Pith. sign in

REVIEW 4 major objections 7 minor 65 references

AI for Regulatory Affairs: Balancing Accuracy, Interpretability, and Computational Cost in Medical Device Classification

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper benchmarks traditional machine learning, deep learning, and large language models for medical device risk classification and finds that no single model optimizes accuracy, interpretability, and computational cost simultaneously.

desk verdict A useful but shaky benchmark: the empirical ranking rests on an unvalidated label-inference step, so the headline accuracy numbers are not yet trustworthy. read the letter →

arxiv 2505.18695 v1 pith:B3WJ4ZQR submitted 2025-05-24 cs.AI

classification cs.AI
keywords medicaldeviceclassificationregulatoryaffairsNMPAUDIdatabasemodelbenchmarkinginterpretabilitycomputationalcostXGBoostlargelanguagemodels
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

Regulatory classification is the first gate a medical device must pass, and it decides how much scrutiny, evidence, and clinical data a product needs; getting it wrong can delay market entry or expose patients to under-regulated devices. The paper's central finding is empirical: on a 42,000-sample registry of Chinese device descriptions, no AI model simultaneously maximizes accuracy, interpretability, and speed. A CNN reaches the highest accuracy (88.28%), XGBoost offers 86% accuracy with low latency, logistic regression is fastest and most transparent, and zero-shot LLMs trail badly on accuracy (about 46%) while producing natural-language rationales. The authors conclude that the right model depends on the use case, and they propose context-dependent hybrid workflows rather than a single best algorithm. That matters because regulators are already piloting AI review tools and need evidence about where each model class can safely be used.

What carries the argument

The load-bearing apparatus is a controlled benchmark pipeline. All classical and deep models consume the same Chinese BERT whole-word-mask embeddings, a pretrained masked-language model that represents each device description as contextual vectors; the [CLS] vector is used for SVM, logistic regression, random forest, and XGBoost, while full token sequences are used for CNN, RCNN, DPCNN, and RNN, so architecture differences rather than representation drive the accuracy gap. A 10-fold stratified cross-validation on the 42,000-sample training set fixes the evaluation protocol, and a separate 1,000-sample stratified subset is used for zero-shot LLM prompting. Interpretability is assessed per model type: coefficients and class-conditional log probabilities for intrinsic models, TreeSHAP (exact Shapley attribution for tree ensembles) for XGBoost and random forest, LIME and Anchors for SVM and logistic regression, and Integrated Gradients for deep networks, with the paper noting that explanation quality has no ground-truth benchmark. The per-class F1 breakdown is what turns the headline accuracy numbers into regulatory risk statements.

What would settle it

Take a random sample of records whose risk class was inferred from the zczbhhzbapzbh identifier field and compare those parsed labels with official NMPA classification records or a manual regulatory review; if the inferred-label error rate is large, or if the accuracy ranking of models changes when only independently verified labels are used, the central trade-off claim would need to be re-quantified.

Watch

Extended reading notes

Core claim

The paper's central claim is that for the downstream task of NMPA risk classification (Class I, II, or III), the achievable frontier is a trade-off: accuracy, interpretability, and computational cost cannot be maximized jointly. CNN achieves the best measured accuracy (88.28% and macro F1 0.6768), XGBoost gives the best classical-model balance (86.00%, macro F1 0.69) at 0.0514 seconds per sample, and logistic regression offers the best speed-transparency combination (82.00%, 0.0208 seconds). Zero-shot LLMs (DeepSeek R1-Distill-Qwen-7B and LLaMA 3.1 8B) reach only about 46% accuracy but generate explanations in natural language; the paper interprets this as plausible but not necessarily legally correct reasoning. It also emphasizes per-class metrics, where all models do poorly on Class I devices (for example, logistic regression Class I F1 = 0.20 and RNN 0.14) while Class III recall is comparatively strong, so aggregate accuracy hides asymmetric risks. The conclusion is not that one model should be deployed, but that selection should follow the stage of the workflow: fast models for triage, accurate models for initial classification, interpretable models for ambiguous and high-risk cases, and LLMs only in supervised auxiliary roles.

Load-bearing premise

The ground-truth labels in the benchmark are partly harvested by character-parsing the NMPA unique identifier field when official risk labels are missing, and those parsed labels are never checked against an independent official classification, so any systematic parsing error would shift every accuracy and F1 score reported.

