Applying GPT-4 to the Plastic Surgery Inservice Training Examination

Abstract

The recent introduction of Generative Pre-trained Transformer (GPT)-4 has demonstrated the potential to be a superior version of ChatGPT-3.5. According to many, GPT-4 is seen as a more reliable and creative version of GPT-3.5. In conjugation with our prior manuscript, we wanted to determine if GPT-4 could be exploited as an instrument for plastic surgery graduate medical education by evaluating its performance on the Plastic Surgery Inservice Training Examination (PSITE). Sample assessment questions from the 2022 PSITE were obtained from the American Council of Academic Plastic Surgeons website and manually inputted into GPT-4. Responses by GPT-4 were qualified using the properties of natural coherence. Incorrect answers were stratified into the consequent categories: informational, logical, or explicit fallacy. From a total of 242 questions, GPT-4 provided correct answers for 187, resulting in a 77.3% accuracy rate. Logical reasoning was utilized in 95.0% of questions, internal information in 98.3%, and external information in 97.5%. Upon separating the questions based on incorrect and correct responses, a statistically significant difference was identified in GPT-4's application of logical reasoning. GPT-4 has shown to be more accurate and reliable for plastic surgery resident education when compared to GPT-3.5. Users should look to utilize the tool to enhance their educational curriculum. Those who adopt the use of such models may be better equipped to deliver high-quality care to their patients.