正常人與中風患者 8通道三模型左右腳分類比較
統一使用 Fp1, Fp2, Fz, C3, C4, Pz, O1, O2，連續資料先做 1–40 Hz zero-phase SOS IIR、降至 200 Hz，取事件後 0–2 秒，並對每個 trial、每個通道 Z-score。未使用 ASR、ICA 或 EA。
群組比較
| 群組 | 資料量 | 模型 | Accuracy mean ± STD | Balanced accuracy |
|---|---|---|---|---|
| Normal | All trials | EEGNet | 66.32% ± 11.73% | 66.32% |
| Normal | All trials | ShallowConvNet | 71.91% ± 17.58% | 71.91% |
| Normal | All trials | Support-CSP + ShallowConvNet | 73.16% ± 15.45% | 73.16% |
| Stroke | All trials | EEGNet | 75.58% ± 17.18% | 75.48% |
| Stroke | All trials | ShallowConvNet | 80.61% ± 15.88% | 80.65% |
| Stroke | All trials | Support-CSP + ShallowConvNet | 81.71% ± 14.84% | 81.69% |
| Normal | 20 Left + 20 Right | EEGNet | 56.90% ± 7.24% | 56.90% |
| Normal | 20 Left + 20 Right | ShallowConvNet | 60.95% ± 9.97% | 60.95% |
| Normal | 20 Left + 20 Right | Support-CSP + ShallowConvNet | 63.95% ± 9.60% | 63.95% |
| Stroke | 20 Left + 20 Right | EEGNet | 65.40% ± 11.04% | 65.40% |
| Stroke | 20 Left + 20 Right | ShallowConvNet | 72.00% ± 14.52% | 72.00% |
| Stroke | 20 Left + 20 Right | Support-CSP + ShallowConvNet | 73.40% ± 14.69% | 73.40% |
比較說明
ShallowConvNet 是和 Support-CSP 使用完全相同 split、試驗數與資料增強的 scratch baseline。Support-CSP + ShallowConvNet 以 outer-train 中的 calibration trials 建立 Mu、low beta、high beta 的 CSP 空間濾波器與類別 prototype，再將 CSP log-variance/prototype 距離特徵與 ShallowConvNet 波形特徵融合。EEGNet 為標準 8通道 EEGNet 對照組。