Crash injury severity analysis using a two-layer Stacking framework.

Area of Science:

• Traffic Safety
• Machine Learning
• Data Science

Background:

• Understanding crash injury severity is crucial for effective traffic management.
• Existing models often struggle to accurately predict the severity of traffic incidents.

Purpose of the Study:

• To develop and validate a novel Stacking framework for predicting crash injury severity.
• To compare the proposed model's performance against traditional machine learning methods.

Main Methods:

• A two-layer Stacking model integrating Random Forests (RF), AdaBoost, and Gradient Boosting Decision Tree (GBDT) in the first layer, with Logistic Regression in the second.
• Model calibration involved systematic grid search for parameter optimization.
• Model validation compared the Stacking model against Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and RF using accuracy and recall metrics.

Main Results:

• The Stacking model demonstrated superior performance in predicting crash injury severity compared to SVM, MLP, and RF.
• Accuracy and recall were key indicators used for performance evaluation.
• Sensitivity analysis identified and categorized factors influencing prediction accuracy.

Conclusions:

• The proposed Stacking framework offers a more accurate approach to predicting traffic crash injury severity.
• The findings can aid traffic management agencies in improving road safety and resource allocation.
• Further analysis of significant factors can lead to targeted safety interventions.