Table 3.
Comparison with other methods.
| Methods | Data | Accuracy | |||
|---|---|---|---|---|---|
| Gait | EEG | HC vs. MCI/AD (%) | MCI vs. AD (%) | Three-way classification (%) | |
| Handcrafted features + SVM | ✓ | 63.64 | 57.73 | 55.45 | |
| Handcrafted feature + RF | ✓ | 81.82 | 57.14 | 68.18 | |
| AST-GCN(ours) | ✓ | 93.09 | 58.41 | 68.51 | |
| standard CNN | ✓ | – | 69.66 | – | |
| EEGnet | ✓ | – | 97.85 | – | |
| ResNet 18 | ✓ | – | 97.59 | – | |
| VGG 13 | ✓ | – | 96.48 | – | |
| ST-CNN(ours) | ✓ | – | 98.63 | – | |
| cascade neural network(ours) | ✓ | ✓ | 93.09 | 98.63 | 91.07 |
Standard CNN represents the model we substitute 2D convolution layers with a kernel size of Ks × Kt for ST-CNN modules. “Handcrafted features + SVM” and “Handcrafted features + RF” indicate the methods using different classifiers with the handcrafted features same as (10). The bold values indicates the best performance that method obtain in that experiment.