Patients don’t describe symptoms the way medical ontologies do.
Someone with Sjögren’s disease doesn’t report “xerostomia with intermittent exacerbation of sicca-related symptoms.” They write that they’re flaring. Someone with narcolepsy doesn’t record “excessive daytime sleepiness.” They describe the day they lost. When that vocabulary doesn’t translate into the standardized medical terminologies that structure healthcare data — SNOMED, RxNorm, ICD — a significant amount of lived experience stays outside the evidence base.
What traditional lexicon work captures — and what it leaves behind.
Qualitative lexicon studies are often conducted with 20 to 60 patients recruited through surveys, structured interviews, or focus groups, and analyzed against predefined category frameworks. Those sample sizes are well-matched to identifying common themes — methodological reviews place theme saturation around 9 interviews and meaning saturation around 24¹. Lexicon work asks a different question — what vocabulary does this community actually use? When studies are designed around predefined frameworks, vocabulary that doesn’t fit — even when it appears frequently in the source material — can fall outside the analysis.
Scaling the signal without losing the voice.
TREND’s natural language processing pipeline was built for this problem. It extracts clinically relevant terms as spans from community-generated text, categorizes them in context, and links them to controlled vocabularies like SNOMED and RxNorm through a lexical knowledge base we maintain internally. The practical effect: patient language can be analyzed at scale without flattening the variation in how patients actually talk.
In our recently published Sjögren’s work in Rheumato, led by Dr. Chiara Baldini of the University of Pisa and co-authored with Sjögren Europe and Sjögren’s UK, that meant analyzing 59,266 documents from 5,025 authors across more than a decade of community conversation.² Flare — a term with no consensus clinical definition in Sjögren’s — was discussed by roughly one in five authors and ranked among the top five symptoms in the community’s own vocabulary.²
Scale matters because lexicon work depends on seeing the full breadth of how a community talks about a condition, including variants that smaller studies aren’t designed to surface. And because community conversation already exists, we can deliver on timelines traditional methods often can’t match. Community vocabulary doesn’t have to be collected. It just has to be heard.
What’s next.
We’re expanding our lexicon capabilities across conditions — surfacing community-specific vocabulary, tracking how terms are used differently across communities, and mapping lexical variants that don’t yet live in any controlled terminology. The language patients actually use deserves to be measured with the same rigor as the endpoints that already have codes.
References
¹ Wutich, A., Beresford, M., Bernard, H. R. Sample Sizes for 10 Types of Qualitative Data Analysis: An Integrative Review, Empirical Guidance, and Next Steps. Field Methods, 2024. journals.sagepub.com/doi/10.1177/16094069241296206
² Baldini, C., Flurie, M., Cline, Z., Flowers, C., Bouillot, C. P., Stone, L. J., Dougherty, L., DeFelice, C., Picone, M. Using Social Media Listening to Characterize the Flare Lexicon in Patients with Sjögren’s Disease. Rheumato, 2025. doi.org/10.3390/rheumato5040014