Pith. sign in

REVIEW 3 major objections 6 minor 46 references

Initial Investigation of LLM-Assisted Development of Rule-Based Clinical NLP System

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read LLMs can do the two hardest early steps of rule-based clinical NLP with near-perfect recall.

desk verdict First clinical NLP study to use LLMs in the development loop of a rule-based system; the snippet recall result is credible, but the 1.0 keyword coverage is a circular artifact and the paper's own limitations concede it — conditional, not reject. read the letter →

arxiv 2506.16628 v1 pith:FY3TMW3G submitted 2025-06-19 cs.CL cs.LG

Jianlin Shi , Brian T. Bucher This is my paper · ORCID
classification cs.CLcs.LG
keywords LLM-assisteddevelopmentrule-basedNLPclinicalsurgicalsiteinfectionsurveillancenamedentityrecognitionsnippetidentificationkeywordextractionchain-of-thoughtprompting
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 paper tries to establish that large language models can take over the two most labor-intensive early steps of building a rule-based clinical NLP system: picking out the clinically relevant snippets from patient notes and listing the key terms that become the rules for named entity recognition. The authors test this on EasyCIE, a rule-based surgical site infection surveillance system, using two instruction-tuned open-weight models. They report snippet-identification recall of 0.98 and 0.99, meaning the models found nearly all relevant snippets, and they report that the generated keywords covered 100% of the snippets the existing rules matched. The point of the approach is that the expensive LLM works only during development, while the finished system stays rule-based, transparent, and cheap to run.

What carries the argument

The machinery is a two-pass prompt protocol wrapped around a rule-based pipeline. Each task has a reasoning prompt that embeds the NSQIP SSI manual as markdown and asks the model to walk through the relevant subtasks (signs and symptoms, treatment, anatomic sites, procedures, infection, wound care), then a verification prompt that embeds the annotation guideline, asks the model to critique the first pass, and requires a JSON output. Snippet identification is run five times and decided by majority vote; keyword extraction asks for both verbatim concepts and expanded synonyms, then prunes certainty, temporality, and experiencer modifiers so downstream rules can handle them. EasyCIE, the target system, is a modular rule-based NLP pipeline whose NER component matches keywords against note text, and the two LLM steps feed exactly that component.

What would settle it

Run the keyword-extraction prompt on a random sample of snippets from the annotated corpus that the existing EasyCIE rules do not match, then measure how many extracted keywords correspond to annotated concepts and how many would cause false matches in a rule-based NER pass. If coverage drops well below 100% or precision approaches zero, the keyword-extraction claim is an artifact of the evaluation design rather than a property of the method.

Watch

Extended reading notes

Core claim

The paper's central claim is that the first two development steps of a rule-based clinical NLP pipeline can be delegated to an LLM without sacrificing the downstream system's rule-based character. Working from the EasyCIE surgical site infection pipeline, the authors cast snippet identification as binary classification: given a sentence-level snippet from a postoperative note, decide whether it carries any cue relevant to an SSI diagnosis. On the held-out test set, Deepseek R1 distilled Qwen 32B reached recall 0.98 and Qwen2.5 Coder 32B reached 0.99, with precision 0.10 and 0.08 respectively. A manual review of 100 errors per model found that most apparent false positives were snippets an annotator had skipped for contextual reasons rather than true noise. For keyword extraction, both models produced keyword sets that covered 100% of the snippets that the existing EasyCIE NER rules matched, and the authors show examples where generated keywords were broadened into more generalizable rule candidates.

Load-bearing premise

The keyword result rests on the assumption that matching the snippets the old rules already catch is a good test of whether the generated keywords would support a new rule-based component; if that test is too easy, the 100% coverage figure does not carry the weight the paper gives it.

