Relative efficacy of interventions in the treatment of second-line non-small cell lung cancer: a systematic review and network meta-analysis

A systematic literature review was undertaken to identify all relevant publications of phase 2/3 randomized controlled trials that were conducted in adult patients who had locally advanced or metastatic NSCLC and whose disease had progressed after first-line chemotherapy through 2 September 2015. There were no date, language, or geographical restrictions on the medical reference database searches. The search strategy involved searches of MEDLINE (using PubMed platform), MEDLINE In-Process (using PubMed platform), Embase (using Elsevier platform), Biosciences Information Service (using the Dialog Platform) and the Cochrane Library (using the Wiley platform). Additionally, the American Society of Clinical Oncology, the European Society of Medical Oncology, and the International Association for the Study of Lung Cancer were searched for relevant new evidence. The search included regimens containing the following interventions: docetaxel (any dose), erlotinib (150 mg), gefitinib (250 mg), gemcitabine (any dose), nintedanib (200 mg), nivolumab (3 mg/kg), pembrolizumab (any dose), pemetrexed (500 mg/m²), ramucirumab (10 mg/kg), vinorelbine (any dose), and best supportive care. Two further interventions also were included from this literature review due to growing interest in the Asian market. These were S-1 (40 mg/m²) and bevacizumab (15 mg/kg). After the systematic literature review had been completed, the trial results for pembrolizumab [18] were published online in December 2015 and made fully available in April 2016. That study was considered in addition to the 30 clinical trials identified for inclusion in the NMA.

Two independent reviewers performed the screening; where there was any uncertainty about inclusion, the study was checked by a third researcher. Data were extracted from full-text versions of studies, where available (i.e., abstracts or posters were not used unless an abstract or poster was the terminal source document). Where mature data from a single trial were reported later, the most recent data were used in the analysis.

The outcomes studied were OS and PFS, which were evaluated in terms of relative mean survival compared to the reference treatment (docetaxel 75 mg/m²). Although clinical trials typically report hazard ratios and median survival times, the fitting of parametric models to survival data lends itself more readily to the estimation of mean survival. Mean survival times are of particular importance to cost-effectiveness analyses, which are part of the health technology assessment process. Mean survival was estimated as the area under the probability of survival curve with a horizon of 60 years. Relative differences were assessed and mean estimates with 95% credible intervals presented. In addition, significance tests were performed on the difference between two distributions. The minimum of the proportions greater or less than 0 was used as a probability and converted to a two-tailed test by multiplying it by 2 [19]. These values can be interpreted in the same way as P values from a frequentist analysis, although they do not have the same statistical definition. Rank probability charts, cumulative probability charts and surface under the cumulative ranking (SUCRA) were also derived following the methods described by Salanti et al. [20]. However, caution is needed when interpreting SUCRA scores or associated rankings because they can be biased when interventions are not represented equally in a network and/or, interventions are not connected to the same reference treatment [21, 22]. The efficacy of each intervention was evaluated for each combination of NSCLC factors, namely EGFR mutation positive (versus EGFR mutation negative), squamous versus nonsquamous, and PD-L1 expression < 5% versus PD-L1 expression ≥5% (PD-L1 expression < 1% vs. PD-L1 expression ≥1% was also studied, but only the full results for the 5% cut-off are presented), which together defined 16 distinct subgroups. Studies investigating TKI regimens included only TKI-naïve patients, and studies investigating nivolumab included only PD-L1 immunotherapy-–naïve patients.

