Pre-crash injury risk prediction with guaranteed confidence level: a conformal and interpretable framework

Area of Science:

• Traffic Safety Engineering
• Machine Learning Applications
• Risk Prediction Modeling

Background:

• Traditional pre-crash injury risk models lack interpretability and reliable uncertainty estimation.
• This limits their effectiveness in proactive safety measures and injury severity mitigation.
• There is a need for interpretable models that provide reliable uncertainty quantification.

Purpose of the Study:

• To develop a novel framework for pre-crash injury risk prediction using only pre-crash data.
• To output potential injury risk distributions with corresponding probabilities at a 90% confidence level.
• To enhance interpretability and provide data-driven guidance for injury mitigation strategies.

Main Methods:

• Utilized data from the National Automotive Sampling System-Crashworthiness Data System and Crash Investigation Sampling System.
• Incorporated 28 pre-crash risk factors and evaluated machine learning models including TabNet.
• Applied conformal prediction methods (naive and class-conditional) for 90% confidence prediction sets and addressed class imbalance with resampling strategies.

Main Results:

• Achieved nearly 90% prediction coverage and a 70.3% recall rate for serious injuries.
• Significantly outperformed existing studies in prediction performance.
• Identified key risk factors such as intersection relevance, crash type, and speed limits influencing injury severity.

Conclusions:

• The proposed framework significantly enhances pre-crash injury risk prediction using conformal prediction techniques.
• It offers improved predictive performance, enhanced interpretability through uncertainty quantification, and identification of key risk factors.
• The framework's validity under distribution shifts and combined uncertainty estimation with interpretability provide a foundation for proactive traffic safety applications and policy development.