This study used an integrated approach that combined bibliometric analysis and a scoping review to provide complementary insights. The bibliometric method examined the current application of DeepSeek in medicine from multiple dimensions, analyzed researcher characteristics and journal distributions, and identified research hot spots and trends. The bibliometric analysis was conducted based on the framework proposed by Cobo et al [17], following the guidelines for reporting bibliometric reviews of biomedical literature (BIBLIO) [18]. This scoping review systematically extracted and synthesized the applications, challenges, and future research directions of DeepSeek in medicine. The study was conducted according to the framework by Arksey and O’Malley [19] and reported following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines (Checklist 1) [20].

To ensure a comprehensive retrieval of the literature on the applications of DeepSeek in medicine, a systematic search was conducted on December 16, 2025, in PubMed, Web of Science Core Collection (WoSCC), and Scopus. The search strategy (Multimedia Appendix 1) used both controlled vocabularies (MeSH, Web of Science Categories, and SUBJAREA) and free-text terms tailored to each database to optimize retrieval.

Given that the public release of DeepSeek’s reasoning model, DeepSeek-R1, on January 20, 2025 [21,22], marked the beginning of subsequent research into its applications, including in medicine, the search encompassed the period from January 20, 2025, to November 30, 2025.

To ensure comprehensive retrieval, the inclusion criteria were as follows: (1) studies investigating the application of DeepSeek in medicine, (2) document types limited to original articles and reviews for bibliometric analysis and original articles only for scoping review, (3) studies published in peer-reviewed academic journals, and (4) no language restrictions.

The exclusion criteria were as follows: (1) duplicate publications; (2) literature that proposed only speculative or hypothetical uses without substantive analysis or findings; (3) non-peer-reviewed journal items, including books, editorials, preprints, commentaries, conference abstracts, case reports, and retracted articles; and (4) studies with insufficient information for bibliometric analysis or whose full text was unavailable for in-depth content extraction during the scoping review.

After receiving professional training, two authors (HZ and DW) independently screened the titles and abstracts and excluded irrelevant studies based on the aforementioned criteria. The interrater agreement was almost perfect (Cohen κ=0.93). Any disagreements during screening were resolved through discussion or, when necessary, arbitration by a third reviewer (GW).

The final bibliometric analysis included 371 papers. Full records of the selected publications were exported and stored in Excel 2021 (Microsoft Corp) and EndNote Desktop (Clarivate). Bibliographic metadata such as authors’ names, affiliations, countries/regions, and keywords were standardized in a uniform format.

Excel 2021 was used to generate tables highlighting the top 10 authors, institutions, and countries/regions based on their publication output, whereas VOSviewer (version 1.6.19) was used for data visualization of bibliometric mapping, including keyword co-occurrence analysis. Keyword co-occurrence analysis examined the fundamental characteristics of keywords, such as their frequency and temporal evolution. This method helped identify research hot spots and track developmental trends within specialized fields. The three common types of visualizations used in the keyword co-occurrence analysis were the network, density, and overlay maps. In the network map, nodes represented keywords, and the connecting lines represented keyword co-occurrence relationships. The size of a node indicates its frequency, the thickness of a line represents the strength of co-occurrence, and the nodes are clustered together by color to reveal distinct research themes or subfields. The overlay map chronologically visualized the keyword trajectories by assigning chromatic codes corresponding to the computationally derived average publication years (APYs). The density map emphasizes the “research density” or concentration of keywords in the knowledge landscape. Areas with numerous closely located keywords appear as warm-colored regions, such as purple, indicating core well-developed research fronts. Cooler-colored areas such as blue or white represent sparser, potentially peripheral, or emerging topics. The centrality of keywords, which reflects their capacity to bridge different parts of the research network, was derived using CiteSpace (version 7.0.0).

