Motor imagery-based brain-computer interfaces (MI-BCIs) rely on interactions between humans and machines. Therefore, the (learning) characteristics of both components are key to understand and improve performances. Data-driven methods are often used to select/extract features with very little neurophysiological prior. Should such approach include prior knowledge and, if so, which one? This paper studies the relationship between BCI performances and characteristics of the subject-specific Most Discriminant Frequency Band (MDFB) selected by a popular heuristic algorithm. First, our results showed a correlation between the selected MDFB characteristics (mean and width) and performances. Then, to investigate a possible causality link, we compared, online, performances obtained with a constrained (enforcing characteristics associated to high performances) and an unconstrained algorithm. Although we could not conclude on causality, average performances using the constrained algorithm were the highest. Finally, to understand the relationship between MDFB characteristics and performances better, we used machine learning to 1) predict MI-BCI performances using MDFB characteristics and 2) select automatically the optimal algorithm (constrained or unconstrained) for each subject. Our results revealed that the constrained algorithm could improve performances for subjects with either clearly distinct or no distinct EEG patterns.