Editorial extensions

If this is right

  • If the two hypotheses hold, the dominant labor cost of rule-based NLP development—reading notes and drafting rules—can be shifted to an LLM, with clinicians editing candidates instead of starting from scratch.
  • Because the LLM is used only at build time, the deployed surveillance system retains the speed, low cost, and interpretability that make rule-based NLP attractive in clinical settings.
  • High recall in snippet identification means the downstream NER and context components receive nearly all relevant evidence; the low precision may be tolerable if later stages filter false positives.
  • The keyword sets, including expanded synonyms, give developers a concrete starting point for NER rules and can make the rules more generalizable than the original hand-built rules.
  • The prompt design, which embeds the manual and guideline directly, points to a route for semi-automated maintenance: when guidelines change, prompts can be regenerated rather than re-engineering rules by hand.

Reading between the lines

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

  • A fairer test of the keyword claim would measure coverage and precision on snippets sampled independently of the existing EasyCIE rules; the reported 100% coverage is measured against snippets the old rules themselves produced, so it is an upper bound on how well the keywords would serve a fresh rule set.
  • If the development-phase-only pattern generalizes, the same workflow could apply in other regulated text-processing domains where explainability is mandatory, such as diagnostic coding, prior authorization, or adverse-event reporting.
  • An end-to-end comparison—building NER rules from the generated keywords, adding the downstream context detector, and comparing final SSI classification against EasyCIE—would settle whether the 100% keyword coverage actually translates into equal or better system performance.
  • The similar recall across two different model families suggests the bottleneck is the prompt-embedded domain knowledge rather than the specific model; a controlled ablation with and without the NSQIP manual in the prompt would test this.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes using large language models (LLMs) during the development phase of a rule-based clinical NLP pipeline, rather than at runtime, and investigates the first two development steps: identifying clinically relevant snippets from patient notes and extracting keywords for a rule-based named entity recognition (NER) component. The authors evaluate their approach on a surgical site infection (SSI) detection task using the EasyCIE system, reporting high recall for snippet identification (Deepseek 0.98, Qwen 0.99) and 100% coverage for keyword extraction. They argue that these results provide initial evidence that LLMs can assist in semi-automating rule-based NLP development.

Significance. If the central feasibility claim were well supported, this would be a valuable contribution: it points to a way of combining the interpretability and efficiency of rule-based systems with the semantic capabilities of LLMs, using LLMs only at development time. The high recall for snippet identification is genuinely suggestive, and the transparent prompt design and error analysis are useful for the community. However, the keyword-extraction result, which is a key pillar of the abstract's claim, rests on an evaluation design that essentially guarantees high coverage and does not measure precision, so the current evidence does not establish that LLM-generated keywords are ready for NER rule construction.