Editorial extensions

If this is right

  • A hybrid workflow follows directly: fast, interpretable models for batch pre-screening, CNN or XGBoost for automatic classification, logistic regression plus SHAP for ambiguous cases, CNN plus TreeSHAP with human review for high-risk Class III devices, and LLMs used only to generate rationales for novel devices.
  • Regulatory agencies can choose a default model by priority: XGBoost for a balanced accuracy-speed profile, logistic regression for audits and low latency, and CNN when maximum accuracy is worth the modestly greater compute.
  • Per-class F1, not aggregate accuracy, should drive acceptance decisions, because all models struggle on Class I devices and asymmetric errors create unequal safety and regulatory-burden consequences.
  • LLM-based assistants in regulatory pipelines should be constrained to explanation and review support, not autonomous classification, since their rationales can be fluent while legally wrong.

Reading between the lines

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

  • Editorial extension: the character-parsed NMPA labels should be independently verified against official classification sources; if parsing noise is systematic, the accuracy rankings in Tables 2-4 could shift, especially for Class I where every model already performs poorly.
  • Editorial extension: the same controlled benchmark protocol could be run on English-language FDA 510(k) or EU MDR device descriptions to test whether the accuracy-interpretability-cost ordering is language- and jurisdiction-specific.
  • Editorial extension: a retrieval-augmented or fine-tuned LLM variant could close the gap between the 46% zero-shot accuracy and the 86-88% supervised models; the paper does not test this, but its hybrid workflow would be the natural place to insert such a variant.
  • Editorial extension: the proposed explanation-utility dimensions (comprehensibility, justifiability, trust alignment) could be turned into a scoring rubric and validated against expert raters, giving interpretability a quantifiable axis alongside accuracy and cost.
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

4 major / 7 minor

Summary. The paper benchmarks a range of AI models—traditional ML, deep learning, and LLMs—on the task of classifying Chinese medical device descriptions from the NMPA UDI registry into three regulatory risk classes. It evaluates models on accuracy (with per-class F1), interpretability (via various explanation tools), and computational cost (inference time), and concludes that no single model dominates all criteria, recommending context-dependent hybrid strategies. A qualitative 'validation' is attempted via a GPT-4-based virtual focus group of expert personas.

Significance. If the empirical results are trustworthy, the paper would provide a useful, practical benchmark for a real regulatory task, with the rare virtue of reporting per-class F1 and making code and data available. The comparison of three model families on the same regulatory dataset is a legitimate contribution, and the proposed decision-tree workflow (Figure 7) and explanation-utility framework are reasonable starting points. However, the central claim—the accuracy ranking CNN > XGBoost > RF/SVM—rests entirely on ground-truth labels whose reliability is not established, and several reported numbers are internally inconsistent, so the headline results are not yet sufficiently secure for a definite accept.

