DISABILITY AND FUNCTIONAL STATUS INVISIBLE IN CLAIMS: CAPTURING LONGITUDINAL EDSS IN MULTIPLE SCLEROSIS WITH A HUMAN-IN-THE-LOOP LARGE LANGUAGE MODEL APPLIED TO NORSTELLALINQ US ELECTRONIC HEALTH RECORDS

Author(s)

Becky Hollenberg, MPH1, Sarah Bach, MSE1, ilan behm, MPH2, Rahul Das, PhD1, Allison Perry, PhD1.
1Norstella, New York, NY, USA, 2Norstella, Englewood, CO, USA.
OBJECTIVES: Administrative claims capture diagnoses, procedures, and dispensed therapies but miss what matters most in multiple sclerosis (MS): disability severity and progression. The Expanded Disability Status Scale (EDSS), the principal measure of MS disability, is typically documented only in narrative clinical notes, leaving severity and longitudinal progression unavailable for claims-based analyses. We evaluated whether a human-in-the-loop large language model (LLM) could extract EDSS at scale from NorstellaLinQ electronic health record (EHR) notes and characterize disability across MS phenotypes over time.
METHODS: Patients with a structured MS diagnosis (ICD-10 G35) in NorstellaLinQ EHR were identified. Free-text clinical notes were processed using human-in-the-loop LLM extraction to identify EDSS scores and MS phenotypes. EDSS values were normalized to the 0-10 scale (0.5-point increments) and time-stamped. We quantified patient and record yield, EDSS distribution (mild ≤3.5, moderate 4.0-5.5, severe ≥6.0), phenotype-specific disability, and longitudinal assessment availability.
RESULTS: Human-in-the-loop LLM extraction identified 43,341 valid EDSS measurements from 38,865 clinical notes across 12,654 patients. Among patients with MS, 12,192 had at least one EDSS score recoverable exclusively from narrative clinical notes. Median EDSS was 3.0 (IQR 2.0-6.0), with 58.3% classified as mild, 14.9% moderate, and 26.8% severe. Median EDSS followed the expected clinical severity gradient across phenotypes (clinically isolated syndrome 1.5, relapsing-remitting 2.5, primary progressive 6.0, secondary progressive 6.5). Overall, 57.5% of patients had at least two longitudinal EDSS assessments over a median follow-up of 22.2 months.
CONCLUSIONS: Human-in-the-loop LLM extraction recovered longitudinal EDSS measurements at scale from routine clinical documentation, generating clinically coherent disability distributions and trajectories across MS phenotypes. These data provide structured measures of disability severity and progression unavailable in claims alone, enabling real-world studies of MS progression, comparative effectiveness, and disability burden.

Conference/Value in Health Info

2026-11, ISPOR Europe 2026, Vienna, Austria

Value in Health, Volume 29, Issue 12S

Code

RWD125

Topic

Real World Data & Information Systems

Topic Subcategory

Health & Insurance Records Systems

Disease

Neurological Disorders

Your browser is out-of-date

ISPOR recommends that you update your browser for more security, speed and the best experience on ispor.org. Update my browser now

×