Table A1.

Applying CNN to Feature Extraction from EEG Signals.

№	Tasks		Dataset	CNN	Learning Type	Steps	Structure	Optimization	Activation Function	Function Loss	Evaluation Metrics	Framework	Ref.
1	Classification task	Sleep stage annotations	Physionet Sleep-EDF dataset	SleepEEGNet	Supervised	Decomposition of data into frequency components and subsequent classification	2DCNN and BiRNN	RMSProp optimizer	ReLU	lMFE	k-fold cross validation. overall accuracy, precision, recall (sensitivity), specificity, and F1-score.	Python 3.7–3.10, TensorFlow 2.8	[256]
2		Emotion recognition	DEAP dataset 39	EEG-Based Emotion Recognition Using a 2D CNN	Supervised	Decomposition of data into frequency components and subsequent classification	2D CNN	Particle Swarm Optimization	LeakyReLU Outpit—Softmax	Cross Entropy	85%	Python 3.7–3.10	[214]
3		Motor Imagery Signals Classification	BCI Competition IV 2a (BCI-IV2a), High Gamma (HGD)	MBEEGSE	Supervised	MBEEGSE architecture. Divided into three branches, each with EEGNet and SE block	EEGNet and Squeeze-and-Excitation (SE) Block	Adam optimizer	Softmax	Cross Entropy	70%	Keras 3.0.4, Python 3.6, 3.7, 3.8, 3.9	[202]
4		Motor Imagery EEG Decoding	BCI Competition 2008 IV 2a Dataset High Gamma Dataset: (HGD)	TS-SEFFNet	Supervised	First, the deep temporal convolution block (DT-Conv block). Second, multispectral convolution block (MS-Conv block) is then run in parallel using multilevel wavelet convolutions. Finally, block (SE-Feature-Fusion block) displays the depth-time and multispectral features into complex pooled feature maps that extract the feature responses across channels.	DT-Conv block, MS-Conv block, SE-Feature-Fusion block	The Optimization Steps of the Proposed TS- SEFFNet Method	Softmax	Custom loss function	93.25%	Torch 1.4, Python 3.8	[217]
5		Sleep stage annotations	Physionet Challenge dataset	Self-supervised learning (SSL)	Unsupervised	The first step is a sampling process by which examples are extracted from the time series S (EEG recording). The following describes a learning process where sample examples are used to train the feature extractor end-to-end.	Relative positioning (RP). Temporal shuffing (TS), Contrastive predective coding (CPC)	Adam optimizer	Rectified Linear Unit (ReLU)	Cross-entropy loss function	72.3%	Torch 1.4, Python 3.9	[218]
6	Prediction task	EEG Imaginary Speech Recognition	Kara One database	-	Supervised	a CNN containing two convolutional layers with 64 and 128 filters connected to a dense layer with 64 neurons for input signal spectrum of a 0.25 s window	2D CNN	Adam optimizer	Linear	Categorical cross-entropy	37%	-	[257]
7		EEG-Speech Recognition	Custom dataset (not available)	-	Supervised	ResNet18/50/101 with 2 layers of managed recurrent units—Gated Recurrent Unit (GRU). And after that ResNet18 operation are fed to the input of a recurrent neural network containing 1024 hidden GRUs.	CNN and RNN	Adam optimizer	Softmax	-	85%	-	[258]
8		EEG speech recognition	Custom dataset (not available)	-	Supervised	The architecture used includes the already trained VGG Net CNN design and the target CNN design, while the already trained VGG Net CNN design extracts global features for general image classification work, and the target CNN design aims at efficient and accurate categorization of EEG signals using already trained Model VGG Net CNN.	Deep Residual–encoder–based VGG net CNN	-	Softmax	Softmax cross-entropy	95%	-	[259]
9		Seizure prediction	CHB-MIT and Kaggle	-	Supervised	a hybrid network that can combine the additional benefits of CNN and Transformer. The CNN is used to extract local information that contains two 3 × 3 convolutions with stride 1 and another 3 × 3 convolution with stride 2 to reduce the size of the input features. Each convolutional layer is followed by a GELU activation and a batch normalization (BN) layer. The model has two stages for extracting multiscale features from the EEG spectrum. Each stage consists of a set of Transformer blocks applied to extract long-term dependencies.	CNN and transformer	Adam	Softmax	Cross-entropy	95%	Torch 1.4, Python 3.8	[260]
10		Predicting Human Intention-Behavior	BCI competition IV Dataset 2b	-	Supervised	The multi-scale CNN model has seven layers, which are one input layer, two convolutional layers, one max-pooling layer, one multi-scale layer, one full connection layer and one softmax output layer. The input layer in the multi-scale CNN model is fed with a time-frequency image with the size of 40 × 32 × 3 after EEG signals are preprocessed by STFT	Multi-Scale CNN Model	-	Linear	Cross-entropy	73.9%	Python 3.8, Keras 3	[261]
11	Artifact Removal	EEG Artifact Detection and Correction	Costum dataset, not available	-	Unsupervised	modification of a feed-forward neural network that uses weight sharing and exhibits translation invariance. Learning in the CNNs operates on the same principle as a traditional feed-forward neural network where an error from output layer is back-propagated through the network and weights of the network are proportionally updated to the gradient of error.	CNN	Adam	-	Cross-entropy	-	Python 3.9, Keras 3	[262]
12		Remove Muscle Artifacts from EEG	EEGdenoiseNet	-	Supervised	CNN for myogenic artifact reduction contains seven similar blocks. In each of the first six blocks, two 1D-convolution layers with small 1*3 kernels, 1 stride, and a ReLU activation function are followed by a 1D-Average pooling layer with pool size equal to two. In the seventh block, two 1D-convolution layers are followed by a flatten layer.The network gradually reduce the EEG signal sampling rate by the 1D-Average pooling layer.	CNN	RMSprop	ReLU	mean squared error (MSE)	-	Python 3.10, Tensorflow 2.8	[263]
13		Denoise EEG signal from artifacts	EEGdenoiseNet	MultiResUNet3+	Supervised	Net3+ consists of full-blown pass-through connections that aggregate connections between encoders and decoders and internal connections between decoder subnets. Instead of directly combining the encoder and decoder functions, the encoder functions go through several convolutional levels with residual connections	CNN, encoders	Adam	Rely	mean squared error (MSE)	-	-	[152]