LG-Sleep: Local and Global Temporal Dependencies for Mice Sleep Scoring

TL;DR

The paper tackles automated, subject-independent sleep staging in mice using EEG, addressing inter-subject variability and data imbalance. It introduces LG-Sleep, a two-stage encoder–decoder architecture that combines local temporal features from a time-distributed CNN with global dynamics captured by an LSTM, trained with a reconstruction loss plus classification loss $\mathcal{L}=\mathcal{L}_c+\mathcal{L}_{mse}$ and later refined with class weights. Evaluations on 16 mice with fourfold subject-out CV show LG-Sleep achieves $0.85\pm0.09$ accuracy and $0.75\pm0.10$ macro F1, outperforming state-of-the-art methods, including under limited labeled data conditions (e.g., $N_l=50\%$ yields $0.76\pm0.06$). The approach demonstrates strong generalization across subjects and resilience to limited annotations, potentially enabling scalable, preclinical sleep studies and cross-subject analyses. The work highlights the value of jointly leveraging local and global temporal information in EEG for robust rodent sleep scoring.

Abstract

Efficiently identifying sleep stages is crucial for unraveling the intricacies of sleep in both preclinical and clinical research. The labor-intensive nature of manual sleep scoring, demanding substantial expertise, has prompted a surge of interest in automated alternatives. Sleep studies in mice play a significant role in understanding sleep patterns and disorders and underscore the need for robust scoring methodologies. In response, this study introduces LG-Sleep, a novel subject-independent deep neural network architecture designed for mice sleep scoring through electroencephalogram (EEG) signals. LG-Sleep extracts local and global temporal transitions within EEG signals to categorize sleep data into three stages: wake, rapid eye movement (REM) sleep, and non-rapid eye movement (NREM) sleep. The model leverages local and global temporal information by employing time-distributed convolutional neural networks to discern local temporal transitions in EEG data. Subsequently, features derived from the convolutional filters traverse long short-term memory blocks, capturing global transitions over extended periods. Crucially, the model is optimized in an autoencoder-decoder fashion, facilitating generalization across distinct subjects and adapting to limited training samples. Experimental findings demonstrate superior performance of LG-Sleep compared to conventional deep neural networks. Moreover, the model exhibits good performance across different sleep stages even when tasked with scoring based on limited training samples.

Figures (3)

• Figure 1: Overview of the proposed LG-Sleep deep neural network.
• Figure 2: Sample EEG trace and the corresponding sleep stages: a) true labels, and b) predicted labels. The EEG trace is shown in black, and the background colors purple, red, and green correspond to the wake, NREM, and REM stages, respectively.
• Figure 3: Confusion matrices of the proposed method and model backbones. W, NR, and R denote wake, NREM, and REM sleep stages.