Pith. sign in

REVIEW 5 major objections 5 minor 40 references

Standard Applicability Judgment and Cross-jurisdictional Reasoning: A RAG-based Framework for Medical Device Compliance

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A retrieval-augmented system that reads a free-text device description and labels candidate standards as Mandatory, Recommended, or Not Applicable reaches 73% sample-level accuracy and 87% Top-5 recall in a medical device compliance…

desk verdict The task framing is new and worth building on; the reported numbers are not yet interpretable, so the paper needs a conditional reject with major revision. read the letter →

arxiv 2506.18511 v1 pith:U4XPXM6B submitted 2025-06-23 cs.AI

Yu Han , Aaron Ceross , Jeroen H.M. Bergmann This is my paper
classification cs.AI
keywords medicaldeviceregulationstandardapplicabilityretrieval-augmentedgenerationregulatorycompliancelargelanguagemodelscross-jurisdictionalreasoningsemanticretrievalscience
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

Deciding which regulatory standards apply to a medical device is a labor-intensive expert judgment that varies by jurisdiction. This paper tries to automate that judgment with a retrieval-augmented generation (RAG) system: given a free-text device description, it retrieves candidate standards from a curated corpus of more than 3,900 Chinese and U.S. standards, then uses GPT-4 to label each candidate Mandatory, Recommended, or Not Applicable with a traceable justification. The authors build a benchmark of 105 expert-annotated device descriptions and report 73% sample-level classification accuracy and 87% Top-5 retrieval recall, well above retrieval-only, rule-based, and zero-shot LLM baselines. If the system works as reported, it could reduce the cost and error rate of cross-border regulatory submissions and make applicability reasoning auditable.

What carries the argument

The central mechanism is the retrieval-augmented reasoning loop: Sentence-BERT encodes both device descriptions and standard scope texts into dense vectors; a FAISS flat index retrieves top-k standards from a structured JSON corpus of over 3,900 Chinese (GB/YY) and U.S. (FDA/ANSI/AAMI/ASTM) standards; GPT-4, prompted with cue-word rules ('shall' implies Mandatory) and few-shot JSON examples, assigns each candidate a Mandatory, Recommended, or Not Applicable label and a justification; post-processing and a compliance-comparison module align standards across regions, compare clauses and justifications semantically, and flag conflicts such as Mandatory in China but Not Applicable in the U.S. The paper argues that grounding the LLM in retrieved regulatory text, rather than relying on the model's memory alone, is what makes the applicability judgment reliable.

What would settle it

Have two independent regulatory experts re-annotate the 105 device descriptions and compute inter-annotator agreement, then recompute applicability accuracy counting every gold standard not in the top-5 as a miss; if agreement is low or the recomputed accuracy drops materially below 73%, the central claim fails.

Watch

Extended reading notes

Core claim

The paper claims that standard applicability judgment—traditionally a heuristic expert task—can be formulated as a retrieval-augmented semantic classification problem, and that the proposed region-aware RAG agent is the first end-to-end system to solve it in the medical device domain. On a manually annotated benchmark of 105 real device descriptions drawn from NMPA and FDA sources, it reports 71% per-standard applicability classification accuracy, 73% sample-level accuracy (at least one correct labeled match per device), 87% Top-5 retrieval recall, and significantly better performance than retrieval-only, rule-based, and zero-shot LLM baselines.

Load-bearing premise

The benchmark labels correctly represent which standards truly apply to each device description, since no inter-annotator agreement or external audit is reported and the metric's treatment of unretrieved gold standards is not fully specified.

Editorial extensions

If this is right

  • Regulatory teams could submit a free-text device description and receive a ranked, labeled list of applicable standards with reasons, shortening the scoping phase of market-entry assessments.
  • Cross-jurisdictional gap analysis becomes semi-automated: the system can flag when a standard is mandatory in China but has no U.S. counterpart, as in the vacuum blood collection tube case study.
  • The 3,900-entry bilingual corpus and the 105-example benchmark become reusable assets for training and evaluating future regulatory AI systems.
  • Because outputs include clause-level justifications linked to source standards, the approach supports the audit trails needed for regulatory decision support rather than black-box prediction.

