8通道 Support-conditioned CSP-ShallowNet
每個 outer fold 只用 train trials 建立 Mu、low beta、high beta shrinkage-CSP filters 與 left/right prototypes。CSP log-variance + prototype distance 會和 0–2 s Shallow waveform features 融合；episodic 版本先在 34 位未使用正常人的 support/query episodes 上訓練。
結果
| 設定 | 模型 | Accuracy mean ± STD | BACC |
|---|---|---|---|
| 20 Left + 20 Right | Global ShallowConvNet | 59.40% ± 10.69% | 59.40% |
| 20 Left + 20 Right | Support-CSP Shallow episodic-pretrained | 61.25% ± 9.86% | 61.25% |
| 20 Left + 20 Right | Support-CSP Shallow scratch | 60.60% ± 10.37% | 60.60% |
| All trials | Global ShallowConvNet | 72.47% ± 17.25% | 72.44% |
| All trials | Support-CSP Shallow episodic-pretrained | 73.03% ± 15.16% | 73.03% |
| All trials | Support-CSP Shallow scratch | 71.61% ± 17.52% | 71.61% |
方法界線
不使用 ASR、ICA、EA 或 test labels。CSP、prototype 與標準化都在每個 outer-train support fit；target validation/test 僅 transform。source 34 位只作 supervised episodic pretraining，沒有參與 target outer test。