Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

DeepSeek in Healthcare: A Survey of Capabilities, Risks, and Clinical Applications of Open-Source Large Language Models

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Open-source AI nearly matches costly rivals at a fraction of the cost

desk verdict A useful but methodologically thin narrative review of DeepSeek-R1 in healthcare; the synthesis is broadly consistent with the literature, but the comparative claims rest on shaky preprint point estimates and a few factual errors. read the letter →

arxiv 2506.01257 v1 pith:T6GZT36L submitted 2025-06-02 cs.CL cs.AI

Jiancheng Ye , Sophie Bronstein , Jiarui Hai , Malak Abu Hashish This is my paper
classification cs.CLcs.AI
keywords DeepSeek-R1largelanguagemodelshealthcareAIchain-of-thoughtreasoningmixtureofexpertsclinicaldecisionsupportsafetybenchmarkevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This survey argues that DeepSeek-R1—an open-source large language model built on mixture-of-experts, chain-of-thought reasoning, and reinforcement learning—can match or approach proprietary models on structured medical and mathematical tasks while costing far less, but that its open weights and reasoning style carry elevated safety, bias, and misinformation risks. Its evidence includes competitive USMLE performance, 87.0% accuracy on pediatric MedQA questions versus ChatGPT-o1's 92.8%, and 82.0% accuracy on ophthalmology cases matching OpenAI o1 at roughly 15-fold lower token cost. The same profile shows outside medicine: 86.7% on AIME 2024 and a 96.3rd-percentile Codeforces rank. The survey therefore presents DeepSeek-R1 as a viable low-cost alternative for resource-limited healthcare settings, provided alignment, bias mitigation, and regulatory safeguards are added. If the comparative picture holds, open-weight models could lower the financial barrier to AI-assisted clinical decision support.

What carries the argument

The load-bearing mechanism is the hybrid architecture: a mixture-of-experts (MoE) network that activates only relevant expert subnetworks per input, Group Relative Policy Optimization (GRPO) reinforcement learning with self-reflection, and chain-of-thought (CoT) reasoning that decomposes tasks into explicit steps. This combination drives the survey's two headline results: high structured-task accuracy (86.7% on AIME, 82% in ophthalmology, 87.0% on pediatric MedQA) and low inference cost (15 to 27 times cheaper than OpenAI's o1) from sparse expert activation, while the same CoT mechanism creates token bloat and a hijacking surface that the survey ties to elevated safety risk. The MoE sparsity makes the cost claims credible, and the CoT visibility is what makes the interpretability claims credible.

What would settle it

Re-run the three load-bearing comparisons—ophthalmology cases, pediatric MedQA, and USMLE—on the same DeepSeek-R1 checkpoint with identical prompt templates and scoring rules; if the accuracy gaps or the 15-fold cost ratio move materially from the cited values, the survey's central cost-performance and parity claims would need revision.

Watch

Extended reading notes

Core claim

The paper's central claim is that an open-source LLM trained with reinforcement learning and chain-of-thought reasoning can be both high-performing and cheap in exactly the domains where healthcare needs help—structured clinical reasoning and quantitative problem solving—while being measurably less safe than closed rivals. On the survey's reading, DeepSeek-R1 delivers about 82% diagnostic accuracy in ophthalmology, equal to OpenAI's o1 but at 15 times lower token cost; 87.0% on pediatric MedQA, slightly below ChatGPT-o1's 92.8%; strong USMLE and MedXpertQA results; and inference priced about 27 times cheaper than o1. The same evidence shows the flip side: roughly 83% of political, racial, and gender prompts produce biased outputs, the model is reported three times more bias-prone than Claude-3 Opus, four times more likely than GPT-4o to propagate misinformation, and susceptible to fine-tuning attacks and jailbreaks that exploit the reasoning chain. The survey concludes that DeepSeek-R1 is a promising but not-yet-safe tool: useful for low-resource clinical settings and research, but requiring hybrid RL+SFT alignment, domain-specific validation, and regulatory clarity before routine use.

Load-bearing premise

The survey's comparative numbers assume the underlying benchmark studies used comparable prompts, model versions, and evaluation protocols; if those differ across papers, the head-to-head accuracy and cost ratios are not trustworthy.

