Not long ago, the "Xiehe Taichu" rare patient AI model jointly developed by Peking Union Medical College Hospital and the Institute of Automation of the Chinese Academy of Sciences officially entered the clinical application stage. The research team regards the ability to effectively suppress artificial intelligence "illusions" as one of its three core advantages. Related reports show that traditional artificial intelligence models often produce "illusions" due to data noise or knowledge blind spots, which can easily lead to erroneous inferences. By integrating authoritative data, dynamically updating knowledge, and adding traceability mechanisms, "Concordia · Taichu" effectively suppresses the occurrence of "illusions" in the model and enhances the credibility of clinical decision-making.
To suppress' hallucinations', we mainly use localized high-quality datasets (such as real diagnosis and treatment records) and a large number of manually validated knowledge bases (such as clinical diagnosis and treatment guidelines, textbooks, and publicly available scientific literature) for model fine-tuning. "Gong Mengchun explained," Through techniques such as prompt word engineering, effective evidence support, traceability verification, criterion verification, etc. are applied in the reasoning process, and feedback and correction are continuously provided. "This can be understood as allowing AI to only learn" good knowledge "and limit the" boundary "of reasoning, or adding a" traceability "mechanism to continuously revise" education "before clinical application.
According to statistics released by relevant professional institutions, DeepSeek has been deployed in top tier hospitals in China. Several hospitals, including Huashan Hospital affiliated with Fudan University, Ruijin Hospital, and Shenzhen People's Hospital, have started joint research and development with top domestic artificial intelligence companies. Through cross disciplinary integration and docking, they have accelerated the promotion of medical intelligence and solved the "bottleneck" problem. The industry believes that utilizing DeepSeek's unique "hierarchical knowledge distillation" technology can reduce the inference and computation workload in complex medical scenarios by 30%.
A large amount of research based on multimodal clinical data at home and abroad has confirmed that the clinical insights provided by artificial intelligence technology may exceed the knowledge and physical sensory limits that doctors can grasp, providing important auxiliary opinions for clinical decision-making. Gong Mengchun said that when it comes to new evidence and judgments provided by AI, we should adopt a "cautious verification and full judgment" attitude, build a multimodal clinical data infrastructure, sort out high-quality queues, and verify various data-driven medical discoveries. Integrating clinical evidence provided by AI into the diagnosis and treatment process in the form of multidisciplinary diagnosis and treatment (MDT) to serve patients is one of the mainstream development trends in the global medical research field.
