US Accident Severity Intelligence: Intelligent Prediction System for US Traffic Accident Severity

This article introduces an open-source machine learning-based project for predicting the severity of traffic accidents in the US. By analyzing real traffic data, the project builds an intelligent prediction pipeline to provide data support for traffic safety management and accident prevention.

Traffic accidents are one of the leading causes of casualties and property damage worldwide. Accurately predicting the severity of accidents is of great value for the rational allocation of emergency response resources, insurance risk assessment, and the formulation of traffic safety policies. The US Accident Severity Intelligence project was developed based on this need; it uses machine learning technology to conduct in-depth analysis of US traffic accident data and build a complete accident severity prediction system.

This project not only demonstrates the practical application of data science in the field of public safety but also provides researchers and practitioners with a reusable machine learning engineering template covering the complete process from data preprocessing to model deployment.

The project is based on the US public traffic accident dataset, which contains traffic accident information recorded across the US over several years. The dataset covers multi-dimensional information such as the spatiotemporal characteristics of accidents, environmental conditions, road conditions, and accident outcomes.

The project extracts and processes the following key features:

Spatiotemporal Features:

• Time of accident (hour, day of week, month)
• Geographic location information (latitude, longitude, city, state)
• Accident duration

Environmental Conditions:

• Weather conditions (sunny, rainy, snowy, foggy, etc.)
• Visibility level
• Wind speed and direction
• Temperature and humidity

Road and Traffic Features:

• Road type (highway, urban road, rural road, etc.)
• Intersection and traffic signal conditions
• Road surface conditions (dry, wet, icy, etc.)
• Traffic flow information

Accident Features:

• Number of vehicles involved
• Accident type (rear-end collision, side collision, rollover, etc.)
• Whether pedestrians or cyclists are involved

The project uses a variety of feature engineering techniques to improve model performance:

• Encoding Processing: One-Hot encoding and label encoding for categorical variables
• Feature Scaling: Standardization and normalization of numerical features
• Feature Selection: Using correlation analysis and feature importance evaluation to select effective features
• Feature Construction: Creating interaction features (e.g., combination of weather and time)

The project builds an automated data preprocessing pipeline:

Data Cleaning Phase:

• Handling missing values (deletion, filling, interpolation)
• Identifying and handling outliers
• Correcting data format inconsistencies
• Removing duplicate records

Data Transformation Phase:

• Feature type conversion (string, numerical, datetime)
• Standardization of geographic coordinates
• Periodic encoding of time features

The project implements multiple machine learning algorithms for comparative experiments:

Traditional Machine Learning Models:

• Logistic Regression
• Random Forest
• Gradient Boosting Trees (XGBoost, LightGBM)
• Support Vector Machine (SVM)

Ensemble Learning Methods:

• Voting Ensemble
• Stacking Ensemble
• Bagging and Boosting strategies

The project establishes a comprehensive model evaluation framework:

Classification Metrics:

• Accuracy
• Precision
• Recall
• F1 Score
• ROC-AUC Curve

Multi-class Evaluation:

• Macro-average and weighted average metrics
• Confusion matrix analysis
• Detailed performance reports for each category