Introduction

Progression-free survival (PFS) indicates tumor growth earlier than overall survival (OS). Death is included as one of the defining events of PFS, and PFS indeed appears to show some correlation with OS. Nonetheless, PFS remains to be proven as a valid surrogate endpoint for OS in most cancers [1]. One meta-analysis investigating the relationship between PFS and OS in gastric cancer revealed a correlation between the two. The correlation was not strong enough, however, to merit PFS being regarded as a surrogate endpoint [2].

Progression-free survival is defined as a composite of a number of events, including new lesions, progression of existing lesions, and death. Therefore, the European Medicines Agency guideline recommends each individual event be analyzed and reported separately [3]. One randomized phase III study of breast cancer demonstrated that the prognosis in patients with new metastatic lesions was less favorable than in those with increased size of pre-existing lesions [4]. This suggests that progression type might likewise affect the prognosis in gastric cancer as well.

In investigating the association between various events occurring in a trial and OS, the standard approach adopted in many studies has been to employ a Cox regression model using time-dependent covariates [4–9]. However, the association between progression type and survival remains to be investigated by this method in gastric cancer.

The aim of the present report is to investigate the association between progression type and OS based on the results of an earlier randomized phase III G-SOX study [10] demonstrating the non-inferiority of S-1 plus oxaliplatin (SOX) to cisplatin plus S-1 (CS) in terms of PFS [hazard ratio (HR) of SOX versus CS, 1.004; 95% confidence interval (CI), 0.840–1.199; non-inferiority margin, 1.30] and OS (HR of SOX versus CS, 0.958; 95% CI, 0.803–1.142) in chemotherapy-naïve advanced gastric cancer (AGC) patients.

Materials and methods

Study source

A total of 685 patients were enrolled in the G-SOX study (SOX group, 343; CS group, 342). The present study included 663 patients (full analysis set), excluding 22 patients because of withdrawal before treatment initiation, violation of GCP principles, or deviation from inclusion/exclusion criteria (SOX, 333; CS, 330). The tumors were radiologically assessed every 6 weeks from baseline to last evaluation before beginning post-study treatment. An independent review committee reviewed all images in accordance with the Response Evaluation Criteria in Solid Tumors (RECIST) v1.0 [11]. The investigators identified progressive disease (PD) in 82% of patients, while the independent review committee identified it in 75%. Agreement between the assessments was 81%. The present study was based on the results of the independent review.

Classification of progression type

According to RECIST, the progression of target or non-target lesions represents the progression of existing lesions, while the development of new lesions indicates new metastases not detected at baseline. Here, these two types of lesion were categorized as new lesions (PD NL+) or no new lesions (PD NL−). When new lesions and enlargement of existing lesions were detected simultaneously, this was classified as PD NL+. When new lesions developed and the size of existing lesions remained unchanged, this was also classified as PD NL+. When no new lesions were detected and existing lesions enlarged, this was classified as PD NL−. Even if the study treatment was discontinued because of clinical PD, adverse events or others, the patients were categorized as either no PD, PD NL+, or PD NL− based on the CT assessments reviewed by the independent review committee since the assessments were continued until commencement of post-study treatment.

Statistical analysis

First, the numbers of patients with no PD, with PD NL−, with PD NL+, and death in the SOX and CS groups were determined. The reasons for treatment discontinuation in each group were also summarized according to progression type. The cumulative incidence function was estimated in PD NL− and PD NL+, with other types of progression considered as a competing risk. The HR of time-to-PD NL− for SOX versus CS and that for time-to-PD NL+ for SOX versus CS was estimated using a Cox regression model. In this analysis, events of other progression types were treated as censored.

Next, a Cox regression model was used to investigate the association between progression type and OS, with PD NL− and PD NL+ as time-dependent covariates, which change their values according to time-to-progression. Using Cox regression with these time-dependent covariates allows the states before and after progression to be compared, even if the events constitute competing risks. The presence or absence of PD NL− or PD NL+ was treated as a time-dependent covariate, as shown in Fig. 1. An illustration of how these time-dependent covariates were encoded in patients with no PD, with PD NL−, and with PD NL+ is provided in Supplementary Figure 1 (Online Resource 1). The state before PD NL− or PD NL+ was denoted as before-PD (0, 0). The state after PD NL− was denoted as after-PD NL− (1, 0). The state after PD NL+ was denoted as after-PD NL+ (0, 1). The inter-group HRs for SOX versus CS were calculated in before-PD (d0 in SOX versus d0 in CS), after-PD NL− (d1 in SOX versus d1 in CS), and after-PD NL+ (d2 in SOX versus d2 in CS). The HRs for after-PD NL− versus before-PD (d1 versus d0) and after-PD NL+ versus before-PD (d2 versus d0) were estimated in each treatment group.

