Indirect comparisons of relative efficacy estimates of zuranolone and selective serotonin reuptake inhibitors for postpartum depression

Abstract

Aims

Estimate relative efficacy of zuranolone, a novel oral, Food and Drug Administration-approved treatment for postpartum depression (PPD) in adults vs. selective serotonin reuptake inhibitors (SSRIs) and combination therapies used for PPD in the United States.

Materials and methods

Randomized controlled trials (RCTs) for zuranolone and SSRIs, identified from systematic review, were used to construct evidence networks, linking via common comparator arms. Due to heterogeneity in placebo responses, matching-adjusted indirect comparison (MAIC) was applied, statistically weighting the zuranolone treatment arm of Phase 3 SKYLARK Study (NCT04442503) to the placebo arm of RCTs investigating SSRIs for PPD. MAIC outputs were applied in Bucher indirect treatment comparisons (ITCs) and network meta-analysis (NMA), using Edinburgh Postnatal Depression Scale (EPDS) and 17-item Hamilton Rating Scale for Depression (HAMD-17) change from baseline (CFB) on Days 3, 15, 28 (Month 1), 45, and last observation (Day 45, Week 12/18).

Results

Larger EPDS CFB was observed among zuranolone-treated vs. SSRI-treated patients from Day 15 onward. Zuranolone-treated (vs. SSRI-treated) patients exhibited 4.22-point larger reduction in EPDS by Day 15 (95% confidence interval: −6.16, −2.28) and 7.43-point larger reduction at Day 45 (−9.84, −5.02) with Bucher ITC. NMA showed EPDS reduction for zuranolone was 4.52 (−6.40, −2.65) points larger than SSRIs by Day 15 and 7.16 (−9.47, −4.85) larger at Day 45. Lack of overlap between study populations substantially reduced effective sample size post-matching, making HAMD-17 CFB analysis infeasible.

Limitations

Limited population overlap between SKYLARK Study and RCTs reduced feasibility of undertaking HAMD-17 CFB ITCs and may introduce uncertainty to EPDS CFB ITC results.

Conclusions

Analysis showed zuranolone-treated patients with PPD experienced greater symptom improvement than SSRI-treated patients from Day 15 onward, with largest mean difference at Day 45. Adjusting for differences between placebo arms, zuranolone may be associated with greater PPD symptom improvement (measured by EPDS) vs. SSRIs.

Keywords:

• Indirect treatment comparison (ITC)
• Bucher ITC
• network meta-analysis
• postpartum depression (PPD)
• MAIC
• zuranolone
• SSRIs

JEL CLASSIFICATION CODES:

• I10
• I1
• I
• I11
• I19
• L65
• L6
• L

This article is related to:

The cost-effectiveness of zuranolone versus selective serotonin reuptake inhibitors for the treatment of postpartum depression in the United States

Introduction

Postpartum depression (PPD) is a serious complication associated with pregnancy^1,2. The American Psychiatric Association defines PPD as a major depressive episode with symptom onset during the peripartum period (during pregnancy or in the 4 weeks following childbirth)³, while other organizations extend the time period for symptom onset to 12 months from childbirth⁴. PPD is commonly occurring, though generally under-diagnosed⁵. In an analysis of data from 2018, an estimated 13.2% of women experiencing a live birth in the United States (US) self-reported postpartum depressive symptoms¹. However, among a sample of US women who self-reported postpartum depressive symptoms, only 14% had received a diagnosis⁵.

PPD can be associated with significant impairment in the mother’s daily functioning, quality of life, and ability to connect with and care for their child^6–9. Depression in either parent may increase the likelihood of depressive symptoms in their partner¹⁰, and PPD can be associated with implications for the physical and psychosocial health of the child, including longer-term effects on behavior and risk of mood disorders^11,12. Given these potential implications for the entire family, and that mental health conditions are one of the leading causes of pregnancy-related mortality in the US¹³, the identification of effective treatments for PPD is an important factor in improving outcomes for both patients and families. Additionally, early and effective treatment of PPD is also an important consideration, as delayed onset of treatment efficacy may be particularly detrimental in postpartum patients, whose depressive symptoms occur during an important period for mother-child bonding and who may be managing increased responsibilities due to caring for a newborn¹⁴.

For patients with PPD, pharmacological and non-pharmacological interventions are available. According to 2023 clinical guidance from the American College of Obstetricians and Gynecologists (ACOG), consideration of selective serotonin reuptake inhibitors (SSRIs), serotonin-norepinephrine reuptake inhibitors (SNRIs), brexanolone, and zuranolone use is recommended for PPD treatment and management^15,16. SSRIs and SNRIs are not approved by the US Food and Drug Administration (FDA) for use in PPD, though they are commonly used off-label for PPD treatment, either alone or with psychotherapy, depending on symptom severity^17–19. Two pharmacotherapy treatments have been approved in the US specifically for PPD by the FDA: brexanolone and zuranolone^20,21. Brexanolone is a positive allosteric modulator of gamma-aminobutyric acid type A (GABA_A) receptors and a neuroactive steroid, and it must be administered intravenously over the course of 60 h in a certified Risk Evaluation and Mitigation Strategy (REMS) inpatient setting^22,23.

Zuranolone is the only oral medication approved for adults with PPD as a once-daily, 14-day treatment course in the US. Zuranolone is a positive allosteric modulator of both synaptic and extrasynaptic GABA_A receptors and a neuroactive steroid^20,24,25. The SKYLARK Study (ClinicalTrials.gov Identifier: NCT04442503), a phase 3, double-blind, randomized, placebo-controlled, parallel-group trial, evaluated the efficacy and safety of treatment with zuranolone (50 mg, orally once per day) vs. placebo for a 14-day treatment course in patients with PPD.