Editorial extensions

If this is right

  • Institutions with limited budgets can run DeepSeek-R1 locally for diagnostic support and medical education, avoiding per-query API fees and keeping patient data on-site.
  • Any such deployment should be coupled with red-teaming and hybrid RL+SFT alignment, because the same open weights enable the fine-tuning attacks the survey documents.
  • Distilled smaller variants extend the same reasoning capabilities to edge devices and low-bandwidth settings, widening the cost advantage beyond server deployments.
  • The token-intensive CoT style creates latency and higher compute load for real-time or high-throughput use, so organizations must weigh reasoning depth against response time.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The survey's cost and accuracy ratios come from different studies with different prompts and model checkpoints; a single standardized benchmark across models would test whether the 15x-27x advantage survives controlled conditions.
  • A natural next step the paper does not run is a prospective deployment in one specialty, such as emergency triage, with physician oversight to measure diagnostic agreement, safety incidents, and actual per-case cost.
  • The reported bias and misinformation ratios come from separate evaluations with different probes, so the most defensible reading is that DeepSeek-R1 is riskier than closed models in tendency, not that the exact 3x/4x figures are directly transferable.
  • Because the recipe of MoE, GRPO, CoT, and distillation appears reproducible, the open architecture could push high-reasoning LLMs toward commodity pricing while leaving safety alignment as the main competitive differentiator.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. This manuscript is a narrative survey of DeepSeek-R1 in healthcare, covering the model's training architecture (MoE, CoT, RL/GRPO), comparisons with open- and closed-source LLMs, clinical applications (pediatrics, ophthalmology, USMLE, patient education, nephrology, drug-drug interaction prediction), strengths and limitations, ethics, and future directions. The central claim is that DeepSeek-R1 is a competitive, cost-effective open-source alternative for resource-constrained healthcare settings, while carrying elevated safety and bias risks. The survey is based on a mixture of peer-reviewed papers, preprints, and news-style sources, with no systematic search or selection methodology.

Significance. If accurate, this survey would be a useful entry point for clinicians and health informatics researchers considering open-source LLMs: it gathers recent evaluations of DeepSeek-R1 across several specialties and highlights the open-source/closed-source trade-off. Its strengths include organizing a fast-moving literature, naming specific benchmarks (USMLE, MedXpertQA, AIME, ophthalmology/pediatrics evaluations), and acknowledging safety concerns with concrete references (fine-tuning attacks, H-CoT, safety benchmarks). However, the survey's value is limited by the absence of a documented methodology, by reliance on unreviewed preprints, and by several factual errors (e.g., GRPO mislabeled as 'Gradient-Based Policy Optimization'; Mixtral described as closed-source; conflicting cost claims of 15-fold vs. 27-fold). These issues do not invalidate the broad qualitative message, but they prevent the paper from being a reliable reference without substantial revision.

