The rise of DeepSeek: technology calls for the “catfish effect”
National Center for Respiratory Medicine
During the 2025 Chinese Spring Festival, a topic that garnered widespread attention was DeepSeek. On January 20, the Hangzhou-based DeepSeek company released its latest large language model, DeepSeek-R1. This release sent shockwaves through the technology sector and attracted attention from top scientific journals such as Nature and Science (1,2). With its powerful performance and open-source characteristics, DeepSeek-R1 has created substantial pressure on existing artificial intelligence (AI) competitors, exemplifying the “catfish effect” in the AI domain. This concept originates from a classical management theory: Norwegian fishermen placed catfish, a natural predator, in sardine transport tanks, significantly reducing mortality rates by triggering the sardines’ survival instincts. By analogy, in other fields, the introduction of strong competitors often activates industry innovation dynamics. DeepSeek’s emergence has injected new momentum into the AI field, driving rapid technological iteration and innovation.
Cost-effectiveness and efficiency: DeepSeek’s technological breakthroughsOther Section
DeepSeek-R1’s development costs are significantly lower than comparable Western products, an advantage that has distinguished it in the global AI market. Its core technology, the “Mixture-of-Experts” architecture, optimizes the training process by reducing parameter quantities and chip requirements, thereby substantially lowering costs. Additionally, the Multi-head Latent Attention mechanism enables the model to store more data while occupying less memory. These innovations significantly enhance model efficiency, making it more competitive in resource-constrained environments.
Innovation under restrictions: the birth context of DeepSeekOther Section
Against the backdrop of U.S. restrictions on high-performance chip exports to China, DeepSeek’s emergence demonstrates China’s autonomous innovation capabilities in AI. Through algorithm optimization and independent research and development, DeepSeek has successfully overcome hardware limitations to achieve technological breakthroughs. This achievement not only showcases China’s strength in the AI field but also provides new perspectives for global AI technological development. We believe that China’s advancements in AI will, in turn, stimulate AI development in the U.S., particularly through open-source products like DeepSeek. For the collective advancement of human technology, we unequivocally advocate for openness and oppose restrictions.
Local deployment and data privacy: DeepSeek’s unique advantagesOther Section
DeepSeek-R1’s support for local deployment offers unique value in terms of privacy protection. Researchers can deploy the model on local systems, thereby maintaining complete control over their data and research outcomes. This innovative design is particularly significant for disciplines involving sensitive data, such as medical research.
The power of open source: promoting transparency and collaboration in AI researchOther Section
The open-source nature of DeepSeek-R1 makes its reasoning processes transparent to researchers, thereby enhancing model interpretability. This transparency not only facilitates understanding of the model’s decision-making mechanisms but also enables potential model improvements. The open-source model simultaneously promotes collaboration within the global research community, allowing competitors to iteratively optimize based on DeepSeek’s methods, thus advancing the entire field.
AI applications in medicine: potential and challengesOther Section
AI (including DeepSeek) has demonstrated significant potential in medical applications, particularly in driving paradigm shifts and innovative thinking in research. For instance, AI can assist physicians in early diagnosis and surgical planning through analysis of medical imaging data (3). AI can also help researchers write code, optimize algorithms, and even revise and refine academic papers (4). Furthermore, AI can accelerate research progress by analyzing vast literature databases to help researchers quickly access the latest research developments. The development process of this article itself illustrates the value of AI assistance—a Chinese draft was repeatedly refined through DeepSeek until the final version was completed. This experience stands in stark contrast to earlier approaches: previously, one may need to hire professional editors for language refinement, which was time-consuming, expensive, and communication-intensive. Today, AI technology not only provides an “omniscient partner” available around the clock but also completes dozens of iterative optimizations with remarkable patience. Had someone predicted this scenario 10 years ago, I would have dismissed it as fantasy.
However, AI still faces numerous challenges in clinical applications. First, data bias and generalization issues have not been fully resolved. Training data biases in models like DeepSeek-R1 and ChatGPT may influence clinical research outcomes. Second, AI has limited adaptability in complex contexts and lacks human emotional capacity and judgment. For example, in scenarios such as end-of-life care or mental health interventions, AI performance remains suboptimal. To standardize AI clinical applications, frameworks such as CONSORT-AI (Consolidated Standards of Reporting Trials-Artificial Intelligence) have emerged, providing new reporting standards for evaluating clinical trials with AI interventions (5).
