Promoting Responsible DeepSeek Deployment in Health Care: Scoping Review Comparing Grey and White Literature

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Medical Technology Assessment

Background:

• DeepSeek, an open-source large language model (LLM), has seen rapid adoption in healthcare, raising concerns about patient safety and outcomes.
• Limited information exists on the specific usage and regulatory frameworks governing DeepSeek in clinical settings.

Purpose of the Study:

• To systematically review the deployment characteristics of DeepSeek in China's top 100 hospitals.
• To compare the reported performance and risks of DeepSeek based on hospital disclosures versus existing research evidence.

Main Methods:

• Searched hospital websites/social media and academic databases (Web of Science, PubMed) for DeepSeek deployments and studies post-January 15, 2025.
• Extracted data on DeepSeek characteristics, applications, evaluations, risks, and regulations using a developed coding framework.
• Assessed risk of bias using the Joanna Briggs Institute checklist.

Main Results:

• 58 DeepSeek models were identified in 48 of the top 100 Chinese hospitals; 27 related studies were found.
• DeepSeek-reasoner (R1) was the most common model, primarily used for clinical decision support (diagnosis, treatment).
• Hospital disclosures showed poor transparency regarding predeployment assessments (36%), results (22%), and risk management (9%).

Conclusions:

• DeepSeek deployment in leading Chinese hospitals is rapid and widespread, mainly for clinical decision support.
• Lack of transparency and rigorous validation in hospital disclosures poses potential risks to patient safety.
• Urgent need for expanded regulations and more transparent reporting policies for responsible LLM use in healthcare.