Table 5 Results for five proposed architectures across hybrid categories utilizing the proposed preprocessing method: FD-F under random split evaluation strategy.
From: EEG imagined speech neuro-signal preprocessing and deep learning classification
Architecture | Category | ||||||||
|---|---|---|---|---|---|---|---|---|---|
20-Class (CharDig) | 30-Class (CharDigObj) | Training time (s) | Testing time (s) | ||||||
FD-F | Vs Full-bands (baseline) | Vs TD-F (baseline) | FD-F | vs. Full-bands (baseline) | vs. TD-F (baseline) | ||||
CNN-1-LSTM (baseline) | Balanced Acc(%) | 98.28 ± 0.45 | + 17.84 | + 18.92 | 97.91 ± 0.34 | + 21.06 | + 22.91 | 532.73 ± 88.67 | 1.17 ± 0.02 |
Accuracy (%) | 98.28 ± 0.44 | 97.90 ± 0.33 | |||||||
Macro F1(%) | 98.28 ± 0.45 | 97.91 ± 0.33 | |||||||
CNN-2-LSTM | Balanced Acc(%) | 99.24 ± 0.26 | + 7.76 | + 6.99 | 99.12 ± 0.11 | + 7.34 | + 7.17 | 689.16 ± 44.25 | 1.42 ± 0.02 |
Accuracy (%) | 99.24 ± 0.26 | 99.12 ± 0.11 | |||||||
Macro F1(%) | 99.24 ± 0.26 | 99.12 ± 0.11 | |||||||
CNN-2-Bi-LSTM | Balanced Acc(%) | 99.33 ± 0.08 | + 7.65 | + 6.68 | 99.38 ± 0.08 | + 9.86 | + 6.42 | 908.22 ± 131.06 | 2.06 ± 0.04 |
Accuracy (%) | 99.33 ± 0.09 | 99.38 ± 0.08 | |||||||
Macro F1(%) | 99.33 ± 0.08 | 99.38 ± 0.08 | |||||||
CNN-3-LSTM | Balanced Acc(%) | 99.16 ± 0.38 | + 8.98 | + 6.19 | 99.90 ± 0.37 | + 8.50 | + 7.78 | 813.53 ± 62.12 | 1.73 ± 0.01 |
Accuracy (%) | 99.16 ± 0.38 | 99.10 ± 0.37 | |||||||
Macro F1(%) | 99.16 ± 0.38 | 99.10 ± 0.37 | |||||||
3-LSTM | Balanced Acc(%) | 99.13 ± 0.24 | + 9.75 | + 7.08 | 99.08 ± 0.15 | + 9.86 | + 8.74 | 968.05 ± 248.35 | 1.74 ± 0.03 |
Accuracy (%) | 99.12 ± 0.24 | 99.09 ± 0.14 | |||||||
Macro F1(%) | 99.13 ± 0.24 | 99.08 ± 0.15 | |||||||
