Classification of Occupational Accident Risks through Machine Learning Algorithms

Abstract

This study proposes a Machine Learning (ML_-based framework for the automatic classification of accident narratives into Risk and Non-Risk categories to support the risk identification phase of occupational risk management. A manually labeled dataset containing 278 accident narratives derived from the publicly available Work Accidents in China dataset was constructed and preprocessed. Four textual representation techniques, including Bag-of-Words (BoW), Term Frequency-Inverse Document Frequency (TF-IDF), Word2Vec, and Bidirectional Encoder Representations from Transformers (BERT) embeddings, were evaluated in combination with four ML classifiers: Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), and Naive Bayes (NB). To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied. Model performance was assessed using stratified 5-fold cross-validation. The experimental results indicate that the choice of textual representation has a significant influence on classification performance. The best overall results were achieved by the SVM + BoW combination, which obtained an accuracy of 93.17 ± 4.16% and an F1-score of 93.10 ± 4.27%. Although BERT embeddings achieved competitive performance, the feature extraction approach adopted in this study did not outperform the best traditional ML configuration on the evaluated dataset. These findings suggest that, for the evaluated dataset, lightweight and interpretable text representations remain highly effective for Risk/Non-Risk classification. They also indicate the potential of the proposed framework to support decision-making during the risk identification phase of occupational risk management.