Abstract
BACKGROUND AND OBJECTIVES: We propose the use of GPT-4 to facilitate initial history-taking in neurology and other medical specialties. A large language model (LLM) could be utilized as a digital twin which could enhance queryable electronic medical record (EMR) systems and provide healthcare conversational agents (HCAs) to replace waiting-room questionnaires. METHODS: In this observational pilot study, we presented verbatim history of present illness (HPI) narratives from published case reports of headache, stroke, and neurodegenerative diseases. Three standard GPT-4 models were designated Models P: patient digital twin; N: neurologist to query Model P; and S: supervisor to synthesize the N-P dialogue into a derived HPI and formulate the differential diagnosis. Given the random variability of GPT-4 output, each case was presented five separate times to check consistency and reliability. RESULTS: The study achieved an overall HPI content retrieval accuracy of 81%, with accuracies of 84% for headache, 82% for stroke, and 77% for neurodegenerative diseases. Retrieval accuracies for individual HPI components were as follows: 93% for chief complaints, 47% for associated symptoms and review of systems, 76% for relevant symptom details, and 94% for histories of past medical, surgical, allergies, social, and family factors. The ranking of case diagnoses in the differential diagnosis list averaged in the 89th percentile. DISCUSSION: Our tripartite LLM model demonstrated accuracy in extracting essential information from published case reports. Further validation with EMR HPIs, and then with direct patient care will be needed to move toward adaptation of enhanced diagnostic digital twins that incorporate real-time data from health-monitoring devices and self-monitoring assessments.