major comments (4)
  1. [Section 3.1, Tables 2–4] The ground-truth labels used to compute every accuracy and F1 score are not reliable as reported. The text states that official NMPA risk labels are used, but 'in cases where labels were missing, we inferred the class from the unique identifier field (zczbhhzbapzbh) via character-level parsing.' No fraction of affected records, no validation of the parsing, and no agreement rate or confusion matrix against official classifications are provided. If the inferred labels are noisy or systematically biased by identifier-format quirks, then every metric in Tables 2–4 is biased, and the relative ranking CNN > XGBoost > Random Forest > SVM could be an artifact. Because all models are trained and evaluated on the same potentially noisy labels, the robustness of the ranking is not demonstrated. Please quantify the fraction of inferred labels and validate the parsing on a random sample against independent official classifications; without this, the central empirical claim is not supported by the evidence in the manuscript.
  2. [Section 3.4 vs. Table 4; Table 3 vs. Table 6; Table 2 vs. Section 6.2] Several load-bearing numbers are internally inconsistent. Section 3.4 says the LLM evaluation uses a '10-sample subset,' while Section 3.1 and Table 4 state 1,000 samples. Table 3 reports CNN inference time as 0.0034 s, but Table 6 reports 0.0024 s for the same model. Table 2 reports XGBoost accuracy as 86.00%, while Section 6.2 quotes 83.95% for XGBoost. These are not presentation-level typos only; they affect the exact figures that support the trade-off narrative. Please reconcile the numbers and specify which values correspond to which experimental run, model configuration, and test set.
  3. [Section 5] The paper describes a 'virtual focus group' of GPT-4 agents as a way to 'validate' the experiment results. This is not an independent validation: the agents are the same family of LLMs as one of the systems under evaluation, and they are prompted to act as experts rather than evaluated as a gold-standard human panel. The thematic findings in Section 6.2 therefore cannot provide independent confirmation of the quantitative results, and the manuscript's claim that these themes 'directly validated or challenged specific results' overstates what such a simulation can establish. Please reframe this section as an exploratory illustration of possible stakeholder perspectives, or replace it with a genuinely independent human-expert evaluation if validation is intended.
  4. [Section 3.2 and Section 3.4] The description of model inputs is contradictory. Section 3.2 says deep models (BERT-LSTM, TextCNN) use the full sequence of token embeddings, while Section 3.4 says 'All models use BERT [CLS] embeddings as input' and that deep models take input shaped (n, 768, 1). Table 1 further states that SVM and LLMs use 'raw text' as input, which is inconsistent with the stated use of BERT [CLS] embeddings for all models. This ambiguity makes it unclear what was actually trained and evaluated, and it impedes reproducibility. Please clarify the exact input representation used by each model family and adjust the methodology text accordingly.
minor comments (7)
  1. [Section 3.1] The class distribution of the 42,000-sample dataset is not reported. Since accuracy and F1 are sensitive to class imbalance, and the paper's discussion emphasizes asymmetric misclassification costs, the proportion of Class I/II/III devices should be stated in the methodology or results.
  2. [Section 3.3] The train/validation/test split is described inconsistently: the text mentions '80/10/10 or 70/15/15' splits, then says models are trained with a '20% validation split' and evaluated with 10-fold cross-validation. Please specify the actual split protocol used for each model.
  3. [Table 1 (star-rating table), Section 6.1] The star ratings for accuracy, interpretability, and computational cost in the comparison table are assigned without a transparent derivation from the experimental results. The criteria in Table 7 are defined, but the mapping from measured values (e.g., accuracy 0.86 for XGBoost) to stars (5) is not justified, and Figure 5 uses these normalized stars as if they were data. Please make the rating procedure explicit or present the raw metrics instead.
  4. [Table 1 and Table 7 (captions)] There are two different tables labeled 'Table 1' (the interpretability-strategy table and the star-rating comparison), and the star-rating table caption repeats 'T able 1.' Renumber the tables to be unique and fix the typo.
  5. [Section 2.4, Section 3.4] The LLM name is inconsistent: Section 2.4 mentions 'DeepSeek and LLaMA,' Section 3.4 says 'DeepSeek-R1-Distill-Qwen-7B,' and Table 4 lists 'DeepSeek R1 7B.' Please use a single consistent model identifier and specify the exact model version and inference framework used.
  6. [Reference [58]] Reference [58] is malformed (placeholder text, incomplete citation). It should be corrected or removed, as it is cited in support of the LLM focus-group methodology.
  7. [Section 4.2, Figure 4] The claim that 'LLMs exhibited the greatest interpretability' is the subjective interpretation of the authors, not a measured result; the paper later acknowledges there is no interpretability benchmark. Please soften this wording to describe the qualitative nature of LLM explanations without implying a comparative advantage over other models.

Circularity Check

1 steps flagged · score 2.0 of 10

Qualitative 'expert validation' is circular—GPT-4 agents are fed the paper's own experimental results before confirming them—but the quantitative benchmark itself is self-contained.

  1. other [Section 5 ('Expert Validation') and Section 6.1 ('Model Tradeoff')]
    "To validate the experiment results identified in this study, we convened a virtual focus group composed of large language model agents in expert roles. ... We showed each agent our experiment numerical results about each model accuracy, Interpretability and cost. ... these themes ... directly validated or challenged specific results from our empirical evaluations."

