Design of a BMI for tetraplegic patients

Recent advances in electronics and signal processing opened the way to better interactions between humans and machines. Today, brain-machine interfaces (BMIs) offer the possibility to replace a manual control simply by thinking. While it can expand human capabilities, it makes perfect sense to use this kind of technology to help people with physical disorders as paralysis. When the brain has been disconnected from some parts of the body, BMIs can short-circuit this path and help to restore lost functions. The interest in this field is growing but big challenges are still to be faced in order to make a reliable product out of it. When designing a BMI, several aspects related to the patient have to be taken into account. The device has to be adapted to the patient’s health and psychological state as well to his motivation, abilities, etc. However, the release of a reliable product will only be possible when a good trade-off between a universal solution and an adaptable one will be achieved. In the branch of motor imagery based on electroencephalogram (EEG) signals, the recognition between two tasks, typically movements of right or left hand, has been widely studied and shows better and better results. However, recognition of three or more tasks is still limited in terms of accuracy. Moreover, most of the studies operate on offline systems with an unlimited amount of resources. A key element in the development of BMIs is to make them portable in order to allow a day-to-day use of the system. This work aims to study the different techniques used in motor imagery classification, from the acquisition of brain activity to the classification, and compare them regarding their accuracy. On this basis, the best performing techniques are implemented on a FPGA-based experimentation system. This BMI could offer some comfort to control a simple equipment, turn on a device, direct a wheelchair, etc. The current system records sensorimotor rhythms and more especially, mu and beta bands with an EEG equipment. Three electrodes are placed over the motor cortex (C3, Cz and C4 positions). After filtering, the influence of pre-processing with independent component analysis is investigated. Then, features need to be generated from the filtered signals. Different methods are used as power bands, wavelet transform and autoregressive coefficients. The classifiers used in this work are the support vector machine and the linear discriminant analysis. A classification accuracy up to 70 % is obtained offline while the online implemented model provides an average accuracy above 62 %. This work paves the way for further studies that could be undertaken starting from the suggestion made at the end of the master thesis.