The DOI matches its Crossref record; 80 of the 80 references it cites were found in a registry or an index; 24 of the 47 sources its claims cite were checked for support.
Retrieval-Augmented Generation (RAG) in Healthcare: A Comprehensive Review
Open accessPublished version of a preprint
Abstract
Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by integrating external knowledge retrieval to improve factual consistency and reduce hallucinations. Despite growing interest, its use in healthcare remains fragmented. This paper presents a Systematic Literature Review (SLR) following PRISMA guidelines, synthesizing 30 peer-reviewed studies on RAG in clinical domains, focusing on three of its most prevalent and promising applications in diagnostic support, electronic health record (EHR) summarization, and medical question answering. Persistent challenges are identified, including retrieval noise, domain shift, generation latency, and limited explainability.
Overview
CoChat Summary
TL;DR
This systematic review synthesizes 30 peer-reviewed studies on retrieval-augmented generation (RAG) in healthcare, focusing on diagnostic support, EHR summarization, and medical question answering. It argues that RAG improves factual grounding and interpretability, but healthcare deployment is still limited by retrieval noise, latency, explainability, privacy, and drift.
Key findings
- The review synthesized 30 peer-reviewed studies and organized them into healthcare use cases including diagnostic support, EHR summarization, medical question answering, chatbots, and specialty-specific systems.
- It identifies persistent challenges in healthcare RAG as retrieval noise, domain shift, generation latency, and limited explainability; later sections also add privacy, multimodal limitations, infrastructure constraints, knowledge drift, bias/fairness, and lack of human oversight.
- The paper highlights clinical evaluation metrics that are more suitable than generic NLP metrics, naming FactScore, RadGraph-F1, and MED-F1 as especially important for factual accuracy, medical validity, and clinical relevance.
- The authors conclude that Advanced RAG appears “most ready for pilot clinical use,” while Modular RAG has long-term potential but is harder to deploy because of higher complexity and computational cost.
Methods
- Systematic literature review following PRISMA 2020 guidelines, prospectively registered in OSF.
- Searched PubMed, IEEE Xplore, Scopus, Web of Science, Google Scholar, and ACM Digital Library for publications from 2022 to 2025 using RAG/healthcare/LLM-related search terms.
- Screening flow: 123 records identified, 46 duplicates removed, 77 screened by title/abstract, 47 excluded, and 30 full-text studies included.
- Studies were classified using a taxonomy of RAG applications and compared across architectural variants, applications, metrics, and challenges.
Limitations
- The paper states that the evolving terminology around RAG may have led to exclusion of some studies.
- It states the review relied solely on reported findings rather than raw experimental data and was constrained by a narrow publication window, which may limit completeness and generalizability.
- Evident from the text: the synthesis depends on English-language, peer-reviewed or high-quality preprint sources only, which the paper itself identifies as a source of selection bias.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
20 Valid · 25 Not valid · 23 Not checked · 119 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, these models required structured inputs and extensive manual feature engineering to extract meaningful signals from clinical text [13].”
What the sources say
- Partly supportedcites [13] Spasic, 2020, Clinical text data in machine learning: Systematic review, JMIR Med. Inform. sourceExperimental — accuracy not yet measured
The Results (Availability; page not provided) say that most reviewed studies relied on manual annotation of clinical text, and the Discussion identifies annotation as a key bottleneck. Manual annotation is not the same as manual feature engineering, however, and the review does not say that machine-learning models required structured inputs. (LLM, conf=0.83)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Introduced in 2020, Retrieval-Augmented Generation (RAG) mitigates these limitations by combining a retriever which accesses external sources such as clinical guidelines or EHRs with a generator that conditions responses on both the query and retrieved content [8].”
What the sources say
- Partly supportedcites [8] Lewis, 2020, Retrieval-augmented generation for knowledge-intensive nlp tasks, Adv. Neural Inf. Process. Syst. sourceExperimental — accuracy not yet measured
The Introduction (p. 2) and Methods (p. 2) describe a retriever that obtains documents and a generator conditioned on both the input and those documents; the Discussion (p. 9) reports benefits over a purely parametric model. The paper’s experiments use Wikipedia as the knowledge source (Experiments, p. 4), not clinical guidelines or EHRs, so the cited paper does not support those examples. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“While interest in RAG continues to grow, most prior work has focused either on architectural innovations [10] or on isolated components such as retriever and generator design [11].”
