Neurological history both twinned and queried by generative artificial intelligence.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) offer potential for enhancing medical record systems and patient interactions.
• Current methods like waiting-room questionnaires can be inefficient.
• Digital twins and healthcare conversational agents (HCAs) are emerging concepts.

Purpose of the Study:

• To evaluate the use of GPT-4 for initial medical history-taking.
• To assess the accuracy of an LLM-based system in extracting patient history data.
• To explore the potential of LLMs in generating differential diagnoses.

Main Methods:

• An observational pilot study using published case reports (headache, stroke, neurodegenerative diseases).
• Three GPT-4 models were employed: a patient digital twin (P), a neurologist model (N) querying P, and a supervisor model (S) synthesizing dialogue.
• Each case was analyzed five times to ensure reliability and consistency.

Main Results:

• Overall accuracy for history of present illness (HPI) content retrieval was 81%.
• Specific accuracies included 84% for headache, 82% for stroke, and 77% for neurodegenerative diseases.
• The LLM-generated differential diagnoses ranked in the 89th percentile.

Conclusions:

• The tripartite LLM model demonstrated significant accuracy in extracting critical information from medical case reports.
• Further validation with electronic medical record (EMR) HPIs and direct patient interaction is necessary.
• Future applications may involve diagnostic digital twins integrating real-time health monitoring data.