A team of researchers from Tsinghua University in Beijing has developed the world's first “AI hospital” in which all doctors , nurses and patients are managed by intelligent agents based on a large language model.
The purpose of the project – improve medical care by training physician agents in a simulated environment. The team claims that this will allow them to develop independently and pump up their abilities. AI doctors can diagnose and prescribe treatment for 10,000 patients in just a few days – a task that would take at least two years to solve in an ordinary clinic. At the same time, the platform can train real doctors, endlessly generating virtual patients with a wide variety of diagnoses and symptoms, adjusting their diagnostic and treatment skills.
In the simulated Agent Hospital environment, there are two main entities: patients and healthcare professionals. GPT-3.5 generates roles for these objects, providing infinite extensibility. The operation of the hospital, which includes several consultation and examination rooms, requires a team of medical intelligent agents. The researchers created a configuration consisting of 14 doctor agents tasked with diagnosing diseases and creating detailed treatment plans, and four nurse agents who perform related tasks.
As a result, this team of agents can simulate all processes of medical clinics: from consultation and examination to diagnosis, treatment and follow-up. According to the developers, AI doctors can cure 10,000 patients in just a few days – a task that would take at least two years for a traditional clinic with live doctors. “Hospital” with AI can model and predict various medical situations, in particular the occurrence, spread and containment of infectious diseases in the conditional attached territory, as well as the containment of infectious diseases in the conditional attached territory.
As a result, the improved agent doctors achieved in Agent Hospital accuracy of 93.06% on the MedQA (US Medical Licensing Examination) dataset covering major respiratory diseases. Even with very good doctors, of which there are few, the accuracy is rarely 90%.
At the same time, Agent Hospital can not only treat, but also educate. The platform offers medical students opportunities for professional development. By simulating a variety of patients with AI, future doctors can develop treatment plans without the risk of harming real patients due to errors in decision-making.
Leader of the research team, Liu Yang, said that after six months of intensive AI research and development – the hospital is close to practical deployment – AI doctors will have access to care for real patients. This should happen very soon, in the second half of 2024.
Despite optimistic forecasts, implementation and promotion of diagnosis and treatment based on artificial intelligence may face obstacles. The researchers emphasized the importance of strict adherence to national medical regulations, careful testing of technological maturity, and the study of effective cooperation mechanisms between AI and humans. These measures are necessary to mitigate the potential risks to public health created by AI technologies.