Background

Breast cancer is the most frequently diagnosed cancer, with an estimated 2.3 million new cases worldwide in 2020, and the leading cause of cancer death among women [1]. Approximately 15–20% of breast cancer cases are molecularly classified as triple-negative breast cancer (TNBC), which is characterized by tumors lacking expression of the estrogen receptor, progesterone receptor, and epidermal growth factor receptor 2 [2, 3]. Compared with women with other molecular subtypes of breast cancer, women with TNBC generally present at younger ages and are more likely to experience early recurrence after treatment and distant metastasis to visceral organs including the brain [3,4,5]. Moreover, women with TNBC exhibit the shortest survival time among all breast cancer patients, with a mortality rate of 40% within 5 years of diagnosis [3, 5].

As TNBC is insensitive to endocrine therapy due to its molecular phenotype, cytotoxic chemotherapy has historically been the mainstay treatment approach. Chemotherapy combinations involving doxorubicin, cyclophosphamide, anthracycline, paclitaxel, or platinum-based agents administered before tumor resection are shown to improve early outcomes, including pathologic complete response (pCR) and event-free survival (EFS), among women with early-stage or locally advanced TNBC [3]. Advances in our understanding of the molecular mechanisms of cancer, however, have led to the development of new targeted therapies that may achieve even greater improvements in TNBC patient outcomes. Of particular interest, immunotherapies that inhibit immune checkpoints such as programmed cell death protein 1 (PD-1) or programmed cell death ligand 1 (PD-L1), which are leveraged by tumor cells to evade recognition and destruction by the immune system, can restore the body’s ability to effectively attack tumors. Indeed, neoadjuvant delivery of the PD-1 inhibitor pembrolizumab + chemotherapy followed by adjuvant pembrolizumab was granted U.S. Food and Drug Administration approval in July 2021 for treating high-risk, early-stage TNBC based on statistically significant and clinically meaningful improvements in pCR and EFS in a phase III randomized controlled trial (RCT) [6, 7]. Furthermore, another phase III RCT shows that neoadjuvant treatment with the PD-L1 inhibitor atezolizumab + chemotherapy improves pCR among early-stage TNBC patients [8].

Network meta-analysis (NMA) is a statistical method that enables indirect comparisons between treatments where head-to-head evidence may not be available, which allows clinicians, guideline developers, and health technology assessment agencies to evaluate evidence on new treatments within the context of all existing evidence [9, 10]. Specifically, NMA can be used to combine direct and indirect evidence regarding any interventions that form a connected network of RCTs wherein each trial has at least one intervention (active or placebo) in common with another trial and all trials are sufficiently similar [11]. Clinical trial evidence suggests that neoadjuvant pembrolizumab + chemotherapy followed by adjuvant pembrolizumab is an effective and safe approach to treating early-stage or locally advanced TNBC; however, this treatment regimen has not been compared against all alternative neoadjuvant treatment regimens in head-to-head trials. As the relative efficacy of various immunotherapy- and chemotherapy-based regimens is of interest to both clinicians and healthcare policymakers, the aim of this analysis was to estimate the comparative efficacy of neoadjuvant pembrolizumab + chemotherapy followed by adjuvant pembrolizumab versus other neoadjuvant treatments for patients with high-risk, early-stage TNBC in terms of pCR, EFS, and overall survival (OS) through a systematic review and NMA.

Methods

Systematic review

The systematic review and NMA were performed in accordance with Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) guidelines [12] and followed a pre-specified protocol. Selection criteria for the population, interventions, comparators, outcomes, and study design (PICOS) are outlined in Table 1. RCTs were included if they enrolled patients with early-stage and locally recurrent non-metastatic TNBC, evaluated interventions of interest, reported outcomes of interest, and were published in English. Due to an anticipated lack of trials conducted solely in TNBC patients, trials were eligible for inclusion if they enrolled TNBC patients exclusively and reported at least one outcome of interest or if they enrolled patients from a broader population of breast cancer patients and reported at least one outcome of interest in a subgroup composed of > 90% TNBC patients. As the primary purpose of this systematic review and NMA was to identify and synthesize evidence from the clinical literature to support health technology assessment submissions, the interventions of interest included only those treatment regimens used in clinical practice in multiple countries, and the protocol was not registered in a systematic review registry.

