Time-locked amplitude modulations of brain signals recorded from the sensorimotor cortex have long been shown to index various processes related to movement planning, execution and imagination (motor imagery). Relative to a baseline period, signal power in the mu (8 – 13 Hz) and beta (13 – 30 Hz) frequency bands decreases prior to the onset of a real, attempted or imagined movement and conversely increases following its completion. These reproducible findings can be used in order to decode motor-related brain activity and are leveraged by several applications. Brain computer interface (BCI) technology refers to applications that allow a user to communicate with a computer through thought, by monitoring and inferring their mental state to issue an appropriate output. Motor imagery BCI (MI-BCI) constitutes a particular non-invasive BCI category that attempts to monitor and utilize signal changes due to the imagination of movements in order to produce commands, usually targeting the aforementioned signal modulations. MI-BCIs form the basis of non-invasive applications that attempt to decode motor intention in patients suffering from various deficits with the goal of restoring some level of communication with and/or control of their environment. However, they have so far been unable to live up to the high expectations and are consequently still far from widespread adoption out of research institutions. Multiple factors have been suggested to affect BCI reliability. In this work, we hypothesize that one of these factors is related to the signal features used. Based on recent literature that highlights the importance of characterizing transient activity dynamics in Chapter 2 - Motivation we suggest that MI-BCI should exploit features other than signal power. Moving beyond many assumptions regarding brain activity and therefore the algorithms used to analyze it, we propose that an analysis of single-trial signal properties that do not conform to the idea of sustained, sinusoidal-like oscillations can provide novel insights in motor-related brain processes and concurrently improve BCI performance. Notably, we argue in favor of analyzing beta band burst activity and mu band non-sinusoidal οscillatory characteristics. Focusing on the burst-like nature of the beta band activity in Chapter 3 – Beta bursts feasibility for BCI we show that like executed movements motor imagery too is characterized by beta bursts. We explore multiple ways to describe this burst activity showing that the burst rate of bursts with certain waveforms comprises the best representation among all in terms of classification accuracy and, importantly, that it is is more informative than signal power in a MI binary classification task. Then, we propose a computationally efficient algorithm to compute waveform-resolved burst rates, suitable for real-time BCI applications. Assuming the existence of calibration data we find waveforms of beta bursts whose rate is expected to be modulated the most by a MI task. We use these waveforms to filter raw brain activity in time domain. The resulting signals can then be exploited by standard signal-processing and classification paradigms used in the field. Compared to signal power we show that these novel features increase classification accuracy and information transfer rate, thus further reinforcing their potential role for MI-BCI. Finally, we address the question of waveform stability across recording sessions. We adopt a similar approach but focus on the impact of transferring waveforms learned in specific recording sessions to the decoding results of other sessions. We show that in contrast to power features classification accuracy is minimally impacted when using beta bursts in transfer learning. Taken together, these results prove the advantages of analyzing transient activity features and establish the feasibility of beta bursts for offline as well as online MI-BCI applications.