Abstract
Objective
Extract infant breast milk feeding status from electronic health record clinical notes to support observational drug-safety studies.
Materials and Methods
Infant notes (19 489; 228 infants) were weakly labeled by large language model (LLM) and validated on 1045 manually reviewed notes. Using these labels, we built a rule-based medspaCy classifier from a curated feeding term set and compared it with ClinPhen and fine-tuned Bidirectional Encoder Representations from Transformers (BERT) document classifiers. Maternal postpartum notes (23 mothers; 616 partitions) were manually labeled to compare BERT with prompt-based language models.
Results
LLM labels achieved 99.62% accuracy (macro F1 99.32%). medspaCy achieved 94.36% accuracy and 90.40% balanced accuracy (macro F1 0.90) and was applied to 52 278 mother–infant dyads. On maternal notes, the prompt-based model improved Positive-class F1 vs BERT (0.96 vs 0.82).
Discussion
LLMs were most effective for scalable labeling; deterministic rules supported transparent large-scale inference.
Conclusion
Weak supervision plus lightweight rules can convert sparse breastfeeding documentation into structured variables at-scale.
Keywords: breast feeding, lactation, natural language processing, documentation
Introduction
Medication safety and efficacy during breastfeeding remains poorly studied,1,2 in part because of ethical and logistical constraints that limit participation in clinical trials and prospective pharmacokinetic studies.3 Breastfeeding is common and recommended,4,5 yet many people who breastfeed require medications for acute or chronic conditions.6–8 A central barrier to post-market evidence generation is the lack of large cohorts with reliable, time-resolved indicators of potential infant exposure to medications via breast milk.
Electronic health records (EHRs) could enable observational cohorts, but infant feeding behavior is inconsistently captured.9 Feeding status evolves over time (exclusive breast milk, mixed feeding, formula only, and weaning) and documentation appears in multiple locations and formats. Free-text notes often contain specific evidence that a mother is breastfeeding,10,11 but terminology is variable and sometimes ambiguous (eg, “bottle feeding” without specifying formula vs expressed breast milk). Consequently, studies frequently rely on coarse proxies (eg, administrative codes) or labor-intensive manual chart review, which is difficult to scale. In a recent study, multiple machine learning methods were explored to extract breastfeeding phenotypes from an academic birth hospital, which reached a 90% F1-score.12 In this present study, our goal was to extend extraction across the full first year of life and to produce a reusable structured variable derived across an academic birthing (The Ohio State University Wexner Medical Center [OSUWMC]) and pediatric hospital (Nationwide Children’s Hospital [NCH]).
Objective
Primary objective: develop and evaluate a Natural Language Processing (NLP) approach, seeded by large language model (LLM)-assisted weak supervision, to classify infant breast milk feeding status from clinical notes during the first year of life and to create document-level labels usable for downstream drug-safety analyses.
Secondary objectives: (1) compare a fine-tuned Bidirectional Encoder Representations from Transformers (BERT) model with a prompt-based local LLM for classifying maternal postpartum notes and (2) illustrate longitudinal breastfeeding timelines by combining maternal and infant classifications.
Methods
Data sources
Infant cohort (NCH)
We identified 228 infants born at OSUWMC (in 2011-2015) with encounters in the NCH network during the first year of life based on the LEAF study.13 NCH uses an Epic EHR14 and archives data in a research data warehouse. We extracted demographics and the text of all notes during the first year of life (19 489 notes; cohort characteristics summarized in Table 1). Document types judged to be instructional or not relevant to feeding behavior (eg, after-visit summaries, discharge instructions, discharge planning, and pharmacy documentation) were excluded, yielding 16 288 notes for analysis (see also Figure 1).
Table 1.
Summary of cohort demographics and engagement with the NCH network (created using tableone15 Python library).