Reading between the lines

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

  • Editorial inference: the same retrieval-plus-reasoning architecture could extend to other fragmented rule domains such as pharmaceuticals, environmental compliance, and financial regulation, since the framework treats applicability as a semantic inference problem over structured text.
  • Editorial inference: the reported accuracy depends on the benchmark's gold labels; a natural next step is to publish an inter-annotator agreement study, since without it the 73% figure cannot be compared across future systems.
  • Editorial inference: if the corpus added ISO/IEC standards and full-text chunking, cross-jurisdictional comparisons would become more complete; the current limitation to national standards likely understates the gap-detection value for globally marketed devices.
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 / 5 minor

Summary. The paper proposes a retrieval-augmented generation (RAG) system for determining whether medical device standards are Mandatory, Recommended, or Not Applicable for a given device description, with structured justifications and cross-jurisdictional (China vs. US) conflict analysis. The system combines a curated regulatory corpus with Sentence-BERT-style retrieval, GPT-based classification, and a rule-based gap-analysis module. The authors construct a benchmark of 105 device descriptions with expert labels and report a Top-5 retrieval recall of 0.87, an applicability classification accuracy of 0.71, and a sample-level accuracy of 0.73, claiming this is the first end-to-end system for standard applicability reasoning. The central claim is that the proposed RAG pipeline substantially outperforms retrieval-only, rule-based, and zero-shot LLM baselines.

Significance. If the reported evaluation is made interpretable and reproducible, the paper could be a useful contribution to regulatory AI: it introduces a domain-specific bilingual corpus, a small manually annotated benchmark, a modular RAG architecture with traceable JSON outputs, and a concrete cross-jurisdictional use case. The authors deserve credit for building the corpus and benchmark from public sources and for comparing against several baselines. However, the headline numbers are currently ambiguous, the benchmark validation is thin, and the exact retrieval configuration is described inconsistently across sections. The contribution is therefore not yet independently verifiable, and the central effectiveness claim needs substantial clarification before the paper can be accepted.

major comments (5)
  1. [§4.1, §4.2, Contributions] The corpus size is stated inconsistently. Section 4.1 says the dataset comprises 150 curated NMPA entries and 150 FDA entries (300 total), while Section 4.2 and the contribution list say the final JSON dataset comprises over 3,900 unique standard entries. The retrieval and benchmark results in Table 1 depend on the actual corpus used, so the authors must reconcile this discrepancy and state exactly which corpus was used in the experiments.
  2. [§5.1, Table 1, Abstract] The abstract's 'classification accuracy of 73%' corresponds to Sample-Level Accuracy (0.73), not to Applicability Classification Accuracy (0.71), which is the metric that directly measures whether the applicability label is correct. Moreover, Top-k Retrieval Recall is defined as 'whether at least one gold standard appears in the top-k results,' so the reported 'Top-5 recall of 87%' is a per-device binary hit rate, not the fraction of gold standards retrieved. Please report standard-level recall and use the applicability accuracy as the primary classification metric, or justify why the sample-level metric is the appropriate headline.
  3. [§5.1, Table 1] The denominator of the Applicability Classification Accuracy is not specified because the LLM only classifies standards that were retrieved in the top-k set. Section 5 does not state whether gold standards that were not retrieved are counted as errors, are excluded from the denominator, or are automatically labeled Not Applicable. These choices materially change the reported 0.71 figure. The authors should define the scoring protocol explicitly and report an end-to-end accuracy over all gold standards for the device.
  4. [§5, benchmark construction] All conclusions rest on the correctness of the 105 manually annotated device-standard mappings, but the paper gives no inter-annotator agreement, no external audit, and no released benchmark.csv despite the contribution list promising a release. Since the gold labels are the ground truth for every metric in Table 1, the authors must provide annotation reliability statistics and make the benchmark available for independent verification.
  5. [§3.3, §4.3, §5.1] The retrieval configuration is described inconsistently: Section 3.3 says the retriever uses bge-m3 embeddings with cross-encoder reranking and a hybrid keyword module, Section 4.3 says the retriever uses paraphrase-MiniLM-L6-v2, and Section 5.1 says evaluation uses paraphrase-multilingual-MiniLM-L12-v2. The numbers in Table 1 cannot be reproduced without knowing which configuration was actually used. Please specify the exact retrieval stack used for the reported results and for the retrieval-only baseline.
minor comments (5)
  1. [§5.1, Table 2] Table 2 contains malformed spacing in the p-values ('1 .30 × 10−12' and '1 .60 × 10−22'), which should be corrected.
  2. [§3.5] Section 3.5 refers to GPT-3.5 and the gpt-4 endpoint in the same paragraph; please use one consistent model name and version.
  3. [§7.1] The discussion in §7.1 cites 'explanation alignment scores' and 'human reviewers reported...' even though §5.1 states that human evaluation was omitted in this version. Please remove these claims or clearly label them as qualitative observations from prior versions.
  4. [§2.1 and §7.4] The paper attributes medical device regulation to the European Medicines Agency; in the EU, medical devices are regulated under the MDR by national competent authorities and the MDCG, so the EMA reference should be corrected or replaced.
  5. [§7.5] The limitations in §7.5 correctly note that international ISO/IEC standards are not in the corpus; however, the case study in §6 cites ISO 15197:2013 as a retrieved standard, which appears inconsistent with that limitation and should be clarified.