major comments (5)
  1. [Abstract and §2.2, §Strengths] The central cost-efficiency claim is internally inconsistent: §2.2 states that DeepSeek-R1 matched GPT-o1's 82% accuracy in ophthalmology 'at a fraction of the computational cost' and §2.5 gives 'almost 15 times lower', while the Strengths section states the inference cost is 'reportedly 27 times cheaper than OpenAI's o1 model'. The manuscript never reconciles these two figures or cites which underlying source supports which number. Because the 'viable low-cost alternative' claim is load-bearing for the survey's conclusion, the authors should either report the figures with their sources and context or drop the unsupported 27-fold claim.
  2. [§2.2 and Table 2] The headline comparative claims rest on single preprint point estimates presented without confidence intervals, sample sizes, or evaluation protocols. For example, the ophthalmology claim ('82.0% accuracy ... 15-fold lower cost') and the pediatric claim ('87.0% vs 92.8%') are drawn from two medRxiv preprints, and Table 2 provides only the accuracy percentages. Since the survey's central conclusion depends on these cross-model comparisons, the authors should explicitly report the number of cases, the statistical uncertainty, and any differences in prompts or test-set construction, or weaken the categorical language in the abstract. As written, the comparison implies a precision that the cited studies likely do not support.
  3. [§1.1] Section 1.1 misnames GRPO as 'Gradient-Based Policy Optimization' in the sentence introducing DeepSeek-R1's reinforcement learning mechanism. GRPO is Group Relative Policy Optimization, as correctly abbreviated in Figure 1's caption. This is a factual error in the technical description of the model's training, and it should be corrected.
  4. [§2.1] Section 2.1 states that 'Mixtral—developed by Mistral AI—is one of the few closed-source models that also uses MoE'. Mixtral is an open-weights model released under the Apache 2.0 license. Placing it in the closed-source category confuses the paper's central open/closed distinction and undermines the comparison in Table 1. The authors should either classify Mixtral correctly or remove the example.
  5. [§1 (Introduction) and overall methodology] The survey states no search strategy, inclusion criteria, or quality appraisal for the studies summarized in Table 2. Given that the manuscript presents itself as a survey of capabilities and risks, the lack of a systematic methodology is a load-bearing limitation: heterogeneous preprints and peer-reviewed studies are treated with equal weight, and no attempt is made to account for evaluation-date, model-version, or prompt differences. The authors should add a short Methods subsection describing how references were identified and selected, and should flag in the text when a specific number comes from a non-peer-reviewed preprint.
minor comments (6)
  1. [§1.2] The paragraph on 'Advanced reasoning and symbolic computation' states that DeepSeek-R1 'has performed competitively on USMLE' without giving a citation in that sentence; the citation appears later in §Clinical Applications (reference 23). Consider citing the USMLE evaluation at its first mention.
  2. [§2.2] The text says 'In a task involving ophthalmology-based questions, DeepSeek-R1 matched GPT-o1's 82% accuracy but at a fraction of the computational cost.' The following sentence claims 'both scoring above 75% on general NLP benchmarks' for GPT-4o and Claude-3 Opus, but no benchmark name or source is given for those numbers. Add a reference or remove the specific figures.
  3. [Table 2] Table 2 has formatting inconsistencies: the country 'Chins' for Zhang W et al. should read 'China', and several rows have inconsistent spacing in the 'Implications' column. Also, reference [40] is listed in the reference list as 'From GPT to DeepSeek' with a typo ('remains' instead of 'remain' in the title).
  4. [§Strengths and Limitations] The limitation paragraph says DeepSeek-R1 is 'reportedly three times more prone to bias than Claude-3 Opus and four times more likely to propagate misinformation than GPT-4o', but it does not say which study produced these ratios or how they were measured. These are strong quantitative claims and need a citation with the specific evaluation context.
  5. [§Ethical Considerations] The sentence 'DeepSeek-R1 exhibits biased behavior in approximately 83% of political, racial, and gender-related prompts' is reported without a citation. The cited reference [11] is about fine-tuning attacks, not a bias-rate audit. The authors should either provide the correct source or soften the claim.
  6. [Figure 1] Figure 1's caption lists GRPO as 'Group Relative Policy Optimization', which is correct, but the body text in §1.1 uses the wrong expansion. Please harmonize the two.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a survey that summarizes external benchmark studies and does not derive claims from its own assumptions or fitted parameters.

full rationale

This manuscript is a literature survey of DeepSeek-R1 in healthcare. It contains no equations, no fitted parameters, no new empirical predictions, and no derivation chain that could reduce to its own inputs. Claims such as competitive USMLE performance, 82% ophthalmology accuracy, and 87.0% vs. 92.8% pediatric accuracy are explicitly attributed to external studies (Mikhail et al. medRxiv 2025; Mondillo et al. medRxiv 2025; de Paiva et al. medRxiv 2025), and Table 2 presents these as reported results from those studies rather than as new findings. The paper's self-citations (e.g., refs. 1, 9, 15, 16, 27, 31, 32, 36, 44, 45, 46) are background citations on clinical informatics, digital health, NLP methods, and telehealth; none of them supplies the load-bearing evidence for the central comparative claims about DeepSeek-R1, and none is invoked as a uniqueness theorem or as an ansatz that forces the survey's conclusions. The central synthesis therefore has independent content: it aggregates externally verifiable benchmark reports, and its strengths/limitations discussion is a qualitative summary of those reports. Concerns about whether the cited preprint point estimates are protocol-comparable or statistically robust are correctness or evidence-quality concerns, not circularity. There is no step where a prediction is defined in terms of the data it is supposed to explain, no fitted parameter is renamed as a prediction, and no central claim is justified only by a self-citation. Accordingly, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

No new entities are proposed; the only postulates are the reliability and comparability of the cited literature.

