Comparative Efficacy and Safety of Tislelizumab in Second-Line Esophageal Squamous Cell Carcinoma: Systematic Literature Review and Simulated Treatment Comparisons.

Key Summary Points

Introduction
Esophageal cancer (EC) has been identified as the seventh most common cause of cancer-related deaths globally with a mortality rate of 5.48 deaths per 100,000 people [1], resulting in approximately 445,000 deaths in 2022 [2]. Esophageal squamous cell carcinoma (ESCC) accounts for approximately 90% of EC cases [3], representing the leading cause of EC-related cancer incidence and deaths [4]. Usually asymptomatic in early stages, ESCC is most frequently diagnosed in symptomatic people with advanced disease [5] aged 64–74 years who have a poor prognosis [6].
Recent data from the Global Cancer Observatory (GLOBOCAN 2022) suggest that male individuals are two to three times more likely than female individuals to be diagnosed with and die of EC [7]. Eastern and South-Central Asia, alongside Northern Europe, rank among the five regions with the highest incidence and mortality rates of EC across both sexes [2]. The aggressive nature of EC is reflected in the low overall survival (OS) rates of all EC cases (both early-stage EC and advanced disease), with the estimated 5-year OS rate being around 10% in Europe, 20% in the USA and Eastern Asia, 30% in China, 31–33% in South Korea, and 36% in Japan [8–11]. However, the corresponding rates for advanced, unresectable, or metastatic EC can be as much as four to six times lower, with representative 5-year OS rates in the USA and South Korea being as low as 3.4% [12] and 7.3% [11], respectively.
The management of ESCC is largely dependent on disease staging based on the American Joint Committee on Cancer tumor, node, metastasis (TNM) system [13]. Locally advanced ESCC is usually treated with curative intent using a combination of surgery, radiotherapy, and chemotherapy [14], while advanced, unresectable ESCC is treated with immunotherapy, chemoradiotherapy, chemotherapy, chemo-immunotherapy, and radiation therapy, depending on patients’ performance status and programmed death-ligand 1 (PD-L1) expression level [15]. In particular, PD-L1 scores in the advanced disease setting have emerged as essential markers for eligibility for immunotherapy [16]. For eligible patients, immune checkpoint inhibitors that disrupt the interaction of PD-L1 with programmed cell death protein 1 (PD-1) by binding to PD-1, such as pembrolizumab [17] and nivolumab [18], are typically added to the therapeutic regimen. The introduction of anti-PD-1 agents has transformed the treatment landscape for immunotherapy-eligible patients with ESCC in the USA over the past few years [19]. According to the latest (2024) update of the guidelines of the National Comprehensive Cancer Network (NCCN), both pembrolizumab and nivolumab + platinum-based chemotherapy are recommended as first-line treatment for patients with advanced, recurrent, or metastatic ESCC [20]. However, recommendations for pembrolizumab combinations at first line are dependent on patients’ PD-L1 combined positive score (CPS), leading to category 1–2A recommendations for CPS ≥ 10 and category 2B for CPS < 10. Although nivolumab and pembrolizumab as monotherapies are also included in the NCCN Guidelines for second-line (2L) ESCC, pembrolizumab is only approved for patients with PD-L1 CPS ≥ 10. In addition, pembrolizumab monotherapy has yet to be approved for use at 2L ESCC in the European Union. Taken together, this body of evidence underscores the requirement for additional anti-PD-1 agents to improve the management of ESCC in the USA as well as regions with even higher disease burden, such as Europe and Asia [2].
Camrelizumab [21] and sintilimab [22] have also been investigated in phase II/III randomized clinical trials (RCTs) as 2L monotherapies for ESCC. Despite the promising results of anti-PD-1 antibodies in phase III trials and the successful use of pembrolizumab (in patients with PD-L1 CPS ≥ 10) and nivolumab in 2L treatment of ESCC, evidence from preclinical studies suggests that the clinical efficacy of anti-PD-1 antibodies could potentially be further improved [23]. One issue that has been shown to compromise the anti-tumor activity of anti-PD-1 antibodies and has been suggested as a potential mechanism of resistance to these agents is their binding to FcγR on macrophages and the downstream activation of antibody-dependent macrophage-mediated killing of effector T cells [24]. To overcome this issue and minimize binding to FcγR on macrophages, the humanized immunoglobulin G4 (IgG4) anti-PD-1 monoclonal antibody tislelizumab (BGB-A317) was developed [25]. In the phase III RATIONALE-302 trial, tislelizumab was shown to significantly improve OS compared with chemotherapy (median OS of 8.6 vs 6.3 months; hazard ratio [HR], 0.70; 95% confidence interval [CI], 0.57–0.85; P = 0.0001) in patients with advanced or metastatic ESCC regardless of PD-L1 expression level, whose tumor progressed after first-line systemic treatment, with a manageable safety profile [26]. This led to the approval of tislelizumab for use in 2L ESCC by the US Food and Drug Administration (FDA) and European Medicines Agency (EMA) [27, 28], and to the inclusion of tislelizumab monotherapy in the latest update to the NCCN Guidelines as a category 1 recommendation for the treatment of 2L ESCC [20].
To date, most RCTs in advanced, unresectable, or metastatic 2L ESCC compared anti-PD-1 antibodies with chemotherapy. In the absence of an RCT that directly compares the efficacy and safety of tislelizumab with that of other anti-PD-1 interventions, inferential analysis was used to identify the evidence on existing 2L treatments for ESCC and derive relative efficacy and safety estimates versus tislelizumab by means of indirect treatment comparisons (ITCs).

