R-TAP: A Cloud-Based Multi-Region Road Traffic Accident Severity Prediction Model

The tremendous impact of road traffic accidents (RTAs) remains a global concern, especially in low- and middle-income countries (LMICs). While most LMICs have limited budget to afford advanced traffic monitoring systems, it is necessary to develop feasible solutions for enhancing road safety in these regions. Recent studies have extensively explored the applicability of machine learning (ML) models to predict RTA severity in different regions. However, there is limited exploration of training and evaluating the performance of ML model from a multi-region perspective. Further investigation is also necessary to formulate a cost-effective implementation strategy for LMICs to deploy ML models for real-time prediction. Hence, this paper contributes by proposing R-TAP, which deploys a TabNet-based model on the cloud for multi-region RTA severity predictions. A web application is also developed to receive inputs and visualise the prediction results. The TabNet model is trained on a multi-region dataset consisting of 42 features, produced by combining the individual RTA datasets collected from Australia, Brazil, Ethiopia, New Zealand, the UK, and the US. Evaluation with random forest and multinomial logistic regression showed that TabNet can achieve performance comparable to random forest. It also has the highest area under the receiver operating characteristics at 0.8857. R-TAP is also simulated using testing samples as real-time input, with an overall execution time of 4.104s per prediction. It is considered satisfactory and is expected to be useful for emergency responders to make informed decisions. This solution is expected to support LMICs in their roadmap towards enhanced road safety. Future research is essential to improve the predictive power and include datasets from more regions for a more inclusive RTA severity prediction.