Synthetizing Published Evidence on Survival by Reconstruction of Patient-Level Data and Generation of a Multi-Trial Kaplan-Meier Curve

Abstract

Introduction In conducting a survival meta-analysis, the typical methodological approach analyses the hazard ratios (HRs) of individual trials and then combines them into a pooled meta-analytical estimate. The length of follow-up of individual trials is not generally accounted for. Recent techniques aimed at individual patient-data reconstruction from Kaplan-Meier graphs represent an important methodological innovation. These techniques permit the combination of the survival curves published in a single clinical trial but are also applicable to more than one trial. In the case of multiple trials, a meta-analysis can be conducted without using any statistical model of meta-analysis. Methods As an example of this new approach, we applied a technique of individual patient data reconstruction to the Kaplan-Meier graphs of overall survival reported in two phase-III trials, which were conducted on patients with locally advanced/advanced non-small cell lung cancer selected according to their PD-L1 expression status, not previously treated for their metastatic disease. Only subjects with PD-L1 $\geq$50{\%} were considered for our analysis. The experimental arms received pembrolizumab monotherapy while the control arms were given platinum-based chemotherapy. The survival graphs were obtained for both trials. For each Kaplan-Meier curve, the graph was firstly digitalized. Then, the Shiny package was used to reconstruct patient-level data. Finally, the pooled survival curves were generated from the reconstructed patient-level data along with the relevant Cox statistics; for this purpose, we used three packages ({\textquotedbl}coxph{\textquotedbl}, {\textquotedbl}survfit{\textquotedbl}, and {\textquotedbl}ggsurvplot{\textquotedbl}) under the R-platform. Results In our pooled analysis based on this procedure, we compared 453 patients given pembrolizumab vs. 451 controls given chemotherapy. The HR estimated from reconstructed patient-level data was 0.670 (95{\%} confidence interval [CI], 0.566 to 0.793). Conclusion The analysis described herein demonstrates the easy applicability of the Shiny technique. This technique was successful in generating a pooled survival graph for the experimental treatment groups vs. controls and efficiently estimated the pooled HR in which the results of the two trials were combined.