Traffic collisions represent a significant public health issue, resulting in extensive human suffering and considerable economic and social burdens. These occurrences not only generate substantial medical and rehabilitation costs but also lead to production declines, property damage, and elevated insurance premiums. Furthermore, the enduring consequences on victims, families, and communities exacerbate their effects. Precise study of traffic accidents and prediction of their severity are crucial for alleviating these expenses. This study underscores the significance of data preprocessing, particularly normalisation, to improve dataset quality. Key variables were identified by clustering, chi-square tests, Cramer’s V, and predictor importance, subsequently organised into effective groups for efficient analysis. Conventional ELM sometimes use L2 regularisation to mitigate overfitting; yet, the 2-RELM model's dependence on manually chosen regularisation parameters is ineffective. To resolve this, it present ARegELM, an adaptive model that substitutes the static regularisation factor with a dynamic function, facilitating automatic selection. The ARegELM model demonstrated an impressive accuracy of 99.28% in forecasting accident severity. These results indicate that ARegELM improved predictive performance and model usability, hence facilitating better informed decision-making in traffic safety and accident prevention.