Predicting higher-risk crash factors and patterns using machine learning-association rule mining

Road traffic crashes remain a significant global public health concern, particularly in low- and middle-income countries, where fatalities and injuries disproportionately affect vulnerable road users. Despite efforts to improve road safety, crash severity levels are conventionally classified based on observable injury and fatality outcomes. This outcome-based approach oversimplifies the complexity of crash risk by ignoring latent hazards embedded in severity-based categories for non-fatal or minor crashes. This study addresses this gap by integrating Machine Learning (ML) and Association Rule Mining (ARM) to predict high-risk crashes and identify crash patterns respectively. Using nine years of historical crash data (2013−2021) from Ghana, the study employed Logistic Regression, Random Forest (RF), and XGBoost for high-risk crash prediction, followed by ARM to identify hidden patterns. RF outperformed the other models, achieving 71.4 % accuracy and a 90.3 % ROC AUC. The identified high-risk crashes depicted a mismatch between crashes classified as high severity and crashes identified as high-risk crashes. Crashes termed low severity dominated the set of crashes classified by the RF model as high-risk crashes. ARM revealed significant hidden patterns, such as rear-end collisions at signalized intersections and road width and non-impaired driving co-occur crashes under optimal environmental and behavioural driving conditions. The study demonstrates the value of combining ML and ARM for actionable insights. The findings emphasize that infrastructural design and driver behaviour both play important roles in high-risk crash outcomes, suggesting a need for holistic road safety strategies, including infrastructure redesign, enhanced traffic control measures, and public awareness campaigns to mitigate complacency in ideal driving conditions. Policymakers and traffic engineers are urged to adopt context-sensitive designs and prioritize non-junction segments, where road width significantly impacts crash risk. • Machine Learning and Association Rule Mining improve the prediction of high-risk crash factors. • Random Forest achieved the highest accuracy (71.4 %) and ROC AUC (90.3 %) in crash prediction. • Driver or rider age, road width, and collision type were identified as determinants of higher-risk crashes. • Rear-end collisions at signalized intersections showed a 70.3 % confidence level in hidden pattern analysis. • Infrastructural design and driver behaviour contribute to road width and non-impaired–related crashes.