(053) Comparative Analysis on incident risk of Anemia and Immune-Mediated Hepatitis in non-small cell lung cancer (NSCLC): Enhancing Cross-Data Source Comparability in PD-1/PD-L1 Therapy

Background: A primary responsibility of a Clinical Safety Scientist involves assessing drug safety signals, relying on information from diverse sources, including clinical trials, real-world data, spontaneous reporting data, and published literature. A major challenge in this role stems from the difficulty in comparing results across varied data sources due to differences in study design, patient populations, drug exposures, and the nature of the databases themselves.

Objectives: This study employs propensity score weighting using SuperLearner along with advanced statistical methods and machine learning techniques to demonstrate the feasibility of comparative analyses across diverse data types.

Methods: This retrospective, secondary data use observational study assessed and compared incident risk of anemia and immune-mediated hepatitis in Stage IIIb/IV NSCLC patients treated with PD-1/PD-L1 drugs. Data from the TAIL study (NCT03285763), the nationwide, electronic health record-derived, de-identified Flatiron Health (FH) NSCLC datamart, and the Roche ARISg database (up to August 31, 2023) were analyzed. MICE/missRanger handled missing data imputation, with Out-Of-Bag mean prediction error indicating imputation performance. Weights were calculated using the SuperLearner method via R's WeightIt package, employing specific submodels: mean model, random forest, xgboost, and LASSO for Hepatitis-specific analysis, and LASSO, xgboost, and recursive partitioning and regression trees for Anemia. Covariate balance was checked using Love plots, and marginal effects were measured with weighted incident risks with 95% percentile bootstrapping confidence intervals (CIs) in the Average Treatment effect among the Treated, and then compared across data sources.

Results: The effectiveness in balancing cohorts fluctuated based on the weighting methods used, model specifics, and the patterns of missing data. Weighting was favored for its potential to yield a higher effective sample size than matching. For Anemia, estimated crude risks were 0.036 (95% CI: 0.033, 0.038) for ARISg, 0.003 (0.003, 0.004) for FH, and 0.126 (0.118, 0.134) for TAIL. The adjusted risks post-weighting were 0.055 (0.045, 0.067) for ARISg and 0.003 (0.002, 0.003) for FH. Regarding hepatitis, estimated crude risks were 0.085 (0.074, 0.098) for ARISg, 0.006 (0.004, 0.008) for FH, and 0.016 (0.009, 0.030) for TAIL, with adjusted risks of 0.068 (0.048, 0.097) for ARISg and 0.006 (0.004, 0.010) for FH.

Conclusions: Traditional approaches to directly compare analysis results from different data sources tend to be entrenched, time-consuming, and often inaccurate; however, it is a crucial task for effective safety signal management and causality assessment. This feasibility study demonstrates that, with the application of appropriate consideration and methodology to control for potential confounding and biases, it is possible to build robust analyses and allow confident decision-making in scenarios where direct comparisons across varied data sources are challenging.