Table 4.
Performance overview of all assessed methods on all datasets.
| No Whitening | Whitening | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Dataset | LDA | rLDAc | LDA | TVLDA | ||||||||||||
| CSP | Channel Sel. | CSP | Channel Sel. | PCAd | PCAd | |||||||||||
| Protocola | Gridsb | ID | Acc. (%) | NF | Acc. (%) | NF | Acc. (%) | Acc. (%) | NF | Acc. (%) | NF | Acc. (%) | NF | Acc. (%) | NF | |
| RPS | Macro | S1 | 63.8 ± 3.0 | 2 | 52.2 ± 3.1 | 4 | 68.8 ± 2.5 | 73.8 ± 2.0 | 1 | 69.1 ± 0.9 | 1 | 65.4 ± 1.5 | 1 | 82.4 ± 2.0 | 1 | |
| RPS | Macro | S6 | 53.6 ± 2.2 | 3 | 57.3 ± 1.5 | 9 | 79.9 ± 0.8 | 76.5 ± 1.4 | 8 | 77.6 ± 1.0 | 8 | 79.2 ± 0.8 | 1 | 91.3 ± 0.9 | 2 | |
| Average | 58.7 ± 2.6 | 2.5 | 54.7 ± 2.4 | 6.5 | 74.4 ± 1.9 | 75.2 ± 1.7 | 4.5 | 73.4 ± 0.9 | 4.5 | 72.3 ± 1.2 | 1.0 | 86.9 ± 1.6 | 1.5 | |||
| RPS | Micro | S2 | 75.9 ± 1.7 | 4 | 77.3 ± 2.2 | 8 | 89.2 ± 1.0 | 91.7 ± 1.2 | 3 | 82.0 ± 2.4 | 11 | 91.6 ± 0.8 | 1 | 99.0 ± 0.3 | 1 | |
| RPS | Micro | S4 | 77.4 ± 2.3 | 6 | 79.8 ± 1.2 | 4 | 67.8 ± 1.5 | 90.7 ± 1.0 | 11 | 90.3 ± 1.0 | 13 | 92.5 ± 0.7 | 2 | 98.9 ± 0.2 | 1 | |
| RPS | Micro | S5 | 87.4 ± 1.8 | 5 | 89.0 ± 1.1 | 7 | 96.3 ± 0.4 | 95.0 ± 0.6 | 4 | 95.7 ± 0.6 | 5 | 98.4 ± 0.3 | 1 | 99.0 ± 0.4 | 1 | |
| Average | 80.2 ± 2.0 | 5.0 | 82.0 ± 1.6 | 6.3 | 84.4 ± 1.1 | 92.5 ± 1.0 | 6.0 | 89.4 ± 1.5 | 9.7 | 94.2 ± 0.7 | 1.3 | 99.0 ± 0.3 | 1.0 | |||
| FTPU | Macro | S7 | 54.9 ± 3.6 | 4 | 60.2 ± 3.0 | 7 | 79.3 ± 1.7 | 65.3 ± 3.5 | 2 | 65.6 ± 2.3 | 4 | 76.0 ± 2.1 | 1 | 89.4 ± 1.3 | 1 | |
| FTPU | Macro | S8 | 56.6 ± 1.7 | 1 | 63.5 ± 2.3 | 4 | 71.8 ± 1.9 | 75.6 ± 2.4 | 2 | 79.4 ± 1.6 | 5 | 69.1 ± 2.2 | 1 | 82.8 ± 1.2 | 1 | |
| FTPU | Macro | S9 | 53.5 ± 3.4 | 1 | 70.8 ± 3.6 | 7 | 83.3 ± 1.1 | 72.9 ± 2.1 | 2 | 75.4 ± 2.1 | 3 | 78.6 ± 2.3 | 1 | 85.7 ± 1.2 | 1 | |
| FTPU | Macro | S10 | 55.8 ± 3.2 | 4 | 62.8 ± 2.1 | 2 | 60.8 ± 1.6 | 57.7 ± 1.9 | 1 | 73.9 ± 1.8 | 2 | 71.0 ± 1.9 | 1 | 77.3 ± 2.0 | 1 | ⋆ |
| FTPU | Macro | S11 | 27.0 ± 2.9 | 1 | 38.0 ± 2.5 | 3 | 50.4 ± 1.6 | 39.3 ± 2.1 | 1 | 50.8 ± 2.7 | 5 | 50.4 ± 1.9 | 1 | 64.5 ± 3.2 | 1 | |
| FTPU | Macro | S12 | 40.0 ± 0.0 | 1 | 53.3 ± 0.0 | 1 | 70.0 ± 0.0 | 60.0 ± 0.0 | 1 | 63.3 ± 0.0 | 1 | 80.0 ± 0.0 | 1 | 90.0 ± 0.0 | 1 | ⋆ |
| FTPU | Macro | S13 | 49.9 ± 2.7 | 3 | 57.4 ± 2.0 | 1 | 74.6 ± 1.8 | 66.0 ± 1.5 | 1 | 72.4 ± 1.8 | 3 | 68.1 ± 2.8 | 1 | 80.1 ± 1.7 | 2 | |