The SKYLARK Study found that zuranolone-treated patients with PPD exhibited significantly greater improvement in depressive symptoms than placebo-treated patients, as measured by the primary endpoint of 17-item Hamilton Rating Scale for Depression (HAMD-17)²⁶ total score change from baseline (CFB) at Day 15²⁵. This improvement occurred rapidly, as measured by the key secondary endpoint of HAMD-17 CFB at Day 3 compared to the placebo group, and the improvement was sustained, as shown by the key secondary endpoint of HAMD-17 CFB at Day 45 compared to placebo. The zuranolone-treated group also met additional secondary endpoints assessing patient-reported outcomes using the Edinburgh Postnatal Depression Scale (EPDS) at Days 3, 8, 15, and Day 45²⁵. In the SKYLARK Study, the most common treatment-emergent adverse events (≥5% and greater than in the placebo group) in zuranolone-treated patients were somnolence, dizziness, diarrhea, fatigue, and urinary tract infection²⁷.

An understanding of the comparative effectiveness and safety of different available treatments can support decision-making for treating patients with PPD. The ROBIN (Phase 3 double-blind, randomized, placebo-controlled trial; zuranolone [30 mg] vs. placebo)²⁸ and SKYLARK Studies investigated the efficacy and safety of zuranolone, but direct comparisons (i.e. head-to-head studies) between zuranolone and other treatments used in PPD have not been conducted. In the absence of a direct head-to-head comparative clinical trial, indirect treatment comparisons (ITCs) are a well-supported analytical method^29–31. Application of indirect comparison methodology are needed to inform the relative effectiveness of zuranolone.

The objective of this study was to estimate the relative efficacy of zuranolone (50 mg) with other oral therapies, to compare treatments with similar routes of administration used for the treatment of PPD. This study discusses assessment of the existing evidence for oral pharmacologic treatments for PPD, the construction of evidence networks, and use of these networks to apply unanchored matching-adjusted indirect treatment comparison (MAIC), Bucher ITC, and network meta-analysis (NMA) methods to evaluate the relative efficacy of zuranolone compared to common, off-label, oral treatments for PPD.

Methods

SKYLARK study: investigating zuranolone treatment for PPD

In the Phase 3 SKYLARK Study, patients were randomized to receive either a 14-day course of zuranolone 50 mg (to be taken orally, once per day, in the evening with fat-containing food) or placebo (n = 98 in each group) and were followed through Day 45 (one month after the end of treatment)²⁵. Patients who did not tolerate 50 mg/day of zuranolone had their dose reduced to 40 mg/day²⁵. Patients returned to the study center for in-person outpatient study visits during the treatment period and follow-up²⁵. At baseline, all patients were between the ages of 18 and 45 years, ≤12 months postpartum, and had PPD (HAMD-17 ≥ 26), with symptom onset during the third trimester of their pregnancy or in the first 4 weeks postpartum²⁵.

Outcomes were assessed on Days 3, 8, 15, 21, 28, and 45 in the SKYLARK Study. The SKYLARK Study met the primary endpoint of HAMD-17²⁶ total score CFB at Day 15, with patients who received zuranolone demonstrating significantly greater improvement in PPD symptoms relative to those who received placebo (least squares mean [LSM] difference: −4.0; 95% confidence interval [CI]: −6.3, −1.7; p = 0.001)²⁵. The zuranolone-treated group also met key secondary endpoints of HAMD-17 CFB at Day 3 and Day 45 compared to placebo (Day 3: LSM difference: −3.4; 95% CI: −5.4, −1.4; p = 0.001; Day 45: LSM difference: −3.5; 95% CI: −6.0, −1.0; p = 0.007)²⁵.

Additional secondary endpoints, which were not adjusted for multiplicity and were reported with nominal p-values, assessed patient-reported outcomes using the EPDS²⁵. The group treated with zuranolone exhibited greater EPDS CFB compared to placebo at Day 3 (LSM difference: −1.5; 95% CI: −2.9, −0.1; nominal p = 0.03), Day 8 (LSM difference: −2.2; 95% CI: −3.8, −0.5; nominal p = 0.01), Day 15 (LSM difference: −2.0, 95% CI: −3.8, −0.1; nominal p = 0.04), and at Day 45, which was approximately 1 month after completing the zuranolone treatment course (LSM difference: −2.4; 95% CI: −4.5, −0.3; nominal p = 0.03)²⁵.

The indirect treatment comparison analyses described below were conducted using HAMD-17 CFB and EPDS CFB data from Days 3, 15, 28, and 45²⁵.

Identification of relevant studies for indirect treatment comparison

To conduct ITCs between zuranolone and other oral treatments for PPD, relevant trials were identified systematically. Following the literature review described in Cooper et al.³², the studies in a 2017 systematic literature review (SLR)³³ and its subsequent update in 2021 (unpublished observations; data on file) were included (48 studies total). Reinhart et al.’s 2017 SLR identified literature from Medline, Embase, PsychInfo, Cochrane Central Register of Controlled Trials, Cochrane Database of Systematic Reviews, and several conferences³³. Their review searched for "postpartum depression", "postnatal depression", and "peripartum depression" in observational or interventional studies among women with PPD age ≥15 who received any pharmacologic therapy³³. This SLR and its subsequent update formed the basis of this study’s literature review; one additional study was identified using a published SLR from 2022³⁴ and two studies from 2023 were also identified, including the SKYLARK Study^25,35.

In total, 51 studies were considered for inclusion (Supplementary Table 1), of which seven studies were ultimately included for comparison. Sixteen of the initial 51 studies were not randomized controlled trials (RCTs) and were thus excluded from further analyses (eight were observational/cohort studies, seven were meta-analyses, one was an open-label pilot study). Of the remaining 35 studies, 22 were excluded because they either investigated only non-pharmacologic treatments, reported no outcomes of interest (i.e. did not report HAMD-17 or EPDS scores as outcomes), or investigated treatments for PPD that were not treatments of interest for these ITCs (Figure 1). Though some other measures of depressive symptoms can be converted into HAMD-17 scores, we omitted any studies that did not directly report HAMD-17 or EPDS outcomes in order to avoid introducing uncertainty by converting between metrics. At this step, treatments that did not involve oral therapies for PPD, such as brexanolone (intravenous [IV] administration)^22,23,36,37 and hormone treatment (transdermal administration)³⁸, were excluded. No eligible RCTs identified in the systematic review investigated SNRIs.

Figure 1. Evidence flowchart following exclusion of non-RCTs. Abbreviations. ITC, indirect treatment comparison; non-PTs, non-pharmacologic therapies; PPD, postpartum depression; PTs, pharmacologic therapies.

Of the remaining 13 studies, six were excluded due to insufficient evidence, a lack of a group to connect the study to the network, investigation of treatments unrelated to the treatments of interest (Saffron³⁹) inconsistencies in data reporting (excluded after unsuccessful attempts to contact the author teams to clarify data), or treatment (zuranolone 30 mg²⁸) outside of the expected prescribing dose in the US. seven studies (including the SKYLARK Study) were included in the final network. Details of the treatments assessed in the included studies are available in Supplementary Table 2.

The outcomes most commonly reported in RCTs were HAMD-17 and EPDS scores over time; as such, HAMD-17 and EPDS scores were used as patient outcomes of interest, such that the HAMD-17 CFB or EPDS CFB could be determined and then compared across studies. HAMD-17 is a common measure in studies investigating depression⁴⁰, and EPDS is often used as a screening metric and assessment of symptoms in practice¹⁷. Since both metrics are commonly used, both outcomes were considered for potential ITCs and were included in the construction of evidence networks. This analysis also considered these two outcomes because matching methodologies, such as those discussed below, may reduce the effective sample size of the matched study population below a robust sample size for analysis. Constructing an evidence network for each outcome therefore provided an additional avenue for analysis in the event that one outcome did not have sufficient overlap between study populations to complete the indirect comparisons.

Evaluation of included studies

After the selection of the seven studies included in the network, the treatments administered were grouped into categories: placebo, SSRIs, and combination therapy (SSRIs + non-pharmacologic interventions). Of note, the construction of this network required assumed equivalence of placebo arms across studies. Arms that included nonpharmacologic interventions such as counseling/listening sessions or clinical management and mothercrafting, alone or in addition to placebo, were assumed to be comparable within the network, since the impact of these interventions on the relative effects within studies was assumed to be low (see Supplementary Table 2 for further details)^41–43. Such arms are referred to as “placebo groups” throughout this manuscript. Similarly, all SSRIs were grouped and all combination therapies were grouped, assuming that all members of each class were equally effective. One of the studies in the evidence base⁴¹ allowed patients in the treatment group to receive any of a range of pharmacotherapies, and so comparisons between zuranolone and each specific SSRI were not conducted.

Studies were assessed for homogeneity and heterogeneity of outcomes, timepoints, method of diagnosis, key baseline characteristics and inclusion/exclusion criteria. The baseline characteristics that were considered included age, race, relationship status, employment, time since childbirth and primipara, comorbidities, previous mental health episodes, concomitant medications, depressive symptom severity, and suicidal ideation (Supplementary Tables 3 and 4). Generally, differences in baseline characteristics across studies were not prominent enough or did not provide enough data to indicate clear potential for outcome bias, or to warrant additional adjustments (Supplementary Tables 3 and 4). Methods related to differences in observation timepoints are discussed below.

Differences across studies in the timepoints for which data were available were reported. Linear interpolation was used to estimate treatment effects at the timepoints not reported in the comparator studies³². This method was applied for both HAMD-17 CFB and EPDS CFB results to infer data at any missing time points. Analyses were conducted on Day 3, Day 15, Day 28 (assumed to be equivalent to Week 4 and Month 1), and Day 45, using linear interpolation where data were not reported. Data at the last observation point (Week 12/18 for SSRIs; Day 45 for zuranolone) of each study were also compared.

Because available HAMD-17 and EPDS data from the studies are often not reported as CFB, but rather as mean ± standard deviation (SD) for each timepoint, the mean CFB was calculated using the difference between the observed score at each timepoint and baseline score. The CFB variance was calculated using the within-patient correlation from the SKYLARK Study to impute all variances, in the absence of within-patient correlation data from studies in the network³². The imputed standard errors did account for the differences in sample sizes across the studies. CFB is continuous and was analyzed using a normal likelihood.

Matching-adjusted indirect comparison

Rationale and comparator study selection

Additional analyses were undertaken of HAMD-17 CFB and EPDS CFB in the placebo arms of the included studies, noting that a greater improvement was observed in the placebo arm of the SKYLARK Study (Figure 2(A,B)). The placebo response in the SKYLARK Study may have been prominent due to the high frequency of in-person visits in the clinical trial⁴⁴ or due to expectations^45–47 regarding the efficacy or speed of onset of zuranolone. As such, unanchored MAIC analyses were conducted^48,49. For the MAICs, the placebo arm of the SKYLARK Study was omitted, and the zuranolone treatment arm was matched to the placebo arm of a comparator study, to provide an assessment of the relative effect of zuranolone vs. placebo.

Figure 2. Plots of mean HAMD-17 CFB and EPDS CFB reported in study placebo arms. (A) HAMD-17 CFB^{25,42,43,51,58}. (B) EPDS CFB^25,41,42. Abbreviations. CFB, change from baseline; HAMD-17, Hamilton Rating Scale for Depression; EPDS, Edinburgh Postnatal Depression Scale. Note. The plot makes the underlying assumption of a linear change in HAMD-17/EPDS CFB between observed timepoints.

Treatment effects were quantified using two metrics of PPD symptoms, HAMD-17²⁶ CFB and EPDS^4,50 CFB. Matching was conducted separately for the evidence networks reporting each outcome (HAMD-17 CFB and EPDS CFB), since different studies were included in each network, and the optimal RCT for matching that reported one outcome may not necessarily be present in the other network.

