Developing tools to support students and learning in a traditional or online setting is a significant task in today's educational environment. The initial steps toward enabling such technologies using machine learning techniques focused on predicting the student's performance in terms of the achieved grades. However, these approaches do not perform as well in predicting poor-performing students. The objective of our work is twofold. First, in order to overcome this limitation, we explore if poorly performing students can be more accurately predicted by formulating the problem as binary classification, based on data provided before the start of the semester. Second, in order to gain insights as to which are the factors that can lead to poor performance, we engineered a number of human-interpretable features that quantify these factors. These features were derived from the students' grades from the University of Minnesota, an undergraduate public institution. Based on these features, we perform a study to identify different student groups of interest, while at the same time, identify their importance. As the resulting models provide us with different subsets of correct predictions, their combination can boost the overall performance.