuations due to adverse events, lurasidone, cariprazine, olanzapine and ziprasidone had similar odds compared to placebo. Aripiprazole and quetiapine had significantly higher odds of discontinuation due to adverse events compared to placebo. According to SUCRA ranking lurasidone and olanzapine ranked first in terms of discontinuation due to adverse events followed by cariprazine, ziprasidone, aripiprazole and quetiapine. In pairwise comparison, for discontinuations due to adverse events there were no significant differences between the atypical antipsychotics (Table 6). For discontinuation due to lack of efficacy, quetiapine and olanzapine had a significantly lower odds of discontinuation than placebo (Table 7). Based on SUCRA values, quetiapine ranked first in terms of discontinuation due to a lack of efficacy followed by olanzapine, cariprazine, lurasidone, aripiprazole and ziprasidone (Table 14). No significant differences were found for discontinuation due to lack of efficacy between placebo, and lurasidone, cariprazine, aripiprazole, and ziprasidone (Table 7).

All atypical antipsychotics except lurasidone and aripiprazole were associated with significantly more weight gain than placebo. During the short-term trials, olanzapine had the largest mean weight gain relative to placebo (2.88 kg), followed by quetiapine (1.17 kg), cariprazine (0.65 kg), lurasidone (0.34 kg), and aripiprazole (0.20 kg). Olanzapine had significantly greater weight gain than all other antipsychotics and quetiapine had significantly greater weight gain than lurasidone and aripiprazole (Table 8). The SUCRA rankings further confirm that aripiprazole and lurasidone were most likely to be associated with smaller changes in weight from baseline followed by cariprazine, quetiapine and olanzapine (Table 14). All atypical antipsychotics except lurasidone and aripiprazole were associated with significantly greater odds of 7% weight gain versus placebo. In pairwise comparison for clinically significant weight gain (≥7%) olanzapine had significantly greater odds of weight gain compared to quetiapine, cariprazine and aripiprazole (Table 8). Descriptively, the rates of clinically significant weight gain were the lowest for lurasidone (2.4%) followed by cariprazine (3.2%), aripiprazole (4.7%)), quetiapine (6.9%)), and olanzapine (20.3%).

There were no significant differences for change in total cholesterol among the atypical antipsychotics, but olanzapine was associated with greater increases in total cholesterol than placebo (Table 9). There were no significant differences from placebo or between atypical antipsychotics for change in triglycerides (Table 9), LDL (Table 10) or glucose (Table 10), during these acute trials.

Lurasidone was associated with higher changes in prolactin than placebo, aripiprazole, and quetiapine; and cariprazine was also associated with higher changes in prolactin than placebo (Table 11). According to SUCRA ranking for changes in prolactin aripiprazole ranked first followed by quetiapine, cariprazine and lurasidone (Table 14). All antipsychotics except lurasidone, cariprazine, and aripiprazole had greater somnolence than placebo, with quetiapine having greater somnolence than all antipsychotics except ziprasidone (Table 12). On SUCRA analysis, lurasidone ranked the best tolerated option in terms of somnolence followed by cariprazine, aripiprazole, olanzapine, quetiapine and ziprasidone (Table 14). Switch to mania for all antipsychotics was comparable to placebo, except for quetiapine which had a significantly lower odds of switching. Quetiapine also had lower odds of switch to mania compared with aripiprazole (Table 12) and ranked the best according to SUCRA values (Table 14). Rates of EPS were higher than placebo for lurasidone, quetiapine, and cariprazine, but there were no significant differences among the atypical antipsychotics (Table 13). Aripiprazole and cariprazine ranked the best tolerated options in terms of EPS followed by quetiapine and ziprasidone (Table 14). Among the atypical antipsychotics where data on akathisia was reported (aripiprazole, quetiapine, and lurasidone), odds were higher than placebo (Table 13). According to SUCRA rankings, cariprazine and lurasidone ranked the best tolerated options with fewer akathisia rates followed by aripiprazole (Table 14).

