Surrogate Endpoints for Overall Survival in Immune-Oncology Trials of Advanced Gastro-Esophageal Carcinoma.

Background: We aimed to assess whether the Response Evaluation Criteria in Solid Tumors (RECIST)-based objective response rate (ORR), disease control rate (DCR) and progression-free survival (PFS) could serve as surrogate endpoints for overall survival (OS) in immune-oncology (IO) trials of advanced gastro-esophageal (GE) carcinoma.
Methods: Randomized controlled trials (RCTs) of IO that reported RECIST-based endpoints and OS in advanced GE carcinoma were screened. Surrogacy of endpoints for OS was assessed based on the correlation between endpoints with OS (arm-level), and between treatment effects on endpoints (trial-level). The correlations were quantified by Pearson correlation coefficient (R). Leave-one-out cross-validation was used to assess the prediction accuracy of surrogate model.
Results: Seventeen RCTs (9,657 subjects) with 20 comparisons were included. The correlations between DCR and OS were not strong at arm- (R = 0.80) and trial-levels (R = 0.45), but strong correlations between ORR (R = 0.91), PFS (R = 0.89) and OS at arm-level were observed. Treatment effect on ORR and PFS (both R = 0.71) was moderately correlated with treatment effect on OS. Leave-one-out cross-validation approach further validated the surrogacy of PFS. Our analysis showed that 3-month PFS could reliably predict 6-month OS, 6-month PFS could reliably predict 12-month OS, and 12-month PFS could reliably predict 18-month OS. The conservative minimum threshold effect of HRPFS was 0.73. Conclusions: PFS may be the appropriate surrogate for OS in IO trials of GE carcinoma. A conservative minimum threshold effect of HRPFS ≤ 0.73 has the potential to predict a significant improvement in OS. Copyright 2022, Li et al.

Introduction

Despite that the incidences of gastric and esophageal carcinoma are broadly declining over the past decades, they remain the fifth (5.7% of total) and seventh (3.2% of total) most common cancer worldwide, respectively. According to the GLOBOCAN 2018 database, gastric and esophageal carcinoma, in total, accounted for 13.5% of all cancer deaths worldwide [1]. Patients with gastro-esophageal (GE) carcinoma commonly have advanced or metastatic disease at initial diagnosis [2, 3], and the treatment strategy is characterized by the use of cytotoxic regimens. However, although several randomized trials have demonstrated that advanced or metastatic GE carcinoma could benefit from systemic chemotherapy, the prognosis of GE carcinoma patients remains dismal, with a median overall survival (OS) of approximately 12 months [4-7]. Therefore, novel drugs are needed to improve clinical outcomes [8].

Over the past decades, immune checkpoint inhibitors (ICIs) that block the programmed death-1 (PD-1) axis have shown promising therapeutic efficacy in various solid tumors, including GE carcinoma [9-11]. So far, ICIs have shown superior survival over chemotherapy as first and later line treatment in advanced GE carcinoma [12-18]. Nonetheless, approximately 40% of GE carcinoma patients treated with ICIs still suffer from intrinsic or acquired drug resistance, and many immune-oncology (IO) trials are required to further improve their prognoses. To accelerate the approval of effective ICIs, development of surrogate endpoint for OS is an optional but promising strategy. In the era of chemotherapy, the conventional RECIST-based endpoints have been widely applied to reflect the antitumor activity and validated as the robust surrogacy for OS in advanced GE carcinoma trials [19]. However, ICIs have distinct mechanisms of action (e.g., delayed clinical benefit [20], pseudoprogression [21] and hyper-progression [22]). Previous meta-analyses have shown that the conventional RECIST-based endpoints cannot serve as a primary endpoint for OS in pan-cancer IO trials [23, 24]. Nonetheless, significant heterogeneities among different solid tumors limit these applications in the IO trials of advanced or metastatic GE carcinoma.

Therefore, we used arm- and trial-level quantitative approaches to evaluate, for the first time, the correlation between RECIST-based endpoints (including progression-free survival (PFS), objective response rate (ORR) and disease control rate (DCR)) and OS in randomized controlled IO trials of GE carcinoma.