Kaplan-Meier charts were digitized using the DigitizeIt software package [23], and the method described by Guyot et al. [24] was used to reconstruct patient-level data. All of the data were plotted to check for evidence of nonproportional hazards and tests of significance performed using the method described by Grambsch and Therneau [25] based on a Kaplan-Meier transformation. The method presented by Jansen [12] was used to fit first- and second-order fractional polynomial models. In addition, hazard ratios and confidence intervals also were extracted and summarized as log (hazard ratios) and standard errors. The log (hazard ratios) were used to estimate Higgin’s I² [26], which provides a measure of the percentage of variance explained by heterogeneity and/or inconsistency. Pairwise meta-analyses of duplicate comparisons and two types of node-splitting (traditional node-splitting [27], which detects only inconsistency for a closed loop in the network of evidence, and an experimental node-splitting [28], which assumes inconsistency is another measure of heterogeneity) also were conducted. The log (hazard ratios) were used to perform NMAs; this followed the method described by Woods et al. [29]. If a study did not provide complete information on hazard ratio estimates, the hazard ratios were estimated using the reconstructed patient-level data where possible; if not available, the hazard ratios were derived from median survival times and sample size following the method described by Hackshaw [30]. All the NMAs were extended using hierarchical exchangeable structures by adapting the code presented by Owen et al. [31]. Hierarchical exchangeable structures allow treatment effects to vary by covariates independently of the other treatments in the network of evidence. The treatment effect remains constant for any treatment not specified within a hierarchical exchangeable structure. The efficacy of EGFR TKIs was allowed to vary by EGFR mutation status (evidence supported by Wang et al. [32]; Lim et al. [33]; Sun et al. [34]; Urata et al. [35]; Vale et al. [17]). Further, pemetrexed was allowed to vary by histology (supported by evidence from Kubota et al. [36]; Scagliotti et al. [37]), nintedanib was allowed to vary by histology (supported by evidence from Reck et al. [38]), and nivolumab was allowed to vary by PD-L1 expression and histology (supported by evidence from Borghaei et al. [39]; Brahmer et al. [40]). Predictions were made for all combinations of these subgroups. In addition, where possible, different doses also were included as a hierarchical structure with an overall treatment class effect. Constraints were imposed to ensure that the efficacy increased with dose intensity (docetaxel 75 mg/m² every 3 weeks → docetaxel 60 mg/m² every 3 weeks → weekly lower-dose docetaxel) and where there was external supporting evidence for the treatment covariate interactions (EGFR TKIs with EGFR mutation status and pemetrexed with histology). Where there was sufficient evidence for subpopulation results within a trial, these were used instead of the overall study-level results. Subgroup data were used in the NMA only where we had a reason to suspect there should be a difference between subgroups, or there was supporting evidence in the literature, or a treatment had an indication for a subgroup. Studies that did not present the relevant proportion of patients with a given tumor subtype were not included in the NMA, i.e., pemetrexed and proportion of patients with nonsquamous tumors not reported, TKIs and proportion of patients with tumors that were EGFR mutation positive not reported, and PD-L1 immunotherapies with proportion of patients with PD-L1 expression not reported. The following study-level covariates were included one at a time to see if they improved model fit: time since publication, proportion of patients with nonsquamous tumors, mean age, proportion of patients with Eastern Cooperative Oncology Group (ECOG) ≥ 1, proportion of patients with metastatic disease, and proportion of Asian patients.

Noninformative priors were used for treatment efficacy and for between-study variances. For the fractional polynomial model, four random-effects models were conducted: first-order fractional polynomial NMA with an additional heterogeneity parameter for scale, first-order fractional polynomial NMA with additional heterogeneity parameters for scale and shape, second-order fractional polynomial NMA with an additional heterogeneity parameter for scale, and second-order fractional polynomial NMA with additional heterogeneity parameters for scale and the two shape parameters. A range of power functions were fitted for first- and second-order fractional polynomial models, which followed that of Jansen [12].

Posterior sampling was performed using Markov Chain Monte Carlo, with three chains of 70,000 iterations per chain, each with a burn-in of 30,000 iterations for the fractional polynomial models and a total of 120,000 iterations for the hazard ratio models. Convergence was assessed through iteration plots, the shape of the posterior distributions, and Gelman-Rubin diagnostics [41]. Bayesian NMAs were performed using JAGS software [42] within R 3.1.3 [43]. Heterogeneity tests were performed using the netmeta package in R [44] and node-splitting was performed using the geMTC package in R [45]. Further details and the JAGS code are provided in the Additional file 1.

Further analyses were required to estimate mean survival times. For the hazard ratio approach, all of the reconstructed patient-level data for the reference data were collated. Bayesian first-order and second-order fractional polynomial models were fitted to these data with heterogeneity parameters for all scale and shape parameters. The same number of iterations and chains was used as stated for the hazard ratio NMA. A grid search for the best-fitting model with up to two covariates was performed.

The efficacy of study treatments in each tumor subgroup was estimated by linear extrapolation of the parameters available for each proportion of the relevant tumor type. If not all of the required information was available, the efficacy of a combination regimen was estimated using the separate information from monotherapy regimens.

We screened 2601 records and 30 trials evaluating 17 interventions for inclusion in the NMA (Fig. 1 and Table 1). Characteristics of included studies are provided in Table 1. The network diagrams for OS and PFS are shown in Figs. 2 and 3, respectively. A bias assessment of the included studies is presented in Additional file 1: Table S1. The bias assessment showed that a high proportion of studies were open label, with physicians and patients not blinded to treatment. An additional cause of concern was treatment switching in the study by Kawaguchi et al. [46], which was the only study that connected docetaxel (60 mg/m²) to the rest of the network. This study used erlotinib (150 mg) as the comparator in a patient population with 16% EGFR mutation-positive tumors in the erlotinib arm and 25% in the docetaxel arm. The treatment switching, therefore, was likely to have caused confounding by potentially increasing the perceived relative efficacy for docetaxel (60 mg/m²), although the use of constraints prevented this reversal from being observed. The data for OS, including hazard ratios, median survival times, and covariate data, are presented in Additional file 1: Table S2. The covariate data from this table shows some differences between studies for the proportion of patients restricted in activity (performance status (ECOG) ≥ 1 = 0.48–0.94), stage of disease (stage IV = 0.53–1.0), and proportion of Asian patients (0.01–1.0); although none of these covariates appeared to improve the fit of the NMAs.

