(C-215) Methodological Approaches to Capture Real-World Adverse Events (AEs) of first-line (1L) Systemic Treatments in Electronic Health Records (EHRs) in a cohort of Patients With Locally Advanced/Metastatic Urothelial Cancer (la/mUC)

Background: Real-world safety data complement clinical trial findings by offering insights in a broader patient population. However, capturing comprehensive safety data in oncology remains challenging due variations in treatment (tx)-emergent AEs (TEAEs) assessments and data collection methods, such as those used for EHR or claims data, which may affect TEAE incidence reporting.

Objectives: This analysis describes 2 methodological approaches used to capture TEAEs occurring during 1L systemic tx in a cohort of patients with la/mUC.

Methods: This retrospective observational cohort study used the nationwide US Flatiron Health EHRs-derived data collected during routine care in cancer clinics and included adult patients with la/mUC who initiated 1L tx with a regimen of interest after Jan 1, 2016. The incidence of 30 prespecified TEAEs was assessed using 2 methods: 1) International Classification of Disease (ICD) codes (versions 9 and 10) from structured data and 2) a natural language processing machine learning (ML) approach to extract clinical documentation from unstructured data. TEAE data were assessed from the start of 1L tx until the earliest of: 90 days after the final 1L tx dose, initiation of subsequent tx, or death.

Results: 5,235 patients met the study criteria. Using the structured ICD code method, TEAE incidence was 4.1% diarrhea, 13.6% fatigue, 0.2% peripheral neuropathy, 1.5% pruritus, 3.4 rash, and 2.5% weight loss. Using the ML approach, TEAE incidence was 30.7% diarrhea, 70.7% fatigue, 5.2% peripheral neuropathy, 14.5% pruritus, 20.3% rash, and 34.6% weight loss. The relative incidence of TEAEs across 1L tx regimens were similar using both approaches. TEAEs incidence based on the ML approach more closely approximated the values reported in clinical trials of respective tx.

Conclusions: This study highlights the effectiveness of the ML approach in extracting evidence from unstructured data and provides a reliable method for capturing TEAEs in an EHR dataset across 1L tx regimens for la/mUC. Compared with the structured ICD code method, which is designed for billing and reimbursement purposes, the ML method may produce incidence estimates that better reflect real-world clinical patterns. Furthermore, many TEAEs, particularly those associated with immune checkpoint inhibitors, currently lack structured ICD codes, which may result in underreporting of TEAE incidence. As new techniques such as secondary data processing ML are introduced, there is a need for increased transparency regarding the quality of ML-extracted variables, how to promote responsible and effective use of ML-extracted real-world TEAE data, and how to select the best methodological approach for data collection.