Materials and Methods

Search strategy and study selection

Two authors (RCN and YW) independently searched Medline (PubMed), Web of Science, Embase, ClinicalTrials.gov and Cochrane Library databases for eligible trials from January 1, 2000 to September 30, 2021, using the following search terms: nivolumab, pembrolizumab, avelumab, atezolizumab, durvalumab, PD-1, PD-L1, checkpoint inhibitors, gastro-esophageal carcinoma and randomized controlled trial. Supplementary Material 1 (www.wjon.org) shows the detailed search terms. Randomized controlled trials (RCTs) investigating anti-PD-1/programmed death ligand-1 (PD-L1) therapy in advanced GE carcinoma that reported treatment effect (hazard ratios (HRs)/odds ratios (ORs)) on OS and surrogate endpoints (PFS/ORR/DCR) were included. We excluded reviews, abstracts, case reports and studies with sample size less than 150 subjects. Conference abstracts of the 2021 American Society of Clinical Oncology (ASCO) annual meeting and the European Society for Medical Oncology (ESMO) Congress 2021 were manually searched to retrieve eligible trials.

Data extraction and endpoints

The following data for each eligible trial were extracted: population, study phase, treatment protocol, sample size, primary endpoint, results of OS and surrogate endpoints (PFS, ORR and DCR). For trials reporting on multiple populations, the largest population with reported primary endpoints was included. The survival rates of OS and PFS at different cut-off time points (3, 6, 9, 12, 15, 18 and 24 months) were measured using the Engauge Digitizer tool V.12.1 (http://markummitchell.github.io/engauge-digitizer/). The HRs for OS and PFS at different cut-off time points were calculated using the Kaplan-Meier curves, according to the description by Parmar et al [25]. OS was defined as the time from randomization to death from any cause. PFS was defined as the time from randomization to disease progression or any death. ORR was defined as the proportion of best confirmed complete response (CR) or partial response (PR). DCR was defined as the percentage of best-confirmed CR, PR or stable disease (SD).

Statistical analysis

Our quantitative evaluation used two correlation approaches (arm- and trial-level) to assess the potential surrogate endpoints for OS, as previously described [26, 27]. The strength of association between the surrogate endpoints (median PFS, ORR and DCR) and median OS of each experimental arm (arm treated with ICIs) at the arm-level was assessed. The correlation between HRs for PFS and ORs for ORR/DCR and HRs for OS at the trial-level was assessed via a linear regression model, weighted by trial arm or trial size. The sample size of trials that reported multiple arms was down-weighted based on the descriptions of A’Hern et al [28]. The arm- and trial-level correlations were quantified by weighted Pearson correlation coefficient (R). According to the criteria of the Institute for Quality and Efficiency in Health Care (IQWIG) [29], the strength of association between endpoints was categorized as weak (R < 0.70), moderate (R = 0.70 - 0.85) and strong (R > 0.85), based on the value of R.

For each meta-analysis, we used the leave-one-out cross-validation analysis to assess the prediction accuracy of the surrogate model [30]. Each trial was left out once and the surrogate model was built using the remaining trials. Predicted HRs for OS with 95% prediction intervals were calculated from the observed HR of PFS of that particular trial. To demonstrate typical conditions, the strength of associations between HRs for 3, 6, 12, and 15-month PFS, and HRs for 6, 12, 18, and 24-month OS were calculated, and several subgroup analyses of tumor type, trials line, treatment strategy and follow-up duration were performed. Statistical analyses were performed using the R software, version 4.2.0 (http://www.r-project.org).

