A comparison of accident causation models (ACMS) and machine learning (ML) for applied analysis within accident reports

Machine learning (ML)-supported accident prediction models appear as an alternative to the much older accident causation models (ACMs). ACMs represent a simplification of accident processes and resulted loss and play an important role in accident investigations and identifying potential risk factors. This effort investigates ACMs and ML results of accident reports analysis in relation to each other and aims at comparing the latter based on their level of causes, the relationship between causes, and the predictability of severity. A framework of understanding of these main processes and their challenges is provided, which is also used as a methodological framework for the comparison. The comparison is based on a desk study of literature and material on the two types of models. ACMs are different in typology, levels of causes, and the logic through which the analysis of the events that have taken place is conducted. Many ML prediction models in construction not only provide predictions but also result into structures of features which work as predictors, e.g., decision trees. ACMs and ML are different in the task they perform. ML models in the literature are focused on predicting the severity of an event while missing the identification of prevention measures. ACMs focus on the occurrence of unwanted events and lack the ranking of important features. Finally, ML analysis of accident reports need ACMs as a theory to shift focus to risks instead of severity, while interpretable ML algorithms (e.g., RF) appear more capable of complex representations of contributing factors. An unsolved issue is the random element involved in most accident processes.