목적: Understanding the physiology and pathophysiology of sleep cycles is clinically significant as humans spend about one-third of their lives asleep. Polysomnography(PSG) is the gold standard in sleep studies. Especially, sleep stage classification is crucial for the effective diagnosis and treatment of sleep-rated disorders. However, manual scoring of polysomnography(PSG) is labor-intensive and subjective, with high inter-rater variability. Although existing machine-learning approaches have made significant progress in recent years, they have not been adopted widely in clinical settings due to black-box skepticism. In this study, we propose a novel vision-based automatic sleep staging model that provides enriched explanations of the model’s predictions. The goal of our method is to assist sleep clinicians with scoring tasks in clinical settings. 방법:Our approach is based on standardized image-based PSG dataset rather than raw signals, to mimic clinician’s scoring as they “see” the waveforms. This dataset encompasses a total of 10,253 patient records, which was collected from four institutes in South Korea between 2013 and 2020. The standard image is a snapshot of what sleep clinicians see in scoring tasks, encompassing not only EEG, EOG, and EMG but also respiratory and movement signals. We also validated our model with publicly available PSG datasets by transforming those signals into images. Our proposed framework adopted Vision Transformer architecture and can be divided into three modules. First, the Intra-epoch Transformer learns the features of single-epoch images. By utilizing the transformer explainability method, we show a detailed attribution map that highlights where the model focuses. After extracting feature vectors from the previous module, we concatenated multiple epochs to generate the input of the Inter-Epoch Transformer. Afterward, the Inter-epoch Transformer learns the context of the adjacent epochs, as clinicians see the preceding and subsequent epochs to score the target one. We also provided Inter-epoch interpretability, which visualizes the relevance of adjacent epochs in predicting the target epoch. Throughout the training phase, we employed image augmentation and 5- fold ensemble method. During inference, we adopted a sliding window scheme to increase the performance. 결과:Our model shows Macro F1 greater than 80%, which is comparable to the state-of-the-art sleep staging models. We evaluate our model to publicly available datasets and verify robust performance. Unlike signal-based models which require tailored pre-processing depending on diverse datasets, our image transformation can be uniformly applied to other datasets. This significantly reduces the training costs by eliminating the necessity for fine-tuning the signal pre-processing. Furthermore, our interpretable visualization shows a significant alignment with the AASM scoring guide rules. For Intra-epoch Interpretability, our results successfully pinpoint salient waveforms such as alpha waves, sleep spindles, K-complexes, and delta waves. The Inter-epoch interpretability shows how the model mirrors the clinicians’ approach of referencing neighboring epoch stages during scoring. We found that REM stages have benefited the most from the Inter-epoch transformer, as clinicians do in scoring tasks. Furthermore, to the best of our knowledge, this is the first attempt to provide an interpretable visualization of how neighboring epochs attend each other. 결론:The proposed framework achieves state-of-the-art performance in terms of the Macro F1 metric and additionally provides the reasoning for the sleep stage prediction. For the interpretability analysis, our method provides the most detailed and accurate visualization aligning with the sleep scoring rules. This breakthrough allows AI models to be seamlessly integrated into practical healthcare settings.