Methods

Systematic Literature Review
A systematic literature review (SLR) was conducted to identify and summarize all published data on clinical efficacy and safety of existing 2L treatment regimens for patients with advanced, unresectable, or metastatic ESCC. Table 1 lists the full inclusion and exclusion criteria based on the population, intervention, comparator, outcome, and study design (PICOS) criteria [29]. The general recommendations of the Cochrane Handbook for Systematic Reviews of Interventions [30], the general principles of the Centre for Reviews and Dissemination (CRD; University of York) [31] for undertaking reviews in healthcare, and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [32] guidelines were followed.
The search was first run in May 2021 and was updated twice, in November 2022 and October 2023, following the same protocol. The Health Technology Assessment database, OVID, OVID SP, conference proceedings, and trial registry websites were searched (full SLR strategy details are available in the Supplementary material). Identified abstracts were screened by two independent reviewers, with discrepancies resolved by a third senior reviewer. The full text of qualifying abstracts was assessed for final inclusion following the same process as in the abstract screening stage. Data extraction followed a piloted template, and included, among others, characteristics of the study design, population, evaluated interventions, and assessed outcomes. The quality of retained RCTs and non-randomized studies was assessed using the Cochrane Collaboration tool for risk of bias 2 [33] and the Downs and Black checklist [34], respectively.

Evidence Synthesis Feasibility Assessment
For the purposes of evidence synthesis, only studies that assessed immunotherapies licensed in the European Union were retained. These included the following: RATIONALE-302 (tislelizumab), ATTRACTION-3 (nivolumab), RAMONA (nivolumab with ipilimumab), KEYNOTE-181 (pembrolizumab), ESCORT (camrelizumab), and ORIENT-2 (sintilimab). An ITC feasibility assessment was conducted to evaluate the distribution of effect modifiers across included studies and recommend ITC methods that best address the characteristics of the available evidence base [35, 36]. The following study attributes were compared: design (e.g., phase, blinding, crossover, stratification, treatment switching, region of enrollment, PD-L1 assay employed, inclusion/exclusion criteria), patient demographics (e.g., Eastern Cooperative Oncology Group Performance Status [ECOG PS], proportions of liver/bone/lung metastasis, PD-L1 expression, prior treatments), control therapies (e.g., mechanism of action, posology, schedule of administration), and outcomes (e.g., endpoint definitions, follow-up duration). Effect-modifying and prognostic variables were determined, separately for each outcome, through a rigorous process that included analyzing the RATIONALE-302 individual-patient data (IPD) using likelihood ratio tests and a stepwise covariate selection process, ensuring consistency with previous ITCs in the same disease area, and consulting clinical experts. The proportional hazards (PH) assumption was evaluated for OS and progression-free survival (PFS) through visual inspection of the log-cumulative hazards and Schoenfeld residuals, as well as the Grambsch–Therneau test [37].

