A Machine Learning Approach to Decode Mental States in Bistable Perception

This work demonstrates the usefulness of machine learning framework in decoding mental states from recorded brain signals. Magnetoencephalogram (MEG) signals were recorded from human participants while they were presented with six different conditions of bistable stimuli. Two internal mental states, transition and maintenance, which are related to switching or maintaining a perception in bistable perception respectively, were decoded. We extracted two types of features using complex Morlet wavelet transform that capture the spatio-temporal dynamics of large scale brain oscillations at global and local scale. Principal component analysis (PCA) was employed to reduce the dimension of the feature vector as well to minimize the redundancy among the features. Support vector machine (SVM) and artificial neural network (ANN) based classifiers were used to predict the mental states on a trial-by-trial basis. We were able to decode the two mental states from pooled data of all six conditions with accuracies of 79.52% and 79.56% using SVM and ANN classifier, respectively from local features which performed better than global features. The results show the effectiveness of signal processing and machine learning based approaches to identify internal mental states.