HealthVer: Biomedical Claim Verification

• HealthVer is an evidence-grounded verification task that determines whether health claims are supported, refuted, or unverified using scientific literature and patient records.
• It leverages benchmarks like HealthVer, HealthFC, and VeriFact to annotate claim–evidence pairs, improving reliability through structured rationale extraction and granular claim decomposition.
• The system integrates retrieval-augmented methods with machine learning architectures, achieving high performance and transparent decision-making in both biomedical and clinical settings.

Searching arXiv for papers directly related to HealthVer and health claim verification. arXiv search query: "HealthVer biomedical claim verification health fact checking"

HealthVer is the task of determining whether a health-related claim is supported, refuted, or cannot be verified with current evidence, and of explaining the decision with transparent, credible sources. In the recent literature, the term is used in two closely related senses: literature-grounded verification of biomedical and consumer-health claims against scientific evidence, and patient-specific health fact verification against a patient’s own electronic health record (EHR). The first sense is exemplified by the HealthVer benchmark of claim–evidence pairs from scientific abstracts, while the second is instantiated by VeriFact’s chart-level verification of Brief Hospital Course narratives against longitudinal EHR notes (Kumar, 11 Jun 2026, Chung et al., 28 Jan 2025).

1. Scope and conceptual boundaries

In its literature-grounded form, HealthVer centers on evidence-based medical fact-checking rather than general web fact-checking. HealthFC defines the task as determining whether a health-related claim is supported, refuted, or cannot be verified with current evidence, with evidence derived from clinical trials and systematic reviews rather than general web content (Vladika et al., 2023). CER likewise frames biomedical fact-checking as grounding veracity decisions in peer-reviewed biomedical evidence, using PubMed abstracts as the primary knowledge base and labels aligned to true, false, and insufficient evidence (Barone et al., 17 Sep 2025).

In its patient-specific form, HealthVer shifts from public literature to intra-chart factuality. VeriFact verifies the factuality of long-form Brief Hospital Course narratives against the patient’s EHR, with the focus on patient-specific, intra-chart consistency rather than general medical claims sourced from the web. Its labels are Supported, Not Supported, and Not Addressed, and its evidence pool is restricted to the patient’s pre-discharge clinical notes (Chung et al., 28 Jan 2025).

This suggests that HealthVer is best understood as a family of evidence-grounded verification tasks rather than a single benchmark. One branch asks whether a biomedical claim is justified by scientific literature; another asks whether a clinical statement is justified by the chart. Both branches emphasize evidence provenance, transparent rationales, and explicit handling of uncertainty.

HealthVer should also be distinguished from virtual-care trust systems that verify identity, credentials, or licensure. VerifyMed addresses identity assurance of participants, validation of credentials/license authorization, and ongoing competence verification via outcomes on Ethereum, but it does not perform evidence-claim veracity assessment in the literature-grounded or chart-grounded sense (Rensaa et al., 2020).

2. Benchmarks, corpora, and annotation protocols

The literature describes several resources that define the operational landscape of HealthVer-style research. HealthVer is a biomedical claim verification dataset with evidence drawn from scientific abstracts. HealthFC is a bilingual benchmark grounded in expert-written evidence articles from Medizin Transparent. VeriFact-BHC extends the problem to chart-level verification by decomposing discharge-summary narratives into propositions and annotating each proposition against the patient’s EHR (Kumar, 11 Jun 2026, Vladika et al., 2023, Chung et al., 28 Jan 2025).

Resource	Evidence basis	Reported composition
HealthVer	Scientific abstracts from PubMed	14,330 evidence–claim pairs; 10,590 train, 1,917 dev, 1,823 test
HealthFC	One expert-written evidence article per claim	750 claims in German and English; supported/refuted/NEI
VeriFact-BHC	Patient EHR clinical notes	13,290 propositions from 100 patients

HealthVer is evaluated in multiple later studies as a three-way natural language inference problem over claim–evidence pairs. One study normalizes model outputs to supports, refutes, and nei, with deterministic mappings such as SUPPORTS/ENTAILMENT $\rightarrow$ supports and NEI/NOT ENOUGH INFO/NEUTRAL $\rightarrow$ nei (Kumar, 11 Jun 2026). MultiVerS uses HealthVer as one of three target scientific claim verification datasets and reconstructs abstract-level instances because sentence-level rationales were not originally provided; sentences with $>50\%$ overlap with the original evidence are marked as rationales, and a small number of claims with both supporting and refuting evidence in the same abstract were removed to conform to the single-stance definition (Wadden et al., 2021).

