China tests AI assistant for neurosurgeons – 03/17/2024 – Science

China is testing an AI assistant for neurosurgeons at seven hospitals in Beijing and other cities in the coming months, one of many initiatives the government is backing to try to harness the technology.

A Hong Kong-based agency under the Chinese Academy of Sciences — the top state-backed scientific institute in China — on Monday introduced an AI model based on Llama 2.0 from Meta Platforms Inc. Researchers trained and fine-tuned the model with articles, medical journals and manuals to act as a surgery consultant for doctors, said Liu Hongbin, executive director of the center.

State bodies are joining private Chinese companies in developing indigenous AI in the mold of OpenAI’s ChatGPT. The Chinese Academy of Sciences’ TaiChu model was one of the first services approved for public launch in August. The technology is considered to have the potential to revolutionize fields from diagnostics to in-person consultations.

The Hong Kong Center for Artificial Intelligence and Robotics used about 100 graphics processing units to train its healthcare-focused model, split evenly between Nvidia Corp.’s cutting-edge A100 chips. and Huawei Technologies Co.’s Ascend 910B, researchers told reporters on Monday.

They expect the AI ​​chatbot — dubbed Cares Copilot 1.0 — to answer questions with quotes based on more than a million academic records. It must also be capable of processing diagnostic data such as MRI, ultrasound or CT scans, as well as images, text and audio, they said.

Eventually, researchers envision AI taking a more active role — warning doctors not to proceed with risky procedures, for example.

“There are obstacles, including restricted computing power due to prohibited access to Nvidia’s advanced chips,” said Feng Ming, chief physician of the neurosurgery department at Peking Union Medical University Hospital, who participated in the model’s development. “However, we can develop a vertical model with our own characteristics with more high-quality data from major hospitals on the mainland, which is not available to OpenAI and many domestic private companies.”