Presentation (CTI)

OBJECTIVES: Psoriasis is a chronic, immune-mediated disease associated with substantial physical, functional, and psychosocial burden. Accurate assessment of affected body surface area (BSA), specific anatomic locations, and involvement of nails and joints is essential for disease severity evaluation, treatment decisions, and real-world outcomes research. However, these clinically meaningful measures are frequently documented only in unstructured free-text clinical notes and inconsistently captured in structured electronic health record (EHR) fields. Scalable natural language processing (NLP) approaches may help bridge this gap.
METHODS: We evaluated the performance of Briya’s AI research environment (AIRE), powered by an internally-trained, large language model (LLM), for extracting psoriasis disease distribution data from unstructured EHR text. The data source included outpatient dermatology visit records from a large, state-mandated healthcare provider in Israel, where a random sample of 100 complete visit notes from patients with a psoriasis diagnosis was selected. Free-text clinical documentation was processed by AIRE using a standardized prompt to identify affected skin regions, nail and joint involvement, and BSA impacted by psoriasis. Each record was reviewed and manually annotated by a senior dermatologist. Model performance was compared with the human-expert reference using agreement metrics.
RESULTS: The cohort included 100 patients diagnosed with psoriasis in 2025 with heterogeneous documentation styles. A total of over 40 body areas impacted by psoriasis were documented, with the majority of patients having a BSA<3%. LLM demonstrated strong agreement with expert annotation across affected anatomical sites, including consistent identification of nail and joint involvement, and robust performance despite variable note structure and terminology.
CONCLUSIONS: A LLM embedded within AIRE can reliably extract clinically relevant psoriasis disease distribution features from unstructured EHR notes. Findings demonstrate scalable natural language understanding (NLU) of unstructured clinical text, with the potential extension to natural language generation (NLG), enabling automated quantification of psoriasis disease features for research, outcomes evaluation, and treatment optimization.