A Risk-Aware Communication Framework for Intelligent Transportation Systems Using Machine Learning-Based Severity Scoring and Message Prioritization

Abstract

In recent years with over 90% of car accidents attributable to human error, especially during intricate driving manoeuvres, road safety remains a crucial challenge in the development of intelligent transportation systems. Current frameworks for accident prediction frequently lack real-time flexibility and communication system integration. This study uses a large dataset including environmental, temporal, and geospatial characteristics to present a machine learning-based severity prediction framework for evaluating the risks of traffic accidents. To forecast severity levels, the system uses a stacking ensemble model in conjunction with many classifiers, including Random Forest, Logistic Regression, and XGBoost. Accuracy, F1-score, and ROC- AUC are used to assess performance; the ensemble model achieved a ROC-AUC of more than 94%. Over 84% accuracy is still difficult to achieve, though, which motivates feature engineering and hyperparameter optimisation for additional advancements. A new Severity Score is presented, which measures the likelihood of collisions by dividing the number of high-impact severity predictions (classes 3 and 4) by the total number of recorded occurrences. In a connected environment, this score is further utilised to filter messages and lower communication overhead between agents. By giving high- severity vehicles priority and maximising message distribution, the suggested paradigm facilitates proactive risk-aware communication. The introduction of scalable and secure autonomous mobility systems is supported by experimental results showing notable gains in communication latency and decision-making efficiency.