MedF2Route: Evidence-Routed Graph Verification for Hallucination Detection in Clinical LVLMs¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Agentic-CliniAI/MedF2Route
Area: Medical NLP
Keywords: Clinical LVLMs, Hallucination Detection, Evidence Routing, Knowledge Graph, Dual-Path Verification
TL;DR¶
MedF2Route casts black-box clinical LVLM hallucination detection as an evidence routing problem, converting free-text case queries and generated responses into structured clinical graphs, dynamically dispatching response triplets across patient history, medical imaging, and external knowledge graphs, and achieving substantial precision and F1 gains while eliminating false alarms.
Background & Motivation¶
Clinical large vision-language models (LVLMs) hold immense promise for auxiliary medical diagnosis, imaging report generation, and automated clinical consultations, yet their clinical adoption is heavily compromised by the vulnerability to hallucinations. In high-stakes medical workflows, hallucinations manifest across two distinct and intertwined dimensions: faithfulness-oriented hallucinations, where the model output departs from or directly contradicts patient-specific evidence (such as hallucinating unmentioned comorbidities, inventing non-existent imaging lesions, or confusing laterality), and factuality-oriented hallucinations, where the diagnostic deductions, etiological assertions, or causal mechanisms conflict with verified medical knowledge and physiological laws (such as attributing benign tumors to malignant processes or asserting anatomically implausible lesion locations). In actual clinical practice, reliable clinical reasoning demands that an output simultaneously anchor to patient-specific facts and conform strictly to universal medical knowledge.
Despite recent advances in hallucination detection, existing paradigms are poorly equipped to tackle this dual-dimensional challenge. The majority of prevailing frameworks operate on an atomic claim-level decomposition pipeline; while isolating short sentences or claims enables fine-grained verification, it fundamentally shatters the contextual narrative and logical dependencies indispensable for clinical reasoning. For instance, natural language inference (NLI) classifiers and visual grounding metrics restrict their scope exclusively to text-grounded faithfulness, while fact-checking frameworks rely solely on text retrieval against general corpora. When prior methods attempt to address both dimensions jointly (such as UniHD), they routinely pool heterogeneous evidence—patient queries, multi-modal imaging, and retrieved web passages—into an undifferentiated prompt. This lack of architectural modularity triggers severe evidence misalignment and signal interference, inevitably causing the detectors to over-flag benign variations as hallucinations and driving precision down to impractical levels.
The central insight of this work is that the root cause of these detection failures lies in an unguided evidence assignment: black-box clinical hallucination detection is intrinsically an evidence routing problem. Different segments of a clinical diagnosis require fundamentally different evidentiary authorities—patient symptoms must be verified against case text, lesion morphology must be grounded in raw medical images, and differential diagnoses must be checked against curated biomedical knowledge graphs. Core idea: formulate clinical hallucination detection as an evidence-routed graph verification paradigm that transforms multimodal queries and responses into structured relational clinical graphs, partitions them into patient-grounding and knowledge-inference subgraphs, and dynamically dispatches response triplets to patient text, imaging, or multi-hop knowledge subgraphs for targeted dual-path verification before aggregating signals into sentence-level decisions.