This scoping review included a total of 353 publications. A data extraction form was created using Excel to extract in-depth content from the papers. This form included items such as paper title, research objectives, key findings, research design types, DeepSeek’s strengths, limitations and challenges, future recommendations, DeepSeek model version, quality tier, and application areas. It should be noted that, although quality assessment is not obligatory for scoping reviews, the methodological quality of all included studies was categorized into 3 tiers (high, moderate, and low) based on the criteria (Multimedia Appendix 2) in order to characterize the strength of the available evidence. Data extraction was conducted independently by 2 authors (HZ and DW). Both authors independently extracted data from all 353 included articles in duplicate using the data extraction form created in Excel. After independent extraction, the 2 authors compared their results. Disagreements were resolved through discussion or by consulting a third author (GW) when consensus could not be reached. The extracted data (Multimedia Appendix 3) were then critically analyzed and organized thematically to address the research question, thereby mapping the key application areas of DeepSeek in medicine. The discussion section elaborates on the challenges, research gaps, and future work for the application of DeepSeek in the medical field.

Since this study was a bibliometric and scoping review of previously published literature, ethical approval from an ethics committee is not required.

A systematic search of PubMed, Scopus, and WoSCC yielded 371 publications on the application of DeepSeek in medicine for bibliometric analysis (Figure 1). Among these, the majority (363/371, 97.8%) were categorized as original articles, while the remaining (8/371, 2.2%) were reviews. In terms of publication languages, 358 papers were written in English, and 13 were written in Chinese.

The monthly publication output increased progressively over time. From January to November 2025, the number of papers rose from 0 to 70, with the highest output (70 papers) observed in November (Figure 2).

Of the 216 journals that published papers on the applications of DeepSeek in medicine, 12 published more than 5 papers each. The 10 most active journals collectively contributed 90 publications, accounting for 24.3% (90/371) of the total output. Cureus was the most productive journal with 19 publications, followed by Scientific Reports (n=10), BMC Oral Health (n=9), International Journal of Medical Informatics (n=9), BMC Medical Education (n=8), Frontiers in Artificial Intelligence (n=7), Frontiers in Public Health (n=7), JMIR Medical Informatics (n=7), Journal of Medical Internet Research (n=7), and Journal of Medical Systems (n=7).

Table 1 presents the top 10 authors, institutions, and countries/regions ranked by their respective number of publications on the applications of DeepSeek in medicine.

Table 2 lists the 10 most-cited publications on the medical applications of DeepSeek: 9 were original articles, while 1 was a review [12,23-31].

A keyword co-occurrence analysis was performed to map predominant research hot spots. Synonyms were consolidated prior to analysis; specifically, “large language model(s)” was standardized as “large language model,” and “generative artificial intelligence/AI” was standardized as “generative artificial intelligence.” The top 10 keywords by frequency are listed in Table 3. Notably, “generative artificial intelligence” ranked seventh in frequency but third in centrality. From an initial set of 968 keywords, 41 occurring more than 4 times were included in the keyword co-occurrence analysis. These formed 7 well-defined clusters, visualized in the network map (Figure 3A).

The temporal overlay map (Figure 3B) illustrates the evolution of research focus, with keywords colored by their APYs. Purple nodes represent earlier themes, while crimson indicates more recent activity. Early research concentrated primarily on medical education. The keywords “retrieval-augmented generation” and “oncology” showed the highest APY, reflecting a rising interest in these areas.

The density map (Figure 3C) displays keywords according to their average frequency of occurrence. Crimson regions correspond to the most frequently occurring keywords, followed by blue and then white areas, in descending order.

Medical research and data analysis were mentioned in 73 articles. Among these, 41 had a cross-sectional design, 6 were descriptive studies, 2 were perspective studies, 9 were proof-of-concept studies, 9 were retrospective studies, 1 had a mixed design, and the remaining 5 used other design types.

