A new study led by investigators from Mass General Brigham has found that ChatGPT was about 72 percent accurate in overall clinical decision making, from coming up with possible diagnoses to making final diagnoses and care management decisions. The large-language model (LLM) artificial intelligence chatbot performed equally well in both primary care and emergency settings across all medical specialties. The research team’s results are published in the Journal of Medical Internet Research.

“Our paper comprehensively assesses decision support via ChatGPT from the very beginning of working with a patient through the entire care scenario, from differential diagnosis all the way through testing, diagnosis, and management,” said corresponding author Marc Succi, MD, associate chair of innovation and commercialization and strategic innovation leader at Mass General Brigham and executive director of the MESH Incubator. “No real benchmarks exists, but we estimate this performance to be at the level of someone who has just graduated from medical school, such as an intern or resident. This tells us that LLMs in general have the potential to be an augmenting tool for the practice of medicine and support clinical decision making with impressive accuracy.”

Changes in artificial intelligence technology are occurring at a fast pace and transforming many industries, including health care. But the capacity of LLMs to assist in the full scope of clinical care has not yet been studied. In this comprehensive, cross-specialty study of how LLMs could be used in clinical advisement and decision making, Succi and his team tested the hypothesis that ChatGPT would be able to work through an entire clinical encounter with a patient and recommend a diagnostic workup, decide the clinical management course, and ultimately make the final diagnosis.
The study was done by pasting successive portions of 36 standardized, published clinical vignettes into ChatGPT. The tool first was asked to come up with a set of possible, or differential, diagnoses based on the patient’s initial information, which included age, gender, symptoms, and whether the case was an emergency. ChatGPT was then given additional pieces of information and asked to make management decisions as well as give a final diagnosis—simulating the entire process of seeing a real patient. The team compared ChatGPT’s accuracy on differential diagnosis, diagnostic testing, final diagnosis, and management in a structured blinded process, awarding points for correct answers and using linear regressions to assess the relationship between ChatGPT’s performance and the vignette’s demographic information.

Artificial Intelligence Homepage