FEASIBILITY OF IDENTIFYING GENOTYPE-DEFINED DUCHENNE MUSCULAR DYSTROPHY POPULATIONS USING HUMAN-IN-THE-LOOP LLM AND LINKED CLINICAL NOTES FROM NORSTELLALINQ EHR
Author(s)
Odis Garrett, II, BS1, Eric Mitchell2, ilan behm, MPH3, Raghu R4, Rahul Das4, Allison Perry5.
1Norstella, New York, NY, USA, 2Brooklyn, NY, USA, 3Norstella, Englewood, CO, USA, 4USA, 5New York, NY, USA.
1Norstella, New York, NY, USA, 2Brooklyn, NY, USA, 3Norstella, Englewood, CO, USA, 4USA, 5New York, NY, USA.
OBJECTIVES: To evaluate the feasibility of identifying genotype-defined Duchenne muscular dystrophy (DMD) populations using linked claims, EHR, and human-in-the-loop LLM analysis of clinical notes, to support downstream treatment access and outcomes research.
METHODS: A retrospective cohort study was conducted using NorstellaLinQ’s US real-world linked structured electronic health record (EHR) data, open claims, and unstructured clinical notes processed via LLM techniques. Patients with a DMD diagnosis (ICD-10: G71.00-G71.11) from January 2017 onward were identified across structured EHR and claims sources. A subset with linked clinical notes was isolated for LLM analysis. A curated entity library was applied to identify treatment-relevant exon 44 genotype language using broad pattern matching and explicit exon-specific documentation. LLM outputs underwent human review to distinguish broad candidate mentions from high-confidence genotype-relevant signals.
RESULTS: A total of 25,315 patients with DMD were identified. Of these, 4,170 (16.5%) had linked clinical notes available, representing 1 in 6 diagnosed patients. Among notes-documented patients, LLM identified exon 44-amenable language in 922 patients (22.1%) using broad pattern matching; high-confidence filtering based on explicit exon-specific documentation yielded 221 patients (5.3%), reflecting expected ambiguity in free-text genotype documentation. These findings demonstrate the feasibility of identifying clinically relevant genotype-defined subgroups within linked clinical notes, establishing the notes-documented cohort as the foundation for financial barrier signal extraction and treatment initiation analysis.
CONCLUSIONS: NorstellaLinQ enables rare disease population identification at scale, with clinical notes available for 16.5% of DMD patients. Human-in-the-loop LLM enabled identification of genotype-relevant language within linked clinical notes, demonstrating the feasibility of stratifying rare disease populations beyond what is possible using structured data alone. This approach provides a scalable foundation for future analyses of treatment pathways, access barriers, and real-world outcomes in genetically defined rare disease populations.
METHODS: A retrospective cohort study was conducted using NorstellaLinQ’s US real-world linked structured electronic health record (EHR) data, open claims, and unstructured clinical notes processed via LLM techniques. Patients with a DMD diagnosis (ICD-10: G71.00-G71.11) from January 2017 onward were identified across structured EHR and claims sources. A subset with linked clinical notes was isolated for LLM analysis. A curated entity library was applied to identify treatment-relevant exon 44 genotype language using broad pattern matching and explicit exon-specific documentation. LLM outputs underwent human review to distinguish broad candidate mentions from high-confidence genotype-relevant signals.
RESULTS: A total of 25,315 patients with DMD were identified. Of these, 4,170 (16.5%) had linked clinical notes available, representing 1 in 6 diagnosed patients. Among notes-documented patients, LLM identified exon 44-amenable language in 922 patients (22.1%) using broad pattern matching; high-confidence filtering based on explicit exon-specific documentation yielded 221 patients (5.3%), reflecting expected ambiguity in free-text genotype documentation. These findings demonstrate the feasibility of identifying clinically relevant genotype-defined subgroups within linked clinical notes, establishing the notes-documented cohort as the foundation for financial barrier signal extraction and treatment initiation analysis.
CONCLUSIONS: NorstellaLinQ enables rare disease population identification at scale, with clinical notes available for 16.5% of DMD patients. Human-in-the-loop LLM enabled identification of genotype-relevant language within linked clinical notes, demonstrating the feasibility of stratifying rare disease populations beyond what is possible using structured data alone. This approach provides a scalable foundation for future analyses of treatment pathways, access barriers, and real-world outcomes in genetically defined rare disease populations.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD29
Topic
Clinical Outcomes, Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Health & Insurance Records Systems
Disease
Neurological Disorders, No Additional Disease & Conditions/Specialized Treatment Areas, Rare & Orphan Diseases