Brain Decoding over the MEG Signals Using Riemannian Approach and Machine Learning

Abstract

Brain decoding is an emerging approach for understanding the face perception mechanism in the human brain. Face visual stimuli and perception mechanism are considered as a challenging ongoing research of the neuroscience field. In this study, face/scrambled face visual stimulations were implemented over the sixteen participants to be decoded the face or scrambled face classification using machine learning (ML) algorithms via magnetoencephalography (MEG) signals. This noninvasive and high spatial/temporal resolution signal is a neurophysiological technique which measures the magnetic fields generated by the neuronal activity of the brain. The Riemannian approach was used as a highly promising feature extraction technique. Then Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN) were employed as deep learning algorithms, Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA) were implemented as shallow algorithms. The improved classification performances are very encouraging, especially for deep learning algorithms. The LSTM and GRU have achieved 92.99% and 91.66% accuracy and 0.977 and 0.973 of the area under the curve (AUC) scores, respectively. Moreover, CNN has yielded 90.62% accuracy. As our best knowledge, the improved outcomes and the usage of the deep learning on the MEG dataset signals from 16 participants are critical to expand the literature of brain decoding after visual stimuli. And this study is the first attempt with these methods in systematic comparison. Moreover, MEG-based Brain-Computer Interface (BCI) approaches may also be implemented for Internet of Things (IoT) applications, including biometric authentication, thanks to the specific stimuli of individual’s brainwaves.