Abstract: Objective To generate a keyword library for drug inspection defects and analyze classifications and risk levels of the defects based on the keywords and their combinations in different settings. Methods The keywords were retrieved from reports about GMP inspection delivered by the Center for Food and Drug Inspection of NMPA using a large language model. Highly automated methods, including traversal algorithms and VBA functions, were employed to calculate the proportions of defect keywords and keyword groups within their respective classification. Results A detailed keyword database for GMP inspection defects was established, along with a risk level table involving defect keywords, keyword groups, their classifications, and risk levels ranked from the highest to the lowest. Conclusion This research has established a model that integrates artificial intelligence technologies with traditional on-site drug inspections, which can significantly reduce the labor and time required for data compilation and analysis. It pioneers a novel paradigm of intelligent regulatory oversight based on large language models, facilitating a shift from experience-based inspections to data-driven decision-making.

Key words: Drug Inspection, Inspection Defects, Defect, Defects Classifications, Defects Risk Levels, Drug Safety