An Interpretable Single-Channel EEG Sleep Staging Model Based on Prototype Matching and Multi-Task Learning

Sleep staging is an essential technique for evaluating sleep quality. Recently, deep learning-based sleep staging methods have achieved high performance. However, existing methods are unable to provide a decision basis for classification, which limits their clinical reliability. To address this issue, we propose a novel model, which integrates prototype matching with multi-task learning to enhance interpretability without sacrificing performance. Our method captures significant waveform prototypes from various stages of sleep EEG as templates. Subsequently, the similarity score is calculated by comparing the deep representations of testing EEG segments with these prototypes as the basis for classification. To further learn the sleep stage transition rules and improve performance, we design a multi-task learning framework. We take the sample level sleep stage classification and sleep transition stage discrimination as two auxiliary tasks, and take the sequence level sleep stage classification as the main task. Compared to other state-of-the-art methods, our method demonstrates competitive performance, achieving an overall accuracy of 85.9% and macro-averaged F1 score of 81.0% on the SleepEDF dataset, and an accuracy of 86.5% and a macro-averaged F1 score of 80.6% on the MASS-SS3 dataset. Importantly, our method provides a visualization of the prototype set, and the salient sleep waveforms in the prototype set representing unique sleep states are labeled by the expert physician. In summary, our method enhances model interpretability while ensuring robust sleep staging performance, promising a reliable and clinically applicable tool for automatic sleep staging.