A comparative analysis for automated information extraction from OSHA Lockout/Tagout accident narratives with Large Language Model

Analysing workplace accidents is crucial for improving occupational safety by understanding causes and preventing recurrence. However, the primary challenge in analysing accident narratives lies in the unstructured nature of the text data. This study examines the effectiveness of Large Language Models (LLMs), specifically GPT-4 Turbo, in extracting information from lockout/tagout (LOTO) accident narratives in the Occupational Safety and Health Administration (OSHA) database. It compares the extracted features, namely the degree of fatality, nature of injury, and employee's occupation, with those recorded by OSHA supervisors. Despite occasional misclassifications and hallucinations, GPT-4 Turbo shows significant potential in automating critical information extraction, reducing reliance on human interpretation. Moreover, the model achieved high accuracy rates for each feature. These findings suggest that LLMs can enhance occupational safety data analysis, though improvements in prompt design and verification are recommended for further accuracy.