    The validation is not independent: the 'expert' agents are generative LLMs (GPT-4), and they are explicitly given the paper's own accuracy, interpretability, and cost numbers before responding. Their thematic endorsements (for example, 'High Accuracy but Limited Interpretability') are therefore restatements of the input tables rather than external confirmation. Calling this a 'validation' of the experimental results reduces the confirmation to the results themselves: the focus group output is constructed from the very figures it is said to validate. The central empirical benchmark in Tables 2–4 does not depend on this step, so the circularity is confined to the qualitative validation layer.

full rationale

The paper's quantitative contribution is a benchmark, not a derivation: models are trained on real NMPA device descriptions with BERT embeddings and evaluated with standard accuracy, F1, and inference-time metrics, so the rankings in Tables 2–4 have independent empirical content. No load-bearing self-citation chain is present; citations to the authors' earlier work ([5], [6], [12], [25]) are contextual related work, not premises of the results. The Section 3.1 label-inference step ('in cases where labels were missing, we inferred the class from the unique identifier field via character-level parsing') is a data-quality and validity concern rather than a circularity, because the labels do not derive from the models under test and are not defined in terms of the predictions. The one genuinely circular element is Section 5's 'expert validation': GPT-4 agents are fed the paper's own numerical results, and their thematic agreement is then presented as cross-validation. That step reduces to its inputs. Because the central empirical ranking survives without that qualitative endorsement, the overall circularity score is low.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central benchmark depends on data quality assumptions (label inference, embedding adequacy) and on the paper's own subjective star ratings for interpretability and cost. No new physical or mathematical entities are introduced.

free parameters (2)
  • Star ratings for interpretability and computational cost = e.g., Rule-Based 5 stars, CNN interpretability 3 stars, DeepSeek cost 5 stars
    Table 7 assigns 1-5 stars to each model based on subjective judgment; these are hand-set values, not measured, and feed Figure 5 and trade-off conclusions.
  • Hyperparameters = SVM C in {0.1,1,10}, XGB depth {4,6,8}, learning rate {0.05,0.1}, CNN 128 filters, LSTM 128 units
    Tuned via grid search/validation; standard practice, not ad hoc to the central claim.
assumptions (4)
  • domain assumption Inferred NMPA labels from identifier parsing are accurate
    Section 3.1 derives some ground-truth labels from the zczbhhzbapzbh field; any parsing errors bias all metrics.
  • domain assumption Pretrained BERT embeddings capture the regulatory semantics of Chinese device descriptions without fine-tuning
    All models use hfl/chinese-bert-wwm-ext embeddings; no fine-tuning is reported, so representational adequacy is assumed.
  • ad hoc to paper LLM agents in a GPT-4 focus group can simulate human expert regulatory judgment
    Section 5 treats LLM-generated thematic opinions as expert validation, an assumption unique to this paper's validation design.
  • standard math 10-fold stratified cross-validation with 80/20 split yields unbiased performance estimates
    Standard practice, but class imbalance is not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI for Regulatory Affairs: Balancing Accuracy, Interpretability, and Computational Cost in Medical Device Classification." pith.science (2026). https://pith.science/paper/B3WJ4ZQR

@misc{pith2026250518695,
  author       = {Pith},
  title        = {Pith review of: AI for Regulatory Affairs: Balancing Accuracy, Interpretability, and Computational Cost in Medical Device Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B3WJ4ZQR}},
  note         = {Machine review of arXiv:2505.18695}
}
read the original abstract

Regulatory affairs, which sits at the intersection of medicine and law, can benefit significantly from AI-enabled automation. Classification task is the initial step in which manufacturers position their products to regulatory authorities, and it plays a critical role in determining market access, regulatory scrutiny, and ultimately, patient safety. In this study, we investigate a broad range of AI models -- including traditional machine learning (ML) algorithms, deep learning architectures, and large language models -- using a regulatory dataset of medical device descriptions. We evaluate each model along three key dimensions: accuracy, interpretability, and computational cost.

Figures

Figures reproduced from arXiv: 2505.18695 by the authors.

Figure 1
Figure 1. Text process, Model Training and Evaluation Pipeline [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Comparative interpretability visualizations across models: (a) SVM with LIME, (b) [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. TreeSHAP-based explanations for BERT embeddings using ensemble models. [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: LLM interpretability While interpretability is often cited as a key requirement in medical regulatory AI systems, there exists no universally accepted benchmark or ground truth to evaluate the quality of ex￾planations. We benchmarked SHAP, LIME, Anchors, and Integrated…
Figure 5
Figure 5. Figure 5: Comparison of Models Based on Accuracy, Interpretability, and Cost [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Theme Analysis of Models 25 [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: AI Algorithm Decision Tree in Different Regulatory Affairs Scenarios [PITH_FULL_IMAGE:figures/full_fig_p029_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 53 canonical work pages

  1. [1]

    Fda announces completion of first ai- assisted scientific review pilot and aggressive agency-wide ai rollout,

