Advancing proactive crash prediction: A discretized duration approach for predicting crashes and severity

Area of Science:

• Transportation Engineering
• Statistical Modeling
• Road Safety

Background:

• Proactive crash prediction models increasingly use machine learning and AI.
• Statistical models offer causal insights and effect size estimation.
• Existing statistical models often use a case-control approach with challenges in ratio definition and severity incorporation.

Purpose of the Study:

• To extend the duration-based modeling framework for predicting traffic crashes and their severity.
• To investigate the trade-off between model performance and estimation time when incorporating crash severities.
• To identify optimal data sampling strategies for robust crash prediction.

Main Methods:

• Developed a novel duration-based framework integrating crash severity.
• Explored data sampling strategies (15% epoch-level sample) to manage computational complexity.
• Conducted stability analysis of predictor variables across different sample sizes.

Main Results:

• A 15% epoch-level sample provides a balanced approach between data size and predictive accuracy.
• Certain variables (e.g., Time of day, Weather, Lighting, Volume) require larger samples for stable estimation.
• Other variables (e.g., Daytime, Terrain, Land use, Number of lanes, Speed) converge with smaller sample increases.
• The model shows better performance on highway segments with shorter crash durations (<100 hours).

Conclusions:

• The proposed duration-based framework effectively predicts crashes and severity.
• Epoch-level data sampling, specifically 15%, is a viable strategy for balancing model performance and computational efficiency.
• Understanding variable sensitivity to sample size aids in optimizing data collection and model development for road safety.