RAG-based Architectures for Drug Side Effect Retrieval in LLMs

Afshin Beheshti; Andre Daniels; Diego Galeano; Pinar Avci; Reza Rassol; Shad Nygren

REVIEW 4 major objections 5 minor 29 references

RAG-based Architectures for Drug Side Effect Retrieval in LLMs

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 GraphRAG retrieves drug side effects with near-perfect accuracy when a Neo4j graph of SIDER 4.1 associations is added to Llama 3 8B.

desk verdict The GraphRAG result is a lookup artifact—the graph is built from the same SIDER pairs used as labels—so the 0.9999 accuracy is expected by construction; the RAG Format B result is the only non-tautological number, and even that is a retrieval benchmark, not a prediction. read the letter →

arxiv 2507.13822 v1 pith:YOUZSYZV submitted 2025-07-18 cs.IR cs.AIcs.CL

Shad Nygren , Pinar Avci , Andre Daniels , Reza Rassol , Afshin Beheshti , Diego Galeano This is my paper · ORCID

classification cs.IRcs.AIcs.CL

keywords retrieval-augmentedgenerationGraphRAGdrugsideeffectspharmacovigilanceSIDER4.1Llama38Bbinaryclassification

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 proposes two ways to give Llama 3 8B reliable drug side effect knowledge: RAG over a vector database and GraphRAG over a graph of known drug side effect associations. The central claim is that GraphRAG retrieves those associations almost perfectly, with accuracy and F1 around 0.9999 on a balanced sample of 19,520 pairs spanning 976 drugs and 3,851 side effect terms. That would matter because standalone LLMs, including ChatGPT 3.5 and 4, answer such questions at little better than chance on the same test. The paper argues that structured knowledge representation, not a bigger model, is what makes LLM-based pharmacovigilance querying practical.

What carries the argument

The load-bearing component is a graph whose nodes are drugs and side effects and whose directed edges are labeled may_cause_side_effect, built directly from SIDER 4.1. An entity-recognition module extracts the drug and side effect from the user question, a Cypher query checks whether the edge exists, and a prompt-engineering module turns the graph answer into a YES/NO instruction for Llama 3 8B. The same prompt strategy is used for the RAG architecture, where the retrieved context comes from a vector database of text chunks, so the comparison isolates the role of graph-structured retrieval versus text retrieval.

What would settle it

Build the same GraphRAG system but with 10% of SIDER associations removed from the graph, then query those removed pairs: if accuracy stays near 0.9999, the LLM is contributing real biomedical knowledge; if accuracy drops to the level of the missing-edge rate, the system is purely a database lookup. A second check is to query side effects added to external adverse event reports after SIDER 4.1 was released and see whether the system correctly flags them as unknown rather than false negatives.

Watch

Extended reading notes

Core claim

The paper's central discovery is that storing SIDER 4.1 drug side effect associations as a graph and letting a Llama 3 8B model answer a binary yes/no question based on a Cypher-generated prompt yields near-perfect classification of known associations: accuracy 0.9999, F1 0.9999, precision 0.9998, sensitivity 0.9999, and specificity 0.9998. In the same evaluation, RAG over per-pair text reaches 0.998 accuracy, RAG over per-drug aggregated lists reaches 0.886, and standalone Llama 3 8B is at 0.529 accuracy with 0.092 sensitivity. The authors interpret this as evidence that augmenting LLMs with structured domain knowledge overcomes hallucination and missing-domain-knowledge limitations in pharmacovigilance.

Load-bearing premise

The evaluation assumes SIDER 4.1 is a complete and correct list of side effects, so every association it does not contain can be treated as a true negative; if SIDER has gaps or errors, the near-perfect scores partly measure how well the system echoes those gaps.

Editorial extensions

If this is right

If GraphRAG's near-perfect numbers hold in production, a clinician-facing LLM could answer 'is this side effect associated with this drug?' with lookup-level reliability for the 976 drugs and 3,851 terms covered.
Data Format B's 0.998 accuracy shows that simple per-pair text retrieval is nearly as good as graph retrieval for single-edge queries, so the marginal value of the graph layer may be small for this narrow task.
The poor standalone performance of Llama 3 8B, ChatGPT 3.5, and ChatGPT 4 suggests that model scale alone will not fix pharmacovigilance knowledge gaps; external structured knowledge is the determining factor.
The balanced evaluation design allows direct comparison of precision, sensitivity, and specificity across all methods on the same 19,520 pairs.