major comments (3)
  1. [Methods – Tasks formulation; Methods – Evaluation] The keyword extraction test set is constructed by running a subset of the existing EasyCIE NER rules and keeping only snippets that produced output, and the coverage metric asks whether any LLM-generated keyword overlaps with the original rule output. This design measures backward compatibility with the existing rules, not the quality or clinical utility of the generated keywords, and it makes 100% coverage trivially achievable by overgenerating broad terms. The authors explicitly acknowledge this in the Discussion ('Not surprisingly, both models excessively generate keywords that covered the snippets of the original rules can match') and in the Limitations ('the keywords generated currently remain distant from the optimized rule set utilized in the NER component'). The abstract's claim of '1.0 in extracting key terms for NER' is therefore not supported by the evidence.
  2. [Results – Table 1; Discussion – Snippet identification error analysis] The reported precision for snippet identification is 0.10 for Deepseek and 0.08 for Qwen, with F1 scores of 0.18 and 0.15. The paper argues that low precision is acceptable because many false positives are actually useful clinical cues, but the reclassification in Table 2 is based on the authors' own judgment and is not validated against an external gold standard or adjudicated by independent annotators. Without such validation, the claim that the false positives are 'largely informative' is not empirically established, and the high recall alone does not demonstrate that the LLMs reliably identify snippets useful for downstream rule development.
  3. [Discussion – Findings of the keyword extraction task; Limitations] Because the keyword test snippets are selected by running the original EasyCIE rules, the evaluation cannot detect whether LLM-generated keywords would help identify relevant text that the original rules miss. The study therefore cannot support the broader claim that LLMs assist in extending or improving a rule-based system beyond the existing rule set, only that they can reproduce some of its outputs. This is a load-bearing limitation for the paper's central argument that LLMs can accelerate rule-based NLP development.
minor comments (6)
  1. [Abstract; Results – Table 1] The abstract describes the recall as 'exceptional' without mentioning the very low precision and F1 scores; reporting these values in the abstract, or at least a balanced summary, would set more accurate expectations.
  2. [Discussion – Findings of the keyword extraction task] The sentence 'Additionally, the expanded keywords instructions seem inspiring to make the rules more generalizable' appears twice in close succession, with slightly different wording, and should be reduced to one occurrence.
  3. [Discussion – Snippet identification error analysis] There is a typo 'if if such a snippet were merely part of a routine follow-up' that should be corrected.
  4. [Appendix – Keywords extraction prompts] The appendix contains 'dictonary' instead of 'dictionary' and uses an incomplete example template that may confuse readers; a careful proofread of the appendix is recommended.
  5. [Results – Table 2] The note 'Some snippets fall into multiple types' makes the counts in Table 2 ambiguous; presenting the number of snippets in each exclusive category, or the total number of snippets classified, would improve clarity.
  6. [Methods – Evaluation] No confidence intervals or statistical significance tests are reported for the recall, precision, or coverage estimates; given the small test sets likely involved, this would help assess the reliability of the results.

Circularity Check

1 steps flagged · score 6.0 of 10

The 1.0 keyword-coverage result is an artifact of a circular evaluation: the test snippets are generated by the original rules and overgeneration is not penalized.

  1. self definitional [Methods, Tasks formulation; Evaluation]
    "we processed these snippets using our previously developed NLP pipeline with a subset of NER rules and extracted all the snippets that produced any output from the NER component. These snippets were then used as the input for the second task: verifying if the keyword set derived by the LLMs can adequately cover the rule set used in the existing NER component. ... For the keyword extraction task, we measured the percentage of the snippets that generated keywords can cover compared with original rules."

    The keyword extraction test set is not an independent sample of clinically relevant text; it consists exclusively of snippets that the existing EasyCIE NER rules already match. The coverage metric then asks whether LLM-generated keywords overlap with those same original rule outputs. Because no precision is measured and the prompts instruct the models to expand keywords with synonyms, broad keyword lists trivially cover every pre-selected snippet. The authors themselves concede that 'both models excessively generate keywords' and that the generated keywords 'remain distant from the optimized rule set utilized in the NER component.' Thus the reported 100% coverage is forced by the evaluation design rather than demonstrating that LLM-extracted keywords are usable for NER rule construction.

full rationale

The snippet identification task is evaluated against independent human annotations, so the reported high recall (Deepseek 0.98, Qwen 0.99) is not circular. The keyword extraction result, however, is circular. The test snippets were created by running a subset of the original NER rules and keeping only snippets that produced output; the coverage metric then measures whether LLM-generated keywords cover those original rule outputs. Since precision was not measured and models were instructed to expand synonyms, broad keyword generation trivially achieves 100% coverage. The authors' own statements—'both models excessively generate keywords that covered the snippets of the original rules can match' and the keywords 'currently remain distant from the optimized rule set utilized in the NER component'—confirm that the 1.0 coverage does not validate NER-ready keyword extraction. No load-bearing self-citation was found; the EasyCIE citations describe the existing system and prior validation, not the LLM result. Overall, one of the two central claims reduces by construction to the evaluation design, warranting a score of 6.