What the sources say
- Partly supportedcites [10] Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Wang, H., and Wang, H. (2023). Retrieval-augmented generation for large language models: A survey. arXiv. sourceExperimental — accuracy not yet measured
The Introduction describes Retrieval and Generation as core components studied in the survey, and the Modular RAG section gives examples of refining retrievers and fine-tuning generators. But the paper also discusses interactions between modules; these passages do not establish that most prior work studied the components in isolation. The architectural-innovations clause is attached to a different citation and is not judged here. (LLM, conf=0.82)
- Partly supportedcites [11] Zhao, P., Zhang, H., Yu, Q., Wang, Z., Geng, Y., Fu, F., Yang, L., Zhang, W., Jiang, J., and Cui, B. (2024). Retrieval-Augmented Generation for AI-Generated Content: A Survey. arXiv. sourceExperimental — accuracy not yet measured
The Background (p. 1) says prior RAG enhancements include optimizations of individual components, and the Preliminary section (p. 3) discusses retrievers and generators. But the Background also describes holistic pipeline enhancements; the paper does not establish that most prior work focused on isolated components. The architectural-innovations clause is attached to a different citation and is not assessed here. (LLM, conf=0.84)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“medical domain-specific LLMs such as Medical Pathways Language Model (Med-PaLM) [2] and Biomedical Generative Pretrained Transformer (BioGPT) [3] have shown improved performance across various clinical natural language processing (NLP) tasks, including question answering, summarization, entity recognition, relation extraction, document classification, and literature mining.”
What the sources say
- Partly supportedcites [2] Singhal, 2025, Toward expert-level medical question answering with large language models, Nat. Med. sourceExperimental — accuracy not yet measured
The Abstract (page unavailable) reports improved Med-PaLM 2 performance on medical question-answering benchmarks, supporting the question-answering part of the claim. The cited paper does not establish improved Med-PaLM performance on the other listed tasks—summarization, entity recognition, relation extraction, document classification, or literature mining. BioGPT has a separate citation in the citing sentence and is not assessed here. (LLM, conf=0.86)
- Partly supportedcites [3] Luo, R., Sun, L., Xia, Y., Qin, T., Zhang, S., Poon, H., and Liu, T.Y. (2022). BioGPT: Generative pre-trained transformer for biomedical text generation and mining. Briefings Bioinform., 23. sourceExperimental — accuracy not yet measured
The Introduction (p. 1) identifies BioGPT as a domain-specific model and reports better performance on biomedical question answering, relation extraction, document classification, and text generation. The Experiments section (p. 4) lists those evaluated tasks, but not summarization or entity recognition; its biomedical-literature evaluations also do not establish improved performance across clinical NLP tasks generally. This verdict concerns BioGPT, not Med-PaLM, which has a separate citation. (LLM, conf=0.88)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Variants of Bidirectional Encoder Representations from Transformers (BERT) [5], such as Biomedical BERT (BioBERT) [6] and ClinicalBERT [7], improve the contextual understanding of medical terminology.”
What the sources say
- Not checked — no full text of the sourcecites [5] Devlin, J., Chang, M.W., Lee, K., and Toutanova, K. (2019, January 2–7). Bert: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Minneapolis, MN, USA. source
- Supportedcites [6] Lee, 2020, BioBERT: A pre-trained biomedical language representation model for biomedical text mining, Bioinformatics source
The Introduction describes BioBERT as an adaptation of BERT to biomedical language, where specialized terms make general-domain models less effective. The Conclusion reports that BioBERT outperforms previous models on biomedical text-mining tasks. Citation [6] is attached to BioBERT; the separate ClinicalBERT citation [7] is not a claim this source needs to support. (LLM, conf=0.90)
- Supportedcites [7] Huang, K., Altosaar, J., and Ranganath, R. (2019). ClinicalBert: Modeling Clinical Notes and Predicting Hospital Readmission. arXiv. source
Our copy identifies ClinicalBERT as a BERT variant specialized to clinical notes ('ClinicalBERT is bert [11] specialized to clinical notes', 1.2 Significance, p. 2; Abstract, p. 1) and shows it improves contextual understanding of clinical/medical terminology: it outperforms bert on clinical language modeling (Table 1, p. 3), and its contextual embeddings correlate with physician-rated similarity of medical concepts (0.670 vs 0.553 Word2Vec and 0.487 FastText, Table 2, p. 4; §3.2.2, p. 4 'ClinicalBERT captures some clinical semantics'). The Background (p. 1) also motivates contextual BERT representations as needed to capture clinical meaning that local-context models miss. BioBERT itself is not discussed in this paper, but that element of the sentence carries its own separate citation ([6]), so it is not part of what is attributed here. (LLM, conf=0.85)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
36 more claims · 2 with their cited sources checked
Create a free account to see all 187 claims and to verify the 13 references still open.
