Design and validation of a real-time, deep learning-based muscle activity detector for clinical and robotic applications

The precise determination of muscle activation timing is important in various fields such as motion analysis, biomechanical assessment in sports science, myoelectric control of prostheses and exoskeletons, diagnosis and follow-up of neuromuscular disorders through the evaluation of altered locomotion patterns and monitoring therapeutic interventions or rehabilitation programs. In spite of that, there is little consensus in literature on the methods for accurately detecting muscle activity onset/offset. The performance of traditional muscle activity detectors is significantly influenced by the signal-to-noise ratio (SNR) of surface electromyographic (sEMG) signals, the features used, and threshold settings. In recent years, machine learning-based methods have shown great potential in this task, encouraging further research in this direction. This thesis aims to validate a detector for muscle activation intervals of sEMG signals based on long short-term memory (LSTM) recurrent neural networks. Initially, the applicability of the LSTM-based muscle activity detector is evaluated through the use of a dataset of simulated sEMG signals, generated to replicate the typical temporal and frequency parameters of real signals. The performance of the LSTM approach is compared with that of two widely used classical approaches: the Teager-Kaiser Energy Operator and the Double Threshold statistical detector. Additionally, two other machine learning-based approaches were used for comparison: one based on 1D-CNN-LSTM layers and another using a Multi Layer Perceptron network trained with linear envelope, root mean square and continuous wavelet transform of sEMG signals. After comparing the proposed approach with the detectors aforementioned, the effect of SNR on the performance of the LSTM-based model was analyzed using simulated signals with different SNR values. The model's behavior was also examined with varying numbers of bursts, considering simulated signals with one or two activations. Finally, the LSTM approach was tested on real sEMG signals of 8 healthy, 6 orthopedic, and 6 neurological subjects. The results obtained demonstrate that the proposed algorithm overcomes the performances of the other implemented approaches, operating directly on the sEMG signals without the need to estimate noise or extract features. The LSTM approach showed superior performance compared to the other approaches, with higher accuracy values (99.2 ± 1.4% for simulated and 87.1 ± 12.4% for real signals), F1-score (98.1 ± 4.4 for simulated and 87.2 ± 11.3% for real signals), and Jaccard similarity index (96.5 ± 7.0% for simulated and 78.9 ± 15.3% for real signals). Moreover, the method proved robust to varying SNR values, maintaining good performance even with low SNR, with accuracy always above 95%. Furthermore, the model presents very low inference times (less than 45 ms), allowing for real-time use. The LSTM-based approach revealed excellent performance compared to others methods evaluated, showing that this method can be considered an effective tool for the accurate recognition of muscle activity, operating directly on raw signals in a user-independent manner, robust to noise and fast, allowing its use in clinical and robotic applications.