    U.S. Food and Drug Administration, “Fda announces completion of first ai- assisted scientific review pilot and aggressive agency-wide ai rollout,” May 2025, ac- cessed: 2025-05-10. [Online]. Available: https://www.fda.gov/news-events/press-announcements/ fda-announces-completion-first-ai-assisted-scientific-review-pilot-and-aggressive-agency-wide-ai

  2. [2]

    Openai and the fda are talking about ai for drug evaluation,

    W. D. Heaven, “Openai and the fda are talking about ai for drug evaluation,” WIRED, May 2025, accessed: 2025-05-10. [Online]. Available: https://www.wired.com/story/openai-fda-doge-ai-drug-evaluation/

  3. [3]

    Mantus and D

    D. Mantus and D. J. Pisano, FDA regulatory affairs. CRC Press, 2014

  4. [4]

    Current regulatory requirements for registration of medicines, compilation and submission of dossier in australian therapeutic goods administration,

    B. S. Kumari, G. S. Hanuja, M. Nagabhushanam, D. N. Reddy, and B. Bonthagarala, “Current regulatory requirements for registration of medicines, compilation and submission of dossier in australian therapeutic goods administration,” International Journal of Advanced Scientific and Technical Research, ISSN , pp. 2249–9954, 2016

  5. [5]

    Regulatory frameworks for ai-enabled medical device software in china: Com- parative analysis and review of implications for global manufacturer,

    Y. Han, A. Ceross, and J. Bergmann, “Regulatory frameworks for ai-enabled medical device software in china: Com- parative analysis and review of implications for global manufacturer,” JMIR AI , vol. 3, p. e46871, 2024

  6. [6]

    Uncovering Regulatory Affairs Complexity in Medical Products: A Qualitative Assessment Utilizing Open Coding and Natural Language Processing (NLP)

    Y. Han, A. Ceross, and J. H. Bergmann, “Uncovering regulatory affairs complexity in medical products: A qualitative assessment utilizing open coding and natural language processing (nlp),” arXiv preprint arXiv:2401.02975 , 2023

  7. [7]

    Fda-cleared artificial intelligence and machine learning-based medical devices and their 510 (k) predicate networks,

    U. J. Muehlematter, C. Bluethgen, and K. N. Vokinger, “Fda-cleared artificial intelligence and machine learning-based medical devices and their 510 (k) predicate networks,” The Lancet Digital Health , vol. 5, no. 9, pp. e618–e626, 2023

  8. [8]

    Medical devices: definition, classification, and regulatory implications,

    J. K. Aronson, C. Heneghan, and R. E. Ferner, “Medical devices: definition, classification, and regulatory implications,” Drug safety , vol. 43, no. 2, pp. 83–93, 2020

Show all 65 references
  1. [9]

    A comprehensive strategy to overhaul fda authority for misleading food labels,

    J. L. Pomeranz, “A comprehensive strategy to overhaul fda authority for misleading food labels,” American journal of law & medicine , vol. 39, no. 4, pp. 617–647, 2013

  2. [10]

    Position statement: the need for eu legislation to require disclosure and labelling of the composition of medical devices,

    A. Herman, W. Uter, T. Rustemeyer, M. Matura, K. Aalto-Korte, J. Duus Johansen, M. Gon¸ calo, I. R. White, A. Balato, A. M. Gim´ enez Arnau et al. , “Position statement: the need for eu legislation to require disclosure and labelling of the composition of medical devices,” Jou...

  3. [11]

    The complexity of medical device regulations has increased, as assessed through data-driven techniques,

    A. Arnould, R. Hendricusdottir, and J. Bergmann, “The complexity of medical device regulations has increased, as assessed through data-driven techniques,” Prosthesis, vol. 3, no. 4, pp. 314–330, 2021

  4. [12]

    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. 30

  5. [13]

    Role of artificial intelligence applications in real-life clinical practice: systematic review,

    J. Yin, K. Y. Ngiam, and H. H. Teo, “Role of artificial intelligence applications in real-life clinical practice: systematic review,” Journal of medical Internet research , vol. 23, no. 4, p. e25759, 2021

  6. [14]

    Ai applications to medical images: From machine learning to deep learning,

    I. Castiglioni, L. Rundo, M. Codari, G. Di Leo, C. Salvatore, M. Interlenghi, F. Gallivanone, A. Cozzi, N. C. D’Amico, and F. Sardanelli, “Ai applications to medical images: From machine learning to deep learning,” Physica medica, vol. 83, pp. 9–24, 2021