Reading between the lines

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

Because the graph's edges and the evaluation labels both come from SIDER 4.1, the near-perfect accuracy is best read as retrieval fidelity against that database, not as evidence that new or unreported side effects can be discovered; the paper itself acknowledges underreporting.
A sharper test would hold out a set of SIDER associations from the graph and see whether GraphRAG still answers them correctly from the LLM alone; on the current design that cannot happen, since the answer key and the graph are the same source.
The single-drug, single-side-effect question format and constrained YES/NO output leave open the clinically useful reverse queries, such as which drugs cause a given side effect, which the authors list as future work; extending the graph traversal to those queries is a natural next test.

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

Read this one so you don't have to. The headline result—GraphRAG at 0.9999 accuracy on 19,520 SIDER pairs—is circular. The graph is built from the same SIDER associations used as positive/negative labels, and the Cypher query is a direct edge lookup. The prompt then tells Llama whether the pair is known, so the LLM is just formatting a database answer. A raw Cypher lookup would score the same; the paper reports no such baseline. That's the load-bearing flaw, and it sinks the 'significant advancement' claim.

The paper does have real content. It builds two concrete retrieval pipelines (Pinecone vector RAG on two text formats, and Neo4j GraphRAG), evaluates them on a balanced subset, and ships code and data. The RAG-Format B result (0.998 accuracy) is genuine and useful: it shows that pair-level text retrieval works well for this lookup task, though even there the filtering module injects the answer into the prompt. The standalone Llama 3-8B and ChatGPT subset numbers are a nice sanity check that raw LLMs are bad at this task, but that's a known result.

The novelty is thin. The paper calls the architectures 'novel' but doesn't cite Lewis et al. on RAG or Edge et al. on GraphRAG, so the contribution reads as an engineering application rather than a method advance. The limitations paragraph is honest about SIDER incompleteness and single-drug queries, but it doesn't acknowledge the evaluation's circularity. There are no error bars across random samples, and the ChatGPT evaluation is only 51 drugs.

If the authors reframed this as a retrieval benchmark rather than a prediction system, and added a database-only lookup baseline, the RAG-Format B result could support a modest, useful paper. The GraphRAG comparison would then be against the lookup itself, not against the LLM. As submitted, the central claim is unsupported.

My call: reject, but send it to peer review rather than desk-rejecting. The evaluation flaw is fixable, the engineering is clear, and the code availability makes it reproducible. A serious referee could extract a solid benchmark paper from this. I'd cite the RAG-Format B number if I needed a reference for SIDER retrieval, but not the GraphRAG claim.

Referee Report

4 major / 5 minor

Summary. The paper describes two retrieval-augmented architectures for drug side effect identification: a text-based RAG pipeline using OpenAI embeddings in Pinecone and a GraphRAG pipeline using a Neo4j graph built from SIDER 4.1 associations, both feeding a Llama 3 8B model that returns binary YES/NO answers. The authors evaluate on 19,520 balanced drug-side effect pairs and report near-perfect GraphRAG metrics (accuracy 0.9999, F1 0.9999, precision 0.9998, sensitivity 0.9999, specificity 0.9998), with RAG Data Format B at 0.998 accuracy and standalone Llama 3-8B at 0.529 accuracy, and claim this demonstrates a significant advancement in pharmacovigilance retrieval.

Significance. If the claims were valid, the paper would show a practical way to make an LLM reliably answer drug side effect questions by grounding it in a structured knowledge graph. The architecture descriptions are clear, and the authors provide a public code repository and a balanced evaluation dataset. However, the central claim is not established because the evaluation is circular: the GraphRAG graph contains exactly the SIDER associations used to define positive and negative labels, and the Cypher query retrieves the answer key before the LLM is prompted. The reported near-perfect accuracy is therefore a lookup artifact rather than evidence of retrieval quality or LLM augmentation. The manuscript also lacks a direct graph-lookup baseline, held-out validation, or external validation, which are necessary to support the claimed contribution.

