Brain computer interfaces: an engineering view. Design, implementation and test of a SSVEP-based BCI.

Abstract

This thesis presents the realization of a compact, yet flexible BCI platform, which, when compared to most commercially-available solution, can offer an optimal trade-off between the following requirements: (i) minimal, easy experimental setup; (ii) flexibility, allowing simultaneous studies on other bio-potentials; (iii) cost effectiveness (e.g. < 1000 €); (iv) robust design, suitable for operation outside lab environments. The thesis encompasses all the project phases, from hardware design and realization, up to software and signal processing. The work started from the development of the hardware acquisition unit. It resulted in a compact, battery-operated module, whose medium-to-large scale production costs are in the range of 300 €. The module features 16 input channels and can be used to acquire different bio-potentials, including EEG, EMG, ECG. Module performance is very good (RTI noise < 1.3 uVpp), and was favourably compared against a commercial device (g.tec USBamp). The device was integrated into an ad-hoc developed Matlab-based platform, which handles the hardware control, as well as the data streaming, logging and processing. Via a specifically developed plug-in, incoming data can also be streamed to a TOBI-interface compatible system. As a demonstrator, the BCI was developed for AAL (Ambient Assisted Living) system-control purposes, having in mind the following requirements: (i) online, self-paced BCI operation (i.e., the BCI monitors the EEG in real-time and must discern between intentional control periods, and non-intentional, rest ones, interpreting the user’s intent only in the first case); (ii) calibration-free approach (“ready-to-use”, “Plug&Play”); (iii) subject-independence (general approach). The choice of the BCI operating paradigm fell on Steady State visual Evoked Potential (SSVEP). Two offline SSVEP classification algorithms were proposed and compared against reference literature, highlighting good performance, especially in terms of lower computational complexity. A method for improving classification accuracy was presented, suitable for use in online, self-paced scenarios (since it can be used to discriminate between intentional control periods and non-intentional ones). Results show a very good performance, in particular in terms of false positives immunity (0.26 min^-1), significantly improving over the state of the art. The whole BCI setup was tested both in lab condition, as well as in relatively harsher ones (in terms of environmental noise and non-idealities), such as in the context of the Handimatica 2014 exhibition. In both cases, a demonstrator allowing control of home appliances through BCI was developed.