Disease-free survival as surrogate for overall survival in real-world settings for esophageal cancer: an analysis of SEER-Medicare data

Abstract

Background: Establishing surrogate endpoints for overall survival (OS) may expedite assessment of new therapies in esophageal cancer (EC) and gastroesophageal junction cancer (GEJC). This study aimed to evaluate disease-free survival (DFS) as a surrogate endpoint for OS.

Methods: Patients from the Surveillance, Epidemiology, and End Results (SEER)-Medicare database aged ≥66 years with resection after primary diagnosis of stage 2 or 3 EC/GEJC between 2009 and 2017 were analyzed (N = 925; median follow-up 26.2 months). Surrogacy was assessed by evaluating individual level associations between DFS and OS using Spearman's rank correlation and the association between treatment effects by Pearson's correlation coefficient. To evaluate the association between treatment effects, patients were classified in synthetic clusters based on treatments received. Propensity score matching addressed imbalances in baseline characteristics between treatment and control groups in the clusters. Predictive performance of the surrogacy equation was assessed internally for the generated clusters via leave-one-out cross-validation and externally via predictions for 26 clinical trials of early-stage EC/GEJC.

Results: Patients were mostly male (84%), non-Hispanic white (89.3%), with median age 71.8 years, and cancer stages 2 (50.4%) and 3 (49.6%). Cancer types were adenocarcinoma (76.1%), squamous cell carcinoma (10.4%), and other types (13.5%). Most patients 766/925 (82.8%) received neoadjuvant therapy (680/766 chemoradiotherapy versus 86/766 chemotherapy alone) while 23.6% of the patients received adjuvant therapy. Within each treatment setting, most [705/766 (92.0%) of neoadjuvant therapy and 178/218 (81.7%) of adjuvant therapy] received multi-agent chemotherapy. The individual level correlation was 0.76 (95% confidence interval 0.70-0.80). The correlation between treatment effects was 0.96 (95% confidence interval 0.80-0.99) with a corresponding surrogate threshold effect of 0.71. Both internal (91%) and external (89%) validation of the model demonstrated high predictive accuracy.

Conclusions: Correlations between DFS and OS were meaningful at both individual and treatment effect level. The derived surrogacy equation enables reliable early assessments of OS benefit from the observed DFS benefit for early-stage EC/GEJC treatments in real-world settings.

Keywords: disease-free survival; esophageal cancer; gastroesophageal junction cancer; meta-analysis; overall survival; real-world data; surrogacy.

Figure 1

WLR model for the linear relationship between natural log-transformed HR_DFSand HR_OSusing primary definition of DFS. The predictive surrogacy equation is graphed as a solid straight line with its corresponding 95% prediction interval boundaries as dashed curves. Red circles represent HR_OS and HR_DFS estimates within the clusters and their sizes are proportional to the sample sizes of the clusters they represent. The surrogate threshold effect (STE) is indicated on the graph where the upper bound of the 95% prediction interval of the surrogacy equation intersects the horizontal HR_OS = 1 line (i.e. an HR_DFS <0.71 is necessary to predict an HR_OS <1 with 95% probability). DFS, disease-free survival; HR, hazard ratio; OS, overall survival; WLR, weighted linear regression.

Figure 2

Results of LOOCV for WLR using the primary definition of DFS. The blue diamond-shaped markers are the computed HR_OS estimates per cluster and the green diamond-shaped markers are the predicted HR_OS measures from the weighted linear regression. Corresponding to the HR_OS measures, the 95% confidence/prediction intervals are presented as black horizontal lines. Rows represent the clusters of patients with indices referring to cluster numbers. DFS, disease-free survival; HR, hazard ratio; LOOCV, leave-one-out cross-validation; OS, overall survival. ^aIndicates the cluster(s) for which the computed HR_OS measure did not fall within the predicted interval.

Figure 3

External validation results using the RCT evidence base from Ajani et al. 2022. The blue diamond-shaped markers are the reported HR_OS in the RCTs and the green diamond-shaped markers are the corresponding predicted HR_OS from the weighted linear regression using data from the Surveillance, Epidemiology, and End Results (SEER)-Medicare database. Corresponding to the HR_OS measures, the 95% confidence/prediction intervals are presented as black horizontal lines. There is one outlier RCT, Thomas 2020, with an outlier HR_OS measure (both observed and predicted) potentially due to a relatively small sample size (n = 30 patients). Nevertheless, the actual HR_OS was relatively close to the predicted HR_OS and within the 95% prediction interval. HR, hazard ratio; OS, overall survival; RCT, randomized clinical trial. ^aIndicates the trial(s) for which the observed HR_OS did not fall within the 95% prediction interval.