Results

After screening 657 reports and conference abstracts, a total of 17 trials were found eligible (Fig. 1) [12-18, 31-41]. Two phase 2 RCTs were excluded because of small sample size [42, 43]. All the eligible studies were phase 3 randomized trials. Table 1 shows the detailed information of the eligible trials. We included the largest primary endpoint population of KEYNOTE-181 (combined positive score (CPS) ≥ 10) [15], CheckMate 649 (CPS ≥ 5) [16, 36], JAVELIN Gastric 100 (all patients) [37], KEYNOTE-062 (CPS ≥ 1) [35], CheckMate 648 (CPS ≥ 1) [31], ORIENT-15 (all patients) [39] and ORIENT-16 (all patients) [40]. CheckMate 649 was comprised of 2,031 patients, of whom 1,581 were randomly assigned (1:1) to nivolumab plus chemotherapy or chemotherapy, and 813 were randomly assigned (1:1) to nivolumab plus ipilimumab or chemotherapy. The former cohort was published in 2021 [16], and the latter was reported in the ESMO congress 2021 [36]; thus, two comparisons were included in our analysis. Likewise, two comparisons of the KEYNOTE-062 [35] (pembrolizumab versus chemotherapy, and pembrolizumab plus chemotherapy versus chemotherapy) and CheckMate 648 [31] (nivolumab plus chemotherapy versus chemotherapy, and nivolumab plus ipilimumab versus chemotherapy) were included in our analysis. Overall, the 17 eligible trials yielded 20 treatment comparisons with a total of 9,657 subjects.

Figure 1

Study flow chart. PD-1: programmed death-1; ASCO: American Society of Clinical Oncology; ESMO: European Society for Medical Oncology; OR: odds ratio; HR: hazard ratio; OS: overall survival.

Table 1

Characteristics of the Included Trials

Studies	Population	Line of treatments	Study phase	Primary endpoint	Treatment arms	N	DCR (%)	ORR (%)	OS		PFS
									Median	HR	Median	HR
ATTRACTION-2 [12]	G/EGJ carcinoma	≥ 3	3	OS	Nivo; placebo	330; 163	40; 25	11; 0	5.26; 4.14	0.63 (0.51 - 0.78)	1.61; 1.45	0.60 (0.49 - 0.75)
JAVELIN Gastric 300 [32]	G/EGJ carcinoma	≥ 2	3	OS	Avelumab; chemotherapy	185; 186	22.2; 44.1	2.2; 4.3	4.6; 5.0	1.1 (0.90 - 1.40)	1.4; 2.7	1.73 (1.40 - 2.20)
KEYNOTE-061 [33]	G/EGJ carcinoma	2	3	OS; PFS	Pembrolizumab; paclitaxel	196; 199	59.5; 29.5	16; 14	9.1; 8.3	0.82 (0.66 - 1.03)	1.5; 4.1	1.27 (1.03 - 1.57)
ATTRACTION-3 [13]	ESCA	2	3	OS	Nivo; chemotherapy	210; 209	37; 66	19; 22	10.9; 8.4	0.77 (0.62 - 0.96)	1.7; 3.4	1.08 (0.87 - 1.34)
