Optimizing the performance of convolutional neural network for enhanced gesture recognition using sEMG.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Deep neural networks (DNNs) show promise for myoelectric control (MEC) but face real-time application delays due to optimization.
• Optimizing DNN hyperparameters is crucial for enhancing MEC system performance and reducing latency.

Purpose of the Study:

• To investigate optimal configurations for convolutional neural network (CNN)-based MEC systems.
• To propose an effective data segmentation technique and a generalized set of hyperparameters for DNN-based MEC.

Main Methods:

• Compared disjoint and overlap data segmentation strategies with varying segment and overlap sizes.
• Employed Bayesian optimization to abstract and solve the DNN hyperparameter optimization problem.
• Collected surface electromyography (sEMG) data from 20 healthy individuals performing 10 grasping movements.

Main Results:

• The overlap segmentation technique, with an optimal segment size of 200 ms and 80% overlap, significantly outperformed disjoint segmentation (p < 0.05).
• Bayesian optimization achieved a mean classification error rate (CER) of 0.08 ± 0.03, outperforming manual, grid, and random search methods.
• A generalized CNN architecture with optimal hyperparameters yielded an overall CER of 0.09 ± 0.03 when tested across all subjects.

Conclusions:

• The overlap segmentation technique is superior for MEC applications.
• Bayesian optimization provides an effective strategy for tuning CNN hyperparameters in MEC systems.
• The proposed generalized CNN architecture and optimized hyperparameters offer practical improvements for prosthetic control and human-computer interfaces.