Method¶
Overall Architecture¶
The MedF2Route framework operates via a four-stage verification pipeline: 1. Clinical Graph Construction: Given a multimodal patient query \(X = \{x_t, x_v\}\) (clinical case text and associated medical images) and an LVLM response sequence \(Y\), an instruction-guided relation extractor parses the unstructured clinical texts into structured query and response relational graphs \(G_{x_t}\) and \(G_y\); 2. Evidence-Aware Routing: A two-stage routing mechanism \(\rho\) partitions the response graph \(G_y\) into a query-following subgraph \(G_y^{\text{qry}}\) and a diagnostic inference subgraph \(G_y^{\text{inf}}\), subsequently dispatching triplets into image-grounded (\(G_y^{\text{qry,img}}\)), text-grounded (\(G_y^{\text{qry,txt}}\)), and knowledge-grounded (\(G_y^{\text{inf,kg}}\)) validation channels; 3. Dual-Path Verification: Routed response triplets undergo evidence-specific validation. In the faithfulness pathway, textual triplets are verified via semantic entailment against \(G_{x_t}\) while visual triplets are checked against medical image \(x_v\); in the factuality pathway, a sufficiency-guided multi-hop expansion over an external medical knowledge graph (LMKG) retrieves a compact evidence subgraph \(G_{\text{ref}}\) against which inference triplets are checked for biomedical validity; 4. Granularity Alignment: By leveraging explicit sentence-to-triplet mappings established during construction, fine-grained triplet-level verification labels are mapped back to sentence-level binary outputs \(\hat{z}\) using a conservative max-pooling strategy.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
In["Input: Multimodal Query X={x_t, x_v} and LVLM Response Y"] --> D1["Clinical Graph Construction<br/>Map free-text cases and responses into relational graphs G_{x_t} and G_y"]
D1 --> D2["Evidence-Aware Routing<br/>Decouple query-following and inference subgraphs; allocate to text/image/KG"]
D2 --> D3["Dual-Path Verification<br/>Faithfulness visual/textual checking + Factuality multi-hop KG verification"]
D3 --> D4["Granularity Alignment<br/>Anchor triplet decisions to sentences and apply conservative max-pooling"]
D4 --> Out["Output: Sentence-level Hallucination Vector z_hat"]
Key Designs¶
1. Clinical Graph Construction: Context Fragmentation -> Modifier-Preserving Relational Extraction -> Coherent Graph Representation
Claim-level decomposition algorithms frequently destroy crucial clinical context by dissecting sentences into disconnected fragments, stripping away vital medical modifiers such as laterality, severity, or negation. To preserve rich clinical semantics, MedF2Route deploys an instruction-guided clinical relation extractor \(\mathcal{G}\) that maps the patient case text \(x_t\) and LVLM response \(Y\) into directed clinical graphs \(G_{x_t} = (\mathcal{C}_{x_t}, \mathcal{E}_{x_t})\) and \(G_y = (\mathcal{C}_y, \mathcal{E}_y)\). The node set \(\mathcal{C}\) covers salient clinical entities (e.g., symptoms, anatomical structures, imaging findings, diagnoses, and procedures), while directed edges \(\mathcal{E}\) capture clinical relations \(r = \langle c_h, \ell, c_t \rangle\). Crucially, the extractor explicitly extracts and maintains attributes of negation, clinical certainty, laterality (e.g., left vs. right hilar lesions), temporal duration, and severity directly within the triplet structure. This ensures that the extracted graphs retain complete clinical validity without suffering from narrative or contextual loss.
2. Evidence-Aware Routing: Heterogeneous Signal Misalignment -> Two-Stage Subgraph Partitioning -> Zero-Interference Dispatch
Existing methods feed heterogeneous multimodal inputs into a single prompt, causing patient-specific clinical facts to cross-contaminate external knowledge checks. MedF2Route introduces an evidence-aware routing function \(\rho\):
$\(\left( G_y^{\text{qry,txt}}, G_y^{\text{qry,img}}, G_y^{\text{inf,kg}} \right) = \rho(G_y)\)$
The first routing step performs semantic subgraph separation, dividing \(G_y\) into a query-following subgraph \(G_y^{\text{qry}}\) (triplets describing patient history, reported symptoms, examination results, and explicit case findings) and an inference subgraph \(G_y^{\text{inf}}\) (triplets stating diagnostic conclusions, underlying etiological explanations, differential diagnoses, or prognostic evaluations). The second routing step refines the evidence targets: triplets in \(G_y^{\text{qry}}\) describing directly observable visual features (such as lesion morphology, density, contour, margin, and anatomical location) are assigned to the image branch \(G_y^{\text{qry,img}}\) pointing to \(x_v\), while non-imaging patient findings are routed to \(G_y^{\text{qry,txt}}\) pointing to \(G_{x_t}\). Triplets in \(G_y^{\text{inf}}\) are directed to the medical knowledge base branch \(G_y^{\text{inf,kg}}\). By strictly circumscribing the reference evidence for each clinical assertion, this design prevents cross-modal interference.