HealthFC uses a different evidence regime. Each of its 750 claims is paired with one expert-written evidence article, a short explanation paragraph, and manually annotated rationale sentences, with up to 5 rationale sentences per article. All 750 claims are available in both German and English, and the dataset is released with 5 cross-validation folds rather than a fixed train/dev/test split (Vladika et al., 2023).

VeriFact-BHC uses a clinician-annotated protocol rather than literature evidence. It contains 13,290 propositions, including 9,108 atomic claim propositions and 4,182 sentence propositions. Each proposition is labeled by three clinicians as Supported, Not Supported, or Not Addressed based on the patient’s EHR, and 478 disagreements were reviewed by two additional physicians with further adjudication to obtain a denoised, adjudicated average clinician ground truth (Chung et al., 28 Jan 2025).

3. Evidence units, decomposition, and representation

A central design choice in HealthVer is the unit of verification. Literature-grounded systems often take a claim–evidence pair as the atomic input, but several papers show that the granularity of the claim and the scope of evidence context materially affect performance.

VeriFact studies two proposition types for Brief Hospital Course narratives: sentence propositions and atomic claims. Atomic claims are LLM-extracted simple Subject–Predicate–Object assertions modeled on first-order logic and logical atomism. The study reports that sentence propositions from human-written text had 19.8% invalid, whereas atomic claims reduced invalids in human-written summaries to 0.4% (Chung et al., 28 Jan 2025). A plausible implication is that decomposition is not merely a preprocessing convenience; it changes both annotation reliability and the attainable ceiling for automated verification.

MultiVerS takes the opposite direction by enlarging the evidence context. It encodes the claim and the full abstract in a single Longformer pass, using the format

$$\langle s \rangle\ c\ \langle /s \rangle\ t\ \langle /s \rangle\ s_1\ \langle /s \rangle_1 \dots s_n\ \langle /s \rangle_n,$$

and optimizes a joint objective

$$L = L_{\text{label}} + \lambda L_{\text{rationale}},$$

with $\lambda = 15$ in the final experiments (Wadden et al., 2021). The rationale head predicts whether each sentence is evidentiary, while the label head predicts Supports, Refutes, or NEI. This design treats rationale extraction and stance prediction as coupled tasks rather than independent stages.

The contrast between these approaches reflects a recurrent tension in HealthVer. Smaller units such as atomic claims improve validity and reduce ambiguity, but full-document encoders preserve qualifiers, population restrictions, and other contextual dependencies that are often lost when evidence is reduced to isolated snippets. This suggests that decomposition and contextualization are complementary rather than mutually exclusive.

4. Verification architectures and methodological paradigms

HealthVer research now spans several methodological families. MultiVerS represents the multitask full-document paradigm: a shared encoding supports both label prediction and rationale identification, and weak supervision is enabled by setting $\lambda = 0$ on weakly labeled instances lacking sentence-level rationales (Wadden et al., 2021).

CER represents a retrieve-reason-classify pipeline. It builds sparse BM25 and dense SBERT indices on PubMed abstracts, retrieves top-$k=20$ passages per claim, selects up to three evidence sentences for prompting, and then feeds claim, evidence, and an LLM-generated justification to a supervised classifier. The fine-tuning objective is cross-entropy over the label set $\{\text{true}, \text{false}, \text{NEI}\}$, and the LLM prompt is explicitly grounded to use only the retrieved scientific context (Barone et al., 17 Sep 2025).

VeriFact couples retrieval-augmented generation with an LLM-as-a-Judge. All EHR clinical notes are transformed into sentence facts or atomic claim facts, stored in a Qdrant vector database, retrieved with dense or hybrid search, optionally reranked with a cross-encoder, and then compared to each proposition by an LLM judge that emits Supported, Not Supported, or Not Addressed in structured JSON. The pipeline also includes proposition validity screening, duplicate fact deduplication, a self-healing loop for invalid JSON, and temporal formatting variants such as Absolute Time and Relative Time (Chung et al., 28 Jan 2025).

PROClaim extends HealthVer-style verification into structured multi-agent deliberation. It introduces specialized roles such as Plaintiff Counsel, Defense Counsel, The Court, Expert Witness, Critic Agent, and a three-model Judicial Panel. Its core retrieval mechanism is Progressive RAG, formalized as

$E_t = E_{t-1} \cup R_t(q_t, E_{t-1}),$

with novelty filtering, admissibility scoring, stance-conditioned retrieval, premise-grounded retrieval, role-switching, self-reflection, and majority-vote judgment (Chowdhury et al., 30 Mar 2026). The framework is explicitly designed to mitigate hallucinations, epistemic bubble effects, and premature convergence.

