Currently the amount of available and variety of data constitute a challenging purpose of decision makers, especially in education field. This article aims to predict a specific kind of students that reveal many concerns in all countries. These students are qualified as low educational skills attainment and named at-risk students in some literature. This objective is pursued given the countries classification by regional and income belonging. As methodology, we have used six machine learning algorithms, six models' performance evaluation metrics, three labels and seven features. The relevance of these features is assessed via feature importance analysis technics in intention to exploit them in models' construction and training. Results shown, first that some models perform very well than other given countries classification either by region or by income according to three performance metrics. Second, models fitted with African countries data are more performant and accurate than the other models given all performance metrics. Third, some features are important than other in the model training phase. Finally, we conclude that using other machine learning parameters and computing powers can lead to performance enhancement of the adopted models.