目前狀態
完整評估已完成；all-trial 目前最佳為 Shallow 0-2 s 71.95%，20+20 目前最佳為 Shallow 0-2 s 59.75%。
200 Hz
0.5 s window / 0.1 s hop
8 ch × 23 freq × 26 frames
both
方法
連續 raw EEG 經四階 1–40 Hz Butterworth SOS zero-phase filtering、降採樣至 200 Hz，再依 event 擷取 −1~2 秒。模型先建立 absolute log-power 與 fixation-relative dB，再依序執行逐通道 temporal convolution、全八通道 spatial projection、frequency convolution、depthwise-separable residual blocks 與 event-aware mean/log-variance pooling。
STFT robust normalizer 僅以各 outer fold 的 train trials 估計；validation/test 不參與。Shallow 同時提供 0~2 秒歷史窗口及 −1~2 秒 matched-window 對照。
原始時頻結構
表示篩選
Absolute、baseline-relative、兩者合併只根據 inner-validation balanced accuracy 鎖定，screening 階段不讀 outer test。
| representation | completed_folds | mean_val_balanced_accuracy | std_val_balanced_accuracy | mean_val_accuracy |
|---|---|---|---|---|
| both | 50 | 62.61% | 9.84% | 62.53% |
| absolute | 50 | 62.03% | 8.46% | 61.95% |
| relative | 50 | 61.99% | 9.26% | 61.86% |
群組結果
| protocol | model | representation | subjects | mean_accuracy | std_subject_accuracy | mean_balanced_accuracy | mean_macro_f1 |
|---|---|---|---|---|---|---|---|
| 20x20 | Event Log-STFT | both | 10 | 53.50% | 7.09% | 53.50% | 48.51% |
| 20x20 | Shallow -1-2 s | waveform | 10 | 59.65% | 9.39% | 59.65% | 56.64% |
| 20x20 | Shallow 0-2 s | waveform | 10 | 59.75% | 10.58% | 59.75% | 56.18% |
| all | Event Log-STFT | both | 10 | 55.07% | 6.72% | 55.05% | 52.87% |
| all | Shallow -1-2 s | waveform | 10 | 71.19% | 16.70% | 71.18% | 70.22% |
| all | Shallow 0-2 s | waveform | 10 | 71.95% | 17.58% | 71.94% | 71.16% |
結果判讀
主要結果：all-trial 下 Event Log-STFT 為 55.07%，低於 Shallow 0–2 秒的 71.95%，低 16.88 個百分點 （subject-level bootstrap 95% CI -26.36 至 -7.59 個百分點；Wilcoxon p=0.0039；1/10 人改善）。這不支持以目前 Log-STFT 架構取代波形型 ShallowConvNet。
少試次：20+20 下 Event Log-STFT 為 53.50%，Shallow 0–2 秒為 59.75%；低 6.25 個百分點 （subject-level bootstrap 95% CI -12.75 至 0.25 個百分點；Wilcoxon p=0.1055；3/10 人改善）。目前時頻表示在小樣本下也沒有呈現較強的歸納偏置。
事件前資訊：Shallow 增加 −1–0 秒後，all-trial 為 71.19%（相對 0–2 秒 -0.76 個百分點），20+20 為 59.65%（-0.10 個百分點）。因此「單純增加事件前一秒」未帶來可見改善；若要繼續使用 baseline，需要更有約束的 task/baseline 對比機制。
模型診斷：occlusion 影響最大的通道為 O2，頻率約 35.9 Hz，時間約為事件後 1.5 秒。這種依賴偏向後枕/高頻與任務後段，而非穩定的中央運動區節律，顯示模型可能依賴不易泛化的非中央區代理特徵。
單一受試者
| protocol | subject_id | model | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 |
|---|---|---|---|---|---|---|
| 20x20 | sub1 | Event Log-STFT | 55.50% | 16.57% | 55.50% | 49.96% |
