The big goal of brain-computer interfaces (BCI) is to open up a communication channel between machines and the brain directly through the human mind. This form of communication will ultimately enable us to not only control our body with thinking processes, but also external devices and machines such as robotic systems, prosthesis, cars and much more! As you may have noticed, applications for BCI systems may range from general assistive technologies for the public to medical devices that are able to completely replace limbs or re-enable communication with locked-in patients.
While BCI systems that are even able to control robotic arms already exist, they often use invasive techniques, e.g., implanting electrodes into the brain. Unfortunately, such invasive techniques come with a lot of risks due to surgical intervention, electrode rejection by the immune system, etc.
On the other hand, safe non-invasive techniques like Magneto- or Electroencephalography (MEG or EEG) that measures brain signals at the scalp suffer heavily from signal superposition and noise caused by the environment and the user. To identify and extract the actual neuronal signals originating from different areas in the brain has proven to be a very hard problem that is most successfully tackled with mathematical signal processing and machine learning techniques.
One major challenge in building a stable and easy to use BCI comes from the fact that the performance and ability to control a BCI varies heavily between subjects and even within a subject across sessions. This problem does not only require subjects to have a "talent" for BCI control at all, but often require further individual training sessions to be conducted before the BCI can actually be used.
The most prominent non-invasive BCI systems are based on EEG signals that measure electrical brain activity caused by neuronal populations in the brain. Knowing the so called Brodmann areas on the cerebral cortex, we can infer paradigms that are likely to cause neuronal activity patterns in specific areas of the brain. Such patterns can be identified with appropriate noise reduction methods and learning algorithms in order to decode a "thought" from the EEG data. A general BCI pipeline system will usually preprocess EEG signals to remove noise and focus on specific regions of interest or neuronal activity sources in the brain, before features are extracted and fed into a trained model for interpretation. One of the most often used paradigms is motor imagery (MI) of the limbs, e.g., the imagination of executing an action with a hand, elbow or foot. MI leads to similar activity patterns in the corresponding motor cortex as in actual movements of that limb, namely a decrease in band power within the μ- and β-rhythm that is exploited by learning algorithms.
A very general BCI pipeline for MI of the left hand, right hand and feet is depicted below: