Road accidents pose a significant global health and economic challenge. This research conducts a comparative analysis of machine learning (ML), deep learning (DL), and Transformer-based natural language processing (NLP) techniques to predict the specific type of fatal casualty in British road accidents, aiming to enhance understanding and prevention efforts. The study utilizes two correlated datasets detailing fatal road accidents and associated casualties in Great Britain from 2006 to 2008, sourced from data.gov.uk. Data preprocessing included exploratory data analysis, handling missing values, feature engineering, and correlation analysis. For traditional ML models (Logistic Regression, k-Nearest Neighbors, Decision Tree, Random Forest, XGBoost, CatBoost), class imbalance was addressed using SMOTE, and models were tuned using Randomized Search. For DL (TabNet, MLP) and Transformer models (BERT, RoBERTa, DistilBERT), a standard train/validation/test split was used without SMOTE, and tabular data was converted into descriptive text sequences for Transformer input. Model performance was assessed using accuracy, F1-score, the area under the receiver operating characteristic curve, and confusion matrices on a held-out test set. Explainability for the best models was explored using LIME. Results indicate that tuned gradient boosting models (Random Forest, XGBoost) achieved the highest accuracy (approx. 87%). Our research highlights the capabilities and trade-offs of diverse modeling approaches for identifying risk factors associated with fatal casualty types, informing targeted road safety strategies.