8 通道跨通道關聯自監督模型

依使用者要求中途停止：已保存 202 folds

資料流程：Raw → IIR 0.5–50 Hz → 200 Hz → 2 秒窗口 → 每通道窗口內 Z-score。
預訓練 30 位，共 33,584 windows；外部 SSL 驗證 4 位；正式微調 10 位。
三種預訓練均固定 36,000 updates；正式比較沿用相同 5-fold assignments。32 通道既有 SSL 65.97% ± 12.64% 僅列為參考。

資料與隔離

dataset subject_id split sex age kept_windows excluded_windows left_trials right_trials
pretrain sub16 train female 23.0 1164 0.0 - -
pretrain sub16 val female 23.0 125 0.0 - -
pretrain sub17 train male 24.0 1146 0.0 - -
pretrain sub17 val male 24.0 124 0.0 - -
pretrain sub18 train male 21.0 1147 0.0 - -
pretrain sub18 val male 21.0 123 0.0 - -
pretrain sub19 train female 23.0 1139 0.0 - -
pretrain sub19 val female 23.0 123 0.0 - -
pretrain sub20 train female 25.0 1170 0.0 - -
pretrain sub20 val female 25.0 126 0.0 - -
pretrain sub21 train male 20.0 1158 0.0 - -
pretrain sub21 val male 20.0 125 0.0 - -
pretrain sub23 train female 24.0 1154 0.0 - -
pretrain sub23 val female 24.0 125 0.0 - -
pretrain sub24 train male 59.0 1136 0.0 - -
pretrain sub24 val male 59.0 122 0.0 - -
pretrain sub25 train female 52.0 1129 0.0 - -
pretrain sub25 val female 52.0 121 0.0 - -
pretrain sub26 train male 79.0 565 0.0 - -
pretrain sub26 val male 79.0 59 0.0 - -
pretrain sub27 train female 74.0 1148 0.0 - -
pretrain sub27 val female 74.0 123 0.0 - -
pretrain sub28 train male 80.0 1207 0.0 - -
pretrain sub28 val male 80.0 130 0.0 - -
pretrain sub29 train female 71.0 1152 0.0 - -
pretrain sub29 val female 71.0 125 0.0 - -
pretrain sub30 train female 58.0 814 0.0 - -
pretrain sub30 val female 58.0 86 0.0 - -
pretrain sub31 train male 58.0 1129 0.0 - -
pretrain sub31 val male 58.0 122 0.0 - -
pretrain sub32 train female 68.0 1164 0.0 - -
pretrain sub32 val female 68.0 126 0.0 - -
pretrain sub33 train male 71.0 1150 0.0 - -
pretrain sub33 val male 71.0 124 0.0 - -
pretrain sub34 train male 65.0 1108 0.0 - -
pretrain sub34 val male 65.0 119 0.0 - -
pretrain sub35 train female 57.0 1152 0.0 - -
pretrain sub35 val female 57.0 124 0.0 - -
pretrain sub36 train male 66.0 1113 0.0 - -
pretrain sub36 val male 66.0 120 0.0 - -
pretrain sub38 train male 62.0 1119 0.0 - -
pretrain sub38 val male 62.0 121 0.0 - -
pretrain sub39 train female 61.0 1142 0.0 - -
pretrain sub39 val female 61.0 123 0.0 - -
pretrain sub40 train male 46.0 1146 0.0 - -
pretrain sub40 val male 46.0 123 0.0 - -
pretrain sub41 train male 64.0 1152 0.0 - -
pretrain sub41 val male 64.0 124 0.0 - -
pretrain sub42 train female 64.0 1125 0.0 - -
pretrain sub42 val female 64.0 121 0.0 - -
pretrain sub43 train female 59.0 1170 0.0 - -
pretrain sub43 val female 59.0 126 0.0 - -
pretrain sub44 train female 61.0 1166 0.0 - -
pretrain sub44 val female 61.0 126 0.0 - -
pretrain sub47 train female 29.0 1116 0.0 - -
pretrain sub47 val female 29.0 120 0.0 - -
pretrain sub48 train male 31.0 1140 0.0 - -
pretrain sub48 val male 31.0 123 0.0 - -
pretrain sub49 train male 21.0 1263 0.0 - -
pretrain sub49 val male 21.0 136 0.0 - -
external sub22 external male 23.0 1263 0.0 - -
external sub37 external female 65.0 1256 0.0 - -
external sub45 external male 66.0 1260 0.0 - -
external sub46 external female 59.0 1277 0.0 - -
labeled sub1 all - - 203 - 100.0 103.0
labeled sub2 all - - 203 - 102.0 101.0
labeled sub3 all - - 200 - 100.0 100.0
labeled sub9 all - - 201 - 100.0 101.0
labeled sub10 all - - 200 - 100.0 100.0
labeled sub11 all - - 200 - 100.0 100.0
labeled sub12 all - - 201 - 100.0 101.0
labeled sub13 all - - 200 - 100.0 100.0
labeled sub14 all - - 200 - 100.0 100.0
labeled sub15 all - - 201 - 100.0 101.0

