DR.EHR: Dense Retrieval for Electronic Health Record with Knowledge Injection and Synthetic Data

Huaiyuan Ying; Sheng Yu; Yue Zhong; Zhengyun Zhao

REVIEW 3 major objections 7 minor 15 references

DR.EHR: Dense Retrieval for Electronic Health Record with Knowledge Injection and Synthetic Data

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 110M-parameter retriever trained on MIMIC-IV notes can beat much larger and proprietary models on the CliniQ EHR benchmark.

desk verdict Well-built pipeline, but the SOTA claims rest on an unaddressed MIMIC-IV/MIMIC-III training/evaluation overlap that could inflate every headline number. read the letter →

arxiv 2507.18583 v1 pith:43LALAI6 submitted 2025-07-24 cs.IR cs.AIcs.CL

Zhengyun Zhao , Huaiyuan Ying , Yue Zhong , Sheng Yu This is my paper · ORCID

classification cs.IRcs.AIcs.CL

keywords electronichealthrecordretrievaldenseknowledgeinjectionbiomedicalgraphsynthetictrainingdataMIMIC-IVCliniQbenchmarkcontrastivelearning

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 aims to show that a retriever purpose-built for electronic health records can close the semantic gap that defeats keyword search and general-purpose embedding models. The authors propose DR.EHR, a dense retriever (a model that ranks text by vector similarity) trained in two stages on MIMIC-IV discharge summaries: first, contrastive knowledge injection that aligns note chunks with medical entities, their abbreviations, synonyms, hypernyms, and related concepts from the BIOS biomedical knowledge graph; second, fine-tuning on large-language-model-generated disease, procedure, and drug query entities. On the CliniQ benchmark, the 110M-parameter DR.EHR-small reaches 89.12 MAP for single-patient retrieval and 64.11 Recall@100 for multi-patient retrieval, beating a proprietary embedding model and 7B-parameter baselines; the 7B DR.EHR-large reaches 67.04 Recall@100. The paper reads these results as evidence that the pipeline supplies both the medical knowledge and the query diversity that existing models lack.

What carries the argument

The load-bearing object is a two-stage contrastive training pipeline built around automatically generated positive examples. In Stage I, each 100-word chunk of a MIMIC-IV discharge summary is the anchor, and positive samples are string-matched entities from BIOS, full names produced by abbreviation reduction, and for each such entity up to two synonyms, two hypernyms, two related entities, and a random synonym of each, averaging 137.9 positives per chunk. In Stage II, the same chunks are paired with synthetic disease, procedure, and drug entities generated by a large language model, adding about 15.8 positives per chunk. Both stages optimize Multi-Similarity Loss with in-batch negatives. The machinery matters because it replaces expensive human relevance judgments with large-scale, knowledge-expanded supervision, letting a 110M encoder learn medical term associations that general-domain and biomedical retrievers trained on other corpora do not capture.

What would settle it

Run a near-duplicate search, using exact chunk overlap or embedding similarity, between the 5.8M MIMIC-IV training chunks and CliniQ's 16.5k MIMIC-III evaluation chunks; then retrain DR.EHR-small after excluding any matching admissions and check whether the reported 89.12 MAP and 64.11 Recall@100 margins survive.

Watch

Extended reading notes

Core claim

DR.EHR is a family of dense retrievers for electronic health records, and its central discovery is that a two-stage training pipeline can manufacture the large-scale, medically grounded relevance data that EHR retrieval was missing. Stage I turns each 100-word chunk of a discharge summary into an anchor and treats as positive examples the medical entities found in the chunk, the full names of their abbreviations, and the synonyms, hypernyms, and related entities that the BIOS knowledge graph associates with them. Stage II uses a large language model to generate disease, procedure, and drug entities that are explicitly mentioned in or inferable from each chunk, following Doc2Query, and trains the model on those synthetic queries. Trained with Multi-Similarity Loss and in-batch negatives, DR.EHR-small outperforms all compared dense retrievers on CliniQ, including the proprietary text-embedding-3-large and 7B models, with near-perfect string matching and the largest gains on abbreviation, synonym, hyponym, and implication matches; DR.EHR-large improves further mainly in multi-patient retrieval.