The study designs and reporting of trials studying PD-L1 immunotherapies also was an issue. Brahmer et al. [40] and Borghaei et al. [39] studied nivolumab versus docetaxel in patient populations without any restriction on PD-L1 expression and presented the results for < 1% versus ≥1%, < 5% versus ≥5%, and < 10% versus ≥10%. However, Herbst et al. [18] included only patients with PD-L1 expression of ≥1% and results were stratified as 1 to 49% and ≥ 50%. The approach undertaken for the NMA relied on being able to estimate the efficacy of regimens for the full range of patient types; in the absence of information on how pembrolizumab performed in a low PD-L1 expression population, the Herbst study was excluded from the NMA. This left us with the choice for which of the cut-offs presented by Brahmer et al. [40] and Borghaei et al. [39] to present in this study. The results for the 5 and 10% cut-offs were closer to each other and appeared to be more consistent than the 1% cut-off results, i.e., efficacy improved for both OS and PFS, with higher PD-L1 expression values for both studies, for the 5 and 10% cut-offs, whereas the 1% cut-off results reported by Brahmer et al. [40] appeared to show a reversal in this pattern, although this was not statistically significant. The 5% cut-off data therefore were selected because they represented the mid cut-off presented and appeared to give more consistent results. The overall results in which nivolumab showed a significant benefit over docetaxel (150 mg) were not affected by the choice between the 1 and 5% cut-offs for either the fixed-effects or random-effects models for OS and PFS.

An assessment of the proportional hazards assumption and heterogeneity and inconsistency is provided in the Additional file 1. The assessment indicated that the proportional hazards assumption was mainly an issue for PFS (2 out of 32 studies had nonproportional hazards for OS, and 7 out of 29 studies had nonproportional hazards for PFS) and that the exchangeable hierarchical structures were needed for all the NMAs fitted. The main evidence for the difference in efficacy of TKIs in the network of evidence came from the subgroup data presented by Urata et al. [35]. For OS, these data were presented only as median survival times compared to all Kaplan-Meier estimates available for PFS. The fitting of fractional polynomial models for OS was problematic because these models were not able to differentiate between the efficacy of TKIs for EGFR mutation-negative and mutation-positive tumor types. However, the hazard ratio NMAs were able to demonstrate a difference in efficacy of TKIs by EGFR status; and because the evidence for nonproportional hazards for OS was weak, the results presented for OS are those based on the hazard ratio NMA. The hazard ratios for OS were applied to fractional polynomial model fitted to the reference treatment (docetaxel 75 mg/m²) from the network of evidence. Details of this model are given in the Additional file 1 and predicted survival curves by tumor histology presented in Additional file 1: Figure S1. The fractional polynomial model for PFS did not have these issues, and so this was the model used to present PFS results. For both these NMAs, the random-effects models did not improve model fit over the fixed-effects results; only the fixed-effects results are presented in the main body of the text; random-effects results are presented in Additional file 1: Tables S3 and S4 for OS and PFS, respectively, for more details.

The fitting of fractional polynomial models to multiple treatment data can result in survival curves that flatten before reaching zero, especially for second-order models with time-varying treatment effects. (All predictions by treatment are shown in Additional file 1: Figure S2.) This was observed in the NMAs fitted using this technique. The hazard rates therefore were estimated up to the end of follow-up for the available evidence by intervention; after this point had been reached, the hazard rates from the reference treatment were used.

The results for the eight subgroups are presented in the following subsections and in the Additional file 1. Only main figures for the subgroup representing the largest population (nonsquamous, EGFR mutation negative, and PD-L1 expression < 5%) are presented in the manuscript. The full set of results for each subgroup can be in found in the Additional file 1: Figures S3–S48). The Additional file 1 provides survival probability curves, mean survival times, and relative differences in mean survival for OS and PFS, by subgroup. A summary table for all the results by subgroups is provided in Table 2. SUCRA scores are not presented due to apparent bias with this network of evidence and because they were found to not be consistent with relative differences and credible intervals. This was particularly the case for treatments represented by a small sample size and located on the periphery of the network of evidence.