major comments (4)

[Graph-Based Retrieval Augmented Generation (GraphRAG); Performance Evaluation] The evaluation is circular by construction. The Neo4j graph edges are built from SIDER 4.1 drug-side effect associations, and the positive and negative evaluation labels are sampled from the same SIDER 4.1 database. The Cypher query MATCH (s)-[r:May_Cause_Side_Effect]->(t) WHERE s.name = 'metformin' AND t.name = 'headache' returns an edge exactly when the queried pair is in the positive label set, and the prompt then states that the drug is or is not known to be associated with the side effect. The Llama model merely formats the lookup result, so the reported 0.9999 accuracy should be expected from a direct database lookup. A baseline consisting of the Cypher query alone, without any LLM, would likely match these metrics, and the absence of such a baseline makes the claimed superiority over standalone LLMs uninformative.
[Retrieval Augmented Generation (RAG) Framework; Performance Evaluation] The near-perfect RAG Data Format B result (accuracy 0.998) suffers from the same issue in a softer form. In Format B each drug-side effect pair is one line, so the top-5 similarity retrieval on the exact query pair will almost always include that exact line, and the filtering module then tells the model whether the pair is present. The final accuracy is therefore dominated by whether the exact pair was found in the vector store, not by whether the LLM can reason about side effects. The paper should report the retrieval hit rate separately from the final binary accuracy and should include a test scenario where the correct pair is not in the retrieved chunks, to show that the LLM adds value beyond lookup.
[Performance Evaluation: Evaluation Dataset Creation; Discussion] The negative labels are defined as side effects not associated with a drug in SIDER 4.1. This treats SIDER as complete ground truth, but the paper's own Discussion acknowledges that SIDER primarily captures reported side effects and that underreporting is pervasive. If SIDER omits a true association, that pair is mislabeled as negative, and the specificity and accuracy metrics are relative to the same incomplete resource used to build the retrieval graphs. This is a load-bearing assumption for the reported near-perfect metrics. The authors should either validate on an external dataset (e.g., FAERS or post-marketing reports) or at minimum quantify how sensitive the conclusions are to label incompleteness.
[Performance Evaluation: Models Assessed] The comparison between standalone Llama 3-8B and GraphRAG is not a fair assessment of augmentation because GraphRAG receives the exact answer in its prompt. The paper needs a proper baseline: for example, a direct Neo4j lookup without an LLM, or a RAG variant where the retrieved context does not explicitly state the association. Without such a baseline, the conclusion that GraphRAG 'significantly outperforms' standalone LLMs is not supported by the data, and the 'stark contrast' in metrics is an artifact of the leakage in the experimental design.

minor comments (5)

[Abstract] The phrase 'two Retrieval Augmented Generation (RAG) and GraphRAG architectures' is grammatically unclear; the paper actually proposes one RAG and one GraphRAG architecture, so the wording should be corrected.
[Results; Figure 1b] The Results text says the raw SIDER data was 'processed into two distinct text formats' and Figure 1b is described as showing three formats, including the graph representation; please reconcile this wording, since the graph representation is a third format but not a text format.
[Figure 3 caption; Supplementary Figure 1 caption; Supplementary Figure 2 caption] The captions state that the data were obtained from 'SIDER 4.0', while the Methods and main text consistently use SIDER 4.1; the captions should be corrected for consistency.
[Performance Evaluation] There is a typo in the sentence 'This evaluation framework in in line with the drug side effect prediction literature' — 'in in' should be 'is in'.
[Performance Evaluation: Models Assessed] The ChatGPT 3.5 and ChatGPT 4 evaluation is described only as using 'a subset of 51 randomly selected drugs'; details such as the number of queries per drug, the exact prompt provided, temperature settings, and the random seed are needed for reproducibility.