To assess which studies were most suitable for matching, the sample size and the number of baseline characteristics, which could potentially be used as variables for matching, were considered. Yonkers et al.⁵¹ was deemed most suitable for matching on HAMD-17 CFB data, based on its sample size, the nature of its placebo arm, and the potential to match the patient populations by age, race, suicidal ideation, and HAMD-17 score at baseline (Supplementary Table 5). A separate assessment was conducted for matching in an analysis of EPDS CFB data. Sharp et al.⁴¹ was determined to be most suitable for matching, due to the potential to match on baseline characteristics such as age, race, employment, suicidal ideation, baseline EPDS score, and marital status (Supplementary Table 6). Additionally, Sharp et al.⁴¹ assessed a much larger sample than the other study which reported EPDS outcomes⁴².

MAIC methodology

Matching was conducted by reweighting the zuranolone treatment arm of the SKYLARK Study using the individual patient data (IPD) available for the SKYLARK Study and the reported aggregate patient and trial characteristics from the comparator study, which could act as treatment-effect modifiers or key clinical prognostic factors. Potential prognostic factors were evaluated based on data availability in the comparator study. MAICs were conducted according to the National Institute for Health and Care Excellence (NICE) Decision Support Unit (DSU) Technical Support Document (TSD) 2018 guidance⁴⁸.

Statistical weights were applied to patients from the zuranolone treatment arm of the SKYLARK Study; these weights adjusted for any difference in representation of an individual patient relative to the comparator study. Average baseline characteristics were then balanced between the arms. Propensity score weighting, proposed by Signorovitch et al.^49,52, was used to develop statistical weights, and a logistic regression model was used to approximate the likelihood of being enrolled in either arm³². Since IPD for the comparator study was not available, a method of moments was used to develop the model, using the IPD for zuranolone-treated patients and published summary data for comparator studies⁴⁸.

Rescaled weights were assessed to determine how specific patients or patient groups were over- or under-represented in the analysis, and the effective sample size was approximated for the matched analysis. The effective sample size represents the sample size of non-weighted individuals that would provide the same level of precision as that of the weighted sample; a small effective sample size suggests that the matched populations have a low degree of overlap⁴⁸.

Patient outcomes were compared across groups using weighted data: for HAMD-17 CFB and EPDS CFB, patient data from the zuranolone-treated group were weighted to produce a weighted mean CFB. The weighted mean CFB for the zuranolone treatment arm and the mean CFB of the selected placebo arm were then used to calculate the mean difference in CFB for use in the ITC analyses. Estimated standard errors and CIs for the weighted treatment effects of zuranolone were generated as previously described^32,53,54: bootstrap datasets were constructed by sampling patients from the zuranolone treatment arm with replacement, and these datasets were used to develop weights (using Signorovitch methodology) and a weighted mean CFB. A distribution of means was generated by repeating this protocol, and this distribution was used to estimate the standard error and CIs³².

The MAIC results were incorporated into the final network such that the link for zuranolone into the network was based on the matched comparison between zuranolone and the selected placebo arm, allowing ITCs to be conducted.

Indirect treatment comparisons (ITCs)

Bucher ITCs

Bucher comparisons can be used to evaluate two treatments via a common comparator⁵⁵. The rationale for these comparisons is presented in Figure 3: in brief, to compare Treatments A and B, studies comparing (1) Treatments A and C and (2) Treatments B and C can be used as links to indirectly compare A and B. Each indirect estimate was calculated via the difference in EPDS CFB or HAMD-17 CFB between studies, using the equations described in Figure 3³² to determine the relative effects between treatments.

Figure 3. Bucher comparison diagram. Note. d_AB is the estimated relative effect between A and B (e.g. log odds ratio, log hazard ratio), var(d_AB) is the variance of the indirect estimate

Network meta-analysis (NMA)

The relative efficacy estimates for each study were compiled into one analysis using standard NMA techniques, which use intra-study differences in CFB, calculated as described above, to compare treatments within an evidence network²⁹. Due to the small size of the networks, frequentist NMAs were used to compare all treatments without losing randomization³². All models were fitted and ran using the “netmeta” package in R software^56,57.

Ethics statement

This analysis uses only data from prior conducted studies and does not include any new investigation of human subjects.

Results

Evaluation of existing evidence

Using the criteria described in the Methods, seven studies were identified for inclusion in the network for ITCs: Hantsoo et al., Yonkers et al., Appleby et al., Misri et al., Sharp et al., O’Hara et al., and the SKYLARK Study^{25,41–43,51,58,59}. The full report from the SKYLARK Study was also available and was used for additional data⁶⁰. The heterogeneity and homogeneity of these studies were assessed. Adjustments were made to match the zuranolone treatment cohort to the placebo arm of another RCT and to interpolate results for missing timepoints.

Of the selected studies, four reported EPDS outcomes (Appleby et al.⁴², Sharp et al.⁴¹, Misri et al.⁵⁹, and the SKYLARK Study^25,60), and six reported HAMD-17 outcomes (Appleby et al.⁴², Hantsoo et al.⁵⁸, Yonkers et al.⁵¹, O’Hara et al.⁴³, Misri et al.⁵⁹, and the SKYLARK Study²⁵) The final network diagrams (following MAIC) for HAMD-17 CFB and EPDS CFB analyses are presented in Figure 4.

Figure 4. Evidence networks. (A) Studies reporting HAMD-17 CFB^{25,42,43,51,58,59}. (B) Studies reporting EPDS CFB^25,41,42,59. Abbreviations. CFB, change from baseline; HAMD-17, Hamilton Rating Scale for Depression; EPDS, Edinburgh Postnatal Depression Scale; MAIC, matching-adjusted indirect comparison; SSRI, selective serotonin reuptake inhibitor. Note. Combination refers to the combination of pharmacologic therapies and non-pharmacologic therapies.

Matching-adjusted indirect comparisons

MAIC results: HAMD-17 CFB outcomes

When conducting the MAIC on HAMD-17 CFB outcomes between the SKYLARK Study and Yonkers et al.⁵¹, several potential prognostic factors were assessed based on available data, which may influence patient outcomes in treated or untreated patients. The use of Mixed Models for Repeated Measures indicated that age, baseline HAMD-17, and suicidal ideation were key prognostic factors to consider in matching (in HAMD-17 multiple covariates analysis, p = 0.0313, p = 0.0004, and p = 0.0001, respectively); race was excluded from the final matching variables, as the covariate analyses indicated it was not a significant prognostic factor (p = 0.6964 for Black vs. White race; p = 0.8181 for Other vs. White race; Supplementary Table 5).