| FTPU | Macro | S14 | 55.7 ± 0.0 | 2 | 67.1 ± 0.0 | 2 | 71.4 ± 0.0 | 60.0 ± 0.0 | 3 | 64.3 ± 0.0 | 5 | 78.6 ± 0.0 | 1 | 81.4 ± 0.0 | 1 | ⋆ |
| FTPU | Macro | S15 | 53.5 ± 1.7 | 1 | 58.9 ± 2.3 | 2 | 58.9 ± 2.0 | 68.7 ± 1.4 | 1 | 68.9 ± 1.8 | 2 | 70.9 ± 1.4 | 1 | 77.5 ± 1.7 | 1 | ⋆ |
| Average | 49.6 ± 2.8 | 2.0 | 59.1 ± 2.6 | 3.2 | 68.9 ± 1.7 | 62.8 ± 2.2 | 1.6 | 68.2 ± 2.1 | 3.3 | 71.4 ± 2.1 | 1.0 | 81.0 ± 1.9 | 1.1 | |||
| FTPU | Micro | S2 | 77.8 ± 1.5 | 5 | 80.9 ± 1.4 | 4 | 80.2 ± 1.3 | 79.8 ± 1.7 | 7 | 85.3 ± 1.2 | 7 | 87.5 ± 1.1 | 1 | 97.2 ± 0.6 | 1 | |
| FTPU | Micro | S3 | 93.3 ± 1.8 | 2 | 83.8 ± 1.1 | 3 | 85.8 ± 1.2 | 95.3 ± 0.6 | 2 | 93.1 ± 0.7 | 3 | 91.0 ± 1.4 | 2 | 93.8 ± 1.4 | 1 | |
| FTPD | Micro | S4 | 67.7 ± 2.4 | 12 | 65.3 ± 1.3 | 8 | 50.5 ± 0.9 | 85.3 ± 0.9 | 5 | 85.0 ± 0.7 | 10 | 89.3 ± 0.5 | 1 | 97.9 ± 0.3 | 1 | |
| FTPU | Micro | S4 | 71.0 ± 1.1 | 7 | 75.0 ± 0.9 | 5 | 51.7 ± 1.0 | 89.0 ± 0.7 | 4 | 88.8 ± 0.6 | 9 | 89.2 ± 0.6 | 1 | 96.6 ± 0.3 | 2 | |
| FTPU | Micro | S5 | 58.1 ± 1.5 | 12 | 55.4 ± 1.0 | 9 | 68.2 ± 0.9 | 64.3 ± 1.1 | 8 | 65.2 ± 0.9 | 12 | 70.6 ± 0.8 | 1 | 85.3 ± 0.8 | 2 | |
| Average | 73.6 ± 1.7 | 7.6 | 72.1 ± 1.2 | 5.8 | 67.3 ± 1.1 | 82.7 ± 1.0 | 5.2 | 83.5 ± 0.9 | 8.2 | 85.5 ± 0.9 | 1.2 | 94.2 ± 0.8 | 1.4 | |||
The table is organized in four blocks, such that the rock-paper-scissor experiments with standard and high-density electrode grids are clustered in the first and second block, respectively. Likewise, the finger-tapping experiments with standard and high-density electrode grids are presented in the third and fourth block, respectively. Accuracies and corresponding number of features NF are representative values, as described in the text. The percentages are given as means ± standard deviation over 20 cross-validation repetitions. If the standard deviation is missing, it was zero due to the low number of trials. Emphasized values are considered the best for each dataset, which lie within the quantization margin of the best result (cf. section 2.9.3).
RPS, rock-paper-scissors; FTPD, finger tapping, palm down; FTPU, finger tapping, palm up.
Macro, standard ECoG grid; Micro, high-density ECoG grid.
rLDA operates on vectorized features, so NF is not applicable (cf. section 2.6.2).
For rows marked with an asterisk (⋆), a regularization of 25% was used for PCA (cf. section 2.7).