assumptions (2)
  • domain assumption The cited primary studies are accurate and their outcome measures are directly comparable.
    The survey makes no independent measurements; all performance and safety claims are inherited from cited papers. The cross-model comparisons in 'Comparisons with Other Models' depend on this.
  • domain assumption DeepSeek-R1's architecture, training pipeline, and MIT license are as stated in the cited technical report and secondary analyses.
    The architectural descriptions in 'Training and Architecture' and the comparisons in 'Comparisons with Other Models' rest on the accuracy of the underlying sources.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepSeek in Healthcare: A Survey of Capabilities, Risks, and Clinical Applications of Open-Source Large Language Models." pith.science (2026). https://pith.science/paper/T6GZT36L

@misc{pith2026250601257,
  author       = {Pith},
  title        = {Pith review of: DeepSeek in Healthcare: A Survey of Capabilities, Risks, and Clinical Applications of Open-Source Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T6GZT36L}},
  note         = {Machine review of arXiv:2506.01257}
}
read the original abstract

DeepSeek-R1 is a cutting-edge open-source large language model (LLM) developed by DeepSeek, showcasing advanced reasoning capabilities through a hybrid architecture that integrates mixture of experts (MoE), chain of thought (CoT) reasoning, and reinforcement learning. Released under the permissive MIT license, DeepSeek-R1 offers a transparent and cost-effective alternative to proprietary models like GPT-4o and Claude-3 Opus; it excels in structured problem-solving domains such as mathematics, healthcare diagnostics, code generation, and pharmaceutical research. The model demonstrates competitive performance on benchmarks like the United States Medical Licensing Examination (USMLE) and American Invitational Mathematics Examination (AIME), with strong results in pediatric and ophthalmologic clinical decision support tasks. Its architecture enables efficient inference while preserving reasoning depth, making it suitable for deployment in resource-constrained settings. However, DeepSeek-R1 also exhibits increased vulnerability to bias, misinformation, adversarial manipulation, and safety failures - especially in multilingual and ethically sensitive contexts. This survey highlights the model's strengths, including interpretability, scalability, and adaptability, alongside its limitations in general language fluency and safety alignment. Future research priorities include improving bias mitigation, natural language comprehension, domain-specific validation, and regulatory compliance. Overall, DeepSeek-R1 represents a major advance in open, scalable AI, underscoring the need for collaborative governance to ensure responsible and equitable deployment.

Figures

Figures reproduced from arXiv: 2506.01257 by the authors.

Figure 1
Figure 1. Workflow of DeepSeek -R1 . Abbreviations: CoT: Chain -of -Thought GRPO: Group Relative Policy Optimization SFT: Supervised Fine Tuning RL: Reinforcement Learning [PITH_FULL_IMAGE:figures/full_fig_p050_1.png] view at source ↗
Figure 2
Figure 2. Ethical considerations of DeepSeek -R1 [PITH_FULL_IMAGE:figures/full_fig_p051_2.png] view at source ↗
Figure 3
Figure 3. Future directions of DeepSeek -R1 [PITH_FULL_IMAGE:figures/full_fig_p052_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Verifiable Manifest Signing and Transparency Enforcement for Secure MCP-Based LLM Pipelines

    cs.CR 2026-01 reject novelty 3.0 of 10

    The paper proposes a security layer that signs, freshness-checks, policy-validates, and Merkle-logs MCP tool-use manifests before LLM execution, reporting near-linear scaling and low overhead.

Reference graph

Works this paper leans on

60 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ye, J. and L.N. Sanchez-Pinto. Three data-driven phenotypes of multiple organ dysfunction syndrome preserved from early childhood to middle adulthood. in AMIA Annual Symposium Proceedings. 2020. American Medical Informatics Association

    2020

  2. [2]

    medRxiv, 2024: p

    Zhang, Z., et al., Echo-Vision-FM: A Pre-training and Fine-tuning Framework for Echocardiogram Videos Vision Foundation Model. medRxiv, 2024: p. 2024.10.09.24315195

    work page 2024

  3. [3]

    arXiv preprint arXiv:2501.12948, 2025

    Guo, D., et al., Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    arXiv 2025

  4. [4]

    Spillard, and D.P

    Mercer, S., S. Spillard, and D.P. Martin, Brief analysis of DeepSeek R1 and its implications for Generative AI. 2025

    work page 2025

  5. [5]

    Guo, D., et al., DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. 2025

    work page 2025

  6. [6]

