Benefits, Limits, and Risks of GPT-4 as an AI Chatbot for Medicine

We evaluated the performance of a large language model called ChatGPT on the United States Medical Licensing Exam (USMLE), which consists of three exams: Step 1, Step 2CK, and Step 3. ChatGPT performed at or near the passing threshold for all three exams without any specialized training or reinforcement. Additionally, ChatGPT demonstrated a high level of concordance and insight in its explanations. These results suggest that large language models may have the potential to assist with medical education, and potentially, clinical decision-making.

Machine learning can help clinicians to make individualized patient predictions only if researchers demonstrate models that contribute novel insights, rather than learning the most likely next step in a set of actions a clinician will take. We trained deep learning models using only clinician-initiated, administrative data for 42.9 million admissions using three subsets of data: demographic data only, demographic data and information available at admission, and the previous data plus charges recorded during the first day of admission. Models trained on charges during the first day of admission achieve performance close to published full EMR-based benchmarks for inpatient outcomes: inhospital mortality (0.89 AUC), prolonged length of stay (0.82 AUC), and 30-day readmission rate (0.71 AUC). Similar performance between models trained with only clinician-initiated data and those trained with full EMR data purporting to include information about patient state and physiology should raise concern in the deployment of these models. Furthermore, these models exhibited significant declines in performance when evaluated over only myocardial infarction (MI) patients relative to models trained over MI patients alone, highlighting the importance of physician diagnosis in the prognostic performance of these models. These results provide a benchmark for predictive accuracy trained only on prior clinical actions and indicate that models with similar performance may derive their signal by looking over clinician’s shoulders—using clinical behavior as the expression of preexisting intuition and suspicion to generate a prediction. For models to guide clinicians in individual decisions, performance exceeding these benchmarks is necessary.