資料表示與模型
magnitude = log1p(|Δtask / MAD(Δbaseline)|)；direction 同時保留 hard sign 與 tanh(slope/2) 軟可信度。斜率限制在 ±50，避免近乎平坦的 baseline 放大雜訊。模型內使用 causal dilated convolution，但輸入仍是 1–40 Hz zero-phase SOS IIR。
| mode | parameters |
|---|---|
| Raw + Sign-Magnitude fusion | 168717 |
| Sign-Magnitude full | 107188 |
| Hard-sign only | 103604 |
| Direction only | 42819 |
| Magnitude only | 39667 |
| Without wavelet init | 107188 |
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%。
雙流 full 尚未高於 Raw Shallow。 Full 相對 Raw Shallow：20+20 -6.84%（95% CI -11.19% 至 -2.86%）；all-trial -15.38%（95% CI -23.61% 至 -6.77%）。
加入 raw 的 fusion：20+20 -0.85%（95% CI -2.35% 至 0.54%，4/10 人改善）；all-trial -3.19%（95% CI -5.88% 至 -0.32%，3/10 人改善）。
主要判讀
這組結果不支持「局部斜率的正負交替是左右腳分類的主要特徵」。Full 在 20+20 為 51.17%，all-trial 為 52.57%；增加訓練 trial 後只增加 1.40%，仍接近 chance，表示瓶頸不是單純模型容量或樣本數。
Magnitude-only 為 52.40% / 52.95%，direction-only 為 50.56% / 52.66%（20+20 / all）。方向流沒有高於 magnitude 流；full 也沒有高於兩個單流，未觀察到 sign–magnitude 解耦後的互補增益。這裡的 direction 是一階斜率符號，不等同 Hilbert phase，因此不能據此否定 EEG 相位特徵整體。
Hard-only 為 51.42% / 54.46%，反而略高於 hard+soft full；no-wavelet 為 51.29% / 54.59%，all-trial 高於 wavelet full。這表示軟方向可信度和短尺度 Ricker/Morlet 初始化都沒有提供穩定樣本效率，all-trial 下小波先驗甚至形成限制。
Raw fusion 是唯一接近基準的模式：20+20 57.16%，相對 Raw 的 CI 跨 0；all-trial 64.76%，仍低 3.19%，CI 完全低於 0。Raw gate 為 0.546 / 0.569，模型略偏向 raw，但 SignMag 分支仍使完整資料表現下降。合理結論是斜率可作個體診斷或輔助資訊，主分類表示仍應保留 raw waveform 與 Mu/Beta 慢變功率。
20+20
| mode | n_subjects | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 | mean_raw_gate |
|---|---|---|---|---|---|---|
| Raw + Sign-Magnitude fusion | 10 | 57.16% | 6.69% | 57.15% | 50.50% | 0.545781 |
| Raw Shallow | 10 | 58.00% | 7.83% | 58.01% | 52.68% | |
| Sign-Magnitude full | 10 | 51.17% | 1.83% | 51.16% | 41.61% | |
| Hard-sign only | 10 | 51.42% | 2.18% | 51.43% | 42.07% | |
| Direction only | 10 | 50.56% | 2.35% | 50.58% | 43.47% | |
| Magnitude only | 10 | 52.40% | 3.14% | 52.44% | 42.94% | |
| Without wavelet init | 10 | 51.29% | 2.48% | 51.34% | 42.89% |
| mode | reference | mean_difference | bootstrap_ci_low | bootstrap_ci_high | wilcoxon_p | subjects_improved | subjects_equal | n_subjects |
|---|---|---|---|---|---|---|---|---|
| Sign-Magnitude full | Raw Shallow | -6.84% | -11.19% | -2.86% | 0.009766 | 2 | 0 | 10 |
| Hard-sign only | Raw Shallow | -6.58% | -11.08% | -2.57% | 0.013672 | 2 | 0 | 10 |
| Without wavelet init | Raw Shallow | -6.71% | -11.08% | -2.62% | 0.027344 | 3 | 0 | 10 |
| Magnitude only | Raw Shallow | -5.60% | -9.63% | -1.98% | 0.019531 | 2 | 0 | 10 |
| Direction only | Raw Shallow | -7.44% | -12.23% | -2.98% | 0.027344 | 3 | 0 | 10 |
| Raw + Sign-Magnitude fusion | Raw Shallow | -0.85% | -2.35% | 0.54% | 0.287109 | 4 | 0 | 10 |
20+20 個別受試者
| Subject | Raw Shallow | Full | Magnitude only | Direction only | Raw + SignMag | Fusion vs Raw |
|---|---|---|---|---|---|---|
| sub1 | 71.07% | 51.21% | 54.03% | 49.27% | 65.31% | -5.76% |
| sub2 | 50.11% | 52.04% | 50.56% | 50.42% | 50.24% | 0.13% |
| sub3 | 59.00% | 50.00% | 51.00% | 49.00% | 58.83% | -0.17% |
| sub9 | 52.56% | 50.41% | 49.78% | 48.45% | 51.92% | -0.64% |
