Sleep stage prediction using multimodal body network and circadian rhythm

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Sleep Science

Background:

• Quality sleep is crucial for physiological restoration and waste removal.
• Sleep disturbances are linked to various health issues, including depression and metabolic problems.
• Understanding sleep cycles and stages is vital for diagnosing and managing sleep disorders.

Purpose of the Study:

• To develop and evaluate models for predicting sleep stages and patterns using data from wearable devices.
• To compare the effectiveness of memoryless (ML) and memory-based (LSTM) models in sleep stage prediction.
• To validate the proposed methodology on diverse datasets, including real-world and benchmarked data.

Main Methods:

• Collected multi-parameter data from subjects using wearable devices.
• Applied various memoryless classifiers (Random Forest, LR, MLP, kNN, SVM) and a memory-based model (LSTM).
• Evaluated model performance using accuracy (ACC) and Cohen Kappa on five custom and two public datasets.

Main Results:

• Random Forest achieved high accuracy (0.96) and Kappa (0.96) among memoryless models.
• LSTM demonstrated strong performance across datasets, with a maximum accuracy of 0.88 and Kappa of 0.82.
• The methodology outperformed original work, with memory-based models showing better sleep pattern capture.

Conclusions:

• Wearable device data combined with advanced modeling can accurately predict sleep stages and patterns.
• Memory-based models, especially LSTM, are more effective in capturing temporal dynamics of sleep compared to memoryless models.
• This approach holds promise for non-invasive sleep monitoring and personalized sleep health management.