DeepSeek models have demonstrated significant utility in accelerating and refining medical research and data analysis workflows. DeepSeek facilitates the reading of medical literature, information extraction, and screening. Several studies have developed AI-powered screening tools using DeepSeek to identify relevant studies for systematic reviews, reporting high accuracy and a significant reduction in manual workload [51-53]. For example, the LitAutoScreener tool, which integrates DeepSeek, achieved high accuracy and significantly improved screening efficiency, reducing the processing time to seconds per article [51]. Similarly, other evaluations have confirmed that DeepSeek-based tools can reduce manual workload while maintaining high recall rates in literature screening for meta-analyses [53]. In fields such as aging research, DeepSeek-R1 is part of a multi-LLM ensemble that successfully extracts protocol details from clinical trial records, doubling the yield of conventional search methods and achieving expert-level accuracy for core data points [54]. Second, DeepSeek assists with generating and refining research topics and study designs. It helps researchers analyze cutting-edge trends, funding guidelines, and successful grant applications, thereby validating the novelty of the proposed research questions [3]. For instance, DeepSeek-R1 has been used to explore novel research ideas and generate systematic review topics in fields such as oral and maxillofacial surgery [55]. Similarly, in biomedical research, DeepSeek models show promise in extracting structured pre-analytical variability data from the scientific literature, facilitating standardized reporting and systematic evaluation [56]. Furthermore, DeepSeek serves as a valuable tool for peer review and for critiquing research proposals. Its capacity to generate high-quality evidence-based responses enables a preliminary assessment of a proposal’s feasibility and soundness. This function is particularly beneficial in multidisciplinary contexts where the model’s ability to synthesize information from diverse sources significantly enhances the evaluation process [57,58]. Third, DeepSeek demonstrated substantial potential as an assistant for drafting, editing, and refining the content of medical research papers. Its capabilities span various domains of medical research and practice, making it a versatile tool for enhancing the quality and efficiency of academic writing. The model’s proficiency at generating structured, clear, and comprehensible content is particularly valuable in medical research, where precision and clarity are paramount [59].

In other application domains, 25 articles were identified, comprising 18 cross-sectional studies, 2 perspective articles, 2 descriptive studies, and 3 proof-of-concept studies.

Beyond the primary domains discussed, DeepSeek has been explored in several niche but critical areas, including treatment outcome prediction, drug development assistance, and suicide risk prediction. Instead of reactive question-answering, DeepSeek is integrated into predictive analytics platforms. It can proactively flag at-risk patients, suggest personalized screening intervals, and predict individual responses to therapies based on electronic health records and real-time data [60,61]. In nasopharyngeal carcinoma, DeepSeek-V3-0324 demonstrated superior performance in treatment response evaluation compared with ChatGPT-4o-latest (96.5% vs 82.9%) and showed stronger agreement with expert annotations [62].

In drug discovery, DeepSeek aids with predicting drug-drug interactions and molecular property modeling, achieving superior performance in regression and classification tasks critical to drug discovery [63,64].

The model’s chain-of-thought enabled analysis of factors associated with correct predictions, such as substance abuse and age-related comorbidities. This application underscores DeepSeek’s potential for mental health risk assessment, though further validation is needed [65].

This integrated bibliometric and scoping review provided a comprehensive early-stage mapping of the rapidly evolving research landscape concerning DeepSeek’s applications in medicine. This field is characterized by explosive growth, global engagement, and exploration across a remarkably diverse spectrum of clinical and operational domains. The findings collectively underscore DeepSeek’s emergence not merely as another LLM but as a potent, open-weight contender with specific capabilities that address critical needs in modern health care, including cost-effectiveness, linguistic accessibility, and scalability.

Bibliometric data showed a research frontier that has been intensively explored. The increased publication output regarding applications of DeepSeek in medicine is clear. Our results align with those of an analysis of the global research profile of another LLM, ChatGPT, conducted by Alessandri-Bonetti et al [66], who revealed explosive growth in publications during the first 7 months after its release. This pattern is also consistent with a broader LLM systematic review by Chen et al [67], which reported that, between January 2022 and September 2025, approximately 3.2 clinical LLM studies were published per day, with a linear increase of 7.04 studies per month following the release of ChatGPT. Notably, DeepSeek was not included in the analysis by Chen et al [67], underscoring the gap and the need for our focused review.

