Nature Sub-Journal | DeepSeek Deploys in Over 750 Hospitals, Tsinghua Team Analyzes AI Medical Supervision Risks}

DeepSeek, a highly popular tech company recently, has seen its flagship model DeepSeek-R1 widely adopted in physical hospitals. Since its launch in January 2025, it has been used in over 750 hospitals across China, with more than 500 deploying local systems as of May 8, 2025.

Despite this rapid deployment, the use of large language models (LLMs) in hospitals remains in a regulatory gray area, as pointed out by Professor Zhang Yi from Tsinghua University. How to improve supervision frameworks for clinical AI to prevent potential risks to patients and hospitals is a pressing issue.

Professor Zhang and his team published an article titled "Rapid deployment of large language model DeepSeek in Chinese hospitals demands a regulatory response" in Nature Medicine on July 30, 2025.

DeepSeek in Hospitals

DeepSeek's widespread use benefits from its high cost-effectiveness and open-source customization, which are crucial for reducing deployment costs in hospitals. Currently, two mainstream integrated commercial systems of DeepSeek-R1 are available at affordable prices for many hospitals.

Compared to earlier LLMs, DeepSeek-R1 achieves higher reasoning performance through a multi-stage training approach, rivaling OpenAI's models, which is critical for complex medical tasks. Its open-source license allows hospitals to modify and develop further, fitting various budgets and IT conditions across China.

Table 1: Scenarios and needs for deploying DeepSeek solutions in hospitals.

Different clinical scenarios have varying requirements for safety and confidentiality. Even the same service may have differentiated demands based on actual needs. For applications like diagnosis and treatment support, the highest standards of reasoning and interpretability are required. The paper praises DeepSeek for excelling in these areas.

Clinical Service Scenarios

• Pre-diagnosis and triage: Chatbots pre-screen symptoms, recommend non-emergency departments, and provide route and policy inquiries.
• Pre-visit history collection: Patients input history and symptoms before consultation, generating preliminary diagnosis reports with high privacy and reasoning reliability.
• Clinical decision support: Integrates multi-source medical data (including imaging) to generate diagnostic suggestions and treatment plans, handling sensitive records with high privacy and interpretability needs.
• Real-time report interpretation: Converts professional terminology and interprets reports, with high privacy for patient records and low reasoning requirements.
• Standardized EMR generation and correction: Automatically extracts data to generate structured medical records, requiring high privacy and reasoning accuracy.
• Other medical document generation: Auto-generates initial drafts of case summaries/discharge notes, with high privacy and terminology compliance.
• Prescription review: Checks prescription compliance and medication risks, with moderate privacy and reasoning needs.

Hospital Operations Management

• Administrative assistant: Integrates rules and processes into a knowledge base, supports staff queries and document generation, with high privacy and low reasoning demand.
• Medical record quality monitoring: Batch monitoring of record quality, requiring access to core databases with high privacy, with reasoning focused on accuracy.

Research and Education

• Research assistant: Analyzes unstructured medical records for modeling and experimental design, requiring high privacy, reasoning, and interpretability.
• Data cleaning and standardization: Uses authoritative guidelines to clean historical data, with high privacy needs and accurate terminology.
• Medical education robots: Support Q&A and simulation teaching, with privacy and reasoning needs varying with task complexity.

Personal Health Management

• Chronic disease monitoring and report interpretation: Uploads medical records to generate risk assessments and advice, with high privacy needs and reasoning adjusted by service type.

Regulation of LLMs

Despite the benefits of widespread LLM use, potential risks remain. DeepSeek-R1's advanced reasoning could produce misleading outputs, leading to misdiagnosis or improper treatment, especially if doctors overly rely on multimodal models.

In clinical scenarios, errors could delay treatment or cause self-medication issues. Ethical, compliance, and data security risks from operational errors or hacking also need attention.

What Should Proper and Effective Supervision Look Like?

The basic step is classifying LLM applications by scenario and risk level, clarifying supervision responsibilities of NHC and NMPA, and establishing coordination mechanisms.

Next, define high-risk application thresholds, distinguishing simple guidance from intelligent medical devices, especially for diagnosis and treatment support.

Build lifecycle management pathways, including evaluation tools like MedBench, real-time validation, post-deployment monitoring, and change control for updates.

Professor Zhang emphasizes that many countries, including China, lack sufficient regulation for hospital LLM applications, especially balancing innovation and patient safety. Establishing a comprehensive supervision framework is urgent and necessary.