Circularity Check

1 steps flagged · score 9.0 of 10

GraphRAG's near-perfect accuracy is a SIDER lookup artifact: the Neo4j edges and the evaluation labels are the same database rows, and the prompt announces the edge's presence before Llama responds.

self definitional [GraphRAG Framework section ('Graph-Based Retrieval Augmented Generation (GraphRAG) for Drug Side Effect Data') and Performance Evaluation section ('Evaluation Dataset Creation')]
"In our GraphRAG framework, drug-side effect associations are precisely modeled as a graph-based representation, leveraging the extensive Side Effect Resource (SIDER) 4.1 database... known relationships are encoded as directed edges, specifically labeled 'may_cause_side_effect'. ... If a match is found, the prompt is modified to state, 'Metformin is known to be associated with headache as a side effect'. ... a positive set was constructed by randomly sampling ten known side effects, and a negative set was generated by sampling an equal number of side effects not associated with that drug."

The Neo4j graph contains exactly the SIDER 4.1 may_cause_side_effect edges, and the evaluation's positive labels are 'known side effects' from the same database while negative labels are 'side effects not associated with that drug' from the same database. The Cypher query MATCH (s)-[r:May_Cause_Side_Effect]->(t) returns an edge if and only if the queried pair is in the label set, and the prompt then tells Llama 'known to be associated' or 'not known to be associated'. Therefore the GraphRAG output is determined by a direct lookup against the answer key before the LLM is consulted; the reported 0.9999 accuracy is expected by construction and would be matched by a raw Cypher lookup baseline.

full rationale

The central claim — GraphRAG achieves near-perfect drug side effect retrieval — is circular in construction. The graph database is populated from SIDER 4.1, the evaluation's positive and negative pairs are sampled from SIDER 4.1, and the prompt supplied to Llama states whether the SIDER edge exists. Thus the measured YES/NO response is a lookup transcript, not a prediction. No fitted parameter or learned quantity is involved, so this is not 'fitted input called prediction'; it is self-definitional: the label and the retrieval source are the same relation. The paper never reports a direct Cypher or exact-match baseline, so the 0.9999 score cannot be attributed to the RAG/GraphRAG architecture. The self-citations (refs 22, 23) are used only to justify framing the task as binary classification and are not load-bearing for the near-perfect result, so they do not add to the score. Because the near-perfect accuracy is forced by the experimental design, the circularity score is 9.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests entirely on existing data and tools: SIDER, OpenAI embeddings, Pinecone, Neo4j, and Llama 3 8B. There are no new physical or conceptual entities. The only choices that affect the result are the free parameters listed above, plus the unstated assumption that SIDER is a valid answer key.

free parameters (3)

RAG retrieval top-k = 5
The top five nearest chunks are retrieved for every query; chosen without ablation, and it affects RAG A and RAG B results.
Side effects sampled per drug = 10
Balanced evaluation builds 10 positives and 10 negatives per drug; no random seed is reported, so the subsets are not exactly reproducible.
Minimum number of side effects per drug = 10
This filter selects the 976 evaluated drugs from SIDER; changing the threshold changes the test population and the reported metrics.

assumptions (4)

domain assumption SIDER 4.1 associations are accurate, complete ground truth for drug side effects.
Used both as labels and as graph content. The paper itself notes underreporting and incomplete post-marketing data in the Discussion, so errors would propagate directly into measured accuracy.
domain assumption Side effects not listed for a drug in SIDER are true negatives.
If SIDER is incomplete, a sampled negative could be a real side effect, biasing the measured specificity and accuracy.
domain assumption The entity recognition module correctly extracts drug and side effect names from free-text queries.
Entity recognition is described but not evaluated separately; errors would corrupt both Cypher queries and the prompts sent to Llama.
domain assumption Llama 3 8B will follow the prompt instruction to answer YES or NO based only on the provided retrieved results.
The final output depends on the LLM's faithfulness to the prompt, and no faithfulness measurement is reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RAG-based Architectures for Drug Side Effect Retrieval in LLMs." pith.science (2026). https://pith.science/paper/YOUZSYZV

@misc{pith2026250713822,
  author       = {Pith},
  title        = {Pith review of: RAG-based Architectures for Drug Side Effect Retrieval in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YOUZSYZV}},
  note         = {Machine review of arXiv:2507.13822}
}

read the original abstract

Drug side effects are a major global health concern, necessitating advanced methods for their accurate detection and analysis. While Large Language Models (LLMs) offer promising conversational interfaces, their inherent limitations, including reliance on black-box training data, susceptibility to hallucinations, and lack of domain-specific knowledge, hinder their reliability in specialized fields like pharmacovigilance. To address this gap, we propose two architectures: Retrieval-Augmented Generation (RAG) and GraphRAG, which integrate comprehensive drug side effect knowledge into a Llama 3 8B language model. Through extensive evaluations on 19,520 drug side effect associations (covering 976 drugs and 3,851 side effect terms), our results demonstrate that GraphRAG achieves near-perfect accuracy in drug side effect retrieval. This framework offers a highly accurate and scalable solution, signifying a significant advancement in leveraging LLMs for critical pharmacovigilance applications.

