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

Area of Science:

• Biomedical Engineering
• Neuroscience
• Rehabilitation Science

Background:

• Accurate temporal analysis of muscle activation is crucial for neurorobotics, patient rehabilitation, and understanding locomotion.
• Existing muscle activity detectors struggle with signal-to-noise ratio (SNR) and feature selection in surface electromyography (sEMG).

Purpose of the Study:

• To introduce and validate a novel Long Short-Term Memory (LSTM) recurrent neural network-based approach for detecting muscle activation intervals from sEMG signals.
• To compare the performance of the proposed LSTM-based muscle activity detector (LSTM-MAD) against established methods.

Main Methods:

• Simulated sEMG signals were used to compare LSTM-MAD against Teager-Kaiser Energy Operator (TKEO) and a statistical detector (Stat).
• The impact of varying Signal-to-Noise Ratio (SNR) on LSTM-MAD performance was assessed using simulated signals.
• The LSTM-MAD was validated on real sEMG data from healthy individuals and patients with orthopedic and neurological conditions during gait.

Main Results:

• LSTM-MAD demonstrated superior performance over TKEO and Stat, achieving high F1-scores (>0.91) and Jaccard similarity (>0.85) with low onset/offset bias (<6 ms).
• The algorithm effectively detects muscle activation directly from sEMG signals, eliminating the need for background noise or SNR estimation.
• LSTM-MAD's advantages were particularly pronounced in low to medium SNR conditions.

Conclusions:

• The LSTM-MAD is a powerful and accurate tool for recognizing muscle activity from sEMG signals, outperforming traditional methods.
• Its robust performance across simulated and real-world data, including pathological gait, highlights its potential in clinical and research applications.
• The LSTM-MAD offers a significant advancement in the temporal analysis of muscle activation.