中榮資料：8通道 Sequential 左右腳分類
資料由另一台 1000 Hz 設備收錄，通道為 Fp1/Fp2/Fz/C3/C4/Pz/O1/O2。資料先連續做 1–40 Hz 四階 Butterworth zero-phase SOS IIR、降至 200 Hz，切 event 前 1 秒至後 2 秒；模型使用 event 後 0–2 秒，trial/channel Z-score。
資料與評估規則
-1 = Left、1 = Right。s3_1 全段排除；s1_5、s2_1~s2_4 依資料記錄排除指定的 target-trial 編號。每位受試者的 session 依序進行：先以更新前 production model 計算 before accuracy；再從該 session 做 stratified 5-fold，於每 fold 僅以 train trials fine-tune 並在 held-out test trials 計算 after accuracy；最後才以該 session 全部有效 trials 更新 production model，供下一段使用。
EEGNet 與 ShallowConvNet 先以 34 位既有正常人資料做 source supervised pretraining；Support-CSP 先以同一批來源資料進行 30 epochs support/query episodic pretraining。s1、s2 的 sequential chain 各自從相同 source base 重新開始。
Sequential 結果
| Subject | Session | Model | Before | After 5-fold adaptation | 有效 trials | Left/Right |
|---|---|---|---|---|---|---|
| s1 | s1_1 | EEGNet | 62.50% | 67.50% ± 18.96% | 40 | 20/20 |
| s1 | s1_1 | ShallowConvNet | 62.50% | 47.50% ± 10.46% | 40 | 20/20 |
| s1 | s1_1 | Support-CSP episodic | - | 67.50% ± 6.85% | 40 | 20/20 |
| s1 | s1_2 | EEGNet | 32.50% | 42.50% ± 14.25% | 40 | 20/20 |
| s1 | s1_2 | ShallowConvNet | 60.00% | 32.50% ± 11.18% | 40 | 20/20 |
| s1 | s1_2 | Support-CSP episodic | 42.50% | 47.50% ± 24.04% | 40 | 20/20 |
| s1 | s1_3 | EEGNet | 45.00% | 45.00% ± 14.25% | 40 | 20/20 |
| s1 | s1_3 | ShallowConvNet | 50.00% | 45.00% ± 11.18% | 40 | 20/20 |
| s1 | s1_3 | Support-CSP episodic | 50.00% | 40.00% ± 24.04% | 40 | 20/20 |
| s1 | s1_4 | EEGNet | 70.00% | 65.00% ± 16.30% | 40 | 20/20 |
| s1 | s1_4 | ShallowConvNet | 50.00% | 50.00% ± 23.39% | 40 | 20/20 |
| s1 | s1_4 | Support-CSP episodic | 57.50% | 47.50% ± 16.30% | 40 | 20/20 |
| s1 | s1_5 | EEGNet | 51.28% | 53.57% ± 23.32% | 39 | 19/20 |
| s1 | s1_5 | ShallowConvNet | 43.59% | 58.57% ± 24.98% | 39 | 19/20 |
| s1 | s1_5 | Support-CSP episodic | 46.15% | 58.57% ± 21.63% | 39 | 19/20 |
| s2 | s2_1 | EEGNet | 54.29% | 60.00% ± 21.19% | 35 | 17/18 |
| s2 | s2_1 | ShallowConvNet | 42.86% | 60.00% ± 11.95% | 35 | 17/18 |
| s2 | s2_1 | Support-CSP episodic | - | 54.29% ± 11.95% | 35 | 17/18 |
| s2 | s2_2 | EEGNet | 33.33% | 38.93% ± 15.59% | 36 | 18/18 |
| s2 | s2_2 | ShallowConvNet | 50.00% | 58.57% ± 12.78% | 36 | 18/18 |
| s2 | s2_2 | Support-CSP episodic | 47.22% | 53.21% ± 16.77% | 36 | 18/18 |
| s2 | s2_3 | EEGNet | 48.39% | 45.71% ± 15.19% | 31 | 15/16 |
| s2 | s2_3 | ShallowConvNet | 54.84% | 53.81% ± 25.46% | 31 | 15/16 |
| s2 | s2_3 | Support-CSP episodic | 54.84% | 48.57% ± 12.21% | 31 | 15/16 |
| s2 | s2_4 | EEGNet | 50.00% | 50.48% ± 14.91% | 34 | 16/18 |
| s2 | s2_4 | ShallowConvNet | 44.12% | 58.57% ± 21.67% | 34 | 16/18 |
| s2 | s2_4 | Support-CSP episodic | 47.06% | 64.76% ± 23.71% | 34 | 16/18 |
受試者摘要
| Subject | Model | Mean after across sessions | Sessions |
|---|---|---|---|
| s1 | EEGNet | 54.71% ± 11.34% | 5 |
| s1 | ShallowConvNet | 46.71% ± 9.45% | 5 |
| s1 | Support-CSP episodic | 52.21% ± 10.81% | 5 |
| s2 | EEGNet | 48.78% ± 8.85% | 4 |