Fig. 1
figure 1

Classification of progression state and time-dependent covariates. Progression status was classified according to progression type as follows: before-PD, after-PD NL−, or after-PD NL+. Value of each time-dependent covariate (PD NL−, PD NL+) is given in brackets. Before-PD includes either the entire period in patients with no PD or the period before PD was observed. PD progressive disease, PD NL− PD with no new lesions, PD NL+ PD with new lesions, p1 Progression with no new lesions (Covariate PD NL− changed from 0 to 1 at time of progression), p2 Progression with new lesions (Covariate PD NL+ changed from 0 to 1 at time of progression). d0, d1, d2 Death

Full size image

Finally, Cox regression adjusted for the baseline background factors (including sex; age, <70 years or ≥70 years; Eastern Cooperative Oncology Group performance status, PS0 or PS1–2; disease status, recurrent or unresectable; tumor histology, intestinal or diffuse; peritoneal metastasis, present or absent; sum of tumor diameters, < median or ≥ median; and alkaline phosphatase, < median or ≥ median) was performed for further exploratory analyses. The median values and 95% CI of survival post-progression (SPP) [OS−PFS] for after-PD NL− and after-PD NL+ were also estimated in each treatment group, referring to Kristine’s report [12], along with their Kaplan-Meier curves. Kaplan-Meier plots on OS according to progression type were also estimated for reference.

Statistical analyses were performed with SAS v9.3 (SAS Institute, Cary, NC, USA).

Results

Incidence of progression type

The numbers of patients with no PD, with PD NL−, and with PD NL+ are shown in Table 1. In the SOX group (333 patients), 167 showed PD NL− and 91 PD NL+. In the CS group (330 patients), 147 showed PD NL− and 95 PD NL+. No clear difference was observed in the site of new lesions between the two groups. The reasons for treatment discontinuation in patients with no PD, with PD NL−, and with PD NL+ are shown in Supplementary Table 1 (Online Resource 2). Among the patients in whom treatment was discontinued because of clinical PD, the number of patients with no PD, with PD NL−, and with PD NL+ was 24, 6, and 2 in the SOX group and 30, 0, and 2 in the CS group, respectively. Descriptions of progression type and new lesions according to baseline metastatic sites are presented in Supplementary Tables 2 and 3 (Online Resources 3 and 4).

Table 1 Number of patients according to progression type
Full size table

Cumulative incidence of progression type

The cumulative incidence functions of PD NL+ and PD NL− are shown in Fig. 2. The cumulative incidence of PD NL− was higher than that of PD NL+. The incidences of PD NL− and PD NL+ were similar between SOX and CS.

Fig. 2
figure 2

Cumulative incidence of progression type. Hazard ratio (HR) and 95% confidence interval (CI) of time-to-PD NL− for SOX versus CS were 1.00 and 0.80–1.25, respectively. HR and 95% CI of time-to-PD NL+ for SOX versus CS were 0.85 and 0.64–1.14, respectively. SOX S-1 plus oxaliplatin; CS cisplatin plus S-1; PD progressive disease; PD NL− PD with no new lesions; PD NL+ PD with new lesions

Full size image

Cox regression with time-dependent covariates

The results of the Cox regression for OS with time-dependent covariates are shown in Table 2 and Fig. 3.

