(281) Patient Characteristics, Outcomes and Potential for Bias in Oncological Real-World Data Studies with Imaging-Derived Response Endpoints

Publication Number: 955

Background: Response endpoints in solid tumor trials are typically assessed by RECIST, based on the measurement of tumor sizes in imaging studies obtained at baseline and during follow-up (FU). Assessments in routine care may be different from clinical trials as the schedules and measurement of imaging studies may be less strictly predefined and little is known about the impact of imaging surveillance and risk for immortal time bias (ITB) by conditioning on patients with FU scans.

Objectives: To assess differences in patient characteristics and overall survival (OS) among patients with and without FU scans.

Methods: Using the nationwide Flatiron Health EHR-derived de-identified database, advanced non-small cell lung cancer patients with any first-line (1L) antineoplastic therapy on or before March 2021 were selected. The index date was defined as the date of 1L initiation. Patients were eligible if they had available radiological images in the picture archiving and communication system (PACS) with a CT or PET baseline scan within 8 weeks prior to index date. Differences in patient characteristics between patients with and without FU scan during 1L were assessed using standardized mean differences (SMD). OS between both cohorts was compared employing Kaplan-Meier estimates and multivariate Cox proportional hazards regression. To empirically assess the magnitude of ITB and the potential influence of imaging surveillance on OS outcomes, naive and ITB-adjusted estimates were compared using time-dependent and landmark models.

Results: Overall, 4,074 patients were eligible of which 2,874 (70.5%) had ≥1 FU scan and 1,200 (29.5%) did not have any FU scan after baseline. Patients with ≥1 FU scan had longer periods to initiate 1L after diagnosis, lower ECOG categories, were less likely to present with metastatic disease at initial diagnosis and were younger (all SMDs>0.1). In the naive analysis, median OS for patients with ≥1 FU scan was significantly longer with 16.1 (95% CI 15.1-17.0) versus 3.1 (95% CI 2.7-3.5) months for patients without FU scan (HR 0.34 [95% CI 0.32-0.37]). After adjusting for ITB, survival differences were less pronounced, however, patients with ≥1 FU scan still showed significantly longer survival across all models (time-dependent model HR 0.62 [95% CI 0.57 - 0.68] and HRs [95% CIs] ranging from 0.89 [0.81-0.98] to 0.66 [0.56-0.79] in 3 and 12 month landmark models, respectively).

Conclusions: Patients with ≥1 FU scan after 1L treatment initiation showed significant differences in baseline characteristics and OS compared to patients without FU scans, even after ITB adjustment. While further aspects may impact this association, conditioning on patients with FU scans introduces ITB and selection bias towards a population with more favorable prognosis.