From July 4th to July 7th, the 2024 World Artificial Intelligence Conference (WAIC 2024) was successfully held in Shanghai. As the most influential industry event in the field of artificial intelligence, WAIC attracted more than 500 companies to participate in the exhibition. The large-scale model-assisted diagnosis and treatment product "AI Doctor Assistant All-in-One Machine" developed by Baidu Health also made its debut here. Based on the large model, the solution opens up the last mile from data to application, and at the same time adopts lower computing power cost while ensuring application performance, which greatly improves the application accessibility of large models.
"AI Doctor Assistant" booth
How to implement the medical model?
Use the "all-in-one machine" to realize the "box" ready to use
At WAIC in 2024, the large model is undoubtedly the focus of the most attention, and thousands of new achievements of the large model will be unveiled here, from the application of large models in medical, financial, agricultural and other industries, to various tools for their landing to "clear the blockage", involving various innovative achievements in the industry. Among them, the landing application of medical models is one of the focuses of this year's exhibition.
However, unlike other fields, the application of large medical models often faces more complex challenges. First of all, medical data involves patient privacy and hospital information security, and data security must be strictly guaranteed, which requires privatization and deployment. Secondly, there are many internal systems in the hospital (such as HIS, LIS, PACS, etc.), and the data standards and formats of each system are different, and the data is very messy, which limits the application. Moreover, the application scenarios of hospitals are very complex, and everyone has different requirements for information systems, which must not only be real-time and accurate, but also convenient and humane, which makes it very difficult for large-scale model products to land; In addition, large models have high requirements for computing power, and hospitals need to pay the additional cost of AI accelerators.
Medical large model all-in-one machine
In order to solve the above problems, Baidu Health has cooperated to develop a CPU version of the large model all-in-one machine based on Intel Xeon Scalable processors and xFastTransformer large model inference acceleration engine, which can take into account the performance required for large model inference, and can also meet various complex and stringent standards for hospital applications: it can not only realize privatized deployment, but also meet real-time or near-real-time standards; In addition, the deployment cost is low, and the hospital does not need to configure additional related hardware, and it also avoids the debugging cost of the model and hardware, and realizes out-of-the-box use. The CPU version has a lower cost and no reduction in performance than the GPU, and the core medical information extraction and medical content generation capabilities are not inferior to the GPU version, or even slightly improved, providing users with more resource choices.
Gang tertiary hospital
Unleash hospital productivity with large models
"Case Essentials" - help the precipitation of classic/difficult/rare/cutting-edge cases
At present, the CPU large model all-in-one machine has been deployed in a top tertiary hospital in Shanghai. The essence of the large-scale model generation of liver cancer cases cooperated by the two parties was highlighted at this exhibition.
Case Essentials screen
The essence of the case is mainly based on the real medical records of the hospital, and the diagnosis and treatment process of patients is summarized and refined through a large model, and sorted out according to the timeline, including patient profile, examination results, diagnosis, diagnosis and treatment, prognosis and treatment subsections. It is an important tool for communicating a patient's condition between members of the care team, ensuring continuity and consistency of information, and is especially important for multidisciplinary teamwork.
In the past, the essence of the case required doctors to sort it out manually, which was time-consuming and labor-intensive, and it was difficult to ensure the quality. Now, doctors can quickly comb through the large model. In addition to fully unleashing physician productivity; It also helps the hospital activate the massive stock data, which is convenient for the precipitation of the hospital's digital assets.
"AI Health Assistant" booth "CDSS-LLM" - large model combined with CDSS integrated into the diagnosis and treatment process
The CPU large model all-in-one machine also provides relevant large model capabilities with standard API interfaces, which is convenient for the integration of third-party information manufacturers, and can also be connected to the CDSS (clinical assisted decision-making) products in the hospital to seamlessly upgrade the large model version, so as to activate the use of CDSS and help doctors improve efficiency services in work scenarios.
Taking the application of case essence as an example, it can be automatically applied through large models in the clinical diagnosis and treatment process, such as automatic recommendation of similar medical records, intelligent question-and-answer extraction of required knowledge, helping experts inherit clinical experience, and further helping clinical teaching. Let the resources of "digital experts" sink to promote medical inclusion.
AI doctor assistant empowers new medical productivity
At present, in addition to the essence of cases, the functions of the "AI Doctor Assistant" also involve the auxiliary generation of medical records, the preview of patient briefings, and the intelligent medical Q&A.
These applications are all derived from Baidu's health care model, based on the domain basic model trained by 100 billion Token high-quality medical and health corpus, and supported by the whole life cycle model development tool chain; Using the industry's original evidence-based AI technology combined with innovative knowledge and data RAG engine support, it minimizes model hallucination; The self-developed optimization of medical knowledge graph and vectorized retrieval technology can effectively improve the model recall rate and reasoning performance, provide doctors with auxiliary support in medicine, teaching, research and management, and provide doctors with a better interactive experience in the diagnosis and treatment process.
Baidu Health has long been committed to using AI + Internet to promote the development of China's medical industry. At present, the overall structure of China's medical resources needs to be optimized, and AI is an excellent means of "allocation", and innovative products such as "AI doctor assistants" are using large model technology to continuously meet clinical needs, promote medical inclusion, and give full play to the value of new quality productivity.
In the future, Baidu Health will further deepen cooperation with Intel to realize the low-cost deployment of "AI doctor assistant" in the form of CPU all-in-one machine, provide services for more hospitals and enterprises, create new medical productivity, and jointly promote the development of China's medical and health industry.
Source: Yigu Network