DARN: An Attention-Based Neural Network Using Residual Blocks for Sleep Micro-Events Detection

Abstract

Sleep micro-events are crucial indicators of sleep quality and neurological health. However, traditional sleep micro-events detectors often suffer from low precision, which leads to a high rate of false positives and misclassification. This affects the reliability of sleep research and clinical diagnosis. In this paper, we propose an attention-based neural network using residual blocks named DARN for precise detection of sleep micro-events. It consists of three main parts: 1) local feature extraction: convolutional neural networks (CNNs) with attention mechanism are used for extracting temporal features; 2) residual module: solving the problems of gradient vanishing and network degradation, for precise feature representations learning; and 3) temporal context modelling: bidirectional long short-term memory (Bi-LSTM) units to better capture the contextual relationships in the sleep electroencephalogram (EEG). By conducting experiments on the publicly accessible datasets MASS2 and DREAMS, the proposed architecture achieves a precision of 0.771 and an F1-score of 0.768 for detecting sleep spindles (MASS2), a precision of 0.632 and an F1-score of 0.608 for detecting K-complexes (MASS2), and a precision of 0.642 and an F1-score of 0.616 for detecting sleep spindles (DREAMS). The results show that our method outperforms existing detection methods, laying the foundation for better clinical utility.