What is PUMCH·Taichu
PUMCH·Taichu is China's first AI large model in the rare disease field, jointly developed by Peking Union Medical College Hospital and the Institute of Automation of the Chinese Academy of Sciences, and has officially entered clinical application. The model is based on years of accumulation in China's rare disease knowledge base and genetic testing data from the Chinese population, and is the world's first rare disease large model that conforms to the characteristics of the Chinese population. It adopts extremely small sample cold-start technology, requiring only a small amount of data fused with medical knowledge to enable full-process assisted decision-making.
Main functions of PUMCH·Taichu
- Initial consultation and recommendations: Through multi-turn interactive consultation with the model, patients can obtain preliminary diagnosis and treatment recommendations in a short time.
- Assisting doctors in decision-making: The model builds a progressive reasoning chain of "symptoms-examination-differential diagnosis," which is highly consistent with doctors' clinical thinking and can help doctors quickly grasp diagnosis and treatment approaches.
- Medical record writing and genetic interpretation: In the future, it will support doctor-side service functions such as medical record writing, genetic interpretation, and genetic counseling.
- Autonomous knowledge iteration: By evaluating the diagnosis and treatment process through interaction records with patients, it enables decision-driven proactive updating and evolution of data, forming a closed loop of "clinical use-data feedback-model iteration."
- Suppressing AI hallucinations: It builds a multi-dimensional traceable knowledge base, effectively suppressing possible "hallucinations" of the model and enhancing the credibility of clinical decision-making.
Technical principles of PUMCH·Taichu
- Extremely small sample cold-start technology: To address the problems of scattered rare disease cases and scarce data, the research team adopted an extremely small sample cold-start approach, requiring only a small amount of data fused with medical knowledge to enable full-process assisted decision-making functions.
- "Data + knowledge" hybrid drive: The model combines years of accumulation in the rare disease knowledge base and genetic testing data from the Chinese population. Through a hybrid drive of "data + knowledge," it enhances the model's decision-making logic and credibility.
- Deep reasoning capability: It introduces the deep reasoning capability of DeepSeek-R1 and builds a progressive reasoning chain of "symptoms-examination-differential diagnosis," which is highly consistent with doctors' clinical thinking and can display the key nodes and branch logic from symptoms to diagnosis.
- Proactive perception interaction and closed-loop iteration: Through multi-turn interaction with patients, the model proactively perceives changes in the condition and updates the decision-making process, forming a closed loop of "clinical use-data feedback-model iteration" and enabling autonomous knowledge iteration.
- Multi-dimensional traceable knowledge base: To suppress AI "hallucinations," the model builds a multi-dimensional traceable knowledge base, integrates authoritative data, dynamically updates knowledge, and enhances the credibility of clinical decision-making.
Project address of PUMCH·Taichu
- Project official website：PUMCH·Taichu
Application scenarios of PUMCH·Taichu
- Patient initial consultation and rapid consultation: Patients can conduct multi-turn interactive consultations with PUMCH·Taichu, obtain preliminary diagnosis and treatment recommendations in a short time, and quickly identify possible rare disease directions.
- Promotion of rare disease diagnosis and treatment collaboration network: The model has been piloted for one year at the Rare Disease Multidisciplinary Clinic of Peking Union Medical College Hospital, with good results. Subsequently, it will be integrated into the online diagnosis and treatment services of Peking Union Medical College Hospital, and gradually promoted to hospitals in the National Rare Disease Collaboration Network, helping to build the hierarchical diagnosis and treatment system.
- Autonomous Knowledge Iteration and Closed-Loop Optimization: Through interaction records with patients, the model evaluates the diagnosis and treatment process, enabling decision-driven proactive updating and evolution of data, forming a closed loop of "clinical use—data feedback—model iteration" to continuously optimize diagnosis and treatment capabilities.