    Kuo, M., et al., H-CoT: Hijacking the Chain-of-Thought Safety Reasoning Mechanism to Jailbreak Large Reasoning Models, Including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking. 2025

    work page 2025

  7. [7]

    Zhang, W., et al., Safety evaluation of deepseek models in chinese contexts. 2025

    work page 2025

  8. [8]

    2025, Elsevier BV

    Krause, D., DeepSeek and FinTech: The Democratization of AI and Its Global Implications. 2025, Elsevier BV

    work page 2025

Show all 60 references
  1. [9]

    JMIR Perioperative Medicine, 2023

    Ye, J., Patient Safety of Perioperative Medication Through the Lens of Digital Health and Artificial Intelligence. JMIR Perioperative Medicine, 2023. 6: p. e34453

    2023

  2. [10]

    Journal of Biomedical Informatics, 2025

    Peng, Y., et al., From GPT to DeepSeek: Significant gaps remain in realizing AI in healthcare. Journal of Biomedical Informatics, 2025. 163: p. 104791

    2025

  3. [11]

    Gardiner, and S

    Xu, Z., J. Gardiner, and S. Belguith, The dark deep side of DeepSeek: Fine-tuning attacks against the safety alignment of CoT-enabled models. 2025

    2025

  4. [12]

    Moran, A.E. and R. Gupta, Implementation of global hearts hypertension control programs in 32 low-and middle-income countries: JACC international. Journal of the American College of Cardiology, 2023. 82(19): p. 1868-1884

    2023

  5. [13]

    Boye, J. and B. Moell, Large language models and mathematical reasoning failures. arXiv preprint arXiv:2502.11574, 2025

    2025 arXiv

  6. [14]

    arXiv preprint arXiv:2501.18576, 2025

    Evstafev, E., Token-Hungry, Yet Precise: DeepSeek R1 Highlights the Need for Multi- Step Reasoning Over Speed in MATH. arXiv preprint arXiv:2501.18576, 2025

    2025 arXiv

  7. [15]

    AMIA Summits on Translational Science Proceedings, 2024

    Ye, J., et al., The role of artificial intelligence for the application of integrating electronic health records and patient-generated data in clinical decision support. AMIA Summits on Translational Science Proceedings, 2024. 2024: p. 459

    2024

  8. [16]

    medRxiv, 2024: p

    Ye, J., et al., Development and Application of Natural Language Processing on Unstructured Data in Hypertension: A Scoping Review. medRxiv, 2024: p. 2024.02. 27.24303468

    2024

  9. [17]

    medRxiv, 2025: p

    Mikhail, D., et al., Performance of DeepSeek-R1 in Ophthalmology: An Evaluation of Clinical Decision-Making and Cost-Effectiveness. medRxiv, 2025: p. 2025.02. 10.25322041

    2025

  10. [18]

    arXiv preprint arXiv:2502.11544, 2025

    Chen, A., et al., Evaluating o1-like llms: Unlocking reasoning for translation through comprehensive analysis. arXiv preprint arXiv:2502.11544, 2025

    2025 arXiv

  11. [19]

    Neha, F. and D. Bhati, A Survey of DeepSeek Models. Authorea Preprints, 2025

    2025

  12. [20]

    Parmar, M. and Y. Govindarajulu, Challenges in Ensuring AI Safety in DeepSeek-R1 Models: The Shortcomings of Reinforcement Learning Strategies. arXiv preprint arXiv:2501.17030, 2025

    2025 arXiv

  13. [21]

    International Journal of Scientific 40 Research in Computer Science, Engineering and Information Technology (IJSRCSEIT),

    Pendyala, S.K., Enhancing Healthcare Pricing Transparency: A Machine Learning and AI-Driven Approach to Pricing Strategies and Analytics. International Journal of Scientific 40 Research in Computer Science, Engineering and Information Technology (IJSRCSEIT),

  14. [22]

    DeepSeek-R1

    Mondillo, G., et al., Comparative Evaluation of Advanced AI Reasoning Models in Pediatric Clinical Decision Support: ChatGPT O1 vs. DeepSeek-R1. medRxiv, 2025: p. 2025.01. 27.25321169

    2025

  15. [23]

    Faray de Paiva, L., et al., How does DeepSeek-R1 perform on USMLE? medRxiv, 2025: p. 2025.02. 06.25321749

