Decoding of semantic categories of imagined concepts of animals and tools in fNIRS.

Objective.Semantic decoding refers to the identification of semantic concepts from recordings of an individual's brain activity. It has been previously reported in functional magnetic resonance imaging and electroencephalography. We investigate whether semantic decoding is possible with functional near-infrared spectroscopy (fNIRS). Specifically, we attempt to differentiate between the semantic categories of animals and tools. We also identify suitable mental tasks for potential brain-computer interface (BCI) applications.Approach.We explore the feasibility of a silent naming task, for the first time in fNIRS, and propose three novel intuitive mental tasks based on imagining concepts using three sensory modalities: visual, auditory, and tactile. Participants are asked to visualize an object in their minds, imagine the sounds made by the object, and imagine the feeling of touching the object. A general linear model is used to extract hemodynamic responses that are then classified via logistic regression in a univariate and multivariate manner.Main results.We successfully classify all tasks with mean accuracies of 76.2% for the silent naming task, 80.9% for the visual imagery task, 72.8% for the auditory imagery task, and 70.4% for the tactile imagery task. Furthermore, we show that consistent neural representations of semantic categories exist by applying classifiers across tasks.Significance.These findings show that semantic decoding is possible in fNIRS. The study is the first step toward the use of semantic decoding for intuitive BCI applications for communication.

Potential pitfalls of widely used implementations of common spatial patterns.

We have uncovered serious flaws in handling EEG signals with a decreased rank in implementations of the common spatial patterns (CSP). The CSP algorithm assumes covariance matrices of the signal to have full rank. However, preprocessing techniques, such as artifact removal using independent component analysis, may decrease the rank of the signal, leading to potential errors in the CSP decomposition. We inspect what could go wrong when CSP implementations do not take this into consideration on a binary motor imagery classification task. We review CSP implementations in open-source toolboxes for EEG signal analysis (FieldTrip, BBCI Toolbox, BioSig, EEGLAB, BCILAB, and MNE). We show that unprotected implementations decreased mean classification accuracy by up to 32%, with spatial filters resulting in complex numbers, for which corresponding spatial patterns do not have a clear interpretation. We encourage researchers to check their implementations and analysis pipelines.

Neural component analysis: source localisation for motor imagery classification.

The electroencephalogram (EEG) records a summed mixture of multiple sources of neural activity distributed throughout the brain. Source separation methods aim to un-mix the EEG in order to recover activity generated by the original sources. However, most current state-of-the-art source separation methods do not take into account the physical locations of sources of EEG activity.We present a new source separation method which uses an accurate model of the head to un-mix the EEG into individual sources based on their physical locations.We apply our method to an EEG dataset recorded during motor imagery and show that it is able to identify sources that are located in distinct physical regions of the brain. We compare our method to independent component analysis and show that our sources have higher spatial specificity and, furthermore, allow higher classification accuracies (a mean improvement in accuracy of 8.6% was achieved p =0.039).