A model-based survival meta-analysis for indirect comparison of immune therapy efficacy in NSCLC
Kirill Zhudenkov (1, 2, 3), Nikolai Katuninks (2), Boris Shulgin (2), Kirill Peskov (1, 2, 3)
(1) M&S Decisions FZ-LLC, Dubai, UAE; (2) Research Center of Model-Informed Drug Development, I.M. Sechenov First Moscow State Medical University, Moscow, Russia; (3) Marchuk Institute of Numerical Mathematics of the Russian Academy of Sciences (RAS)
Introduction:
Overall survival (OS) is a key endpoint for the efficacy assessment in oncology clinical trials [1]. However, indirect comparison of the study outcomes based on OS measures may be challenging or not informative if it is based on the standard pairwise meta-analysis. It is due to a high heterogeneity in inclusion criteria, disease and patient characteristics as well as limited observation time in particular studies.
Objectives:
The aim of this research was to perform an indirect comparison of cancer immune therapy efficacy based on the published clinical study outcomes in advanced non-small cell lung cancer (NSCLC) involving the techniques of model-based meta-analysis (MBMA) by means of flexible survival models.
Methods:
The published Kaplan-Meier survival curves were digitized via an updated approach by Wei et al. [3]. A set of covariates was introduced to represent the data – treatment type for NSCLC (chemotherapy, PD-1 therapy, PD-L1 therapy, PD-1+chemo and PD-L1+chemo combinations), treatment line (first line and second line+), PD-L1 status (positive for at least 1% expression via Tumor Proportion Score). For optimal parameter estimation and representative forward simulations (up to 36 months) the studies with observation time of at least 24 months were selected. To perform the MBMA, flexible parametric survival models from mexhaz package for R were utilized to describe patient survival [2].
These models incorporate baseline hazard functions using cubic splines, proportional or not proportional hazards models for covariate effects (when a covariate effect changes over time) as well as random effects that represent between-study variability. Model selection was performed via a stepwise covariate search algorithm towards Akaike Information Criterion minimization, parameter identifiability assessment and a set of the visual predictive model diagnostics.
Current research extended the previous analysis [4] introducing the updated NSCLC study data, applying covariate search techniques for model building, testing non-proportional hazards effects, and introducing the survival prediction for combined patient cohorts.
Results:
The dataset incorporated the OS data of 20 phase II and III clinical studies (11626 subjects). The data on chemotherapy as well as immune therapy and its combinations (nivolumab, atezolizumab, pembrolizumab, avelumab, durvalumab) were collected. Covariate search supplied the final model with the following hazard function: h(t|Z) ~ exp(ran.eff + PD-L1_status + (PD-L1/chemo) + (PD-1/chemo) + spline.basis*(PD-1 + PD-L1 + Line)), where the baseline hazard is represented by cubic spline functions (with knot positions at 5.3 and 12.5 months) dependent on immune monotherapy type and treatment line. It implied the description of non-proportional hazard component for these covariates. The optimal model also incorporated the impact of PD-L1/chemo or PD-1/chemo combined therapy covariates as well as PD-L1 status as proportional hazards.
The outcomes of the analysis suggest that the immune monotherapy vs. chemotherapy demonstrates higher hazard ratios (HR) at first 2 months of observation while for all further times the HR stays below 1. HR = 1.36 (1.21-1.52), 0.81 (0.76-0.87), 0.59 (0.54-0.65), 0.7 (0.62-0.78), 0.9 (0.72-1.11) for 2, 6, 12, 24, 36 months were obtained, respectively, for PD-1 therapy vs. placebo (point estimates and 95% CI). However, the PD-1/chemo provides a constant HR = 0.68 (0.61-0.75) for all observation times. Similar results correspond to PD-L1 inhibitors: the HR estimates are quite close to the provided above. A high unexplained variability shows that PD-1 and PD-L1 immunotherapies can be hardly differentiated. Furthermore, according to the simulated scenarios for a combined cohort of naïve patients containing 20% of PD-L1 positive subjects the median survival times of 11.5 (9.6-14.8), 14.9 (11.8-21.2), 15.7 (12.3-21.1), 16.7 (13.4-22.0), 15.3 (11.5-20.8) months were predicted for chemotherapy, PD-1, PD-L1, PD-1/chemo and PD-L1/chemo treatments, respectively.
Conclusions: The created MBMA survival methodology based on the flexible parametric survival models with constant and time-varying covariate effects provides a comprehensive basis for the exhaustive analysis and indirect comparison of time-to-event outcome measures. Finally, application of the methodology for the NSCLC immunotherapy shows noninferiority of PD-1 vs. PD-L1 immune checkpoint inhibitors.
References:
[1] Clinical Trial Endpoints for the Approval of Cancer Drugs and Biologics. Guidance for Industry. FDA CDER/CBER, 2018.
[2] Charvat, H., & Belot, A. (2021). mexhaz: An R Package for Fitting Flexible Hazard-Based Regression Models for Overall and Excess Mortality with a Random Effect. Journal of Statistical Software, 98(14), 1–36.
[3] Wei Y, Royston P. Reconstructing time-to-event data from published Kaplan-Meier curves. Stata J. 2017;17(4):786-802.
[4] Sergey Gavrilov et al. Survival Model-based Meta-Analysis Framework for the Indirect Comparison of Anti-Cancer Therapy Efficacy. Page Meeting 2021. Abstr 9697.
