Abstract

Road traffic accidents remain a persistent public safety challenge, responsible for over 1.3 million deaths annually according to the World Health Organization. While considerable effort has gone into understanding accident patterns, predicting the severity of an accident before or immediately after it occurs remains a difficult problem — particularly when the underlying data is messy, imbalanced, and largely categorical in nature. This study applies three supervised machine learning classification algorithms — Logistic Regression, Random Forest, and Gradient Boosting — to predict the severity of road traffic accidents using a dataset of 12,316 records collected by the Addis Ababa Sub-city Police De- partment in Ethiopia. The dataset contains 32 variables spanning driver characteristics, vehicle information, road conditions, and casualty details, with Accident severity as the target variable across three classes: Slight Injury, Serious Injury, and Fatal Injury. Prior to modelling, the data underwent a structured cleaning process in which four variables exceeding 20% missing data were dropped, and the remaining gaps were filled using mode imputation. A Pearson correlation analysis was then conducted to examine relationships between features, revealing that Casualty severity holds the strongest linear association with the target variable (r= 0.42). Feature importance analysis from the Random Forest model further identified Cause of accident, Day of week, and Type of vehicle as the three most influential predictors, suggesting that behavioural and contextual factors carry more predictive weight than road surface or weather conditions alone. Among the three models evaluated, Gradient Boosting achieved the best overall performance with an accuracy of 84.94% and a weighted F1-score of 0.79. It was also the only model to correctly identify any Fatal Injury cases — a finding that reflects both the strength of the boosting approach and the challenge posed by severe class imbalance in the dataset. Logistic Regression served as a useful baseline but struggled to predict minority classes, while Random Forest showed moderate improvement over the baseline. The findings of this study demonstrate that machine learning methods, and Gradient Boost- ing in particular, offer a practical framework for accident severity prediction from routine police record data. The results also highlight the importance of addressing class imbalance in future work, where techniques such as SMOTE or cost-sensitive learning could further improve the detection of serious and fatal accident cases.

Publication Date

9-2026

Document Type

Thesis

Student Type

Graduate

Degree Name

Professional Studies (MS)

Department, Program, or Center

Professional Studies

Advisor

Ioannis Karamitsos

Campus

RIT Dubai

Download

Share

COinS