Simulated Treatment Comparisons
Aligned with recent methods guidance updates, simulated treatment comparisons (STCs) are preferred over matching-adjusted indirect comparisons (MAICs) for population adjustment in this anchored setting [38]. The STC implementation was based on NICE DSU TSD18 [39] but was extended to allow adjustment for binary and multilevel categorical characteristics and to enable the estimation of population-average treatment effects [40, 41]. To that extent, for each comparison, a dummy population of 10,000 patients was simulated on the basis of the reported baseline characteristics for the effect modifiers of interest in each comparator trial. Intra-trial correlations across covariates were preserved and reflected those in RATIONALE-302, because correlations were not reported in the comparator trials. Using the IPD in RATIONALE-302, a binomial generalized linear model was fit to treatment-related adverse events (TRAEs) grade ≥ 3 and a Cox model to OS and PFS. Effect modifiers were included as interactions with the treatment arm [39]. The fitted STC model was used to predict the outcome for every simulated patient and the predictions were marginalized for each treatment group before deriving the counterfactual relative effect that would have been observed in RATIONALE-302 had it enrolled the population of the comparator trial. Bootstrapping was used to estimate the variance of the counterfactual relative effect. In the final step, the relative effect of interest (i.e., tislelizumab vs comparator) was calculated with standard Bucher methods using the reported relative effect in the comparator study and the counterfactual relative effect in RATIONALE-302.

Ethics Approval and Consent to Participate
This article is based on previously conducted studies and does not contain any new studies with human participants or animals performed by any of the authors.

Results

Systematic Literature Review
Figure 1 illustrates the number of records identified and included in the original SLR and subsequent updates, with the stand-alone PRISMA charts of each update provided in the Supplementary material (Figs. S1–S3). A total of 3480 records were identified in the original search (update 1: 1455, update 2: 2113). Among those, 232 records were removed from the original search (update 1: 308, update 2: 926) as duplicate records, and 3190 records were excluded as non-eligible at the title-abstract screening stage (update 1: 1082, update 2: 1167), resulting in 58 records for full-text screening (update 1: 65, update 2: 20). Upon full-text assessments, 19 unique records were eventually included from the original SLR (update 1: 13, update 2: 3). During the subsequent SLR update, 10 conference abstracts of the original SLR were superseded by full-text records and only the latter were retained. Overall, the total number of unique records identified across the original search and subsequent updates was 25, pertaining to 13 unique trials.
Among the 13 included studies, five assessed a chemotherapy regimen in both arms [42–46], one evaluated chemotherapy against a tyrosine kinase inhibitor (TKI) [47], one compared chemotherapy with a combination of an anti-PD-1 agent and a TKI [48], five assessed chemotherapy against an anti-PD-1 agent [21, 22, 26, 49, 50], and one compared the combination of an anti-PD-1 antibody with another immunotherapy versus a single anti-PD-1 agent [51].
Of the six immunotherapy studies of interest, four were phase III RCTs [21, 26, 49, 50], one was a phase II RCT [22], and one was a non-randomized study [51]. All five RCTs [21, 22, 26, 49, 50] displayed a similar risk of bias profile (Table S1), with some concerns raised primarily as a result of the open-label nature of the studies, and the methodologic quality of the non-randomized study [51] was rated as “fair” (Table S2). On the basis of these findings, and since the purpose of the SLR was to inform the comparative efficacy of anti-PD-1 agents on 2L ESCC, all six studies that included anti-PD-1 agents were considered in the ITC feasibility assessment.

