From knowledge generation to knowledge verification: examining the biomedical generative capabilities of ChatGPT

The generative capabilities of LLM models offer opportunities for accelerating tasks but raise concerns about the authenticity of the knowledge they produce. We present a computational approach that evaluates the factual accuracy of biomedical knowledge generated by an LLM. Our approach consists of generating disease-centric associations and verifying them using biomedical ontologies. Using ChatGPT, we designed prompt-engineering processes to establish linkages between diseases and related drugs, symptoms, and genes, and assessed consistency across multiple ChatGPT models (e.g., GPT-4, GPT-4o, and GPT-4o-mini). Results demonstrate high accuracy in identifying disease terms (88%-97%), drug names (90%-91%), and genetic information (88%-98%). Symptom term identification was lower (49%-61%) due to informal symptom descriptions. Verification reveals coverage of 89%-91% for disease-drug and disease-gene pairs; symptom-related associations show lower coverage (49%-62%). Despite high term accuracy, generated IDs were often invalid or redundant. GenAI tools can be reliable if used with care. Retrieval Augmented Generation (RAG) may enhance reliability.