    2025

  16. [24]

    arXiv preprint arXiv:2501.18362, 2025

    Zuo, Y., et al., MedXpertQA: Benchmarking Expert-Level Medical Reasoning and Understanding. arXiv preprint arXiv:2501.18362, 2025

    2025 arXiv

  17. [25]

    Zhao, M., et al., Evaluating the Readability and Quality of AI-Generated Scoliosis Education Materials: A Comparative Analysis of Five Language Models. 2025

    2025

  18. [26]

    arXiv preprint arXiv:2503.04153, 2025

    Long, Y., et al., KidneyTalk-open: No-code Deployment of a Private Large Language Model with Medical Documentation-Enhanced Knowledge Database for Kidney Disease. arXiv preprint arXiv:2503.04153, 2025

    2025

  19. [27]

    Ye, J. and S. Bronstein, Artificial intelligence and digital health in Anesthesiology, in The Digital Doctor. 2025, Elsevier. p. 291-307

    2025

  20. [28]

    Ferrucci, and A

    De Vito, G., F. Ferrucci, and A. Angelakis, LLMs for Drug-Drug Interaction Prediction: A Comprehensive Comparison. arXiv preprint arXiv:2502.06890, 2025

    2025 arXiv

  21. [29]

    Cureus, 2025

    Temsah, A., et al., DeepSeek in healthcare: revealing opportunities and steering challenges of a new open-source artificial intelligence frontier. Cureus, 2025. 17(2)

    2025

  22. [30]

    Shahsavar, and H

    Choudhury, A., Y. Shahsavar, and H. Shamszare, User Intent to Use DeepSeek for Healthcare Purposes and their Trust in the Large Language Model: Multinational Survey Study. arXiv preprint arXiv:2502.17487, 2025

    2025 arXiv

  23. [31]

    He, and M

    Ye, J., L. He, and M. Beestrum, Implications for implementation and adoption of telehealth in developing countries: a systematic review of China’s practices and experiences. npj Digital Medicine, 2023. 6(1): p. 174

    2023

  24. [32]

    JAMIA Open, 2023

    Ye, J., et al., Leveraging natural language processing and geospatial time series model to analyze COVID-19 vaccination sentiment dynamics on Tweets. JAMIA Open, 2023. 6(2): p. ooad023

    2023

  25. [33]

    Hayder, W.A., Highlighting DeepSeek-R1: Architecture, Features and Future Implications. 2025

    2025

  26. [34]

    Scientific Reports, 2024

    Ding, S., et al., Distilling the knowledge from large-language model for health event prediction. Scientific Reports, 2024. 14(1): p. 30675

    2024

  27. [35]

    arXiv preprint arXiv:2502.12659, 2025

    Zhou, K., et al., The hidden risks of large reasoning models: A safety assessment of r1. arXiv preprint arXiv:2502.12659, 2025

    2025

  28. [36]

    Ye, J. and Z. Ren, Examining the impact of sex differences and the COVID-19 pandemic on health and health care: findings from a national cross-sectional study. JAMIA Open, 2022

    2022

  29. [37]

    Arrieta, A., et al., o3-mini vs DeepSeek-R1: Which One is Safer? arXiv preprint arXiv:2501.18438, 2025

    2025 arXiv

  30. [38]

    arXiv preprint arXiv:2503.13806, 2025

    Zhang, W., et al., Organ-aware Multi-scale Medical Image Segmentation Using Text Prompt Engineering. arXiv preprint arXiv:2503.13806, 2025

    2025

  31. [39]

    Nature medicine, 2025: p

    Tordjman, M., et al., Comparative benchmarking of the DeepSeek large language model on medical tasks and clinical reasoning. Nature medicine, 2025: p. 1-1

    2025

  32. [40]

    2025, Elsevier

    Peng, Y., et al., From GPT to DeepSeek: Significant gaps remains in realizing AI in healthcare. 2025, Elsevier. p. 104791

    2025

  33. [41]

    Nature Medicine, 2025: p

    Sandmann, S., et al., Benchmark evaluation of DeepSeek large language models in clinical decision-making. Nature Medicine, 2025: p. 1-1

    2025

  34. [42]

    The Lancet Digital Health, 2024

    Ong, J.C.L., et al., Ethical and regulatory challenges of large language models in medicine. The Lancet Digital Health, 2024. 6(6): p. e428-e432. 41

    2024

  35. [43]

    arXiv preprint arXiv:2112.04359, 2021

    Weidinger, L., et al., Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359, 2021