Evidence Synthesis Feasibility Assessment
All six studies compared an anti-PD-1 antibody with a chemotherapy regimen that varied across trials, but clinical opinion did not expect this to induce substantial heterogeneity, and it was therefore deemed suitable for creating a common anchor. The population overlap across RATIONALE-302 and comparator studies for the available effect-modifying variables is illustrated in Fig. 2. The KEYNOTE-181 trial enrolled a mixed histology population (squamous and non-squamous) but reported histology-specific baseline characteristics and outcomes. The RAMONA trial substantially differed from RATIONALE-302 in study design and inclusion criteria, preventing the possibility of an ITC. In contrast, all the remaining studies were deemed to be adequately similar to RATIONALE-302 in most effect-modifying aspects, with a few notable differences in important baseline characteristics (ECOG PS, disease and PD-L1 expression status, liver and lung metastases), which implied the need for population adjustment (see Table S3 in Supplementary material). The full list of variables considered for their effect-modifying status and deemed important for adjustment in the base case and scenario analyses is provided in Table S4 in the Supplementary material. The PH assumption held for OS in RATIONALE-302, suggesting that ITC analyses employing constant HRs are feasible. However, the PH assumption did not hold for PFS, as it is typically for immune-oncologic agents [52], likely as a result of delayed curve separation. Nevertheless, a common methodologic framework across endpoints driven by the primary outcome (OS) was considered most appropriate for consistency.
The STCs adjusted for all variables deemed important, if they were reported in the comparator trials. Scenario analyses adjusted for an extended set of variables, heeding the likely uncertainty inflation. Subgroup analyses were pursued for the primary outcome (OS) and separately considered ECOG 0 and ECOG 1 as well as PD-L1-positive and -negative categories defined on the basis of tumor positivity score (TPS) and CPS (positive: TPS ≥ 1% or CPS ≥ 10; negative: TPS < 1% or CPS < 10). All analyses were conducted with both available RATIONALE-302 data cutoffs (DCOs): December 1, 2020, used in the submission to the EMA and henceforth termed simply “early DCO,” and December 28, 2022, termed “latest DCO”; for comparators, the latest available follow-up was used throughout.

Simulated Treatment Comparisons
The results of the STC analyses using the latest DCO and the early DCO are provided in Tables 2 and 3, respectively. Notably, the relative effect estimates pertaining to different comparisons (e.g., tislelizumab vs pembrolizumab, tislelizumab vs nivolumab) are not directly comparable to each other because each STC separately adjusts the RATIONALE-302 population characteristics to the population of each comparator study. Hence, relative effects against different comparators pertain to different target populations.
Overall, for both OS and PFS, most analyses numerically favored tislelizumab, but there was no statistically significant difference between tislelizumab and any of the comparator treatments either before or after population adjustment. Scenario and subgroup analyses concurred with the base case. For TRAEs grade ≥ 3, tislelizumab exhibited a similar safety profile to nivolumab, camrelizumab, and sintilimab (although the point estimates substantially deviated from 1) and was more favorable to camrelizumab when the extended list of effect modifiers was adjusted for under the latest DCO and across all analyses under the early DCO. For tislelizumab and pembrolizumab, TRAEs were not compared because the relevant data for the squamous subgroup were not provided in KEYNOTE-181. The PFS results should be interpreted with caution because of the likely violation of the PH assumption.

