資料與模型
梯度定義為 asinh(Δtask / (1.4826 × MAD(Δbaseline)))，不再做逐通道 task Z-score。Gradient-TCN 使用 3/7/15-sample stems、dilation 1/2/4/8，並分別彙整 signed、energy 與 C3/C4 lateral features。Dual model 保留原始 Raw Shallow branch，再以 learned gate 融合 gradient branch。
| mode | parameters |
|---|---|
| MAD-gradient Shallow | 3082 |
| Gradient-TCN full | 52707 |
| TCN without energy | 46435 |
| TCN without lateral | 45635 |
| TCN without signed pool | 40163 |
| Raw + Gradient dual | 122524 |
結果摘要
20+20 最佳為 Raw Shallow：58.00% ± 7.83%。All-trial 最佳為 Raw Shallow：67.94% ± 14.59%。Dual branch 在 20+20 的平均 raw gate 為 0.541；越接近 1 代表越依賴 raw branch。
主要判讀
梯度包含部分受試者的任務資訊，但目前不是穩定的跨受試者分類表示。20+20 的 MAD-gradient Shallow 相對 Raw Shallow 為 -4.21%（95% CI -8.10% 至 -0.66%）；all-trial 為 -9.26%（95% CI -16.05% 至 -2.74%）。增加資料量沒有消除差距，表示問題不只在模型容量或小樣本。
Dual branch 最接近 raw 基準，但仍沒有提升：20+20 差 -1.68%（95% CI -3.28% 至 -0.14%，3/10 人改善）；all-trial 差 -1.03%（95% CI -3.32% 至 1.55%，4/10 人改善）。平均 raw gate 分別為 0.541 與 0.571，模型略偏向 raw branch，但沒有學到可穩定超越 raw 的互補資訊。
Full TCN 與移除 signed、energy、lateral 後的差異小且在 20+20/all-trial 間方向不一致，因此目前不能把任何一個 pooling component 視為可靠機制。較合理的結論是：一階差分突顯短暫高頻變化，對少數人有辨識力，但也刪除了 Raw Shallow 所利用的慢變頻帶能量。
20+20
| mode | n_subjects | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 | mean_raw_gate |
|---|---|---|---|---|---|---|
| MAD-gradient Shallow | 10 | 53.80% | 5.49% | 53.79% | 47.10% | |
| Gradient-TCN full | 10 | 52.32% | 2.38% | 52.30% | 44.15% | |
| TCN without energy | 10 | 51.12% | 2.32% | 51.10% | 41.60% | |
| TCN without lateral | 10 | 52.07% | 3.61% | 52.05% | 44.49% | |
| TCN without signed pool | 10 | 51.92% | 3.26% | 51.93% | 44.61% | |
| Raw + Gradient dual | 10 | 56.32% | 6.08% | 56.34% | 49.56% | 0.541259 |
| Raw Shallow | 10 | 58.00% | 7.83% | 58.01% | 52.68% |
相對 Raw Shallow
| mode | reference | mean_difference | bootstrap_ci_low | bootstrap_ci_high | wilcoxon_p | subjects_improved | subjects_equal | n_subjects |
|---|---|---|---|---|---|---|---|---|
| MAD-gradient Shallow | Raw Shallow | -4.21% | -8.10% | -0.66% | 0.130859 | 3 | 0 | 10 |
| Gradient-TCN full | Raw Shallow | -5.68% | -9.85% | -2.11% | 0.003906 | 1 | 0 | 10 |
| TCN without signed pool | Raw Shallow | -6.08% | -9.65% | -2.83% | 0.001953 | 0 | 0 | 10 |
| TCN without energy | Raw Shallow | -6.88% | -11.56% | -2.74% | 0.013672 | 2 | 0 | 10 |
| TCN without lateral | Raw Shallow | -5.93% | -10.35% | -2.17% | 0.009766 | 1 | 0 | 10 |
| Raw + Gradient dual | Raw Shallow | -1.68% | -3.28% | -0.14% | 0.130859 | 3 | 0 | 10 |
20+20 個別受試者
| Subject | Raw Shallow | MAD-gradient Shallow | Gradient-TCN full | Raw + Gradient dual | Dual vs Raw |
|---|---|---|---|---|---|
| sub1 | 71.07% | 58.41% | 52.55% | 64.89% | -6.17% |
| sub2 | 50.11% | 50.07% | 49.28% | 49.59% | -0.52% |
| sub3 | 59.00% | 53.00% | 54.33% | 58.33% | -0.67% |
| sub9 | 52.56% | 52.23% | 52.57% | 50.57% | -1.99% |
