中榮資料：ASR cutoff 10 品質診斷與三模型獨立評估
此頁只處理中榮新設備的 8 通道資料；34 位既有正常人的 source pretraining 權重沒有重新做 ASR，也沒有改變。
處理流程
連續 raw EEG（1000 Hz）→ 四階 Butterworth 1–40 Hz zero-phase SOS IIR → ASR cutoff=10 → 降採樣至 200 Hz → event 後 0–2 s → 每 trial、每通道 Z-score。ASR 在連續濾波資料上擬合與修正；不使用 ICA、EA 或累積 session 訓練。
評估為每一 session 獨立從同一份 34 人 source base 出發，於該 session 做 stratified 5-fold adaptation。Raw IIR 與 ASR10 使用相同有效 trial、模型架構與 source checkpoint。
品質診斷
ASR calibration 比例表示連續訊號中 ASR 判定可作為乾淨校正參考的窗口比例。所有 session 都有約 25–56% 可用校正窗口，但 ASR 的 correction RMS 仍高，尤其 s1_1、s1_2、s2_1；代表資料污染並非輕微。
| Session | ASR calibration | 1-40 Hz SD before | SD after ASR10 | ASR correction RMS |
|---|---|---|---|---|
| s1_1 | 39.57% | 49.56 nV | 9.13 nV | 48.85 nV |
| s1_2 | 24.90% | 48.06 nV | 9.73 nV | 47.24 nV |
| s1_3 | 55.51% | 33.06 nV | 8.80 nV | 32.02 nV |
| s1_4 | 47.88% | 36.49 nV | 11.98 nV | 35.08 nV |
| s1_5 | 53.16% | 35.21 nV | 9.69 nV | 34.13 nV |
| s2_1 | 25.99% | 48.18 nV | 15.51 nV | 45.96 nV |
| s2_2 | 44.52% | 28.67 nV | 16.46 nV | 24.00 nV |
| s2_3 | 54.64% | 29.06 nV | 22.30 nV | 19.13 nV |
| s2_4 | 48.12% | 28.19 nV | 21.57 nV | 18.69 nV |
模型結果：ASR10 沒有一致改善
結論：雖然 ASR10 明顯降低連續訊號振幅，但目前 9 段、各 31–40 trials 的獨立 5-fold 結果沒有顯示穩定分類提升。EEGNet 幾乎持平（平均 -0.50 pp）；ShallowConvNet 與 Support-CSP 反而平均下降。這表示 cutoff 10 對此資料可能過於積極，或清掉了同時被模型使用的任務相關結構。
| Model | Raw IIR mean | ASR10 mean | ASR10 - raw | Win / tie / loss |
|---|---|---|---|---|
| EEGNet | 52.42% | 51.92% | -0.50 pp | 5/1/3 |
| ShallowConvNet | 50.86% | 48.81% | -2.05 pp | 4/1/4 |
| Support-CSP episodic | 52.42% | 48.78% | -3.64 pp | 2/0/7 |
Win/tie/loss 是在 9 段 session 中，ASR10 after accuracy 相對 raw IIR 的上升／相同／下降次數。這僅是描述性比較；每段樣本非常少，不能做強結論。
逐段 5-fold held-out 結果
| Subject | Session | Model | Trials | Raw IIR | ASR10 | Difference |
|---|---|---|---|---|---|---|
| s1 | s1_1 | EEGNet | 40 | 67.50% | 67.50% | +0.00 pp |
| s1 | s1_2 | EEGNet | 40 | 45.00% | 52.50% | +7.50 pp |
| s1 | s1_3 | EEGNet | 40 | 50.00% | 45.00% | -5.00 pp |
| s1 | s1_4 | EEGNet | 40 | 52.50% | 55.00% | +2.50 pp |
| s1 | s1_5 | EEGNet | 39 | 48.21% | 40.71% | -7.50 pp |
| s2 | s2_1 | EEGNet | 35 | 65.71% | 51.43% | -14.29 pp |
| s2 | s2_2 | EEGNet | 36 | 38.57% | 41.79% | +3.21 pp |
| s2 | s2_3 | EEGNet | 31 | 48.57% | 54.76% | +6.19 pp |
| s2 | s2_4 | EEGNet | 34 | 55.71% | 58.57% | +2.86 pp |
| s1 | s1_1 | ShallowConvNet | 40 | 45.00% | 45.00% | +0.00 pp |
| s1 | s1_2 | ShallowConvNet | 40 | 30.00% | 52.50% | +22.50 pp |
| s1 | s1_3 | ShallowConvNet | 40 | 42.50% | 40.00% | -2.50 pp |
| s1 | s1_4 | ShallowConvNet | 40 | 45.00% | 37.50% | -7.50 pp |
| s1 | s1_5 | ShallowConvNet | 39 | 56.07% | 58.93% | +2.86 pp |
| s2 | s2_1 | ShallowConvNet | 35 | 57.14% | 40.00% | -17.14 pp |
| s2 | s2_2 | ShallowConvNet | 36 | 58.21% | 63.93% | +5.71 pp |
| s2 | s2_3 | ShallowConvNet | 31 | 48.10% | 51.43% | +3.33 pp |
| s2 | s2_4 | ShallowConvNet | 34 | 75.71% | 50.00% | -25.71 pp |
| s1 | s1_1 | Support-CSP episodic | 40 | 60.00% | 55.00% | -5.00 pp |
| s1 | s1_2 | Support-CSP episodic | 40 | 40.00% | 37.50% | -2.50 pp |
| s1 | s1_3 | Support-CSP episodic | 40 | 40.00% | 32.50% | -7.50 pp |
| s1 | s1_4 | Support-CSP episodic | 40 | 32.50% | 40.00% | +7.50 pp |
| s1 | s1_5 | Support-CSP episodic | 39 | 61.79% | 59.29% | -2.50 pp |
| s2 | s2_1 | Support-CSP episodic | 35 | 60.00% | 42.86% | -17.14 pp |
| s2 | s2_2 | Support-CSP episodic | 36 | 56.07% | 55.71% | -0.36 pp |
| s2 | s2_3 | Support-CSP episodic | 31 | 51.90% | 55.24% | +3.33 pp |
| s2 | s2_4 | Support-CSP episodic | 34 | 69.52% | 60.95% | -8.57 pp |
檔案
quality summary · raw IIR results · ASR10 results · comparison CSV · ASR preparation code · evaluation code