Enhancing road safety with machine learning: Current advances and future directions in accident prediction using non-visual data

Road traffic accident (RTA) poses a significant road safety issue due to the increased fatalities worldwide. To address it, various artificial intelligence solutions are developed to analyze RTA characteristics and make predictions. This study contributed to a systematic review of machine learning (ML) RTA prediction, specifically focusing on the non-visual approaches. It provides insights into recent advancements and introduces potential future research directions for the non-visual approach to better improve road safety. There are 95 studies shortlisted for non-visual ML-based RTA prediction, covering different modeling categories such as risk level, occurrence, frequency, and severity level. The findings revealed that government departments and open-access data portals are the trending sources for RTA datasets, while sources such as sensors and surveys can be used to collect additional RTA information. Moreover, conventional ML approaches are more prominent than statistical regression for RTA prediction due to their ability to handle RTA datasets with complex relationships. Meanwhile, deep learning and hybrid approaches are emerging techniques that are worth investigating in future research. Specifically, hybrid models can boost performance by combining multiple algorithms to focus on different processes such as feature extraction and model tuning. Eventually, several potential directions are highlighted for future research, such as multitask RTA prediction, data quality issues, ethical concerns for real-world implementation, model transferability, and the adoption of a standardized framework for model development. These recommendations are believed to be the catalyst that promotes further research advancement in the context of RTA prediction with non-visual data.