DOSED: A deep learning approach to detect multiple sleep micro-events in EEG signal

Area of Science:

• Neuroscience
• Sleep Medicine
• Artificial Intelligence

Background:

• Electroencephalography (EEG) is crucial for sleep disorder identification, involving manual annotation of sleep stages and micro-architecture events.
• Manual annotation of sleep events like spindles, K-complexes, and arousals is time-consuming, subjective, and prone to inter-expert variability.
• Existing automated methods are event-specific and rely on hand-crafted features, limiting their generalizability.

Purpose of the Study:

• To introduce a novel deep learning architecture, Dreem One Shot Event Detector (DOSED), for automated detection of sleep micro-architecture events.
• To enable joint prediction of event location, duration, and type directly from raw EEG signals.
• To overcome limitations of current event-specific and feature-dependent detection algorithms.

Main Methods:

• Developed DOSED, a deep learning model inspired by computer vision object detectors (YOLO, SSD).
• Utilized a convolutional neural network to extract features from raw EEG signals.
• Incorporated localization and classification modules for comprehensive event detection.

Main Results:

• DOSED was evaluated on four distinct datasets.
• The model demonstrated proficiency in detecting three types of sleep micro-architecture events: spindles, K-complexes, and arousals.
• Performance was benchmarked against current state-of-the-art detection algorithms.

Conclusions:

• The DOSED approach shows significant versatility in analyzing EEG time series.
• The novel deep learning architecture achieves improved performance compared to existing state-of-the-art methods.
• DOSED offers a promising automated solution for sleep micro-event detection, enhancing diagnostic efficiency and accuracy.