| 20x20 | sub10 | Event Log-STFT | 49.50% | 17.49% | 49.50% | 43.16% |
| 20x20 | sub11 | Event Log-STFT | 50.00% | 13.01% | 50.00% | 46.12% |
| 20x20 | sub12 | Event Log-STFT | 50.00% | 14.88% | 50.00% | 43.37% |
| 20x20 | sub13 | Event Log-STFT | 71.00% | 13.37% | 71.00% | 69.87% |
| 20x20 | sub14 | Event Log-STFT | 54.00% | 10.66% | 54.00% | 47.44% |
| 20x20 | sub15 | Event Log-STFT | 47.00% | 15.00% | 47.00% | 41.44% |
| 20x20 | sub2 | Event Log-STFT | 46.50% | 20.89% | 46.50% | 43.00% |
| 20x20 | sub3 | Event Log-STFT | 56.00% | 14.49% | 56.00% | 50.45% |
| 20x20 | sub9 | Event Log-STFT | 55.50% | 18.43% | 55.50% | 50.28% |
| 20x20 | sub1 | Shallow -1-2 s | 71.50% | 13.27% | 71.50% | 70.27% |
| 20x20 | sub10 | Shallow -1-2 s | 54.50% | 18.00% | 54.50% | 51.51% |
| 20x20 | sub11 | Shallow -1-2 s | 69.50% | 19.12% | 69.50% | 66.88% |
| 20x20 | sub12 | Shallow -1-2 s | 59.00% | 15.10% | 59.00% | 56.87% |
| 20x20 | sub13 | Shallow -1-2 s | 63.00% | 15.51% | 63.00% | 60.11% |
| 20x20 | sub14 | Shallow -1-2 s | 51.50% | 14.12% | 51.50% | 47.13% |
| 20x20 | sub15 | Shallow -1-2 s | 50.00% | 17.31% | 50.00% | 45.93% |
| 20x20 | sub2 | Shallow -1-2 s | 50.00% | 16.93% | 50.00% | 46.23% |
| 20x20 | sub3 | Shallow -1-2 s | 74.50% | 14.65% | 74.50% | 73.38% |
| 20x20 | sub9 | Shallow -1-2 s | 53.00% | 15.43% | 53.00% | 48.07% |
| 20x20 | sub1 | Shallow 0-2 s | 69.50% | 17.34% | 69.50% | 67.49% |
| 20x20 | sub10 | Shallow 0-2 s | 60.50% | 17.56% | 60.50% | 58.47% |
| 20x20 | sub11 | Shallow 0-2 s | 74.00% | 14.40% | 74.00% | 72.58% |
| 20x20 | sub12 | Shallow 0-2 s | 60.50% | 16.41% | 60.50% | 58.28% |
| 20x20 | sub13 | Shallow 0-2 s | 63.50% | 15.28% | 63.50% | 59.90% |
| 20x20 | sub14 | Shallow 0-2 s | 51.50% | 12.14% | 51.50% | 45.10% |
| 20x20 | sub15 | Shallow 0-2 s | 49.00% | 12.46% | 49.00% | 43.86% |
| 20x20 | sub2 | Shallow 0-2 s | 49.50% | 13.25% | 49.50% | 44.33% |
| 20x20 | sub3 | Shallow 0-2 s | 74.00% | 15.28% | 74.00% | 72.32% |
| 20x20 | sub9 | Shallow 0-2 s | 45.50% | 14.83% | 45.50% | 39.45% |
| all | sub1 | Event Log-STFT | 57.43% | 10.74% | 57.37% | 56.30% |
| all | sub10 | Event Log-STFT | 56.83% | 8.94% | 56.83% | 52.89% |
| all | sub11 | Event Log-STFT | 54.50% | 5.36% | 54.50% | 52.77% |
| all | sub12 | Event Log-STFT | 53.08% | 6.85% | 53.04% | 52.37% |
| all | sub13 | Event Log-STFT | 72.50% | 6.05% | 72.50% | 72.36% |
