Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 28 of the 28 references it cites were found in a registry or an index; 5 of the 17 sources its claims cite were checked for support.

Unlocking electronic health records: a hybrid graph RAG approach to safe clinical AI for patient QA

Open access
Thio, Samuel, Lewis, Matthew, Denaxas, Spiros, Dobson, Richard J. B. · 2026 · Frontiers in Digital Health

Abstract

MediGRAF combines Neo4j Text2Cypher structured retrieval with vector embeddings over unstructured clinical notes for patient-level QA on MIMIC-IV.

Overview

CoChat Summary

## TL;DR

MediGRAF is a hybrid graph RAG system for patient-level EHR question answering that combines Text2Cypher graph retrieval with vector search over unstructured notes. In a 10-patient MIMIC-IV evaluation, it achieved 100% recall on factual queries and strong expert-rated performance on complex questions, with 0/10 unsafe responses.

## Key findings

  • On simple queries (n=100), the Cypher-only system achieved 80% accuracy, precision 0.8, recall 0.8, and F1 0.8; the hybrid system achieved 100% accuracy and recall, with precision/F1 not applicable because outputs were enriched responses rather than discrete answers.
  • On medium-complexity queries (n=31), the Cypher-only system achieved 51.6% accuracy, precision 0.806, recall 0.688, and F1 0.742; the hybrid system achieved 100% accuracy and recall.
  • For 10 complex queries evaluated by two clinical reviewers, the hybrid system scored high on factual accuracy and completeness: Evaluator 1 rated both at 4.40±1.02 and 4.40±0.92, and Evaluator 2 rated both at 4.90±0.30.
  • No complex-query responses were flagged as unsafe by either evaluator (0/10 cases). The paper also reports a mean expert quality score of 4.25/5 for complex inference tasks and 100% recall for factual queries.

## Methods

  • The paper presents MediGRAF, a hybrid Graph RAG system combining Neo4j Text2Cypher for structured retrieval with OpenAI text-embedding-3-small vectors for semantic search over discharge summaries and radiology reports.
  • Evaluation used MIMIC-IV version 3.1 data from 10 patients, producing 5,973 nodes and 5,963 relationships; the cohort included 4 single-encounter patients and 6 multi-encounter patients (2–9 admissions), with 89 free-text documents embedded.
  • The evaluation dataset comprised 141 QA pairs manually curated by three independent clinicians; simple/medium queries were assessed with precision/recall/F1/accuracy, and complex queries were judged by two hospital physicians using 5-point Likert scales plus a binary safety metric.

## Limitations

  • The authors state the evaluation scope was only 10 patients and two evaluators, which limits generalisability; they explicitly call for more patients, more clinical experts, and multiple institutions.
  • The paper notes context window limitations for hospital-wide queries and source-attribution ambiguity from context flattening; this limits transparency and requires filtering/pagination or future citation-aware generation.
  • The use of Likert scales for complex queries is subjective because there was no discrete ground truth for inference-based responses; this is stated in the paper.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fdgth.2026.1780700

