How Brain-Controlled Devices Actually Work: Decoding EEG Signals

Once you have acquired the samples of the EEG signals for the different user commands, you will need a mechanism by means of which the device can distinguish between the signals corresponding to the different commands. For example, if you have a brain-controlled toy car that moves left or right as per the thoughts of the operator, you will need to collect samples of the brain signals of the operator when he thinks about moving the car towards the right as well as when he thinks about moving it towards the left. Once you have these samples, you will need to develop an algorithm that classifies them into the Right and Left categories accurately.

You will first need to identify what properties of the two classes of signals vary the most drastically and consistently from each other. Some of the common properties used for the classification of EEG signals include the Mean Absolute Value and the relative power of various frequency bands compared to each other and to the total power. Selection of the properties that help with the most accurate classification is usually done on the basis of experimental results obtained by using different combinations of these properties as an input for the classification algorithm during the training stage.

In the case of binary classification, where the inputs need to be mapped to one of two valid outcomes (as is the case in the toy car problem mentioned above), two popular classification methods are Neural Networks and Support Vector Machines (SVM). Software packages for both of these are available online, and you can utilise the same to feed different inputs or properties to the neural network/SVM, to identify the properties with which the software can generate the most accurate results. Since this is the training data and you are aware of the desired output for each input, you can automate the entire process in software to come up with a robust classification algorithm.

The situation becomes a little trickier when the desired outputs are more than two in number, say you can move the car in four different directions, right, left, forward, and backward. This leads to a multi-class classification problem and you may need a more sophisticated algorithm to solve the same. The training data will now correspond to the brain waves captured for the four different commands, the desired output for which can be one of the four directions. You will need to recognise the properties for classification as in the binary problem described above, but since the output classes are more than two in number, you will need a classification model that can handle more than two outputs. Such multi-class classification problems can be solved using Multi-Class SVM and Random Forests, among other classification techniques.

Once fully designed, the output of this classification algorithm is what will drive the brain-controlled device or application. Based on the complexity of the problem statement, these devices may require one-time training that can cater to all users or may need a round of training with each new user who wants to use the device. The more robust the training algorithm, the lesser is the time/data required for training and the higher is the accuracy of the brain-controlled device.

In the next section, we will discuss the current EEG-based technologies available in the market and the scope of further research in this domain.