On February 16, 2025, PUMCH-GENESIS, an AI large model for rare diseases jointly developed by PUMCH and the Institute of Automation, Chinese Academy of Sciences (CASIA), officially began clinical use. Built on China's extensive rare disease knowledge base and population genetic data, this model is the world's first rare disease large model tailored to the Chinese population. It will enhance physicians' ability to diagnose rare diseases accurately and efficiently, reducing diagnostic delays and addressing nationwide care disparities. This breakthrough places China at the forefront of rare disease AI technology and represents a significant advancement in the “Chinese approaches” to rare disease management.
While individual rare diseases are uncommon, they collectively affect approximately 400 million patients worldwide. The shortage of rare disease specialists compounds this challenge, with over 70% of cases facing misdiagnosis or delayed diagnosis, underlining an urgent need to strengthen primary care capabilities. Despite AI's promising potential in healthcare, traditional models have struggled with rare diseases due to data scarcity, unreliable knowledge bases, and opaque decision-making. Confronting these clinical challenges head on, PUMCH experts collaborated with CASIA in a two-year initiative, integrating PUMCH's extensive rare disease expertise with big data from the National Rare Disease Collaborative Network and the Quality Control Center to develop PUMCH-GENESIS.
The interdisciplinary team developed innovative features including active perception interaction, differential diagnosis capabilities, and a hybrid "data + knowledge" driven approach. By incorporating DeepSeek-R1's powerful reasoning capabilities, they established an intelligent foundation for rare disease diagnosis and treatment. The model thus offers three core advantages: clinical-oriented decision logic, effective control of AI hallucinations, and autonomous knowledge iteration. Its user-friendly interface enables patients to receive preliminary guidance on diagnosis and treatment through interactive consultations with the chatbot “Xiexiaochu” within minutes.
"We aim to establish PUMCH-GENESIS as a world-class model in terms of technical capabilities while ensuring it remains China's leading platform for AI-assisted disease diagnosis and treatment", stated Professor Zhang Shuyang, President of PUMCH and Director of the Rare Disease Medical Center, sharing the vision she and Professor Xu Bo from CASIA established at the project's inception. With that consensus in mind, the team designed a progressive reasoning chain from "symptoms to examination to differential diagnosis", closely matching physicians' clinical thinking patterns. The model displays key nodes and decision branches throughout the symptom-to-diagnosis pathway. This visualized decision-making process serves dual purposes: providing patients with scientifically-validated diagnostic recommendations while enabling physicians to rapidly comprehend the reasoning pathway.
To combat potential AI hallucinations, PUMCH-GENESIS is equipped with a multi-layered knowledge base with full traceability. Where traditional AI models often generate unreliable inferences due to data noise or knowledge gaps, PUMCH-GENESIS employs a robust system of authoritative data integration, dynamic knowledge updates, and enhanced traceability mechanisms to effectively minimize such errors and increase the reliability of clinical decision-making.
To overcome the inherent challenge of limited and scattered rare disease data - a factor that typically hampers AI model training and development, the team pioneered an innovative approach, minimal-sample cold start. This novel method enables whole-process assisted decision-making by strategically combining limited clinical data with expert medical knowledge. Additionally, the model evaluates the care delivery process using interaction records with patients and enables proactive updates of decision-driving data, creating a closed loop of "clinical application - data feedback - model iteration".
The development of PUMCH-GENESIS received substantial backing from the National Health Commission, Beijing Municipal Health Commission, and the Ministry of Science and Technology's 2030 "New Generation Artificial Intelligence" Major Project initiative. The National Rare Disease Collaborative Network, the National Rare Disease Quality Control Center, and the Beijing Rare Disease Quality Control Center provided solid platform support and data resources for the model's development.
Following its release in February 2024, PUMCH-GENESIS has demonstrated remarkable effectiveness during its year-long pilot phase at PUMCH's multidisciplinary joint clinic for rare diseases. Now the optimized large model is being rolled out for patients first, with plans to integrate it into the online consultation platform of the joint clinic for rare diseases at PUMCH. Looking ahead, the model will be gradually implemented across the entire network of hospitals within the National Rare Disease Collaborative Network.
"PUMCH-GENESIS marks a pivotal breakthrough in China's approach to rare disease diagnosis and treatment", Professor Zhang emphasized, "we view user feedback as essential for its continued refinement". PUMCH remains committed to deepening the integration between AI technology and clinical needs, making new technology an accelerator to overcome both clinical challenges and development bottlenecks. This initiative aims to enhance primary healthcare capabilities and strengthen China's tiered healthcare system, bringing hope to families affected by rare diseases.
Written by Jin Ye, Chen Xiao, and Duan Wenli