PD-1 inhibitors combined with paclitaxel and cisplatin in first-line treatment of esophageal squamous cell carcinoma (ESCC): a network meta-analysis

We conducted a comprehensive search of Medline, Web of Science, Pubmed, the Cochrane Library and Embase databases up to April 2023. In addition, we systematically reviewed all abstracts from the American Society of Clinical Oncology (ASCO) and the European Society of Medical Oncology (ESMO) Congress between 2012 and 2023. Our search strategy was designed to identify published randomized controlled trials that evaluated first-line treatment options for patients with esophageal cancer.

All RCTs assessing first-line treatments for esophageal cancer were included in this systematic review, without any restrictions on publication date, location or language. Eligible studies had to meet the following criteria: 1) prospective phase III randomized controlled trials; 2) included patients with metastatic, unresectable, or recurrent squamous of the esophagus; 3) first-line treatment setting; 4) compared at least two arms that consisted of the following agents: FP(5-fluorouracil, cisplatin), TP (paclitaxel, cisplatin), and PD-1 inhibitors.

Two authors (Zhao and Zhang) independently scrutinized the titles and abstracts of retrieved RCTs. They reviewed the full texts of selected RCTs to evaluate eligibility criteria for inclusion in NMA, extracted study characteristics and outcome data. Qu resolved any disagreements between the authors if necessary.

The Cochrane tools were utilized to evaluate the quality and risk of bias. HRs with their corresponding 95% CI for PFS and OS were extracted from various studies, while OR and its 95% CI were employed to indicate the frequencies of ORR. In cases where data on HR and its 95% CI could not be obtained, Kaplan–Meier curves were digitized using Engauge Digitizer (www.​digitizer.​sourceforge.​net) followed by hazard ratio calculation in R.

Pairwise meta-analyses were conducted using the JAGS ‘Gemtc’ package in R software (version 4.0.3) [16, 17]. Heterogeneity among studies was evaluated by means of the Q test and I² statistic [18]. The fixed-effect or random-effects model was selected based on the value of I² (< 50% or > 50%, respectively). Results from pairwise meta-analysis were presented as HR with 95% CI for OS and PFS, and as OR with 95% CI for ORR and ≥ 3TRAEs. Network plots were generated using R (version 4.0.3) to compare different treatments and depict network geometry. Network meta-analysis (NMA) was performed under the Bayesian framework using JAGS and the "gemtc" package in R [17]. Both random effects and consistency models were utilized in NMA, with four independent Markov chains automatically generated for posterior distribution estimation through 5000 adaptation iterations and 20,000 inference iterations per chain.Run lengths were extended if the Brooks-Gelman-Rubin diagnostic or time series plots indicated that the Markov chains had not converged (Supplementary Fig. 2). The NMA results were presented as HRs with 95% credible intervals (CIs) for OS and DFS, and ORs with 95% CIs for ORR and ≥ 3TRAEs. The probability of each treatment regarding survival outcomes was ranked according to the HRs and the posterior probabilities. Two-sided p < 0·05 indicates statistical significance.

Consistency, global inconsistency assessment, and local inconsistency assessment were conducted to evaluate the study. The evaluation of global inconsistency was based on comparing the fit of consistency and inconsistency models using deviance information criterion (DIC). Similar DIC values among different models indicate good consistency [19, 20]. The local inconsistency was evaluated through comparing the direct and indirect evidence generated under the Bayesian framework using the node-splitting analysis, where p < 0·05 indicates significant inconsistency [21].

Eventually, after screening 1786 articles, only five RCTs met the eligibility criteria for this network meta-analysis (NMA) [4, 11‐14] (Fig. 1). However, due to the inability to compare treatments simultaneously in a network among these RCTs, we identified a retrospective control study during our search that could link RCTs to a network and provide data on PFS, OS, ORR, and ≥ 3TRAEs [22]. Therefore, this study was included in our NMA. All clinical trials included in the present meta-analysis are listed in Table 1. A total of 3360 subjects were included, all of them received first-line treatment. Four treatment regimens (PD-1 + TP, PD-1 + FP, TP, FP) were included. KEYNOTE-590 and CheckMate-648 were compared with PD-1 + FP and FP. In KEYNOTE-590, the eligible patients had “unresectable or metastatic adenocarcinoma or squamous cell carcinoma of the esophagus or Siewert type 1 gastro-esophageal junction adenocarcinoma” [11]. In this NMA, only the squamous cell sub-group was included. CheckMate-648 included three arms, among which, the third arm (I + P) did not meet the standards, so was not included. The comparison between PD-1 + TP and TP was conducted via Escort1-st and JUPITER06 studies. In Orient-15, TP was administered as the base-line regimen to 94.5% of subjects. Therefore, the two arms of Orient-15 were considered PD-1 + TP and TP in this NMA. The hazard ratios (HRs) for the PFS and OS were not reported by Liu Y and Ren Z. The Kaplan–Meier curves were digitized using Engauge Digitizer (www.​digitizer.​sourceforge.​net), and HRs were calculated in R. The OS and PFS were analyzed across all six trials, while only five trials were included for the ORR and ≥ 3TRAEs. Data on the ORR in the squamous subgroup were not provided in Keynote590.

The primary outcome was the OS, while secondary endpoints included the PFS, ORR and ≥ 3TRAEs. Study quality was assessed using the Cochrane Risk of Bias tool, version 5.1.0, with items scored as low, high or unknown risk of bias. Based on I² < 50%, there was no significant heterogeneity observed among the trials in the network.

A network consisting of four treatments, namely PD-1 + TP, PD-1 + FP, TP, and FP (Fig. 2), was established. The results of the comparison among all treatments are summarized in Table 2. Given that there was no statistical heterogeneity in this network (I² = 0%), a fixed effects model was adopted to report the findings. In terms of PFS, PD-1 + TP demonstrated superior efficacy compared with other treatments. The HRs for PD-1 + TP compared to PD-1 + FP, TP and FP were 0.59 (95% confidence interval [CI]: 0.44–0.80), 0.56 (95% CI: 0.51–0.61), and 0.45 (95% CI: 0.37–0.56), respectively (Table 2A). In terms of the OS, PD-1 + TP exhibited superior efficacy compared to other treatments. The HRs for PD-1 + TP compared to PD-1 + FP, TP and FP were 0.74 (95% CI: 0.56, 0.96), 0.64 (95% CI: 0.57, 0.71), 0.53 (95% CI: 0.43, 0.67) respectively (Table 2A). Due to the lack of data on the ORR in the squamous subgroup in Keynote590, only five trials were included for analysis of the ORR and ≥ 3TRAEs (Supplementary Fig. 1). No significant difference in the ORR was observed between PD-1 + TP and PD-1 + FP, as indicated by RRs of 1.2 (95% CI:0.69, 2.11). However, PD-1 + TP showed a significant advantage over TP and FP with ORs of 2.5 (95% CI: 2.04, 3.04) and 2.95 (95% CI: 1.91, 4.63), respectively (Table 2B). For security, PD-1 + TP had no significant difference with PD-1 + FP and TP, ORs of ≥ 3TRAEs were 0.75(95% CI:0.41,1.4) and 0.89(95% CI:0.73, 1.1). The results suggest that FP may have had superior security, with an OR of 2.23 (1.35, 3.69) (Table 2B). Forest plots were generated to visualize the network estimates (Fig. 3).