3. Dual-Path Verification: Modality Blind Spots -> Multimodal Faithfulness & Multi-Hop KG Factuality -> Rigorous Grounding
Each routed triplet is evaluated by dedicated verification modules tailored to its allocated evidence source. For faithfulness verification: textual triplets \(r \in G_y^{\text{qry,txt}}\) are matched against query graph \(G_{x_t}\) via prompted semantic entailment, tolerating lexical synonyms while strictly flagging any unsubstantiated patient facts; visual triplets \(r \in G_y^{\text{qry,img}}\) are submitted alongside raw medical image \(x_v\) to a vision-language judge to verify whether the claimed morphological attributes (e.g., well-circumscribed dense lesion vs. destructive mass) are visually corroborated. For factuality verification: inference triplets \(r \in G_y^{\text{inf,kg}}\) are verified against an external biomedical knowledge graph (LMKG). Salient concepts are identified as seeds, filtered via BioLORD-2023 semantic embeddings (threshold \(\delta = 0.7\)), and expanded iteratively via sufficiency checking up to a depth budget \(T = 5\) to construct a compact reference subgraph \(G_{\text{ref}}\). Triplets are evaluated against \(G_{\text{ref}}\) to detect medically false, anatomically implausible, or contradictory assertions, decoupling clinical reasoning correctness from case-specific observations.
4. Granularity Alignment: Granularity Mismatch -> Traceable Sentence Anchoring -> Conservative Max-Pooling
A major challenge in deploying triplet-level graph verification to clinical benchmarks is bridging the gap to sentence-level evaluation without introducing assignment ambiguity. During initial graph parsing, every extracted response triplet \(r\) is explicitly indexed with its source sentence identifier \(s_t\), defining the triplet set \(R(s_t)\). Following dual-path verification where each triplet receives a binary label \(\hat{y}_r \in \{0, 1\}\), MedF2Route enforces a safety-critical conservative aggregation rule:
$\(\hat{z}_t = \max_{r \in R(s_t)} \hat{y}_r\)$
Under this formulation, a response sentence \(s_t\) is classified as hallucinated (\(\hat{z}_t = 1\)) if any single constituent clinical relation is unsupported by patient evidence, visually contradicted by imaging, or medically fallacious according to the knowledge base. This conservative guarantee ensures zero-tolerance for erroneous clinical relations in safety-critical medical deployment while providing precise triplet-level explainability.
Benchmark & Evaluation Protocol¶
To benchmark hallucination detection under realistic clinical diagnostic scenarios, the authors introduced MedF2Bench, comprising 500 multimodal clinical cases derived from MultiCaRe (originating from PubMed Central case reports). Spanning 5 imaging modalities (CT, ECG, MRI, Ultrasound, and X-ray) with 585 medical images and 2,383 annotated response sentences, the benchmark incorporates controlled sentence-level perturbations following FactCC and HaluEval protocols. Both faithfulness-oriented errors (unsupported clinical findings or conflicting imaging details) and factuality-oriented errors (incorrect medical mechanisms or implausible differential diagnoses) are introduced under rigorous stratified auditing (98% intended error precision, 100% type accuracy). All models are evaluated under a case-averaged protocol across sentence-level Precision (P), Recall (R), F1-score, Accuracy (Acc), and macro-averaged F1 (Macro-F1).
Key Experimental Results¶
Main Results¶
The performance of MedF2Route was evaluated across three distinct vision-language model backbones (GPT-5.4-nano, InternVL-3.5, and Gemini-3.1-flash-lite) against representative sentence-level (VLM-as-Judge, SelfCheckGPT), claim-level (FActScore, UNIHD), and graph-structured (FactSelfCheck) baselines on MedF2Bench.