Following matching, the effective sample size of the reweighted SKYLARK Study zuranolone treatment arm was 8.3, which indicated a 91.8% reduction from the original sample size of 98. This effective sample size indicates that the weighted population provides a similar level of precision as a non-weighted sample with n = 8.3. While matching will always lead to a reduction from the original sample size, this large change indicates that the weights are likely variable due to heterogeneity and a lack of overlap at baseline between the populations in the two studies^48,61. Considerable uncertainty can be introduced in the analysis and results when the absolute effective sample size is small, since the loss of information while matching leads to a reliance on a few patients for outcome projections^48,61; as such, the true efficacy of zuranolone was unlikely to be well-represented by this small effective sample size.

The 91.8% reduction in effective sample size suggests that there may be an underlying difference in the populations of the SKYLARK Study and the study by Yonkers et al.⁵¹, which may make it inappropriate to compare their results. Indeed, the mean (SD) HAMD-17 score for the placebo group in the Yonkers et al. study was 24.7 (5.0)⁵¹, which is less than the minimum HAMD-17 score in the SKYLARK Study inclusion criteria (HAMD-17 ≥ 26)²⁵. The mean (SD) HAMD-17 score for the zuranolone-treated group in the SKYLARK Study was 28.6 (2.5)²⁵, and the populations in the SKYLARK Study²⁵ and the RCT used for matching⁵¹ had only a low level of overlap in terms of baseline HAMD-17 score (Supplementary Table 7). Due to this lack of overlap between the study populations, indirect comparisons utilizing HAMD-17 CFB results would have a high level of uncertainty and would not provide robust estimates of relative efficacy. Therefore, HAMD-17 CFB results were not considered for further analysis beyond this MAIC (Supplementary Tables 8 and 9).

MAIC results: EPDS CFB outcomes

When conducting the MAIC on EPDS CFB outcomes between the SKYLARK Study and Sharp et al.⁴¹, the use of Mixed Models for Repeated Measures indicated that age, baseline HAMD-17, and suicidal ideation were key prognostic factors to consider (in EPDS multiple covariates analysis, p = 0.0007, p < 0.0001, and p = 0.0003, respectively). Marital status was identified as a variable of borderline importance (p = 0.0930), and employment and race were identified as non-significant factors (p = 0.6429 for employment status; p = 0.5909 for Black vs. White race; p = 0.1588 for Other vs. White race; Supplementary Table 6). Since baseline HAMD-17 and EPDS score would likely be correlated, the final variables used for matching were age, baseline EPDS, suicidal ideation, and marital status (Supplementary Table 6).

The effective sample size of the reweighted SKYLARK Study zuranolone treatment arm, after matching, was 38.8 (compared to 98 in the unmatched population), suggesting that the overlap between the SKYLARK Study and Sharp et al.⁴¹ study populations in terms of baseline EPDS score would increase the precision of the estimates (Supplementary Table 10). After matching, the mean EPDS CFB for zuranolone-treated patients decreased across all timepoints, and these adjusted CFB values were used to assess the relative effect of zuranolone vs placebo, by comparing the adjusted SKYLARK Study CFB values to those of the placebo group in Sharp et al.⁴¹ (Supplementary Tables 11 and 12). This adjusted relative effect was utilized in Bucher ITCs and NMAs with respect to EPDS CFB outcomes.

Indirect treatment comparisons using MAIC results

Bucher ITCs

Bucher ITCs were performed for EPDS CFB on Day 3, Day 15, Day 28/Month 1, Day 45, and the last observation timepoint. These comparisons allowed for an evaluation of the effects of zuranolone in relation to those of SSRIs and combination treatments. The studies included in the evidence network that reported both EPDS CFB outcomes and placebo arms were Sharp et al.⁴¹ and Appleby et al.⁴² (Figure 4(B)).

The first analysis compared the EPDS CFB for zuranolone and SSRI treatments, as presented in Sharp et al.⁴¹. Sharp et al.⁴¹ utilized a pragmatic study design and did not administer one specific SSRI treatment or dosage⁴¹; this aggregate category was then compared to zuranolone in this analysis. Overall, there was a smaller EPDS CFB for the SSRI arm than the zuranolone treatment arm across all timepoints at Day 15 and later, as indicated by the negative differences in CFB (Figure 5). A comparison of particular interest included that on Day 15, which marks the end of the zuranolone treatment period. The prominent difference in EPDS CFB at Day 15 indicates that patients receiving zuranolone treatment experienced a greater reduction in EPDS from baseline to Day 15 than those receiving SSRIs (mean difference [MD]: −4.22). The 95% CI for this comparison, which falls entirely below 0 (95% CI: −6.16, −2.28), indicates that this difference is significant.

Figure 5. Forest plots displaying the results of Bucher ITCs and NMA (including MAIC) assessing EPDS CFB for a 14-day course of zuranolone treatment vs. ongoing treatment with SSRIs or placebo. Abbreviations. CFB, change from baseline; CI, confidence interval; EPDS, Edinburgh Postnatal Depression Scale; MAIC, matching-adjusted indirect comparison; MD, mean difference; NMA, network meta-analysis; SSRI, selective serotonin reuptake inhibitor. Note. A negative mean difference favors zuranolone; this indicates that patients in the other treatment groups observed a smaller change from baseline than those in the zuranolone treatment arm. The zuranolone treatment course lasted for 14 days, with off-treatment follow-up through Day 45; SSRIs could be administered through the end of the study period (Week 12 in Appleby et al.⁴², Week 18 in Sharp et al.⁴¹) Last observation occurred at Day 45 for zuranolone, Week 18 for placebo, and Weeks 12 and 18 for SSRIs (Appleby et al.⁴² and Sharp et al.⁴¹, respectively). Results for the SSRI MAIC + Bucher treatment compare the zuranolone treatment arm to Sharp et al.⁴¹; results from the Placebo treatment were determined using MAIC + NMA. The NMA considered zuranolone, SSRI, placebo, and combination treatments, as outlined in Figure 4(B).

