Predicting Likelihood of Multiple Secondary Crashes Using Ordinal Neural Networks

Document Type

Article

Publication Date

11-2025

Publication Title

2025 IEEE International Conference on Future Machine Learning and Data Science (FMLDS)

Abstract

A secondary crash involves a crash that is a result of a primary crash and plays a great role in injury severity and traffic operations on the roadways. It is estimated that secondary crashes are likely to be 3 to 3.5 times more severe than primary crashes and take more time to clear. While previous studies have focused on single secondary crashes or did not distinguish between single and multiple secondary crashes, the impact of multiple secondary crashes is hypothesized to be severe, thus understanding and distinguishing between the two would benefit traffic operators. Therefore, this study introduces the use of Ordinal Neural Networks (ONN) in predicting multiple secondary crash occurrences which accounts for the ordinal nature of the severity of the crashes and compares its accuracy with the traditional multiclass neural network (MNN). The MNN and ONN models were fitted with similar architecture (i.e., number of neurons and hidden layers) to facilitate accurate comparison performance measures such as balanced accuracy, and confusion matrix were used in the analysis. The ONN model has demonstrated high performance in predicting multiple secondary incidents compared to MNN, and this can be used for proactive measures to minimize the occurrence and impact of secondary crashes.

Comments

Paper presented at 2025 IEEE International Conference on Future Machine Learning and Data Science (FMLDS) held in Los Angeles, CA, November 2-5, 2025

DOI

10.1109/FMLDS67896.2025.00054

Link to Full Text

Share

COinS