Abstract 6199: Defining and capturing progression in glioma by harnessing NLP in unstructured electronic health records

Abstract

Abstract Gliomas exhibit nearly uniform recurrence and poor prognosis. Management of high-grade tumors is surgery followed by chemoirradiation (CRT). Radiographic progression is assessed using contrast-enhanced MRI with reporting captured in Electronic Health Records (EHR). The ability to harness large scale EHR data is limited by the Response Assessment in Neuro-Oncology (RANO) tumor progression criteria, which defines progression as an aggregate of clinical and/or radiographic parameters requiring clinician judgement. As a result, glioma progression is not captured systematically in large data sets, limiting Progression Free Survival (PFS) as an outcome endpoint in data analysis. We developed an AI-based method using natural language processing to capture PFS parameters for analysis of MRI radiology reports. 1088 available brain MRI radiology reports for 81 patients with a pathologically confirmed diagnosis of glioblastoma (GBM) were aggregated in the NIH Integrated Data Analysis Platform. MRI reports were systematically analyzed and PFS manually captured using RANO criteria as ground truth. Common report terms indicating progression were compiled and included in task prompts applied to Large Language Model (LLM) Extraction Tools. The words progression (n=1047) and stable (n=1133) did not necessarily indicate either overall progression or stability in a report. Yet, in the 24 (30%) patients who progressed within 3 months of CRT, recurrence and mass effect occurred in 3.7% and 53% of reports, respectively, which was comparable to the term frequencies of 5.8% and 54% in patients with stable disease. Thus, mass effect and recurrence could not necessarily be used alone as terms to signal overall progression. The analysis identified other commonly employed radiology report terms relevant to determining tumor progression including enhancement (n=2445), perfusion (n=2085), enhancing (n=1785), resection (n=1576), and increased (n=1331), which allowed for experimentation with more detailed LLM extraction task prompts. Our method flagged radiographic progression prior to clinician-coded progression in 43 (53%) patients. Future directions include handling further clinical context such as surgical features, radiation therapy dates, progress notes, and the coadministration of agents such as bevacizumab and steroids. These results indicate the feasibility of LLMs to identify tumor progression dates using EHR text with potential transferability to large glioma data sets pending further optimization and validation. Citation Format: Shreya Chappidi, Hawon Lee, Sarisha Jagasia, Casey Syal, George Zaki, Dylan Junkin, Nathan Golightly, Patrick Chitwood, Kevin Camphausen, Andra Krauze. Defining and capturing progression in glioma by harnessing NLP in unstructured electronic health records [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 6199.