Scholarlink

<p>This study applies Bayesian Probabilistic Structural Equation Modeling (BPSEM), implemented as a Bayesian network with latent factors, to examine pedestrian crash severity in Abu Dhabi. We analyzed 2,630 pedestrian injury records from 2013 to 2019. The framework uses four steps: unsupervised network learning to identify probabilistic dependencies among observed variables, hierarchical variable clustering, data clustering to construct latent factors, and target optimization for fatal injury severity. Five latent domains were identified: spatial/infrastructure, driver-vehicle, road/intersection, pedestrian demographics and transport, and environmental conditions. Among recorded pedestrian crashes, the baseline fatality probability was 13.54%. Evidence states with the highest posterior fatality probabilities were high-speed roads (120 to 140&#xa0;km/h; approximately 33 to 35%), external roads (approximately 32%), metallic median barriers (approximately 30%), and crossings outside designated lines (approximately 15% individually and influential in multiple-evidence scenarios). Joint-probability analysis highlighted prevalent crash contexts, including male pedestrians and non-intersection locations; these findings are interpreted as exposure-influenced crash profiles rather than causal hazards. Cross-validation showed stable network structure and reasonable agreement within coarse predicted risk tiers (mean calibration index = 89.02%) but modest discrimination (fatal-class ROC-AUC = 0.61; R-squared = 0.063), and maximum-likelihood classification did not recover fatal cases. Therefore, the findings are presented as conditional associations within recorded crashes, not exposure-adjusted risks or causal effects. The results support targeted safety audits of external/high-speed corridors, median design, pedestrian access control, and designated crossing provision.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Understanding pedestrian crash patterns using bayesian probabilistic structural equation modeling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Md Didarul Alam,
  • Luqman Ali,
  • Hamad AlJassmi

摘要

This study applies Bayesian Probabilistic Structural Equation Modeling (BPSEM), implemented as a Bayesian network with latent factors, to examine pedestrian crash severity in Abu Dhabi. We analyzed 2,630 pedestrian injury records from 2013 to 2019. The framework uses four steps: unsupervised network learning to identify probabilistic dependencies among observed variables, hierarchical variable clustering, data clustering to construct latent factors, and target optimization for fatal injury severity. Five latent domains were identified: spatial/infrastructure, driver-vehicle, road/intersection, pedestrian demographics and transport, and environmental conditions. Among recorded pedestrian crashes, the baseline fatality probability was 13.54%. Evidence states with the highest posterior fatality probabilities were high-speed roads (120 to 140 km/h; approximately 33 to 35%), external roads (approximately 32%), metallic median barriers (approximately 30%), and crossings outside designated lines (approximately 15% individually and influential in multiple-evidence scenarios). Joint-probability analysis highlighted prevalent crash contexts, including male pedestrians and non-intersection locations; these findings are interpreted as exposure-influenced crash profiles rather than causal hazards. Cross-validation showed stable network structure and reasonable agreement within coarse predicted risk tiers (mean calibration index = 89.02%) but modest discrimination (fatal-class ROC-AUC = 0.61; R-squared = 0.063), and maximum-likelihood classification did not recover fatal cases. Therefore, the findings are presented as conditional associations within recorded crashes, not exposure-adjusted risks or causal effects. The results support targeted safety audits of external/high-speed corridors, median design, pedestrian access control, and designated crossing provision.