Prognostic score models for acquired resistance to frontline anti-programmed cell death protein 1/programmed death-ligand 1 therapy in metastatic non-small-cell lung cancer

Acquired resistance (AR) to immunotherapy following initial response occurs frequently in metastatic non-small-cell lung cancer (mNSCLC). We developed and validated easy-to-apply prognostic score models to identify patients with mNSCLC at greater risk of AR to frontline (1L) anti-programmed cell death protein 1/programmed death-ligand 1 [anti-PD-(L)1] therapy.

Electronic health records of patients with mNSCLC who responded to 1L anti-PD-(L)1 (diagnosed 2015-2023) in the ConcertAI Patient360™ database were analyzed. Multivariable Cox models (monotherapy versus combination) assessing association between baseline prognostic factors and real-world duration of response (rwDoR) were used to develop prognostic scores and stratify patients into low- and high-risk groups. Risk stratification was evaluated using Kaplan-Meier curves, performance metrics, and decision curves. External validation utilized the COTA NSCLC dataset.

The majority of patients (n = 3181) did not respond to anti-PD-(L)1, leaving 1891 1L immunotherapy-responding patients (94% partial responders, 6% complete responders), with a median (interquartile range) time from anti-PD-(L)1 initiation to first response of 2.5 months (1.84-3.19 months). Across the monotherapy and combination therapy cohorts, median rwDOR was significantly shorter in high-risk (7.6 and 5.2 months) versus low-risk groups (18.9 and 12.0 months), suggesting model differentiation. Decision curves were above reference lines at predicted AR probabilities 180-day after initial response (monotherapy: 35.5%; combination therapy: 45.4%) corresponding to the cut-off for high-risk patients, indicating greater net benefits when employing risk grouping. External dataset validation demonstrated that model-based risk grouping significantly distinguished rwDOR and exhibited consistent performance metrics.

Proposed prognostic score models may facilitate identifying patient subsets at higher risk of AR to 1L anti-PD-(L)1.