BAYESIAN MULTI-PARAMETER EVIDENCE SYNTHESIS FOR INFORMED INDIRECT TREATMENT COMPARISONS OF EXTRAPOLATED SURVIVAL OUTCOMES
Author(s)
Daniel J. Sharpe, PhD1, Ashley E. Tate, PhD2, Tuli De, PhD3, Jackie Vanderpuye-Orgle, MSc, PhD3;
1Parexel International Ltd, London, United Kingdom, 2Parexel International Ltd, Amsterdam, Netherlands, 3Parexel International Ltd, Durham, NC, USA
1Parexel International Ltd, London, United Kingdom, 2Parexel International Ltd, Amsterdam, Netherlands, 3Parexel International Ltd, Durham, NC, USA
OBJECTIVES: Relative efficacy estimates from network meta-analyses may be biased by differences in follow-up duration between studies. Supplementing immature trial data with historical trial data can address this bias while allowing more reliable estimation of longer-term treatment benefit. We extended the Bayesian multi-parameter evidence synthesis (B-MPES) framework for survival extrapolation to perform indirect treatment comparisons.
METHODS: We analyzed reconstructed interim overall survival data from four phase III studies of first-line immunotherapy plus tyrosine kinase inhibitor (IO+TKI) vs TKI monotherapy (sunitinib) in metastatic renal cell carcinoma (Bosma 2022; median follow-up range 18.1-30.6 months). Extrapolations were informed by a historical phase III trial of dual IO therapy vs sunitinib (median follow-up 55 months), via additional likelihood contributions derived from 6-month conditional survival in the historical trial, with the reference trial chosen for closest similarity in distribution across baseline risk strata. Treatments were ranked by 5-year restricted mean survival time (RMST). Results from a B-MPES model based on independent cubic splines and from a naïve dependent spline model were compared.
RESULTS: The B-MPES model predicted that the pembrolizumab+lenvatinib, pembrolizumab+axitinib, and nivolumab+cabozantinib regimens were closely competitive (respective surface under the cumulative ranking curve [SUCRA]: 0.84, 0.75, 0.60; vs 0.01 for sunitinib). The naïve model (SUCRA: 0.68, 0.65, 0.78) and standard meta-analysis of hazard ratios (Bosma 2022) likewise identified these three treatments as preferred, but slightly favored nivolumab+cabozantinib, for which available follow-up was shortest. The B-MPES model had lower uncertainty than the naïve model (e.g., difference in 5-year RMST for pembrolizumab+lenvatinib vs sunitinib: 5.64 [95% credible interval: 5.01-17.0] vs 4.67 [0.59-16.6] months) and better captured the strongly non-proportional hazard trends.
CONCLUSIONS: Employing a B-MPES framework with historical trial data to inform survival extrapolations for a connected network of studies attenuated bias disfavoring the treatments in trials with more mature data, which arose from relapse effects observed with longer follow-up.
METHODS: We analyzed reconstructed interim overall survival data from four phase III studies of first-line immunotherapy plus tyrosine kinase inhibitor (IO+TKI) vs TKI monotherapy (sunitinib) in metastatic renal cell carcinoma (Bosma 2022; median follow-up range 18.1-30.6 months). Extrapolations were informed by a historical phase III trial of dual IO therapy vs sunitinib (median follow-up 55 months), via additional likelihood contributions derived from 6-month conditional survival in the historical trial, with the reference trial chosen for closest similarity in distribution across baseline risk strata. Treatments were ranked by 5-year restricted mean survival time (RMST). Results from a B-MPES model based on independent cubic splines and from a naïve dependent spline model were compared.
RESULTS: The B-MPES model predicted that the pembrolizumab+lenvatinib, pembrolizumab+axitinib, and nivolumab+cabozantinib regimens were closely competitive (respective surface under the cumulative ranking curve [SUCRA]: 0.84, 0.75, 0.60; vs 0.01 for sunitinib). The naïve model (SUCRA: 0.68, 0.65, 0.78) and standard meta-analysis of hazard ratios (Bosma 2022) likewise identified these three treatments as preferred, but slightly favored nivolumab+cabozantinib, for which available follow-up was shortest. The B-MPES model had lower uncertainty than the naïve model (e.g., difference in 5-year RMST for pembrolizumab+lenvatinib vs sunitinib: 5.64 [95% credible interval: 5.01-17.0] vs 4.67 [0.59-16.6] months) and better captured the strongly non-proportional hazard trends.
CONCLUSIONS: Employing a B-MPES framework with historical trial data to inform survival extrapolations for a connected network of studies attenuated bias disfavoring the treatments in trials with more mature data, which arose from relapse effects observed with longer follow-up.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR217
Topic
Methodological & Statistical Research
Disease
SDC: Oncology