8通道七方法統一比較

正常人、患者與中榮 | Zero-phase SOS IIR | 無 ASR / ICA / EA

實驗範圍

共同通道
Fp1、Fp2、Fz、C3、C4、Pz、O1、O2
前處理
1-40 Hz fourth-order zero-phase SOS IIR，200 Hz
窗口
正常 / 患者 0-2 s；中榮 2-4 s
評估
單一受試者 / session 5-fold，inner validation

正常人與患者各跑 all-trial 和固定 20+20；中榮部分 session 少於每類 20 筆，因此只跑 all-trial。全部方法共用逐 fold 完全相同的 train、validation、test trial。

總覽

七方法 group accuracy

正常人：all trials：Support-CSP + Residual-gated 76.81%；正常人：20 Left + 20 Right：Support-CSP + ShallowConvNet 62.00%；中風患者：all trials：Support-CSP + Residual-gated 82.95%；中風患者：20 Left + 20 Right：Support-CSP + Residual-gated 75.50%；中榮：各 session all trials：Support-CSP + Residual-gated 75.24%

相對 ShallowConvNet

paired difference

95% CI 以受試者層級配對 bootstrap；中榮 session 並非獨立樣本，只作描述。

單一受試者 / session

unit accuracy heatmaps

正常人：all trials

最高平均為 Support-CSP + Residual-gated：76.81%。

方法Accuracy mean +/- STDBalanced accuracyMacro-F1
CSP + shrinkage LDA57.36% +/- 8.09%57.36%0.570
FDCSP + shrinkage LDA58.89% +/- 5.26%58.87%0.585
CSP sliding features + RF61.22% +/- 7.82%61.22%0.610
EEGNet64.49% +/- 13.07%64.48%0.638
ShallowConvNet71.39% +/- 18.23%71.40%0.696
Support-CSP + ShallowConvNet73.57% +/- 14.86%73.57%0.732
Support-CSP + Residual-gated76.81% +/- 13.77%76.81%0.767

相對 ShallowConvNet

candidate_label相對 Shallow95% bootstrap CI改善單位
CSP + shrinkage LDA-14.03 pp[-21.15, -6.57] pp2/10
FDCSP + shrinkage LDA-12.50 pp[-21.09, -4.02] pp3/10
CSP sliding features + RF-10.17 pp[-18.22, -2.12] pp4/10
EEGNet-6.90 pp[-12.72, -0.95] pp2/10
Support-CSP + ShallowConvNet+2.18 pp[-0.55, +5.55] pp6/10
Support-CSP + Residual-gated+5.42 pp[+2.50, +9.34] pp9/10

STD 為受試者間標準差；差異和 CI 先在每位受試者內平均 5 folds 後配對。

正常人：20 Left + 20 Right

最高平均為 Support-CSP + ShallowConvNet：62.00%。

方法Accuracy mean +/- STDBalanced accuracyMacro-F1
CSP + shrinkage LDA55.25% +/- 9.09%55.25%0.531
FDCSP + shrinkage LDA56.75% +/- 13.18%56.75%0.547
CSP sliding features + RF58.00% +/- 11.29%58.00%0.568
EEGNet54.25% +/- 7.08%54.25%0.510
ShallowConvNet59.25% +/- 12.64%59.25%0.561
Support-CSP + ShallowConvNet62.00% +/- 12.18%62.00%0.606
Support-CSP + Residual-gated60.25% +/- 12.88%60.25%0.584

相對 ShallowConvNet

candidate_label相對 Shallow95% bootstrap CI改善單位
CSP + shrinkage LDA-4.00 pp[-12.00, +3.75] pp5/10
FDCSP + shrinkage LDA-2.50 pp[-11.25, +5.75] pp5/10
CSP sliding features + RF-1.25 pp[-12.50, +9.25] pp5/10
EEGNet-5.00 pp[-12.75, +2.25] pp4/10
Support-CSP + ShallowConvNet+2.75 pp[-3.00, +9.50] pp4/10
Support-CSP + Residual-gated+1.00 pp[-5.75, +7.25] pp6/10

STD 為受試者間標準差；差異和 CI 先在每位受試者內平均 5 folds 後配對。

中風患者：all trials

最高平均為 Support-CSP + Residual-gated：82.95%。

方法Accuracy mean +/- STDBalanced accuracyMacro-F1
CSP + shrinkage LDA68.03% +/- 15.04%67.79%0.676
FDCSP + shrinkage LDA68.83% +/- 13.40%68.72%0.685
CSP sliding features + RF72.62% +/- 13.52%72.35%0.722
EEGNet76.20% +/- 19.22%76.31%0.760
ShallowConvNet80.98% +/- 17.48%80.93%0.808
Support-CSP + ShallowConvNet82.00% +/- 15.53%81.93%0.819
Support-CSP + Residual-gated82.95% +/- 14.67%82.88%0.828