Circularity Check

1 steps flagged · score 6.0 of 10

Evaluation fallback maps gold labels into predictions, making part of the reported accuracy correct by construction.

  1. self definitional [Section 3.4, LLM-Based Applicability Classification, paragraph on parsed outputs and fallback]
    "Parsed outputs are post-processed to restore region and clause information and to ensure validity. A fallback pseudo-labeling routine is used during evaluation by mapping gold labels in benchmark.csv for consistency tests."

    The evaluation computes Applicability Classification Accuracy and Sample-Level Accuracy by comparing model predictions against the gold labels in benchmark.csv. The quoted sentence states that a fallback pseudo-labeling routine maps those same gold labels into the evaluation pipeline. If the fallback substitutes a gold label for a missing or invalid model output, then for every such entry the 'predicted' label equals the annotated gold label by construction, so that entry is guaranteed to be scored as correct. The paper does not report how often the fallback is invoked or what 'consistency tests' means precisely, but as written this makes at least part of the reported 71% applicability accuracy and 73% sample-level accuracy self-fulfilling rather than measured from the model's own outputs.

full rationale

Most of the paper is a standard RAG evaluation rather than a formal derivation chain, and the bulk of the pipeline is not circular: retrieval and classification use off-the-shelf components (SBERT, FAISS, GPT) that are not fitted to the benchmark labels; the reported metrics compare outputs to expert annotations; and the authors' self-citations are contextual background, not load-bearing evidence for the performance claims. The corpus and benchmark were both built by the authors, and the few-shot examples may be hand-crafted, but these are self-influence concerns rather than definitional circularity. The one concrete circular element is the Section 3.4 fallback that maps gold labels from benchmark.csv during evaluation. Since benchmark.csv is also the source of the gold labels used to score the system, any fallback entry is correct by construction, and the paper does not quantify its usage. Separately, the abstract's '73% classification accuracy' is actually the Sample-Level Accuracy metric, defined as whether at least one correctly labeled match was found per device, which is a lenient reporting choice but not itself circular. The headline metrics should be re-reported with the fallback removed or quantified, and with standard-level rather than sample-level denominators, before the effectiveness claims can be fully interpreted.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim is an empirical system performance claim. It rests on the correctness of the benchmark annotations and the completeness of the corpus, plus the assumption that SBERT similarity and LLM cue-word reasoning capture regulatory applicability. The only hand-chosen numbers that directly affect results are k=5 and temperature 0.3. No new physical or conceptual entities are introduced.

free parameters (4)
  • top_k retrieval cutoff = 5
    Chosen as k=5 for all reported metrics; affects recall and accuracy.
  • LLM temperature = 0.3
    Set to 0.3 to balance output stability; hand-chosen, not fitted to data.
  • synonym expansion dictionary = hand-crafted, not enumerated
    Used in the hybrid keyword module to expand queries (e.g., glucose to sugar, CGM); the full dictionary is not provided.
  • clause conflict similarity threshold = not specified
    The gap analysis computes semantic similarity between justifications, but the threshold for flagging divergence is not stated in the paper.
assumptions (5)
  • domain assumption Sentence-BERT embeddings capture semantic relevance between device descriptions and standard scopes.
    Retrieval and reranking depend on cosine similarity in embedding space throughout Sections 3.2 and 3.3.
  • domain assumption Regulatory cue words such as shall or must indicate mandatory applicability.
    The prompt instructions in Section 3.4 rely on these cues for classification.
  • domain assumption The expert-annotated labels in benchmark.csv are correct ground truth.
    All metrics compare against these labels; no inter-annotator agreement is reported in Section 5.
  • domain assumption GPT-4 outputs are reliable enough for classification after JSON parsing.
    The system uses an external model via API; there is no verification of outputs beyond syntactic parsing and fallback routines.
  • domain assumption The regulatory corpus is complete and correctly parsed.
    Missing or mismatched standards are listed as an error source in Section 5.2, implying completeness is assumed in the main evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Standard Applicability Judgment and Cross-jurisdictional Reasoning: A RAG-based Framework for Medical Device Compliance." pith.science (2026). https://pith.science/paper/U4XPXM6B