| Overall | Train | Validation | Test | ||
|---|---|---|---|---|---|
| N | 228 | 114 | 57 | 57 | |
| Sex, n (%) | Female | 120 (52.6) | 64 (56.1) | 28 (49.1) | 28 (49.1) |
| Male | 108 (47.4) | 50 (43.9) | 29 (50.9) | 29 (50.9) | |
| Race, n (%) | Asian | 1 (0.4) | 1 (0.9) | 0 (0.0) | 0 (0.0) |
| Black or African American | 167 (73.2) | 80 (70.2) | 47 (82.5) | 40 (70.2) | |
| Multiple race | 13 (5.7) | 7 (6.1) | 2 (3.5) | 4 (7.0) | |
| Unknown | 3 (1.3) | 1 (0.9) | 0 (0.0) | 2 (3.5) | |
| White | 44 (19.3) | 25 (21.9) | 8 (14.0) | 11 (19.3) | |
| Ethnicity, n (%) | Hispanic or Latino | 5 (2.2) | 0 (0.0) | 2 (3.5) | 3 (5.3) |
| Not Hispanic or Latino | 221 (96.9) | 112 (98.2) | 55 (96.5) | 54 (94.7) | |
| Other | 1 (0.4) | 1 (0.9) | 0 (0.0) | 0 (0.0) | |
| Unknown | 1 (0.4) | 1 (0.9) | 0 (0.0) | 0 (0.0) | |
| Visits, median [Q1,Q3] | 14 [7,22] | 14 [5,22] | 13 [8,19] | 15 [8,24] | |
| Total notes, median [Q1,Q3] | 41 [22,70] | 40 [17,64] | 39 [22,63] | 49 [25,90] | |
| Filtered notes, median [Q1,Q3] | 29 [15,53] | 29 [12,46] | 25 [16,43] | 37 [17,60] | |
| First visit age (days), median [Q1,Q3] | 7 [5,16] | 8 [5,22] | 6 [5,11] | 7 [4,16] | |
| Last visit age (days), median [Q1,Q3] | 325 [281,350] | 326 [274,349] | 319 [284,348] | 327 [289,352] | |
Abbreviation: NCH, Nationwide Children’s Hospital.
Figure 1.

Flowchart showing the data filtering and splitting process at Nationwide Children’s Hospital.
Maternal postpartum cohort (OSUWMC)
For a subset of infants, maternal Epic EHR records were available within OSUWMC that were linked to the infant EHR as described previously.16,17 We extracted maternal notes from delivery through 365 days postpartum to capture lactation documentation that is often unavailable in the infant chart during the delivery hospitalization. The cohort included 23 mothers (26 deliveries) and 720 OSUWMC EHR notes across 7 note types (progress notes, lactation notes, labor, and delivery notes, discharge summaries, emergency department notes, operating room pre-operative notes, and operating room nursing notes). Lactation documentation was concentrated between delivery and maternal discharge; we restricted modeling these 2 note types, yielding 291 notes for modeling. Both infant cohort and maternal postpartum cohorts were accessed under an sIRB approved protocol, Breast Feeding Phenotype from Electronic Medical Record (2024H0282).
Preprocessing and splits
Infant notes were minimally normalized (eg, XML entity replacement and sentence boundary normalization). Infants were split into development (50%), validation (25%), and test (25%) sets; all notes for an infant were assigned to the same split.
Maternal notes often exceeded the 512-token input limit of BERT models. We partitioned each selected maternal note into contiguous segments using a character-based strategy that preserved original note order while reducing the likelihood of splitting semantically related content.18,19 Partitions averaged 333 tokens and served as the modeling unit for both the BERT and LLM pipelines. This produced 616 partitions from 291 notes. Splits were performed at the delivery level so that all partitions from a delivery were contained within a single split to prevent leakage.
Reference labels and annotation
Infant notes: LLM-assisted weak supervision
Manual labeling of 16 288 infant notes was infeasible, so we used the Anthropic Claude Sonnet 3.5 LLM20 to label each note as 1 of set {Positive, Negative, Absent}. To support downstream use of these labels for cohort selection in prospective pharmacokinetic studies, we decided to apply coarse labels indicating whether the note contained textual evidence for the receipt of breast milk. Positive indicated evidence that the infant is currently receiving breast milk (direct or expressed; exclusive or mixed). Negative indicated evidence that the infant is not currently receiving breast milk; as a simplifying assumption, mention of current formula feeding without mention of breast milk was treated as negative. Absent indicated no relevant feeding information or ambiguity insufficient to infer the current status. We validated and refined labeling via manual review of 1045 notes from 20 infants (10 with at least 1 Positive note and 10 with no Positive notes). Four data scientists (C.G., A.A.A., P.W., and J.L.) did the primary review and a pediatric epidemiologist (S.A.K.) who specializes in mother–infant nutrition and breastfeeding outcomes provide assistance and reviewed a subset. Discordances were resolved using an iterative process that modified prompts and clarified definitions until concordance between manual and LLM labels exceeded 98% per class. The resulting LLM labels were used as reference labels for pipeline development and evaluation. Claude Sonnet 3.5 was used only for weakly supervised labeling of the much larger infant-note corpus to rapidly generate high-concordance reference labels, but due to the cost, this method was not used for classification at scale. Anthropic models were accessed under a Business Associate Agreement that permits the use of services for processing in accordance with HIPAA requirements.
