Deep-Learning-Based Prediction of High-Risk Taxi Drivers Using Wellness Data

Abstract

Background: Factors related to the wellness of taxi drivers are important for identifying high-risk drivers based on human factors. The purpose of this study is to predict high-risk taxi drivers based on a deep learning method by identifying the wellness of a driver, which reflects the personal characteristics of the driver. Methods: In-depth interviews with taxi drivers are conducted to collect wellness data. The priorities of factors affecting the severity of accidents are derived through a random forest model. In addition, based on the derived priority of variables, various combinations of inputs are set as scenarios and optimal artificial neural network models are derived for each scenario. Finally, the model with the best performance for predicting high-risk taxi drivers is selected based on three criteria. Results: A model with variables up to the 16th priority as inputs is selected as the best model; this has a classification accuracy of 86% and an F1-score of 0.77. Conclusions: The wellness-based model for predicting high-risk taxi drivers presented in this study can be used for developing a taxi driver management system. In addition, it is expected to be useful when establishing customized traffic safety improvement measures for commercial vehicle drivers.