Open-source large language models are quietly reshaping how clinicians, educators, and researchers interact with medical information — and the field is moving faster than most oversight frameworks can track. Understanding which AI tools are gaining traction, in which clinical domains, and with what demonstrated limitations has become a prerequisite for responsible adoption. This first systematic bibliometric and scoping review of DeepSeek's medical applications provides that map.
Drawing on 371 publications indexed across PubMed, Web of Science, and Scopus from January through November 2025, the analysis found a sharply rising publication curve concentrated heavily in China, which accounted for 163 of the identified papers. The scoping component synthesized 353 original articles, cataloging DeepSeek's primary application domains — spanning clinical decision support, medical education, administrative workflow, and research assistance — alongside recurring strengths such as customizability and lower computational cost compared with proprietary models, and consistent limitations including hallucination risk, regulatory uncertainty, and variable performance across medical specialties.
This review arrives at a critical inflection point. Proprietary models like GPT-4 and Claude have dominated English-language medical AI discourse, but open-source alternatives such as DeepSeek introduce a different risk-benefit calculus: greater institutional control over fine-tuning and data handling, but less third-party safety validation. The geographic concentration of the literature in China is itself a significant finding, suggesting that Western clinical contexts remain underrepresented, which limits generalizability of performance claims. The review methodology — combining bibliometrics with a PRISMA-ScR scoping framework — is appropriate for a rapidly evolving field where randomized evidence is scarce, though it cannot establish causal efficacy for any specific clinical outcome. For health systems evaluating LLM adoption, this synthesis offers a rare structured baseline, though the eleven-month publication window means the landscape it captures is already evolving. Incremental but strategically important.