用于运动单元活动追踪的动作电位卷积 神经网络分类方法仿真研究

To solve the problem of low accuracy of motor unit action potential (MUAP) waveform similarity classification method caused by MUAP waveform variation under real conditions, which reduces the tracking performance of the motor unit (MU), a MUAP classification method based on convolutional neural network (CNN) was proposed and tested using the simulated high-density electromyogram (EMG) recordings in this paper. The network was designed with the Kcras deep-learning framework and had a simple structure with the MUAP waveform called MU "fingerprints" as the input. The network was trained using the early-stop strategy in order to realize the classification and matching of different MUs, and thus longitudinal tracking of the discharge activities of given MUs. The results showed that the proposed CNN-based method could reach a classification accuracy of (89. 41% + 3. 72%) among (30. 23 + 5. 37) MUs on average even when only (54. 49 + 29. 28) MUAP waveform samples were available for individual MUs. The performance of the proposed method was significantly higher than the conventional MUAP waveform similarity classification method with an accuracy of (49. 66% + 6. 12%). These results demonstrated that the proposed method was more robust to the variation of MUAP waveforms. The proposed method has the potential to promote the studies in which longitudinal tracking of MU discharge activities is needed, and it is of great significance for multiple studies such as the exploration of the neural mechanisms underlying motor control.