Table 2 Cox regression for overall survival with time-dependent covariates
Full size table
Fig. 3
figure 3

Cox regression with two time-dependent covariates (progression type). Before-PD includes either the entire period in patients with no PD or the period before PD was observed. Each HR is conditional on progression status and cannot be interpreted as a simple treatment effect. aSOX vs. CS in before-PD: HR 0.64 (95% CI, 0.43–0.97). bSOX vs. CS in after-PD NL−: HR 1.11 (95% CI, 0.87–1.42). cSOX vs. CS in after-PD NL+: HR 0.81 (95% CI, 0.60–1.09). SOX S-1 plus oxaliplatin; CS cisplatin plus S-1; PD progressive disease; PD NL− PD with no new lesions; PD NL+ PD with new lesions; HR hazard ratio; CI confidence interval

Full size image

In the SOX group, the HR for after-PD NL− versus before-PD was 4.66 (95% CI, 3.21–6.77), while that for after-PD NL+ versus before-PD was 7.26 (95% CI, 4.89–10.80). In the CS group, the HR for after-PD NL− versus before-PD was 2.71 (95% CI, 1.95–3.75), while that for after-PD NL+ versus before-PD was 5.78 (95% CI, 4.13–8.08).

The results of the Cox regression adjusted for the baseline background factors are presented in Supplementary Table 4 (Online Resource 5). The inter-group HRs (SOX vs. CS) for each progression type and intra-group HRs for after-PD NL− versus before-PD and after-PD NL+ versus before-PD were similar to those described above.

Survival post-progression

In the SOX group, the median SPPs were 9.7 months (95% CI, 8.0–11.3) for after-PD NL− and 5.8 months (95% CI, 3.7–7.4) for after-PD NL+, while in the CS group, they were 9.0 months (95% CI, 7.5–11.3) for after-PD NL− and 5.3 months (95% CI, 4.0–6.7) for after-PD NL+. Kaplan-Meier plots of SPP and OS are presented in Supplementary Figures 2 and 3 (Online Resources 6 and 7).

Discussion

The results of the present study showed that the negative impact of PD NL+ on OS, in particular, was higher than that of PD NL−. The association between progression type and OS was almost comparable between the two treatment groups (SOX and CS).

One earlier meta-analysis suggested a moderate correlation between OS and PFS in AGC [2]. However, heterogeneity of progression was not taken into consideration in that report. We hypothesized that heterogeneity of progression affects OS in gastric cancer, although this has not been suggested in earlier studies. The present results indicate that differences in how each progression type affects OS influence the correlation between OS and PFS. This suggests that an exploratory analysis of progression type would be one feasible approach to understanding any inconsistencies that might arise between PFS and OS in comparative clinical trials.

One earlier study of breast cancer suggested that a new endpoint, defined as the time from randomization to progression with new metastasis or death, might improve the predictivity on OS [4]. It has been noted that it can be difficult to identify progression and measure lesions accurately by assessing pre-existing lesions, especially when the baseline lesions are small [13]. This suggests that endpoints focused on new lesions might be more adequate and predictive. This indicates that progression type is one candidate worth investigating in developing optimal endpoints. An evaluation of progression type, therefore, should be recommended in future clinical studies of gastric cancer.

Here, Cox regression with time-dependent covariates was selected to explore the association between progression type and OS. It has also been used in earlier studies in other cancers [4, 5]. We believe that this is the best approach in this setting. Information on post-progression survival may be of use in determining how progression type is likely to affect prognosis, but is limited by the number of patients and observation period. Moreover, Kaplan-Meier plots categorizing patients according to progression type are also inadequate as they do not reflect the time point of progression.

This study had two major limitations: first, it was a retrospective analysis of just one clinical trial, and SPPs for after-PD NL− and after-PD NL+ were not directly compared in a prospective manner; second, here we focused on progression observed during chemotherapy. Therefore, progression cannot be interpreted as a prognostic factor before treatment. Additionally, there is also a potential risk that the limited total number of target lesions representative of all involved organs proposed in the RECIST guidelines does not accurately reflect the tumor burden [5]. Even within pre-existing lesions, target or non-target lesions may have a different effect on prognosis. However, they were not investigated in the present study.

The present findings showed that the appearance of new lesions or an increase in the extent of tumors indicated a poorer prognosis. In clinical practice, therefore, progression type may provide useful information in determining a prognosis on termination of treatment. It may also help in determining subsequent treatment options.

As a rule, progression is usually treated irrespective of specific type when considering PFS. However, the present results revealed that progression accompanied by new metastatic lesions had a strong negative impact on OS in patients treated with S-1 and platinum for AGC. This suggests that progression type may constitute an important factor in evaluating the association between PFS and OS. The present findings, however, need to be validated in future studies.