Table 1 PICOS criteria to identify trials for the systematic literature review
Full size table

Relevant trials were identified by searching Excerpta Medica DataBASE (EMBASE), Medical Literature Analysis and Retrieval System Online (MEDLINE), and Cochrane Central Register of Controlled Trials (CENTRAL) through the OVID platform on April 21, 2022 (Additional Tables 1-3). EMBASE and MEDLINE searches were limited to RCTs using the Scottish Intercollegiate Guidelines Network (SIGN) filter (https://www.sign.ac.uk/what-we-do/methodology/search-filters/). Manual searches American Society of Clinical Oncology (2021–2022), European Society of Medical Oncology (2021), and San Antonio Breast Cancer Symposium (2021) conference proceedings were conducted to identify RCTs that had not yet been published in full-text form. In addition, the U.S. National Institute of Health Clinical Trials Registry (clinicaltrials.gov) and European Union Clinical Trial Registry (clinicaltrialsregister.eu) were searched to identify completed RCTs with results available that had not yet been published.

Two reviewers performed abstract selection, full-text selection, and data extraction in duplicate. Any unresolved discrepancies between the two reviewers were resolved by involving a third reviewer and reaching consensus. The Cochrane Collaboration’s Risk of Bias tool was used to assess the risk of bias of included RCTs [13].

Following completion of the systematic review, the comparators component of the PICOS was broadened to include any intervention that was compared to two or more interventions of interest, and a targeted literature review was conducted to identify additional RCTs to form a connected network for the NMA.

Feasibility assessment and network meta-analysis

An extension of pairwise meta-analysis, NMA allows indirect comparisons of interventions that have not been evaluated in head-to-head trials [11]. As the validity of any NMA is based on whether there are systematic differences among trials included in the network across treatment comparisons [11, 14,15,16,17], a feasibility assessment was conducted before proceeding with the NMA [14].

As only one trial connected each treatment in the networks of evidence, between-study heterogeneity could not be reliably estimated; therefore, NMA was performed with a fixed-effects assumption. The NMA of reported hazard ratios (HRs) in terms of EFS and OS assuming proportional hazards between treatments was performed using a regression model with a contrast-based normal likelihood for the log HR and corresponding standard error of each trial in the network [18]. Normal non-informative prior distributions for the parameters were estimated with a mean of 0 and a variance of 10,000.

NMA of reported Kaplan–Meier (KM) curves in terms of EFS and OS assuming time-varying hazards between treatment was performed using a fractional polynomial model [11, 19]. Weibull, Gompertz, and second-order fractional polynomial models were considered using a multivariant NMA framework. Reported KM curves were digitized for each treatment arm included in the NMA using DigitizeIt (http://www.digitizeit.xyz). Goodness-of-fit was compared across competing survival models using the deviance information criterion, with a model having a better trade-off between fit and parsimony having a lower deviance information criterion. Relative treatment effects were expressed as odds ratios for pCR and HRs for EFS and OS with 95% credible intervals (CrIs), which reflect a 95% probability that the estimate is within the specified range. Additionally, to allow for time-varying HRs, NMAs with fractional polynomial models representing different survival distributions were fit to the data under a variety of different assumptions about the shape of the hazard function. All analyses were performed using R version 4.0.3 (http://www.r-project.org/) and OpenBugs version 3.2.3 (OpenBUGS Project Management Group) [20].

Results

Systematic review

The study selection process for identifying trials of interest is outlined in Fig. 1. Thirteen citations pertaining to seven unique trials were identified in the systematic review (Additional Table 4) [6, 8, 21,22,23,24,25,26,27,28,29,30]. As the trials identified by the systematic review did not form a connected network of evidence, a targeted literature search was conducted to include any intervention that was compared to two or more treatments of interest, which resulted in the inclusion of five additional citations pertaining to three unique trials. Trials that were not connected in a network, including IMpassion031 evaluating atezolizumab + chemotherapy, were excluded. As a result, a total of eight citations pertaining to five unique trials were included in the feasibility assessment and NMA (Additional Table 5) [22, 23, 31,32,33,34,35].

