Sleep Stage Classification in Children Using Self-Attention and Gaussian Noise Data Augmentation

Area of Science:

• Medical Informatics
• Machine Learning
• Sleep Medicine

Background:

• Accurate sleep stage analysis is crucial for diagnosing and treating pediatric sleep disorders.
• Data imbalance and the difficulty in identifying minority sleep stages pose significant challenges in automated sleep analysis.

Purpose of the Study:

• To develop and evaluate a novel sleep stage classification method that addresses data imbalance and improves the identification of minority sleep stages.
• To assess the performance of the proposed method on both pediatric and adult sleep datasets.

Main Methods:

• Applied Gaussian Noise Data Augmentation (GNDA) to balance sleep stage data distribution in polysomnography recordings.
• Developed a DeConvolution- and Self-Attention-based Model (DCSAM) to extract local and correlational features for improved sleep stage distinction.
• Evaluated the DCSAM-GNDA method on a custom pediatric dataset and the public Sleep-EDFX dataset for adult sleep analysis.

Main Results:

• The DCSAM-GNDA method achieved 90.26% accuracy and 86.51% macro F1-score on the pediatric dataset.
• On the Sleep-EDFX dataset, DCSAM demonstrated comparable performance to state-of-the-art methods, with high accuracies for various sleep stage classifications (e.g., 95.30% for three-stage).

Conclusions:

• The proposed DCSAM model combined with GNDA effectively overcomes data imbalance and feature extraction challenges in sleep stage classification.
• This approach shows significant potential for improving performance in diverse medical domains dealing with imbalanced time-series data.