Transparent deep machine learning framework for predicting traffic crash severity

Analysis of crash injury severity is a promising research target in highway safety studies. A better understanding of crash severity risk factors is vital for the proactive implementation of suitable countermeasures. In literature, crash injury severity was widely studied using statistical models. Though these models have a sound theoretical basis and interpretability, they were based on several unrealistic assumptions, which, if flouted, may yield biased model estimations. To overcome the limitations of statistical models, applied machine learning has rapidly emerged on the horizon of highway safety analysis. This study aims to model injury severity of motor vehicle crashes using three advanced machine learning approaches, i.e., vanilla multi-layer perceptron (MLP) using Keras, MLP with embedding layers, and TabNet. Among the three models, TabNet may be considered a fairly complex framework which is based on attention-based network for tabular data. To improve the predictive performance of proposed models, hyperparameter tuning was carried out using the Bayesian optimization technique. Different evaluation metrics (i.e., accuracy, precision, recall, F−1 score, AUC, and training time) were utilized to compare all the models' injury severity classification performance. Experimental results showed that all the models yielded similar and adequate performance based on most of the evaluation metrics. However, based on training time, the Keras (MLP) model outperformed other models with a training time of 3.45 s which represents a reduction of 51% and 93% compared to MLP with embedding layers and TabNet, respectively. Feature importance analysis conducted using TabNet revealed that predictors such as number of vehicles involved, number of casualties, speed limit, junction location, vehicle type, and road type are the most sensitive variables aggravating the injury severity. The proposed supervised deep learning models supported by feature importance analysis make the modeling framework transparent and interpretable. The outcome of this study could provide essential guidance for practitioners for taking timely and concrete steps to improve highway safety. Moreover, this research will allow trauma and emergency centers to predict possible damage from a traffic accident and deploy the necessary emergency units to offer appropriate emergency treatment.