FINANCIAL AND ACCESS BARRIERS FROM CLINICAL NOTES USING HUMAN-IN-THE-LOOP LLM: EVIDENCE FROM A PCSK9 INHIBITOR-TREATED COHORT USING NORSTELLALINQ
Author(s)
Odis Garrett, II, BS1, Rahul Das2, Raghu R2, Allison Perry3, ilan behm, MPH4, Eric Mitchell5.
1Norstella, New York, NY, USA, 2USA, 3New York, NY, USA, 4Norstella, Englewood, CO, USA, 5Brooklyn, NY, USA.
1Norstella, New York, NY, USA, 2USA, 3New York, NY, USA, 4Norstella, Englewood, CO, USA, 5Brooklyn, NY, USA.
OBJECTIVES: To characterize the prevalence, distribution, and treatment status associations of financial and access barrier signals in unstructured clinical notes using a human-in-the-loop large language model (LLM) applied to a PCSK9 inhibitor-treated cohort in NorstellaLinQ.
METHODS: A retrospective cross-sectional analysis was conducted using NorstellaLinQ's real-world US linked structured EHR data with free-text clinical notes processed through the LLM engine. Among 95,998 patients with documented PCSK9 inhibitor encounters, a curated LLM entity library extracted and harmonized barrier signals across seven pre-specified domains: out-of-pocket cost burden, insurance coverage status, prior authorization and access barriers, adherence barriers, treatment pending reasons, discontinuation reasons, and switch reasons. Domains are not mutually exclusive; multi-barrier burden was quantified by counting co-occurring domains per patient. Treatment status was categorized as active, pending, discontinued, or denied.
RESULTS: Of 95,998 patients, 50,036 (52.1%) had at least one barrier domain documented. Treatment pending reasons were most prevalent in 33,122 patients (34.5%), insurance-related reasons driving 19,527 cases. Cost barriers were present in 12,038 (12.5%); prohibitive cost cited in 6,227. Access barriers, dominated by prior authorization, were present in 8,415 (8.8%). Nearly one in four (23.9%) carried two or more co-occurring domains. Among those with at least one barrier, 19.6% had active treatment recorded, 32.3% remained pending, 11.1% discontinued, and 7.6% had an insurance denial documented, suggesting many never initiated or sustained therapy, though longitudinal follow-up was not assessed.
CONCLUSIONS: Human-in-the-loop LLM enables scalable extraction of clinically meaningful financial and access barriers not routinely captured in structured real-world data. Barrier documentation frequently coincided with non-active treatment status, with insurance-related barriers and prohibitive cost as the most frequently documented contributors to treatment disruption. These signals are largely unavailable in structured claims or dispensing records alone. These findings establish a scalable human-in-the-loop LLM framework for high-cost therapeutic areas, with implications for patient support, utilization management, and health technology assessment.
METHODS: A retrospective cross-sectional analysis was conducted using NorstellaLinQ's real-world US linked structured EHR data with free-text clinical notes processed through the LLM engine. Among 95,998 patients with documented PCSK9 inhibitor encounters, a curated LLM entity library extracted and harmonized barrier signals across seven pre-specified domains: out-of-pocket cost burden, insurance coverage status, prior authorization and access barriers, adherence barriers, treatment pending reasons, discontinuation reasons, and switch reasons. Domains are not mutually exclusive; multi-barrier burden was quantified by counting co-occurring domains per patient. Treatment status was categorized as active, pending, discontinued, or denied.
RESULTS: Of 95,998 patients, 50,036 (52.1%) had at least one barrier domain documented. Treatment pending reasons were most prevalent in 33,122 patients (34.5%), insurance-related reasons driving 19,527 cases. Cost barriers were present in 12,038 (12.5%); prohibitive cost cited in 6,227. Access barriers, dominated by prior authorization, were present in 8,415 (8.8%). Nearly one in four (23.9%) carried two or more co-occurring domains. Among those with at least one barrier, 19.6% had active treatment recorded, 32.3% remained pending, 11.1% discontinued, and 7.6% had an insurance denial documented, suggesting many never initiated or sustained therapy, though longitudinal follow-up was not assessed.
CONCLUSIONS: Human-in-the-loop LLM enables scalable extraction of clinically meaningful financial and access barriers not routinely captured in structured real-world data. Barrier documentation frequently coincided with non-active treatment status, with insurance-related barriers and prohibitive cost as the most frequently documented contributors to treatment disruption. These signals are largely unavailable in structured claims or dispensing records alone. These findings establish a scalable human-in-the-loop LLM framework for high-cost therapeutic areas, with implications for patient support, utilization management, and health technology assessment.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD81
Topic
Health Service Delivery & Process of Care, Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Health & Insurance Records Systems
Disease
Biologics & Biosimilars, Cardiovascular Disorders (including MI, Stroke, Circulatory), No Additional Disease & Conditions/Specialized Treatment Areas