MultiGraph-vet: A multimodal knowledge-augmented decision-support framework for safe dairy cow disease assessment on...

Rapid and reliable disease assessment is essential for dairy cow health management, but field-level clinical inputs are often colloquial, noisy, visually ambiguous, and context-dependent. These characteristics make large language model (LLM)-based veterinary decision-support systems vulnerable to factual errors, unsafe recommendations, and cross-modal hallucinations. To address this problem, we propose MultiGraph-Vet, a multimodal knowledge-augmented decision-support framework for safer dairy cow disease assessment. MultiGraph-Vet integrates a bovine veterinary knowledge graph, visual lesion parsing, colloquial symptom normalization, environmental context fusion, and a multimodal safety gate. It performs cross-modal consistency checking across text, image-derived lesion evidence, and environmental context, retrieves candidate diseases through a dual-path knowledge graph strategy, and generates evidence-grounded diagnostic suggestions through rule-constrained reasoning. We evaluated MultiGraph-Vet on a curated bovine clinical benchmark containing 338 cases, including 275 non-safety diagnostic decision-support cases, 132 real image-associated cases, and 63 safety-challenge scenarios involving contradictory, incomplete, non-bovine, or adversarial inputs. Compared with generic multimodal LLM baselines, MultiGraph-Vet achieved 80.36% diagnostic accuracy, 87.27% candidate recall@K, 96.83% safe interception rate, and 3.35% hallucination-related commitment rate. These results indicate that explicit veterinary knowledge grounding, visual lesion standardization, and multimodal conflict checking can improve reliability and safety control in AI-assisted dairy cow disease decision support. MultiGraph-Vet is intended to support preliminary assessment and risk-aware decision-making rather than replace professional veterinary diagnosis.