Heterogeneous Datasets for TinyUStaging

Nowadays, more and more machine learning models have emerged in the field of sleep staging. However, they have not been widely used in practical situations, which may be due to the non-comprehensiveness of these models' clinical and subject background and the lack of persuasiveness and guarantee of generalization performance outside the given datasets. Meanwhile, polysomnogram (PSG), as the gold standard of sleep staging, is rather intrusive and expensive. In this paper, we propose a novel automatic sleep staging architecture called TinyUStaging using single-lead EEG and EOG. The TinyUStaging is an efficient U-Net with multiple attention modules, including Channel and Special Joint Attention (CSJA) block and Squeeze and Excitation (SE) block. Besides, we design sampling strategies and propose a class-aware Sparse Weighted Dice and Focal (SWDF) loss function. The results show that it significantly improves the recognition rate for minority classes and hard samples such as N1 sleep. Noteworthily, we select seven highly heterogeneous datasets covering 9,970 records with above 2w hours among 7,226 subjects spanning 950 days for training, validation and evaluation. Additionally, two hold-out sets containing healthy and sleep-disordered subjects are considered to verify the model's generalization. The results demonstrate that our model outperforms state-of-the-art methods, achieving an average overall accuracy, macro F1-score and kappa of 84.62%, 0.796, 0.764 on heterogeneous datasets, providing a solid foundation for out-of-hospital sleep monitoring.