PHENOTYPING TENOSYNOVIAL GIANT CELL TUMOR (TGCT) AT POPULATION SCALE: VALIDATED LARGE LANGUAGE MODEL (LLM) EXTRACTION FROM US REAL-WORLD CLINICAL NOTES
Author(s)
Atharva R. Manjrekar, MS, Nicole K. Hobbs, PhD, Shrinal Patel, MD, Rahul K. Das, PhD.
Norstella, New York, NY, USA.
Norstella, New York, NY, USA.
OBJECTIVES: Localized tenosynovial giant cell tumor (TGCT) has no diagnosis code, so structured data alone under-ascertain the population. We aimed to show that unstructured clinical notes can identify and characterize TGCT patients at population scale, validated against established clinical knowledge and an independent modality.
METHODS: Within NorstellaLinQ — a US real-world dataset linking open claims, structured EHR, clinical notes, and pharmacy fills — human-in-the-loop large language model (LLM) extraction of notes identified 4,378 confirmed TGCT patients (2019-June 2026); subtype was resolved in 2,724 (62%): diffuse 1,458, localized 1,266. Core fields—diagnosis, subtype, surgical history and candidacy, and adverse events—were clinical-expert-validated at mean F1 ≥ 95%, and corroborated against an independent claims-and-EHR procedure stream.
RESULTS: Subtype tracked known recurrence biology: among subtype-resolved patients, 30.9% of diffuse had ≥1 synovectomy and 5.8% a recurrence proxy (≥2 dates), versus 2.5% and 0.4% of localized. Claims and EHR independently documented ≥1 synovectomy in 599 patients; of 3,409 surgery-documented patients, 82.7% appeared only in notes—historical or out-of-system operations. Notes captured surgical candidacy: not a candidate in 31.7% of diffuse versus 11.4% of localized, consistent with diffuse disease's lower resectability. Adverse events matched the CSF1R class profile: among 247 on an approved CSF1R inhibitor (pexidartinib or vimseltinib; imatinib-only excluded), 55.9% had a documented AE, including dermatologic (26.7%), fatigue (24.7%), and class-characteristic periorbital edema (11.3%). Systemic therapy concentrated in diffuse disease—11.6% versus 0.9% of localized on a systemic CSF1R agent—consistent with guideline-reserved positioning. Among treated diffuse, 50.3% had multiple surgeries. No TGCT diagnosis in claims or EHR was recorded for 53.4% of patients—including 78% of localized cases—identifiable only from notes.
CONCLUSIONS: LLM-based extraction of clinical notes phenotyped TGCT, enabling payers and HTA bodies to size the population, anticipate recurrence-driven surgical burden, and flag the CSF1R-class tolerability signal—actions a code-based view cannot support. Findings are descriptive; AEs reflect co-occurrence, not incidence.
METHODS: Within NorstellaLinQ — a US real-world dataset linking open claims, structured EHR, clinical notes, and pharmacy fills — human-in-the-loop large language model (LLM) extraction of notes identified 4,378 confirmed TGCT patients (2019-June 2026); subtype was resolved in 2,724 (62%): diffuse 1,458, localized 1,266. Core fields—diagnosis, subtype, surgical history and candidacy, and adverse events—were clinical-expert-validated at mean F1 ≥ 95%, and corroborated against an independent claims-and-EHR procedure stream.
RESULTS: Subtype tracked known recurrence biology: among subtype-resolved patients, 30.9% of diffuse had ≥1 synovectomy and 5.8% a recurrence proxy (≥2 dates), versus 2.5% and 0.4% of localized. Claims and EHR independently documented ≥1 synovectomy in 599 patients; of 3,409 surgery-documented patients, 82.7% appeared only in notes—historical or out-of-system operations. Notes captured surgical candidacy: not a candidate in 31.7% of diffuse versus 11.4% of localized, consistent with diffuse disease's lower resectability. Adverse events matched the CSF1R class profile: among 247 on an approved CSF1R inhibitor (pexidartinib or vimseltinib; imatinib-only excluded), 55.9% had a documented AE, including dermatologic (26.7%), fatigue (24.7%), and class-characteristic periorbital edema (11.3%). Systemic therapy concentrated in diffuse disease—11.6% versus 0.9% of localized on a systemic CSF1R agent—consistent with guideline-reserved positioning. Among treated diffuse, 50.3% had multiple surgeries. No TGCT diagnosis in claims or EHR was recorded for 53.4% of patients—including 78% of localized cases—identifiable only from notes.
CONCLUSIONS: LLM-based extraction of clinical notes phenotyped TGCT, enabling payers and HTA bodies to size the population, anticipate recurrence-driven surgical burden, and flag the CSF1R-class tolerability signal—actions a code-based view cannot support. Findings are descriptive; AEs reflect co-occurrence, not incidence.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR153
Topic
Epidemiology & Public Health, Methodological & Statistical Research, Study Approaches
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
Rare & Orphan Diseases