Enhancing Road Safety Evaluation with AI: Causal Discovery and Reasoning in Road Traffic Accident Analysis

Road traffic accidents result from a complex interplay of various conditions, making it difficult to pinpoint key factors that influence their occurrence and severity. This study employs Bayesian Belief Networks (BBNs) to construct robust predictive models capable of handling incomplete data via probabilistic inference, providing transparent and interpretable insights. Unlike traditional machine learning methods, which often struggle with missing values, BBNs efficiently estimate probabilities from available evidence, offering a cohesive framework for assessing road safety risks. Using the UK government's 2022 traffic accident dataset, we developed a BBNbased accident probability model, identifying significant variables such as road type, speed limits, junction configurations, lighting, and weather conditions through expert consultation and literature reviews. Each variable was categorized, quantified by frequency, converted into probabilities, and subsequently integrated multiplicatively to approximate accident likelihood. These probabilities were discretized into ordinal classes, serving as input states for the BBN. The model was validated using the 2021 accident dataset, demonstrating its predictive reliability across temporal variations. Further, we extended our model to accident severity classification. The combined probability-severity BBN approach offers actionable insights for targeted road safety interventions, such as improved lighting or drainage. This methodology exemplifies a practical, interpretable, and scalable solution for guiding policymakers in optimizing road safety interventions.