PSG-MAE: Robust Multitask Sleep Event Monitoring Using Multichannel PSG Reconstruction and Inter-Channel Contrastive Learning

Polysomnography (PSG) signals are essential for studying sleep processes and diagnosing sleep disorders. With the advancement of deep neural networks (DNNs), automated analysis of PSG data has become increasingly feasible. However, the limited availability of data for certain sleep events often restricts DNNs to single-task learning on a single-source dataset, limiting their ability to generalize to new events and reducing robustness across datasets. To address these challenges, we propose PSG-MAE, a pretraining framework based on the masked autoencoder (MAE). By leveraging self-supervised learning on large volumes of unlabeled PSG data, PSG-MAE trains a robust feature extraction network applicable to diverse sleep event monitoring tasks. Unlike conventional MAEs, PSG-MAE applies complementary masking across PSG channels, integrates a multichannel signal reconstruction mechanism, and incorporates an inter-channel contrastive learning (ICCL) strategy. This design enables the encoder to capture temporal features from each channel while simultaneously modeling latent inter-channel relationships, thereby enhancing the utilization of multichannel information. Experimental results demonstrate that PSG-MAE effectively learns both temporal details and inter-channel dependencies from PSG signals. When the pretrained encoder is fine-tuned with downstream sleep event monitoring networks, it achieves a macro-averaged F1-score of 81.0% for sleep staging and 82.6% for apnea detection on the SHHS dataset. Cross-dataset validation on the MESA dataset further confirms the framework's robustness and broad applicability. The code of PSG-MAE is available at https://github.com/yfw-scut/PSG-MAE.git.