ATTRACTION-4 [34]	G/EGJ carcinoma	1	3	OS; PFS	Nivo + chemotherapy; placebo + chemotherapy	362; 362	71.8; 68.5	57.5; 47.8	17.45; 17.15	0.90 (0.75 - 1.08)	10.45; 8.34	0.68 (0.51 - 0.90)
ESCORT [14]	ESCA	2	3	OS	Camrelizumab; chemotherapy	228; 220	44.7; 34.5	20.2; 6.4	8.3; 6.2	0.71 (0.57 - 0.87)	1.9; 1.9	0.69 (0.56 - 0.86)
KEYNOTE-590 [18]	ESCA	1	3	OS; PFS	Pembrolizumab + chemotherapy; placebo + chemotherapy	373; 376	79.3; 75.6	45; 29.3	12.4; 9.8	0.73 (0.62 - 0.86)	6.3; 5.8	0.65 (0.55 - 0.76)
KEYNOTE-181 [15]*	ESCA	2	3	OS (CPS ≥ 10)	Pembrolizumab; chemotherapy	107; 115	49.5; 47.0	21.5; 6.1	9.3; 6.7	0.69 (0.52 - 0.93)	2.6; 3.0	0.73 (0.54-0.97)
CheckMate 649 [16, 36]**	G/EGJ carcinoma	1	3	OS; PFS (CPS ≥ 5)	Nivo + chemotherapy; chemotherapy; Nivo + ipi; chemotherapy	473; 482; 234; 239	88; 79; 54; 83	60; 45; 27; 47	14.4; 11.1; 11.2; 11.6	0.71 (0.59 - 0.86); 0.89 (0.71 - 1.10)	7.7; 6.0; 2.8; 6.3	0.68 (0.56 - 0.81); 1.42 (1.14 - 1.76)
JAVELIN Gastric 100 [37]***	G/EGJ carcinoma	1	3	OS (all patients or TPS ≥ 1)	Avelumab; chemotherapy	249; 250	50.2; 61.2	13.3; 14.4	10.4; 10.9	0.91 (0.74 - 1.11)	3.2; 4.4	1.04 (0.85 - 1.28)
KEYNOTE-062 [35]****	Gastric carcinoma	1	3	OS; PFS (CPS ≥ 1 or ≥ 10)	Pembrolizumab + chemotherapy; pembrolizumab; chemotherapy	257; 256; 250	78; 42; 79	49; 15; 37	13.9; 10.6; 11.1	0.85 (0.70 - 1.03); 0.91 (0.74 - 1.10)	6.9; 2.0; 6.4	0.84 (0.70 - 1.02); 1.66 (1.37 - 1.51)
RATIONALE 302 [38]	ESCA	2	3	OS	Tislelizumab; chemotherapy	256; 256	47; 41.8	20.4; 9.8	8.6; 6.3	0.70 (0.57 - 0.85)	1.6; 2.1	0.83 (0.67 - 1.01)
CheckMate 648 [31]*****	ESCA	1	3	OS; PFS (CPS ≥ 1)	Nivo + chemotherapy; Nivo + ipi; chemotherapy	158; 158; 157	78; 63; 66	53; 35; 20	15.4; 13.7; 9.1	0.54 (0.37 - 0.80); 0.64 (0.46 - 0.90)	6.9; 4.0; 4.4	0.65 (0.46 - 0.92); 1.02 (0.73 - 1.43)
ESCORT-1st [17]	ESCA	1	3	OS; PFS	Camrelizumab + chemotherapy; chemotherapy	298; 298	91.3; 88.9	72.1; 62.1	15.3; 12.0	0.70 (0.56 - 0.88)	6.9; 5.6	0.56 (0.46 - 0.68)
ORIENT-15 [39]***	ESCA	1	3	OS (CPS ≥ 10 and all patients)	Sintilimab + chemotherapy; chemotherapy	327; 332	NR	66.1; 45.5	16.7; 12.5	0.63 (0.51 - 0.78)	7.2; 5.7	0.56 (0.46 - 0.68)
ORIENT-16 [40]***	G/EGJ carcinoma	1	3	OS (CPS ≥ 5 and all patients)	Sintilimab + chemotherapy; chemotherapy	327; 323	NR	58.2; 48.4	15.2; 12.3	0.76 (0.63 - 0.94)	7.1; 5.7	0.64 (0.53 - 0.77)
JUPITER-06 [41]	ESCA	1	3	OS; PFS	Toripalimab + chemotherapy; chemotherapy	257; 257	NR	NR	17.0; 11.0	0.58 (0.43 - 0.78)	5.7; 5.5	0.58 (0.46 - 0.74)

