結論摘要
20+20 最佳模式為 ShallowConvNet scratch，跨受試者平均 58.00% ± 7.83%。
Staged joint fine-tune 在 20+20 為 55.94%，比 matched Shallow 低 2.06%；all-trial 為 66.54%，仍低於 Shallow 的 67.94%。本實驗未達成相對 Shallow +3 pp、CI 下界高於 0 或至少 7/10 人改善的成功標準。
去噪器先以 30 位連續 EEG 選擇訓練步數，再用全部 34 位重訓；正式 10 位及 outer test 均未參與預訓練或 checkpoint 選擇。
結果判讀
去噪器保留了大部分 MI 頻帶結構：Mu/Beta power correlation 為 0.9954，C3/C4 covariance correlation 為 0.9931，側化方向保持率為 95.66%。然而，未污染輸入的 residual RMS 為 0.135 SD，超過 0.10 SD 安全門檻；固定合成污染的 waveform MSE 也由 0.0440 上升至 0.0602（+0.0162）。因此本版屬於保留頻帶大方向、但存在過度修正且未真正降低污染誤差，不能宣稱去噪有效，分類負結果與此診斷一致。
all-trial 明顯高於 20+20，表示這個資料集的主要限制仍較像個人化標記 trial 數量與 subject-specific variability，而不是單純增加模型容量即可解決。
20+20 主結果
| mode | n_subjects | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 | corrupted_accuracy |
|---|---|---|---|---|---|---|
| Attention-Residual scratch | 10 | 54.90% | 4.97% | 54.87% | 46.48% | 54.45% |
| Frozen denoiser + classifier | 10 | 54.41% | 5.55% | 54.47% | 46.00% | 53.68% |
| Staged joint fine-tune | 10 | 55.94% | 6.44% | 55.94% | 50.67% | 55.44% |
| ShallowConvNet scratch | 10 | 58.00% | 7.83% | 58.01% | 52.68% | 57.47% |
相對 Shallow 的配對檢定
| mode | reference | mean_difference | bootstrap_ci_low | bootstrap_ci_high | wilcoxon_p | subjects_improved | n_subjects |
|---|---|---|---|---|---|---|---|
| Attention-Residual scratch | shallow_scratch | -3.11% | -7.49% | 0.41% | 0.232422 | 4 | 10 |
| Frozen denoiser + classifier | shallow_scratch | -3.59% | -8.41% | 1.06% | 0.160156 | 2 | 10 |
| Staged joint fine-tune | shallow_scratch | -2.06% | -5.74% | 1.27% | 0.275391 | 4 | 10 |
統計單位是受試者，bootstrap 95% CI 與 Wilcoxon 均由 subject-level mean accuracy 計算。
All-trial 容量上限
| mode | n_subjects | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 | corrupted_accuracy |
|---|---|---|---|---|---|---|
| Attention-Residual scratch | 10 | 66.66% | 12.27% | 66.65% | 65.27% | 66.17% |
| Frozen denoiser + classifier | 10 | 65.71% | 13.10% | 65.68% | 63.19% | 64.66% |
| Staged joint fine-tune | 10 | 66.54% | 12.92% | 66.53% | 64.97% | 65.81% |
| ShallowConvNet scratch | 10 | 67.94% | 14.59% | 67.96% | 65.39% | 65.60% |
去噪安全檢查
| Check | Value | Criterion | Pass |
|---|---|---|---|
| Identity residual RMS | 0.135087 | < 0.10 SD | False |
| Mu/Beta power correlation | 0.995448 | > 0.95 | True |
| C3-C4 lateral direction | 0.956551 | > 95% | True |
| Corrupted MSE before | 0.043999 | diagnostic | |
| Corrupted MSE after | 0.060218 | lower is better | False |
流程
1–40 Hz zero-phase SOS IIR → 200 Hz → 0–2 s → trial/channel Z-score。分類訓練增強僅有 0.95–1.05 振幅、每通道小量 DC 與 ±50 ms shift。未使用 ASR、ICA、EA、Rest 或 pseudo-clean teacher。