Uncoded Clinical Features from Multilingual Electronic Health Records in Catalonia: Development and Validation Study

The study presents a local hybrid system for extracting unstructured clinical information from electronic health records in Catalonia. The system combines a lightweight language model Phi4-mini with regular expression rules to identify six urinary tract infection symptoms (fever, nitrites, leukocytes, lumbar pain, abdominal pain, and haematuria). The research included 15,498 medical records from 2,962 patients. In clinical validation, the system achieved 93.8% accuracy, 83.6% sensitivity, and 96.6% specificity. Patients who progressed to acute pyelonephritis showed a higher symptom burden (71.3% versus 55.8%). The system operates entirely within institutional servers, protecting patient privacy and avoiding cloud service costs.