SpindleX: A Multi-Scale Event Detection Window and CBAM Attention-Based Model for Sleep Spindle Detection

Sleep spindles, characteristic waveforms in stage 2 of non-rapid eye movement (NREM) sleep on EEG, are linked to a multitude of diseases. Although numerous sleep spindle detection methods exist, several challenges persist: 1)How to effectively utilize spindle properties, such as their duration and prominence in EEG signals? 2)How to capture spindles without threshold settings? 3)How to make spindle detection resemble object detection without converting signals into images? To address these challenges, we introduce SpindleX, a novel end-to-end sleep spindle detection algorithm. It comprises four key modules: a local feature extraction module for capturing fine-grained characteristics; a sequence feature extraction module for analyzing temporal patterns; an attention module to highlight important features across channel and spatial dimensions; and a localization and classification module to precisely pinpoint and categorize spindles. Convolutional neural networks and bidirectional long short-term memory networks are used to extract common spindle characteristics. The attention module then captures more details from the signal's channel and spatial aspects. To precisely locate spindles, multi-scale event detection windows are employed to fit the varying spindle durations, aiming to capture all spindles in the EEG sequence. Experiments on the widely-used sleep dataset MASS show that our method is significantly efficient.Clinical relevance- This study proposes a reliable automated spindle detection method, which achieves an average F1-score of 78.9% on public datasets. This method provides clinicians with an efficient and practical tool, reducing the time cost of manual scoring by experts while mitigating the inconsistency caused by subjectivity. As a result, it offers support for clinical applications such as sleep quality assessment and disease prognosis analysis.