Graph-based RAG Architecture for Healthcare Fake News Detection using LLM-Driven Classification and Multi-Hop Knowledge Graph Evidence Analysis

Abstract

The rapid dissemination of healthcare information in digital platforms is a menace to health and involves the need to have effective, reliable and explainable systems of detection. A novel graph-based Retrieval-Augmented Generation (Graph-RAG) model of healthcare fake news detection is a large language model (LLM)-based classification system with multi-hop evidence analysis based on knowledge graphs. The hybrid models allows the combination of the structured knowledge inferences and semantic cognition for the hallucination concerns and contextual verification associated with LLM reasoning models and text categorization. The purpose of this study is to generate an evidence-based system that can be further used to classify and confirm the healthcare claims. By referring the calculated study data, there are 24 characteristics, 7,588 healthcare claims entries, pre-processing stage, and finally a multi-level feature engineering phase that uses the TF-IDF, medical entity recognition, and contextual embedding (RoBERTa/BioBERT). In the study the Logistic Regression Model is used to do an accurate predictions regarding the baselines, and a hybrid model is used to incorporate transformer-based models for the purpose of context analysis. Whenever, it’s time to talk about the treatments, diseases, and authoritative resources such as WHO (World Health Organization) and the Centers for Disease Control and Prevention. The medical knowledge graph facilitates the connecting entities and multi-hop reasoning to improve the dependability. Analyzing the evidence to either refute or support the Graph-RAG module, to ensure that the decision should be considered in the Support, Refute, or in the insufficient Evidence. The study result shows that the transformer models will show a 96% of classification rate. The Graph-RAG verification shows itself superior to the baseline models of the transformer, by achieving the 79% accuracy as opposed to 50% accuracy and no hallucination for the baseline models. The study conclusion shows that the organized knowledge graphs improve the interpretability of LLM judgments, risk of miscommunication and the dependability of the taken decisions.