IMPROVING REAL-WORLD IDENTIFICATION AND MOLECULAR CHARACTERIZATION OF DIFFUSE MIDLINE GLIOMA USING HUMAN-IN-THE-LOOP LLM EXTRACTION OF NORSTELLALINQ UNSTRUCTURED ELECTRONIC HEALTH RECORD CLINICAL NOTES
Author(s)
Bharath P, MPH1, Allison Perry, MPH2, Eric Mitchell, MPH3, ilan behm, MPH4.
1Norstella, Bengaluru, India, 2Norstella, New York, NY, USA, 3Norstella, Brooklyn, NY, USA, 4Norstella, Englewood, CO, USA.
1Norstella, Bengaluru, India, 2Norstella, New York, NY, USA, 3Norstella, Brooklyn, NY, USA, 4Norstella, Englewood, CO, USA.
OBJECTIVES: Diffuse midline glioma (DMG), an H3K27-altered WHO Grade IV central nervous system tumor, is an ultra-rare malignancy with a median overall survival of approximately 9-13 months. Accurate identification of DMG in real-world data is challenging because structured diagnosis coding incompletely captures evolving molecular nomenclature (e.g., H3K27M, H3K27-altered), legacy terminology (e.g., diffuse intrinsic pontine glioma [DIPG]), and disease-specific anatomical descriptors. This study evaluated whether human-in-the-loop large language model (LLM) extraction of unstructured clinical notes could improve real-world identification and molecular characterization of patients with DMG.
METHODS: A retrospective observational study was conducted using NorstellaLinQ US unstructured electronic health record clinical notes collected between 2016 and 2026. A comprehensive terminology framework incorporating molecular markers, historical disease terminology, WHO classification, and anatomical descriptors was developed for human-in-the-loop LLM extraction and concept normalization. Patients with confirmed DMG-related documentation were included. Molecular status (H3K27M-positive, H3K27M-negative, indeterminate, wildtype, and tested only) was extracted from free-text clinical documentation.
RESULTS: Human-in-the-loop LLM extraction identified 2,513 unique patients with documented DMG. Among these patients, 2,141 (85.2%) had documentation consistent with H3K27M-positive disease, 183 (7.3%) were H3K27M-negative, 14 (0.6%) had indeterminate results, and 7 (0.3%) were documented as H3-wildtype; an additional 102 patients had documented H3K27M testing without definitive molecular classification. The identified cohort represents substantial capture of this ultra-rare malignancy within a national real-world database and demonstrates that unstructured clinical documentation contains clinically relevant molecular, historical, and anatomical disease information not routinely available through structured diagnosis coding alone.
CONCLUSIONS: Human-in-the-loop LLM extraction of unstructured clinical notes enables scalable identification and molecular characterization of patients with DMG beyond what is typically available in structured electronic health record data. Incorporating unstructured clinical documentation into real-world evidence workflows may improve rare disease cohort construction, clinical trial feasibility, epidemiologic studies, and treatment pattern research for molecularly defined central nervous system malignancies.
METHODS: A retrospective observational study was conducted using NorstellaLinQ US unstructured electronic health record clinical notes collected between 2016 and 2026. A comprehensive terminology framework incorporating molecular markers, historical disease terminology, WHO classification, and anatomical descriptors was developed for human-in-the-loop LLM extraction and concept normalization. Patients with confirmed DMG-related documentation were included. Molecular status (H3K27M-positive, H3K27M-negative, indeterminate, wildtype, and tested only) was extracted from free-text clinical documentation.
RESULTS: Human-in-the-loop LLM extraction identified 2,513 unique patients with documented DMG. Among these patients, 2,141 (85.2%) had documentation consistent with H3K27M-positive disease, 183 (7.3%) were H3K27M-negative, 14 (0.6%) had indeterminate results, and 7 (0.3%) were documented as H3-wildtype; an additional 102 patients had documented H3K27M testing without definitive molecular classification. The identified cohort represents substantial capture of this ultra-rare malignancy within a national real-world database and demonstrates that unstructured clinical documentation contains clinically relevant molecular, historical, and anatomical disease information not routinely available through structured diagnosis coding alone.
CONCLUSIONS: Human-in-the-loop LLM extraction of unstructured clinical notes enables scalable identification and molecular characterization of patients with DMG beyond what is typically available in structured electronic health record data. Incorporating unstructured clinical documentation into real-world evidence workflows may improve rare disease cohort construction, clinical trial feasibility, epidemiologic studies, and treatment pattern research for molecularly defined central nervous system malignancies.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD102
Topic
Epidemiology & Public Health, Real World Data & Information Systems, Study Approaches
Topic Subcategory
Health & Insurance Records Systems
Disease
Neurological Disorders, Oncology, Rare & Orphan Diseases