Data analysis methods for neuroimaging data pre-processing to decode cognitive tasks using logistic regression for BCI applications

Brain-Computer Interfaces (BCI) permit neural activity to be directly interpreted and used for applications, like therapeutic replacement of lost function (e.g. stroke) or to supplement existing function (e.g. handsfree applications). Two major challenges for BCI are accurate interpretation of neural activity and signal processing speed for real-time applications i.e. correctly decode a user’s intent and the timely execution of that intent. Magnetoencephalography (MEG) has advantages over Electroencephalography (EEG) with respect to spatial and temporal resolution which could potentially allow better decoding of brain activity. High spatial and temporal resolution using MEG generates a large volume of data which must be rapidly preprocessed and classified correctly for practical realtime BCI. This paper presents a simple data processing technique to clean, normalise and reduce data dimensionality, for optimal class label decoding using a simple Logistic Regression classifier. Good decoding performance was achieved using an off-line MEG dataset, with or without data dimensionality reduction, comparable to more complex data pre-processing methods and classifiers already studied.