Maternal notes: manual consensus labels
For the maternal postpartum corpus, 2 annotators independently labeled each partition using the same conceptual definitions. Disagreements were resolved by discussion and guideline refinement. Because explicit negative evidence of no breastfeeding was rare during the delivery hospitalization, we merged Negative and ambiguous/insufficient categories into a single non-Positive class, yielding binary labels (Positive vs non-Positive) for modeling.
Classification pipelines
Infant document classification (medspaCy named entity and attribute recognition)
We curated a 114-term infant feeding term set and organized terms ontologically. Using medspaCy,20 we implemented a deterministic named entity and attribute recognition pipeline to detect breast milk and formula mentions and contextual attributes (eg, negation and whether feeding was current vs historical/intentional). Detected entities were mapped to parent concepts (“breast milk” or “formula”) and aggregated into a single document-level label (Positive, Negative, Absent) aligned with the reference definitions. We also evaluated an ontology-based extraction approach (ClinPhen)21 and several BERT-based document classifiers;22 these comparisons and implementation details are provided in the Supplementary material.
Maternal partition classification (bio_ClinicalBERT vs MedGemma)
For supervised learning, we fine-tuned Bio_ClinicalBERT23 to classify each maternal partition as Positive or non-Positive. For a few-shot learning LLM approach, we used MedGemma-27B24 with prompt engineering. Prompts were developed using partitions from 3 deliveries and then applied to held-out deliveries. The prompt instructed the model to identify lactation-related mentions (eg, latch, pumping, colostrum, and milk transfer) and determine whether each mention constituted evidence of current breastfeeding activity for the indexed infant.
Evaluation and deployment at-scale
For infant notes, we quantified LLM–manual agreement on the review set and evaluated pipelines against LLM reference labels on validation and test sets. Metrics included accuracy, balanced accuracy, macro F1, and class-specific precision, recall, and F1. For maternal partitions, we evaluated Bio_ClinicalBERT23 and MedGemma24 against manual consensus labels and reported class-specific precision, recall, and F1. To demonstrate scalability, we applied the infant medspaCy20 pipeline to 52 278 mother–infant dyads across OSUWMC and NCH (from years 2010-2025) in order to create a combined document-level breastfeeding label set into a structured variable for future downstream applications.
Results
Infant notes: LLM concordance and medspaCy agreement
On the infant review set (1045 notes), the LLM achieved 99.62% accuracy and 99.32% macro F1 relative to manual labels. Three notes were discordant: 2 labeled Absent by the LLM that manual reviewers labeled Positive and Negative, and 1 labeled Negative that manual reviewers labeled Absent. Across the full infant corpus, LLM-assigned labels were 81.03% Absent, 13.70% Negative, and 5.27% Positive, consistent with sparse and episodic feeding documentation in general pediatric notes.
Among automated infant pipelines, the medspaCy20 classifier achieved the strongest agreement with the LLM (Supplement). On the held-out test set, medspaCy20 achieved 94.36% overall accuracy and 90.40% balanced accuracy, with a macro precision of 89.67%, macro recall of 90.40%, and macro F1 of 89.97%. Class-specific F1 scores for Positive, Negative, and Absent each exceeded 0.80 (Figure 2).
Figure 2.

medspaCy performance vs LLM and performance by note class.
Scale-out and cross-institution use
We applied the medspaCy20 pipeline to 52 278 mother–infant dyads across OSUWMC and NCH and stored document-level breastfeeding labels as a structured variable, enabling linkage to medication orders, diagnoses, and outcomes. The pipeline was shared with OSUWMC and applied to postpartum maternal notes; predicted label distributions across NCH infant notes and OSUWMC maternal notes are shown in Figure 3.
Figure 3.

Detailed breakdown of lactation-related content across note type and time.
Maternal postpartum notes: Bio_ClinicalBERT vs MedGemma
The maternal postpartum modeling set comprised 616 partitions from 291 notes (26 deliveries), with an approximate Positive:non-Positive (negative/absent) ratio of 15:85 in the maternal notes. Bio_ClinicalBERT achieved a Positive precision of 0.80, recall of 0.83, and F1 of 0.82, and non-Positive precision of 0.96, recall of 0.95, and F1 of 0.95 on its held-out test deliveries. MedGemma-27B achieved a Positive precision/recall/F1 of 0.96 and a non-Positive precision/recall/F1 of 0.99 on held-out deliveries (Table 2). In qualitative review, Bio_ClinicalBERT was consistent on explicit breastfeeding statements, whereas MedGemma more often captured implicit or context-dependent cues (eg, latch attempts, pumping with uncertain transfer, or counseling paired with contemporaneous feeding evidence).