@misc{pith2026250618511,
  author       = {Pith},
  title        = {Pith review of: Standard Applicability Judgment and Cross-jurisdictional Reasoning: A RAG-based Framework for Medical Device Compliance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U4XPXM6B}},
  note         = {Machine review of arXiv:2506.18511}
}
read the original abstract

Identifying the appropriate regulatory standard applicability remains a critical yet understudied challenge in medical device compliance, frequently necessitating expert interpretation of fragmented and heterogeneous documentation across different jurisdictions. To address this challenge, we introduce a modular AI system that leverages a retrieval-augmented generation (RAG) pipeline to automate standard applicability determination. Given a free-text device description, our system retrieves candidate standards from a curated corpus and uses large language models to infer jurisdiction-specific applicability, classified as Mandatory, Recommended, or Not Applicable, with traceable justifications. We construct an international benchmark dataset of medical device descriptions with expert-annotated standard mappings, and evaluate our system against retrieval-only, zero-shot, and rule-based baselines. The proposed approach attains a classification accuracy of 73% and a Top-5 retrieval recall of 87%, demonstrating its effectiveness in identifying relevant regulatory standards. We introduce the first end-to-end system for standard applicability reasoning, enabling scalable and interpretable AI-supported regulatory science. Notably, our region-aware RAG agent performs cross-jurisdictional reasoning between Chinese and U.S. standards, supporting conflict resolution and applicability justification across regulatory frameworks.

Figures

Figures reproduced from arXiv: 2506.18511 by the authors.

Figure 1
Figure 1. Perception, Retrieval, Context Construction, Reasoning, Compliance Output, and [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. User-facing interface of the proposed system, demonstrating automated regulatory [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 31 canonical work pages

  1. [1]

    More than red tape: exploring complexity in medical device regulatory affairs,

    Y. Han, A. Ceross, and J. Bergmann, “More than red tape: exploring complexity in medical device regulatory affairs,” Frontiers in Medicine, vol. 11, p. 1415319, 2024

    work page 2024

  2. [2]

    Fda ai use cases in medical device regulation: Current progress and future challenges,

    Q. Li, X. Wang, and H. Chen, “Fda ai use cases in medical device regulation: Current progress and future challenges,” Journal of Regulatory Science and Technology, vol. 3, no. 2, pp. 45–58, 2021

    work page 2021

  3. [3]

    Lexglue: A benchmark dataset for legal language understanding in english,

    I. Chalkidis, A. Jana, D. Hartung, M. Bommarito, I. Androutsopoulos, D. M. Katz, and N. Aletras, “Lexglue: A benchmark dataset for legal language understanding in english,” arXiv preprint arXiv:2110.00976 , 2021

    arXiv 2021

  4. [4]

    Ema regulatory guidance for ai in medical devices,

    E. M. Agency, “Ema regulatory guidance for ai in medical devices,” 2024, available at https://www.ema.europa.eu/en

    work page 2024

  5. [5]

    Regulatory frameworks for ai-enabled medical de- vice software in china: Comparative analysis and review of implications for global manu- facturer,

    Y. Han, A. Ceross, and J. Bergmann, “Regulatory frameworks for ai-enabled medical de- vice software in china: Comparative analysis and review of implications for global manu- facturer,” JMIR AI , vol. 3, p. e46871, 2024. 22

    work page 2024

  6. [6]

    Transforming Medical Regulations into Numbers: Vectorizing a Decade of Medical Device Regulatory Shifts in the USA, EU, and China

    Y. Han and J. Bergmann, “Transforming medical regulations into numbers: Vectorizing a decade of medical device regulatory shifts in the usa, eu, and china,” arXiv preprint arXiv:2411.00567, 2024

    work page Pith review arXiv 2024

  7. [7]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein et al. , “On the opportunities and risks of foundation models,” arXiv preprint arXiv:2108.07258, 2022

    arXiv 2022

  8. [8]