| sub10 | 57.50% | 50.33% | 52.00% | 52.33% | -5.17% |
| sub11 | 68.67% | 52.17% | 52.67% | 64.83% | -3.83% |
| sub12 | 55.75% | 57.38% | 52.57% | 56.24% | 0.49% |
| sub13 | 64.83% | 66.33% | 57.17% | 63.00% | -1.83% |
| sub14 | 51.50% | 47.83% | 51.33% | 52.33% | 0.83% |
| sub15 | 49.05% | 50.23% | 48.74% | 51.06% | 2.00% |
Gradient-TCN 消融
Contribution 定義為 full TCN accuracy 減去移除該 component 後的 accuracy；正值代表該 component 有幫助。
| Component | Full accuracy | Without accuracy | Contribution |
|---|---|---|---|
| Signed pooling | 52.32% | 51.92% | 0.40% |
| Gradient energy | 52.32% | 51.12% | 1.20% |
| C3/C4 lateral | 52.32% | 52.07% | 0.25% |
All-trial
| mode | n_subjects | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 | mean_raw_gate |
|---|---|---|---|---|---|---|
| MAD-gradient Shallow | 10 | 58.68% | 7.58% | 58.68% | 54.99% | |
| Gradient-TCN full | 10 | 55.06% | 6.53% | 55.17% | 50.65% | |
| TCN without energy | 10 | 54.61% | 5.64% | 54.67% | 50.06% | |
| TCN without lateral | 10 | 55.86% | 6.00% | 55.90% | 50.00% | |
| TCN without signed pool | 10 | 56.51% | 6.54% | 56.58% | 52.71% | |
| Raw + Gradient dual | 10 | 66.91% | 12.96% | 66.90% | 65.86% | 0.570921 |
| Raw Shallow | 10 | 67.94% | 14.59% | 67.96% | 65.39% |
| mode | reference | mean_difference | bootstrap_ci_low | bootstrap_ci_high | wilcoxon_p | subjects_improved | subjects_equal | n_subjects |
|---|---|---|---|---|---|---|---|---|
| MAD-gradient Shallow | Raw Shallow | -9.26% | -16.05% | -2.74% | 0.037109 | 2 | 0 | 10 |
| Gradient-TCN full | Raw Shallow | -12.88% | -19.16% | -6.77% | 0.001953 | 0 | 0 | 10 |
| TCN without signed pool | Raw Shallow | -11.43% | -17.68% | -5.38% | 0.007812 | 1 | 1 | 10 |
| TCN without energy | Raw Shallow | -13.34% | -20.32% | -6.33% | 0.009766 | 2 | 0 | 10 |
| TCN without lateral | Raw Shallow | -12.08% | -18.74% | -5.47% | 0.009766 | 2 | 0 | 10 |
| Raw + Gradient dual | Raw Shallow | -1.03% | -3.32% | 1.55% | 0.390625 | 4 | 0 | 10 |
All-trial 個別受試者
| Subject | Raw Shallow | MAD-gradient Shallow | Gradient-TCN full | Raw + Gradient dual | Dual vs Raw |
|---|---|---|---|---|---|
| sub1 | 84.21% | 63.50% | 57.55% | 84.77% | 0.56% |
| sub2 | 56.63% | 50.21% | 49.74% | 54.74% | -1.89% |
| sub3 | 80.00% | 64.00% | 60.50% | 73.50% | -6.50% |
| sub9 | 50.28% | 55.72% | 49.80% | 58.22% | 7.94% |
| sub10 | 70.50% | 56.50% | 53.00% | 68.00% | -2.50% |
| sub11 | 86.50% | 57.00% | 60.00% | 83.50% | -3.00% |
| sub12 | 75.10% | 64.65% | 53.27% | 75.12% | 0.02% |
| sub13 | 75.00% | 73.00% | 68.50% | 71.00% | -4.00% |
| sub14 | 48.00% | 54.50% | 47.00% | 51.00% | 3.00% |
| sub15 | 53.21% | 47.76% | 51.26% | 49.24% | -3.96% |
判讀原則
若 gradient-only 模式低、dual branch接近或高於 Raw Shallow，表示梯度是輔助資訊而非可替代表示；若移除某個 pooling 反而提高準確率，表示該統計量在目前樣本量下主要帶來噪聲。所有比較均以 subject-level mean accuracy 做 bootstrap CI 與 Wilcoxon。