OBJECTIVES: Immune checkpoint inhibitors have recently been approved for resectable non-small cell lung cancer (rNSCLC): neoadjuvant nivolumab plus chemotherapy (CT) (neoNIVO+CT), and adjuvant atezolizumab (adjATEZO) as treatment following resection and adjuvant CT (adjCT). Head-to-head comparisons are unavailable and differences in "time-zero" across trials preclude traditional indirect treatment comparisons (ITC). Our objective was to develop a method that adjusts for differences in time-zero, to incorporate into an ITC of event-free survival (EFS) between neoNIVO+CT and adjATEZO among patients with stage II-IIIA rNSCLC whose tumours express programmed death ligand-1 (PD-L1) ≥ 50% or PD-L1 ≥ 1%.

METHODS: We designed a time-zero-adjusted ITC by incorporating time-varying and mixture modeling components into a traditional ITC. The evidence network was formed by: CheckMate 816 (neoNIVO+CT vs. neoadjuvant CT [neoCT]), IMpower010 (adjATEZO vs. adjuvant best supportive care; both initiated after adjCT), and NATCH (neoCT vs. adjCT). We set the analysis time-zero to align with CheckMate 816 (in which randomization occurred prior to neoadjuvant treatment) and with NATCH, and estimated time-zero in IMpower010 to occur 7 months later. We used piecewise constant hazard ratios (HRs), partitioned at 7 months. For adjATEZO, we set HR_0to7mo vs. adjCT to 1 (no difference), while HR_7to36mo was obtained from IMpower010. A mixture model was implemented to address the intent-to-treat principles (i.e., trial dropout/ineligibility for adjATEZO during 0 to 7 months). Bootstrapping was used to estimate 95% confidence intervals (CI), and sensitivity analyses tested modeling assumptions.

RESULTS: Using a common time-zero commencing prior to neoadjuvant treatment or immediate surgery, the EFS HR_0to36mo between neoNIVO+CT vs. adjATEZO was 0.29 (95% CI: 0.11, 0.75) for PD-L1 ≥ 50% and 0.45 (0.23, 0.81) for PD-L1 ≥ 1%. Estimates were robust to sensitivity analyses.

CONCLUSIONS: This time-zero-adjusted ITC approach addresses bias arising from differences in time-zero across a network of trials, although additional assumptions should be evaluated through sensitivity analysis.