Discussion
The objective of this analysis was to derive estimates of relative efficacy and safety of tislelizumab versus other anti-PD-1 treatments for patients with 2L ESCC. In the absence of head-to-head RCT data, this objective was achieved through STCs using IPD from the RATIONALE-302 clinical trial and published aggregate-level evidence for the comparator treatments (nivolumab, pembrolizumab, camrelizumab, and sintilimab). Among the 13 unique studies identified in the SLR, six assessed treatments licensed in the European Union were considered in the ITC feasibility assessment. Except for the RAMONA trial, which evaluated nivolumab with ipilimumab against nivolumab monotherapy, the five remaining studies were largely alike in most aspects and compared anti-PD-1 therapies with similar chemotherapy regimens that could be used for anchored analyses. However, some imbalances in key effect-modifying covariates were noted that warranted the need for population adjustment. We followed a thorough process for determining the effect-modifying covariates that maintained consistency with previous population-adjustment analyses in 2L ESCC [53] but also corroborated their findings through statistical analyses in RATIONALE-302 and clinical opinion.
Overall, tislelizumab performed similarly to other anti-PD-1 treatments in OS, PFS, and TRAEs grade ≥ 3. The only exception was the TRAE results against camrelizumab, which indicated that tislelizumab may be more favorable. The results were generally consistent across base-case, scenario, and subgroup analyses and the two DCOs used from RATIONALE-302.
Our results are also consistent with previous ITCs in 2L ESCC that included tislelizumab among the compared treatments, namely, a network meta-analysis by Gao et al. [54] and an MAIC by Yu et al. [53]. In the analysis by Gao and colleagues, tislelizumab had a statistically significantly better safety profile than camrelizumab and was comparable with nivolumab, while in OS, tislelizumab was found to be comparable to pembrolizumab, nivolumab, camrelizumab, and sintilimab. However, Gao and colleagues found that tislelizumab was more favorable to sintilimab, and this result was not corroborated in our analysis, potentially as a result of the additional uncertainty introduced by the population adjustment. In Yu et al., tislelizumab was compared with camrelizumab, and the results aligned with our study, with the two treatments showing similar efficacy in OS.
The main strengths of our analyses are the thorough evidence identification process and the comprehensive evaluation of the between-study heterogeneity through a detailed feasibility assessment. In addition, our STC implementation builds on TSD 18 [39] and makes further methodologic advancements to ensure that the target estimand remains the marginal effect, which is the most relevant quantity for population-level policymaking.
It should, however, be acknowledged that the produced results are not directly comparable across comparators because each STC considered, by definition, a slightly different population. That is, each STC separately adjusted the RATIONALE-302 IPD to the comparator study enrolled population, which was not the same across comparator studies. The validity of the STC results is contingent on the assumption that the reported baseline characteristics sufficiently account for all effect modifiers; unmeasured confounders could still bias the outcomes. While the included studies were deemed adequately similar for an anchored comparison, residual heterogeneity in their design, regional enrolment, and control regimens introduces uncertainty that may not be fully mitigated by population adjustment. Also, the need to borrow intra-trial covariate correlations from the RATIONALE-302 dataset for the simulation process, a necessary step due to the lack of such data from comparator trials, may not perfectly reflect the true distributions in those populations and represents another source of uncertainty. Furthermore, the conducted analyses relied on the PH assumption, which held for OS but did not seem to hold for PFS. Yet, more advanced analyses, such as fractional polynomials, were not pursued since they introduce challenges with the validity of long-term prediction and cannot accommodate population adjustment. It is also crucial to note that most comparisons did not reach statistical significance, which may be partly due to the reduced statistical power inherent in population-adjustment methods and the wide confidence intervals.
Our study was inevitably subject to data and scope limitations. A primary limitation is the reliance on published aggregate-level data for all comparator trials, as IPD was only available for RATIONALE-302, which limits the ability to provide inference for any relevant population and introduces assumptions regarding the target population. Additionally, the trial populations were predominantly of Asian descent, which may influence the generalizability of these findings to other racial or ethnic groups. The scope of this analysis was limited to immunotherapies licensed in the European Union, thereby excluding other potentially relevant global or emerging treatments, which could affect the broader applicability of the findings. Finally, as the systematic literature search was last updated in October 2023, any trials more recently published are not included in this analysis.

Conclusions
In RATIONALE-302, tislelizumab achieved a survival benefit over chemotherapy across all prespecified subgroups, including region, race, and PD-L1 expression level [26]. Compared with other anti-PD-1 therapies (nivolumab, pembrolizumab, camrelizumab, and sintilimab), tislelizumab performed similarly in OS, PFS, and TRAEs of grade ≥ 3 for patients receiving 2L treatment for ESCC. However, tislelizumab may be associated with a more favorable TRAE grade ≥ 3 safety profile than camrelizumab, though this finding should be interpreted with caution because of the lack of direct information. Results were consistent across primary, sensitivity, and subgroup analyses. Conclusions should be interpreted with caution considering the uncertainty surrounding the estimates and the limitations inherent to the methodology employed to obtain the indirect comparisons.

Supplementary Information
Below is the link to the electronic supplementary material.