Methods for brain–computer interfaces utilizing MEG and motor imagery

Brain–computer interfaces (BCI) are systems that translate the user's brain activity into commands for external devices in real time. Magnetoencephalography (MEG) measures electromagnetic brain activity noninvasively and can be used in BCIs. The aim of this thesis was to develop an MEG-based BCI for decoding hand motor imagery. The BCI could eventually serve as a therapeutic method for patients recovering from e.g. cerebral stroke. Here, we validated machine-learning methods for decoding motor imagery (MI)-related brain activity with healthy subjects' MEG measurements. In addition, we studied the effect of different BCI feedback modalities on the subjects' brain function related to MI. In Study I, we compared feature extraction methods for classifying left- vs right-hand MI, and MI vs rest. We found that spatial filtering and further extraction of bandpower features yielded better classification accuracy than time–frequency features extracted from MEG channels above the parietal area. Furthermore, prior spatial filtering improved the discrimination capability of time–frequency features. The training data for a BCI are typically collected in the beginning of each measurement session. However, as this can be time-consuming and exhausting for patients, data from other subjects' measurements could be used for training as well. In Study II, methods for across-subject classification of MI were compared. The results showed that a classifier based on multi-task learning with a l2,1-norm regularized logistic regression was the best method for across-subject decoding for both MEG and electroencephalography (EEG). In Study II, we also compared the decoding results of simultaneously measured EEG and MEG data, and investigated whether MEG responses to passive hand movements could be used to train a classifier to detect MI. MEG yielded slightly better results than EEG. Training the classifiers with the subject's own or other subjects' passive movements did not result in high accuracy. Passive movements should thus not be used for calibrating an MI-BCI. In Study III, we investigated how the amplitude of sensorimotor rhythms (SMR) changes while the subjects practise hand MI with a BCI. We compared the effect of visual and proprioceptive feedback on brain functional changes during a single measurement session. In subjects receiving proprioceptive feedback, the power of SMR increased linearly over the session in motor cortical regions, while similar effect was not observed in subjects receiving purely visual feedback. According to these results, proprioceptive feedback should be preferred over visual feedback especially in BCIs aiming at recovery of hand functions. The methods presented in this thesis are suitable for an MEG-based BCI. The decoding results can be used as a benchmark when developing classifiers specifically for MI-related MEG data.