An Evidence-Guided Reasoning Framework for Explainable Chinese Health Misinformation Detection

Abstract

Online health misinformation has become a persistent challenge for public health communication, medical decision-making, and digital information governance. Effective health fact-checking requires evidence-based verification rather than claim-only classification. However, health evidence is often heterogeneous, conditional, or insufficient, while structured Chinese health fact-checking resources containing both evidence and explanations remain scarce. These limitations make it difficult for large language model (LLM)-based systems to produce reliable predictions and auditable explanations. To address this challenge, this study constructs CHHMD, a Chinese health fact-checking dataset comprising 7296 structured claim–evidence–label–explanation instances. Building on this dataset, this study develops EGER-LLM, an integrated framework for evidence-guided reasoning. EGER-LLM combines heterogeneous evidence organization, evidence-conditioned LLM adaptation, uncertainty-aware decision-boundary refinement, and evidence-grounded explanation generation within a unified workflow. This design enables more reliable and explainable detection of health misinformation. Experiments on CHHMD show that EGER-LLM achieves an accuracy of 0.8247 and a Macro-F1 score of 0.7766 on the three-way classification task. Further evaluation on the public HealthFC dataset yields an accuracy of 0.9333 and a Macro-F1 score of 0.9139, outperforming the compared methods. These results demonstrate the effectiveness and generalizability of EGER-LLM for reliable, evidence-guided, and explainable health misinformation detection, providing a practical foundation for deployable health fact-checking systems.