In order to assess the EPDS CFB when all treatments had been administered for sufficient time to observe results, EPDS CFB at the last observation (timepoint varied by study; last observation values carried forward and compared) were also compared, finding a mean difference of −4.15 (95% CI: −5.70, −2.61). The difference in CFB was smaller at this timepoint than on Day 15, Day 28/Month 1 (MD: −4.93; 95% CI: −6.61, −3.24), or Day 45 (MD: −7.43; 95% CI: −9.84, −5.02). However, comparison of the SKYLARK Study data to Sharp et al.⁴¹ still showed a significantly larger EPDS CFB, indicating that at the last observation, patients who received a 14-day treatment course of zuranolone had a larger change from initial EPDS scores than those who received ongoing SSRI treatment over their study period.

While the SKYLARK Study assessed patient outcomes as early as Day 3, no data were available at Day 3 for any of the included SSRIs arms. As such, all ITCs were conducted using interpolated estimates for SSRI-treated patients, and Bucher ITCs for outcomes at Day 3 are presented in Supplementary Figure 1. ITCs that compare zuranolone treatment to combination treatment are presented in Supplementary Figure 2. Additional results by timepoint and by study from Bucher ITCs comparing zuranolone treatment to SSRIs and combination treatment are presented in Supplementary Tables 13 and 14.

Network meta-analysis

NMAs were used, with MAIC results as inputs, to provide a unified analysis comparing efficacy of zuranolone, SSRIs, placebo, and combination treatment. As in the Bucher ITCs, comparisons were drawn based on EPDS CFB on Day 3, Day 15, Day 28/Month 1, Day 45, and last observation. Forest plots displaying the mean difference in EPDS CFB indicate that patients treated with a 14-day course of zuranolone experienced a significantly greater change from baseline EPDS score than those receiving up to 12–18 weeks of SSRIs, placebo, or combination treatment at all timepoints on or following Day 15 (Figure 5; Supplementary Figure 1).

At Day 15 (end of zuranolone treatment period), the EPDS CFB for zuranolone was 6.32 points greater than placebo (95% CI: −7.83, −4.81) and 4.52 greater than that of SSRIs (95% CI: −6.40, −2.65). At the time of the last observation, the EPDS CFB for zuranolone surpassed that of placebo and SSRIs by 4.75 (95% CI: −6.17, −3.33) and 4.10 (95% CI: −5.64, −2.56), respectively.

A summary of EPDS CFB for oral treatments, assessed via Bucher ITC (with MAIC) and NMA (with MAIC), is available in Supplementary Table 15.

Discussion

This study aimed to provide an estimation of the relative effects of oral treatments and combination treatments for PPD by comparing patient outcomes observed with a novel, US FDA-approved treatment for adults with PPD, zuranolone, to SSRIs, SNRIs, and combination treatments. Using the eligible studies identified, this analysis assessed the comparative effectiveness of zuranolone, SSRIs (including fluoxetine, sertraline, paroxetine, citalopram, or escitalopram), placebo, and combination treatment (SSRIs [including fluoxetine and paroxetine] + non-pharmacologic interventions), using EPDS CFB as a measure of patient outcomes.

Findings from these ITCs suggest that patients treated with zuranolone experienced greater improvement in depressive symptoms, as measured by the EPDS, than those treated with SSRIs. Indeed, in this analysis, zuranolone-treated patients experienced a significantly larger improvement in EPDS score, compared to those treated with SSRIs, by Day 15, the timepoint marking the end of the 14-day treatment course of zuranolone. SSRIs, on the other hand, may require 6–12 weeks before some patients begin to experience their full effects⁶². SSRIs may require multiple dosing adjustments or up-titration for some patients before an effective therapeutic dose is achieved, which can contribute to delayed onset of efficacy¹⁶. These comparisons thus can provide clinical value in informing real-world patient care and potentially optimizing the time to treatment response. Early and effective treatment is particularly important for postpartum patients, who experience a crucial period of bonding with their infant and may experience parenting stress, which may impact depressive symptoms^14,63. Furthermore, improvement in PPD symptoms is critical in helping to address the health of the entire family unit, including maternal, infant, and partner/caregiver outcomes^7,12,64–67. While these ITCs focused on relative treatment efficacy, patients and providers should also consider the safety profile of pharmacological interventions when deciding on a treatment course for PPD, which should also consider the safety of each treatment option during breastfeeding. Research to date has shown that zuranolone transfer into the breast milk of healthy, lactating adult females was low^27,68.

In this analysis, to provide accurate comparisons and estimations of relative efficacy, the placebo effects observed across studies were evaluated. Placebo effects are widely observed in studies investigating mental health⁶⁹ and can also be influenced by both the visit frequency⁴⁴ and patient expectations^45–47. Inflation of the placebo effect could lead to an underestimation of the relative efficacy of treatment vs. placebo. The comparison of the response observed in the placebo arm across studies suggested a general trend in which those studies with more frequent study visits tended to have greater placebo responses. The placebo arm of the SKYLARK Study showed a much larger CFB in the outcomes of interest than those of the other studies in the network, possibly due to expectations of the rapidity and magnitude of the effect associated with zuranolone²⁸ or the high frequency of visits in the study period^25,44. As visits were conducted during the COVID-19 pandemic, when in-person contact may have been particularly valuable, the placebo effect may have been amplified in the SKYLARK Study, thereby underestimating the relative efficacy of zuranolone. To adjust for this effect, an unanchored MAIC was conducted, in which the zuranolone treatment arm of the SKYLARK Study was matched to a placebo arm from another study, using statistical weighting derived from key baseline characteristics.