Load-bearing premise

The central result assumes that the MIMIC-IV discharge summaries used for training contain no near-copies of the MIMIC-III patient summaries used in CliniQ, and the paper reports no de-duplication or patient-exclusion step.

Editorial extensions

If this is right

A 110M-parameter retriever can outperform much larger and proprietary models on EHR retrieval, which makes high-accuracy clinical search feasible at lower serving cost.
The two-stage recipe supplies the missing training data for EHR retrieval, so other clinical text collections could be turned into retrieval training corpora without manual relevance annotation.
Removing the knowledge-injection stage drops semantic-match scores by roughly five points and Recall@100 from 64.11 to 60.42, so KG-backed positive samples, not model size, carry much of the gain.
Models trained only on single-entity queries still transfer to natural-language questions, including multi-entity ones, as shown by the MRR improvements on adapted WhyQA, emrQA, and RxWhyQA sets.
Combining different synthetic query types in Stage II improves all query types beyond what each type's data alone achieves, suggesting positive transfer across disease, procedure, and drug queries.

Reading between the lines

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

A direct test of the contamination question would be to retrain the pipeline after deleting training chunks that are near-duplicates of CliniQ evaluation chunks; if the margin collapses, part of the reported gain is memorization rather than retrieval skill.
The same pipeline could be applied to other clinical note types and languages, since nothing in the method depends on MIMIC-IV specifically beyond the availability of entity-indexed notes and a knowledge graph.
The '1+1+1>3' transfer effect suggests that multi-task synthetic data may be more valuable than simply maximizing positives per note; a curriculum that deliberately mixes query types could be explored further.
If the gains hold under de-duplication, the result suggests that parameter count is secondary to task-aligned training signal for specialized retrieval, challenging the assumption that 7B embeddings are needed for clinical search.

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.

Desk Editor's Note

read the letter

DR.EHR is a well-engineered, clearly written paper. The two-stage pipeline—BIOS knowledge-graph expansion and abbreviation reduction for Stage I, Llama-derived Doc2Query positives for Stage II—is a sensible combination of established ingredients, and the ablations genuinely support each component's contribution. The QA generalization experiments, though supplementary, are a thoughtful extension, and the manual validation of synthetic data (85% clinically acceptable) is a good sign. This is the first EHR-specific dense retriever trained on MIMIC-IV discharge summaries, so the models themselves are new artifacts.

The soft spot is the one the review process should force out into the open. The central SOTA claim depends on the 5.8M training chunks being effectively disjoint from CliniQ's 16.5k MIMIC-III evaluation chunks. MIMIC-IV and MIMIC-III cover overlapping admission periods at the same hospital, and the paper reports no patient-exclusion, de-duplication, or near-duplicate analysis. Given 100-word chunks with 10-word overlap, identical chunks can easily appear in both sets. Because Stage II generates queries from the chunk itself, a near-copy in training gives the model the exact query-chunk pair to memorize. The largest gains are exactly in string and abbreviation matches, which is where memorization helps most, though semantic matches also improve. The supplementary n2c2-derived QA sets are MIMIC-III notes, so they inherit the same concern. The authors also claim to release models without publishing any code or data artifact link; that is a smaller but real reproducibility gap.

I don't think the paper is fraudulent. The ablations are consistent, the writing is careful, and the limitation section is honest about evaluation scope. But the magnitude of the reported gains—from 80.21 to 89.12 MAP, 51.54 to 64.11 R@100—is exactly what contamination would produce, and the onus is on the authors to show it did not.

This paper deserves a serious referee. The pipeline is worth publishing even if the absolute numbers come down. What it needs is an explicit analysis of train/eval overlap, ideally with a de-duplicated split, plus release of the trained models so others can verify. I would ask the editor to send it out, with the contamination issue as the central referee question. The impact is applied rather than deep, but the method is sound and the results, if verified, would matter for clinical search and cohort selection.

