破解罕见病诊断难题，全球首个医学循证推理智能体上线

Core Insights - The article discusses the development of DeepRare, the world's first intelligent rare disease diagnostic system, which addresses the global challenges of difficult diagnosis and high misdiagnosis rates in rare diseases [1][2]. Group 1: System Features - DeepRare employs a "central-embodiment" architecture that surpasses traditional medical AI by linking vast medical literature and real clinical data, enabling deep understanding and internalization of medical knowledge [2]. - Unlike traditional AI, DeepRare exhibits "slow thinking" capabilities similar to human doctors, allowing it to ask questions, validate hypotheses, and iteratively refine diagnostic clues [3]. - The system provides a complete evidence chain for each diagnostic conclusion, enhancing trust in AI by making the reasoning process transparent [3]. Group 2: Diagnostic Performance - DeepRare achieved a first-position accuracy rate of 57.18% in phenotype diagnosis, improving by 23.79 percentage points over the previous best international model [4]. - In retrospective human-machine comparison experiments, DeepRare outperformed rare disease specialists with ten years of clinical experience in diagnostic recall rates [4]. - When gene sequencing data is included, DeepRare's comprehensive first-position diagnostic accuracy exceeds 70.6%, significantly better than the current international standard tool, Exomiser, which has a 53.2% accuracy rate [4]. Group 3: Real-World Application - DeepRare has been deployed in Xinhua Hospital and is undergoing internal testing, functioning as a digital quality control tool to ensure thorough patient evaluations [7]. - The system has launched an online rare disease diagnostic platform that has attracted over 1,000 professional users and covers more than 600 medical and research institutions globally [8]. - A "ten-thousand clinical validation plan" has been initiated to conduct real-world clinical validations of rare diseases, aiming to create a global intelligent diagnostic network [8].