| sub10 | 57.50% | 51.00% | 51.33% | 47.33% | 54.67% | -2.83% |
| sub11 | 68.67% | 51.67% | 52.33% | 50.83% | 70.17% | 1.50% |
| sub12 | 55.75% | 51.09% | 52.40% | 51.58% | 55.37% | -0.38% |
| sub13 | 64.83% | 55.67% | 60.67% | 55.83% | 61.67% | -3.17% |
| sub14 | 51.50% | 49.33% | 51.33% | 51.67% | 51.67% | 0.17% |
| sub15 | 49.05% | 49.25% | 50.58% | 51.26% | 51.74% | 2.69% |
雙流消融
Difference 為 full 減去替代模式；正值代表保留該設計較好。Magnitude-only 與 direction-only 同時也用來判斷兩路是否具有互補性。
| Component | 20+20 full | 20+20 alternative | 20+20 difference | All full | All alternative | All difference |
|---|---|---|---|---|---|---|
| Soft-direction confidence | 51.17% | 51.42% | -0.25% | 52.57% | 54.46% | -1.90% |
| Wavelet initialization | 51.17% | 51.29% | -0.12% | 52.57% | 54.59% | -2.02% |
| Direction stream | 51.17% | 52.40% | -1.23% | 52.57% | 52.95% | -0.39% |
| Magnitude stream | 51.17% | 50.56% | 0.60% | 52.57% | 52.66% | -0.10% |
All-trial
| mode | n_subjects | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 | mean_raw_gate |
|---|---|---|---|---|---|---|
| Raw + Sign-Magnitude fusion | 10 | 64.76% | 12.82% | 64.73% | 61.18% | 0.568723 |
| Raw Shallow | 10 | 67.94% | 14.59% | 67.96% | 65.39% | |
| Sign-Magnitude full | 10 | 52.57% | 3.30% | 52.66% | 45.26% | |
| Hard-sign only | 10 | 54.46% | 4.40% | 54.39% | 47.24% | |
| Direction only | 10 | 52.66% | 3.11% | 52.64% | 48.02% | |
| Magnitude only | 10 | 52.95% | 4.54% | 52.99% | 46.40% | |
| Without wavelet init | 10 | 54.59% | 4.00% | 54.60% | 49.01% |
| mode | reference | mean_difference | bootstrap_ci_low | bootstrap_ci_high | wilcoxon_p | subjects_improved | subjects_equal | n_subjects |
|---|---|---|---|---|---|---|---|---|
| Sign-Magnitude full | Raw Shallow | -15.38% | -23.61% | -6.77% | 0.013672 | 1 | 0 | 10 |
| Hard-sign only | Raw Shallow | -13.48% | -21.54% | -4.93% | 0.027344 | 3 | 0 | 10 |
| Without wavelet init | Raw Shallow | -13.35% | -21.26% | -5.56% | 0.013672 | 1 | 0 | 10 |
| Magnitude only | Raw Shallow | -14.99% | -22.68% | -7.38% | 0.013672 | 2 | 0 | 10 |
| Direction only | Raw Shallow | -15.28% | -23.13% | -7.43% | 0.009766 | 2 | 0 | 10 |
| Raw + Sign-Magnitude fusion | Raw Shallow | -3.19% | -5.88% | -0.32% | 0.083984 | 3 | 0 | 10 |
All-trial 個別受試者
| Subject | Raw Shallow | Full | Magnitude only | Direction only | Raw + SignMag | Fusion vs Raw |
|---|---|---|---|---|---|---|
| sub1 | 84.21% | 52.71% | 56.65% | 50.70% | 79.29% | -4.91% |
| sub2 | 56.63% | 52.68% | 57.15% | 52.71% | 49.77% | -6.87% |
| sub3 | 80.00% | 52.00% | 57.50% | 52.50% | 71.50% | -8.50% |
| sub9 | 50.28% | 49.26% | 49.24% | 51.79% | 56.18% | 5.90% |
| sub10 | 70.50% | 48.00% | 46.50% | 49.50% | 65.50% | -5.00% |
| sub11 | 86.50% | 53.50% | 51.00% | 54.00% | 83.50% | -3.00% |
| sub12 | 75.10% | 52.76% | 55.24% | 57.68% | 76.07% | 0.98% |
| sub13 | 75.00% | 58.00% | 58.50% | 58.00% | 66.00% | -9.00% |
| sub14 | 48.00% | 57.50% | 50.50% | 49.00% | 49.50% | 1.50% |
| sub15 | 53.21% | 49.27% | 47.26% | 50.76% | 50.24% | -2.96% |
判讀原則
若 direction-only 接近或高於 magnitude-only，才支持正負交替本身具辨識力；若 full 高於兩個單流，才支持解耦後融合具有互補性；若 raw fusion 才能接近基準，表示斜率是輔助訊息，不能取代原始 EEG 的慢變頻帶功率。