The geographical and institutional productivity led by China, followed by Turkey and the United States, reflects widespread international interest of DeepSeek’s potential, with major academic medical centers driving early investigations. Papers on DeepSeek’s applications in medicine have been published in various journals, ranging from well-known open-access journals such as Cureus and Scientific Reports to professional medical informatics and medical education journals such as the Journal of Medical Internet Research. This publication pattern indicates that the research reaches both broad scientific and specialized clinical audiences. Keyword co-occurrence analysis effectively identified the core themes of this research trend. The temporal overlay, which revealed a shift from foundational medical education topics toward more specialized areas such as “retrieval-augmented generation” and “oncology,” illustrates the field’s rapid maturation and deepening focus. Synthesizing the scoping review findings, DeepSeek as a medical tool initially gained attention for its strength in democratizing medical information. For instance, in patient education, it can generate outputs with higher readability than its counterparts, such as ChatGPT.

Perhaps the most striking finding is that DeepSeek has demonstrated competitive and sometimes superior performance compared with existing proprietary models in clinical decision support tasks. The bibliometric analysis revealed that “clinical decision support” formed the largest cluster, while the scoping review further indicated that these studies primarily focused on three specific tasks: “aiding diagnosis,” “differential diagnosis,” and “treatment plan formulation.” The evidence that DeepSeek-V3 can match senior radiologists at specialized diagnostic classifications or that DeepSeek-R1 rivals GPT-4 and OpenAI o1 in diagnostic accuracy across ophthalmology and complex clinicopathological cases challenges the assumption that superior capability is the exclusive domain of closed, commercial models. This “performance parity” achieved through an open-weight architecture has profound implications. Specifically, it suggests a pathway toward breaking the monopoly of advanced AI in clinical support, potentially fostering innovation, reducing costs, and allowing for better adaptation to local health care contexts and linguistic needs.

The utility of this model in medical education and benchmarking further supports its position as a disruptive and cost-effective tool [68]. For institutions and learners worldwide, particularly in resource-constrained settings, DeepSeek offers a viable, high-quality alternative for exam preparation, simulation, and curriculum development, potentially lowering the barriers to accessing advanced medical training aids.

Beyond its direct clinical and educational applications, this review highlighted DeepSeek’s transformative potential across broader health care operations. Documented case studies have demonstrated reductions in documentation error rates in nursing and lower specimen return rates in gynecological examinations and enabled large-scale patient follow-up. By automating a vast array of low-complexity tasks, DeepSeek can free human resources to provide higher quality care and reduce systemic inefficiencies across the health care continuum.

Of the 353 papers included in this scoping review, only 6.8% (24/353) met the criteria for high quality, whereas the majority (235/353, 66.6%) were classified as low quality, consisting predominantly of invalidated benchmarking using examination questions, single-center convenience samples, and proof-of-concept studies. This distribution reflects a critical gap in the current literature: The rapid proliferation of DeepSeek in medicine has been accompanied by an abundance of exploratory studies with limited external validity. Although such benchmarking studies offer valuable insights into the model’s technical capabilities and serve as initial performance indicators, they do not directly inform real-world diagnostic accuracy, patient safety, or clinical utility [69].

In head-to-head comparisons with other LLMs, DeepSeek demonstrated predominantly favorable or comparable performance: Positive outcomes (126/283, 44.5%) were more frequent than negative ones (73/283, 25.8%), and a substantial proportion of studies (84/283, 29.7%) showed no clear superiority of either model. However, because these results derived predominantly from low-quality (235/353, 66.6%) or moderate-quality evidence, with only 6.8% (24/353) meeting high methodological standards, performance claims should be considered preliminary and hypothesis-generating rather than definitive. Clinically, it excels in open-source accessibility, low cost, readability, Chinese language proficiency, and structured reasoning; nonetheless, limitations, including occasional inaccuracies, lower reliability in certain tasks, and the absence of prospective clinical trials, necessitate continued validation and human oversight.

To contextualize the novel and distinct contributions of our work, we compared this review with existing reviews of other LLMs in medicine, such as ChatGPT, GPT-4, LLaMA, and Gemini. Several prior reviews have documented the rapid adoption of proprietary LLMs in health care, highlighting their utility in clinical reasoning, medical education, and patient communication [66,70,71]. However, most existing reviews have primarily focused on closed-source models, which are characterized by limited transparency, restricted capacity for local deployment, and substantial cost barriers. These limitations hinder their scalability and reduce their adaptability across diverse institutional settings. In contrast, this review specifically focused on DeepSeek, an open-weight LLM, and identified several distinctive features that differentiate it from the patterns reported in previous LLM reviews.