Figures

Figures reproduced from arXiv: 2507.13822 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

[1]

Naghavi, M. et al. Global, regional, and national age-sex specific mortality for 264 causes of death, 1980–2016: a systematic analysis for the Global Burden of Disease Study 2016. The Lancet 390, 1151–1210 (2017)

work page 2017
[2]

Sunshine, J. E. et al. Association of adverse effects of medical treatment with mortality in the United States: A secondary analysis of the Global Burden of Diseases, Injuries, and Risk Factors Study. JAMA network open 2, e187041–e187041 (2019)

work page 2019
[3]

& Oshio, T

Ma, X. & Oshio, T. The impact of social insurance on health among middle-aged and older adults in rural China: a longitudinal study using a three-wave nationwide survey. BMC Public Health 20, 1842 (2020)

work page 2020
[4]

Ernst, F. R. & Grizzle, A. J. Drug-related morbidity and mortality: updating the cost-of- illness model. J Am Pharm Assoc (Wash) 41, 192–199 (2001)

work page 2001
[5]

K., Mosier, M

Ziegler, D. K., Mosier, M. C., Buenaver, M. & Okuyemi, K. How much information about adverse effects of medication do patients want from physicians? Arch Intern Med 161, 706–713 (2001)

work page 2001
[6]

Rothenberger, D. A. Physician Burnout and Well-Being: A Systematic Review and Framework for Action. Dis Colon Rectum 60, 567–576 (2017)

work page 2017
[7]

(McGrawHill, Medical Publ

Handbook of Clinical Drug Data. (McGrawHill, Medical Publ. Division, New York, 2002)

work page 2002
[8]

Tayefi, M. et al. Challenges and opportunities beyond structured data in analysis of electronic health records. WIREs Computational Statistics 13, e1549 (2021)

work page 2021

Show all 29 references

[9]

Pharmacovigilance and drug safety: Fair prescribing and clinical research

Montastruc, J.-L. Pharmacovigilance and drug safety: Fair prescribing and clinical research. Therapie 77, 261–263 (2022)

2022
[10]

& Shakir, S

Hazell, L. & Shakir, S. A. W. Under-reporting of adverse drug reactions : a systematic review. Drug Saf 29, 385–396 (2006)

2006
[11]

Mao, J. J. et al. Online discussion of drug side effects and discontinuation among breast cancer survivors. Pharmacoepidemiol Drug Saf 22, 256–262 (2013)

2013
[12]

Chang, Y. et al. A Survey on Evaluation of Large Language Models. ACM Trans. Intell. Syst. Technol. 15, 39:1-39:45 (2024)

2024
[13]

Thirunavukarasu, A. J. et al. Large language models in medicine. Nat Med 29, 1930– 1940 (2023)

2023
[14]

Landman, R. et al. Using large language models for safety-related table summarization in clinical study reports. JAMIA Open 7, ooae043 (2024)

2024
[15]

Singhal, K. et al. Large language models encode clinical knowledge. Nature 620, 172– 180 (2023)

2023
[16]

& Ramasamy, R

Golan, R., Reddy, R. & Ramasamy, R. The rise of artificial intelligence-driven health communication. Transl Androl Urol 13, 356–358 (2024)

2024
[17]

Van Veen, D. et al. Adapted large language models can outperform medical experts in clinical text summarization. Nat Med 30, 1134–1142 (2024)

2024
[18]

Kung, T. H. et al. Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models. PLOS Digit Health 2, e0000198 (2023)

2023
[19]

& Tong, W

Wang, X., Xu, X., Liu, Z. & Tong, W. Bidirectional Encoder Representations from Transformers-like large language models in patient safety and pharmacovigilance: A comprehensive assessment of causal inference implications. Exp Biol Med (Maywood) 248, 1908–1917 (2023)

