GPT pilot to write draft discharge letters at the UMC Utrecht

The AI for Health Team at the UMC Utrecht has developed an application that utilizes GPT to generate draft discharge letters for the ICU, NICU and Cardiology department. Once inspected and completed by the treating physicians, these letters are provided to a patient's general practitioner or next physician upon discharge, and typically include a summary of the patient's stay based on their medical records.
 
Large Language Models like GPT hold significant promise for the healthcare sector, for example in reducing the administrative burden experienced by physicians. However, to harness this potential for clinical applications, the quality of the generated clinical notes must be meticulously evaluated. The initial step in this process is to validate the generated text, which involves systematically assessing the quality of the model's output.
 
In collaboration with the Data Science Plan of the Julius Center, the application underwent rigorous validation for use in the neonatal and intensive care unit departments at the UMC Utrecht. The initial validation process revealed that GPT omitted facts at least once in approximately 40% of the draft discharge letters and hallucinated facts at least once in less than 30% of the drafts. The levels of omissions and hallucinations in the draft letters were deemed acceptable to continue to the next evaluation stage by the physicians involved in the study. Usability scores revealed that GPT’s draft letters were comparable in usability to physician’s letters.