    The use of chatbots as supportive agents in healthcare,

    L. Ogilvie and J. Prescott, “The use of chatbots as supportive agents in healthcare,” Eu- ropean Addiction Research, vol. 28, no. 6, pp. 405–418, 2022

    work page 2022

Show all 40 references
  1. [9]

    Calculating quality management costs in regulatory compliance,

    M. Crudeli, “Calculating quality management costs in regulatory compliance,” Technology Record, 2020

    2020

  2. [10]

    Ai in regulatory compliance: Applications and challenges,

    Y. Zhang et al. , “Ai in regulatory compliance: Applications and challenges,” Journal of Regulatory Science, vol. 10, no. 2, pp. 45–58, 2022

    2022

  3. [11]

    Chemcrow: Augmenting large-language models with chemistry tools,

    A. Bran and A. White, “Chemcrow: Augmenting large-language models with chemistry tools,” arXiv preprint arXiv:2304.05376 , 2023

    2023 arXiv

  4. [12]

    Open-source ai chatbots are booming,

    D. Castelvecchi, “Open-source ai chatbots are booming,” Nature, 2023

    2023

  5. [13]

    Query understanding in the age of large language models,

    A. Anand and V. Setty, “Query understanding in the age of large language models,” arXiv preprint arXiv:2306.16004, 2023

    2023 arXiv

  6. [14]

    Retrieval-augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, A. Fan, V. Chaudhary, H. Schwenk, F. Guzm´ anet al. , “Retrieval-augmented generation for knowledge-intensive nlp tasks,” Advances in Neural Information Processing Systems , vol. 33, pp. 9459–9474, 2020

    2020

  7. [15]

    Grove: Retrieval-augmented complex story generation framework,

    Z. Wen, Z. Tian, W. Wu et al. , “Grove: Retrieval-augmented complex story generation framework,” arXiv preprint arXiv:2310.05388 , 2023

    2023 arXiv

  8. [16]

    Retrieve anything to augment large language models,

    P. Zhang, S. Xiao, Z. Liu et al. , “Retrieve anything to augment large language models,” arXiv preprint arXiv:2310.07554 , 2023

    2023

  9. [17]

    Lora: Low-rank adaptation of large language models,

    E. Hu, Y. Shen, P. Wallis et al. , “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685 , 2021

    2021 arXiv

  10. [18]

    Rogue scores: Evaluation challenges for generative models,

    M. Grusky, “Rogue scores: Evaluation challenges for generative models,” ACL, 2023

    2023

  11. [19]

    Leveraging passage retrieval with generative models for open domain question answering,

    G. Izacard and E. Grave, “Leveraging passage retrieval with generative models for open domain question answering,” in Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics , 2021, pp. 874–880

    2021

  12. [20]

    Deep medicine: How artificial intelligence can make healthcare human again,

    E. Topol, “Deep medicine: How artificial intelligence can make healthcare human again,” Nature Medicine, vol. 25, pp. 44–46, 2019

    2019

  13. [21]

    Ai-driven compliance for medical devices,

    E. IP and R&D, “Ai-driven compliance for medical devices,” https://iprd.evalueserve.com/ blog/ai-driven-compliance-for-medical-devices/, 2024. 23

    2024

  14. [22]

    Artificial intelligence and machine learn- ing in software,

    U.S. Food and Drug Administration, “Artificial intelligence and machine learn- ing in software,” https://www.fda.gov/medical-devices/software-medical-device-samd/ artificial-intelligence-and-machine-learning-software-medical-device, 2025

    2025

  15. [23]

    Jec-qa: A legal-domain question answering dataset,

    H. Zhong et al., “Jec-qa: A legal-domain question answering dataset,” in Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , 2020, pp. 4318– 4331

    2020

  16. [24]

    How nlp improves clinical evaluation reports for medical devices,

    CapeStart, “How nlp improves clinical evaluation reports for medical devices,” https://capestart.com/resources/blog/ how-nlp-improves-clinical-evaluation-reports-for-medical-devices/, 2021

    2021

  17. [25]

    Leveraging natural language processing to aggregate field safety notices,

    Nature, “Leveraging natural language processing to aggregate field safety notices,” npj Digital Medicine, vol. 7, p. Article 137, 2024