Referee Report

3 major / 7 minor

Summary. This paper presents DR.EHR, a series of dense retrieval models for electronic health record (EHR) retrieval. The authors propose a two-stage training pipeline based on MIMIC-IV discharge summaries: stage I extracts medical entities and injects knowledge from the BIOS knowledge graph, and stage II uses Llama-3.1-8B-Instruct to generate synthetic query-entity pairs. They train two variants (110M and 7B parameters) and evaluate on the CliniQ benchmark, reporting substantial improvements over general-domain and biomedical dense retrievers, including a 110M model that outperforms 7B general-purpose embedding models. They also report supplementary experiments on EHR QA datasets to demonstrate generalizability. The central claims are state-of-the-art performance on CliniQ and the effectiveness of each pipeline component.

Significance. If the results are robust, this is a practically important contribution: it shows that a relatively small dense retriever, trained on EHR notes with knowledge-graph injection and synthetic data, can outperform much larger general-purpose embedding models on a clinical retrieval benchmark. The paper provides useful ablations, a manual validation of synthetic data quality, and a clear training recipe. However, the headline SOTA claim rests on the assumption that the MIMIC-IV training corpus is effectively disjoint from the MIMIC-III evaluation notes in CliniQ. That assumption is not established in the manuscript, and because MIMIC-III and MIMIC-IV overlap in hospital and time period, the reported gains could be inflated by near-duplicate training and evaluation chunks. The paper also relies on a benchmark and a knowledge graph from the same research group, which limits the independence of the evaluation. These issues must be resolved before the central claims can be accepted.

major comments (3)

[Sections 3 and 4.3] The training corpus (332k MIMIC-IV discharge summaries) and the CliniQ evaluation corpus (1k MIMIC-III patient summaries) are drawn from overlapping versions of the same hospital database, yet no de-duplication or patient-exclusion step is described anywhere in the manuscript. MIMIC-III covers admissions from 2001-2012 and MIMIC-IV from 2008-2022 at the same institution, so the same admission can appear in both databases, and discharge summaries for the same patient can be near-identical. With 100-word chunks, a training chunk can be an exact or near-exact copy of an evaluation chunk. Because the positive labels in both training stages are derived locally from each chunk (string-matched entities, abbreviation reductions, and LLM-generated entities), a near-duplicate training chunk would give the model the exact query-chunk pair used at evaluation. This could plausibly explain the large gains over external baselines, especially the near-perfect string match (97.34 vs. 87.34) and abbreviation (83.37 vs. 72.13) scores in Table 4. The authors must report overlap statistics between the MIMIC-IV training notes and the CliniQ MIMIC-III notes (e.g., shared subject_id or hadm_id), conduct a near-duplicate analysis (exact or MinHash-level), and present results after excluding overlapping patients or notes. Without this, the central claim that DR.EHR-small outperforms all existing dense retrievers is not established.
[Section 5.5 and Appendix E] The supplementary QA generalization experiments (WhyQA, emrQA, RxWhyQA) are all derived from n2c2 clinical notes, which are from the same MIMIC-III source as CliniQ. The train/evaluation overlap risk therefore applies to these generalization claims as well. If the training corpus contains near-copies of these notes, the reported MRR improvements on natural-language questions would also be inflated. The authors should either exclude overlapping notes from the QA evaluation or, ideally, evaluate on notes from a different institution or a non-overlapping time period to substantiate the claim of generalizability beyond MIMIC.
[Section 4.3 and benchmark provenance] The main evaluation relies entirely on CliniQ (Zhao et al., 2025), a benchmark introduced by the same research group as the present paper (overlapping first and last authors), and the training pipeline in stage II uses the same entity taxonomy (disease, procedure, drug) and the same chunking scheme as CliniQ. This is not a circularity in the label-generation sense, but it means the evaluation is not independent of the design choices. The authors should explicitly disclose this overlap, explain how CliniQ was used (if at all) in designing the training pipeline, and provide at least one external validation result or a clear statement that no CliniQ-specific tuning was performed. This is necessary for the reader to assess the external validity of the state-of-the-art claim.