相對 ShallowConvNet

candidate_label相對 Shallow95% bootstrap CI改善單位
CSP + shrinkage LDA-12.95 pp[-26.63, -0.30] pp1/5
FDCSP + shrinkage LDA-12.15 pp[-23.22, -0.20] pp1/5
CSP sliding features + RF-8.35 pp[-17.79, +1.01] pp1/5
EEGNet-4.78 pp[-8.13, -1.43] pp0/5
Support-CSP + ShallowConvNet+1.03 pp[-0.55, +2.93] pp3/5
Support-CSP + Residual-gated+1.98 pp[-1.43, +5.67] pp2/5

STD 為受試者間標準差；差異和 CI 先在每位受試者內平均 5 folds 後配對。

中風患者：20 Left + 20 Right

最高平均為 Support-CSP + Residual-gated：75.50%。

方法Accuracy mean +/- STDBalanced accuracyMacro-F1
CSP + shrinkage LDA64.00% +/- 10.55%64.00%0.618
FDCSP + shrinkage LDA66.50% +/- 6.75%66.50%0.654
CSP sliding features + RF69.00% +/- 13.06%69.00%0.677
EEGNet63.50% +/- 16.45%63.50%0.617
ShallowConvNet71.50% +/- 11.81%71.50%0.693
Support-CSP + ShallowConvNet71.00% +/- 15.06%71.00%0.701
Support-CSP + Residual-gated75.50% +/- 11.10%75.50%0.749

相對 ShallowConvNet

candidate_label相對 Shallow95% bootstrap CI改善單位
CSP + shrinkage LDA-7.50 pp[-15.50, -2.00] pp0/5
FDCSP + shrinkage LDA-5.00 pp[-12.50, +2.50] pp1/5
CSP sliding features + RF-2.50 pp[-9.50, +5.50] pp1/5
EEGNet-8.00 pp[-20.50, +3.50] pp1/5
Support-CSP + ShallowConvNet-0.50 pp[-6.00, +4.50] pp3/5
Support-CSP + Residual-gated+4.00 pp[-1.00, +9.00] pp3/5

STD 為受試者間標準差；差異和 CI 先在每位受試者內平均 5 folds 後配對。

中榮：各 session all trials

最高平均為 Support-CSP + Residual-gated：75.24%。

方法Accuracy mean +/- STDBalanced accuracyMacro-F1
CSP + shrinkage LDA68.17% +/- 20.31%68.43%0.666
FDCSP + shrinkage LDA67.61% +/- 15.69%67.78%0.664
CSP sliding features + RF73.44% +/- 20.88%73.52%0.720
EEGNet60.98% +/- 16.13%61.11%0.575
ShallowConvNet69.55% +/- 25.86%69.63%0.675
Support-CSP + ShallowConvNet72.51% +/- 23.02%72.50%0.715
Support-CSP + Residual-gated75.24% +/- 21.65%75.46%0.746

相對 ShallowConvNet

candidate_label相對 Shallow95% bootstrap CI改善單位
CSP + shrinkage LDA-1.38 pp[-11.23, +7.13] pp5/9
FDCSP + shrinkage LDA-1.94 pp[-11.52, +7.92] pp3/9
CSP sliding features + RF+3.89 pp[-3.37, +10.73] pp5/9
EEGNet-8.57 pp[-20.82, +3.04] pp4/9
Support-CSP + ShallowConvNet+2.96 pp[-0.56, +7.13] pp4/9
Support-CSP + Residual-gated+5.69 pp[-0.46, +12.94] pp5/9

中榮表格以 session 描述；同一受試者多天 session 並非獨立樣本，CI 僅供探索。

中榮多日資料層級

s1 與 s2 才是兩位受試者；先在各 session 內做 5-fold mean，再於同一人跨天平均。兩人不足以做可靠群體顯著性推論。

受試者跨日彙整

中榮受試者方法SessionsAccuracyBalanced accuracy
s1CSP + shrinkage LDA583.21%83.17%
s1FDCSP + shrinkage LDA579.29%79.17%
s1CSP sliding features + RF589.43%89.33%
s1EEGNet568.36%68.33%
s1ShallowConvNet591.00%91.00%
s1Support-CSP + ShallowConvNet591.00%91.00%
s1Support-CSP + Residual-gated591.50%91.50%
s2CSP + shrinkage LDA449.38%50.00%
s2FDCSP + shrinkage LDA453.01%53.54%
s2CSP sliding features + RF453.45%53.75%
s2EEGNet451.76%52.08%
s2ShallowConvNet442.74%42.92%
s2Support-CSP + ShallowConvNet449.40%49.37%
s2Support-CSP + Residual-gated454.91%55.42%

