Table 5 Reviewed papers that used GAN in other EEG applications
From: Generative adversarial networks in EEG analysis: an overview
Study | Purpose | Dataset | GAN type | Evaluation metrics | Results (with GAN) | ||||
|---|---|---|---|---|---|---|---|---|---|
Khadijah et al. 2019 [18] | Improve the accuracy, convergence rate, and generalization capabilities of the model | Private dataset  +  Nao dataset | DCGAN  +  WGAN | Classification accuracy | G + R data train DCGAN: + 3% WGAN: + 2% | G data train DCGAN: + 1% WGAN: + 3% | G data train generalization DCGAN: + 12% WGAN: + 11% | ||
Yao et al., 2020 [80] | Specify several standards for operating on EEG data to protect users’ privacy | From UCI, Neuro-dynamics Laboratory at the State University of New York | ResNet generator + patchGAN feature filter | Privacy indicator | Filter out over 90% of alcoholism information on average from EEG signals, with an average of only 4.2% useful feature accuracy lost | ||||
Hazra and Byun, 2020 [82] | Eliminate confidentiality concerns of medical data | Siena Scalp  +  private dataset | SynSigGAN | • Pearson Coefficient • MAE • RMSE • PRD • FD | • 0.997 • 0.0475 • 0.0314 • 5.985 • 0.982 | ||||
Fan et al., 2020 [83] | Address the challenges of automatic sleep staging models such as the inherent class imbalance problem | Montreal Archive of Sleep Studies (MASS)  +  Sleep-EDF-SC | DCGAN | • Classification accuracy • F1 score • Cohen Kappa | MASS • 3.79% • 3.48% • 5.43% | Sleep-EDF • 4.51% • 3.14% • 5.8% | |||
Zeng et al., 2021 [86] | Address the issue of the different distribution of EEG across subjects | Private dataset | GDANN | • Classification accuracy • Precision • F1Score • Recall | •  + 11.9% •  + 9.34% •  + 9.64% •  + 10% | ||||
Hazra et al., 2021 [87] | Develop a cost-effective system for cognitive state classification using ambulatory EEG signals | Private dataset | DCGAN Classifier | Classification accuracy | Compared to CNN • GTCC – MFCC: + 0.6% • GTCC – MFCC – CNN: + 0.3% • GTCC: – 1.33% | ||||
Yin et al., 2021 [78] | Use multivariate time series data in the process of prediction | NASDAQ100 + SML2010 + Energy + EEG + Air Quality | MAGAN | • MAE • SMAPE | ɛ = 50 w.r.t MARNN • − 0.0455 • − 0.0244 | ||||
Yin et al.., 2021 [79] | Improve the accuracy of the long-term prediction | NASDAQ + SML + Energy + EEG + KDDCUP | VAEcGAN | • MAE • RMSE | ɛ = 120 w.r.t LSTM • − 0.1529 • − 0.1334 | ||||
Tazrin et al. 2021 [88] | Solve computational and energy resource issues of IoT devices with EEG headbands/headsets | Confused student dataset | DCGAN | Classification accuracy | CNN & DNN >  + 20% | ||||
Cheon et al. 2021 [90] | Overcome issues of gathering a large dataset of EEG | Confused student dataset | CTGAN + TGAN | • Basic statistics • Correlation column correlations • Mean correlation • 1 – MAPE • Similarity score | CTGAN • 0.9963 • 0.9476 • 0.9393 • 0.7250 • 0.9021 | TGAN • 0.9876 • 0.0881 • 0.9351 • 0.8552 • 0.7165 | |||
Lee et al. 2021 [91] | Improve classification of sleep stages | Publicly Sleep-EDF database (PSG) test | Sig-GAN | • Classification accuracy • IS • FID | • 65.67% with only first 30-s signals (Real data 82.85%) • − 0.51 (w.r.t real data)  • + 39.53 | ||||