news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Shanghai Technology recently launched

"View·What to Say" Series

Following the release of the Ruijin Medical Big Model last year, Ruijin Hospital released the Ruijin AI Doctor at the 2024 Artificial Intelligence Conference in July this year. Huang Feiyue, Chief Technology Officer of the Shanghai Digital Medicine Innovation Center of Ruijin Hospital, introduced that based on the high-quality medical data of Ruijin Hospital, the team has built a group of medical big models for benefiting the people and assisting doctors.

Huang Feiyue said: "AI has developed very rapidly in the past decade, but based on my observations in the AI ​​and medical industries over the past few years, medical AI is still a field that has not made any particularly breakthrough progress."

In the era of AI explosion,

How can medical AI become part of ordinary people’s lives?

Huang Feiyue summarized the main problems encountered in the development of medical AI into four points: "quantity", "accuracy", "completeness" and "security". That is, traditional AI relies on a large amount of finely labeled data, but in fact medical data is relatively scarce; there are high demands for the accuracy of medical AI, so "accuracy" is still lacking; there is insufficient intelligence and there are also deficiencies at the product level; the security of medical data is taken very seriously, which also limits large-scale research and development.

In Huang Feiyue's view: In the past two years, the new paradigm of AI research has been initialized through pre-training and large models. These research paradigms will have better solutions to problems, which is also a good hope. Large model technologies such as GPT have developed very rapidly over the years and have actually entered the medical field. It can be seen that the mainstream research paradigm of medical large models is to achieve full-scene, zero-sample universality through multi-modal, multi-task, weak supervision, and pre-training methods.

"Last year, an article published in Nature

This research and development idea was also mentioned from a medical perspective.

That is to say, through the medical big model

Towards artificial intelligence in general practice.”

Huang Feiyue said.”

After nearly 20 years of accumulation, Ruijin Hospital has accumulated 320 million CDR medical records, among which EMR data that records detailed treatment processes is most suitable for medical pre-training. Huang Feiyue introduced that, based on statistics from the whole hospital, the total number of high-quality EMR records exceeds 80 million, and the average number of tokens per record is about 3k, so the total amount of pre-training is about 240B. For this reason, Ruijin's base model was developed based on this series of medical data using the pre-training enhancement method.

We have built a group of medical agents through the unified scheduling of the medical big model and the supervision and fine-tuning based on the clinical doctor's instructions of Ruijin Hospital. Different medical agents can improve the actual use effect of the application scenario through collaboration. This series of agent experts continue to collaborate and evolve. The past ten years have been more about medical informatization. I believe that with this series of designs and developments in the future, it can also better help our medical information system to move from informatization to intelligence. "

Huang Feiyue said:

“Ruijin’s application model

From the perspectives of symptom identification and medical advice

Come in,

Characterized by its commitment to promoting

Practical application to real clinical medical tasks.””

The current Ruijin big model can identify hundreds of thousands of abnormal symptoms and recommend the most appropriate follow-up department and accurate medical advice to patients. In the top ten departments of Ruijin, the big model also demonstrated very good generalization performance and excellent technical capabilities, with relevant technical indicators reaching more than 95%.

The views of companies and experts do not represent the official position

Edited by: zy