Optimizing Accident Severity Prediction: A Comparative Study of LSTM, CatBoost, and CNN Models with Feature Selection

Abstract

Road accidents have recently emerged as a significant threat, ranking as the ninth leading cause of fatalities globally. The high cost of traffic-related deaths and property damage is particularly burdensome in developing countries. Investigating the factors contributing to accidents and accurately predicting accident severity are crucial steps in mitigating future incidents. Traditional methods for predicting road accident severity have relied on shallow models and statistical approaches, with limited exploration of deep learning techniques. This study conducts a comparative analysis of LSTM, CatBoost, and CNN models for predicting road accident severity, aiming to identify the model that most accurately forecasts accident severity based on road accident data. To address the challenges of small datasets, limited coverage, and real-time applicability, we applied these models to both Balanced and Unbalanced US Accident datasets. Additionally, three feature selection algorithms Random Forest, Decision Tree, and CatBoost were employed to extract the most relevant features from the datasets. Our results demonstrate that the combined approach of CatBoost feature selection and LSTM modeling outperforms standalone models, achieving an accuracy of 98.57%.