Asymmetric verification for detecting hallucinated claims in clinical information extraction
Year: 2026
Type: article
Abstract: Although large language models are capable of turning clinical notes into structured claims they might also create, modify or exaggerate clinical details. This study looks at a simplified asymmetric verification method for claims that have already been written; it does not assess the entire end-to-end extraction system. The concept of the Asymmetric Agentic Network (AAN) involves separating the validation process which is based on the source from the review of clinical coherence that is independent of the source, and then combining the two types of signals using an asymmetric consensus score. The main controlled adversarial stress test included 123 paired evaluations from 28 cases of held-out PMC-Patients. The AAN rejected 69 out of the 71 hallucinated claims and kept 51 of the 52 correct ones, giving an accepted-hallucination rate of 2.8% (95% CI 0.0–6.9%), a precision of 0.986, a recall of 0.972 and an F1 score of 0.979. Naive retrieval-augmented generation (RAG) also rejected the same 69 of the 71 hallucinated claims and retained all 52 correct claims, achieving an F1 score of 0.986; thus the AAN did not perform better than the simpler naive-RAG comparator on the main outcome measure. In the matched pooled analysis, the accepted-hallucination rate was 5.1% for AAN and 3.7% for naive RAG (the historically reported paired p value is 0.360), with false-rejection rates of 15.0% and 2.1% respectively. Since the primary analysis was underpowered, the fact that the comparison was not significant should not be taken to mean that the methods are equivalent. The AAN greatly outperformed the unverified baseline, the symmetric voting method and the call-count-matched comparator, while clinical entity substitutions that are similar to each other remained more difficult to detect than explicit numeric or negation perturbations. These findings describe a controlled verification stress test and illustrate a safety-coverage trade-off; they do not prove that asymmetric verification is better than retrieval-grounded verification nor do they validate the full conceptual architecture. Clinical trial number Not applicable.
Source: Discover Computing
Authors: Hemachandran B, Priyadarsini K
Institution: SRM Institute of Science and Technology
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $1,520
OpenAlex ID: W7225626643