*Patients with CPS ≥ 10 were included. **Patients with CPS ≥ 5 were included. ***All randomly assigned patients were included. ****Patients with CPS ≥ 1 were included. *****Patients with CPS ≥ 1 were included. G: gastric; EGJ: esophagogastric junction; ESCA: esophageal carcinoma; Nivo: nivolumab; Ipi: ipilimumab; NIVO3: nivolumab 3 mg/kg; NIVO1 + IPI3: nivolumab 1 mg/kg plus ipilimumab 3 mg/kg; NIVO3 + IPI1: NIVO3 plus ipilimumab 1 mg/kg; NR: not reported; CPS: combined positive score; TPS: tumor positive score; ORR: objective response rate; DCR: disease control rate; OS: overall survival; PFS: progression-free survival; HR: hazard ratio.

First, a total of 20 available arms were included to derive the arm-level correlations between potential endpoints and OS. ORR and DCR showed strong and moderate correlations with median OS (R = 0.91, P < 0.001, Fig. S1A; R = 0.80, P < 0.001, Fig. S1B) (Supplementary Material 2, www.wjon.org). Similarly, median PFS was strongly correlated with median OS (R = 0.87, P < 0.001, Fig. 2a).

Figure 2

Performance of PFS as surrogate endpoint for OS in immuno-oncology trials of advanced gastro-esophageal carcinoma. (a) Correlation between PFS and OS at arm-level. Each dot represents one of the experimental arms of the phase 3 clinical trials, with size of the dot being proportional to the sample size. (b) Correlation between HRs for PFS and OS at trial-level. Size of dots is proportional to weighted sample size. The blue line represents the upper and lower 95% confidence intervals of the regression line (red line). Trials are colored based on whether the endpoint results were statistically significant. Nivo: nivolumab; Pembro: pembrolizumab; Ipi: ipilimumab; C: chemotherapy; GC: gastric carcinoma; ESCA: esophageal carcinoma; HR: hazard ratio; OS: overall survival; PFS: progression-free survival; NS: not significant; R: weighted Pearson correlation coefficient.

We then derived the degree of association between treatment effect on potential endpoints and OS at trial-level. Since none of the 131 patients in the placebo group had an objective response in the ATTRACTION-2 trial [12], the OR for ORR (infinite) in the ATTRACTION-2 trial was not available. Eighteen comparisons of ORs for ORR and HRs for OS were available, among which 10 reported improvements in both ORR (lower limit of CI for OR > 1.0) and OS (upper limit of CI for HR < 1.0). Correlation between ORORR and HROS was moderate (R = 0.71, P < 0.001, Fig. S1C) (Supplementary Material 2, www.wjon.org). Including the ATTRACTION-2 trial, the correlation between ORORR and HROS was not significant. Sixteen pairs of ORs for DCR and HRs for OS were available, and the correlation between ORDCR and HROS was weak (R = 0.45, P = 0.069, Fig. S1D) (Supplementary Material 2, www.wjon.org). Twenty pairs of HRs for PFS and OS were available. Apart from the comparison of the ATTRACTION-4 trial, other 10 comparisons that showed improvement in PFS reported improvement in OS (Table 1, Fig. 2b). Correlation between HRPFS and HROS was moderate (R = 0.71, P < 0.001, Fig. 2b). A conservative minimum threshold effect of HRPFS less than 0.73 demonstrated the potential to predict a significant improvement in OS.

Further, leave-one-out cross-validation analyses were performed to evaluate the accuracy of PFS in predicting OS. It was noted that the observed HR for OS fell within the 95% prediction intervals in 19 of 20 comparisons, indicating that the treatment effect on PFS could be a potential predictor of OS (Fig. 3).

Figure 3

Leave-one-out cross-validation analysis of the prediction of OS by treatment effect on PFS. Predicted HRs for OS (blue circles) with 95% prediction intervals (vertical grey lines) were calculated from the observed HR on PFS of that particular trial and the surrogate model built on the remaining trials. Observed HRs are shown for OS (red squares). Nivo: nivolumab; Pembro: pembrolizumab; Ipi: ipilimumab; C: chemotherapy; HR: hazard ratio; OS: overall survival; PFS: progression-free survival.

Figure 4 shows the strength of association between PFS and OS at different cut-off time points. The arm- (Fig. 4a) and trial-level (Fig. 4b) correlations showed that 3-month PFS were strongly correlated with 6-month OS (R = 0.92, R = 0.90), 6-month PFS strongly correlated with 12-month OS (R = 0.88, R = 0.94), and 12-month PFS strongly correlated with 18-month OS (R = 0.86, R = 0.86). The strength of association was weakened as the OS increased.

Figure 4

Correlation between PFS and OS at different cut-off time points. (a) Correlation between PFS and OS at arm-level. Bottom right: PFS at 6 months to predict OS at 12 months. (b) Correlation between HRs for PFS and OS at trial-level. Bottom right: HRs for PFS at 6 months to predict HRs for OS at 12 months. Nivo: nivolumab; Pembro: pembrolizumab; Ipi: ipilimumab; C: chemotherapy; GC: gastric carcinoma; ESCA: esophageal carcinoma; HR: hazard ratio; OS: overall survival; PFS: progression-free survival; NS: not significant; R: weighted Pearson correlation coefficient.

Finally, subgroup analyses were performed to evaluate the correlation between treatment effect on PFS and OS in different tumor types, trial lines, treatment strategy and follow-up duration (Table 2). The strength of association between HRPFS and HROS remained moderate in gastric or GE junction cancer (R = 0.71), but weak in esophageal cancer (R = 0.47). Notably, the correlation between HRPFS and HROS became strong in trials of ≥ 2 lines (R = 0.96), monotherapy (R = 0.89) and shorter follow-up duration (R = 0.91).

