Pith. sign in

REVIEW 4 major objections 5 minor 36 references

DENSE: Longitudinal Progress Note Generation with Temporal Modeling of Heterogeneous Clinical Notes Across Hospital Visits

T0 review · 4 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 DENSE is a pipeline that turns scattered hospital notes into temporally coherent SOAP progress notes, and the paper reports that its generated notes exceed original notes on a longitudinal alignment ratio of 1.089.

desk verdict A plausible longitudinal note-generation system whose headline evaluation metric is self-confirmatory; the direction is worth a major revision, not a desk reject. read the letter →

arxiv 2507.14079 v1 pith:SLY6HAQV submitted 2025-07-18 cs.CL cs.AIcs.IRcs.LG

Garapati Keerthana , Manik Gupta This is my paper
classification cs.CLcs.AIcs.IRcs.LG
keywords progressnotegenerationtemporalcoherenceretrieval-augmentedclinicalcategorizationlongitudinalEHRSOAPnoteslargelanguagemodelshealthrecordnarratives
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

The paper introduces DENSE, a system that generates progress notes for hospital visits even when no progress note was written, by reorganizing the scattered clinical notes in a public critical care dataset into structured, chronological, visit-level input. The key idea is to mimic how a physician consults prior encounters: for every visit after the first, the system retrieves clinically relevant evidence and also conditions the language model on a summary of the previously generated note. Evaluated on 56 patients with more than 1,100 visits, DENSE reports high semantic similarity to gold notes (0.74), perfect SOAP structure coverage, and a temporal alignment ratio of 1.089, meaning adjacent generated notes look more semantically coherent than adjacent clinician-authored notes. If correct, the system offers a scalable way to restore longitudinal patient narratives in under-documented electronic health records and to support downstream summarization and prediction tasks.

What carries the argument

The central mechanism is the temporal alignment and dual-mode prompting pipeline: noisy note labels are reclassified into 16 consistent clinical types, notes are pivoted into visit-centric rows, and each note is cleaned and sectioned; retrieval then operates over semantically embedded chunks with SOAP-aligned queries, returning evidence from both the current and prior visits. For the first visit, the prompt uses only retrieved evidence, but for every later visit the prompt augments current evidence with a concise summary of the previously generated progress note. This autoregressive conditioning is what carries the longitudinal claim, and the temporal alignment ratio, the quotient of mean pairwise cosine similarity between adjacent generated notes and adjacent gold notes, is the evaluation device that quantifies it.

What would settle it

Compute the same temporal alignment ratio for a baseline that copies the previous visit's note forward, or that repeats the prior generated note verbatim; if that baseline matches or exceeds 1.089, the metric cannot distinguish longitudinal coherence from redundancy. Additionally, blind clinician ratings of adjacent-note continuity and clinical usefulness would settle whether higher alignment means better notes.

Watch

Extended reading notes

Core claim

DENSE claims that longitudinal progress note generation is achievable by combining clinically informed retrieval with autoregressive conditioning on a summary of the prior visit's note. Across 56 patients and over 1,100 hospital visits, the generated notes achieved a mean adjacent-note semantic similarity of 0.877 versus 0.807 for the original notes, yielding a temporal alignment ratio of 1.089; the paper interprets this as the synthetic notes maintaining narrative continuity across visits better than real documentation. The authors further report that the notes preserve core meaning (0.74 semantic similarity), follow the SOAP structure completely, and are only slightly longer than gold-standard notes, supporting their claim that DENSE restores narrative coherence across fragmented documentation.

Load-bearing premise

The headline result assumes that a higher mean cosine similarity between adjacent generated notes, relative to adjacent gold notes, indicates better longitudinal clinical documentation; because the generator conditions on its own previous note, that similarity can rise from repetition rather than clinical quality, and no human or clinical validation links the metric to utility.

Editorial extensions

If this is right

  • Hospitals could generate missing progress notes for historical visits, reconstructing longitudinal narratives for research, audit, and medicolegal review.
  • The dual-mode prompt design gives a practical template for LLM-based scribes that reference prior encounters rather than treating each visit in isolation.
  • The 16-type note taxonomy offers a reusable normalization layer for noisy, inconsistently labeled EHR note categories.
  • The longitudinal benchmark of 56 patients and over 1,100 visits provides a concrete evaluation setup for future note-generation models that aim to preserve temporal coherence.
  • Because the generated notes preserve SOAP structure and semantic content, they could serve as synthetic data for training downstream predictive models on under-documented patient populations.

