OBJECTIVES: Radiology imaging is critical to diagnose and monitor cancer. However, real-world radiographic imaging data availability and timing for patients may vary, and selecting real-world patients based on imaging availability may introduce biases, potentially impacting generalizability of results. Therefore, we assessed the representativeness of imaging-derived cohorts relative to a broader real-world oncology target population.

METHODS: We used the nationwide Flatiron Health electronic health record (EHR)-derived de-identified Research Database (FHRD) linked to retrospective radiographic imaging data to select two broad study cohorts, patients with advanced non-small cell lung cancer (aNSCLC) and patients with diffuse large B-cell lymphoma (DLBCL), both with the receipt of first-line (1L) treatment as of November 30, 2021. For each respective cohort, our comparisons were based on imaging availability at various time points: any time point, baseline (pre-1L initiation), baseline and post-baseline (during 1L duration). Differences in baseline demographic, clinical, and treatment characteristics between groups were measured through the absolute standardized mean difference (ASMD), with a threshold of 0.1.

RESULTS: There were 12,170 and 1,508 patients in the FHRD imaging-linked aNSCLC and DLBCL cohorts. For both diseases, patients with scans at any time point or in baseline window were broadly representative to the respective target. However, requiring scans at both baseline and post-baseline windows resulted in some covariate and outcome differences with the target cohort. For the aNSCLC cohort, longer median rwOS (+4 months) was observed. For the DLBCL cohort, we saw an increase in lower ECOG performance scores (+20%).

CONCLUSIONS: Requiring scans at any time point or in the baseline window resulted in the cohort being representative relative to the broader real-world cohort; however, requiring scans in the post-baseline window may introduce selection and immortal time bias. Further work is needed to assess the application of existing methods to account for such biases in this context.