Table 2.
Performance on positive class (minority) comparison of BERT and LLM models on different test sets.
| Model | Test set | Precision | Recall | F1 score |
|---|---|---|---|---|
| BERT | 6 deliveries | 0.80 | 0.83 | 0.82 |
| LLM | 20 moms | 0.96 | 0.96 | 0.96 |
Abbreviations: BERT, Bidirectional Encoder Representations from Transformers; LLM, large language model.
Longitudinal breastfeeding timelines
We combined maternal partition labels and infant document labels to construct breastfeeding timelines for 23 infants. Timelines showed dense lactation documentation during the immediate postpartum hospitalization in maternal notes and sparse, intermittent feeding mentions in infant notes over the first year, indicating that maternal notes complement infant-chart documentation early in life (Figure 2).
Discussion
This Brief Communication demonstrates a practical approach for extracting breastfeeding status from EHR free text that couples LLM-assisted weak supervision with a lightweight, auditable rule-based pipeline. LLM labeling produced near-manual agreement on a dedicated review set, enabling rapid refinement of class definitions and scalable reference labeling of over 16 288 infant notes.25 Using these labels, a deterministic medspaCy20 pipeline achieved strong test performance and produced a structured variable that can support large observational studies, including linkage of breastfeeding status to medication exposures and outcomes.3
In this study, the LLM was most effective for rapidly generating high-quality reference labels where agreement with manual review was near perfect, enabling scalable annotation and refinement of class definitions. In contrast, the rule-based medspaCy pipeline was the better choice for routine, large-scale deployment, achieving strong held-out performance while remaining transparent and straightforward to run in privacy-constrained environments. Our deterministic pipeline achieved comparable performance to machine learning approaches,12 while preserving interpretability. A classification from a precise NER approach can be explained by the entities and context attributes, while the explainability of a machine learning classifier is less straightforward. Put differently, the LLM performed best as a labeling and development accelerator, whereas the rule-based pipeline performed best as the operational classifier. The main tradeoff is that rule-based systems are sensitive to local documentation conventions and may miss lexical variability (misspellings, abbreviations) or pragmatics such as negation scope, temporality, and future intentions.26,27 As a result, maintaining performance over time or across new note styles may require periodic updates to the term set and rules.
In the maternal postpartum corpus, we directly compared inference-time classifiers and found that prompt-based MedGemma24 outperformed fine-tuned Bio_ClinicalBERT,23 particularly for the minority Positive class (Table 2). In the infant corpus, we did not evaluate an LLM as the primary inference model; instead, an LLM was used to generate and validate high-concordance reference labels, and we compared deployable classifiers against those labels. In that infant comparison, the rule-based medspaCy pipeline achieved the best validation performance (see Supplementary material) and strong held-out test performance, whereas BERT-based document classifiers did not surpass the rule-based approach in this setting. Because the maternal BERT and LLM evaluations used different held-out deliveries, the maternal results should be interpreted as supportive rather than a definitive head-to-head comparison.
Limitations are important to note for this study. First, infant reference labels were produced by an LLM, and although validated with high concordance, residual errors may bias performance estimates for some downstream applications. Second, breastfeeding status was classified at the document or partition level and does not directly encode exclusivity or duration, as such. Trajectory-level inference requires temporal aggregation which may also incur reconciliation of conflicting documentation in small temporal windows. Third, maternal data were available for a small subset at a single institution and focused on the delivery-to-discharge interval, thus, while broader evaluation is needed, other types of notes were not found to be high yield and are expected to carry little additional information.
While human-derived labels are preferable, it is infeasible to manually label all notes within a hospital. LLM-based weak supervision enables the development of transparent algorithms that can extend the reach of both clinicians and LLMs to provide an evidence-based classification. Future work will align breastfeeding status trajectories with maternal prescriptions to define exposure windows, expand classification to maternal notes beyond discharge, and evaluate portability across additional health systems.
Conclusion
LLM-assisted weak supervision combined with a deployable rule-based pipeline can convert sparse breastfeeding documentation in EHR narratives into structured variables at scale. Integrating maternal postpartum and infant documentation improves temporal coverage and supports large retrospective studies of medication safety during lactation.