First, methodologically, we combined bibliometric analysis with a scoping review to provide both quantitative mapping of research trends and qualitative synthesis of applications and challenges of DeepSeek in medicine, a dual approach rarely applied in prior LLM reviews, which have tended to rely on either bibliometric or narrative synthesis alone [66,71].

Second, geographically, the research landscapes differ substantially. For ChatGPT, early publications were predominantly led by institutions in the United States and Europe, with a wide distribution across high-income countries [66,70]. In contrast, our analysis identified China as the dominant contributor to DeepSeek medical research (163 papers), followed by Turkey and the United States. This pattern aligns with DeepSeek’s country of origin and its rapid deployment across Chinese tertiary hospitals [2,9]. Notably, the early and substantial involvement of Turkish researchers (52 papers) in DeepSeek research is a distinctive feature not observed in early ChatGPT literature.

Third, previous reviews focused predominantly on proprietary models such as ChatGPT, GPT-4, LLaMA, and Gemini. In contrast, our study addressed a significant gap by examining an open-source alternative with distinct architectural advantages and greater deployment flexibility. In terms of real-world deployment, the deployment of DeepSeek across nearly 90 tertiary hospitals in China has resulted in measurable improvements in workflow efficiency and documentation quality. This scale of implementation has not been reported in similar reviews of other LLMs, which have largely focused on simulated or benchmarking studies [6,9]. In terms of application areas, prior work on ChatGPT and other proprietary LLMs identified medical education, clinical decision support, and patient communication as core areas [70,71]. Our keyword co-occurrence analysis confirmed that these are also central themes for DeepSeek. However, DeepSeek’s open-weight architecture introduces distinctive features not emphasized in proprietary LLM reviews: on-premises deployability, data privacy, cost-effectiveness, and superior performance in Chinese-language medical tasks. These features represent unique contributions of DeepSeek to the medical LLM landscape and are not simply typical of any newly introduced LLM. Regarding performance and utility, our findings demonstrated that DeepSeek achieved competitive or superior performance compared with proprietary models in clinical diagnostics, medical licensing examinations, and patient education while substantially reducing costs, advantages that prior reviews have identified as critical unmet needs in AI integration [71,72].

Guided by an ethical framework, the efficacy and safety of any medical intervention must be carefully calibrated in modern medical practice [73]. As aforementioned, DeepSeek demonstrates significant potential for enhancing medical workflows, medical education, and research. However, its application faces numerous challenges in terms of effectiveness and safety, including accuracy issues, data privacy concerns, ethical uncertainties, and diverse global regulations governing AI.

Although DeepSeek has demonstrated diagnostic accuracy comparable to that of specialist clinicians and proprietary models in certain areas [40,74-76], its overall efficacy remains inconsistent [42,77,78]. The model exhibits strong zero-shot and few-shot learning capabilities in general tasks; however, the rapid evolution of medical knowledge necessitates continuous pretraining on extensive volumes of high-quality, domain-specific data. In data-scarce specialties, particularly those lacking sufficient fine-tuning datasets, DeepSeek often fails to effectively acquire new features and patterns, leading to model hallucinations, defined as the generation of seemingly plausible but factually incorrect or unsupported information [36,79]. Such limitations are particularly severe in domains involving rare diseases and complex, nonclassical clinical scenarios, where available pretraining data are often insufficient and clinically unvalidated [36,37,80,81]. Furthermore, as a fundamentally text-based model, DeepSeek exhibits inherent limitations in processing specialized nontextual medical data, such as medical images, complex laboratory metrics, and genomic data [74,82-84]. These constraints collectively contribute to inconsistent model performance across specific medical domains and hinder its generalization.

