Analyzing feature importance for older pedestrian crash severity: A comparative study of DNN models, emphasizing road and vehicle types with SHAP interpretation

Abstract

Recognizing the importance of road safety modeling, the study explores Deep Neural Networks (DNN) with features like hidden layers, batch normalization, Rectified Linear Unit (ReLU) activation, and dropout to predict crash severity, interpreting decisions using SHapley Additive exPlanations (SHAP) for crashes involving older pedestrians. The objective is to understand features influencing crashes involving older pedestrians, including vehicle attributes, road and environmental conditions, and temporal parameters. The analysis focused on 1808 pedestrian crashes involving individuals aged 65 and over at intersections in Victoria, Australia. This dataset comprises 6.14% fatalities, 52.38% serious injuries, and 41.48% incidents with other injuries. The study evaluated three DNN models for crash severity prediction, with the two hidden layers DNN model excelling in precision metrics and achieving a perfect Area Under the Receiver Operating Characteristics curve for fatalities. Compared to XGBoost, the DNN models demonstrated superior performance in predicting severe outcomes. SHAP analysis on the two hidden layers DNN model highlighted key factors influencing crash severity, offering insights into the nuanced relationships between features and predictions. The analysis highlighted the significance of variables like Traffic Control, Vehicle Type, and Movement in predicting fatalities and serious injuries. This study emphasizes the importance of considering Road and Vehicle Types to understand their roles in accident severity and identify interventions to reduce risks. Neglecting these factors may lead to incomplete or biased conclusions about crash outcomes. This research provides valuable insights for improving road safety, highlighting the effectiveness of SHAP force plots, bars, beeswarm plots, and dependency plots in enhancing clarity and understanding of DNN model predictions. These tools help identify the impact of features on crash severity.