Comparative effectiveness of nivolumab versus clinical practice for advanced gastric or gastroesophageal junction cancer

This study was conducted by Precision Xtract (Vancouver, Canada) and was funded by Bristol-Myers Squibb. It has been reported in accordance with the REporting of studies Conducted using Observational Routinely-collected Data (RECORD) and STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) statements [9,10].

Two different methodological approaches were used to achieve the study objective (Table 1). In the main approach, simulated treatment comparison (STC) methodology was used to estimate the comparative effectiveness between nivolumab, as evaluated in the CheckMate 032 trial, and RCP, as available in the Flatiron Health database, with respect to overall survival. In an STC, outcomes are modeled for a treatment of interest as a function of relevant patient-related factors developed from the individual patient data (IPD) of an index trial. This model is then used to predict outcomes that would be expected in another population [11]. The second approach was exploratory and involved an adjustment of the overall survival outcomes, as reported in the ATTRACTION-2 trial, by reweighting the trial’s patient characteristics to match those of a US-based cohort.

To support variable selection in the development of statistical models, literature reviews were conducted to identify potential prognostic factors and treatment effect modifiers. Studies of any design, that aimed to identify prognostic factors, were eligible for the systematic literature review of prognostic factors, while a targeted review of effect modifiers was restricted to randomized controlled trials (RCTs).

Using indirect and population reweighting approaches to estimate treatment effect in a US population, our study showed a beneficial effect of nivolumab over RCP for the treatment of individuals with gastric or GEJ cancer refractory to or intolerant of two or more previous chemotherapy regimens. The relative effect size from our main STC model, HR: 0.50 (95% CI: 0.36–0.68), was of similar magnitude to the reweighted analysis of the ATTRACTION-2 RCT (HR: 0.57; 95% CI: 0.26, 1.23) and was robust to alternate modeling assumptions, as explored in the sensitivity analyses.

There are limited effective treatment options for patients with advanced gastric or GEJ cancer who have failed multiple lines of therapy. Until recently, there were no US FDA-approved or NCCN-recommended therapies for third or later lines of therapy [21]. Pembrolizumab is the only FDA-approved therapy, having been granted accelerated approval in late 2017 based on a favorable benefit–risk profile. The FDA considered that the modest yet durable response observed among a subset of PD-L1-expressing patients in the single-arm KeyNote 059 trial was predictive of a meaningful benefit to patients in the form of improved survival [22,23]. The findings of our study demonstrating that nivolumab is associated with a doubling in median overall survival relative to a US-based cohort of patients receiving RCP provides further support for the use of PD-L1 checkpoint inhibitors in heavily pretreated patients with advanced gastric or GEJ cancer. Although safety and health-related quality-of-life were not in the scope of this research, previous interpretation of the ATTRACTION-2 RCT suggests that the safety profile of nivolumab is manageable and similar to that reported in patients with other advanced solid tumors [24,25]. In that RCT, treatment-related adverse events occurred with a frequency that was similar to placebo. If this relationship holds across populations, then the benefit of nivolumab could be obtained without the additional cost of adverse events.

The beneficial effect of nivolumab over RCP observed in the Asia-based ATTRACTION-2 RCT has been used to support regulatory licensing in Japan, where nivolumab is indicated for the treatment of unresectable advanced or recurrent gastric cancer, which has progressed after chemotherapy. Although our reweighted analysis could not adjust for all immune and inflammation signatures that may differ between Asian and non-Asian populations, this study provides insights into the relative effect size that might have been observed had the RCT been conducted in a non-Asian setting because of the adjustments for other key prognostic and potentially effect modifying differences between the populations such as age and tumor location [5,7]. The trend toward a more protective effect in the US population compared with the Asian setting aligns with subgroup analyses from KeyNote 061 and ATTRACTION-2, which demonstrated trends toward larger treatment effect for immunotherapy among patients with GEJ tumors, which tend to be more common in the USA than in Asia [7,26].

The use of real-world evidence confers a unique advantage in that the dataset can be specified to any subset of gastrointestinal cancer patient. Because of this, we were able to generate a cohort representative of the CheckMate 032 trial to facilitate our STC approach. Our approach serves as a framework for estimating the comparative effectiveness between treatments by merging trial data and real-world evidence. Data source and cohort definition are important aspects of any pharmacoepidemiological study design and constitute a key element in the FDA’s recently released framework for evaluating real world evidence [27]. The Flatiron Health database was selected for this study as it is comprehensive and research-oriented, representing over 200 cancer practices and practice groups providing care for over 750,000 patients with cancer in the USA. Several approaches to deriving patient cohorts were considered and three were evaluated for the purposes of this study to explore potential threats to internal validity due to selection bias. While analyses based on the three cohorts evaluated in this study generated consistent conclusions, there were some variations in the patient characteristics which may have impacted survival outcome estimates and the generalizability of findings. The main analysis was based on the random cohort, which selected patients weighted to match the frequencies of lines of therapies observed in the CheckMate 032 trial. The alternative cohorts, evaluated in sensitivity analyses, were considered to carry some bias. In the prevalence cohort, patients were more likely to have been initiating a third line of therapy at cohort entry, compared with the line of therapy at the baseline of the CheckMate 032 trial. This may have distorted comparability with the patients enrolled in CheckMate 032 toward longer survival times in the RCP cohort and, consequently, estimated a smaller magnitude of treatment effect. Conversely, selection of patients for the last-line cohort was based on knowledge that patients received no further lines of therapy. This may have biased the cohort toward a shorter time-to-death and estimated a larger magnitude of treatment effect. It is essential that real-world evidence-based analyses are thoroughly vetted through sensitivity analyses and that potential limitations and the direction and magnitude of biases are explicitly evaluated.

