Development of a predictive classification model for determining the category of fatal mine accidents using text...

Objectives. The mining sector is a hazardous work environment with high accident rates, causing human suffering and financial losses. Examining incident reports is crucial for preventing comparable events; however, manual classification of large accident datasets is labour-intensive and time-consuming. Methods. This work presents a novel methodology for classifying fatal accident reports using text-mining techniques. It uses natural language processing techniques to transform textual data into a vector representation. Stratified 10-fold cross-validation ensures training and test sets maintain equivalent class distributions from the original dataset, improving model performance. Six supervised machine learning models - logistic regression, support vector machine (SVM), random forest, naïve Bayes, decision tree and multilayer perceptron (MLP) - were employed to classify 1308 fatal accident records into eight accident types. Results. The analysis demonstrates strong model accuracy. The MLP model achieved superior overall performance with a 0.84 weighted average F1 score, followed by logistic regression (0.83), SVM (0.81), random forest (0.73), naïve Bayes (0.65) and decision tree (0.57). Conclusion. This study presents an automated system for identifying accident types from incident reports. The proposed approach reduces misclassifications and mitigates human biases in incident report analysis, offering a reliable tool for mining safety management.