Cites

28 of 28 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Comprehensive transformation pipeline. The diagram illustrates how heterogeneous data sources are processed: structured data (Diagnoses) undergoes synonym enrichment, while unstructured narratives (Radiology Reports, Discharge Notes) undergo vector embedding. All entities are linked via the central Admission node, enabling hybrid retrieval. Note: Other structured entities (e.g., Lab Events, Medications) undergo similar standardization but are omitted here for visual clarity.
Figure 1 Comprehensive transformation pipeline. The diagram illustrates how heterogeneous data sources are processed: structured data (Diagnoses) undergoes synonym enrichment, while unstructured narratives (Radiology Reports, Discharge Notes) undergo vector embedding. All entities are linked via the central Admission node, enabling hybrid retrieval. Note: Other structured entities (e.g., Lab Events, Medications) undergo similar standardization but are omitted here for visual clarity.CC BY · PubMed Central
Figure 2 Neo4j schema of MIMIC-IV data with clinical nodes and relationships.
Figure 2 Neo4j schema of MIMIC-IV data with clinical nodes and relationships.CC BY · PubMed Central
Figure 3 A workflow diagram showing the MediGRAF architecture from query input through graph retrieval to LLM augmentation and response generation.
Figure 3 A workflow diagram showing the MediGRAF architecture from query input through graph retrieval to LLM augmentation and response generation.CC BY · PubMed Central
Figure 4 Overview of the MediGRAF system interface showing the main query input area and configurable retrieval parameters. The left sidebar provides filtering options for patient selection, note type filtering, and advanced settings for both vector search (document retrieval limits) and graph database queries (maximum records to return).
Figure 4 Overview of the MediGRAF system interface showing the main query input area and configurable retrieval parameters. The left sidebar provides filtering options for patient selection, note type filtering, and advanced settings for both vector search (document retrieval limits) and graph database queries (maximum records to return).CC BY · PubMed Central
Figure 5 Comparison of mean expert evaluation scores across four dimensions. Error bars represent standard deviation. Note the significant divergence in “Relevance & Conciseness” scores.
Figure 5 Comparison of mean expert evaluation scores across four dimensions. Error bars represent standard deviation. Note the significant divergence in “Relevance & Conciseness” scores.CC BY · PubMed Central
Figure 6 Proposed deployment pipeline illustrating the integration of MediGRAF with the NHS CogStack ecosystem. CogStack handles real-time ingestion and NLP extraction (via MedCAT) to dynamically populate the Neo4j graph, enabling MediGRAF to query live clinical data.
Figure 6 Proposed deployment pipeline illustrating the integration of MediGRAF with the NHS CogStack ecosystem. CogStack handles real-time ingestion and NLP extraction (via MedCAT) to dynamically populate the Neo4j graph, enabling MediGRAF to query live clinical data.CC BY · PubMed Central

Verification

Claims the paper makes

5 Valid · 2 Not valid · 9 Not checked · 51 Own results and interpretations

22 more claims

Create a free account to see all 67 claims and to verify the 3 references still open.

See all claims — free account
  • It may state two separate claims as one. 29
  • Probably not a substantial claim. 15
  • Borderline: one of the tests for a substantial claim is unclear. 1

Create a free account to see these claims.

Citations

Cites

Show all 17 references

Cited by

See all 6

Metadata

Authors
Thio, S., Lewis, M., Denaxas, S., & Dobson, R. J. B.
Affiliations
  • Engineering and Physical Sciences Research Council (GB)
  • University College London (GB)
  • National and Kapodistrian University of Athens (GR)
  • University of London (GB)
Corresponding author
Thio, Samuel
Venue
Frontiers in Digital Health
Published
2026-03-11
Volume
8
Pages
1780700-1780700
ISSN
2673-253X (ISSN-L)
OpenAlex
W7135012900
Cited by
6
References
17
Citations by year
2026: 6
FWCI
11.72
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Digital Health (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Research Portal (King's College London) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
Funders
  • UK Research and Innovation
  • National Institute for Health and Care Research
  • University College London
  • NIHR Maudsley Biomedical Research Centre
  • Engineering and Physical Sciences Research Council (EP/Y035216/1)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Health Professions › Health Information Management › Electronic Health Records Systems
Keywords
electronic health records, graph-based retrieval-augmented generation, medical information retrieval, large language models, hallucination mitigation, vector embeddings, semantic search, MIMIC-IV

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).

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: Unlocking electronic health records: a hybrid graph RAG approach to safe clinical AI for patient QA. https://app.cochat.ai/papers/10.3389/fdgth.2026.1780700
The paper
Thio, S., Lewis, M., Denaxas, S., & Dobson, R. J. B. (2026). Unlocking electronic health records: a hybrid graph RAG approach to safe clinical AI for patient QA. Frontiers in Digital Health. https://doi.org/10.3389/fdgth.2026.1780700

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.