  7. [15]

    A comparison of rule-based and machine learning methods for medical information extraction,

    O. Imaichi, T. Yanase, and Y. Niwa, “A comparison of rule-based and machine learning methods for medical information extraction,” in The first workshop on natural language processing for medical and healthcare fields , 2013, pp. 38–42

  8. [16]

    Artificial intelligence in pharmaceutical regulatory affairs,

    R. S. Patil, S. B. Kulkarni, and V. L. Gaikwad, “Artificial intelligence in pharmaceutical regulatory affairs,” Drug Discovery Today, p. 103700, 2023

  9. [17]

    Post-market surveillance of medical devices using ai,

    T. Khinvasara, N. Tzenios, and A. Shanker, “Post-market surveillance of medical devices using ai,” Journal of Com- plementary and Alternative Medical Research , vol. 25, no. 7, pp. 108–122, 2024

  10. [18]

    Automatic induction of rule based text categorization,

    D. M. Kumar, “Automatic induction of rule based text categorization,” International Journal of Computer Science & Information Technology (IJCSIT) , vol. 2, no. 6, 2010

  11. [19]

    A review of machine learning algorithms for text-documents classification,

    A. Khan, B. Baharudin, L. H. Lee, and K. Khan, “A review of machine learning algorithms for text-documents classification,” Journal of advances in information technology , vol. 1, no. 1, pp. 4–20, 2010

  12. [20]

    Rule-based semantic relation extraction in regulatory documents

    A. Korger and J. Baumeister, “Rule-based semantic relation extraction in regulatory documents.” in L WDA, 2021, pp. 26–37

  13. [21]

    Support vector machines for classification,

    M. Awad, R. Khanna, M. Awad, and R. Khanna, “Support vector machines for classification,” Efficient learning machines: Theories, concepts, and applications for engineers and system designers , pp. 39–66, 2015

  14. [22]

    Na ¨ ıve bayes

    G. I. Webb, E. Keogh, and R. Miikkulainen, “Na ¨ ıve bayes.” Encyclopedia of machine learning , vol. 15, no. 1, pp. 713–714, 2010

  15. [23]

    Semantic text classification for supporting automated compliance checking in construction,

    D. M. Salama and N. M. El-Gohary, “Semantic text classification for supporting automated compliance checking in construction,” Journal of Computing in Civil Engineering , vol. 30, no. 1, p. 04014106, 2016

  16. [24]

    Semantic nlp-based information extraction from construction regulatory documents for automated compliance checking,

    J. Zhang and N. M. El-Gohary, “Semantic nlp-based information extraction from construction regulatory documents for automated compliance checking,” Journal of Computing in Civil Engineering , vol. 30, no. 2, p. 04015014, 2016

  17. [25]

    A machine learning approach for medical device classification,

    A. Ceross and J. Bergmann, “A machine learning approach for medical device classification,” in Proceedings of the 14th International Conference on Theory and Practice of Electronic Governance , 2021, pp. 285–291

  18. [26]

    Large language modeling and classical ai methods for the future of healthcare,

    S. Banerjee, P. Dunn, S. Conard, and R. Ng, “Large language modeling and classical ai methods for the future of healthcare,” Journal of Medicine, Surgery, and Public Health , vol. 1, p. 100026, 2023

  19. [27]

    Regulating ai adaptation: An analysis of ai medical device updates,

    K. Wu, E. Wu, K. Rodolfa, D. E. Ho, and J. Zou, “Regulating ai adaptation: An analysis of ai medical device updates,” arXiv preprint arXiv:2407.16900 , 2024. [Online]. Available: https://arxiv.org/abs/2407.16900

  20. [28]

    Towards regulatable ai systems: Technical gaps and policy recommendations,

    X. Shen, H. Brown, J. Tao, M. Strobel, Y. Tong, A. Narayan, H. Soh, and F. Doshi-Velez, “Towards regulatable ai systems: Technical gaps and policy recommendations,” arXiv preprint arXiv:2306.12609 , 2023. [Online]. Available: https://arxiv.org/abs/2306.12609

  21. [29]

    A comprehensive survey on on-device ai models,

    e. a. Wang, “A comprehensive survey on on-device ai models,” arXiv preprint arXiv:2503.06027 , 2025. [Online]. Available: https://arxiv.org/abs/2503.06027