Fig. 1
figure 1

PRISMA flow diagram

Full size image

Feasibility assessment

Two trials were multinational (KEYNOTE-522 and BrighTNess), two trials were conducted in the U.S. (CALGB 40603 and NeoSTOP), and one trial was conducted in Germany (GeparSepto-GBG 69). KEYNOTE-522 employed quadruple-blind masking, whereas the other trials were open-label (Table 2). All trials exclusively enrolled patients with early-stage, locally advanced non-metastatic TNBC. Patients were enrolled in all trials irrespective of PD-L1 status; however, KEYNOTE-522 measured PD-L1 status using the PD-L1 IHC 22C3 pharmDx test (Dako North America, Inc.) and reported outcomes for PD-L1 subgroups. All trials employed comparable treatment dosing and administration schedules (Additional Table 6).

Table 2 Characteristics of trials included in the feasibility assessment and network meta-analysis
Full size table

Although baseline patient characteristics were not reported for most trials, the available data on patient characteristics and enrollment criteria suggest no important between-trial differences (Additional Table 7). Among the trials reporting baseline patient characteristics, median age ranged from 48 to 54 years, 99.9–100% of patients were female, and 69–74% of patients were Caucasian. Based on trial eligibility criteria, four trials enrolled patients with an ECOG performance score of 0 or 1, whereas one trial (GeparSepto-GBG 69) enrolled patients with a Karnofsky performance score of 70–80% or better.

All five trials reported pCR. Survival outcomes for CALGB 40603 were only available from the US National Institutes of Health Clinical Trial Registry, which reported identical HRs and 95% confidence intervals for both OS and recurrence-free survival (which was defined similarly as EFS in the other included trials, Additional Table 8). As this appeared to be a reporting error, CALGB 40603 was excluded from the primary NMA of OS and EFS; sensitivity analyses including this trial are reported in Additional Fig. 4 and Additional Tables 19–20. In addition, as GeparSepto-GBG 69 did not provide KM curves for OS, this trial was only included in the constant HR NMA model for OS. Data sources for each outcome measure are presented in Additional Table 9.

Network meta-analysis

Networks were constructed to compare neoadjuvant pembrolizumab + chemotherapy followed by adjuvant pembrolizumab to other neoadjuvant treatment regimens. A fundamental assumption was made in the inclusion of trials evaluating chemotherapy regimens without analyses by PD-L1 subgroups: PD-L1 expression levels only influenced the PD-L1-directed therapy-containing regimen (i.e., pembrolizumab). Also, as doxorubicin and epirubicin have similar efficacy profiles [36], networks of evidence were constructed by treating cohorts assigned to doxorubicin or epirubicin as equivalent (classified as anthracycline). Five trials were included in the pCR network (Fig. 2), and four trials were included in the OS and EFS networks (Fig. 3).

Fig. 2
figure 2

Network of evidence for pathologic complete response. Arrows (→) indicate where treatment was administered sequentially, with treatments to the left of the arrow administered first. The orange circle denotes the primary treatment regimen of interest. Anthra includes doxorubicin and epirubicin, which were assumed to be equivalent. Anthra = anthracycline; bev = bevacizumab; carb = carboplatin; cyclo = cyclophosphamide; doc = docetaxel; nab-pac = nab-paclitaxel; pac = paclitaxel; pembro = pembrolizumab; veli = veliparib

Full size image
Fig. 3
figure 3

Network of evidence for event-free survival and overall survival. Arrows (→) indicate where treatment was administered sequentially, with treatments to the left of the arrow administered first. The orange circle denotes the primary treatment regimen of interest. Anthra includes doxorubicin and epirubicin, which were assumed to be equivalent. Anthra = anthracycline; bev = bevacizumab; carb = carboplatin; cyclo = cyclophosphamide; doc = docetaxel; nab-pac = nab-paclitaxel; pac = paclitaxel; pembro = pembrolizumab; veli = veliparib

Full size image

For pCR, neoadjuvant pembrolizumab + paclitaxel + carboplatin followed by pembrolizumab + anthracycline + cyclophosphamide followed by adjuvant pembrolizumab showed statistically favorable improvements in pCR versus paclitaxel + carboplatin followed by anthracycline + cyclophosphamide, paclitaxel followed by anthracycline + cyclophosphamide, and paclitaxel + bevacizumab followed by anthracycline + cyclophosphamide + bevacizumab (Table 3).