minor comments (7)

[Section 5.3] Typo: 'approxiamately' should be 'approximately'.
[Table 12] Typos in footnotes: 'abbreivation' should be 'abbreviation', and 'hypolycemic' should be 'hypoglycemic'.
[Abstract] Subject-verb agreement: 'our models significantly outperforms' should be 'our models significantly outperform'.
[Table 3] Missing space in the footnote: 'MedCPThas' should be 'MedCPT has'.
[Table 2 caption] Extra word: 'used in in the second training stage' should be 'used in the second training stage'.
[Section 4.2] The claim that 'No middle-sized models are included since they generally perform worse than bge-base' is unsupported; either provide a citation or soften the statement.
[Sections 4.2 and 6] The paper states that the models are released but provides no URL, model identifier, or link to code/checkpoints. This hampers reproducibility and should be added.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training pipeline and CliniQ evaluation labels are independent; self-citations are external resources rather than fitted targets.

full rationale

The derivation chain is self-contained with respect to the stated evaluation. DR.EHR is trained on MIMIC-IV discharge summaries, and its positive samples are produced by string matching, Llama-based abbreviation reduction, and BIOS knowledge-graph expansion; evaluation on CliniQ uses labels constructed from MIMIC-III structured codes and GPT-4o annotations, independent of the training pipeline. No equation in Section 3 defines a training target in terms of a CliniQ score, and no CliniQ metric is optimized or fitted. The use of CliniQ and BIOS, both from the same research group, is a self-citation, but it is not load-bearing in a circular sense: BIOS is an externally published knowledge graph whose relations are not derived from DR.EHR's outputs, and CliniQ labels are not produced by the model or its synthetic-data pipeline. The ablations in Table 4 demonstrate per-component contributions to independently labeled match-type categories rather than predicting those categories from fitted category definitions. The possible MIMIC-III/MIMIC-IV overlap is a data-contamination risk, not a circularity: it does not make any claimed result equivalent to its inputs by construction.

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

No new theoretical entities are introduced. The load-bearing extras are data assumptions and tuned hyperparameters, with the MIMIC-IV versus MIMIC-III overlap being the most consequential unresolved assumption.

free parameters (5)

Multi-Similarity Loss margin epsilon = 0.1
Used in Equations 1 and 2 to define informative positives and negatives; determined by grid search in Section 3.3.
MSL alpha = 2
Positive pair weighting exponent in Equation 3; grid-searched in Section 3.3.
MSL beta = 50
Negative pair weighting exponent in Equation 3; grid-searched in Section 3.3.
MSL lambda = 0.5
Pair similarity offset in Equation 3; grid-searched in Section 3.3.
KG expansion caps = at most 2 synonyms, 2 hypernyms, 2 related entities, plus 1 random synonym
Section 4.1 states these caps were chosen because a pilot study showed no significant gain from more entities; this affects the positive sample distribution.

assumptions (5)

domain assumption MIMIC-IV training summaries and MIMIC-III CliniQ evaluation summaries are disjoint
Not stated or verified; the two databases are versions of the same hospital system with overlapping admissions, as implied by Sections 4.1 and 4.3.
ad hoc to paper BIOS KG relations such as 'may treat', 'may cause', and 'is a' are valid proxies for retrieval relevance
Section 3.1 uses these relations to expand positive samples; CliniQ's match-type categories overlap with these relations, so this is partly training to the benchmark semantics.
domain assumption Llama-3.1-8B-Instruct generated entities accurately reflect relevant queries
Section 3.2 reports 85% clinical validity on 50 chunks, leaving 15% noise; the assumption is needed for stage II positives.
domain assumption CliniQ relevance labels generated by GPT-4o are accurate
Section 4.3 describes the benchmark; no manual verification rate is reported for the benchmark labels.
domain assumption MIMIC-IV is representative of EHR discharge summaries beyond the source hospital
Used to claim generalizability to EHR retrieval; no external hospital data is evaluated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DR.EHR: Dense Retrieval for Electronic Health Record with Knowledge Injection and Synthetic Data." pith.science (2026). https://pith.science/paper/43LALAI6

