Using Machine Learning for Road Accident Severity Prediction and Optimal Rescue Pathways

Road traffic accidents remain a critical challenge for urban planners, transportation authorities, and public safety stakeholders worldwide. Despite advancements in road infrastructure and traffic management, the frequency and severity of accidents continue to strain emergency response systems and compromise public safety. Accident hotspots, are particularly concerning, as they often exhibit recurring patterns of accidents due to factors such as poor road design, high traffic density, and adverse weather conditions. This thesis addresses the challenge of predicting accident severity and identifying accident-prone areas through a data-driven approach. Using a large dataset of over 7.7 million accident records containing geographical, environmental, and temporal features, this thesis develops machine learning models to forecast accident severity and detect spatial clusters of high-risk zones. By integrating historical accident data with real-time factors like weather and road conditions, the thesis aims to create a system that informs proactive interventions and optimizes emergency response strategies. The methodology involves data preprocessing to ensure the dataset's quality and suitability for analysis. Exploratory Data Analysis (EDA) shows insights into accident patterns, highlighting the influence of adverse weather, low visibility, and temporal factors on severity. Accident hotspots are identified using the DBSCAN clustering algorithm, which effectively identifies high-density zones based on geographical coordinates. For severity prediction, Random Forest, XGBoost, Support Vector Machines, and Decision Tree models are trained and evaluated. The Random Forest model achieves the highest accuracy (89.6%) and provides actionable insights into key predictors, such as weather conditions, visibility, and location-specific factors. To further enhance interpretability, cluster-specific decision trees generate localized rules for mitigating accidents in high-risk areas. The findings of this research underscore the importance of integrating spatial and predictive analyses for traffic safety. The identification of urban intersections and poorly managed highways as primary hotspots provides a basis for targeted interventions, such as improved road lighting, better signage, and optimized traffic signals. The severity prediction models demonstrate strong performance, particularly in forecasting moderate accidents, while highlighting the challenges of predicting rare but severe accidents. Cluster-specific rules offer localized solutions, enabling authorities to prioritize resource allocation and implement context-sensitive measures.