Online EMG onset detection during sit-to-stand motion using the time derivative of sample entropy

Low-latency detection of muscle activation onset is essential for online movement analysis and for many assistive and rehabilitation applications, enabling early inference of user motion intent. Surface electromyography (sEMG) are promising signals for this purpose. However, prior work has highlighted that designing robust online sEMG onset detectors remains challenging due to substantial inter- and intra-subject variability, and many approaches are primarily evaluated offline, limiting their applicability in real-world settings. As a step toward addressing these challenges, we propose an online sEMG onset detector designed for demanding activities of daily living and capable of handling a wide range of signal characteristics. The method is evaluated using a pseudo-online framework providing near real-life testing conditions.

The proposed detector relies on the time derivative of fuzzy sample entropy (SampEn) computed on sliding sEMG windows. Validation used two datasets collected for this study: (i) experimental sEMG comprising 1676 samples from 335 sit-to-stand motions recorded from six healthy participants, and (ii) synthetic sEMG signals generated by a physiologically-based model with controlled baseline noise. Evaluation followed a pseudo-online procedure, emulating real-time windowing on streaming data and accounting for filter-induced latency, without prior knowledge of signal amplitude or baseline noise. Performance was compared with two established baseline detectors, respectively based on the Extended Teager-Kaiser Energy Operator (ETKEO) and the fuzzy SampEn, both reimplemented under identical pseudo-online constraints. Metrics included detection delay, missed and false detection rates, and working ranges across signal and noise conditions.

On experimental data, muscle activation was detected 36±70 ms before motion onset on average, with pre-motion detection in 77% of samples. Detection delay was reduced by at least 13% (up to 71%) compared with the online-compatible baselines, and missed detections were reduced from 4% to 1% compared to a non-derivative SampEn detector. Across both datasets, the proposed method exhibited the widest working range across signal amplitudes and baseline noise levels.

Our method enables earlier and more robust online sEMG onset detection for knee extensor activity under near real-life variability, supporting integration into EMG-based online systems. Both datasets and the real-time implementation are released open-source.