EEG Signal Classification — Literature Review Table

Ref #	Author(s)	Year	Focus	Preprocessing / Signal Processing	Feature Extraction	ML / DL Algorithm(s)	Best Acc.	Key Notes
1	Muhammad & Al-Ahmadi	2022	Anxiety	Bandpass Filter Ocular Artifact Removal Muscular Artifact Removal Statistical Channel Selection DASPS database (23 subjects); EEG noise filtered; statistically significant electrodes selected for binary & 4-level classification	Frequency Domain Features Theta Band Power Beta Band Power Feature Selection (subset)	Random Forest (RF) SVM KNN LDA Naïve Bayes	94.90% (binary)	PLoS ONE (2022). Best: RF with only 9–10 features. 2-level & 4-level anxiety. Theta+Beta bands outperformed all others.
2	Aldayel & Al-Nafjan	2024	Anxiety	Standard EEG Preprocessing Artifact Removal Epoching DASPS dataset; HAM-A & SAM labeling; signal segmented into epochs for feature extraction	DWT (Discrete Wavelet Transform) PSD (Power Spectral Density) Valence-Arousal Features Time-Frequency Features	Random Forest (RF) AdaBoost Bagging Gradient Bagging SVM KNN LDA	~90%+ (ensemble)	PeerJ CS (2024). Compared HAM-A vs SAM labeling. DWT outperformed PSD. Ensemble methods > classical classifiers. Two anxiety levels (anxious vs calm).
3	Siuly, Li & Zhang	2016	EEG Analysis (Book)	Bandpass Filtering Notch Filter (50/60 Hz) ICA (Independent Component Analysis) EOG/EMG Artifact Removal Re-referencing	FFT (Fast Fourier Transform) Wavelet Transform AR (Autoregressive) Model Time-domain Stats Entropy Features	SVM ANN LDA Bayesian Classifier KNN	N/A (book)	Springer Book (2016). Comprehensive EEG processing pipeline covering BCI, epilepsy, sleep, emotion. Canonical reference for EEG ML methods.
4	Rajya Lakshmi, Prasad & Chandra Prakash	2014	Survey	CAR (Common Average Reference) Surface Laplacian (SL) ICA PCA Adaptive Filtering (LMS)	ICA Components PCA Dimensions Frequency Band Power	ICA-based Classification PCA-based Classification BCI Paradigms	Survey (no single acc.)	IJARCSSE 2014. Compares ICA vs CAR vs SL vs PCA for artifact removal & feature extraction. ICA: best for large datasets; PCA: best dimensionality reduction.
5	Islam, Ahmed et al.	2023	Review	Bandpass / Notch Filtering ICA / PCA Wavelet Denoising ASR (Artifact Subspace Reconstruction) Epoching & Baseline Correction	Time-domain (mean, variance, skewness) Frequency-domain (PSD, FFT) Time-Frequency (DWT, STFT, CWT) Nonlinear (entropy, fractal) Connectivity Features	SVM RF CNN LSTM CNN-LSTM Hybrid Transformer	Review (various)	Sensors 2023. Most comprehensive modern EEG review. Covers full pipeline: acquisition → denoising → feature engineering → classification. Identifies DL as dominant trend.
6	Al-Ezzi et al.	2020	Anxiety / ERP	Bandpass Filter (0.1–40 Hz) ICA Artifact Removal Re-referencing Epoch Segmentation	Event-Related Potentials (ERP) ERP Components (N200, P300) Brain Connectivity (coherence) Frequency Band Power (Alpha, Theta)	SVM LDA Logistic Regression ANN	~85% (varies)	Biomedical Signal Processing & Control (2020). Reviews EEG+ERP for Social Anxiety Disorder (SAD). Alpha asymmetry & theta power most discriminative features.
7	Al-Nafjan & Aldayel	2022	Attention	EEG Signal Segmentation Bandpass Filtering Artifact Rejection	PSD (Power Spectral Density) FFT-based Attention Index Theta/Alpha/Beta Band Ratios	Random Forest (RF) SVM KNN	96% (RF)	Cited in PeerJ CS (2024). Online student attention detection. RF outperformed SVM & KNN. Theta/Alpha ratio key attention biomarker.
8	Baghdadi, Aribi et al.	2019–21	Anxiety	DASPS Database Psychological Stimulation Protocol Standard EEG Preprocessing Artifact Removal	DWT (Time-Frequency) PSD (Frequency Domain) Time-domain Features Valence-Arousal Dimensions	SVM KNN ANN RF	~83% (DWT+SVM)	Creators of the DASPS dataset. Found DWT + frequency features best for anxiety. Dataset used by refs 1 & 2. 4 anxiety levels labeled.
9	Various (Deep Learning Era)	2021–23	Emotion / Anxiety	Bandpass Filter (1–45 Hz) ICA Euclidean Alignment Epoch Segmentation (2s windows) Normalization	Raw EEG (end-to-end DL) STFT Spectrograms CWT Maps Graph-based Connectivity Spatial-Temporal Features	CNN LSTM CNN-LSTM Hybrid BiLSTM Graph Neural Network (GNN) Transformer / Attention	96–97% (DEAP, STC-CNN)	State-of-art DL approaches on DEAP/SEED datasets. CNN handles spatial, LSTM handles temporal features. GNNs model inter-channel connectivity. Transformer-based models emerging.

① Preprocessing Pipeline

Standard Steps:
1. Bandpass filter (typically 0.5–40 Hz or 1–45 Hz)
2. Notch filter (50/60 Hz power line)
3. Artifact removal (ICA, PCA, or threshold-based)
4. EOG/EMG artifact rejection
5. Re-referencing (average, linked mastoid, or CSD)
6. Epoch segmentation (1–4s windows with overlap)
7. Baseline correction

② Feature Extraction Methods

Frequency Domain: PSD, FFT, band power (δ,θ,α,β,γ)
Time-Frequency: DWT, CWT, STFT spectrograms
Time Domain: mean, variance, skewness, kurtosis
Nonlinear: sample entropy, Hjorth parameters
Connectivity: coherence, phase-locking value
DL (end-to-end): raw signal or 2D spectrogram input

③ ML / DL Classification

Classical ML: SVM (most common), KNN, LDA, Naïve Bayes, RF
Ensemble: Random Forest, AdaBoost, Gradient Boosting
Deep Learning: CNN (spatial), LSTM/BiLSTM (temporal)
Hybrid: CNN-LSTM, CNN-BiLSTM
Advanced: Graph NN, Transformers, Attention mechanisms
Key bands for anxiety: Theta (4–8 Hz) + Beta (13–30 Hz)
Key bands for attention: Theta/Alpha ratio