Answers are grounded in this paper and cite their sources.
Surface electromyography (sEMG) supports personalized movement recognition, but feature utility must be interpreted together with preprocessing conditions. We evaluated a task-specific integration of a smooth fast Fourier transform (FFT) subband bank, nearest-neighbor entropy (NNE), time-domain (TD) descriptors, and participant-specific extreme gradient boosting (XGBoost). Nine-channel recordings from 40 Shenzhen Institute of Advanced Technology Lower Limb Motion Dataset (SIAT-LLMD) participants were evaluated for 16-class recognition using repetition-grouped five-fold cross-validation. Under complete-record amplitude normalization, the 216-dimensional fusion achieved 95.63% accuracy and 94.54% macro-F1, exceeding TD-XGB by 1.043 percentage points; random and repeated-feature expansion did not reproduce this gain. With the classifier configurations held fixed, macro-F1 was 94.85% with record-specific references fitted to all eligible samples, 95.58% with record-specific training-only references, and 75.31% with shared participant-level training references. The record-specific controls require movement-associated recording identity and do not establish prospective unknown-action performance. This gain was accompanied by median single-thread CPU computation of 34.210 ms per conditioned window for Fusion-XGB, compared with 5.392 ms for TD-XGB (approximately sixfold). Across the evaluated amplitude-reference conditions, Fusion-XGB retained a macro-F1 improvement over TD-XGB, while absolute recognition performance varied with the reference definition.
No takes yet. Share an insight, caveat, or question.
Chang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: