Crash Injury Severity Prediction Using an Ordinal Classification Machine Learning Approach.

Shengxue Zhu,Chongyi Li,Ke Wang

doi:10.3390/ijerph182111564

Shengxue Zhu, Chongyi Li + Show 1 more

Open Access

https://doi.org/10.3390/ijerph182111564

Copy DOI

Abstract

In many related works, nominal classification algorithms ignore the order between injury severity levels and make sub-optimal predictions. Existing ordinal classification methods suffer rank inconsistency and rank non-monotonicity. The aim of this paper is to propose an ordinal classification approach to predict traffic crash injury severity and to test its performance over existing machine learning classification methods. First, we compare the performance of the neural network, XGBoost, and SVM classifiers in injury severity prediction. Second, we utilize a severity category-combination method with oversampling to relieve the class-imbalance problem prevalent in crash data. Third, we take advantage of probability calibration and the optimal probability threshold moving to improve the prediction ability of ordinal classification. The proposed approach can satisfy the rank consistency and rank monotonicity requirement and is proved to be superior to other ordinal classification methods and nominal classification machine learning by statistical significance test. Important factors relating to injury severity are selected based on their permutation feature importance scores. We find that converting severity levels into three classes, minor injury, moderate injury, and serious injury, can substantially improve the prediction precision.

Highlights

The prediction and cause analysis of traffic crashes has always been an important topic for scholars in traffic safety
Iranitalab and Khattak [5] compared the performance of a statistical model, Multinomial Logit (MNL), with three machine learning methods including Nearest
The gap between the performance of XGBoost and Multi-layer Perceptron (MLP) may be caused by the data characteristic that most variables are categorical

Summary

Introduction

The prediction and cause analysis of traffic crashes has always been an important topic for scholars in traffic safety. A statistical model usually specifies the mathematical relationship between explanatory variables and crash severity. Based on strict assumptions of uncertainty distribution and hypothesis tests, the statistical model can isolate the effects of explanatory variables on crash severity [1,2]. Cerwick et al [3] used the mixed logit model and the latent class multinomial logit model to predict crash severity. Haghighi et al [4] used standard ordered logit (SOL) and Multilevel ordered logit (MOL) to analyze the effect of roadway geometric features on crash severity. Statistical models are usually weaker in making predictions than machine learning methods. Iranitalab and Khattak [5] compared the performance of a statistical model, Multinomial Logit (MNL), with three machine learning methods including Nearest

Objectives

Methods

Results

Conclusion