Uncovering patterns in ferry near misses: A topic modelling approach with GPT-assisted labelling

Near-miss incident reports provide valuable but underutilised insights for improving maritime safety. However, existing research emphasises failures, with limited focus on positive safety practices. This study addresses this gap by examining connections between incidents and preventive measures to identify what works and what doesn’t. To overcome the limitations of traditional topic modelling approaches, this study applies BERTopic, a transformer-based method. Using this method, we analysed 4360 near-miss reports spanning five years from a ferry operator to identify recurring risk patterns and the preventative measures linked to them. GPT-assisted labelling improved interpretability, with outputs validated through structured multi-expert review. Over 75 distinct incident topics were identified across operational, environmental, and human factors domains, aligning with existing categories and uncovering previously unclassified risk themes. Time-series analysis highlighted seasonal variations, with passenger-related risks peaking in summer and infrastructure failures more common in winter. Prevention measures were modelled separately, producing 87 topics, with association rule mining identifying clear linkages between incident types and mitigations; for example, navigation-related incidents were frequently linked to improved bridge communication. Overall, combining unsupervised topic modelling with LLM-assisted labelling produces interpretable and operationally meaningful outputs. This approach offers a scalable framework for prioritising preventative measures using data-driven evidence.