Intra- and inter-channel deep convolutional neural network with dynamic label smoothing for multichannel biosignal...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Multichannel biosignal processing is crucial for healthcare and human-machine interaction.
• Deep convolutional neural networks (CNNs) show promise but face challenges in feature extraction, model complexity, and generalization.
• Existing methods often require extensive preprocessing or have high parameter counts.

Purpose of the Study:

• To develop an efficient deep learning model for multichannel biosignal processing.
• To address limitations in spatial-temporal feature extraction, performance-complexity trade-offs, and model generalization.
• To improve the applicability of biosignal processing in real-world scenarios.

Main Methods:

• Proposed a 1D-based deep intra and inter channel (I2C) convolutional neural network.
• Introduced an I2C convolutional block to replace standard layers for effective feature extraction with fewer parameters.
• Integrated a branch model with dynamic label smoothing to enhance generalization ability and domain adaptation.

Main Results:

• The proposed I2C network demonstrated significant competitive advantages in accuracy.
• Achieved a favorable trade-off between performance and model complexity.
• Showcased improved generalization ability across different biosignal datasets (ISRUC-S3, HEF, Ninapro-DB1).

Conclusions:

• The 1D I2C CNN offers an effective solution for multichannel biosignal processing.
• The method provides a balance of high accuracy, computational efficiency, and robust generalization.
• This approach holds potential for advancing healthcare and human-machine interaction applications.