VERIRAG inserts a methodological audit into retrieval-augmented verification. Each source is scored with the Veritable 11-point checklist, and the intrinsic document quality is

$\rightarrow$0

where $\rightarrow$1 corresponds to Fail, Uncertain, and Pass, and $\rightarrow$2 is the applicability mask. Quality is combined with a redundancy penalty to form $\rightarrow$3, and the final Hard-to-Vary score is

$\rightarrow$4

A Dynamic Acceptance Threshold $\rightarrow$5 then determines whether the claim is accepted or rejected (Mohole et al., 23 Jul 2025). This formulation makes evidence quality, rather than topical relevance alone, part of the decision rule.

A further methodological branch emphasizes compact fine-tuned decoders. Small-LLM work on HealthVer fine-tunes Phi-3-mini, Qwen2.5-3B, and Mistral-7B with QLoRA on claim–evidence–label triples, with 4-bit NF4 quantization, LoRA rank $\rightarrow$6, $\rightarrow$7, dropout $\rightarrow$8, max sequence length 1,024, and best-checkpoint selection by validation macro-F1 (Kumar, 11 Jun 2026). This line treats HealthVer as a direct supervised inference problem when annotated evidence is already available.

5. Empirical results and benchmark findings

Reported results show that HealthVer performance depends strongly on evidence quality, retrieval design, and dataset construction. In patient-specific verification, VeriFact achieves up to 92.7% agreement on LLM-written sentence propositions when compared to a denoised and adjudicated average clinician ground truth; the highest agreement between clinicians was 88.5%, and best agreement on LLM-written atomic claims was approximately 88.8%, comparable to individual clinician agreement (Chung et al., 28 Jan 2025).

In literature-grounded HealthVer, MultiVerS reports strong fully supervised performance. On HealthVer, its full-document multitask model obtains abstract-level F1 of 77.6 and sentence-level “selection+label” F1 of 69.1, compared with 71.5 and 64.9 for ParagraphJoint and 72.6 and 63.3 for Vert5Erini (Wadden et al., 2021). CER, evaluated on adjacent biomedical fact-checking datasets rather than HealthVer itself, reports macro-F1 of 69.90 on HealthFC, 95.20 on BioASQ-7b, and 61.14 on SciFact for its best fine-tuned configurations, and its ablations show large drops when evidence or justification is removed (Barone et al., 17 Sep 2025).

Small-LLM experiments provide a direct HealthVer comparison under matched training size. With 1,008 HealthVer training examples, Mistral-7B QLoRA achieves 66.0 accuracy and 65.2 macro-F1 on the official HealthVer test set; Phi-3-mini QLoRA reaches 66.1 accuracy and 65.1 macro-F1; BioLinkBERT trained on the same 1,008-example subset achieves 65.5 accuracy and 62.4 macro-F1; and BioLinkBERT trained on the full HealthVer training set reaches 82.7 accuracy and 81.9 macro-F1 (Kumar, 11 Jun 2026).

PROClaim’s main experiments are on Check-COVID, where it reaches 81.7% accuracy and macro F1 $\rightarrow$9, but it also reports an external HealthVer proof-of-concept: 72.0% accuracy and 0.713 F1 on a random 100-claim sample, using a PubMed-based retriever rather than a fully optimized medical pipeline (Chowdhury et al., 30 Mar 2026). VERIRAG evaluates binary healthcare claim verification under evolving temporal evidence scenarios and reports macro F1 from 0.5325 to 0.6542, outperforming the next-best method by 10 to 14 points in each scenario (Mohole et al., 23 Jul 2025).

System	Setting	Reported result
VeriFact	LLM-written sentence propositions	Up to 92.7% agreement
MultiVerS	HealthVer, fully supervised	77.6 abstract-level F1; 69.1 sentence-level F1
Mistral-7B QLoRA	HealthVer, 1,008 training examples	66.0 accuracy; 65.2 macro-F1
PROClaim	HealthVer, 100-claim sample	72.0% accuracy; 0.713 F1
VERIRAG	Four temporal corpora	Macro F1 0.53–0.65

A recurring empirical conclusion is that evidence handling is the main bottleneck. CER shows that removing scientific evidence reduces F1 from 69.90 to 40.61 on HealthFC and from 61.14 to 43.02 on SciFact (Barone et al., 17 Sep 2025). HealthFC similarly shows a large oracle gap: DeBERTa’s veracity prediction rises from 67.5 macro-F1 with model-selected evidence to 81.9 with gold evidence sentences (Vladika et al., 2023).

