LLM-curated clinical notes reveal lower real-world semaglutide and tirzepatide persistence than prescription data

The real-world effectiveness of glucagon-like peptide-1 receptor agonists (GLP-1RA) in managing metabolic health and weight hinges on actual treatment initiation and sustained patient exposure. However, traditional structured prescription records often capture prescribing intent rather than confirmed use, leading to an overestimation of adherence. This gap prevents a true understanding of patient journeys, including barriers to therapy like cost, insurance issues, or adverse events. Leveraging large language models (LLMs) to analyze unstructured clinical notes offers a novel approach to bridge this data gap, providing a more accurate picture of real-world medication use.

Researchers deployed an evidence-linked large language model (LLM) pipeline across a federated, de-identified US EHR platform encompassing over 29 million patients. The LLM's accuracy was rigorously evaluated against an independent physician-adjudicated reference standard, achieving 98.4% for structural chunk triage and 88.2% for medication-status adjudication. The study cohort included 553,073 adults with at least one semaglutide (n=376,697) or tirzepatide (n=176,376) prescription and baseline weight data. They analyzed documented initiation within ±3 months of the first structured prescription and assessed treatment persistence over an 18-month period in a subset of 69,976 patients identified with "frictionless initiation." The primary endpoint was comparing clinical note-derived versus structured prescription-derived ascertainment of initiation and persistence.

Documented initiation within ±3 months of the first structured prescription was identified in 141,189 semaglutide patients and 62,040 tirzepatide patients. Among these, "frictionless starts" accounted for 70.1% of semaglutide and 77.9% of tirzepatide initiations. Initiation after documented friction (e.g., prior authorization delays) was observed in 17.2% of semaglutide and 15.9% of tirzepatide patients, while early interruption within 3 months occurred in 12.7% and 6.2%, respectively. A critical finding emerged from the persistence analysis over 18 months:

Clinical note-based ascertainment revealed significantly lower persistence rates compared to structured prescription data, with 42.3% of semaglutide and 43.1% of tirzepatide patients continuing therapy, versus 55.4% for semaglutide and 58.3% for tirzepatide based on prescription records (log-rank P < .001 for both comparisons). The LLM also identified common barriers to persistence from notes, including insurance, cost, perioperative holds, and gastrointestinal adverse events.