Neural network architectures and normalization techniques for automated sleep stage classification using rodent EEG...

Area of Science:

• Neuroscience
• Computational Biology
• Sleep Research

Background:

• Manual sleep stage classification in animal models is time-consuming and inconsistent.
• Artificial neural networks (ANNs) offer automated solutions for sleep staging.
• Limited cross-model comparisons exist for ANNs in animal sleep research.

Purpose of the Study:

• To systematically evaluate and compare three neural network architectures for automated rodent sleep stage classification.
• To assess the impact of different data preprocessing techniques on model performance.
• To identify the most suitable models and strategies for accurate animal sleep staging.

Main Methods:

• Evaluation of 1-dimensional convolutional neural network (1D-CNN), AccuSleep (2D-CNN), and DeepSleepNet (CNN-LSTM) using rodent EEG and EMG data.
• Comparison of performance under within-subject and cross-subject validation.
• Assessment of raw input, z-scoring, and mixture z-scoring normalization methods.

Main Results:

• 1D-CNN and DeepSleepNet outperformed AccuSleep, especially in classifying Rapid Eye Movement (REM) sleep.
• Non-Rapid Eye Movement (NREM) classification was robust across models; AccuSleep struggled with REM and Wake.
• Optimal normalization varied by model; raw data favored 1D-CNN and DeepSleepNet.

Conclusions:

• Simpler CNN architectures are well-suited for rodent sleep stage classification.
• Preprocessing strategies should align with specific model architectures and data characteristics.
• DeepSleepNet shows greater advantages over 1D-CNN in human datasets compared to rodent datasets.