6. Applications and deployment settings

HealthVer has direct applications in consumer health information, biomedical literature analysis, and clinical documentation workflows. HealthFC was designed for consumer health information retrieval and verification, quality assurance in generative health assistants, and benchmarking components for clinical decision support, while retaining short explanation paragraphs and rationale sentences for transparent output (Vladika et al., 2023).

CER provides a deployment-oriented blueprint for biomedical literature verification: index PubMed abstracts, retrieve top-$>50\%$0 passages, constrain the LLM to use only retrieved evidence, and delegate the final decision to a supervised classifier over claim, evidence, and rationale (Barone et al., 17 Sep 2025). PROClaim extends this to controversial claims by keeping a complete case record of transcripts, evidence, and judge opinions, which supports auditability and uncertainty communication (Chowdhury et al., 30 Mar 2026).

VeriFact is explicitly oriented toward clinical workflow integration. Its per-proposition verdicts and reasons, plus an aggregated score sheet, can feed acceptance thresholds, trigger revision workflows, or flag high-risk sections for clinician review. The paper identifies automated validation of discharge summaries before signing and decision support to highlight contradictions or unsupported assertions across longitudinal notes as concrete uses (Chung et al., 28 Jan 2025).

At the same time, HealthVer is not synonymous with trust in telehealth identity or credentialing. VerifyMed shows that a health verification system may instead be concerned with evidence of authority, evidence of experience, and evidence of competence, implemented via role-gated Ethereum smart contracts, treatment records, and patient-reported experience measures (Rensaa et al., 2020). This is a distinct deployment axis: it verifies actors and governance states rather than scientific or chart-level claims.

7. Limitations, controversies, and open problems

One major controversy concerns the meaning of NEI and related negative labels. In HealthVer, NEI examples include present-but-inconclusive evidence, so NEI cannot be identified from the absence of evidence. A later study identifies a previously unreported structural artifact in SciFact: all NEI examples have empty evidence fields, while all supports/refutes examples have non-empty evidence. The paper argues that this inflates in-domain scores and shows that SciFact-trained models exhibit catastrophic NEI collapse when tested on HealthVer (Kumar, 11 Jun 2026). This suggests that dataset structure can determine whether a system learns biomedical reasoning or merely a shortcut.

A second issue is calibration under evidential insufficiency. VeriFact finds that the distinction between Not Supported and Not Addressed is harder for both clinicians and the system, and that binarization improves agreement by 10–15% compared to the three-label task (Chung et al., 28 Jan 2025). CER’s error analysis similarly emphasizes retrieval misses, nuanced negation, broad under-specified claims, and overgeneralization even under grounding (Barone et al., 17 Sep 2025).

A third unresolved problem is evidence quality. Standard RAG retrieves semantically relevant sources but cannot assess whether those sources are rigorous, powered, free of reporting bias, or even retracted. VERIRAG treats this as methodological blindness and introduces the Veritable checklist, HV Score, and Dynamic Acceptance Threshold to make source quality part of the verification procedure (Mohole et al., 23 Jul 2025). A plausible implication is that retrieval relevance alone is insufficient for high-stakes biomedical verification.

Generalizability also remains unsettled. VeriFact is built on MIMIC-III, which lacks non-ICU inpatient and outpatient notes, and the paper notes information asymmetry because human authors may have access to bedside or external information not present in the chart (Chung et al., 28 Jan 2025). PROClaim’s HealthVer evaluation is explicitly described as a proof-of-concept with PubMed-based retrieval rather than a fully optimized medical pipeline (Chowdhury et al., 30 Mar 2026). HealthFC notes that verdicts reflect knowledge up to end of 2022 and may evolve as new trials and systematic reviews appear (Vladika et al., 2023).

Future directions named across the literature are convergent: better calibration for negative labels, richer claim units such as molecular facts, larger or more diverse evidence pools, fine-tuned medical-domain judges, broader full-text retrieval, stronger contradiction detection, and multi-institution or cross-domain generalization (Chung et al., 28 Jan 2025, Kumar, 11 Jun 2026). Across both literature-grounded and chart-grounded variants, the field is moving toward systems that are not only evidence-aware but also provenance-aware, temporally aware, and explicitly uncertainty-aware.