單一 session

受試者 / session方法Accuracy
s1_1CSP + shrinkage LDA87.50%
s1_1FDCSP + shrinkage LDA85.00%
s1_1CSP sliding features + RF90.00%
s1_1EEGNet50.00%
s1_1ShallowConvNet85.00%
s1_1Support-CSP + ShallowConvNet82.50%
s1_1Support-CSP + Residual-gated85.00%
s1_2CSP + shrinkage LDA85.00%
s1_2FDCSP + shrinkage LDA82.50%
s1_2CSP sliding features + RF90.00%
s1_2EEGNet55.00%
s1_2ShallowConvNet90.00%
s1_2Support-CSP + ShallowConvNet90.00%
s1_2Support-CSP + Residual-gated82.50%
s1_3CSP + shrinkage LDA100.00%
s1_3FDCSP + shrinkage LDA80.00%
s1_3CSP sliding features + RF100.00%
s1_3EEGNet60.00%
s1_3ShallowConvNet87.50%
s1_3Support-CSP + ShallowConvNet87.50%
s1_3Support-CSP + Residual-gated95.00%
s1_4CSP + shrinkage LDA75.00%
s1_4FDCSP + shrinkage LDA72.50%
s1_4CSP sliding features + RF80.00%
s1_4EEGNet90.00%
s1_4ShallowConvNet92.50%
s1_4Support-CSP + ShallowConvNet95.00%
s1_4Support-CSP + Residual-gated95.00%
s1_5CSP + shrinkage LDA68.57%
s1_5FDCSP + shrinkage LDA76.43%
s1_5CSP sliding features + RF87.14%
s1_5EEGNet86.79%
s1_5ShallowConvNet100.00%
s1_5Support-CSP + ShallowConvNet100.00%
s1_5Support-CSP + Residual-gated100.00%
s2_1CSP + shrinkage LDA57.14%
s2_1FDCSP + shrinkage LDA60.00%
s2_1CSP sliding features + RF37.14%
s2_1EEGNet45.71%
s2_1ShallowConvNet40.00%
s2_1Support-CSP + ShallowConvNet37.14%
s2_1Support-CSP + Residual-gated37.14%
s2_2CSP + shrinkage LDA44.64%
s2_2FDCSP + shrinkage LDA63.93%
s2_2CSP sliding features + RF55.71%
s2_2EEGNet50.36%
s2_2ShallowConvNet41.43%
s2_2Support-CSP + ShallowConvNet55.71%
s2_2Support-CSP + Residual-gated58.21%
s2_3CSP + shrinkage LDA54.29%
s2_3FDCSP + shrinkage LDA41.90%
s2_3CSP sliding features + RF64.76%
s2_3EEGNet57.62%
s2_3ShallowConvNet48.10%
s2_3Support-CSP + ShallowConvNet50.95%
s2_3Support-CSP + Residual-gated54.29%
s2_4CSP + shrinkage LDA41.43%
s2_4FDCSP + shrinkage LDA46.19%
s2_4CSP sliding features + RF56.19%
s2_4EEGNet53.33%
s2_4ShallowConvNet41.43%
s2_4Support-CSP + ShallowConvNet53.81%
s2_4Support-CSP + Residual-gated70.00%

方法定義

CSP + LDA：8-30 Hz shrinkage CSP，inner validation 選每端 1/2/3 個極端 eigenvalue filters 與 covariance shrinkage 0.1/0.3，分類器為 lsqr + shrinkage=auto LDA。

FDCSP：由 Mu、low beta、high beta 的複數 FFT 建立 integrated cross-spectral covariance，取實部後做 regularized CSP；這是本報告採用的明確 FDCSP 定義。

CSP sliding features + RF：CSP 投影後保留時序，以 0.5 s window / 0.25 s hop 萃取 log-variance、Hjorth mobility、trace-normalized off-diagonal covariance，再以 300-tree、depth 8、balanced bootstrap Random Forest 分類。

Support-CSP：34 位未納入正式評估的正常人做 episodic source training；每個 target fold 的 CSP、feature standardization 與 prototype 僅由 target train trials 建立。Residual-gated 以 CSP feature 有界調節 residual waveform feature。

完整性稽核

{
  "expected_fold_rows": 1365,
  "actual_fold_rows": 1365,
  "duplicate_keys": 0,
  "max_split_hash_variants_across_methods": 1,
  "train_val_test_overlap_count": 0,
  "finite_metrics": true,
  "complete": true
}

fold_results.csv · unit_summary.csv · group_summary.csv · paired_vs_shallow.csv · classical_parameter_selection.csv · fold_assignments.csv · audit.json · run_config.json