  22. [30]

    Improving support vector machine classifiers by modifying kernel functions,

    S.-i. Amari and S. Wu, “Improving support vector machine classifiers by modifying kernel functions,” Neural networks, vol. 12, no. 6, pp. 783–789, 1999

  23. [31]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pp. 785–794, 2016

  24. [32]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” Advances in Neural Information Processing Systems, vol. 30, 2017

  25. [33]

    Convolutional neural networks for sentence classification,

    Y. Kim, “Convolutional neural networks for sentence classification,” in Proceedings of the 2014 Conference on Em- pirical Methods in Natural Language Processing (EMNLP) . Association for Computational Linguistics, 2014, pp. 1746–1751

  26. [34]

    Comparative study of cnn and rnn for natural language processing,

    W. Yin, K. Kann, M. Yu, and H. Sch¨ utze, “Comparative study of cnn and rnn for natural language processing,”arXiv preprint arXiv:1702.01923, 2017. 31

  27. [35]

    Deep pyramid convolutional neural networks for text categorization,

    R. Johnson and T. Zhang, “Deep pyramid convolutional neural networks for text categorization,” in Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . Association for Computational Linguistics, 2017, pp. 562–570

  28. [36]

    Recurrent convolutional neural networks for text classification,

    S. Lai, L. Xu, K. Liu, and J. Zhao, “Recurrent convolutional neural networks for text classification,” in Proceedings of the 29th AAAI Conference on Artificial Intelligence , 2015, pp. 2267–2273

  29. [37]

    Axiomatic attribution for deep networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” International Conference on Machine Learning (ICML) , pp. 3319–3328, 2017

  30. [38]

    Why should i trust you? explaining the predictions of any classifier,

    M. T. Ribeiro, S. Singh, and C. Guestrin, “Why should i trust you? explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . ACM, 2016, pp. 1135–1144

  31. [39]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , vol. 30, 2017

  32. [40]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

  33. [41]

    Pre-training with whole word masking for chinese bert,

    Y. Cui, W. Che, T. Liu, B. Qin, S. Wang, and G. Hu, “Pre-training with whole word masking for chinese bert,” arXiv preprint arXiv:1906.08101, 2019

  34. [42]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi` ere, N. Goyal, E. Hambro, F. Azhar et al. , “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971 , 2023

  35. [43]

    Deepseek-vl: Scaling vision-language alignment with decoupled pretraining,

    D. AI, “Deepseek-vl: Scaling vision-language alignment with decoupled pretraining,” arXiv preprint arXiv:2401.14185, 2024

  36. [44]

    Towards faithfully interpretable nlp systems: How should we define and evaluate faith- fulness?

    A. Jacovi and Y. Goldberg, “Towards faithfully interpretable nlp systems: How should we define and evaluate faith- fulness?” Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , pp. 4198–4205, 2020

  37. [45]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,

    C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,” Nature Machine Intelligence , vol. 1, no. 5, pp. 206–215, 2019

  38. [46]

    Anchors: High-precision model-agnostic explanations,

    M. T. Ribeiro, S. Singh, and C. Guestrin, “Anchors: High-precision model-agnostic explanations,” Proceedings of the AAAI Conference on Artificial Intelligence , vol. 32, no. 1, 2018

  39. [47]

    Explainable machine-learning predictions for the prevention of hypoxaemia during surgery,

    S. M. Lundberg, B. Nair, M. Vavilala, M. Horibe, M. J. Eisses, T. Adams, D. E. Liston, D. K. Low, S.-F. Newman, J. Kim et al. , “Explainable machine-learning predictions for the prevention of hypoxaemia during surgery,” Nature Biomedical Engineering, vol. 2, no. 10, pp. 749–760, 2018

  40. [48]

    Local explanation methods for tree-based models: A unified approach,

    S. M. Lundberg, G. Erion, H. Chen, A. DeGrave, J. M. Prutkin, B. Nair, R. Katz, J. Himmelfarb, N. Bansal, and S.-I. Lee, “Local explanation methods for tree-based models: A unified approach,” Nature Machine Intelligence , vol. 2, no. 1, pp. 56–67, 2020

  41. [49]

    Energy and policy considerations for deep learning in nlp,

    E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for deep learning in nlp,” arXiv preprint arXiv:1906.02243, 2019

