Presentation (CTI)

OBJECTIVES: To compare the risk of immune related adverse events between commonly used immunotherapy regimens in advanced NSCLC and to assess whether treatment associated irAE risk varies across patients using heterogeneous treatment effect (HTE) methods applied to real world data.
METHODS: We conducted a retrospective cohort study of 174 patients with advanced NSCLC receiving platinum-based chemotherapy in combination with either pembrolizumab (n=71) or ipilimumab-nivolumab (n=103). Immune-related adverse events were defined as the occurrence of ≥1 clinically relevant toxicity (including dermatologic, gastrointestinal, pulmonary, renal, hepatic, or endocrine events). To adjust for non-random treatment assignment, propensity scores were estimated using baseline demographics, disease characteristics, biomarkers, and clinical factors, and overlap weighting was applied to balance covariates between treatment groups. The primary estimand was the propensity-adjusted average treatment effect (ATE) on the absolute risk scale. To assess heterogeneity, a causal machine-learning model was used to estimate conditional average treatment effects (CATEs). A global permutation-based ABC test evaluated whether observed heterogeneity exceeded what would be expected under a constant treatment effect. When heterogeneity was detected, SHapley Additive exPlanations (SHAP) were used descriptively to identify baseline features contributing to variability in predicted treatment-associated irAE risk.
RESULTS: After propensity adjustment, ipilimumab-nivolumab was associated with a higher irAE risk than pembrolizumab, with an absolute risk difference of 0.22 (95% CI: 0.06-0.37). The ABC test indicated statistically significant heterogeneity in treatment-associated irAE risk across patients (p=0.046). Model explanation analyses suggested that tumor mutational burden, smoking status, sex, and contralateral lung involvement contributed to variability in predicted irAE risk differences, with several features associated with larger predicted risk increases under ipilimumab-nivolumab.
CONCLUSIONS: Real-world causal machine-learning methods reveal substantial heterogeneity in immunotherapy-associated irAE risk beyond average treatment differences in advanced NSCLC. Identifying patient subgroups with elevated treatment-specific toxicity risk may support more personalized treatment selection, monitoring strategies, and value assessments in oncology.