Use of Sensitivity Analysis to Address Potential Missing Data and Selection Bias from Linked Genomics Datasets in a Real-World Study of Patients Receiving Immune-Checkpoint Inhibitors for Metastatic Melanoma

OBJECTIVES: As genomic testing in melanoma has evolved, BRAF and MAPK pathway mutations have become a key focus for real-world evidence (RWE). However, integrating genomic information into RWE presents challenges as data often originate outside the electronic health record (EHR). Single-gene testing and variability in documentation of wild-type results adds further complexity. Additionally, linked datasets can introduce selection bias since complete data may only be available for a subset of patients. This study investigates the use of sensitivity analysis to address such bias from a linked genomics database in patients treated with immune-checkpoint inhibitors (ICI) for metastatic melanoma.
METHODS: This was a retrospective observational cohort study of patients with metastatic melanoma, indexed at initiation of first-line ICI between 1/1/16-6/30/22 in The US Oncology Network and non-Network practices (followed through 6/30/24). A structured EHR database was linked with a genomics dataset from select next-generation sequencing labs. The target population (Cohort A) included patients with one mutation of interest (BRAF, NRAS, NF-1, and c-KIT) or confirmed quadruple-negative. A sensitivity analysis was conducted among patients with documented testing for all four mutations of interest (Cohort B). Patient characteristics were descriptively assessed in each cohort.
RESULTS: Overall, 990 and 291 patients were included in Cohorts A and B with mean (SD) ages of 64 (14) and 67 (12) years, respectively. Cohort A included more patients indexed between 2016-2018 (n=340, 34%) relative to Cohort B (n=32, 11%). No substantial differences in melanoma subtype, primary or metastatic sites, body mass index, performance status or lactate dehydrogenase were observed between cohorts.
CONCLUSIONS: Sensitivity analysis revealed that partial linked data better represented the full study period with minimal differences in prognostic factors relative to complete case analysis. Inclusion of such explorations can be an important tool for assessing selection bias in RWE studies and supporting generalizability of results when using linked datasets.