Embracing AI, welcoming competition, maintaining vigilanceOther Section
DeepSeek’s emergence has injected new vitality into the AI field, driving rapid technological advancement. This “catfish effect” not only stimulates competition but also provides new momentum for global AI technological development. However, AI technology is still in the developmental stage and requires human guidance and regulation. We should actively embrace AI while remaining vigilant, ensuring that its applications in medicine and other fields truly benefit humanity.
Journal
Journal of Thoracic Disease
Method of Research
Commentary/editorial
Subject of Research
Not applicable
Article Title
The rise of DeepSeek: technology calls for the “catfish effect”
COI Statement
Conflicts of Interest: The author has completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025b-02/coif). The author has no conflicts of interest to declare.
DeepSeek: the “Watson” to doctors—from assistance to collaboration
DeepSeek is an artificial intelligence (AI) platform built on deep learning and natural language processing (NLP) technologies. Its core products include the DeepSeek-R1 and DeepSeek-V3 models. Leveraging an efficient Mixture of Experts (MoE) architecture, multimodal data fusion capabilities, and significantly reduced training costs (over 90% lower than comparable models), DeepSeek achieves performance on par with OpenAI’s GPT-4o-mini. Through cutting-edge NLP techniques, DeepSeek can rapidly and accurately extract valuable insights from massive datasets, providing users with intelligent information retrieval and analytical solutions. These capabilities unlock new possibilities in healthcare, offering efficient support to both doctors and patients.
In the medical field, clinicians are akin to the legendary detective Sherlock Holmes, while DeepSeek plays the role of “John Watson”—not just an assistant, but also a complement to their thinking and a bridge to humanized care. With its powerful data analysis and reasoning capabilities, DeepSeek helps doctors identify potential “blind spots”, provides comprehensive diagnostic recommendations, optimizes doctor-patient communication, and enhances medical efficiency. Specifically, DeepSeek offers the following benefits: (I) supplementing professional knowledge: through deep learning and multimodal data processing, DeepSeek delivers big data-driven insights, enabling doctors to better understand complex medical conditions. (II) Humanizing healthcare interactions: by enhancing patient education and doctor-patient communication, DeepSeek facilitates more effective interactions, improving the overall healthcare experience. (III) Identifying blind spots: much like Watson, who is often humorously credited with “spotting the blind spots”, DeepSeek’s comprehensive algorithms help doctors uncover details that might otherwise be overlooked. For instance, in diagnosing rare diseases or analyzing complex cases, DeepSeek can identify potential disease features or risk factors, offering doctors a supplementary perspective. (IV) Documentation and organization: just as Watson documented Holmes’ deductive processes for posterity, DeepSeek generates medical record templates and follow-up plans, helping doctors efficiently manage patient information while ensuring data integrity and traceability. Thus, DeepSeek is not merely an intelligent assistant but a collaborative partner in medical practice.
Doctors can utilize DeepSeek for diagnostic and therapeutic assistance while ensuring information privacy protection. Hospitals can also deploy DeepSeek locally and customize its development, training AI models with real-world data to improve the accuracy of diagnoses and treatment recommendations. Below are specific application scenarios for DeepSeek in medical practice: (I) enhancing clinical decision-making efficiency: in clinical settings, doctors can input patient symptoms, test results, and other relevant information to receive diagnostic suggestions and treatment plans from DeepSeek. For instance, when dealing with complex cases, DeepSeek integrates patient medical history and the latest medical research to provide comprehensive decision support. It can also tailor personalized treatment plans based on individual patient characteristics and needs. Additionally, DeepSeek’s data collection and processing capabilities can automatically generate structured medical record drafts, significantly reducing the documentation burden on doctors. (II) Supporting research and academia: DeepSeek can rapidly retrieve the latest medical literature and clinical guidelines, assisting doctors in conducting literature reviews and analyzing research data. This enables them to identify solutions for challenging cases and even propose new scientific questions and research directions. Through human-AI collaboration, doctors can expand their cognitive boundaries, while DeepSeek’s “white-box” interpretability design enhances their understanding of AI reasoning processes and facilitates learning from its insights. (III) Optimizing patient management: DeepSeek can generate personalized patient education programs, helping doctors communicate more effectively with patients and improve treatment adherence. Through its intelligent analysis, doctors can more efficiently manage patient follow-ups and rehabilitation plans, ensuring continuity and effectiveness of care.