Reading between the lines

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

  • An editorial caution: a trivial baseline that simply copies the previous visit's note forward would likely inflate adjacent-note cosine similarity, so the 1.089 ratio should be compared against such a baseline before interpreting it as clinical continuity rather than redundancy.
  • A testable extension is to have clinicians blind-rate adjacent DENSE notes and gold notes for continuity and usefulness; this would show whether the temporal alignment ratio tracks human-perceived quality.
  • The same visit-pivoting and temporal conditioning machinery could plausibly be applied to generate discharge summaries or nursing shift handoffs, not only progress notes.
  • Applying the reclassification heuristics to other EHR datasets would reveal whether the 16-type taxonomy generalizes or requires site-specific customization.
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 / 5 minor

Summary. The paper proposes DENSE, a pipeline that reclassifies MIMIC-III clinical note types into 16 categories, pivots notes into visit-level rows, retrieves evidence using a CLI-RAG adapted framework, and prompts an LLM to generate SOAP-style progress notes for each hospital visit. For visits after the first, the prompt includes a concise summary of the previously generated note. The authors evaluate on a curated cohort of 56 patients with 1,100 visits, reporting lexical, semantic, structural, and temporal metrics, including a novel temporal alignment ratio defined as the mean pairwise cosine similarity between adjacent generated notes divided by that of gold notes. They report a ratio of 1.089, which they interpret as DENSE outperforming clinician-authored notes in longitudinal consistency.

Significance. If validated, DENSE would address a real gap in longitudinal clinical documentation by generating progress notes from scattered evidence. The paper's strengths include a clearly motivated problem, a deterministic note reclassification strategy, a longitudinal benchmark cohort, and the use of a publicly available dataset. However, the headline result is not credible as evidence because the temporal alignment metric is structurally biased by the generation setup, and no baselines, significance tests, or human evaluations are provided. The contribution is therefore currently unsupported.

major comments (4)
  1. [V-B, Eq. (1)] The temporal alignment ratio is inflated by design. Section IV-D states that for all visits beyond the first, the prompt includes a concise summary of the immediately preceding generated progress note. Adjacent generated notes therefore share content by construction, whereas the gold notes were not produced under this autoregressive constraint. A ratio above 1.0 is expected for any system that merely restates the previous summary, and the paper provides no baseline such as copy-forward, no-conditioning, or summary-only generation to calibrate the metric. The claim that DENSE 'outperformed the original notes' is thus not established.
  2. [V-B] No statistical significance or confidence intervals are reported for the temporal consistency scores (0.877 vs. 0.807) across the 56 patients. Given the small cohort and the metric's construction, this difference could easily arise from noise. The metric is also not externally validated; the paper provides no evidence that higher adjacent cosine similarity corresponds to better clinical documentation or to the narrative continuity that clinicians value.
  3. [V] The experimental section includes no comparison to any baseline generation method. Without ablations (e.g., removing the temporal conditioning, using only current-visit evidence, or fine-tuning a clinical LLM) or comparisons to simpler approaches such as concatenation or single-visit summarization, the contribution of DENSE's specific design choices is unclear. The paper's central claim of improved temporal coherence requires at least one such comparison.
  4. [V-B, SOAP Structure Score] The SOAP structural completeness score is reported as a perfect 4.0/4.0 for all generated notes, but the evaluation protocol is not described. No information is given about the number of annotators, their clinical expertise, whether they were blinded to the system, or inter-rater reliability. Since the prompts are designed to elicit SOAP sections, this metric likely reflects prompt instruction-following rather than clinical quality.
minor comments (5)
  1. [III-A] The deterministic re-mapping strategy is described at a high level; please provide examples of the regular expressions or a description of the semantic pattern matching used, and report the mapping accuracy on a sample.
  2. [IV-D] The step that 'parses and summarizes' retrieved chunks to remove redundancy is underspecified; clarify whether this is an LLM-based summarization step and how redundancy is detected.
  3. [Table IV] Consider reporting standard deviations or ranges for the mean scores, particularly for the temporal consistency measures, to give the reader a sense of patient-level variability.
  4. [References] Some references appear out of numerical order in the text (e.g., [17] and [18], [20]), and the relation to the authors' prior CLI-RAG work [14] should be stated more explicitly to clarify the novelty.
  5. [Figure 4] The patient-level alignment scores in Figure 4 are difficult to interpret; a histogram or box plot with summary statistics would be more informative.

