Evaluating Convolution Neural Network Architecture for Neural Drive Decoding from High-Density Surface
Summary
This study compares 1D and 3D CNN models for decoding neural drive from HD-sEMG signals. 1D CNNs excel with larger windows, while 3D CNNs offer lower latency but higher computational cost.
Area of Science:
- Biomedical Engineering
- Neuroscience
- Machine Learning
Background:
- Convolutional Neural Network (CNN) models, including 1D and 3D CNNs, show promise in decoding neural drive from high-density surface electromyography (HD-sEMG) signals.
- The comparative performance of 1D versus 3D CNNs for neural drive decoding using identical datasets remains underexplored.
Purpose of the Study:
- To assess and compare the performance of 1D CNN and 3D CNN models in extracting neural drive as a cumulative spike train (CST).
- To investigate the influence of critical parameters, specifically window and step sizes, on the decoding accuracy of both CNN dimensionalities.
- To evaluate the computational cost associated with 1D and 3D CNN models for neural drive decoding.
Main Methods:
- Utilized an experimental HD-sEMG dataset from the gastrocnemius medialis muscle of three participants.
- Trained and validated 1D CNN and 3D CNN models to decode neural drive into a cumulative spike train (CST).
- Compared model performance using F1 score and correlation coefficient against the convolution kernel compensation (CKC) algorithm across various window and step sizes.
Main Results:
- 1D CNN achieved peak performance (F1 score: 0.84, correlation: 0.94) with larger window sizes (80-120 samples).
- 3D CNN reached peak performance (F1 score: 0.83, correlation: 0.92) with smaller window sizes (20-40 samples), suggesting lower latency.
- Both models showed performance degradation with increased step sizes, and 3D CNN exhibited significantly higher computational requirements (938G FLOPs vs. 60G FLOPs for 1D CNN).
Conclusions:
- Significant differences exist between 1D and 3D CNN architectures for HD-sEMG based neural drive decoding.
- Optimal parameter selection (window/step size) is crucial and differs between 1D and 3D CNNs for maximizing accuracy and minimizing latency.
- The findings provide guidance for selecting appropriate CNN models and parameters for precise and real-time neural drive decoding applications.