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

No numeric parameters were fit; the load-bearing assumptions are about gold-standard validity and test-set construction. The most consequential choice is using the original rules as the gold standard for keyword extraction.

assumptions (4)
  • domain assumption Human snippet-level annotations in the NSQIP SSI dataset are a complete and correct gold standard.
    All snippet recall metrics assume these annotations correctly separate informative from uninformative text.
  • ad hoc to paper The existing EasyCIE NER rules are a valid reference for evaluating generated keywords.
    The keyword coverage metric uses the original rules as the target, which presupposes their completeness and correctness.
  • domain assumption Snippets that produced output from the existing NER component are representative of all snippets requiring keyword extraction.
    The task 2 test set was built from these snippets, which may not reflect all clinically relevant snippets.
  • domain assumption The subset of NER rules used in the pipeline is representative of the full rule set.
    The paper does not specify which subset was used or its size, so the coverage result may not generalize to the full rule set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Initial Investigation of LLM-Assisted Development of Rule-Based Clinical NLP System." pith.science (2026). https://pith.science/paper/FY3TMW3G

@misc{pith2026250616628,
  author       = {Pith},
  title        = {Pith review of: Initial Investigation of LLM-Assisted Development of Rule-Based Clinical NLP System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FY3TMW3G}},
  note         = {Machine review of arXiv:2506.16628}
}
read the original abstract

Despite advances in machine learning (ML) and large language models (LLMs), rule-based natural language processing (NLP) systems remain active in clinical settings due to their interpretability and operational efficiency. However, their manual development and maintenance are labor-intensive, particularly in tasks with large linguistic variability. To overcome these limitations, we proposed a novel approach employing LLMs solely during the rule-based systems development phase. We conducted the initial experiments focusing on the first two steps of developing a rule-based NLP pipeline: find relevant snippets from the clinical note; extract informative keywords from the snippets for the rule-based named entity recognition (NER) component. Our experiments demonstrated exceptional recall in identifying clinically relevant text snippets (Deepseek: 0.98, Qwen: 0.99) and 1.0 in extracting key terms for NER. This study sheds light on a promising new direction for NLP development, enabling semi-automated or automated development of rule-based systems with significantly faster, more cost-effective, and transparent execution compared with deep learning model-based solutions.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 44 canonical work pages

  1. [1]

    & Reiss, F

    Chiticariu, L., Li, Y. & Reiss, F. R. Rule-Based Information Extraction is Dead! Long Live Rule-Based Information Extraction Systems! in Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (eds. Yarowsky, D., Baldwin, T., Korhonen, A., Livescu, K. & Bethard, S.) 827–832 (Association for Computational Linguistics, Seattle...

    work page 2013

  2. [2]

    Chiang, C.-C. et al. A large language model–based generative natural language processing framework fine-tuned on clinical notes accurately extracts headache frequency from electronic health records. Headache J. Head Face Pain 64, 400–409 (2024)

    work page 2024

  3. [3]

    Wornow, M. et al. Zero-Shot Clinical Trial Patient Matching with LLMs. NEJM AI 2, AIcs2400360 (2025)

    work page 2025

  4. [4]

    Peng, C. et al. Model tuning or prompt Tuning? a study of large language models for clinical concept and relation extraction. J. Biomed. Inform. 153, 104630 (2024)

    work page 2024

  5. [5]

    & Wang, Y

    Sivarajkumar, S., Kelley, M., Samolyk-Mazzanti, A., Visweswaran, S. & Wang, Y. An Empirical Evaluation of Prompting Strategies for Large Language Models in Zero-Shot Clinical Natural Language Processing: Algorithm Development and Validation Study. JMIR Med. Inform. 12, e55318 (2024)

    work page 2024

  6. [6]

    P., Virvou, M

    Panagoulias, D. P., Virvou, M. & Tsihrintzis, G. A. Augmenting Large Language Models with Rules for Enhanced Domain-Specific Interactions: The Case of Medical Diagnosis. Electronics 13, 320 (2024)

    work page 2024

  7. [7]

    Wang, X. et al. RuleAlign: Making Large Language Models Better Physicians with Diagnostic Rule Alignment. Preprint at https://doi.org/10.48550/arXiv.2408.12579 (2024)

    work page Pith review arXiv doi:10.48550/arxiv.2408.12579 2024

  8. [8]