Circularity Check

1 steps flagged · score 6.0 of 10

The temporal alignment ratio (Eq. 1) is inflated by construction: DENSE conditions each note after the first on a summary of the previous generated note, and the headline metric then measures adjacent-note cosine similarity, so the 1.089 result rewards the system for echoing its own conditioning input.

  1. self definitional [Section IV-D and Section V-B, Eq. (1)]
    "In visits beyond the patient’s first, we incorporate a longitudinal aspect by referencing the previously generated progress note from the immediately preceding visit. ... we compute the average pairwise cosine similarity between adjacent generated progress notes using embeddings from a Sentence-BERT model ... The final temporal alignment ratio is computed as the quotient of the generated score over the gold score ... DENSE outperformed the original notes with a mean generated temporal consistency of 0.877 vs. 0.807, yielding an alignment ratio of 1.089."

    The metric's numerator measures cosine similarity between adjacent generated notes, but adjacent generated notes are not independent: each note after the first is generated from a prompt that explicitly contains a summary of the immediately previous generated note. Thus high adjacent similarity is written into the generation procedure; the model is scored on how well it reproduces its own conditioning input. Gold notes were never produced under this autoregressive constraint, so the 1.089 ratio is expected for any system that echoes the injected summary.

full rationale

The paper's central quantitative claim is that DENSE-generated notes 'outperform ground truth in temporal consistency' via the temporal alignment ratio of 1.089. That claim is not established as an independent result: the metric in Eq. 1 measures exactly the property that the generation pipeline forces. Because Section IV-D states that each subsequent visit's prompt includes a concise summary of the previously generated note, the adjacent-note similarity measured in Section V-B is partly a reflection of the model echoing its own input, not an independent measure of longitudinal clinical coherence. Gold notes were not generated under that same constraint, so the comparison is structurally biased. The paper also provides no significance testing, no baseline such as copy-forward or summary-only generation, and no human evaluation linking adjacent cosine similarity to documentation quality; the Discussion's future-work call for 'extension to MIMIC-IV and real-world EHR systems for external validation' confirms the absence of external validation. Other components, such as the note-type reclassification, retrieval setup, and SOAP-structure evaluation, are engineering choices and are not circular in themselves, but they do not rescue the headline metric. Overall, the paper exhibits partial circularity: one central 'prediction' reduces by construction to the design of its evaluation metric.

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

The central claim rests on heuristic note classification, an unvalidated semantic-similarity proxy for clinical continuity, and the questionable assumption that more pairwise similarity between adjacent notes is always better. No new physical or data entities are introduced; the 'temporal alignment ratio' is a new metric, but it is not an entity.

free parameters (2)
  • chunk_size = 3000 characters
    Chunking size for retrieval is chosen by hand without justification or experimentation.
  • chunk_overlap = 300 characters
    Overlap between chunks is another hand-picked hyperparameter that affects retrieval quality but is not investigated.
assumptions (4)
  • domain assumption MIMIC-III note labels can be reliably mapped to 16 clean categories using regex and manual curation.
    The entire pipeline depends on the accuracy of this reclassification, but the paper does not report any validation of the mapping.
  • domain assumption Cosine similarity of sentence embeddings from all-mpnet-base-v2 is a valid proxy for clinical content continuity.
    The temporal alignment metric relies on this assumption, which is not independently validated for clinical text.
  • ad hoc to paper A temporal alignment ratio greater than 1 indicates improved longitudinal quality.
    This assumption is loaded into the evaluation: higher similarity is interpreted as better continuity without considering that repetition or copy-forward is clinically undesirable.
  • domain assumption LLMs generate clinically faithful progress notes from retrieved evidence chunks.
    The system assumes the generation model produces clinically coherent text, but the paper does not assess clinical accuracy or hallucination risk through expert review.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DENSE: Longitudinal Progress Note Generation with Temporal Modeling of Heterogeneous Clinical Notes Across Hospital Visits." pith.science (2026). https://pith.science/paper/SLY6HAQV

@misc{pith2026250714079,
  author       = {Pith},
  title        = {Pith review of: DENSE: Longitudinal Progress Note Generation with Temporal Modeling of Heterogeneous Clinical Notes Across Hospital Visits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SLY6HAQV}},
  note         = {Machine review of arXiv:2507.14079}
}
abstract