    2024

  18. [26]

    Intro to retrieval-augmented generation (rag) in legal tech,

    T. Reuters, “Intro to retrieval-augmented generation (rag) in legal tech,” https://legal. thomsonreuters.com/blog/retrieval-augmented-generation-in-legal-tech/, 2024

    2024

  19. [27]

    Reduce ai hallucinations with this neat software trick,

    Wired, “Reduce ai hallucinations with this neat software trick,” https://www.wired.com/ story/reduce-ai-hallucinations-with-rag, 2024

    2024

  20. [28]

    Autonomous ai agents: Leveraging llms for adaptive decision making,

    IEEE Computer Society, “Autonomous ai agents: Leveraging llms for adaptive decision making,” https://www.computer.org/publications/tech-news/community-voices/ autonomous-ai-agents, 2025

    2025

  21. [29]

    Guiding principles on the use of large language models in regulatory science,

    European Medicines Agency, “Guiding principles on the use of large language models in regulatory science,” https://www.ema.europa.eu/en/documents/other/guiding-principles, 2024

    2024

  22. [30]

    Sentence-bert: Sentence embeddings using siamese bert- networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert- networks,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Lan- guage Processing, 2019, pp. 3982–3992

    2019

  23. [31]

    Billion-scale similarity search with gpus,

    J. Johnson, M. Douze, and H. J´ egou, “Billion-scale similarity search with gpus,” IEEE Transactions on Big Data , vol. 7, no. 3, pp. 535–547, 2019

    2019

  24. [32]

    Gpt-4 technical report,

    OpenAI, “Gpt-4 technical report,” OpenAI Technical Report, 2023, https://openai.com/ research/gpt-4

    2023

  25. [33]

    C. National Medical Products Administration (NMPA), YY 1234-2023: General technical requirements for vacuum blood collection tubes, Chinese National Medical Device Standard- ization Administration Std., 2023, in Chinese. Specifies sterility, additive use, sealing, and biocompa...

    2023

  26. [34]

    Defines chemical composition and validation protocols for anticoagulant additives used in vacuum tubes

    ——, YY/T 0612-2022: Requirements for additives used in vacuum blood collection systems, Chinese National Medical Device Standardization Administration Std., 2022, in Chinese. Defines chemical composition and validation protocols for anticoagulant additives used in vacuum tubes

    2022

  27. [35]

    Aligned with ISO 10993-1

    ——, YY/T 0314-2021: Biological evaluation of medical devices — Part 1: Evaluation and testing within a risk management process, Chinese National Medical Device Standardization 24 Administration Std., 2021, in Chinese. Aligned with ISO 10993-1. Provides guidance on biocompatibi...

    2021

  28. [36]

    21 cfr 862.1345 - glucose test system,

    U. Food and D. A. (FDA), “21 cfr 862.1345 - glucose test system,” https://www.ecfr.gov/ current/title-21/chapter-I/subchapter-H/part-862/section-862.1345, 2024, defines regula- tory classification and performance requirements for glucose test systems. Not specific to collection tubes

    2024

  29. [37]

    I. O. for Standardization (ISO), ISO 15197:2013 - In vitro diagnostic test systems — Re- quirements for blood-glucose monitoring systems for self-testing in managing diabetes mel- litus, ISO Std., 2013, applicable to glucose meters. Specifies clinical evaluation protocols, sam...

    2013

  30. [38]

    The use of readability metrics in legal text: A systematic literature review,

    Y. Han, A. Ceross, and J. H. Bergmann, “The use of readability metrics in legal text: A systematic literature review,” arXiv preprint arXiv:2411.09497 , 2024

    2024 arXiv

  31. [39]

    Regulator-manufacturer ai agents modeling: Mathematical feedback- driven multi-agent llm framework,

    Y. Han and Z. Guo, “Regulator-manufacturer ai agents modeling: Mathematical feedback- driven multi-agent llm framework,” arXiv preprint arXiv:2411.15356 , 2024

    2024 arXiv

  32. [40]

    Explainable legal judgment prediction via learning to rank,

    R. Jha, N. Chhaya, K. Sankaranarayanan, S. Ghosh, V. Madhavan, and A. Srivastava, “Explainable legal judgment prediction via learning to rank,” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) . Association for Computational Ling...

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.