Table 3 Results of network meta-analysis for pathologic complete response
Full size table

For EFS, neoadjuvant pembrolizumab + paclitaxel + carboplatin followed by pembrolizumab + anthracycline + cyclophosphamide followed by adjuvant pembrolizumab showed statistically favorable improvements in EFS versus paclitaxel followed by anthracycline + cyclophosphamide, paclitaxel + carboplatin followed by anthracycline + cyclophosphamide, and paclitaxel + carboplatin + veliparib followed by anthracycline + cyclophosphamide (Table 4).

Table 4 Results of network meta-analysis for event-free survival
Full size table

For OS, neoadjuvant pembrolizumab + paclitaxel + carboplatin followed by pembrolizumab + anthracycline + cyclophosphamide followed by adjuvant pembrolizumab showed statistically favorable OS results versus paclitaxel followed by anthracycline + cyclophosphamide (Table 5).

Table 5 Results of network meta-analysis for overall survival
Full size table

As there were no major violations of the assumption that HRs were proportional over time (determined by log–log and Schoenfeld residual plots), the best-fitting models were determined to be constant HR models. Further, the best-fitting time-varying models for both EFS and OS did not show statistically meaningful changes in HR over time for any treatment (Additional Figs. 1-3, Additional Tables 10–17). Therefore, the constant HR results provided the best combination of fit and parsimony for all treatments.

Discussion

NMA including five RCTs of neoadjuvant immunotherapy or chemotherapy regimens for patients with early-stage or locally advanced, non-metastatic TNBC demonstrated that neoadjuvant pembrolizumab + chemotherapy followed by adjuvant pembrolizumab had more favorable treatment outcomes than certain other neoadjuvant treatment regimens. In particular, the relative efficacy of neoadjuvant pembrolizumab + chemotherapy followed by adjuvant pembrolizumab was statistically favorable to paclitaxel followed by anthracycline + cyclophosphamide in terms of pCR, EFS, and OS; to paclitaxel + carboplatin followed by anthracycline + cyclophosphamide in terms of pCR and EFS; to paclitaxel + bevacizumab followed by anthracycline + cyclophosphamide + bevacizumab in terms of pCR; and to paclitaxel + carboplatin + veliparib followed by anthracycline + cyclophosphamide in terms of EFS. Thus, neoadjuvant pembrolizumab + chemotherapy followed by adjuvant pembrolizumab may be more efficacious than other neoadjuvant treatment regimens for patients with high-risk, early-stage TNBC.

The validity of NMA findings are dependent on the quality of the individual RCTs included in the network and the degree to which these RCTs are similar in terms of populations and methodology [9,10,11]. In an NMA of RCTs involving multiple treatment comparisons, randomization holds only within individual RCTs and not across RCTs. Thus, if the direct comparisons between treatments in the network involve systematic between-trial differences in study or patient characteristics, and these differences are treatment effect modifiers, then the estimates of indirect comparisons will be biased. In this study, the feasibility assessment demonstrated that the distribution of most potential treatment effect modifiers was balanced across the RCTs. However, KEYNOTE-522 was the only trial that evaluated a treatment regimen spanning both the neoadjuvant and adjuvant phases, which may have served to influence the relative treatment effects.

Some factors limit the conclusions that can be drawn from this NMA. First, ambiguity in the reporting of recurrence-free survival and OS for CALGB 40603 precluded the inclusion of this trial in the survival analyses. Second, because only one study connected each treatment in the network of evidence, between-study heterogeneity could not be estimated, and the NMA was performed with a fixed-effects assumption. Third, the systematic review did not identify sufficient data to answer questions such as which biomarkers predict pCR, which patients may achieve pCR without pembrolizumab, or the ideal dosage of pembrolizumab to achieve an ideal risk/benefit ratio, which could be investigated in future studies.

Despite these limitations, this study has several strengths that maximize its comprehensiveness and rigor. In particular, highly sensitive systematic searches in the peer-reviewed literature, recent conferences, and clinical trial registries were employed to identify all published evidence from RCTs of neoadjuvant immunotherapy and chemotherapy treatments for early-stage, locally advanced, non-metastatic TNBC. In addition, the review process was guided by pre-defined eligibility criteria, and data quality was ensured through the involvement of two independent reviewers in the study selection and data extraction processes.

Conclusions

The results of this systematic review and NMA suggest that neoadjuvant pembrolizumab + chemotherapy followed by adjuvant pembrolizumab is an effective treatment compared with other neoadjuvant treatments for patients with previously untreated, locally advanced, non-metastatic TNBC.