From Patient Language to Terminological Biomarker: Leveraging the Human Phenotype Ontology to Characterize Long COVID in Primary Care

Background: Long COVID presents with complex and multisystemic symptoms
that are difficult to recognize and document using traditional diagnostic classifica-
tions in primary care.
Objective: To explore how HPO can be used to index and analyze patient narra-
tives in general practice and to propose the concept of a "terminological biomarker"
to describe the syndrome.
Design and Setting: A four-year observational study (2021–2025) conducted in
a Belgian general practice, combining narrative interviews, ontology mapping, and
a large language models (ChatGPT).
Method: Patient narratives were transcribed and indexed using ChatGPT-assisted
prompts. HPO terms were extracted and validated using semantic similarity meth-
ods, and combined with clinical metadata and functional outcome scores.
Results: In a cohort of 307 patients, 1320 distinct HPO terms were identified. Fa-
tigue, memory impairment, and exertional intolerance were most frequent. Manual
verification confirmed the reliability of the LLM-HPO matching. A subset of 50
patients showed transcriptomic evidence of viral persistence.
Conclusion: HPO enables structured representation of complex symptoms in Long
COVID and supports narrative-informed documentation. The proposed "termino-
logical biomarker" bridges lived experience and clinical semantics, providing a re-
producible signal for emerging syndromes.