Traffic Crash Analysis of Three-Wheelers in Rawalpindi: Prediction Modeling using Machine Learning and Emerging Hotspot Analysis

Abstract

The three-wheeled motorized rickshaw (3-MR) is the dominant mode of transportation in developing countries for short trips with passenger carrying capacity of four to six and movement of goods at a small-scale level. These 3-MRs are often associated with road traffic crashes causing significant socioeconomic and public health concerns. Early research mostly focused on the safety analysis of two and four-wheelers, leading to a research gap in exploring the safety dynamics associated with 3-MR crashes. Few recent studies investigated the effect of different factors on injury severity of 3-MR crashes using statistical and machine learning models, but no study evaluated the spatiotemporal dimension of the 3-MR crashes. This study consists of two parts. In first part study aims to identify crash hotspots using 3-MR crash data for sixteen months (January 2022 to April 2023) in Rawalpindi, Pakistan. The study employs cutting-edge spatial analytic methods, such as emerging hotspot analysis with a space-time cube and global and local spatial autocorrelation. Significant clustering in the form of hotspots is observed, employing Moran's index for global spatial autocorrelation detection, and local spatial autocorrelation was assessed using Getis Ord Gi*, which supports a recurring pattern in the 3-MR crashes. Incremental spatial autocorrelation was also employed, indicating 272 meters as the optimal distance for clustering detection. Further utilizing the emerging hotspot analysis technique, substantial spatiotemporal clustering is identified. Critical hotspots were identified near major bus stops, commercial areas, intersections, hospitals, airports, and high-density residential units. In second part Artificial neural network model with back propagation algorithm was developed to predict injury severity of 3-MR traffic crashes using some key variables like emergency response time, that were not considered in past xv research. Synthetic Minority Oversampling Technique (SMOTE) was also employed to deal with class imbalance between Severe and Minor injury severity. Model achieved an overall accuracy of 74.5% and Precision of 75.1% with true positive ratio of 73.7% and false positive ratio of 75.3%. Receiver Operating Characteristics (ROC) curve was also analyzed, area under the curve value of 0.809 for both severe and minor injury classes depicted ability of model to distinguish between both classes. Sensitivity Index further ranked predictors according to variable importance and their impact on injury severity. Model identified higher response times, over speeding, distracted driving behaviors, and higher temperature increases the risk of severe injury crash. The findings of our study have significant practical implications for stakeholders, providing them with a foundation and pinpoint locations in the form of hotspots and key factors contributing to crash injury severity, to update and refine policies specifically aimed at addressing recurrent 3-MR crashes and enhancing overall road traffic safety for vulnerable road users.