| s2 | ShallowConvNet | 57.74% ± 2.70% | 4 |
| s2 | Support-CSP episodic | 55.21% ± 6.83% | 4 |
注意：每段只有 31–40 個有效 trials，單 fold test 約 6–8 trials，故 after accuracy 的標準差大；這是跨設備與小樣本 adaptation 的初步結果，不適合解讀為穩定的模型排序。
不累積訓練比較
Independent 版中，每一段都從同一份 34 人 source base model 重新開始，不帶入前一段的模型權重或 CSP reference；after 同樣是該 session 的 5-fold adaptation。
| Model | Sequential after mean | Independent after mean | Independent - Sequential |
|---|---|---|---|
| EEGNet | 52.08% | 52.42% | +0.34 pp |
| ShallowConvNet | 51.61% | 50.86% | -0.75 pp |
| Support-CSP episodic | 53.54% | 52.42% | -1.12 pp |
整體差異很小：EEGNet 幾乎持平；ShallowConvNet 與 Support-CSP 的累積式平均略高。以目前僅兩位受試者、每段 31–40 trials 的規模，不能視為穩定優勢。
不累積訓練：逐段結果
每一列都從 source base 重新開始。Support-CSP 的 before 為 -，因為 source base 尚未具有此新 session 可用的 CSP/prototype context；after 為該 session 的 5-fold held-out accuracy。
| Subject | Session | Model | Before source base | After 5-fold adaptation | 有效 trials | Left/Right |
|---|---|---|---|---|---|---|
| s1 | s1_1 | EEGNet | 62.50% | 67.50% ± 18.96% | 40 | 20/20 |
| s1 | s1_1 | ShallowConvNet | 62.50% | 45.00% ± 11.18% | 40 | 20/20 |
| s1 | s1_1 | Support-CSP episodic | - | 60.00% ± 5.59% | 40 | 20/20 |
| s1 | s1_2 | EEGNet | 42.50% | 45.00% ± 18.96% | 40 | 20/20 |
| s1 | s1_2 | ShallowConvNet | 42.50% | 30.00% ± 11.18% | 40 | 20/20 |
| s1 | s1_2 | Support-CSP episodic | - | 40.00% ± 25.62% | 40 | 20/20 |
| s1 | s1_3 | EEGNet | 47.50% | 50.00% ± 17.68% | 40 | 20/20 |
| s1 | s1_3 | ShallowConvNet | 40.00% | 42.50% ± 22.71% | 40 | 20/20 |
| s1 | s1_3 | Support-CSP episodic | - | 40.00% ± 10.46% | 40 | 20/20 |
| s1 | s1_4 | EEGNet | 52.50% | 52.50% ± 22.36% | 40 | 20/20 |
| s1 | s1_4 | ShallowConvNet | 37.50% | 45.00% ± 18.96% | 40 | 20/20 |
| s1 | s1_4 | Support-CSP episodic | - | 32.50% ± 30.10% | 40 | 20/20 |
| s1 | s1_5 | EEGNet | 51.28% | 48.21% ± 19.60% | 39 | 19/20 |
| s1 | s1_5 | ShallowConvNet | 53.85% | 56.07% ± 28.05% | 39 | 19/20 |
| s1 | s1_5 | Support-CSP episodic | - | 61.79% ± 17.11% | 39 | 19/20 |
| s2 | s2_1 | EEGNet | 54.29% | 65.71% ± 12.78% | 35 | 17/18 |
| s2 | s2_1 | ShallowConvNet | 42.86% | 57.14% ± 14.29% | 35 | 17/18 |
| s2 | s2_1 | Support-CSP episodic | - | 60.00% ± 15.65% | 35 | 17/18 |
| s2 | s2_2 | EEGNet | 44.44% | 38.57% ± 17.20% | 36 | 18/18 |
| s2 | s2_2 | ShallowConvNet | 41.67% | 58.21% ± 10.38% | 36 | 18/18 |
| s2 | s2_2 | Support-CSP episodic | - | 56.07% ± 21.26% | 36 | 18/18 |
| s2 | s2_3 | EEGNet | 48.39% | 48.57% ± 16.97% | 31 | 15/16 |
| s2 | s2_3 | ShallowConvNet | 54.84% | 48.10% ± 14.72% | 31 | 15/16 |
| s2 | s2_3 | Support-CSP episodic | - | 51.90% ± 8.81% | 31 | 15/16 |
| s2 | s2_4 | EEGNet | 47.06% | 55.71% ± 3.19% | 34 | 16/18 |
| s2 | s2_4 | ShallowConvNet | 38.24% | 75.71% ± 23.47% | 34 | 16/18 |
| s2 | s2_4 | Support-CSP episodic | - | 69.52% ± 21.46% | 34 | 16/18 |
檔案
sequential session results · independent session results · comparison CSV · sequential fold results · session manifest · trial manifest · run code