Learning temporal-frequency features of physionet EEG signals using deep convolutional neural network

Abstract

Since EEG signals encode an individual’s intent of executing an action, scientists have extensively focused on this topic. Motor Imagery (MI) signals have been used by researchers to assistance disabled persons, for autonomous driving and even control devices such as wheelchairs. Therefore, accurate decoding of these signals is essential to develop a Brain–Computer interface (BCI) systems. Due to dynamic nature, low signal-to-noise ratio and complexity of EEG signals, EEG decoding is not simple task. Extracting temporal and spatial features from EEG is accessible via Convolution neural network (CNN). However, enhanced CNN models are required to learn the dynamic correlations existing in MI signals. It is found that good features are extracted via CNN in both deep and shallow models, which indicate that various levels related features can be mined. In this case, spatial patterns from multi-scaled data in different frequency bands are learnt at first and then the temporal and frequency band information from projected signals is extracted. Here, to make use of neural activity phenomena, the feature extraction process employed is based on Multi-scale FBCSP (MSFBCSP). In CNN, the envelope of each spatially filtered signal is extracted in time dimension by performing Hilbert transform. However, to access common morphologies, the convolutional operation across time is performed first and then another convolution layer across channels in the frequency band is used to represent the carried information in a more compact form. Moreover, Bayesian approach is used for mapping hyperparameters to a probability of score on the objective function. The prominent feature of the proposed network is the high capacity of preserving and utilizing the information encoded in frequency bands. Our proposed method significantly improves the efficiency of current classification method in specific dataset of the physionet. According to empirical evaluations, strong robustness and high decoding classification are two distinctive characteristics of our proposed work.

Hybrid fuzzy deep neural network toward temporal-spatial-frequency features learning of motor imagery signals

“…Values in parenthesis are kappa value and the nonparanthesis are accuracy values. Work Methods Activation function Hyperparameters tuning method ACC (Kappa) Sakhavi et al (2018) 9 FBCSP + SVM ReLU Coordinate-descent 71.18(0.616) Sakhavi et al (2018) 9 FBCSP + CNN ReLU Coordinate-descent 74.46 (0.659) Sorkhi et al (2022) 20 MSFBCSP + TFCNN ReLU Bayesian Optimization 75.51 (0.673) Schirrmeister et al (2017) 1 Shallow CNN Deep CNN ELU Sampling strategies 73.7 (0.65) 70.9 (0.61) Riyad et al, (2020) 15 Inception + EEG-net ELU Cross Validation 74.7 (0.66) Lu et. al (2019) 36 1LAYER-CNN + LSTM ReLU Cross Validation 76.62 (0.68) Zhang et.al (2019) 18 Over-FBCSP + HDNN ReLU Constant based on prestudeies 85(0.80) Zhang et.al (2021) 19 Over-FBCSP + HDNN-TL ReLU Fine-tune with 0.001 learning rate 85(0.81) Wang et.…”

“…Each row demonstrate level of significance for a model against different models. Metric FBCSP + CNN 9 MSFBCSP + TFCNN 20 Over-FBCSP + HDNN 18 SCCRNN 24 EEG-CLFCNet(BO) FBCSP + CNN 9 Non 0.0117 0.112 0.015 0.0312 MSFBCSP + TFCNN 20 0. 0117 Non 0.008 0.021 0.0117 Over-FBCSP + HDNN 18 0.112 0.…”

“…Regardless of the success of deep learning, extractions of spatial, temporal and frequency features are rarely done simultaneously and the fuzzy approach is not applicable in these hybrid methods. Lately, Sorkhi et al applied MSFBCSP, CNN, and tune hyperparameters with Bayesian optimization in their work 20 .…”

“…Furthermore, dissimilar extraction orders might cause diverse classification outcomes. In this case, a hybrid neural network containing LSTM and CNN networks has been exploited in some studies to simultaneous excerpt temporal and spatial features [18][19][20][21][22][23][24] .…”

TBEEG: A Two-Branch Manifold Domain Enhanced Transformer Algorithm for Learning EEG Decoding

“…A graph convolution neural network (M-GCN) proposed by [39], involves temporal-frequency processing performed through modified S-transform (MST). M Sorkhi et al [40] presented a Multi-scale FBCSP (MSFBCSP) method and spatial patterns from multi-scaled data in different frequency bands that are learnt, and then, the temporal and frequency band information from projected signals is extracted . [41] proposed a novel model called Hybrid-Scale Spatial-Temporal Dilated Convolution Network (HS-STDCN) for EEG-based imagined speech recognition.…”