    & Kierner, Z

    Kierner, S., Kucharski, J. & Kierner, Z. Taxonomy of hybrid architectures involving rule-based reasoning and machine learning in clinical decision systems: A scoping review. J. Biomed. Inform. 144, 104428 (2023)

    work page 2023

Show all 46 references
  1. [9]

    & Beam, A

    Kompa, B., Snoek, J. & Beam, A. L. Second opinion needed: communicating uncertainty in medical machine learning. Npj Digit. Med. 4, 1–6 (2021)

    2021

  2. [10]

    Cemri, M. et al. Why Do Multi-Agent LLM Systems Fail? Preprint at https://doi.org/10.48550/arXiv.2503.13657 (2025)

    2025 doi arXiv

  3. [11]

    Bucher, B. T. et al. Portable Automated Surveillance of Surgical Site Infections Using Natural Language Processing: Development and Validation. Ann. Surg. 272, 629 (2020)

    2020

  4. [12]

    Shi, J. et al. Identifying Patients Who Meet Criteria for Genetic Testing of Hereditary Cancers Based on Structured and Unstructured Family Health History Data in the Electronic Health Record: Natural Language Processing Approach. JMIR Med. Inform. 10, e37842 (2022)

    2022

  5. [13]

    Shi, J. et al. Natural language processing for the surveillance of postoperative venous thromboembolism. Surgery 170, 1175–1182 (2021)

    2021

  6. [14]

    & Hurdle, J

    Shi, J. & Hurdle, J. F. Trie-based rule processing for clinical NLP: A use-case study of n-trie, making the ConText algorithm more efficient and scalable. J. Biomed. Inform. 85, 106–113 (2018)

    2018

  7. [15]

    An efficient pipeline for high-performance information extraction of support clinical natural language processing

    Shi, J. An efficient pipeline for high-performance information extraction of support clinical natural language processing. (University of Utah, 2019)

    2019

  8. [16]

    Eyre, H. et al. Launching into clinical space with medspaCy: a new clinical text processing toolkit in Python. AMIA. Annu. Symp. Proc. 2021, 438–447 (2022)

    2022

  9. [17]

    Wei, J. et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Preprint at https://doi.org/10.48550/arXiv.2201.11903 (2023)

    2023 doi arXiv

  10. [18]

    & Chen, C

    Jiang, R., Liu, L. & Chen, C. MoPE: Mixture of Prompt Experts for Parameter-Efficient and Scalable Multimodal Fusion. Preprint at https://doi.org/10.48550/arXiv.2403.10568 (2025). Appendix Note: the few shot examples are omitted here because of page limit. Snippet identificati...

    2025 doi arXiv

  11. [19]

    Is it valid or not?

    Verify the opinion expressed by another surgeon. Is it valid or not?

  12. [20]

    Do you agree with the other surgeon's opinion? Why?

  13. [21]

    Summarize your own opinion which will be the final decision

  14. [22]

    conclusion

    Generate a json-format output: either {{"conclusion":"yes"}} or {{"conclusion":"no"}} Keywords extraction prompts: Reasoning prompt: {SSI annotation guideline} -------------------------------- # Your Role: Clinical Informatist # Task: Identify the keywords from the given snipp...

  15. [23]

    Identify all anatomic sites or body parts or mentions indicating anatomic sites that potentially can be a surgical site (can potentially be used to link a procedure or infection mentioned outside the given snippet)

  16. [24]

    Identify all surgical or invasive procedures (can potentially be used to link an infection mentioned outside the given snippet)

  17. [25]

    Identify all infection or infection treatment

  18. [26]

    Identify all wound care

  19. [27]

    Identify all other infections that can be used to explain signs, symptoms, or lab abnormalities that might or might not be caused by SSI

  20. [28]

    Also, you need to double-check to make sure they keywords themselves are not over- generalized to get massive matches in other notes

