An interpretable deep machine learning model for crash severity prediction-Use of international big data

This study sets a goal to pave the path towards a generalized (universal) crash severity prediction model through the synergy between deep learning (DL) model and international crash data. Such a model can be applied for predicting crash severity for countries which have less or no crash data of their own. This was achieved through the collection of crash data from the open sources of four countries namely: USA, GB, Australia and Canada. The international crash database consisted of 276,000 evenly distributed among all countries. Then, seven common machine learning (ML) models were selected and developed. A confusion matrix was used to find the accuracy, precision, recall and F1 score of each model, other evaluation metrics included ROC-AUC and feature importance. Lastly, a comparative analysis was conducted between the eight models on measures of accuracy. The results of this study indicate that a development of a generalized model with high prediction accuracy is possible using DL model. The model showed that vehicle type, speed limit, accident type and light conditions are the most important variables, affecting the crash severity. Models for different countries highlighted the fact that each country may also have, in addition to the above, different parameters contributing to the severity of crashes. It is expected that the model developed in this study could be used by the authorities to identify road conditions, environmental conditions and driver profile which are contributing to severity of crashes, consequently, allowing effective design modifications and policy measures to mitigate severity of crashes.