MEG Neural Decoding Pipeline: The Issues Residing Within The Data and Methods to Improve Your Decoding Accuracy

This study suggests a new analysis pipeline of MEG data, uniquely designed for neural decoding of small-sized datasets. It combines classic methods that assume stationarity of the data together with non-stationary methods to compensate for the distortions created by the classic approach. Popular Fourier-based methods are applied in a classic fashion, followed by additional filters using empirical mode decomposition and principal component analysis to further clean the data. An automated approach for epoch rejection is proposed as well. In this work, we propose a novel approach for data augmentation. Unlike most other solutions, combinations' averaging technique can be used on real data rather than synthetic one, making it more reliable from the neuroscientific point of view. It is also shown that this approach does not create any unnatural patterns within the augmented data. The proposed approach allows for application of machine learning algorithms on small-sized datasets. This broadens the list of available analyses for datasets with limited number of recorded examples. An image naming task was used for in-subject neural decoding estimations. In this work, we propose and compare four different machine learning designs. It is shown that a careful selection of the used channels, reduction of the feature dimensions, and averaging of the recorded epochs may significantly increase the accuracy of neural decoding.