The results of the MAIC were then used to conduct Bucher ITCs, as well as an NMA. Both methods indicated that zuranolone-treated patients experienced a significantly larger EPDS CFB than those treated with SSRIs at all timepoints from Day 15 and onward. Comparison of patient outcomes at Day 45 and at the last observation timepoints indicate that zuranolone-treated patients experienced significantly greater EPDS CFB even after a sufficient window for SSRI onset of efficacy and 30 days after cessation of zuranolone treatment. The improvement in EPDS score among zuranolone-treated patients when compared to SSRIs and combination therapy supports both the speed of onset of action of zuranolone and its sustained outcomes. These findings also support a previous indirect treatment comparison assessing the relative efficacy of another neurosteroid (brexanolone) and SSRIs in patients with PPD and found that those treated with brexanolone exhibited greater EPDS and HAMD-17 CFB than those who received SSRIs³².

In the absence of a direct head-to-head clinical trial, Bucher ITCs are a transparent method to use to undertake pairwise comparison, providing a comparison that depends on fewer assumptions, while NMA provide comparisons across an adjustable range of treatments. The level of agreement between the results of the Bucher ITCs and NMA increases confidence in these findings. However, while NMA can be limited by the required assumption of equivalence between the placebo arms of each trial, Bucher ITCs are not reliant on this assumption. The application of two different methods for indirect comparison indicates that more robust conclusions can be drawn due to the consistency in their findings, though both methods are sensitive to bias introduced during the MAIC.

Limitations

The study findings should be considered in light of several potential limitations. ITCs rely on several assumptions of exchangeability, homogeneity, and comparability between studies⁷⁰, which can be challenging to achieve when investigating treatments for PPD due to the limited evidence base of RCTs investigating those treatments¹⁷, their generally small sample sizes, and the paucity of evidence regarding many baseline characteristics. Of the 51 studies considered for inclusion, only seven studies ultimately met all inclusion criteria and could be used in this analysis, suggesting the paucity of evidence from RCTs demonstrating the use of SSRIs in PPD.

Differences across studies in patient eligibility criteria and matching on specific baseline patient characteristics limited the feasibility of undertaking indirect comparisons using HAMD-17 CFB, due to insufficient overlap in key patient characteristics, such as baseline HAMD-17 score. Such differences across studies may have also impacted the comparisons using EPDS CFB, and a detailed assessment of cross-trial heterogeneity is key for contextualizing ITCs. A systematic review and meta-analysis of SSRIs in the treatment of postnatal depression also suggested that baseline characteristics may not always be of significance in such comparisons⁷¹, but this study also aimed to mitigate some cross-trial differences and account for disparities in clinical covariates by utilizing patient-level data for the SKYLARK Study. To account for differences in study design and mitigate the effects of heterogenous placebo responses, this analysis also omitted the SKYLARK Study placebo arm from comparative analyses and matched the SKYLARK Study zuranolone treatment arm to the placebo arm of another RCT.

Several assumptions were required to conduct this analysis. While some differences in study design were noted, studies were assessed to be comparable given the limited data availability and the anticipated effects of such variation. The NMA also assumed that placebo arms that included non-pharmacological interventions, such as counseling, were comparable to placebo arms without these interventions. Additionally, all SSRIs were grouped and all combination therapies were grouped, assuming that all members of each category were equally effective. Further assumptions of equivalence included the treatment of geometric means reported by Appleby et al.⁴² as comparable to the arithmetic means reported in other studies, and the imputation of CFB variances across all studies using the SKYLARK Study IPD. Patients in the SKYLARK Study were also allowed concurrent treatment with antidepressants in addition to the study treatment, if patients were on a stable antidepressant dose for at least 30 days prior to their first study treatment dose²⁵, which could influence comparison to patients only receiving SSRIs, though it does mirror the real-world clinical option for the use of zuranolone as an adjunctive therapy alongside other oral antidepressants¹⁵.

Assumptions are of particular importance given that differences across studies required use of an unanchored MAIC to compare the effects of zuranolone to those of SSRIs. Assessment of EPDS and HAMD-17 CFB among the placebo groups of included studies suggested that the placebo effect may have been overestimated in the SKYLARK Study, which would impact the estimation of relative efficacy when comparing across studies. As such, these data suggest that the SKYLARK Study placebo arm is not suitable as a common comparator in ITCs for zuranolone, and the SKYLARK Study treatment arm was matched to a placebo arm from a comparator study using statistical weighting to adjust for key baseline characteristics. This method of dropping a placebo arm and considering the SKYLARK Study as a single-arm trial has been utilized in previous literature³², based on methodologies recognized in health technologies assessment⁴⁸. However, this method does present some limitations: considering the SKYLARK Study as a single-arm trial may not account for any effects in the remaining arm (zuranolone treatment group) which may have contributed to the overestimated placebo effect in the other arm, such as the frequency of study visits, the potential value of in-person visits during the COVID-19 pandemic, or expectations of efficacy. Accordingly, there is a level of uncertainty introduced with the use of unanchored MAIC, as in the absence of a common comparator (placebo), within-trial effects are not accounted for, and the comparison may be sensitive to bias. The estimation of between-trial differences is dependent on the strength of the matching between study populations and assumes all key factors have been incorporated. The findings of this ITC and the feasibility of comparisons in these population may also not be generalizable to other ITCs, if the PPD population differed for the included studies.

Since the zuranolone treatment arm was linked to the network via matching to one study, it is also possible that selection of a different study for matching could have provided different results. Notably, the mean EPDS score at baseline was higher in the SKYLARK Study (21.1)²⁵ than in Sharp et al. (17.7)⁴¹, indicating that PPD severity and the choice of a different study for matching could have impacted these findings. The low effective sample size observed after matching for HAMD-17 CFB outcomes may similarly have been driven by the range of HAMD-17 scores included in the two trials. The mean (SD) HAMD-17 score for the placebo group in the Yonkers et al. study was 24.7 (5.0)⁵¹, which fell below the minimum HAMD-17 score (≥26) required for inclusion in the SKYLARK Study²⁵. We observed a 91.8% reduction from the original to effective sample size of the SKYLARK Study zuranolone arm. In relation to other studies utilizing MAIC, this reduction was larger than the median reduction in sample size (74.2%) observed in an assessment of 16 technology appraisals (9 reporting effective sample size) that used MAIC⁶¹. As such, a robust number of patients were not available for matching because the study populations were distinct in terms of baseline HAMD-17 score, and no further valid analyses using HAMD-17 CFB outcomes could be conducted. As none of the other included studies reporting HAMD-17 results had a higher placebo group mean baseline HAMD-17 score, it is unlikely that matching to a different study would have made this analysis feasible.

