TLDR: CER (Combining Evidence and Reasoning) is a novel framework for biomedical fact-checking that integrates scientific evidence retrieval, large language model (LLM) reasoning, and supervised veracity prediction. It aims to combat healthcare misinformation by grounding LLM outputs in verifiable, evidence-based sources, thereby mitigating the risk of hallucinations. Evaluations on expert-annotated datasets (HealthFC, BioASQ-7b, SciFact) demonstrate state-of-the-art performance and promising cross-dataset generalization, highlighting its effectiveness in providing accurate and reliable claim verification.
Misinformation in healthcare, ranging from vaccine hesitancy to unproven treatments, poses significant risks to public health and erodes trust in medical systems. While automated fact-checking has advanced with machine learning and natural language processing, validating complex biomedical claims remains a unique challenge due to specialized terminology, the need for domain expertise, and the critical importance of grounding information in scientific evidence.
To address these challenges, researchers have introduced CER (Combining Evidence and Reasoning), a novel framework designed specifically for biomedical fact-checking. This system integrates three core components: systematic scientific evidence retrieval, reasoning capabilities powered by large language models (LLMs), and supervised veracity prediction.
The CER framework begins with a Scientific Evidence Retrieval module. This module interfaces with extensive scientific knowledge bases, primarily PubMed, to extract domain-specific claims. It focuses on article abstracts, which provide concise yet comprehensive summaries of research findings. The system employs both Sparse Retrieval (using BM25) and Dense Retrieval (using a pre-trained SBERT model) to identify relevant sentences from the indexed database. For each claim, up to three pieces of evidence are extracted and structured with the original claim for the next stage.
Next, the LLM Reasoning phase leverages large language models, such as Mixtral-8x22B-Instruct-v0.1, as reasoning assistants. This design choice is crucial for mitigating the risk of hallucinations often associated with LLMs when used for standalone fact-checking. The LLM’s role is twofold: to assess the claim’s veracity based on the provided scientific evidence and to generate a detailed justification for this assessment. This process is guided by a specific prompt template that combines the claim with the retrieved evidence, often assigning the LLM a ‘Doctor’ role to enhance its reasoning context.
Finally, the Veracity Prediction module acts as a dedicated verification layer. It evaluates both the LLM’s reasoning and the underlying evidence to produce more reliable classifications. This module assigns one of three labels: “true,” “false,” or “insufficient evidence.” The framework explores two approaches for this task: zero-shot classification, where a language model directly classifies based on its pre-trained knowledge, and fine-tuning, where the model is adapted to the specific task using a smaller, domain-specific dataset. Fine-tuning generally leads to enhanced accuracy for specialized tasks.
Evaluations of CER on expert-annotated datasets like HealthFC, BioASQ-7b, and SciFact have demonstrated state-of-the-art performance, showing consistent improvements over existing methods. For instance, the fine-tuned CER achieved an F1 score of 69.90% on HealthFC and 95.20% on BioASQ-7b. Ablation studies confirmed the critical impact of scientific evidence retrieval, with its removal leading to substantial performance degradation. The choice between dense and sparse retrieval methods showed marginal differences, indicating the framework’s robustness. Furthermore, the impact of LLM reasoning was significant, with the full prompt structure (including role assignment, scientific evidence, and justification requirement) yielding the best results.
The framework also demonstrated promising cross-dataset generalization, suggesting its adaptability across diverse biomedical domains. This innovative approach balances interpretability and precision, providing transparent, evidence-based insights crucial for safeguarding public health. The code and data for CER are released for transparency and reproducibility, available at https://github.com/PRAISELab-PicusLab/CER.
Also Read:
- A New AI System for Verifying Health Information Across Text and Video
- Evaluating AI’s Medical Accuracy: A New Benchmark for Chinese Healthcare Texts
Future work aims to expand CER’s evidence retrieval to additional biomedical databases for richer context and to enhance domain generalization through adaptive training or the creation of more diverse datasets.