Progress notes are among the most clinically meaningful artifacts in an Electronic Health Record (EHR), offering temporally grounded insights into a patient's evolving condition, treatments, and care decisions. Despite their importance, they are severely underrepresented in large-scale EHR datasets. For instance, in the widely used Medical Information Mart for Intensive Care III (MIMIC-III) dataset, only about $8.56\%$ of hospital visits include progress notes, leaving gaps in longitudinal patient narratives. In contrast, the dataset contains a diverse array of other note types, each capturing different aspects of care. We present DENSE (Documenting Evolving Progress Notes from Scattered Evidence), a system designed to align with clinical documentation workflows by simulating how physicians reference past encounters while drafting progress notes. The system introduces a fine-grained note categorization and a temporal alignment mechanism that organizes heterogeneous notes across visits into structured, chronological inputs. At its core, DENSE leverages a clinically informed retrieval strategy to identify temporally and semantically relevant content from both current and prior visits. This retrieved evidence is used to prompt a large language model (LLM) to generate clinically coherent and temporally aware progress notes. We evaluate DENSE on a curated cohort of patients with multiple visits and complete progress note documentation. The generated notes demonstrate strong longitudinal fidelity, achieving a temporal alignment ratio of $1.089$, surpassing the continuity observed in original notes. By restoring narrative coherence across fragmented documentation, our system supports improved downstream tasks such as summarization, predictive modeling, and clinical decision support, offering a scalable solution for LLM-driven note synthesis in real-world healthcare settings.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

  1. [1]

    Characterizing the source of text in electronic health record progress notes,

    M. D. Wang, R. Khanna, and N. Najafi, “Characterizing the source of text in electronic health record progress notes,” JAMA internal medicine, vol. 177, no. 8, pp. 1212–1213, 2017

    work page 2017

  2. [2]

    Learning to write case notes using the soap format,

    S. Cameron and I. Turtle-Song, “Learning to write case notes using the soap format,” Journal of Counseling & Development , vol. 80, no. 3, pp. 286–292, 2002. [Online]. Available: https: //onlinelibrary.wiley.com/doi/abs/10.1002/j.1556-6678.2002.tb00193.x

    work page arXiv 2002

  3. [3]

    Length and redundancy of outpatient progress notes across a decade at an academic medical center,

    A. Rule, S. Bedrick, M. F. Chiang, and M. R. Hribar, “Length and redundancy of outpatient progress notes across a decade at an academic medical center,” JAMA Network Open , vol. 4, no. 7, pp. e2 115 334– e2 115 334, 2021

    work page 2021

  4. [4]

    Prediction of emergency department patient disposition based on natural language processing of triage notes,

    N. W. Sterling, R. E. Patzer, M. Di, and J. D. Schrager, “Prediction of emergency department patient disposition based on natural language processing of triage notes,” International journal of medical informatics, vol. 129, pp. 184–188, 2019

    work page 2019

  5. [5]

    Hierarchical annotation for building a suite of clinical natural language processing tasks: progress note understanding,

    Y . Gao, D. Dligach, T. Miller, S. Tesch, R. Laffin, M. M. Churpek, and M. Afshar, “Hierarchical annotation for building a suite of clinical natural language processing tasks: progress note understanding,” in LREC... International Conference on Language Resources & Evalua- tion:[proceedings]. International Conference on Language Resources & Evaluation, vo...

    work page 2022

  6. [6]

    Leveraging medical knowledge graphs into large language models for diagnosis prediction: Design and application study,

    Y . Gao, R. Li, E. Croxford, J. Caskey, B. W. Patterson, M. Churpek, T. Miller, D. Dligach, and M. Afshar, “Leveraging medical knowledge graphs into large language models for diagnosis prediction: Design and application study,” Jmir Ai, vol. 4, p. e58670, 2025

    work page 2025

  7. [7]

    Attention-based clinical note summarization,

    N. Kanwal and G. Rizzo, “Attention-based clinical note summarization,” in Proceedings of the 37th ACM/SIGAPP symposium on applied com- puting, 2022, pp. 813–820

    work page 2022

  8. [8]

    A multimodal transformer: Fusing clinical notes with structured ehr data for interpretable in-hospital mortality prediction,

    W. Lyu, X. Dong, R. Wong, S. Zheng, K. Abell-Hart, F. Wang, and C. Chen, “A multimodal transformer: Fusing clinical notes with structured ehr data for interpretable in-hospital mortality prediction,” in AMIA Annual Symposium Proceedings , vol. 2022, 2023, p. 719

    work page 2022

Show all 36 references
  1. [9]

    Reducing redundancy in clinical documentation: a study of progress note content and structure,