@misc{pith2026250718583,
  author       = {Pith},
  title        = {Pith review of: DR.EHR: Dense Retrieval for Electronic Health Record with Knowledge Injection and Synthetic Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/43LALAI6}},
  note         = {Machine review of arXiv:2507.18583}
}

read the original abstract

Electronic Health Records (EHRs) are pivotal in clinical practices, yet their retrieval remains a challenge mainly due to semantic gap issues. Recent advancements in dense retrieval offer promising solutions but existing models, both general-domain and biomedical-domain, fall short due to insufficient medical knowledge or mismatched training corpora. This paper introduces \texttt{DR.EHR}, a series of dense retrieval models specifically tailored for EHR retrieval. We propose a two-stage training pipeline utilizing MIMIC-IV discharge summaries to address the need for extensive medical knowledge and large-scale training data. The first stage involves medical entity extraction and knowledge injection from a biomedical knowledge graph, while the second stage employs large language models to generate diverse training data. We train two variants of \texttt{DR.EHR}, with 110M and 7B parameters, respectively. Evaluated on the CliniQ benchmark, our models significantly outperforms all existing dense retrievers, achieving state-of-the-art results. Detailed analyses confirm our models' superiority across various match and query types, particularly in challenging semantic matches like implication and abbreviation. Ablation studies validate the effectiveness of each pipeline component, and supplementary experiments on EHR QA datasets demonstrate the models' generalizability on natural language questions, including complex ones with multiple entities. This work significantly advances EHR retrieval, offering a robust solution for clinical applications.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 1 canonical work pages

[1]

Israel Alonso and David Contreras. 2016. Evaluation of semantic similarity metrics applied to the automatic retrieval of medical documents: An umls approach. Expert Systems with Applications, 44:386--399

2016
[2]

Rodrigo Bonacin, J \'u lio Cesar dos Reis, Edemar Mendes Perciani, and Olga Nabuco. 2018. https://api.semanticscholar.org/CorpusID:86736454 Exploring intentions on electronic health records retrieval. studies with collaborative scenarios . Ing \'e nierie des Syst \`e mes d Inf. , 23:111--135

2018
[3]

Cheriton

David R. Cheriton. 2019. https://api.semanticscholar.org/CorpusID:208612557 From doc2query to doctttttquery

2019
[4]

Hall, and Ming-Wei Chang

Zhuyun Dai, Vincent Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, and Ming-Wei Chang. 2022. https://api.semanticscholar.org/CorpusID:252519173 Promptagator: Few-shot dense retrieval from 8 examples . ArXiv, abs/2209.11755

arXiv 2022
[5]

Tracy Edinger, Aaron M Cohen, Steven Bedrick, Kyle Ambert, and William Hersh. 2012. Barriers to retrieving patient information from electronic health record data: failure analysis from the trec medical records track. In AMIA annual symposium proceedings, volume 2012, page 180. American Medical Informatics Association

2012
[6]

Jungwei Fan. 2019. https://doi.org/10.18653/v1/W19-1913 Annotating and characterizing clinical sentences with explicit why- QA cues . In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 101--106, Minneapolis, Minnesota, USA. Association for Computational Linguistics

work page doi:10.18653/v1/w19-1913 2019
[7]

Goodwin and Sanda M

Travis R. Goodwin and Sanda M. Harabagiu. 2017. https://api.semanticscholar.org/CorpusID:9689139 Knowledge representations and inference techniques for medical question answering . ACM Transactions on Intelligent Systems and Technology (TIST), 9:1 -- 26

2017
[8]

Shashi Kant Gupta, Aditya Basu, Bradley Taylor, Anai Kothari, and Hrituraj Singh. 2024. https://arxiv.org/abs/2404.06680 Onco-retriever: Generative classifier for retrieval of ehr records in oncology . Preprint, arXiv:2404.06680

arXiv 2024

Show all 15 references

[9]

Hanauer, Qiaozhu Mei, James Law, Ritu Khanna, and Kai Zheng

David A. Hanauer, Qiaozhu Mei, James Law, Ritu Khanna, and Kai Zheng. 2015. https://api.semanticscholar.org/CorpusID:15425858 Supporting information retrieval from electronic health records: A report of university of michigan's nine-year experience in developing and using the ...