2023
[20]

Liu, Z. et al. AI-based language models powering drug discovery and development. Drug Discov Today 26, 2593–2607 (2021)

2021
[21]

Sun, Z., Pergola, G., Wallace, B. C. & He, Y. Leveraging ChatGPT in Pharmacovigilance Event Extraction: An Empirical Study. Preprint at https://doi.org/10.48550/arXiv.2402.15663 (2024)

2024 doi arXiv
[22]

& Paccanaro, A

Galeano, D., Li, S., Gerstein, M. & Paccanaro, A. Predicting the frequencies of drug side effects. Nat Commun 11, 4575 (2020)

2020
[23]

& Paccanaro, A

Galeano, D. & Paccanaro, A. Machine learning prediction of side effects for drugs in clinical trials. Cell Rep Methods 2, 100358 (2022)

2022
[24]

& Radnov, A

Martin, L., Hutchens, M., Hawkins, C. & Radnov, A. How Much Do Clinical Trials Cost? (Nature Publishing Group, 2017)

2017
[25]

Kuhn, M., Letunic, I., Jensen, L. J. & Bork, P. The SIDER database of drugs and side effects. Nucleic acids research 44, D1075–D1079 (2015)

2015
[26]

Ke, Y. et al. Development and Testing of Retrieval Augmented Generation in Large Language Models -- A Case Study Report. Preprint at https://doi.org/10.48550/arXiv.2402.01733 (2024)

2024 doi arXiv
[27]

R., Mahajan, V., Khajuria, V

Tandon, V. R., Mahajan, V., Khajuria, V. & Gillani, Z. Under-reporting of adverse drug reactions: a challenge for pharmacovigilance in India. Indian J Pharmacol 47, 65–71 (2015)

2015
[28]

García-Abeijon, P. et al. Factors Associated with Underreporting of Adverse Drug Reactions by Health Care Professionals: A Systematic Review Update. Drug Saf 46, 625–636 (2023)

2023
[29]

J., Heneghan, C

Onakpoya, I. J., Heneghan, C. J. & Aronson, J. K. Post-marketing withdrawal of 462 medicinal products because of adverse drug reactions: a systematic review of the world literature. BMC medicine 14, 10 (2016). Supplementary Materials Supplementary Figure 1. Binary classificati...

2016

Pith tools

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

[1] [1]

Naghavi, M. et al. Global, regional, and national age-sex specific mortality for 264 causes of death, 1980–2016: a systematic analysis for the Global Burden of Disease Study 2016. The Lancet 390, 1151–1210 (2017)

work page 2017

[2] [2]

Sunshine, J. E. et al. Association of adverse effects of medical treatment with mortality in the United States: A secondary analysis of the Global Burden of Diseases, Injuries, and Risk Factors Study. JAMA network open 2, e187041–e187041 (2019)

work page 2019

[3] [3]

& Oshio, T

Ma, X. & Oshio, T. The impact of social insurance on health among middle-aged and older adults in rural China: a longitudinal study using a three-wave nationwide survey. BMC Public Health 20, 1842 (2020)

work page 2020

[4] [4]

Ernst, F. R. & Grizzle, A. J. Drug-related morbidity and mortality: updating the cost-of- illness model. J Am Pharm Assoc (Wash) 41, 192–199 (2001)

work page 2001

[5] [5]

K., Mosier, M

Ziegler, D. K., Mosier, M. C., Buenaver, M. & Okuyemi, K. How much information about adverse effects of medication do patients want from physicians? Arch Intern Med 161, 706–713 (2001)

work page 2001

[6] [6]

Rothenberger, D. A. Physician Burnout and Well-Being: A Systematic Review and Framework for Action. Dis Colon Rectum 60, 567–576 (2017)

work page 2017

[7] [7]

(McGrawHill, Medical Publ

Handbook of Clinical Drug Data. (McGrawHill, Medical Publ. Division, New York, 2002)

work page 2002

[8] [8]

Tayefi, M. et al. Challenges and opportunities beyond structured data in analysis of electronic health records. WIREs Computational Statistics 13, e1549 (2021)

work page 2021