6E-05 - Effect-based modeling of individualized treatment effects to predict benefit-harm balance: example using immune checkpoint inhibitors

Background: The benefit-harm balance of medications varies across individuals and subgroups. Predictive modeling of treatment-related benefits and harms can be a useful tool to predict individual responses to medications and guide clinical decision-making.

Objectives: To develop and validate predictive benefit and harm models and compare benefit-harm tradeoffs of immune checkpoint inhibitor + chemotherapy (ICI + chemo) relative to single-agent ICI among patients with non-small cell lung cancer (NSCLC).

Methods: This retrospective cohort study used SEER-Medicare data from 2012-2019 and included adults aged >66 years old with NSCLC cancer diagnosis who were not treated with systemic therapy prior to ICI. We developed and validated multivariable prediction models assessing the effect of ICI + chemo relative to single-agent ICI on overall survival (benefit) and severe irAE toxicity (harm). We used geographically distinct subsets of the patient sample from 11 cancer registries for training and 10 registries for validation. We modeled individual treatment effects comparing ICI + chemo to single-agent ICI in the validation cohort and created quartiles of benefit (difference in one-year restricted mean survival time [RMST]) and harm (difference in one-year probability of severe immune-related adverse events [irAE]). We used descriptive statistics to characterize individuals classified in low-benefit /high-harm, high-benefit/high-harm, and high-benefit /low-harm categories.

Results: Overall, 3,837 patients (2,190 training, 1,647 validation) received ICI + chemo and 2,566 (1,496 training, 1,070 validation) received single-agent ICI. Predicted extended 1-year survival (benefit) ranged from 14 to 102 days; predicted excess risk of severe irAE (harm) ranged from 0.0% to 15.1%. The 5-fold cross-validation C-statistics were 0.56 (internal) and 0.53 (external) for the harm model and 0.62 (internal) and 0.62 (external) for the benefit model. The higher-benefit/lower-harm (n = 176) and lower-benefit/higher-harm (n = 219) groups differed (p < 0.05) in terms of average age (73 vs. 77 years), non-Hispanic White (47% vs. 94%), average NCI index (0.5 vs. 0.8), distant stage diagnosis (92% vs. 25%), adenocarcinoma (95% vs. 19%) and autoimmune disease at baseline (25% vs. 6%).

Conclusions: We implemented a novel predictive benefit and harm approach to support individualized treatment decisions for NSCLC. Given significant heterogeneity of treatment effects of these regimens across NSCLC patients, treatment decisions should be tailored to individual characteristics and preferences to ensure optimal balance between benefits and harms.