Cross-continental transportability of phase III trial results to LMIC cohort with locally advanced NSCLC: A g-computation analysis with layered decomposition.

276

Background: Clinical trial results are frequently extrapolated to populations that were never part of Western RCTs, such as those in lower- and middle-income countries (LMIC), yet formal testing of transportability is almost never performed. We performed a causal transportability analysis to determine if survival predictions from the RTOG 0617 trial generalize to an LMIC cohort. Methods: We harmonised individual patient data from RTOG 0617 (n=258, 60Gy and 74Gy chemoradiation arms, without cetuximab) and a single-institution cohort (n=123) for stage IIIA–IIIB NSCLC. The Indian cohort was restricted to ECOG 0–1 and standard platinum-doublet chemotherapy resulting in 91 patients. The overall survival (OS) restricted mean survival time (RMST) gap at 24 months served as the primary transportability metric, quantified as transported minus observed RMST. We used G-computation with bootstrap shrinkage-corrected Western Cox models using a parsimonious set of harmonizable clinical covariates (age, ECOG, histology, stage, consolidation chemotherapy), to predict outcomes in the Indian cohort. We employed layered decomposition to isolate baseline hazard contributions from covariate effects. Calibration was assessed via slopes and Expected-to-Observed (E/O) ratios. Robustness was tested via three pre-specified sensitivity analyses (SA): SA1 restricted the source to the 60 Gy arm; SA2 removed consolidation chemotherapy; and SA3 added sex and radiotherapy technique (IMRT vs 3D-CRT). Results: Substantial covariate imbalances existed between cohorts (standardised mean differences >0.50 for age, ECOG, smoking, stage, and consolidation). In the primary analysis, the 24-month OS RMST gap was −1.5 months (95% CI −6.8 to +1.7, p=0.412)(observed = 19.2 months, transported = 17.7 months), suggesting plausible exchangeability. However, the E/O ratio of 1.35 indicated the Western model over-predicted Indian mortality by 35%. Sensitivity analyses revealed this result was fragile. Restricting to the 60 Gy arm (SA1) widened the OS gap to −3.6 months (95% CI −10.0 to +0.7, p=0.132). Removing consolidation (SA2) resulted in a collapsed model [calibration slope deviation (7.95, p=0.025)], confirming that consolidation chemotherapy acted as a confounder masking true population differences. Adding IMRT and gender (SA3) pushed the OS gap to near-significance (−2.9 months, 95% CI −6.8 to +0.1, p=0.060). The OS RMST gap was negative in all four analyses, with p-values tightening monotonically as compensatory modelling features were removed. Conclusions: NSCLC trial data provide guidance but not quantitative predictions, and application of trial-derived prognostic tools in LMIC settings requires recalibration of outcome data. High quality registries and equitable access to IPD trial data are needed to support local prognostic model development.