    L. Huang, R. Shah, and K. Hegde, “Reducing redundancy in clinical documentation: a study of progress note content and structure,” npj Digital Medicine, vol. 7, no. 1, p. 12, 2024

    2024

  2. [10]

    Copy, paste, and cloned notes in electronic health records,

    J. M. Weis and P. C. Levy, “Copy, paste, and cloned notes in electronic health records,” Chest, vol. 145, no. 3, pp. 632–638, 2014

    2014

  3. [11]

    Clinical documentation in the 21st century: executive summary of a policy position paper from the american college of physicians,

    T. Kuhn, P. Basch, M. Barr, T. Yackel, and M. I. C. of the American College of Physicians*, “Clinical documentation in the 21st century: executive summary of a policy position paper from the american college of physicians,” Annals of internal medicine, vol. 162, no. 4, pp. 301...

    2015

  4. [12]

    Mimic-iii, a freely accessible critical care database,

    A. E. W. 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, no. 1, p. 160035, 2016. [Online]. Available: https://doi.org/...

    2016 doi

  5. [13]

    Redundancy of progress notes for serial office visits,

    M. R. Hribar, A. Rule, A. E. Huang, H. Dusek, I. H. Goldstein, B. Henriksen, W.-C. Lin, A. Igelman, and M. F. Chiang, “Redundancy of progress notes for serial office visits,” Ophthalmology, vol. 127, no. 1, pp. 134–135, 2020

    2020

  6. [14]

    Cli-rag: A retrieval-augmented framework for clinically structured and context aware text generation with llms,

    G. Keerthana and M. Gupta, “Cli-rag: A retrieval-augmented framework for clinically structured and context aware text generation with llms,” arXiv preprint arXiv:2507.06715 , 2025

    2025 arXiv

  7. [15]

    Are synthetic clinical notes useful for real natural language processing tasks: A case study on clinical entity recognition,

    J. Li, Y . Zhou, X. Jiang, K. Natarajan, S. V . Pakhomov, H. Liu, and H. Xu, “Are synthetic clinical notes useful for real natural language processing tasks: A case study on clinical entity recognition,” Journal of the American Medical Informatics Association , vol. 28, no. 10...

    2021

  8. [16]

    Towards automatic generation of share- able synthetic clinical notes using neural language models,

    O. Melamud and C. Shivade, “Towards automatic generation of share- able synthetic clinical notes using neural language models,” arXiv preprint arXiv:1905.07002, 2019

    1905 arXiv

  9. [17]

    Progress notes classification and keyword extraction using attention- based deep learning models with bert,

    M. Tang, P. Gandhi, M. A. Kabir, C. Zou, J. Blakey, and X. Luo, “Progress notes classification and keyword extraction using attention- based deep learning models with bert,”arXiv preprint arXiv:1910.05786, 2019

    1910 arXiv

  10. [18]

    Toward relieving clinician burden by automatically generating progress notes using interim hospital data,

    S. Soni and D. Demner-Fushman, “Toward relieving clinician burden by automatically generating progress notes using interim hospital data,” in AMIA Annual Symposium Proceedings , vol. 2024, 2025, p. 1059

    2024

  11. [19]

    Intelligent clinical documentation: Har- nessing generative ai for patient-centric clinical note generation,

    A. Biswas and W. Talukdar, “Intelligent clinical documentation: Har- nessing generative ai for patient-centric clinical note generation,” arXiv preprint arXiv:2405.18346, 2024

    2024 arXiv

  12. [20]

    Clinicalt5: A generative language model for clinical text,

    Q. Lu, D. Dou, and T. Nguyen, “Clinicalt5: A generative language model for clinical text,” inFindings of the Association for Computational Linguistics: EMNLP 2022 , 2022, pp. 5436–5443

    2022

  13. [21]

    Assessing electronic note quality using the physician documentation quality instru- ment (pdqi-9),

    P. D. Stetson, S. Bakken, J. O. Wrenn, and E. L. Siegler, “Assessing electronic note quality using the physician documentation quality instru- ment (pdqi-9),” Applied clinical informatics, vol. 3, no. 02, pp. 164–174, 2012

    2012

  14. [22]

    Synthea: An approach, method, and software mechanism for generating synthetic patients and the synthetic electronic health care record,

    J. Walonoski, M. Kramer, J. Nichols, A. Quina, C. Moesel, D. Hall, C. Duffett, K. Dube, T. Gallagher, and S. McLachlan, “Synthea: An approach, method, and software mechanism for generating synthetic patients and the synthetic electronic health care record,” Journal of the Amer...