- It may state two separate claims as one. 85
- Probably not a substantial claim. 119
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“RAG-based knowledge grounding in all three agents reduces hallucination rates and supports factual reliability claims that would be required in device submissions (Neha et al., 2025).”
From A generative AI multi-agent framework with integrated XAI governance for cancer diagnostics: from multi-omics interpretation to lifestyle risk stratification (2026) · cites this paper as Neha et al., 2025
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“The LLM component, grounded in clinical epidemiology databases through RAG, provides causal annotation for each SHAP-ranked factor, distinguishing established causal risk factors from confounded associations and calibrating recommendation strength accordingly (Neha et al., 2025).”
From A generative AI multi-agent framework with integrated XAI governance for cancer diagnostics: from multi-omics interpretation to lifestyle risk stratification (2026) · cites this paper as Neha et al., 2025
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“It has been observed that LLMs are effective zero-shot or one-shot clinical information extractors.6”
From A Scalable Method for Validated Data Extraction from Electronic Health Records with Large Language Models (2025) · cites this paper as [6]
Experimental — accuracy not yet measured
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
BioBERT: a pre-trained biomedical language representation model for biomedical text mining
Lee, J., Yoon, W., Kim, S., Kim, D., Kim, S., So, C. H., & Kang, J. (2020) · Bioinformatics · doi:10.1093/bioinformatics/btz682 · Cited by 7,218 · cited 1×
“Variants of Bidirectional Encoder Representations from Transformers (BERT) [5], such as Biomedical BERT (BioBERT) [6] and ClinicalBERT [7], improve the contextual understanding of medical terminology.”
Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network
Sherstinsky, A. (2020) · Physica D: Nonlinear Phenomena · doi:10.1016/j.physd.2019.132306 · Cited by 5,680 · cited 1×
“Deep learning further advanced the field by enabling end-to-end processing of unstructured data, with early implementations using recurrent neural networks (RNNs) and Long Short-Term Memory (LSTM) networks [14].”
Dense passage retrieval for open-domain question answering
Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., & Yih, W.-T. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.550 · Cited by 3,575 · cited 1×
“Popular dense retrievers include dense passage retrieval (DPR) [21].”
MIMIC-IV, a freely accessible electronic health record dataset
Johnson, A. E. W., Bulgarelli, L., Shen, L., Gayles, A., Shammout, A., Horng, S., Pollard, T. J., Hao, S., Moody, B., Gow, B., et al. (2023) · Scientific Data · doi:10.1038/s41597-022-01899-x · Cited by 3,401 · cited 2×
“When pretrained on large-scale biomedical corpora such as the medical information mart for intensive care (MIMIC)-III clinical data [17] or medical literature, LLMs acquire domain-specific language understanding.”
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-T., Rocktäschel, T., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.11401 · Cited by 3,101 · cited 1×
“Introduced in 2020, Retrieval-Augmented Generation (RAG) mitigates these limitations by combining a retriever which accesses external sources such as clinical guidelines or EHRs with a generator that conditions responses on both the query and retrieved content [8].”
Domain-specific language model pretraining for biomedical natural language processing
Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., & Poon, H. (2022) · ACM Transactions on Computing for Healthcare · doi:10.1145/3458754 · Cited by 2,170 · cited 1×
“These systems use biomedical embeddings (e.g., BioBERT, PubMedBERT [30]), specialized corpora (e.g., PubMed abstracts and clinical notes), and domain-specific knowledge bases (e.g., unified medical language system (UMLS), MIMIC).”
