EKFNet: edge-based Kalman filter network for real-time EEG signal denoising.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Deep Learning

Background:

• Deep learning methods for electroencephalogram (EEG) signal denoising are computationally intensive.
• Existing models face deployment challenges on edge-based portable or wearable (P/W) electronics.
• High computational complexity limits the application of advanced denoising techniques on resource-constrained devices.

Purpose of the Study:

• To propose an edge-based lightweight Kalman filter network (EKFNet) for efficient EEG signal denoising.
• To develop a deep learning model that eliminates the need for manual prior knowledge estimation.
• To enable the deployment of advanced signal denoising on portable and wearable devices.

Main Methods:

• Constructed a multi-scale feature fusion module to capture and implicitly compute prior knowledge.
• Designed an adaptive gain estimation module using LSTM and sequential channel attention for dynamic Kalman gain prediction.
• Implemented an optimization strategy with operator fusion and constant folding to reduce computational overhead and memory footprint.

Main Results:

• EKFNet reduced the sum of the square of the distances by at least 12% compared to state-of-the-art methods.
• Improved cosine similarity by at least 2.2% over existing denoising techniques.
• Model optimization achieved an approximate 3.3× reduction in inference time.

Conclusions:

• The EKFNet effectively integrates Kalman filtering with deep learning for EEG signal denoising.
• This approach overcomes parameter-setting challenges of traditional algorithms while minimizing computational and memory requirements.
• EKFNet offers a favorable balance between algorithmic performance and computing power for edge device applications.