Descriptive estimates of the number needed to treat (NNT) to achieve one additional responder relative to placebo and the number needed to harm (NNH) based on one additional all-cause discontinuation and discontinuation due to adverse events relative to placebo were calculated for each treatment arm. Lurasidone (5) had the lowest NNT value for response (highest responder rates) followed by quetiapine (6), olanzapine (10), cariprazine (11), aripiprazole (50) and ziprasidone (100). For remission defined as MADRS ≤12 lurasidone (6) and quetiapine (6) had lowest NNT followed by olanzapine (13) and ziprasidone (250). For remission defined as MADRS ≤10 lurasidone (8) had lower NNT than cariprazine (13). The NNH values for all-cause discontinuations were the highest for quetiapine (500) (lowest discontinuation rate) followed by lurasidone (100), cariprazine (100), olanzapine (15), ziprasidone (15) and aripiprazole (10). The NNH for discontinuation due to adverse events was the highest for lurasidone (250) followed by cariprazine (50), olanzapine (50), ziprasidone (50), aripiprazole (17) and quetiapine (15).

Dose stratified sensitivity analysis results were largely similar to the base case findings. For cariprazine, only the 1.5 mg and 3.0 mg cariprazine treatment arms used once daily were significantly greater than placebo, whereas all the stratified doses for lurasidone, olanzapine, and quetiapine were significantly more efficacious than placebo.

Overall, the quality of evidence for the primary outcome was high for direct evidence with quality reduced for NMA evidence, primarily because of indirectness of results and imprecisions. For all other outcomes, in general, lurasidone vs placebo and cariprazine vs placebo comparisons had higher quality, whereas ziprasidone vs placebo, and aripiprazole vs placebo comparisons had lower quality owing to limitations in the risk of bias. GRADE results are presented in Additional file 1: Table 7a-b.

Overall, the risk of bias was low, with 6 studies showing some concerns related to the randomization process (e.g., incomplete description of allocation concealment). Results of the risk of bias assessment is presented in Additional file 1: Figure 2.

For the primary outcome change from baseline in MADRS, comparison-adjusted funnel plots of the network meta-analysis did not suggest that small studies gave different results from larger studies. This was true for all other outcomes evaluated in this network meta-analysis, except for continuous outcomes change from baseline in CGI-BP-S overall, and triglycerides, and dichotomous outcomes of response, ≥7% weight gain, and EPS. Potential asymmetry was detected in the funnel plots for these 5 outcomes, suggesting a possibility of reporting bias. Funnel plots can be found in the Additional file 1: Figure 3a-d. Assessment of transitivity showed most of the studies and comparisons had minimal variation in mean age, sex, MADRS and CGI-BP-S depression score at baseline results. Other effect modifiers such as age at onset, weight, and BMI at baseline were reported in few studies only. Most of the studies included only BPD-I patients, whereas 7 (39%) studies included both BPD-I and BPD-II patients. Detailed results are presented in Additional file 1: Figure 4a, b. Heterogeneity was assessed by the median between-study variance (Tau²) and ranged from 0 to 12.57, with some considered moderate to high (Additional file 1: Table 8).

There were several limitations to this NMA. The quetiapine trials included bipolar II patients which may confound the results. While the baseline patient characteristics that were examined from the different trials appeared largely similar, unmeasured confounders could exist. Another limitation was inconsistent reporting of outcome variables. Some outcome variables such as EPS symptoms were not reported for all trials or not reported consistently [35, 36, 39, 44‐46]. In addition, the metabolic laboratory values from the different studies were not all specified as fasting measurements. To increase consistency in study design and reported outcomes, this NMA was limited to atypical antipsychotics used as monotherapy to treat patients with bipolar depression. Inclusion of other treatments such as lithium, lamotrigine, and divalproex as well as combination treatments was beyond the scope of the current analysis. Although the included studies were deemed comparable in transitivity assessment, moderate to high heterogeneity was observed for some outcomes in the NMA, compared with the predictive distribution of heterogeneity [31]. Similar to the original NMA, meta-regression, which can potentially adjust for effect modifiers, was not performed because of the absence of a sufficiently large number of trials per comparison that is required to render meta-regression feasible at the aggregate level [15]. Random effects models were applied to account for between-study variance for the NMA, but the presence of heterogeneity should be acknowledged when interpreting the findings. For the included evidence, despite the statistical tests showing no small-study effects for most outcomes, we found some potential asymmetry of comparison-adjusted funnel plots in this network meta-analysis. All the identified trials were placebo controlled RCTs leading to star shaped networks, and therefore indirect evidence and hierarches should be interpreted with caution.