GPT-Assisted Enhancement of Differential Diagnostics in Emergency Medicine

Emergency Department (ED) settings demand rapid and accurate diagnostic assessments to ensure optimal patient outcomes and efficient healthcare delivery. Traditional differential diagnosis relies heavily on clinical acumen, which is susceptible to variability and error. Advances in artificial intelligence, particularly in natural language processing, offer a novel approach to support clinical decision-making processes. This study introduces a methodology utilizing a Generative Pretrained Transformer (GPT) model to analyze electronic health record (EHR) notes within the first 24 hours of ED admission. The model processes unstructured text to generate a ranked list of differential diagnoses. We evaluated the model's first prediction against actual discharge diagnoses to determine its accuracy and potential utility as a diagnostic aid. The GPT model's top prediction matched the discharge diagnosis in body system level with over 76% accuracy, indicating a robust capacity for accurate early diagnosis prediction. Incorporating GPT into the ED may substantially augment diagnostic accuracy, reduce time to diagnosis, and potentially lower healthcare costs through the optimization of resource allocation. By providing an immediate, ranked list of potential diagnoses, the GPT model assists clinicians in quickly formulating a targeted diagnostic strategy. This AIdriven approach could lead to improvements in patient outcomes by accelerating appropriate medical interventions.