Supplementary Material
Acknowledgments
The authors would like to acknowledge Rajesh Ganta for acquiring data for the study.
Contributor Information
Connor Grannis, Office of Data Sciences, Nationwide Children’s Hospital, Columbus, OH 43205, United States.
Ping Wei, Department of Biomedical Informatics, The Ohio State University, Columbus, OH 43210, United States.
Austin A Antoniou, Office of Data Sciences, Nationwide Children’s Hospital, Columbus, OH 43205, United States.
David M Gordon, Office of Data Sciences, Nationwide Children’s Hospital, Columbus, OH 43205, United States.
Jinyu Xu, Information Technology Research & Innovation, Nationwide Children’s Hospital, Columbus, OH 43205, United States.
Jianing Liu, Department of Biomedical Informatics, The Ohio State University, Columbus, OH 43210, United States.
Peter White, Office of Data Sciences, Nationwide Children’s Hospital, Columbus, OH 43205, United States; The Steve and Cindy Rasmussen Institute for Genomic Medicine, Nationwide Children’s Hospital, Columbus, OH 43205, United States; Department of Pediatrics, The Ohio State University, Columbus, OH 43210, United States.
Sarah A Keim, Department of Pediatrics, The Ohio State University, Columbus, OH 43210, United States; Center for Biobehavioral Health, The Abigal Wexner Research Institute, Nationwide Children’s Hospital, Columbus, OH 43205, United States; Division of Epidemiology, The Ohio State University, Columbus, OH 43210, United States.
Lang Li, Department of Biomedical Informatics, The Ohio State University, Columbus, OH 43210, United States.
Christopher W Bartlett, Office of Data Sciences, Nationwide Children’s Hospital, Columbus, OH 43205, United States; The Steve and Cindy Rasmussen Institute for Genomic Medicine, Nationwide Children’s Hospital, Columbus, OH 43205, United States; Department of Pediatrics, The Ohio State University, Columbus, OH 43210, United States.
Author contributions
Connor Grannis (Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing—original draft), Ping Wei (Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Writing—original draft, Writing—review & editing), Austin A. Antoniou (Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Writing—original draft), David M. Gordon (Supervision), Jinyu Xu (Project administration, Supervision), Jianing Liu (Project administration, Supervision), Peter White (Resources, Supervision, Writing—review & editing), Sarah A. Keim (Conceptualization, Validation), Lang Li (Conceptualization, Funding acquisition, Resources, Supervision, Writing—review & editing), and Christopher W. Bartlett (Conceptualization, Funding acquisition, Project administration, Resources, Supervision, Writing—review & editing)
Supplementary material
Supplementary material is available at JAMIA Open online.
Funding
This work was supported by P30HD106451 (Support to Li). This publication was also supported, in part, by The Ohio State University Clinical and Translational Science Institute (CTSI) and the National Center for Advancing Translational Sciences of the National Institutes of Health under Grant Number UM1TR004548. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Conflicts of interests
The authors have no conflicts to disclose.
Data availability
Due to the highly sensitive nature of clinical text, the underlying data for this study cannot be shared.
References
- 1. Byrne JJ, Spong CY. “Is it safe?”: the many unanswered questions about medications and breast-feeding. Obstetric Anesthesia Digest. 2019;39:157. 10.1097/01.aoa.0000575244.99434.4b [DOI] [PubMed] [Google Scholar]
- 2. Goodrum H, Gudala M, Misra A, et al. Extraction of lactation frames from drug labels and LactMed. In: Proceedings of the 18th BioNLP Workshop and Shared Task. Florence, Italy: Association for Computational Linguistics 2019:191-200.