At the same time, doctors should actively guide patients in using DeepSeek to improve healthcare efficiency and health management capabilities. This includes: (I) assisting healthcare decision-making: patients can use DeepSeek to preliminarily understand potential diseases corresponding to their symptoms, as well as relevant tests and treatment recommendations. For example, before visiting a doctor, patients can use DeepSeek to interpret lab reports and gain initial diagnostic insights. Additionally, patients can access health science knowledge through DeepSeek to enhance their health literacy and even evaluate the expertise of potential doctors and their suitability for the condition. (II) Optimizing doctor-patient communication: by using DeepSeek, patients gain a deeper understanding of their conditions, enabling more efficient communication with doctors and reducing unnecessary misunderstandings and anxiety. For instance, DeepSeek can generate condition summaries to help patients clearly describe their symptoms to doctors. It can also explain medical terminology, allowing patients to better comprehend diagnoses and treatment plans.
Despite its immense potential in healthcare applications, DeepSeek still faces several challenges that require further optimization: (I) data quality issues: the completeness and accuracy of medical data are critical factors influencing DeepSeek’s effectiveness. Irregular data entry and inconsistent collection standards may lead to biases in model analysis and diagnosis. To address this, DeepSeek needs to integrate professional medical databases (such as PubMed) and incorporate more high-quality cases and information to ensure data diversity and representativeness. Additionally, data cleaning and standardization processes should be implemented to enhance data quality. (II) Algorithm stability and accuracy: when dealing with rare diseases (e.g., those with an incidence rate of <1/100,000) or complex cases, the scarcity of relevant data may prevent DeepSeek from accurately identifying disease characteristics, resulting in unreliable diagnoses. To improve performance, DeepSeek can expand training datasets for rare diseases or adopt transfer learning techniques. Furthermore, rigorous testing and validation are necessary to ensure the reliability of model updates and optimizations in clinical practice. (III) Multimodal data fusion: healthcare involves diverse data types (e.g., images, text, physiological signals), but DeepSeek still faces technical bottlenecks in multimodal data fusion, limiting its ability to integrate data effectively for comprehensive diagnostic and treatment recommendations. To overcome this, hierarchical attention mechanisms (e.g., cross-modal attention weight allocation between images and text) can be employed, along with time-series modeling to capture dynamic features of disease progression. (IV) Lack of automated information collection: currently, DeepSeek relies on user-initiated input, creating barriers to usability. To reduce user burden and improve system accessibility, an active information collection approach is needed. For example, clinical decision tree models can be introduced to develop proactive questioning capabilities during doctor-patient interactions, guiding patients to provide key information and enhancing data collection efficiency. (V) Dynamic update delays: the integration of the latest clinical guidelines and research findings often requires manual review, leading to delays of 1–3 months, which may compromise the timeliness of the model. To address this, continuous learning mechanisms should be implemented, enabling the model to dynamically incorporate the most recent data.
In summary, DeepSeek has brought significant efficiency improvements and convenience to the healthcare sector. However, its application must be integrated with professional medical knowledge and real-world scenarios to ensure safety and accuracy. Looking ahead, DeepSeek is poised to make breakthroughs in personalized medicine, telemedicine, and public health management. By continuously improving data quality, privacy protection, technical optimization, ethical and legal compliance, and industry acceptance, DeepSeek will become a reliable partner for both doctors and patients, driving the intelligent transformation of healthcare. Just as Watson grew under Holmes’ guidance, DeepSeek will continue to refine itself in medical practice, ultimately achieving the leap from assistance to collaboration.