    2018

  15. [23]

    Advancing informatics with electronic medical records bots (emrbots),

    U. Kartoun, “Advancing informatics with electronic medical records bots (emrbots),” Software Impacts, vol. 2, p. 100006, 2019

    2019

  16. [24]

    Gen- erating multi-label discrete patient records using generative adversarial networks,

    E. Choi, S. Biswal, B. Malin, J. Duke, W. F. Stewart, and J. Sun, “Gen- erating multi-label discrete patient records using generative adversarial networks,” in Machine learning for healthcare conference . PMLR, 2017, pp. 286–305

    2017

  17. [25]

    Clinicalbert: Modeling clinical notes and predicting hospital readmission,

    K. Huang, J. Altosaar, and R. Ranganath, “Clinicalbert: Modeling clinical notes and predicting hospital readmission,” arXiv preprint arXiv:1904.05342, 2019

    1904 arXiv

  18. [26]

    Publicly shareable clinical large language model built on synthetic clinical notes,

    S. Kweon, J. Kim, J. Kim, S. Im, E. Cho, S. Bae, J. Oh, G. Lee, J. H. Moon, S. C. Youet al., “Publicly shareable clinical large language model built on synthetic clinical notes,” arXiv preprint arXiv:2309.00237 , 2023

    2023 arXiv

  19. [27]

    The dynamic embedded topic model,

    A. B. Dieng, F. J. Ruiz, and D. M. Blei, “The dynamic embedded topic model,” arXiv preprint arXiv:1907.05545 , 2019

    1907 arXiv

  20. [28]

    Etm: Enrichment by topic modeling for automated clin- ical sentence classification to detect patients’ disease history,

    A. Bagheri, A. Sammani, P. G. van der Heijden, F. W. Asselbergs, and D. L. Oberski, “Etm: Enrichment by topic modeling for automated clin- ical sentence classification to detect patients’ disease history,” Journal of Intelligent Information Systems , vol. 55, no. 2, pp. 329–349, 2020

    2020

  21. [29]

    Dynamic topic models,

    D. M. Blei and J. D. Lafferty, “Dynamic topic models,” in Proceedings of the 23rd international conference on Machine learning , 2006, pp. 113–120

    2006

  22. [30]

    A systematic review of large language model (llm) evaluations in clinical medicine,

    S. Shool, S. Adimi, R. Saboori Amleshi, E. Bitaraf, R. Golpira, and M. Tara, “A systematic review of large language model (llm) evaluations in clinical medicine,” BMC Medical Informatics and Decision Making , vol. 25, no. 1, p. 117, 2025

    2025

  23. [31]

    Evaluating measures of redundancy in clinical texts,

    R. Zhang, S. Pakhomov, B. T. McInnes, and G. B. Melton, “Evaluating measures of redundancy in clinical texts,” in AMIA annual symposium proceedings, vol. 2011, 2011, p. 1612

    2011

  24. [32]

    Quantifying clinical narrative redundancy in an electronic health record,

    J. O. Wrenn, D. M. Stein, S. Bakken, and P. D. Stetson, “Quantifying clinical narrative redundancy in an electronic health record,” Journal of the American Medical Informatics Association, vol. 17, no. 1, pp. 49–53, 2010

    2010

  25. [33]

    “note bloat

    J. Liu, D. Capurro, A. Nguyen, and K. Verspoor, ““note bloat” impacts deep learning-based nlp models for clinical prediction tasks,” Journal of biomedical informatics, vol. 133, p. 104149, 2022

    2022

  26. [34]

    Modeling local coherence: An entity-based approach,

    R. Barzilay and M. Lapata, “Modeling local coherence: An entity-based approach,” Computational Linguistics, vol. 34, no. 1, pp. 1–34, 2008

    2008

  27. [35]

    Neural net models of open-domain discourse coherence,

    J. Li and D. Jurafsky, “Neural net models of open-domain discourse coherence,” in Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing , M. Palmer, R. Hwa, and S. Riedel, Eds. Copenhagen, Denmark: Association for Computational Linguistics, Sep...

    2017

  28. [36]

    As- sessing the quality of ai-generated clinical notes: A validated evaluation of a large language model scribe,

    E. Palm, A. Manikantan, M. E. Pepin, H. Mahal, and S. S. Belwadi, “As- sessing the quality of ai-generated clinical notes: A validated evaluation of a large language model scribe,” arXiv preprint arXiv:2505.17047 , 2025

    2025 arXiv

Pith tools

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