Abstract: Road safety modeling is a valuable strategy for promoting safe mobility, enabling the development of crash prediction models (CPM) and the investigation of factors contributing to crash occurrence. This modeling has traditionally used statistical techniques despite acknowledging the limitations of this kind of approach (specific assumptions and prior definition of the link functions), which provides an opportunity to explore alternatives such as the use of machine learning (ML) techniques. This study reviews papers that used ML techniques for the development of CPM. A systematic literature review protocol was conducted, that resulted in the analysis of papers and their systematization. Three types of models were identified: crash frequency, crash classification by severity, and crash frequency and severity. The first is a regression problem, the second, a classificatory one and the third can be approached either as a combination of the preceding two or as a regression model for the expected number of crashes by severity levels. The main groups of techniques used for these purposes are nearest neighbor classification, decision trees, evolutionary algorithms, support-vector machine, and artificial neural networks. The last one is used in many kinds of approaches given the ability to deal with both regression and classification problems, and also multivariate response models. This paper also presents the main performance metrics used to evaluate the models and compares the results, showing the clear superiority of the ML-based models over the statistical ones. In addition, it identifies the main explanatory variables used in the models, which shows the predominance of road-environmental aspects as the most important factors contributing to crash occurrence. The review fulfilled its objective, identifying the various approaches and the main research characteristics, limitations, and opportunities, and also highlighting the potential of the usage of ML in crash analyses.