Prediction of hazards in laboratory work using deep learning models learnt from past laboratory accidents

Abstract

Laboratory-based research activities frequently involve hazardous materials and operations, which are prone to accidents or injuries. While risk assessment is a necessary step in the research plan, it is often difficult to recognize all potential hazards involved in laboratory work, including those in the preparation and clean-up phases. In this study, we investigate the performance of deep learning models in predicting potential hazards in laboratory work. As a training dataset, actual laboratory accident reports collected from national universities in Japan are labeled with the most suitable hazards. The trained models read a text of laboratory work plan as an input and predict the possible primary and secondary hazards as outputs. The model that combines Bidirectional Encoder Representations from Transformers (BERT) with Bidirectional Long Short-Term Memory (BiLSTM) shows higher performance compared to that of its constituent models individually. The trained models can therefore potentially be used as a core component of risk assessment tool and safety training, enabling junior researchers and students to recognize potential hazards and assess possible accident risks in their laboratory work, thus reducing the frequency of accidents and injuries.