There are several strengths to this study. A comprehensive and systematic approach was taken to identify prognostic factors, coupled with a targeted review to identify potential treatment effect modifiers, thereby leveraging the literature to better inform our analyses. Given the novelty of the methods used in our study and the assumptions that were made, such as selection of cohorts from real-world evidence, we conducted a series of alternate and sensitivity analyses. Conclusions were broadly consistent irrespective of the cohort definition and supported the findings of our primary analyses. Moreover, our study was intended to complement the findings of studies that have used other methodological approaches, such as a recent study conducted using a frequency-matched approach, as well as the adjusted Cox model and population-adjusted analysis of ATTRACTION-2 presented herein [13]. Again, conclusions were consistent across methodological approaches, and all demonstrated a similar trend toward longer overall survival associated with nivolumab. However, the methodological literature for STCs is sparse, particularly in the context of time-to-event survival data [28,29]. Therefore, our methodological approach featured several new innovations and multiple different analyses to explore the consistency of our findings. First, we used smooth splines in the STC methodology, which led to more stable estimations that were less affected by variability in hazards over time, especially given small sample sizes. Second, we applied both proportional hazard and time-varying hazard approaches, and used both frequentist and Bayesian frameworks. Third, we estimated relative effect size using both the Flatiron and the CheckMate 032 patients as target populations.

Despite the rigorous approach taken in model development and the evaluations of robustness, there are additional limitations to consider. First, the comparisons made in our study were outside of a randomized trial and, thus subject to bias. For example, while we took a rigorous approach to identify prognostic factors and effect modifiers, we were limited by the quality of the evidence identified. Second, the STC represented an ‘unanchored’ comparison, for which the major assumption is that all effect modifiers or prognostic variables are accounted for. We were only able to account for measured factors and could not account for unmeasured covariates or capture the complex interactions that may exist. Therein lies the major assumption of the unanchored STC that all effect modifiers or prognostic variables are accounted for. That assumption could not be formally tested; however, given the large treatment effect size and the degree of prognostic significance of other variables identified in our systematic review but not included in our analysis, we expect that imbalances in prognostic factors between the nivolumab and RCP cohorts would not be sufficiently large to reverse the direction of association [30,31]. However, this approach was necessary as there were no common comparators. While the RCP comparator group included patients who received active treatment, heavily pretreated patients may decline further active therapies and may instead opt for palliative care. While these patients were not included in the comparator group of our STC, their inclusion may have led to even larger estimated treatment effect sizes. Finally, while the evidence applied to our models was robust, analyses were ultimately based on small-sample sizes.

Despite the absence of head-to-head data and the inherent limitations of using nonrandomized comparisons of clinical trial data and real-world evidence, this study suggests that using nivolumab for third or later lines of therapy may confer an overall survival benefit relative to RCP for patients with advanced gastric or GEJ cancer in the US setting. The methodological approach taken in this study serves as a framework for estimating the comparative effectiveness between treatments using single-arm trial data and real-world evidence.

To view the supplementary data that accompany this paper please visit the journal website at: Supplementary Material

Each author listed above participated sufficiently in the work to take responsibility for the content, and that all those who qualify are listed. Each author provided substantial contributions to the conception or design of the work; or the acquisition, analysis or interpretation of data for the work; and drafting the work or revising it critically for important intellectual content; and final approval of the version to be published; and agreement to be accountable for all aspects of the work are appropriately investigated and resolved.

The authors acknowledge National Health Service funding to the National Institute for Health Research Biomedical Research Centre at the Royal Marsden NHS Foundation Trust and The Institute of Cancer Research. D Ayers, S Goring and S Cope are employees of or subcontractors to Precision Xtract; Precision Xtract was paid by Bristol-Myers Squibb for conducting this work. B Korytowsky and P Abraham are employees of Bristol-Myers Squibb. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Medical writing assistance was provided by Michael Zoratti.

The authors state that they have followed the principles outlined in the Declaration of Helsinki for all human or animal experimental investigations.

The authors certify that this manuscript reports the secondary analysis of clinical trial data that have been shared with them, and that the use of this shared data is in accordance with the terms agreed upon their receipt. The source of this data is: Bristol-Myers Squibb and Flatiron Health.

This work is licensed under the Attribution-NonCommercial-NoDerivatives 4.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc-nd/4.0/