8通道 Filter-Bank Spatial ShallowNet

在同一組 subject-dependent folds 上，將 Mu (8–13)、low beta (13–20)、high beta (20–30 Hz) 保留為可微分的時空 log-variance 分支；不是先把頻帶平均成 PSD 特徵。

資料與候選選擇

Raw continuous 1–40 Hz SOS zero-phase IIR → 200 Hz → event[-1,+2] → task window 逐通道 trial Z-score。screen 僅讀 inner validation：選定 filterbank_spatial、0-2 s。

模型	window	Validation BACC	folds
filterbank_only	0-2 s	67.29% ± 16.51%	100
filterbank_only	0.2-2 s	68.04% ± 15.11%	100
filterbank_spatial	0-2 s	70.75% ± 16.78%	100
filterbank_spatial	0.2-2 s	66.08% ± 17.23%	100
global_shallow	0-2 s	74.79% ± 16.19%	100

Outer-test 結果

設定	模型	Accuracy mean ± STD	BACC	Macro-F1
All trials	Global ShallowConvNet	72.47% ± 17.25%	72.44%	0.709
All trials	FilterBank Spatial ShallowNet	68.71% ± 14.56%	68.69%	0.676
20 Left + 20 Right	Global ShallowConvNet	59.40% ± 10.69%	59.40%	0.557
20 Left + 20 Right	FilterBank Spatial ShallowNet	55.80% ± 5.96%	55.80%	0.501

判斷：Filter-bank fusion does not yet provide a stable few-shot gain; retain it as a physiologically interpretable comparator rather than a replacement.

配對比較：FilterBank − Shallow

設定	平均差	Bootstrap 95% CI	Wilcoxon p	改善受試者
20x20	-3.60 pp	[-6.90, -0.25] pp	0.0684	3/10
all	-3.76 pp	[-6.05, -0.93] pp	0.0371	1/10

ME embedding routing 診斷

每個 outer fold 的 Global Shallow ME embedding 只在 train trials 做 StandardScaler、PCA、K=2 KMeans；本頁只判斷 task-state 路由是否值得進一步訓練 experts，沒有用這些群集來提高本次分類成績。

設定	Silhouette median	Bootstrap ARI median	Label NMI median	每 cluster/class 最少 trials median
20x20	0.089	0.116	0.037	2.0
all	0.111	0.158	0.031	13.0

下載

fold results · subject summary · paired statistics · ME embedding diagnostics · model and evaluation code