8通道 Shallow Transformation Recognition SSL
結果判斷：Transformation SSL 未超過 matched scratch；這表示去除重建並不足以解決跨受試者預訓練的負遷移。
Group accuracy
| regime | model | mean_accuracy | std_accuracy | mean_balanced_accuracy | mean_macro_f1 |
|---|---|---|---|---|---|
| 20+20 | Shallow scratch | 61.21% | 11.13% | 61.19% | 56.36% |
| 20+20 | Transformation SSL | 56.48% | 7.60% | 56.44% | 55.06% |
| all | Shallow scratch | 73.24% | 16.53% | 73.23% | 72.45% |
| all | Transformation SSL | 70.02% | 15.25% | 70.02% | 69.61% |
本實驗的 matched scratch 依計畫使用 calibration 內部切分選 epoch，再以完整 calibration refit；因此其 20+20 數值不應與先前固定 56 epochs 的 Shallow 結果直接混用。SSL 效果以同表 matched scratch 為準。
20+20 單一受試者
Paired statistics
| regime | comparison | mean_delta_pp | bootstrap_ci_low_pp | bootstrap_ci_high_pp | wilcoxon_p | ssl_better_subjects |
|---|---|---|---|---|---|---|
| 20+20 | Transformation SSL vs Shallow scratch | -4.7248 | -8.4358 | -1.1666 | 0.0273 | 1 |
| all | Transformation SSL vs Shallow scratch | -3.2191 | -5.4842 | -1.2122 | 0.0273 | 3 |
Transformation pretraining
外部四位受試者在訓練強度範圍的 balanced accuracy 為 79.73%，但擾動改為較輕微後降至 63.98%。這表示 encoder 能辨識明顯轉換，卻未形成足以轉移至左右腳分類的 transformation-robust representation。
| intensity | class_name | precision | recall | f1 | overall_balanced_accuracy | overall_macro_f1 |
|---|---|---|---|---|---|---|
| training_range | original | 58.69% | 62.50% | 60.54% | 79.73% | 80.06% |
| training_range | gaussian_noise | 98.58% | 83.17% | 90.22% | 79.73% | 80.06% |
| training_range | channel_scaling | 77.26% | 75.96% | 76.61% | 79.73% | 80.06% |
| training_range | time_reversal | 69.57% | 77.75% | 73.44% | 79.73% | 80.06% |
| training_range | segment_permutation | 99.75% | 99.25% | 99.50% | 79.73% | 80.06% |
| mild_holdout | original | 39.72% | 60.34% | 47.90% | 63.98% | 64.16% |
| mild_holdout | gaussian_noise | 98.33% | 42.55% | 59.40% | 63.98% | 64.16% |
| mild_holdout | channel_scaling | 57.23% | 43.75% | 49.59% | 63.98% | 64.16% |
| mild_holdout | time_reversal | 57.09% | 74.50% | 64.64% | 63.98% | 64.16% |
| mild_holdout | segment_permutation | 99.75% | 98.75% | 99.25% | 63.98% | 64.16% |
fold results · trial manifest · group summary · subject summary · paired statistics · pretraining history · run config · code