Table 2

Subgroup Analysis of the Correlation Between PFS and OS as Trial Level

Subgroup analysis	No. of comparisons	Weighted correlation coefficients, R (95% CI)	P value
Tumor type
ESCA [13-15, 17, 18, 31, 38, 39, 41]	10	0.47 (0.00 - 0.99)	0.174
G/EGJ cancer [12, 16, 32-37, 40]	10	0.71 (0.22 - 0.99)	0.021
Trials line
First-line [16-18, 31, 34-37, 39-41]	13	0.57 (0.08 - 0.99)	0.043
≥ 2 lines [12-15, 32, 33, 38]	7	0.96 (0.72 - 0.99)	< 0.001
Treatment strategy
Monotherapy [12-15, 32, 33, 35, 37, 38]	9	0.89 (0.56 - 0.99)	0.001
Combinational therapy [16-18, 31, 34-36, 39-41]	11	0.41 (0.00 - 0.99)	0.215
Median follow-up
≥ 10 months [13, 16-18, 31, 32, 34-37, 39, 40]	14	0.71 (0.31 - 0.99)	0.005
< 10 months [12, 14, 15, 33, 38, 41]	6	0.91 (0.51 - 0.99)	0.011

G: gastric; EGJ: esophagogastric junction; ESCA: esophageal carcinoma; CI: confidence interval.

Discussion

This is the first study to comprehensively evaluate the candidate surrogate endpoints for OS in IO trials of advanced or metastatic GE carcinoma. In the present study, we found that RECIST-based DCR could not serve as appropriate surrogate endpoint for OS. However, RECIST-based ORR and PFS correlated strongly with OS at arm-level and moderately with OS at trial-level. The leave-one-out cross-validation approach also confirmed that the effects observed on PFS were adequate to predict the treatment effect on OS. Therefore, we proposed the use of PFS as potential surrogate endpoint for OS in IO trials of advanced or metastatic GE carcinoma.

Recently, the KEYNOTE-590 [18], and ESCORT-1st [17] trials demonstrated that a combination of ICIs with chemotherapy was more effective than chemotherapy alone in previously untreated esophageal carcinoma. Furthermore, early reports from the CheckMate 648 trial in ASCO 2021 [31] suggest that chemo-free regimen (nivolumab plus ipilimumab) could represent a novel standard first-line treatment for esophageal carcinoma. Despite the unsuccessful exploration of pembrolizumab in second [33] and first-line [35] in gastric cancer, the CheckMate 649 trial [16] showed that nivolumab plus chemotherapy improved survival compared with chemotherapy alone. Therefore, the emerging of anti-PD-1/PD-L1 agents has unprecedentedly changed the treatment landscape of advanced GE carcinoma. However, not all patients have clinical response to ICIs, and several critical issues are required to be clarified, namely, identification of responders before the initial use of ICIs, and improvement of the therapeutic effect of ICIs through effective combination modality. Consequently, several randomized trials were in process to investigate the therapeutic effect of combinational regimens, such as a combination of ICIs with chemotherapy (KEYNOTE-859; RATIONAL-305; NCT03958890), anti-angiogenic (NCT03813784; NCT04949256), and targeted agents (KEYNOTE-811).

It is well recognized that OS is the golden standard primary endpoint for clinical trials of solid tumors. To reduce the sample size, shorten the follow-up duration and accelerate the approval of effective regimens, identification of surrogate endpoint for OS is an optional but important surrogate. Several clinical trials had set PFS (NCT03958890) as the unique primary endpoint or PFS and OS [16, 18, 33, 34] as the dual primary endpoints. Indeed, in the era of chemotherapy, RECIST-based endpoints had been commonly used as surrogate endpoints for OS in GE carcinoma; however, the use of these endpoints for OS in IO trials remains debatable because of the distinct anti-tumor mechanism of ICIs [20, 44], such as low-quality progression and delayed response [21]. Two previous meta-analyses showed that weak correlations did not support the surrogacy of RECIST-based endpoints for OS in pan-cancer advanced IO trials [23, 24]. Despite this, heterogeneity is pervasive and enormous across various cancer types [45], and the response patterns of cancer types treated with ICIs are diverse. Thus, the correlations in pan-cancer advanced IO trials cannot extrapolate to trials of particular cancer type [46]. Therefore, exploration of surrogate endpoints for OS in IO trials of GE carcinoma is still important.

