Neural network architectures and normalization techniques for automated sleep stage classification using rodent EEG
Jinyoung Choi1, Hankil Oh2, Minkyu Ahn2
1Department of Anesthesiology, Mass General Brigham, Department of Anaesthesia, Harvard Medical School, Boston, Massachusetts, United States of America.
None:
Accurate sleep stage classification in animal models is crucial for translational sleep research, enabling the study of mechanistic pathways and therapeutic interventions. Because manual scoring is labor-intensive and variable, artificial neural networks are increasingly used for automation. However, few models are tailored for animal sleep staging, and direct cross-model comparisons under consistent conditions remain limited. We presents a systematic evaluation of three representative neural architectures for automated sleep stage classification using rodent electroencephalogram and electromyogram: a conventional 1-dimensional convolutional neural network (1D-CNN), a 2-dimensional convolutional neural network (AccuSleep), and a convolutional neural network combined with bidirectional long short-term memory (DeepSleepNet). Performance was assessed under within-subject and cross-subject validation frameworks, comparing raw input, z-scoring, and mixture z-scoring. Both 1D-CNN and DeepSleepNet consistently outperformed AccuSleep, particularly for Rapid Eye Movement (REM), where AccuSleep exhibited marked deficits plausibly attributable to class imbalance. Class-wise analysis confirmed stable Non-Rapid Eye Movement (NREM) classification across models, while AccuSleep showed reduced robustness in REM and Wake. Normalization effects were model-dependent: raw data yielded superior outcomes for 1D-CNN and DeepSleepNet, whereas AccuSleep showed modest improvement in Wake detection under mixture z-scoring. Comparison with human electroencephalogram literature indicated that DeepSleepNet's advantage over 1D-CNN is more pronounced in human datasets (especially NREM 1), likely reflecting differences in sleep architecture. These findings highlight the suitability of simpler CNNs for rodent sleep stage classification and underscore the importance of aligning preprocessing strategies with model architecture and data characteristics.