  42. [50]

    A survey on efficient training of transformers,

    Z. Wu, X. Liu, D. F. Zhou, Y. He, and J. Gao, “A survey on efficient training of transformers,” arXiv preprint arXiv:2202.06930, 2022

  43. [51]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,

    V. Sanh, L. Debut, J. Chaumond, and T. Wolf, “Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,” arXiv preprint arXiv:1910.01108 , 2019

  44. [52]

    Towards practical trade-offs between interpretability and performance: A composite interpretability metric for trustworthy ai,

    C. Zhang, Y. Liu, X. Yang, K. Gu, J. Shen, X. Li, and J. Bian, “Towards practical trade-offs between interpretability and performance: A composite interpretability metric for trustworthy ai,” 2024

  45. [53]

    Trust and transparency in human-ai interaction: A survey of trust calibration, user understanding, and interpretability,

    H. Kaur, M. Desai, M. Eslami, and J. Forlizzi, “Trust and transparency in human-ai interaction: A survey of trust calibration, user understanding, and interpretability,” in Proceedings of the 2022 ACM Conference on Fairness, Ac- countability, and Transparency (F AccT), 2022, p...

  46. [54]

    A taxonomy of interpretability in human-centered ai: From explanations to user experience,

    S. Chen, T. Wu, and Q. V. Lee, “A taxonomy of interpretability in human-centered ai: From explanations to user experience,” IEEE Transactions on Visualization and Computer Graphics , 2024

  47. [55]

    Udi database,

    National Medical Products Administration, “Udi database,” https://udi.nmpa.gov.cn/download.html, n.d., accessed: 2024-09-25. 32

  48. [56]

    Split the data between the training data and test data using sklearn,

    Tasos, “Split the data between the training data and test data using sklearn,” 2019, accessed: 2025-05-04. [Online]. Available: https://datascience.stackexchange.com/questions/51236/ split-the-data-between-the-training-data-and-test-data-using-sklearn

  49. [57]

    Scikit-learn: Machine learning in python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and ´E. Duchesnay, “Scikit-learn: Machine learning in python,” Journal of Machine Learn...

  50. [58]

    Focusagent: [article title],

    A. F. N. L. N. Liu, “Focusagent: [article title],” Journal Name , vol. Volume Number, no. Issue Number, p. Page Range, 2024

  51. [59]

    Focus agent: Llm-powered virtual focus group,

    T. Zhang, X. Zhang, R. Cools, and A. Simeone, “Focus agent: Llm-powered virtual focus group,” in Proceedings of the 24th ACM International Conference on Intelligent Virtual Agents , 2024, pp. 1–10

  52. [60]

    A training algorithm for optimal margin classifiers,

    B. E. Boser, I. M. Guyon, and V. N. Vapnik, “A training algorithm for optimal margin classifiers,” Proceedings of the fifth annual workshop on Computational learning theory , pp. 144–152, 1992

  53. [61]

    On the robustness of interpretability methods,

    D. Alvarez-Melis and T. S. Jaakkola, “On the robustness of interpretability methods,” arXiv preprint arXiv:1806.08049, 2018

  54. [62]

    Proposal for a regulation on a european approach for artificial intelligence,

    E. Commission, “Proposal for a regulation on a european approach for artificial intelligence,” 2021, available at: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52021PC0206

  55. [63]

    Good machine learning practice for medical device development: Guid- ing principles,

    U. Food and D. A. (FDA), “Good machine learning practice for medical device development: Guid- ing principles,” 2021, available at: https://www.fda.gov/medical-devices/software-medical-device-samd/ good-machine-learning-practice-medical-device-development-guiding-principles

  56. [64]

    Counterfactual explanations without opening the black box: Automated decisions and the gdpr,

    S. Wachter, B. Mittelstadt, and L. Floridi, “Counterfactual explanations without opening the black box: Automated decisions and the gdpr,” Harvard journal of law & technology , vol. 31, p. 841, 2017

  57. [65]

    Mimic-iii, a freely accessible critical care database,

    A. E. Johnson, T. J. Pollard, L. Shen, L.-w. H. Lehman, M. Feng, M. Ghassemi, B. Moody, P. Szolovits, L. A. Celi, and R. G. Mark, “Mimic-iii, a freely accessible critical care database,” Scientific data , vol. 3, p. 160035, 2016. 33

Pith tools

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