In the present study, we applied rigid criteria and included a total of 17 large phase 3 trials with 9,657 patients to solve this issue. Firstly, we found that DCR and ORR did not strongly correlate with OS at both arm- and trial-level. We considered that not only the evaluation of targeted lesions, but also the follow-up duration is critical. In addition, the DCR and ORR at extreme condition (e.g., 0% and 100%) could not effectively predict outcome of OS. We found that the correlations between PFS and OS at arm- and trial-level were strong and moderate, respectively. The leave-one-out cross-validation analysis further confirmed the potential surrogacy of PFS for OS. Our study indicated a conservative minimum threshold effect of HRPFS ≤ 0.73 to highly predict a significant improvement in OS. It is believed that the acceptable correlation between PFS and OS in IO trials of GE carcinoma is largely ascribed to the condition that limited subsequent lines of therapy if patients with advanced GE carcinoma progressed after treating with ICIs. However, we should note that the heterogeneity is still obvious, including the heterogeneity of multiple cancer types (gastric cancer, gastroesophageal junction adenocarcinoma and esophageal squamous cell carcinoma) and line treatment. Therefore, our study should be interpretated cautiously.

In future IO trial, interest could be focused on predicting the treatment effects on OS by observing the effects on PFS at earlier time points. Kok et al reported that 6-month PFS could effectively predict 12-month OS in IO trials [47]. Similarly, our study found that 3-month PFS could reliably predict 6-month OS, 6-month PFS could reliably predict 12-month OS, and 12-month PFS could reliably predict 18-month OS in IO trials of advanced GE carcinoma. However, we noted weakened correlations between HRPFS and HROS as the follow-up duration increased. We considered that this phenomenon could be mainly attributed to the disproportionate increase of HRPFS and HROS because of delayed responses in the experimental arms.

Our study had several limitations. First, despite that the treatment modalities of gastric and esophageal carcinoma are similar, potential heterogeneity in terms of tumor type should be noted in our study. The combination of first line, later line and different treatment modalities also contributed to certain level of heterogeneity within eligible trials. Although we performed subgroup analyses to reduce these biases, the small number of comparisons (range: 6 - 13) in each analysis indicated a low power for statistical analysis. In addition, several endpoints modified based on RECIST criteria may better reflect the response pattern of ICIs, such as irRC [48], irRECIST [49] and iRECIST [50] criteria. However, the included trials of our studies had not reported these endpoints; thus, we could not explore the surrogacy of these endpoints for OS in IO trials of GE carcinoma. Lastly, our analysis was performed at arm- and trial-levels, and lacked patients-level analysis.

Conclusions

RECIST-based PFS may be the appropriate surrogate for predicting OS in IO trials of GE carcinoma. A conservative minimum threshold effect of HRPFS less than 0.73 has the potential to predict a significant improvement in OS.

PubMed search terms.

Click here for additional data file.

Performance of ORR and DCR as surrogate endpoint for OS in immuno-oncology trials of advanced gastro-esophageal carcinoma. Correlation between ORR (A) and DCR (B) and OS at arm-level. Each dot represents one of the experimental arms of the phase 3 clinical trials, with size of the dot being proportional to the sample size. Correlation treatment effects on ORR (C) and DCR (D) and OS at trial-level. Size of dots is proportional to weighted sample size. The blue line represents the upper and lower 95% confidence intervals of the regression line (red line). Trials are colored based on whether the endpoint results were statistically significant. Nivo: nivolumab; Pembro: pembrolizumab; Ipi: ipilimumab; C: chemotherapy; GC: gastric carcinoma; ESCA: esophageal carcinoma; OR: odds ratio; HR: hazard ratio; ORR: objective response rate; DCR: disease control rate; OS: overall survival; NS: not significant; R: weighted Pearson correlation coefficient.

Click here for additional data file.