架構驗證

Smoke test passed；input [2, 8, 400]；tokens [2, 8, 20, 128]；mask [80, 80]；relation [2, 28, 4]；encoder 1,594,496 params。

等計算量預訓練消融

Pretraining ablation

XChannel 遮罩重建

XChannel reconstruction

外部跨通道關聯預測

Relation diagnostics

左右腳個人化分類

Subject accuraciesGroup accuracies

結果判斷：依使用者要求中途停止：目前只有 202/350 筆初篩 folds，公平七模式比較僅納入 5/10 位完整受試者，且未執行 200-fold 多 seed confirmation。下列 mean、CI 與 p 值均為暫時描述，不能作正式架構優劣結論。

中途結果解讀

完整 5 位受試者中，Recon FT 為 63.00%，XChannel FT 為 58.64%，Latent FT 為 56.93%，scratch 為 52.35%。XChannel 相對 scratch 暫高 6.29 個百分點，但相對 Recon 為 -4.36 點，相對 Latent 只有 1.70 點。外部 relation prediction 的平均 r：broadband 0.776、Mu 0.376、low beta 0.365、high beta 0.376。這表示跨通道關聯目標確實能預測部分未見資料的 relation pattern，但目前沒有轉成優於單純重建的左右腳分類；較合理的結論是 transfer benefit 不明確，而不是已證明 XChannel 有效。

初篩群組摘要

model n_subjects mean_accuracy std_accuracy
Recon FT 5 63.00% 7.56%
XChannel FT 5 58.64% 8.67%
Latent FT 5 56.93% 8.39%
Recon LP 5 55.63% 5.59%
XChannel LP 5 55.05% 9.59%
Scratch 5 52.35% 3.85%
Latent LP 5 51.95% 5.47%

配對檢定

analysis analysis_status first_mode second_mode n_subjects first_mean_accuracy second_mean_accuracy mean_paired_difference positive_subject_fraction bootstrap_ci95_low bootstrap_ci95_high wilcoxon_statistic wilcoxon_p_value holm_p_value
screen_seed_20260716 provisional_partial recon_finetune scratch 5 63.00% 52.35% 10.65% 100.00% 5.99% 15.65% 0.0 0.0625 0.3750
screen_seed_20260716 provisional_partial latent_finetune scratch 5 56.93% 52.35% 4.59% 80.00% -0.21% 9.70% 3.0 0.3125 0.9375
screen_seed_20260716 provisional_partial xchannel_finetune scratch 5 58.64% 52.35% 6.29% 100.00% 2.59% 11.70% 0.0 0.0625 0.3750
screen_seed_20260716 provisional_partial latent_finetune recon_finetune 5 56.93% 63.00% -6.06% 20.00% -13.98% 0.49% 2.0 0.1875 0.7500
screen_seed_20260716 provisional_partial xchannel_finetune latent_finetune 5 58.64% 56.93% 1.70% 80.00% -2.40% 5.31% 4.0 0.4375 0.9375
screen_seed_20260716 provisional_partial xchannel_linear scratch 5 55.05% 52.35% 2.70% 80.00% -1.98% 7.70% 3.0 0.3125 0.9375

輸出檔案

cuda_run.logdata_manifest.csvfold_assignment_validation.jsonfold_results.csvgroup_summary.csvlatent_20260716_external_validation.csvlatent_20260716_pretrain_history.csvpaired_comparisons.csvpartial_run_status.jsonpreprocessing_validation.jsonrecon_20260716_external_validation.csvrecon_20260716_pretrain_history.csvrun_config.jsonscreen_complete_subjects_fold_results.csvscreen_supervisor.logseed_stability.csvsmoke_test.jsonsubject_summary.csvxchannel_20260716_external_validation.csvxchannel_20260716_pretrain_history.csvxchannel_20260716_relation_diagnostics.csv