As HAMD-17 is not regularly used in clinical practice, the real-world applications of such comparisons may have been limited even if they could be conducted; however, there are also limitations associated with assessing outcomes in terms of EPDS CFB alone. While EPDS is a common screening tool in clinical practice⁷², its use here in quantifying PPD symptom severity may extend beyond its intended scope. The EPDS may not fully capture all dimensions of symptom severity and change over time, particularly as it is limited in its ability to capture somatic symptoms^73,74 and effects on parent-child bonds⁷³.

Differences between treatments led to variations in study design and timepoints reported in each trial. Several adjustments were required to compare the results of this trial to the comparator studies. Linear interpolation was used to estimate the effects of SSRIs at timepoints in the SKYLARK Study not reported in the comparator studies. Given that SSRIs may take 6–12 weeks for some patients to experience effects^62,75, such interpolation likely overestimated their potential effects in the early treatment course and likely provides a conservative estimate of the relative effect of zuranolone compared to other treatments at early timepoints. In a real-world context, patients receiving long-term treatment with SSRIs also may discontinue pharmacotherapy at later timepoints.

The follow-up period extended through Day 45 in the SKYLARK Study, as this marked 1 month following the end of the treatment course. The last observation date varied across comparator studies, ranging up to approximately 18 weeks. As such, the effects of treatments at the last observation were carried forward and compared, to assess the relative effects between zuranolone and SSRIs when all treatments had been administered for a sufficient time to observe effects. However, this comparison is limited, as it assumes the efficacy of all treatments would be maintained from their last observation to the final comparator date (the latest follow-up observation time, or approximately 18 weeks).

This assumption may over-estimate the efficacy of zuranolone at the last observation timepoint, since follow-up data from the SKYLARK Study did not extend beyond Day 45. Additionally, not all studies investigating SSRIs used the same final observation timepoint; the NMA incorporated data from both Appleby et al.⁴² and Sharp et al.⁴¹, which had last observation timepoints of Week 12 and Week 18, respectively. The NMA assessment of SSRIs vs. zuranolone thus also assumes that the effects of SSRI treatment remain stable from Week 12 to Week 18. As such, heterogeneity in the included studies and the limitations of the studies in the network may impact the MAIC ITC results and their generalizability.

Conclusions

This study presents multiple ITCs, following MAIC analysis, to compare the relative effects of zuranolone treatment for PPD to SSRIs, placebo, and combination treatment (pharmacologic and non-pharmacologic treatment). The results of the ITCs indicate that, by Day 15, following completion of the 14-day treatment course, patients with PPD who received zuranolone experienced a significantly larger improvement in the reduction in depressive symptoms measured using EPDS score than those who received SSRIs. Zuranolone-treated patients continued to improve post-treatment period and also exhibited significantly greater improvement at later timepoints, including a comparison of the last observations from all studies. These findings support prior findings that zuranolone could represent a novel treatment for PPD with both a rapid onset of action and sustained outcomes and suggest that zuranolone may support greater improvement in depressive symptoms relative to SSRIs. Though ITCs can be limited by available baseline data and variations in study design, the comparative data presented in this study provide a useful starting point for considerations that can inform clinical decision-making in the absence of head-to-head data. By providing insights on the comparative timing and level of response associated with oral PPD treatments, such findings can be considered alongside further investigations to optimize patient care and more effectively manage PPD symptoms.

Transparency

Author contributions

All authors were involved in study conception/design or data analysis and interpretation. All authors were involved in writing/critical review of draft versions of this manuscript, and all approved the final version to be submitted for publication.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Previous presentations

The authors have no prior presentations of this work to report.

Acknowledgements

Clinical trials used in the evidence networks were described in peer-reviewed literature; patient-level data were provided by Sage Therapeutics, Inc. The authors would like to thank Lasair O'Callaghan (Sage Therapeutics, Inc.) for her input and guidance during analytical methods development. Medical writing assistance was provided by Julie Bevilacqua and Francie Moehring-Moskal, employees of Boston Strategic Partners, Inc (funded by Sage Therapeutics, Inc. and Biogen Inc.).

Declaration of financial/other interests

SS and NR are employees of Lumanity Inc and report consulting fees from Sage Therapeutics, Inc. and Biogen Inc. MEG, RT, and YT are employees of Sage Therapeutics, Inc., and may hold stock or stock options. CM and SC are employees of Biogen Inc. and may hold stock. SMB reports grants and other research funding from Sage Therapeutics, Inc., awarded to the University of North Carolina at Chapel Hill during the conduct of the study; and grants from the NIH and PCORI. She also reports advisory board or consulting fees from Embarck Neuro, Modern Health, and WebMD/MedScape outside the submitted work. KD serves as a consultant to Brii Biosciences, Inc.; Gerbera Therapeutics; GH Research Ltd.; Neuroscience Software, Inc.; Reunion Neuroscience; and Sage Therapeutics, Inc.; reports grants from Sage Therapeutics, Inc., awarded to Zucker Hillside Hospital/Feinstein Institutes for Medical Research during the conduct of the brexanolone injection and zuranolone clinical trials; and received grants from the NIH, Premier Healthcare, and Woebot Health and royalties from an NIH employee invention outside of the submitted work.

Additional information

Funding

This study was funded by Sage Therapeutics, Inc. and Biogen Inc. Support for medical writing assistance was funded by Sage Therapeutics, Inc. and Biogen Inc. The funding source was involved in the design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, and approval of the manuscript. The publication of study results was not contingent on the funding source’s approval or censorship of the manuscript.