Abstract
Traffic safety is a major public health issue worldwide. The Kingdom of Saudi Arabia (KSA) is facing alarming road safety concerns with traffic-related injuries being the third leading cause of fatalities in the country. A better understanding of factors influencing injury severity outcomes of traffic crashes is vital for the proactive and effective implementation of suitable countermeasures. To address the drawbacks of statistical models lately, the application of various non-parametric machine learning (ML) based techniques have been increasingly used for modeling the injury severity of traffic crashes. The dissertation proposes the application of six different ML algorithms for injury severity prediction of traffic crashes based on three years data on interstate rural highways in KSA. Injury severity predictive performance of proposed algorithms was assessed in terms of various evaluation indices. Experimental results revealed that some models outperformed others based on the selected evaluation metrics. To address the ML model’s interpretability issue, feature importance and SHAP analysis were also employed. An analysis of several significant factors which aggravate the chances of fatal and severe injuries was done. The study also proposes the application of the Information Root Node Variation (IRNV) technique for extraction of significant decision rules highlighting the circumstances for the categorization of specific crash injury severity instances. For comparison purposes, multinomial logit (MNL) models were also developed to investigate the consistency of injury severity risk factors between MNL and ML models. The outcomes and findings of the current study can yield valuable insights to safety practitioners for the timely and effective implementation of suitable mitigation measures.