Get started
Sign in

Leveraging NLP to characterize real-world triptan use and adherence from unstructured EHR data

by | Nov 9, 2025 | Research, Research Insights

Authors: Katherine Brown, PhD, MSN, RN ⊕Truveta, Inc, Bellevue, WA, Sarah Platt, MS ⊕Truveta, Inc, Bellevue, WA, Agnes Pastwa ⊕Truveta, Inc, Bellevue, WA, Sunny Guin, PhD ⊕Truveta, Inc, Bellevue, WA, Emily Webber, PhD ⊕Truveta, Inc, Bellevue, WA

Banner image with the text “Characterizing real-world triptan use” on a gradient blue background, featuring Truveta’s logo with a dew-covered leaf pattern. Represents research on migraine treatment patterns using real-world electronic health record (EHR) data.
  • Advanced natural language processing (NLP) applied to unstructured EHR notes reveals real-world triptan use and adherence patterns in migraine patients.
  • Triptan adherence is variable, with 26% of patients discontinuing due to ineffectiveness (55%) or side effects (28%).
  • The study highlights demographic disparities and the value of large-scale, ontology-aligned data for optimizing migraine care.

This blog extends findings from “Leveraging NLP to characterize real-world triptan use and adherence from unstructured EHR” presented at ISPOR Europe 2025.

Migraines impose a significant burden on health systems, with high costs stemming from underdiagnosis and suboptimal treatments. Triptans are a standard acute therapy, but their real-world use and patient adherence are highly variable. Traditional structured EHR data often miss the nuanced information needed to understand treatment behavior at scale (1).

Recent advances in artificial intelligence (AI) and natural language processing (NLP) have made it possible to extract clinically meaningful data from unstructured text within EHRs—unlocking insights that were previously hidden in clinical notes. This study used Truveta Data to examine real-world triptan adherence patterns among patients with migraine, demonstrating how NLP-based approaches can bridge critical evidence gaps and enhance understanding of real-world migraine care.

Methods

We developed a natural language processing (NLP) model to extract triptan use and adherence concepts from unstructured clinical notes. Using a subset of Truveta Data, clinical notes were selected based on migraine or headache diagnoses and mentions of triptan medications. Using the Truveta Language Model (TLM), the system extracted medication mentions, treatment attributes, and adherence concepts. These extracted concepts were then mapped to standard clinical ontologies through zero-shot normalization (2). Model performance was validated against expert annotations, with accuracy metrics evaluated using precision and recall.

Results

More than 310,000 clinical notes were evaluated. The model achieved a precision of 86.8%, recall of 86.1%, and overall accuracy of 76.1% when identifying and normalizing triptan adherence concepts. These results indicate strong model performance and reliability in capturing complex, narrative clinical data related to migraine treatment.

Demographic analysis showed that the majority of patients represented in the notes were female (86.3%), White (77.5%), and aged 30–49 years, aligning with established patterns of migraine prevalence.

However, disparities were observed, with notable gaps among older and racially diverse populations, suggesting potential issues related to documentation and access to care. Regarding treatment adherence, 26% of patients discontinued triptan therapy, most commonly due to perceived ineffectiveness (55%) and side effects (28%).

Demographics and adherence status
Table summarizing patient demographics and adherence for 312,559 triptan users identified in Truveta Data. Age distribution: 30–39 (24%), 40–49 (22%), 50–59 (15%). 82% female, 17% male. 94% of triptan patients actively taking medication, 27% stopped, 0.5% not taking, 0.01% unknown. Compares triptan users with a 10-million-patient migraine or headache cohort.

Discussion

This study demonstrates the potential of AI-driven NLP to transform unstructured EHR data into actionable clinical insights. By applying the Truveta Language Model, researchers extracted detailed information on triptan adherence—a key but often undocumented aspect of migraine care.

Applying advanced NLP to unstructured clinical data enables high-fidelity capture of real-world treatment behaviors. This scalable, ontology-aligned approach supports pharmacovigilance, patient segmentation, and population health management. The findings offer new opportunities to inform health policy and optimize migraine care. Disparities in adherence and documentation highlight areas for further research and intervention.

Clinically, these findings emphasize the importance of real-world data in understanding how patients use migraine therapies and where adherence challenges persist. Technologically, they highlight how innovations in AI and language modeling can extend the reach of real-world evidence, offering a scalable pathway to more personalized and data-informed care strategies.

These findings are consistent with data accessed on November 10, 2024.  They are preliminary research findings and not peer reviewed; data are constantly changing and updating.

Citations

  1. Seng EK, Robbins MS, Nicholson RA. Acute migraine medication adherence, migraine disability and patient satisfaction: A naturalistic daily diary study. Cephalalgia. 2017;37(10):955-964. doi:10.1177/0333102416663459
  2. Roberts K, Hamilton J, Bernstam E, Blanco E. A Natural Language Processing Tool for Extracting Medication Adherence Information from Electronic Health Records. Patient-Centered Outcomes Research Institute (PCORI); 2024. doi:10.25302/11.2024.ME.2018C110963