| all | sub14 | Event Log-STFT | 51.17% | 6.61% | 51.17% | 48.77% |
| all | sub15 | Event Log-STFT | 48.27% | 5.05% | 48.15% | 44.86% |
| all | sub2 | Event Log-STFT | 50.67% | 7.95% | 50.74% | 48.49% |
| all | sub3 | Event Log-STFT | 53.83% | 9.40% | 53.83% | 50.94% |
| all | sub9 | Event Log-STFT | 52.38% | 6.55% | 52.32% | 48.99% |
| all | sub1 | Shallow -1-2 s | 92.60% | 3.51% | 92.58% | 92.59% |
| all | sub10 | Shallow -1-2 s | 73.00% | 7.39% | 73.00% | 72.88% |
| all | sub11 | Shallow -1-2 s | 95.50% | 4.14% | 95.50% | 95.48% |
| all | sub12 | Shallow -1-2 s | 79.41% | 5.30% | 79.40% | 79.34% |
| all | sub13 | Shallow -1-2 s | 74.17% | 5.15% | 74.17% | 73.86% |
| all | sub14 | Shallow -1-2 s | 50.67% | 7.70% | 50.67% | 48.73% |
| all | sub15 | Shallow -1-2 s | 51.57% | 5.49% | 51.59% | 49.16% |
| all | sub2 | Shallow -1-2 s | 55.97% | 8.24% | 55.89% | 52.80% |
| all | sub3 | Shallow -1-2 s | 82.50% | 5.75% | 82.50% | 82.34% |
| all | sub9 | Shallow -1-2 s | 56.52% | 6.82% | 56.53% | 55.04% |
| all | sub1 | Shallow 0-2 s | 94.09% | 3.04% | 94.07% | 94.08% |
| all | sub10 | Shallow 0-2 s | 74.00% | 5.49% | 74.00% | 73.93% |
| all | sub11 | Shallow 0-2 s | 94.17% | 3.36% | 94.17% | 94.16% |
| all | sub12 | Shallow 0-2 s | 82.92% | 3.10% | 82.94% | 82.85% |
| all | sub13 | Shallow 0-2 s | 76.67% | 5.32% | 76.67% | 76.53% |
| all | sub14 | Shallow 0-2 s | 50.33% | 6.74% | 50.33% | 46.65% |
| all | sub15 | Shallow 0-2 s | 50.09% | 7.08% | 50.09% | 49.71% |
| all | sub2 | Shallow 0-2 s | 57.48% | 3.43% | 57.41% | 56.38% |
| all | sub3 | Shallow 0-2 s | 85.50% | 5.76% | 85.50% | 85.42% |
| all | sub9 | Shallow 0-2 s | 54.24% | 3.98% | 54.23% | 51.86% |
配對比較
| protocol | comparison | subjects | mean_difference | bootstrap_ci_low | bootstrap_ci_high | improved_subjects | wilcoxon_p |
|---|---|---|---|---|---|---|---|
| 20x20 | event_log_stft - shallow_task | 10 | -6.25% | -12.75% | 0.25% | 3 | 0.10546875 |
| 20x20 | event_log_stft - shallow_full | 10 | -6.15% | -11.70% | -0.70% | 3 | 0.060546875 |
| all | event_log_stft - shallow_task | 10 | -16.88% | -26.36% | -7.59% | 1 | 0.00390625 |
| all | event_log_stft - shallow_full | 10 | -16.12% | -25.68% | -7.44% | 1 | 0.00390625 |
模型是否自行找到重要通道
空間權重與 occlusion 完全由模型學習，不預先指定 C3、C4 或左右差值。
可重現檔案
fold_results.csv · subject_summary.csv · group_summary.csv · paired_comparisons.csv · learned_spatial_weights.csv · occlusion_importance.csv · 程式碼快照