6D-04 - Assessing Current Biomarker Testing Rates Leveraging Structured and Unstructured Real-World Data (RWD): A Case Study of Programmed Death-Ligand 1 (PD-L1) Testing Trends for Metastatic Non-Small Cell Lung Cancer (mNSCLC)

Background: Surveillance of biomarker testing is required to actualize the benefits of precision medicine, particularly in the ever-evolving oncology treatment landscape. While RWD can provide critical insights into biomarker testing patterns, often this information requires access to unstructured data, which may limit or delay its use for research.

Objectives: To evaluate PD-L1 testing documentation in electronic health records (EHR) using multiple data capture methods, including a natural language processing (NLP) algorithm.

Methods: This retrospective, observational study analyzed adult patients diagnosed with mNSCLC initiating first-line treatment between 11/01/2022 and 08/31/2024 in a large network of US community oncology clinics. Data were sourced from the iKnowMed EHR through structured and unstructured data extraction. A previously validated NLP algorithm, based on GPT-4o with Azure Document, captured PD-L1 documentation from unstructured records and the findings were compared with manual abstraction (gold standard) for 100 patients without structured records of PD-L1 testing selected through a stratified random sampling method. Data were analyzed descriptively.

Results: Among 2,148 study-eligible patients, PD-L1 testing was documented in structured EHR records for 75% (n=1,607). Among a stratified random sample of 100 patients without structured records of PD-L1 testing, 62% were found to have unstructured records of testing through chart abstraction and NLP identified evidence of PD-L1 testing among 62% (sensitivity 89% [95% confidence interval (CI) 79%-95%]; specificity 86% [95% CI 71%-95%]; F1 score 90%). In the broader population of patients without structured records of PD-L1 testing (n=541), 58% (n=313) had documentation of PD-L1 testing using NLP. Across the study population, based on structured EHR data and NLP, an estimated 89% (n=1,920) had evidence of PD-L1 testing.

Conclusions: The observed PD-L1 testing rate climbed from 75% to 89% with multiple data sources, underscoring the criticality of leveraging unstructured EHR data, especially in oncology. Additionally, NLP effectively captured PD-L1 testing among patients that lacked structured records, suggesting that it can potentially replace or supplement manual abstraction for this population.