Max Melamed
Research Mentor(s): Brian Athey
Mentor Department: Computational Medicine & Bioinformatics
Authors: Max Melamed, Pranjal Srivastava, Ivan Rozhkov, Greg Farnum, Brian Athey
Session: Session 1 (9:00am – 9:50am)
Presentation Type: Poster 99
Abstract
Due to the fact that doctors have to see and treat countless patients a day, they don’t have as much time as they would like to write detailed public health records for every patient. Additionally, they only have a limited amount of time between each patient’s visit, meaning that even if there was a lengthy, detailed transcript for every patient, they wouldn’t have the necessary time to intently read the whole thing. This study serves to create artificial public health records using generated medical histories of fictitious people with realistic symptoms and prescribed medications. Using this data, we created and tested an abundance of prompts that would allow for ChatGPT to provide an accurate output that reflected an easier-to-read public health record from the perspective of a medical professional. We found that while many prompts generate similar outputs, there were certain prompts that resulted in near-perfect reports, in the sense that no information was left out as well as the report itself made sense and could be understood by anyone who may need to utilize said information. We are working to create a system in which the sole prompt that is selected is used alongside hundreds, if not thousands of new artificial public health records to further test for consistency and accuracy. The end product will serve to assist those greatly in clinical roles, more specifically scribes, medical assistants, and other similar roles in medical fields, by reducing time spent on recording patient notes themselves.
