Comparison of machine learning methods for traffic accident injury severity prediction

Road traffic accidents constitute a substantial worldwide public health emergency, leading to millions of deaths and injuries each year. Traditional statistical models have been used commnly to analyze accident severity, but they are not very useful since they cannot accurately capture complicated data patterns and non-linear connections. This thesis investigates the use of machine learning (ML) as an alternative method for forecasting the severity of vehicle accidents. The objective is to address these limitations and enhance the precision of forecasts. The research used the UK's STATS19 dataset, which comprises records of traffic incidents that result in bodily harm. The dataset includes so many parameters related to accident circumstances and casualties, including sociodemographic variables, naturalistic environmental components, and contributory factors. The inquiry starts by meticulously preprocessing the dataset, ensuring the integrity of the data via the execution of cleaning procedures and the appropriate manipulation of various data types using relevant feature engineering techniques. The approaches include category encoding, data binning, timestamp extraction, and spatial data transformation. Subsequently, the process of feature selection involves first using filtering approaches that depend on Pearson correlation to identify and remove redundant characteristics. Subsequently, wrapper techniques using Recursive Feature Elimination (RFE) and a Logistic Regression classifier are used. Several techniques for dealing with unbalanced data were compared, and it was found that the Random Oversampling tend to be the most suit with the models. The study primarily aims to predict the levels of injury severity, namely "Slight" and "Serious/Fatal" and this done by investigate car accident severity prediction models using accident records by applying machine learning methods and comparing the performances of diffrenet algorithms: Random Forest, Decision Tree, Logistic Regression, Support Vector Classifier, K-Neighbors Classifier, Extreme Gradient Boosting (XGBoost), and Naïve Bayes, to understand which ones are best suited for road crash severity prediction models. The research assesses each model using many metrics, including accuracy, precision, recall, F1-score, G-Mean, and ROC AUC. The evaluation is conducted both with and without the use of data imbalance managing approaches. The research concludes that Random Oversampling is the optimal strategy for mitigating the data imbalance in the STATS19 dataset. The K-Neighbors classifier presents as the most efficient model, performing the greatest G-Mean score and displaying robust discriminating performance according to the ROC AUC study. The thesis finishes by identifying crucial factors that influence accident severity. It also highlights the relevance of data-driven techniques for enhancing road safety.