Current features used in motor imagery-based Brain-Computer Interfaces (BCI) rely on local measurements that miss the interactions among brain areas. Such interactions can manifest as bursts of activations, called neuronal avalanches. To track their spreading, we used the avalanche transition matrix (ATM), which contains the probability that an avalanche would consecutively recruit any two brain regions. Here, we proposed to use ATMs as a potential alternative feature. We compared the classification performance resulting from ATMs to a benchmark model based on Common Spatial Patterns. In both sensor-and source-spaces, our pipeline yielded an improvement of the classification performance associated with reduced inter-subject variability. A correspondence between the selected features with the elements of the ATMs that showed a significant condition effect led to higher classification performance, which speaks to the interpretability of our findings. In conclusion, working in the sensor space provides enough spatial resolution to classify. However the source space is crucial to precisely assess the involvement of individual regions.