    Make sure the keywords can actually be found in the original text and capture the key information that can be used for downstream analysis. Also, you need to double-check to make sure they keywords themselves are not over- generalized to get massive matches in other notes

  21. [30]

    'concerning for,' as these modifiers will be identified in downstream NLP components

    Prune the keywords to remove certainty, temporality, and experiencer modifiers, e.g. 'concerning for,' as these modifiers will be identified in downstream NLP components

  22. [31]

    concepts

    Summarize your findings in a json dictonary using template: {{"concepts": [], "expanded_concepts":[]}} -------------------------------- # For example: ### Given the input snippet: She underwent CT scan of the abdomen that showed an appendix dilated to 11 mm with adjacent fat s...

  23. [32]

    **Anatomic Sites or Body Parts (Surgical Sites):** - **Appendix:** This is a clear anatomic site that could be a surgical site in the context of appendectomy

  24. [33]

    It is more relevant for identifying anatomical findings and conditions rather than surgical interventions

    **Surgical or Invasive Procedures:** - **CT scan:** This is an imaging procedure, but it is not an invasive procedure. It is more relevant for identifying anatomical findings and conditions rather than surgical interventions

  25. [34]

    However, it is not specific to surgical site infections (SSIs)

    **Infection or Infection Treatment:** - **Inflammation:** This term indicates the presence of an inflammatory response, which could be related to an infection. However, it is not specific to surgical site infections (SSIs)

  26. [35]

    **Wound Care:** - **Not mentioned:** There is no mention of wound care in the snippet

  27. [36]

    It could be related to a non- surgical infection, such as appendicitis

    **Other Infections:** - **Inflammation:** This can be a sign of other infections, not specifically SSI. It could be related to a non- surgical infection, such as appendicitis

  28. [37]

    They are specific enough to capture relevant information without being overly generalized

    **Keywords in Original Text:** - **Appendix, inflammation, CT scan:** These are the key terms that can be used for downstream analysis. They are specific enough to capture relevant information without being overly generalized

  29. [38]

    **Expanded Concepts with Synonyms:** - **Appendix:** Appendix, cecum, right lower quadrant - **Inflammation:** Inflammation, inflammatory response - **CT scan:** CT scan, computed tomography

  30. [39]

    concepts

    **Pruned Keywords:** - Remove any certainty, temporality, or experiencer modifiers. The terms are already quite specific. ### Summary in JSON Dictionary: ```json {{ "concepts": ["Appendix", "Inflammation", "CT scan"], "expanded_concepts": ["Appendix", "Cecum", "Right lower qua...

  31. [40]

    Also, you need to double check to make sure the keywords themselves are not over generalized to get massive matches in other notes

    Identify the keywords directly from the snippet, make sure the keywords can actually found in original text and capture the key information that can be used to determine if a patient has SSI, even if the snippet doesn't have complete information. Also, you need to double check...

  32. [41]

    Based on your clinical knowledge, expand the keywords with synonyms, so that same meaning in a different snippet can also be found

  33. [42]

    'concerning for', as these modifiers will be identified in downstream NLP components

    Prune the keywords to remove certainty, temporality, and experiencer modifiers, e.g. 'concerning for', as these modifiers will be identified in downstream NLP components

  34. [43]

    Elaborate with detailed justification

  35. [44]

    # Your task: You will be given the clinical note snippet and the analysis from anther clinical informatist:

    Summarize your findings in a json dictionary. # Your task: You will be given the clinical note snippet and the analysis from anther clinical informatist:

  36. [45]

    Pay attention to the expanded keywords, make sure none of them is too broad

    Verify the analysis to check if it follows the above instructions step by step. Pay attention to the expanded keywords, make sure none of them is too broad

  37. [46]

    Double check the derived keywords list in json format

  38. [47]

    concepts

    Finalize the keywords in json format using the same structure: {{"concepts": [], "expanded_concepts":[]}} {text}

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.