| Backbone | Case Type | Method | Hall. P (%) | Hall. R (%) | Hall. F1 (%) | Overall Acc. (%) | Macro-F1 (%) |
|---|---|---|---|---|---|---|---|
| GPT-5.4-nano | Faithfulness | VLM-as-Judge | 29.79 | 61.33 | 37.76 | 64.01 | 53.67 |
| GPT-5.4-nano | Faithfulness | SelfCheckGPT | 20.24 | 48.00 | 26.41 | 53.87 | 43.13 |
| GPT-5.4-nano | Faithfulness | FActScore | 20.47 | 70.22 | 30.97 | 38.60 | 34.60 |
| GPT-5.4-nano | Faithfulness | UNIHD | 24.83 | 91.96 | 37.74 | 36.60 | 32.21 |
| GPT-5.4-nano | Faithfulness | FactSelfCheck | 21.02 | 99.11 | 34.63 | 21.99 | 18.57 |
| GPT-5.4-nano | Faithfulness | MedF2Route (Ours) | 36.68 | 66.67 | 45.03 | 67.44 | 59.64 |
| GPT-5.4-nano | Factuality | VLM-as-Judge | 42.96 | 84.73 | 54.11 | 68.13 | 62.76 |
| GPT-5.4-nano | Factuality | SelfCheckGPT | 42.94 | 86.18 | 53.55 | 63.36 | 59.19 |
| GPT-5.4-nano | Factuality | FActScore | 25.46 | 80.36 | 37.36 | 42.87 | 39.12 |
| GPT-5.4-nano | Factuality | UNIHD | 29.73 | 96.70 | 43.59 | 41.41 | 38.24 |
| GPT-5.4-nano | Factuality | FactSelfCheck | 21.63 | 99.27 | 35.42 | 22.40 | 18.80 |
| GPT-5.4-nano | Factuality | MedF2Route (Ours) | 48.40 | 93.45 | 60.84 | 70.36 | 67.32 |
| InternVL-3.5 | Faithfulness | VLM-as-Judge | 27.44 | 36.00 | 29.85 | 78.44 | 57.74 |
| InternVL-3.5 | Faithfulness | MedF2Route (Ours) | 44.85 | 56.00 | 48.33 | 79.73 | 67.20 |
| Gemini-3.1-flash-lite | Faithfulness | VLM-as-Judge | 31.13 | 76.89 | 42.30 | 57.27 | 51.44 |
| Gemini-3.1-flash-lite | Faithfulness | MedF2Route (Ours) | 46.20 | 73.78 | 54.61 | 76.46 | 68.58 |
| Gemini-3.1-flash-lite | Factuality | VLM-as-Judge | 43.04 | 91.64 | 55.53 | 64.27 | 61.12 |
| Gemini-3.1-flash-lite | Factuality | MedF2Route (Ours) | 42.94 | 71.64 | 51.62 | 74.45 | 66.05 |
Ablation Study¶
The contribution of each individual component was systematically assessed using GPT-5.4-nano, where modules were ablated while preserving context for downstream evaluation:
| Variant | Sent. Mac.F1 (%) | Joint Mac.F1 (%) | Fact. Mac.F1 (%) | Faith. Mac.F1 (%) | Architectural Impact & Analysis | |---|---|---|---|---| | Full MedF2Route | 63.86 | 67.12 | 66.93 | 41.38 | Optimal holistic balance across all clinical dimensions | | w/o Evidence-aware Routing | 51.91 (-11.95) | 60.27 (-6.85) | 58.07 (-8.86) | 44.83 (+3.45) | Severe loss of dynamic coordination and knowledge factuality | | w/o KG Retrieval | 60.60 (-3.26) | 65.28 (-1.84) | 64.24 (-2.69) | 41.38 (0.00) | Impairs external biomedical fact injection | | w/o Multi-hop KG Expansion | 63.18 (-0.68) | 66.56 (-0.56) | 65.86 (-1.07) | 41.38 (0.00) | Drops bridging concepts for complex reasoning paths | | w/o Visual Evidence Selection | 60.95 (-2.91) | 63.69 (-3.43) | 66.93 (0.00) | 20.02 (-21.36) | Catastrophic drop in image-grounded faithfulness checking |
Key Findings¶
- Resolving the High-Recall False Positive Pathology: Standard claim-level baselines (e.g., UNIHD, FactSelfCheck) suffer from an acute false-alarm bias, achieving near 99% recall at the cost of dismal precision (~20%-25%). In clinical environments, such indiscriminate flagging degrades trust and overwhelms clinician reviewers. MedF2Route dramatically suppresses false alarms, lifting precision to 36.68% (Faithfulness) and 48.40% (Factuality) on GPT-5.4-nano, and reaching 44.85% on InternVL-3.5 (+17.41% over VLM-as-Judge).
