ANTI-HALLUCINATION GROUNDING · NCBI PUBMED & CROSSREF

PubMed Citation & Statistical Effect Size (HR/OR/95% CI) Checker

Generic LLMs frequently fabricate non-existent PMIDs or misquote Hazard Ratios from real trials. Paste your manuscript excerpt below to cross-check every PMID/DOI and inline statistical claim against authentic NCBI PubMed records.

1. Paste Manuscript Paragraph & References

Manuscript Excerpt (Demo includes: 1 verified [1], 1 mismatched HR [2], 1 hallucinated PMID [3])

In patients with heart failure and reduced ejection fraction (HFrEF), the DAPA-HF trial demonstrated that dapagliflozin reduced the primary composite outcome of worsening heart failure or cardiovascular death by 26% (HR 0.74, 95% CI 0.65–0.85, P<0.001) [1].
However, in patients with mildly reduced or preserved ejection fraction (the DELIVER trial), an AI-assisted draft mistakenly claimed a 42% relative risk reduction (HR 0.58, 95% CI 0.45–0.71) [2].
Furthermore, a generic LLM review cited a non-existent 2024 multicenter trial (PMID: 99999999) [3].

References:
[1] McMurray JJV, Solomon SD, Inzucchi SE, et al. Dapagliflozin in Patients with Heart Failure and Reduced Ejection Fraction. N Engl J Med. 2019;381:1995-2008. PMID: 31535829
[2] Solomon SD, McMurray JJV, Claggett B, et al. Dapagliflozin in Heart Failure with Mildly Reduced or Preserved Ejection Fraction. N Engl J Med. 2022;387:1089-1098. PMID: 36027570
[3] Fabricated AI Author, et al. Fictional Trial of SGLT2i in Acute Myocarditis. Lancet. 2024. PMID: 99999999

2. Traffic-Light Grounding Report

How Heurion Eliminates Hallucinated Medical Citations

Studies evaluating general-purpose LLMs in biomedical writing show that 20% to 40% of generated references contain fabricated DOIs, mismatched PMIDs, or inaccurate effect sizes (e.g., inventing a p-value or swapping primary and secondary endpoint Hazard Ratios). Heurion enforces a strict two-layer verification architecture: (1) Metadata Grounding via NCBI Entrez E-utilities (PubMed/PMC), Europe PMC, Crossref, and OpenAlex; and (2) Claim-Level Numerical Verification comparing extracted HR/OR/RR and 95% confidence intervals against the XML abstract of the source publication.