GraphSleepFormer: a multi-modal graph neural network for sleep staging in OSA patients.

Area of Science:

• Artificial Intelligence
• Sleep Medicine
• Biomedical Engineering

Background:

• Obstructive sleep apnea (OSA) is a common sleep disorder impacting health.
• Accurate sleep staging is crucial for diagnosing sleep disorders and assessing sleep quality.
• Existing methods may not fully capture complex dependencies in sleep data.

Purpose of the Study:

• Introduce the novel GraphSleepFormer (GSF) network for enhanced sleep staging.
• Improve the capture of global dependencies and node characteristics in graph-structured sleep data.
• Evaluate the GSF's effectiveness in sleep staging for OSA patients.

Main Methods:

• Developed the GraphSleepFormer (GSF) network incorporating centrality and spatial coding.
• Employed adaptive learning of adjacency matrices for spatial encoding of head channel data.
• Utilized Shapley Additive Explanations (SHAP) to assess channel contributions and multimodal data necessity.

Main Results:

• Achieved 80.10% overall accuracy in sleep staging using polysomnography data from 28 OSA patients.
• Demonstrated performance comparable to state-of-the-art methods on ISRUC database subsets.
• Ablation experiments confirmed the effectiveness of centrality and spatial encoding methods.

Conclusions:

• The GSF network accurately identifies sleep periods, crucial for OSA diagnosis and treatment.
• This novel approach advances sleep medicine by improving sleep staging accuracy.
• Multimodal data and advanced network architectures are essential for robust sleep analysis.