Temporal context informed myoelectric feature extraction uncovers frequency invariance in EMG-based gesture recognition

Human-machine interfaces based on Electromyographic (EMG) armbands are commonly utilized for gesture recognition using cross-sectional feature extraction (FE) schemes, those typically ignoring long-and short-term activity trends. This approach lacks the ability to capture the different movements’ context and often generates spurious decisions based on short windows of non-stationary EMG signals. The current study builds upon recent advances in spatial information extraction, as represented by our Phasor-based Multi-signal Waveform Length (MSWL) features, by encapsulating these features within a temporal context framework. Two streams of information are concatenated: a short-term memory component emphasizing partial correlation with previous analysis windows and a long-term component emphasizing the trend of the features belonging to the specific movements. The proposed method was evaluated on EMG datasets from: 1) twenty-two subjects using two simultaneously placed armbands with different sampling frequencies (200Hz MYO and 1000Hz 3DC), and 2) six transradial amputees following the NinaPro protocol and using the MYO armband for 17 movement classes. Our findings using the LibEMG toolbox show that context-aware EMG feature extraction achieves sampling frequency invariance in gesture pattern recognition. Despite literature favoring higher frequency armbands, our method delivers similar average accuracy (91%, p-value>0.05) across both high- and low frequency armbands. Notably, our method outperforms 58 FE methods from the LibEMG toolbox, this is further supported by the findings on the amputees’ database highlighting its efficacy in context-sensitive EMG pattern recognition.