A project in China developing an autonomous and self-evolving virtual healthcare environment will go public next year.
This is confirmed Healthcare IT news by Yang Liu, professor at the Department of Computer Science and Technology of Tsinghua University and co-research lead of the Agent Hospital project. The virtual hospital concept, developed by researchers at the university Institute for AI Industry Research (AIR) simulates the real cycle of the hospital treatment process, from the onset of the disease to the follow-up. The institute claims that the concept is the first of its kind worldwide. The findings from this study were first published in May arXiv, Cornell University’s open-access online repository for research papers.
WHY IT’S IMPORTANT
All virtual actors in Agent Hospital, including patients, nurses and doctors, are generated via a large language model (LLM). These AI characters will represent real people once the system goes live to the public in the first half of 2025. A public pilot, run by AIR’s spin-off startup Tairex, will begin sometime in the first quarter, Prof. Yang said.
For the virtual hospital concept, researchers proposed a design method called MedAgent-Zero, which enables AI doctors to continuously learn and improve and become accurate in performing clinical tasks by interacting with patients, reviewing medical literature and gathering experience in dealing with both successful and unsuccessful cases.
Their research results showed that thanks to this new method, AI doctors achieved an accuracy of 88%, 95.6% and 77.6% respectively in examining, diagnosing and treating patients.
“The doctor agent is able to complete the diagnosis and treatment of tens of thousands of patients in a few days, which would normally take at least two years for a human doctor,” the researchers also noted.
Meanwhile, an AI doctor was also found to be up to 93% accurate in answering a subset of the MedQA dataset – based largely on the competitive United States Medical Licensing Examination, which included questions about major respiratory diseases.
As part of developing the concept, researchers plan to expand the reach of disease coverage and expand it to more medical departments. The virtual platform currently has 42 AI doctors in 21 medical departments, including emergency medicine, respiratory sciences and cardiology.
They also plan to include more features, including medical position promotions, changes in disease distribution over time, and historical patient medical records.
There is also a plan to optimize the selection and implementation of the basic LLM. OpenAI’s ChatGPT model versions 3.5 and 4 are currently used in their research. “We will use the latest and most advanced LLM,” said Prof. Yang.
THE BIG TREND
Other research initiatives in China have also developed medical LLMs to support clinical decisions. A project at Tongji University School of Medicine built a model called MedGowhich was trained using 6,000 medical textbooks and has since been integrated and used in the affiliated Shanghai East Hospital.
An AI-focused institute under the Chinese Academy of Sciences – one of China’s national research centers – introduced the CARES Copilot chatbot based on Meta’s Llama 2 LLM, which helps doctors make medical diagnoses and treatments.