From: Generative adversarial networks in EEG analysis: an overview
Study | Purpose | Dataset | GAN Type | Evaluation metrics | Results (with GAN) | |
---|---|---|---|---|---|---|
Abdelfattah et al., 2018 [26] | Enhance the model classification performance | PhysioNet | RGAN | Reconstruction accuracy | Reconstruction: • + 19.9% (w.r.t VAE) • + 34.8% (w.r.t AE) | |
Classification accuracy | Classification (25% dataset): • DNN + 36.1% • SVM + 14.1% • RFT + 13.1% Classification (50% dataset): • DNN + 39.1% • SVM + 12.8% • RFT + 12.6% | |||||
Hartmann et al., 2018 [27] | Achieve stabilization of the training | Private dataset | WGANs | • Classification accuracy • IS • FID • ED • SWD | • 91.25% • 1.363 • 9.523 • − 0.056 • 0.078 | |
Corley and Huang 2018 [28] | Produce high spatial resolution EEG data from low-resolution samples | Berlin BCI Competition III, Dataset V | WGANs | • Classification accuracy | Scale 2: 3.87 < HR Scale 4: 5.75 < HR | |
• MSE • MAE | Scale 2: • − 37,497,940 • − 3885.34 | Scale 4: • − 72,991,320 • − 6385.61 | ||||
Fahimi, Zhang et al., 2019 [29] | Enhance the model classification performance | A public dataset collected by [40] | DCGANs | Classification accuracy | NA | |
Fahimi, Dosen et al., 2021 [30] | Enhance the model classification performance | Private dataset + BCI competition III, Dataset IVa | DCGANs | Classification accuracy | • Diverted attention: + 7.32% (p < 0.01) • Focused attention: + 5.45% (p < 0.01) • IVa: + 3.57% (p < 0.02) | |
Li and Yu 2020 [32] | Enhance the model classification performance | BCI competition IV, Dataset 2b | cWGAN-GP | Classification accuracy | (w.r.t raw data) • Shallow + 1.65% • Deep4 + 2.89% | |
Debie et al., 2020 [33] | Protect EEG brain signals against illegal disclosure | BCI competition IV, Dataset 2a | GAN with differential privacy | Classification accuracy | NP-GAN (max:150 trial) • SVM + 5.74% • RF + 3.43% • LDA + 6.39% • LR + 9.54% PP-GAN (max: 50 trial) • SVM − 1.05% • RF + 0.36% • LDA − 0.19% • LR + 0.18% | |
Zhang et al., 2020 [15] | Avoid overfitting and enhance the model classification performance | BCI competition IV (datasets 1  + 2b) | CNN-DCGAN | Classification accuracy | • D1: + 8.7% (1:3) • D2b: + 12.6% (1:3) | |
kappa value | • D1: + 0.1622% • D2b: + 0.1981% | |||||
Luo et al., 2020 [35] | Reconstruct EEG signal with high sampling rates and sensitivity | Private dataset +  Lucid et al., 2014 +  BCI competition IV, dataset 2a | WGAN  +  (TSF-MSE) loss function | Classification accuracy | • MI: + 2.03% • AO: + 4.1% • GAL: + 4.11% | |
Reconstruction accuracy | • MI: + 3.2% • AO: + 4.5% • GAL: + 5.38% | |||||
Yang et al., 2021 [38] | Address the challenge of insufficient MI data | Private dataset  +  BCI competition IV, Dataset 2a | cVAE- GAN | Classification accuracy | 86.14% D1 mean ~  + 4% D2 mean ~  + 1.5% | |
• IS • FID • SWD | w.r.t Real • − 0.121 •  + 11.364 •  + 0.067 | |||||
Xie et al., 2021 [21] | Address the challenge of insufficient MI data | BCI competition IV, datasets 2a + 2b | LGANs (augmentation)  +  MoCNN (classification)  +  attention network | Classification accuracy | D1 (w.r.t raw data) • LGAN + 8.23% • Att-LAGN + 9.34% D2 (w.r.t raw data) • Att-LAGN: + 5.64% − 6.6% | |
Xu et al., 2021 [39] | Enhance the model classification performance | Private dataset | CycleGAN | Classification accuracy |  + 18.3% |