DeepSeek is an artificial intelligence (AI) platform built on deep learning and natural language processing (NLP) technologies. Its core products include the DeepSeek-R1 and DeepSeek-V3 models. Leveraging an efficient Mixture of Experts (MoE) architecture, multimodal data fusion capabilities, and significantly reduced training costs (over 90% lower than comparable models), DeepSeek achieves performance on par with OpenAI’s GPT-4o-mini. Through cutting-edge NLP techniques, DeepSeek can rapidly and accurately extract valuable insights from massive datasets, providing users with intelligent information retrieval and analytical solutions. These capabilities unlock new possibilities in healthcare, offering efficient support to both doctors and patients.
In the medical field, clinicians are akin to the legendary detective Sherlock Holmes, while DeepSeek plays the role of “John Watson”—not just an assistant, but also a complement to their thinking and a bridge to humanized care. With its powerful data analysis and reasoning capabilities, DeepSeek helps doctors identify potential “blind spots”, provides comprehensive diagnostic recommendations, optimizes doctor-patient communication, and enhances medical efficiency. Specifically, DeepSeek offers the following benefits: (I) supplementing professional knowledge: through deep learning and multimodal data processing, DeepSeek delivers big data-driven insights, enabling doctors to better understand complex medical conditions. (II) Humanizing healthcare interactions: by enhancing patient education and doctor-patient communication, DeepSeek facilitates more effective interactions, improving the overall healthcare experience. (III) Identifying blind spots: much like Watson, who is often humorously credited with “spotting the blind spots”, DeepSeek’s comprehensive algorithms help doctors uncover details that might otherwise be overlooked. For instance, in diagnosing rare diseases or analyzing complex cases, DeepSeek can identify potential disease features or risk factors, offering doctors a supplementary perspective. (IV) Documentation and organization: just as Watson documented Holmes’ deductive processes for posterity, DeepSeek generates medical record templates and follow-up plans, helping doctors efficiently manage patient information while ensuring data integrity and traceability. Thus, DeepSeek is not merely an intelligent assistant but a collaborative partner in medical practice.
Doctors can utilize DeepSeek for diagnostic and therapeutic assistance while ensuring information privacy protection. Hospitals can also deploy DeepSeek locally and customize its development, training AI models with real-world data to improve the accuracy of diagnoses and treatment recommendations. Below are specific application scenarios for DeepSeek in medical practice: (I) enhancing clinical decision-making efficiency: in clinical settings, doctors can input patient symptoms, test results, and other relevant information to receive diagnostic suggestions and treatment plans from DeepSeek. For instance, when dealing with complex cases, DeepSeek integrates patient medical history and the latest medical research to provide comprehensive decision support. It can also tailor personalized treatment plans based on individual patient characteristics and needs. Additionally, DeepSeek’s data collection and processing capabilities can automatically generate structured medical record drafts, significantly reducing the documentation burden on doctors. (II) Supporting research and academia: DeepSeek can rapidly retrieve the latest medical literature and clinical guidelines, assisting doctors in conducting literature reviews and analyzing research data. This enables them to identify solutions for challenging cases and even propose new scientific questions and research directions. Through human-AI collaboration, doctors can expand their cognitive boundaries, while DeepSeek’s “white-box” interpretability design enhances their understanding of AI reasoning processes and facilitates learning from its insights. (III) Optimizing patient management: DeepSeek can generate personalized patient education programs, helping doctors communicate more effectively with patients and improve treatment adherence. Through its intelligent analysis, doctors can more efficiently manage patient follow-ups and rehabilitation plans, ensuring continuity and effectiveness of care.
At the same time, doctors should actively guide patients in using DeepSeek to improve healthcare efficiency and health management capabilities. This includes: (I) assisting healthcare decision-making: patients can use DeepSeek to preliminarily understand potential diseases corresponding to their symptoms, as well as relevant tests and treatment recommendations. For example, before visiting a doctor, patients can use DeepSeek to interpret lab reports and gain initial diagnostic insights. Additionally, patients can access health science knowledge through DeepSeek to enhance their health literacy and even evaluate the expertise of potential doctors and their suitability for the condition. (II) Optimizing doctor-patient communication: by using DeepSeek, patients gain a deeper understanding of their conditions, enabling more efficient communication with doctors and reducing unnecessary misunderstandings and anxiety. For instance, DeepSeek can generate condition summaries to help patients clearly describe their symptoms to doctors. It can also explain medical terminology, allowing patients to better comprehend diagnoses and treatment plans.