Toward expert-level medical question answering with large language models
Singhal, K., Tu, T., Gottweis, J., Sayres, R., Wulczyn, E., Amin, M., Hou, L., Clark, K., Pfohl, S. R., Cole-Lewis, H., et al. (2025) · Nature Medicine · doi:10.1038/s41591-024-03423-7 · Cited by 1,734 · cited 1×
“To address these challenges, medical domain-specific LLMs such as Medical Path-ways Language Model (Med-PaLM) [2] and Biomedical Generative Pretrained Transformer (BioGPT) [3] have shown improved performance across various clinical natural language processing (NLP) tasks, including question…”
Research works printed but not linked: 3
- Gaur, M. (2022). Knowledge-Infused Learning. [Ph.D. Thesis, University of South Carolina].
- Kirubakaran, S.S., Kathrine, J.W.G., Kanaga, G.M., Raja, M.J., Singh, R.G., and Yuvaraajan, E. (2024, January 24–26). A RAG-based Medical Assistant Especially for Infectious Diseases. Proceedings of the 2024 International Conference on Inventive Computation Technologies (ICICT), Lalitpur, Nepal.
- Yu, H., Guo, P., and Sano, A. (2023, January 10). Zero-Shot ECG Diagnosis with Large Language Models and Retrieval-Augmented Generation. Proceedings of the 3rd Machine Learning for Health Symposium, PMLR, Orleans, LA, USA.
Cited by
Simou, E. (2025) · International Medical Education · doi:10.3390/ime4040050 · Cited by 35
“By retrieving verifiable evidence and displaying inline citations, RAG architectures further enhance epistemic trust and accountability, enabling learners to appraise the provenance, coherence, and ethical dimensions of AI-generated reasoning critically [159,160].”
InfluentialWołk, K. (2025) · Electronics · doi:10.3390/electronics14214227 · Cited by 26
“The system requires both retrieval and generation operations to finish within 2 s for clinical workflow integration [7].”
Stuhlmiller, T. J., Rabe, A., Rapp, J., Awawda, A., Kouser, H., Lui, K., Salamon, H., Chuyka, D., Mahoney, W., Furgason, J. M., et al. (2025) · medRxiv · doi:10.1101/2025.02.25.25322898 · Cited by 13
Gkousis, S., Vasilaki, V., & Katsou, E. (2025) · Renewable and Sustainable Energy Reviews · doi:10.1016/j.rser.2025.116577 · Cited by 14
Saidu, F. A., & Wall, J. (2026) · Electronics · doi:10.3390/electronics15030555 · Cited by 9
“In line with the G7 AI Code of Conduct, which advocates continuous risk assessment and transparency, the study prioritised transparency in its methodologies and results [36,37].”
Metadata
- Authors
- Neha, F., Bhati, D., & Shukla, D. K.
- Affiliations
- Kent State University (US)
- Rutgers, The State University of New Jersey (US)
- Corresponding author
- Neha, Fnu
- Venue
- AI
- Publisher
- MDPI AG
- Published
- 2025-09-11
- Volume
- 6
- Issue
- 9
- Pages
- 226
- ISSN
- 2673-2688 (ISSN-L)
- OpenAlex
- W4414128336
- Cited by
- 117
- References
- 55
- Citations by year
- 2026: 103 · 2025: 14
- FWCI
- 27.92
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- AI (publishedVersion, cc-by)
- Preprints.org (acceptedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Kent State University’s Open Access APC Support Fund
- Topics
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
- Keywords
- retrieval-augmented generation, health care, large language models, hallucination mitigation, medical question answering, domain shift, explainability
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Retrieval-Augmented Generation (RAG) in Healthcare: A Comprehensive Review. https://app.cochat.ai/papers/10.3390/ai6090226
- The paper
- Neha, F., Bhati, D., & Shukla, D. K. (2025). Retrieval-Augmented Generation (RAG) in Healthcare: A Comprehensive Review. AI. https://doi.org/10.3390/ai6090226
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.