    2021 arXiv

  36. [44]

    AMIA Summits on Translational Science Proceedings, 2024

    Ye, J., et al., Multimodal data hybrid fusion and natural language processing for clinical prediction models. AMIA Summits on Translational Science Proceedings, 2024. 2024: p. 191

    2024

  37. [45]

    Analytical Methods, 2017

    Ye, J., et al., A portable urine analyzer based on colorimetric detection. Analytical Methods, 2017. 9(16): p. 2464-2471

    2017

  38. [46]

    Design and development of an informatics-driven implementation research framework for primary care studies

    Ye, J. Design and development of an informatics-driven implementation research framework for primary care studies. in AMIA Annual Symposium Proceedings. 2021. American Medical Informatics Association

    2021

  39. [47]

    medRxiv, 2025: p

    Zhang, S., et al., Machine Learning-Based Mortality Prediction in Critically Ill Patients with Hypertension: Comparative Analysis, Fairness, and Interpretability. medRxiv, 2025: p. 2025.04. 05.25325307

    2025

  40. [48]

    Minssen, and G

    Gerke, S., T. Minssen, and G. Cohen, Ethical and legal challenges of artificial intelligence-driven healthcare, in Artificial intelligence in healthcare. 2020, Elsevier. p. 295-336

    2020

  41. [49]

    Alghamdi, H. and A. Mostafa, Advancing EHR analysis: Predictive medication modeling using LLMs. Information Systems, 2025: p. 102528

    2025

  42. [50]

    arXiv preprint arXiv:2503.04765, 2025

    Aydin, O., et al., Generative AI in Academic Writing: A Comparison of DeepSeek, Qwen, ChatGPT, Gemini, Llama, Mistral, and Gemma. arXiv preprint arXiv:2503.04765, 2025

    2025 arXiv

  43. [51]

    Olin Business School Center for Finance & Accounting Research Paper, 2023(2023/18)

    Chen, J., et al., ChatGPT, stock market predictability and links to the macroeconomy. Olin Business School Center for Finance & Accounting Research Paper, 2023(2023/18)

    2023

  44. [52]

    Dong, B. and Q. Wang, Evaluating the Performance of the DeepSeek Model in Confidential Computing Environment. arXiv preprint arXiv:2502.11347, 2025

    2025 arXiv

  45. [53]

    Fu, J., et al., LLMPopcorn: An Empirical Study of LLMs as Assistants for Popular Micro- video Generation. 2025

    2025

  46. [54]

    Gupta, G.K. and P. Pande, LLMs in Disease Diagnosis: A Comparative Study of DeepSeek-R1 and O3 Mini Across Chronic Health Conditions. arXiv preprint arXiv:2503.10486, 2025

    2025 arXiv

  47. [55]

    Authorea Preprints, 2025

    Gupta, R., Comparative Analysis of DeepSeek R1, ChatGPT, Gemini, Alibaba, and LLaMA: Performance, Reasoning Capabilities, and Political Bias. Authorea Preprints, 2025

    2025

  48. [56]

    Available at SSRN 5169443

    Liang, J., et al., Evaluation of LLMs Accuracy and Application in Oncology Principles and Practice. Available at SSRN 5169443

  49. [57]

    arXiv preprint arXiv:2503.11068, 2025

    Lin, L., et al., DeepSeek Powered Solid Dosage Formulation Design and Development. arXiv preprint arXiv:2503.11068, 2025

    2025 arXiv

  50. [58]

    arXiv preprint arXiv:2503.04691, 2025

    Qiu, P., et al., Quantifying the reasoning abilities of llms on real-world clinical cases. arXiv preprint arXiv:2503.04691, 2025

    2025 arXiv

  51. [59]

    arXiv preprint arXiv:2502.17947, 2025

    Xu, P., et al., Deepseek-r1 outperforms gemini 2.0 pro, openai o1, and o3-mini in bilingual complex ophthalmology reasoning. arXiv preprint arXiv:2502.17947, 2025

    2025 arXiv

  52. [60]

    rambling

    Zhan, Z., et al., An evaluation of deepseek models in biomedical natural language processing. arXiv preprint arXiv:2503.00624, 2025. 1 Table 1. Comparison of key characteristics of DeepSeek-R1 and other recent LLMs. 1 Table 2. Summary of selected studies. Authors Country Domai...

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.