- 3. Amir LH, Grzeskowiak LE, Kam RL, et al. Ethical issues in use of medications during lactation. J Hum Lact. 2020;36:34-39. 10.1177/0890334419888156 [DOI] [PubMed] [Google Scholar]
- 4.Centers for Disease Control and Prevention. Breastfeeding Report Card United States, 2022. Centers for Disease Control and Prevention; 2022. Accessed October 26, 2026. https://www.cdc.gov/breastfeeding-data/media/pdfs/2024/06/2022-Breastfeeding-Report-Card-508.pdf
- 5. World Health Organization. Infant and young child feeding: model chapter for textbooks for medical students and allied health professionals. 2025. https://iris.who.int/server/api/core/bitstreams/2966cbcb-d5a0-4ee5-b39f-4dc1bde543ba/content [PubMed]
- 6. Schirm E, Schwagermann MP, Tobi H, et al. Drug use during breastfeeding. A survey from The Netherlands. Eur J Clin Nutr. 2004;58:386-390. 10.1038/sj.ejcn.1601799 [DOI] [PubMed] [Google Scholar]
- 7. Stultz EE, Stokes JL, Shaffer ML, et al. Extent of medication use in breastfeeding women. Breastfeed Med. 2007;2:145-151. 10.1089/bfm.2007.0010 [DOI] [PubMed] [Google Scholar]
- 8. Byrne JJ, Saucedo AM, Spong CY, et al. Evaluation of drug labels following the 2015 pregnancy and lactation labeling rule. JAMA Netw Open. 2020; 3:e2015094. 10.1001/jamanetworkopen.2020.15094 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. UC San Diego. Breastfeeding measurement in the outpatient electronic health record: current practices and future possibilities. UC San Diego; 2016. [Google Scholar]
- 10. Wang Y, Wang L, Rastegar-Mojarad M, et al. Clinical information extraction applications: a literature review. J Biomed Inform. 2018;77:34-49. 10.1016/j.jbi.2017.11.011 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Jensen K, Soguero-Ruiz C, Oyvind Mikalsen K, et al. Analysis of free text in electronic health records for identification of cancer patient trajectories. Sci Rep. 2017;7:46226. 10.1038/srep46226 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Lemas DJ, Du X, Rouhizadeh M, et al. Classifying early infant feeding status from clinical notes using natural language processing and machine learning. Sci Rep. 2024;14:7831. 10.1038/s41598-024-58299-x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Klebanoff MA, Fried P, Yeates KO, et al. Lifestyle and early achievement in families (LEAF) study: design of an ambidirectional cohort study of prenatal marijuana exposure and child development and behaviour. Paediatr Perinat Epidemiol. 2020;34:744-756. 10.1111/ppe.12693 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Huerta TR, Bartlett CW, Alain G, et al. Operationalizing a research-oriented learning healthcare system across covered entities: cross-institutional strategies and innovations. NPJ Health Syst. 2025;2:47. 10.1038/s44401-025-00051-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Pollard TJ, Johnson AEW, Raffa JD, et al. Tableone: an open source python package for producing summary statistics for research papers. JAMIA Open. 2018;1:26-31. 10.1093/jamiaopen/ooy012 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Rogerson CM, Bartlett CW, Price J, et al. Derivation and validation of an algorithm for maternal–child linkage in electronic health records. J Am Med Inform Assoc. 2026;33:451-456. 10.1093/jamia/ocaf177 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Epic Systems Corporation. Epic. 2025.
- 18. Pollard TJ, Johnson AEW, Raffa JD, et al. Tableone: an open source python package for producing summary statisti cs for research papers. JAMIA Open. 2018;1:26-31. 10.1093/jamiaopen/ooy012 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Kong H-J. Managing unstructured big data in healthcare system. Health Inform Res. 2019;25:1-2. 10.4258/hir.2019.25.1.1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Anthropic. Claude. 2024.
- 21. Eyre H, Chapman AB, Peterson KS, et al. Launching into clinical space with medspaCy: a new clinical text processing toolkit in Python. In: AMIA Annual Symposium Proceedings, 2022:438. [PMC free article] [PubMed]
- 22. Deisseroth CA, Birgmeier J, Bodle EE, Undiagnosed Diseases Network, et al. ClinPhen extracts and prioritizes patient phenotypes directly from medical records to expedite genetic disease diagnosis. Genet Med. 2019;21:1585-1593. 10.1038/s41436-018-0381-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Warner B, Chaffin A, Clavié B, et al. Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference. In: Proceedings of the 63rd annual meeting of the association for computational linguistics (volume 1: Long papers) 2025:2526-2547.
- 24. Alsentzer E, Murphy J, Boag W, Weng WH, Jindi D, Naumann T, McDermott M. Publicly available clinical BERT embeddings. In: Proceedings of the 2nd clinical natural language processing workshop 2019:72-78. [Google Scholar]
- 25. Sellergren A, Kazemzadeh S, Jaroensri T, et al. Medgemma technical report. arXiv preprint arXiv:2507.05201. 2025.
- 26. Hsu E, Roberts K. Leveraging large language models for knowledge-free weak supervision in clinical natural language processing. Sci Rep. 2025;15:8241. 10.1038/s41598-024-68168-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Chiticariu L, Li Y, Reiss F. 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 2013:827-832. [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Due to the highly sensitive nature of clinical text, the underlying data for this study cannot be shared.
