Brain–computer interfaces (BCI) have recently gained interest both in basic neuroscience and clinical interventions. The majority of noninvasive BCIs measure brain activity with electroencephalography (EEG). However, the real-time signal analysis and decoding of brain activity suffer from low signal-to-noise ratio and poor spatial resolution of EEG. These limitations could be overcome by using magnetoencephalography (MEG) as an alternative measurement modality. The aim of this thesis is to develop an MEG-based BCI for decoding hand motor imagery, which could eventually serve as a therapeutic method for patients recovering from e.g. cerebral stroke. Here, machine learning methods for decoding motor imagery -related brain activity are validated with healthy subjects’ MEG measurements. The first part of the thesis (Study I) involves a comparison of feature extraction methods for classifying left- vs right-hand motor imagery (MI), and MI vs rest. It was found that spatial filtering and further extraction of bandpower features yield better classification accuracy than time–frequency features extracted from parietal gradiometers. Furthermore, prior spatial filtering improved the discrimination capability of time–frequency features. The training data for a BCI is typically collected in the beginning of each measurement session. However, as this can be time-consuming and exhausting for the subject, the training data from other subjects’ measurements could be used as well. In the second part of the thesis (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 EEG. In Study II, we also compared the decoding results of simultaneously measured EEG and MEG data, and investigated whether the MEG responses to passive hand movements could be used to train a classifier to detect MI. MEG yielded altogether slightly, but not significantly, better results than EEG. Training the classifiers with subject’s own or other subjects’ passive movements did not result in high accuracy, which indicates that passive movements should not be used for calibrating an MI-BCI. The methods presented in this thesis are suitable for a real-time MEG-based BCI. The decoding results can be used as a benchmark when developing other classifiers specifically for motor imagery -related MEG data.
|Tila||Julkaistu - 2019|