Using large language models to identify pre-diagnostic clinical features of ovarian cancer from healthcare records: a population-based case-control study.

BACKGROUND: Most women with ovarian cancer are diagnosed after developing symptoms. However, symptoms are often recorded as free text within Electronic Healthcare Records (EHRs) which is not readily accessible for research. AIM: To use EHRs to examine associations between coded and Large Language Model (LLM) extracted free text clinical features with ovarian cancer diagnosis. DESIGN: Population based case-control study using EHR and cancer registry data. SETTING: Women attending primary care, outpatient and emergency clinics associated with the University of Washington. 136 ovarian cancer cases (diagnosed 2012-2019) were matched (age, clinic type) to 1360 controls. METHODS: 12 months of pre-diagnosis coded and free-text data were extracted from EHRs. LLMs were tested on annotated notes, before extracting information on 17 pre-specified clinical features. Univariate conditional logistic regression analyses were used to identify clinical features associated with ovarian cancer. RESULTS: 14 clinical features were more commonly identified from free text using LLMs than from codes in both cases and controls. 14 features were significantly associated with ovarian cancer when using codes than LLM extracted data, but only 8 features were significant using codes alone. Using both coded and LLM extracted data, 11 features had OR's >2. 13 features were significantly associated when restricting analysis to early-stage (I-II) diagnosis. CONCLUSIONS: There is an identifiable ovarian cancer symptom signature within EHRs, with LLM-based NLP approaches enabling extraction of key non-coded symptom information. LLM's could support EHR-based research, while LLM-based Clinical Decision Support Tools may improve identification of patients with symptoms of possible cancer.