Few-Shot Personalization of Lower-Limb EMG Waveform Prediction Using Task–Phase Templates: A Subject-Disjoint Benchmark Study

Lower-limb EMG waveform prediction is difficult to interpret because a model may appear accurate by reproducing average task- and phase-dependent activation patterns rather than learning subject-specific neuromuscular information. This secondary analysis used the processed Gait120 dataset, including 120 healthy adult male participants, 6882 paired kinematic–EMG cycles, seven movement tasks, and 12 processed EMG envelope channels. Under five-fold subject-disjoint validation, a zero-shot task–phase template estimated from training subjects achieved RMSE × 103 = 0.0188, MAE × 103 = 0.0088, R2 = 0.454, and NRMSE_SD = 0.727 in held-out subjects. For k-shot personalization, 1, 3, 5, or 10 EMG-labelled calibration cycles from each held-out subject were used to estimate task-specific shrinkage residual corrections, while separate cycles from the same subject were reserved for evaluation. Ten-shot personalization improved performance to RMSE × 103 = 0.0138, MAE × 103 = 0.0062, R2 = 0.708, and NRMSE_SD = 0.535. Task-specific correction outperformed global residual correction, whereas a fixed-parameter linear Ridge residual comparator did not improve zero-shot performance beyond the template. These findings support task–phase templates as demanding reference baselines for processed EMG envelope prediction, but only under offline conditions with known task labels, known normalized phase, and within-subject EMG-labelled calibration.