8通道 KMeans 硬路由 + ShallowConvNet Experts

KMeans 僅使用 outer-train 的 event 前 -1~0 秒 baseline；task 0~2 秒送入標準 ShallowConvNet。比較 global、每群 scratch 與 global-init 後群內微調三種模式。

主要結果

設定	模型	Accuracy mean ± STD	Balanced accuracy	Macro-F1
All trials	Global Shallow	71.95% ± 17.58%	71.94%	0.712
All trials	Hard-K2 scratch experts	67.71% ± 15.67%	67.68%	0.668
All trials	Hard-K2 global-init experts	71.70% ± 17.88%	71.69%	0.713
20 Left + 20 Right	Global Shallow	59.75% ± 10.58%	59.75%	0.562
20 Left + 20 Right	Hard-K2 scratch experts	56.30% ± 5.84%	56.30%	0.527
20 Left + 20 Right	Hard-K2 global-init experts	60.40% ± 11.48%	60.40%	0.576

判斷：Hard routing did not produce a stable improvement over the matched global Shallow baseline.

受試者配對結果

設定	比較	平均差	Bootstrap 95% CI	Wilcoxon p	改善受試者
All trials	hard_k2_scratch - global_shallow	-4.23 pp	[-6.18, -2.35] pp	0.0039	1/10
All trials	hard_k2_global_init - global_shallow	-0.25 pp	[-0.79, +0.22] pp	0.8457	6/10
20 Left + 20 Right	hard_k2_scratch - global_shallow	-3.45 pp	[-6.65, -0.25] pp	0.1055	2/10
20 Left + 20 Right	hard_k2_global_init - global_shallow	+0.65 pp	[-0.95, +2.00] pp	0.3047	6/10

分群與 fallback

不平衡群完全保留。只有 train cluster 缺類或 validation cluster 為空時，該 expert 才回退至同 fold 的 global Shallow。

設定	Fallback clusters	群內 train trial 中位數	Silhouette 中位數
20 Left + 20 Right	73/500	12.5	0.096
All trials	3/300	64.5	0.137

防止資料洩漏

StandardScaler → PCA → KMeans 均只 fit outer-train baseline；validation/test 只呼叫 transform/predict。KMeans 不讀 task waveform 或左右腳標籤。所有模型共用既有 folds、trial selections、Shallow 架構、task normalization 與 augmentation。

檔案

fold results · cluster diagnostics · expert results · subject summary · paired comparisons · 模型與訓練程式