Empathy and clarity in GPT-4-Generated Emergency Department Discharge Letters

Background and AimThe potential of large language models (LLMs) like GPT-4 to generate clear and empathetic medical documentation is becoming increasingly relevant. This study evaluates these constructs in discharge letters generated by GPT-4 compared to those written by emergency department (ED) physicians. MethodsIn this retrospective, blinded study, 72 discharge letters written by ED physicians were compared to GPT-4-generated versions, which were based on the physicians follow-up notes in the electronic medical record (EMR). Seventeen evaluators, 7 physicians, 5 nurses, and 5 patients, were asked to select their preferred letter (human or LLM) for each patient and rate empathy, clarity, and overall quality using a 5-point Likert scale (1 = Poor, 5 = Excellent). A secondary analysis by 3 ED attending physicians assessed the medical accuracy of both sets of letters. ResultsAcross the 72 comparisons, evaluators preferred GPT-4-generated letters in 1,009 out of 1,206 evaluations (83.7%). GPT-4 letters were rated significantly higher for empathy, clarity, and overall quality (p < 0.001). Additionally, GPT-4-generated letters demonstrated superior medical accuracy, with a median score of 5.0 compared to 4.0 for physician-written letters (p = 0.025). ConclusionGPT-4 shows strong potential in generating ED discharge letters that are empathetic and clear, preferable by healthcare professionals and patients, offering a promising tool to reduce the workload of ED physicians. However, further research is necessary to explore patient perceptions and best practices for leveraging the advantages of AI together with physicians in clinical practice.