Tencent Health announced a large-scale medical model and released a new multi-scenario AI product matrix such as intelligent question and answer, family doctor assistant, and digital medical imaging platform.
According to the introduction, the medical large model is based on the Tencent Hunyuan large model base, adding more than 2.85 million medical entities, 12.5 million medical relationships, and more than 98% of medical knowledge knowledge graphs and medical literature. After 30 million including patients, doctors, medicines, etc. Multi-task fine-tuning of Q&A dialogues in scenes such as factories and medical procedures, as well as reinforcement learning of data labeled by 360,000 expert doctors.
At present, Tencent's large medical model includes scenario large models such as copywriting generation, intelligent question and answer, medical record structuring and retrieval, imaging reports and auxiliary diagnosis. It can be embedded in the entire medical process, including department guidance, doctor recommendation, pre-consultation, doctor-patient consultation, etc. Achieve comprehensive improvements in the level and quality of medical services in application scenarios such as dialogue, automatic generation of medical records, and intelligent hospital customer service.
At present, Shanghai Renji Hospital has taken the lead in using Tencent's medical model to create a highly anthropomorphic virtual digital image "Nurse Xiaowei" focusing on aging-friendly services in Internet hospitals.
In terms of AI-driven drug discovery, Tencent’s “iDrug” platform has the ability to accelerate the discovery of both small molecule drugs and large molecule drugs. Liu Wei, technical director of Tencent AIDD, introduced that in terms of protein structure prediction, the "iDrug" platform has developed a new algorithm framework tFold ; in terms of predicting the ADMET properties of drugs, 70+ ADMET properties have been developed and launched; development Two framework transition molecule generation algorithms were developed to discover nM-level lead compounds, which were effectively proven in 3-4 projects. Reinforcement learning technology was also introduced to achieve 97% of generated molecules meeting the requirements in the generation of small drug molecules.