2015
[10]

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

2021 arXiv
[11]

Qiao Jin, Won Kim, Qingyu Chen, Donald C Comeau, Lana Yeganova, W John Wilbur, and Zhiyong Lu. 2023. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics, 39(11):btad651

2023
[12]

Qiao Jin, Chuanqi Tan, Zhengyun Zhao, Zheng Yuan, and Songfang Huang. 2021. https://api.semanticscholar.org/CorpusID:247849810 Alibaba damo academy at trec clinical trials 2021: Exploringembedding-based first-stage retrieval with trialmatcher . In Text Retrieval Conference

2021
[13]

Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, et al. 2023. Mimic-iv, a freely accessible electronic health record dataset. Scientific data, 10(1):1

2023
[14]

Vladimir Karpukhin, Barlas O g uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906

2020 arXiv
[15]

Bevan Koopman, Guido Zuccon, Peter Bruza, Laurianne Sitbon, and Michael Lawley. 2016. Information retrieval as semantic inference: A graph inference model applied to medical search. Information Retrieval Journal, 19:6--37

2016

Pith tools

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

[1] [1]

Israel Alonso and David Contreras. 2016. Evaluation of semantic similarity metrics applied to the automatic retrieval of medical documents: An umls approach. Expert Systems with Applications, 44:386--399

2016

[2] [2]

Rodrigo Bonacin, J \'u lio Cesar dos Reis, Edemar Mendes Perciani, and Olga Nabuco. 2018. https://api.semanticscholar.org/CorpusID:86736454 Exploring intentions on electronic health records retrieval. studies with collaborative scenarios . Ing \'e nierie des Syst \`e mes d Inf. , 23:111--135

2018

[3] [3]

Cheriton

David R. Cheriton. 2019. https://api.semanticscholar.org/CorpusID:208612557 From doc2query to doctttttquery

2019

[4] [4]

Hall, and Ming-Wei Chang

Zhuyun Dai, Vincent Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, and Ming-Wei Chang. 2022. https://api.semanticscholar.org/CorpusID:252519173 Promptagator: Few-shot dense retrieval from 8 examples . ArXiv, abs/2209.11755

arXiv 2022

[5] [5]

Tracy Edinger, Aaron M Cohen, Steven Bedrick, Kyle Ambert, and William Hersh. 2012. Barriers to retrieving patient information from electronic health record data: failure analysis from the trec medical records track. In AMIA annual symposium proceedings, volume 2012, page 180. American Medical Informatics Association

2012

[6] [6]

Jungwei Fan. 2019. https://doi.org/10.18653/v1/W19-1913 Annotating and characterizing clinical sentences with explicit why- QA cues . In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 101--106, Minneapolis, Minnesota, USA. Association for Computational Linguistics

work page doi:10.18653/v1/w19-1913 2019

[7] [7]

Goodwin and Sanda M

Travis R. Goodwin and Sanda M. Harabagiu. 2017. https://api.semanticscholar.org/CorpusID:9689139 Knowledge representations and inference techniques for medical question answering . ACM Transactions on Intelligent Systems and Technology (TIST), 9:1 -- 26

2017

[8] [8]

Shashi Kant Gupta, Aditya Basu, Bradley Taylor, Anai Kothari, and Hrituraj Singh. 2024. https://arxiv.org/abs/2404.06680 Onco-retriever: Generative classifier for retrieval of ehr records in oncology . Preprint, arXiv:2404.06680

arXiv 2024