Long short-term memory (LSTM) recurrent neural network for muscle activity detection

Background: The accurate temporal analysis of muscle activation is of great interest in many research areas, spanning
from neurorobotic systems to the assessment of altered locomotion patterns in orthopedic and neurological
patients and the monitoring of their motor rehabilitation. The performance of the existing muscle activity detectors
is strongly affected by both the SNR of the surface electromyography (sEMG) signals and the set of features used to
detect the activation intervals. This work aims at introducing and validating a powerful approach to detect muscle
activation intervals from sEMG signals, based on long short-term memory (LSTM) recurrent neural networks.
 

Methods: First, the applicability of the proposed LSTM-based muscle activity detector (LSTM-MAD) is studied
through simulated sEMG signals, comparing the LSTM-MAD performance against other two widely used approaches,
i.e., the standard approach based on Teager–Kaiser Energy Operator (TKEO) and the traditional approach, used in
clinical gait analysis, based on a double-threshold statistical detector (Stat). Second, the effect of the Signal-to-Noise
Ratio (SNR) on the performance of the LSTM-MAD is assessed considering simulated signals with nine different SNR
values. Finally, the newly introduced approach is validated on real sEMG signals, acquired during both physiological
and pathological gait. Electromyography recordings from a total of 20 subjects (8 healthy individuals, 6 orthopedic
patients, and 6 neurological patients) were included in the analysis.

Results: The proposed algorithm overcomes the main limitations of the other tested approaches and it works
directly on sEMG signals, without the need for background-noise and SNR estimation (as in Stat). Results demonstrate
that LSTM-MAD outperforms the other approaches, revealing higher values of F1-score (F1-score > 0.91) and Jaccard
similarity index (Jaccard > 0.85), and lower values of onset/offset bias (average absolute bias < 6 ms), both on simulated
and real sEMG signals. Moreover, the advantages of using the LSTM-MAD algorithm are particularly evident for
signals featuring a low to medium SNR.

Conclusions: The presented approach LSTM-MAD revealed excellent performances against TKEO and Stat. The
validation carried out both on simulated and real signals, considering normal as well as pathological motor function
during locomotion, demonstrated that it can be considered a powerful tool in the accurate and effective recognition/
distinction of muscle activity from background noise in sEMG signals.