Motor Imagery-based Brain-Computer Interfaces (MI-BCI) allow users to control a computer for various applications using their brain activity alone, which is usually recorded by an electroencephalogram (EEG). Although BCI applications are numerous, their use outside laboratories is still scarce due to their poor accuracy. Some users cannot use BCIs, a phenomenon sometimes called "BCI illiteracy", which impacts around 10% to 30% of BCI users, who cannot produce discriminable EEG patterns. By performing neurophysiological analyses, and notably by identifying neurophysiological predictors of BCI performance, we may understand this phenomenon and its causes better. In turn, this may also help us to better understand and thus possibly improve , BCI user training. Therefore, this paper presents statistical models dedicated to the prediction of MI-BCI user performance, based on neurophysiological users' features extracted from a two minute EEG recording of a "relax with eyes open" condition. We consider data from 56 subjects that were recorded in a 'relax with eyes open' condition before performing a MI-BCI experiment. We used machine learning regression algorithm with leave-one-subject-out cross-validation to build our model of prediction. We also computed different correlations between those features (neurophysiological predictors) and users' MI-BCI performances. Our results suggest such models could predict user performances significantly better than chance (p ≤ 0.01) but with a relatively high mean absolute error of 12.43%. We also found significant correlations between a few of our features and the performance, including the previously explored µ µ µ-band predictor, as well as a new one proposed here: the µ µ µ-peak location variability. These results are thus encouraging to better understand and predict BCI illiteracy. However, they also require further improvements in order to obtain more reliable predictions.