- Indispensability of Evidence Routing: Bypassing evidence routing induces the single largest performance plunge across the entire framework, shaving 11.95 points off sentence-level Macro-F1 and 8.86 points off factuality Macro-F1. While unguided models show an artificial +3.45 point drift on simple faithfulness, their capacity to scrutinize complex medical facts collapses without explicit routing.
- Orthogonal Functional Decoupling: Ablating visual evidence selection triggers an immediate 21.36-point plunge in triplet-level faithfulness (from 41.38% down to 20.02%) while leaving factuality completely unharmed (66.93%). Conversely, removing KG retrieval degrades factuality while leaving faithfulness intact. This confirms that the framework successfully isolates independent error modalities without cross-talk.
- Inference Efficiency: According to the cost analysis (Table 14), MedF2Route achieves top-tier detection accuracy with an average of 12.92 API calls and 16,735 tokens per case under GPT-5.4-nano. This compares favorably against multi-sample approaches like SelfCheckGPT (23.06 calls) and document-retrieval methods like FActScore (34,203 tokens), providing a practical computational footprint.
Highlights & Insights¶
- Evidence Routing as a Foundational Paradigm: Rather than treating clinical hallucination detection as generic string matching or monolithic prompting, the paper formalizes detection as an evidence-routed navigational problem, systematically mapping diverse diagnostic assertions to their rightful authorities.
- Modifier-Preserving Graph Extraction: Formulating clinical graph extraction to explicitly retain negation, laterality, severity, and certainty directly eliminates the semantic flips and context loss that routinely plague naive token-level and claim-level decomposition methods.
- Sufficiency-Guided Multi-Hop Subgraph Expansion: Instead of blindly flooding the LLM context with broad KG subgraphs, the expansion dynamically terminates once sufficient explanatory paths connecting head and tail entities are established, maintaining both brevity and high factual coverage.
Limitations & Future Work¶
- Backbone Verification Bottlenecks: Error analysis on InternVL factuality disagreements reveals that 69.90% of misclassifications stem directly from intrinsic limitations in the underlying vision-language model's clinical reasoning rather than pipeline failures. Future efforts should explore specialized medical foundation models as judges.
- Knowledge Graph Incompleteness: Local knowledge graphs (LMKG) cover approximately 70% of clinical entities in MedF2Bench, with knowledge gaps accounting for 14.29% of residual factuality errors. Integrating dynamically refreshed biomedical literature corpora could mitigate coverage ceilings.
- Expansion Beyond Diagnostic Workflows: The current framework focuses primarily on differential diagnosis; extending graph routing to prescriptive therapy, pharmacogenomics, and adverse drug-drug interaction monitoring represents an important frontier.
Related Work & Insights¶
- vs FActScore / UniHD: FActScore is confined to open-domain text and relies on unguided document retrieval without multimodal visual grounding; UniHD lumps text, web, and images into flat prompts, causing severe evidence cross-contamination. MedF2Route enforces explicit topological subgraphs to physically decouple patient observations from medical common sense.
- vs GraphEval / FactSelfCheck: GraphEval operates purely on unimodal text without imaging capability; FactSelfCheck relies on costly sampling to generate pseudo-graphs prone to collective hallucination. MedF2Route grounds assertions directly in real imaging pixel distributions and validated biomedical knowledge graphs, ensuring verifiable, transparent, and interpretable clinical auditing.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ [Pioneers the evidence routing formulation for clinical LVLM hallucination detection and introduces structured patient/knowledge subgraph decoupling]
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive validation across 500 cases spanning 5 imaging modalities, three VLM backbones, multiple baseline paradigms, stratified audits, and statistical significance tests]
- Writing Quality: ⭐⭐⭐⭐⭐ [Exceptionally rigorous formulation, precise mathematical definitions, clean structural diagrams, and lucid prose]
- Value: ⭐⭐⭐⭐⭐ [Addresses the critical clinical challenge of false-alarm reduction and context preservation, establishing an essential building block for trustworthy medical AI systems]