Despite its immense potential in healthcare applications, DeepSeek still faces several challenges that require further optimization: (I) data quality issues: the completeness and accuracy of medical data are critical factors influencing DeepSeek’s effectiveness. Irregular data entry and inconsistent collection standards may lead to biases in model analysis and diagnosis. To address this, DeepSeek needs to integrate professional medical databases (such as PubMed) and incorporate more high-quality cases and information to ensure data diversity and representativeness. Additionally, data cleaning and standardization processes should be implemented to enhance data quality. (II) Algorithm stability and accuracy: when dealing with rare diseases (e.g., those with an incidence rate of <1/100,000) or complex cases, the scarcity of relevant data may prevent DeepSeek from accurately identifying disease characteristics, resulting in unreliable diagnoses. To improve performance, DeepSeek can expand training datasets for rare diseases or adopt transfer learning techniques. Furthermore, rigorous testing and validation are necessary to ensure the reliability of model updates and optimizations in clinical practice. (III) Multimodal data fusion: healthcare involves diverse data types (e.g., images, text, physiological signals), but DeepSeek still faces technical bottlenecks in multimodal data fusion, limiting its ability to integrate data effectively for comprehensive diagnostic and treatment recommendations. To overcome this, hierarchical attention mechanisms (e.g., cross-modal attention weight allocation between images and text) can be employed, along with time-series modeling to capture dynamic features of disease progression. (IV) Lack of automated information collection: currently, DeepSeek relies on user-initiated input, creating barriers to usability. To reduce user burden and improve system accessibility, an active information collection approach is needed. For example, clinical decision tree models can be introduced to develop proactive questioning capabilities during doctor-patient interactions, guiding patients to provide key information and enhancing data collection efficiency. (V) Dynamic update delays: the integration of the latest clinical guidelines and research findings often requires manual review, leading to delays of 1–3 months, which may compromise the timeliness of the model. To address this, continuous learning mechanisms should be implemented, enabling the model to dynamically incorporate the most recent data.
In summary, DeepSeek has brought significant efficiency improvements and convenience to the healthcare sector. However, its application must be integrated with professional medical knowledge and real-world scenarios to ensure safety and accuracy. Looking ahead, DeepSeek is poised to make breakthroughs in personalized medicine, telemedicine, and public health management. By continuously improving data quality, privacy protection, technical optimization, ethical and legal compliance, and industry acceptance, DeepSeek will become a reliable partner for both doctors and patients, driving the intelligent transformation of healthcare. Just as Watson grew under Holmes’ guidance, DeepSeek will continue to refine itself in medical practice, ultimately achieving the leap from assistance to collaboration.
Journal
Journal of Thoracic Disease
Journal of Thoracic Disease
DOI
Method of Research
Commentary/editorial
Commentary/editorial
Subject of Research
Not applicable
Not applicable
Article Title
DeepSeek: the “Watson” to doctors—from assistance to collaboration
DeepSeek: the “Watson” to doctors—from assistance to collaboration
Article Publication Date
28-Mar-2025
28-Mar-2025
COI Statement
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025b-03/coif). N.Z. serves as the Editor-in-Chief of Journal of Thoracic Disease. J.H. serves as the unpaid Executive Editor-in-Chief of Journal of Thoracic Disease. Y.L. is from Guangzhou KingMed Diagnostics Group Co., Ltd. The other authors have no conflicts of interest to declare.
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025b-03/coif). N.Z. serves as the Editor-in-Chief of Journal of Thoracic Disease. J.H. serves as the unpaid Executive Editor-in-Chief of Journal of Thoracic Disease. Y.L. is from Guangzhou KingMed Diagnostics Group Co., Ltd. The other authors have no conflicts of interest to declare.
