DNN Filter Bank Improves 1-Max Pooling CNN for Single-Channel EEG Automatic Sleep Stage Classification

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Neuroscience

Background:

• Automatic sleep stage classification is crucial for diagnosing sleep disorders.
• Existing deep learning models often require complex architectures.
• Time-frequency representations of electroencephalogram (EEG) signals are effective for sleep analysis.

Purpose of the Study:

• To develop an efficient convolutional neural network (CNN) for automatic sleep stage classification.
• To improve feature extraction from time-frequency image representations of EEG data.
• To achieve state-of-the-art performance on a benchmark sleep dataset.

Main Methods:

• A simplified CNN architecture with convolutional kernels of varying sizes for multi-resolution feature learning.
• Implementation of a 1-max pooling strategy to enhance shift-invariance in EEG signal processing.
• Development of a deep neural network (DNN) to learn a discriminative frequency-domain filter bank for data preprocessing.

Main Results:

• The proposed 1-max pooling CNN demonstrates performance comparable to deeper architectures on the Sleep-EDF dataset.
• Preprocessing time-frequency features with the learned filter bank significantly boosts classification accuracy.
• The combined approach establishes a new state-of-the-art performance for automatic sleep stage classification on this dataset.

Conclusions:

• An efficient CNN with multi-resolution capabilities and 1-max pooling offers a competitive alternative for sleep stage classification.
• A learned filter bank significantly enhances the discriminative power of time-frequency features for EEG analysis.
• This methodology sets a new benchmark for accuracy in automatic sleep stage classification.