The integration of DeepSeek into medical practice raises ethical challenges that implicate all 4 foundational principles of biomedical ethics, namely autonomy, nonmaleficence, beneficence, and justice, which were originally proposed by Beauchamp and Childress in 1979 [85].

In addition to challenges such as accuracy, variable performance across medical domains and specialties, and medical ethics and safety issues, the application of DeepSeek in medicine faces other obstacles, including the redesign of clinical workflows, delineation of liability, regulatory lag, and trust and adoption. The deployment of DeepSeek challenges some clinicians’ work habits and creates a demand for professionals who understand both clinical practice and AI. A shortage of talent limits its wider adoption. When errors in DeepSeek-assisted decision-making lead to medical incidents, how should legal responsibility be defined? Should it fall on the operating physician, the hospital that adopted the AI, or the model developers? Currently, global regulations in this field generally lag, and this uncertainty greatly dampens hospitals’ willingness to implement such technologies. Trust remains another challenge; although DeepSeek is easy to use, concerns about risks affect its acceptance [87].

Based on the aforementioned challenges, future research and development should prioritize the directions highlighted in the following sections to advance the reliable, ethical, and equitable integration of DeepSeek into medical practice.

Several limitations of this study should be considered when interpreting the findings. First, the review covered literature published over a relatively short and recent timeframe. Consequently, the observed surge in publications may reflect early enthusiasm rather than sustained scientific progress. Second, although a language-agnostic search strategy was used, most included studies were published in English, with only a small number (n=13) in Chinese. This linguistic imbalance, coupled with the predominance of contributions from researchers based in China, indicates a notable geographical concentration of the available evidence. As a result, the findings may not be directly generalizable to health care systems operating within different regulatory, cultural, or infrastructural contexts. Third, the included studies exhibited substantial heterogeneity in methodologies, medical specialties, evaluation metrics, comparator models, and DeepSeek model versions—for example, R1 versus V3, which differ in parameter counts, training data, and reasoning depth. This variability precluded quantitative synthesis of outcomes and hindered direct cross-study comparisons. Although we reported version-specific findings where available, direct comparisons of performance should be interpreted with caution. Future research should adopt standardized version reporting and benchmark against fixed model checkpoints to enhance comparability and reproducibility. Finally, much of the evidence is derived from benchmarking studies, simulated cases, or retrospective analyses, with a formal quality appraisal showing that 66.6% (235/353) of included original articles were of low quality and only 6.8% (24/353) met the criteria to be considered high quality. Prospective clinical trials or RCTs assessing DeepSeek’s impact on tangible patient health outcomes in real-world clinical settings remain notably scarce. Consequently, the overall quality of the evidence base is inherently preliminary, and the reviewed corpus carries a high risk of bias. The reported strengths of DeepSeek should be interpreted with caution, as these findings predominantly derive from low-quality, controlled, nongeneralizable settings.

This integrated bibliometric and scoping review synthesized the available evidence on DeepSeek’s applications in medicine. The bibliometric analysis revealed a progressive increase in publication output from January 2025 through November 2025, with China, Turkey, and the United States as the leading contributors. Keyword co-occurrence analysis formed 7 clusters; the 3 most frequent keywords were “large language model,” “artificial intelligence,” and “patient education.”

The scoping review found that DeepSeek has been evaluated across 5 primary application domains: patient education and communication, clinical decision support and treatment planning, medical education and benchmarking, clinical workflow optimization, and medical research and data analysis. In these domains, DeepSeek demonstrated variable but often competitive performance compared with proprietary models, with documented strengths in readability of patient education materials, diagnostic accuracy in select specialties, cost-efficiency, and local deployability. Nevertheless, it should be noted that most included studies were of moderate or low quality, and the evidence base is predominantly composed of benchmarking and simulation studies, with a notable scarcity of prospective clinical trials or RCTs assessing patient-relevant outcomes. Additionally, the review identified consistent limitations, including variable performance across medical specialties, model hallucinations, ethical concerns, data privacy challenges, and regulatory gaps. Future integration will require robust prospective clinical validation, expansion of multimodal capabilities, bias mitigation strategies, human-in-the-loop governance frameworks, and equitable access strategies.