Graph-based learning for sleep microarchitecture: a hybrid graph autoencoder and graph attention network approach
Background: Sleep plays a vital role in cognitive function, memory consolidation, and overall neurological health. Analysis of sleep microarchitecture including features such as sleep spindles, K-complexes, slow waves, and EEG bandpower components provides critical insights into sleep disorders and...
Gespeichert in:
| Veröffentlicht in: | International journal of research in medical sciences Jg. 13; H. 11; S. 4696 - 4702 |
|---|---|
| Hauptverfasser: | , |
| Format: | Journal Article |
| Sprache: | Englisch |
| Veröffentlicht: |
30.10.2025
|
| ISSN: | 2320-6071, 2320-6012 |
| Online-Zugang: | Volltext |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | Background: Sleep plays a vital role in cognitive function, memory consolidation, and overall neurological health. Analysis of sleep microarchitecture including features such as sleep spindles, K-complexes, slow waves, and EEG bandpower components provides critical insights into sleep disorders and genetic diseases. However, the complex interactions between sleep architecture and underlying genetic abnormalities remain underexplored. This study aims to investigate these interactions by leveraging advanced graph-based deep learning methods to uncover hidden relationships within EEG signals. Methods: We developed a graph autoencoder (GAE) combined with a Graph attention network (GAT) to analyze polysomnography (PSG) data from the National Children's Hospital (NCH) dataset. EEG epochs were modelled as graph nodes, while edges were constructed based on bandpower similarity between epochs, enabling dynamic representation of sleep activity. The GAE learned latent embeddings that capture subtle patterns in sleep microarchitecture, and the GAT applied attention mechanisms to classify and interpret relationships between EEG events, sleep disorders, and genetic abnormalities. Three core analyses were conducted: (1) identifying differences in sleep microarchitecture across sleep disorders, (2) detecting EEG event changes associated with genetic disorders, and (3) exploring shared patterns linking sleep and genetic abnormalities. Results: The model achieved classification accuracies of 92.4%, 91.2%, and 88.6% across the three tasks, respectively. The approach successfully identified distinct EEG event patterns in subjects with co-occurring sleep disorders. Conclusions: This work presents a scalable, automated, and interpretable framework for analyzing the interplay between sleep microarchitecture, sleep disorders, and genetic disorders. |
|---|---|
| ISSN: | 2320-6071 2320-6012 |
| DOI: | 10.18203/2320-6012.ijrms20253587 |