Reducing calibration time in motor imagery based brain-computer interface : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in School of Fundamental Sciences at Massey University, Palmerston North, New Zealand

Motor imagery (MI) based Electroencephalogram (EEG) Brain-computer interface (BCI) detects neural activity generated due to kinesthetic imagination of limbs from brain scalp and translate it into control commands for external devices. MI-BCIs are indeed very promising for people suffering from neuromuscular disorder, but still lack adoption as access modalities outside laboratories. The main reason that prevents EEG based MI-BCIs from being widely used is there long calibration time. Due to considerable inter-subject/inter-session and intra-session variations, a large number of training trials are collected to calibrate systems at the beginning of each MI-BCI session. This time consuming calibration is required to achieve good performance with the BCI system but causes fatigue to user and leaves less time for online BCI interactions. This thesis focuses on developing reliable signal processing and classification pipeline that reduce MI-BCI calibration time while keeping accuracy in an acceptable range. In the first part of the study, we have provided an extensive review of current state of art in designing a EEG based MI-BCI system. In doing so, I have created an architectural framework which brings together interdisciplinary concepts under a unified umbrella. We used this framework to identify key signal processing, features extraction and learning algorithms and their limitation that must be taken into consideration while designing novel pipeline for reducing calibration in MI-BCI. This architecture is also useful to understand current issues in BCI and to visualize the gaps to be filled by future studies in order to further improve BCI usability. In the second part of the study, we address long calibration issue in MI-BCI under two scenarios. First, when there is only few training trials from new subject (user) is available and no training data from previous sessions or other users is available. Second, reducing (inter-subjects/sessions non-satationarity) calibration time of new subject when there is previous sessions or other subjects data is available along with few trials from new subject. In order to contribute to the progress of reducing calibration in MI-BCI, we proposed novel signal processing and classification pipeline that uses spatial, spectral, temporal and geometrical properties of subject’s trial from EEG signals and achieve acceptable performance under reduce calibration setting