Personalized Sleep Prediction via Deep Adaptive Spatiotemporal Modeling and Sparse Data.

Area of Science:

• Computational neuroscience
• Biomedical engineering
• Health informatics

Background:

• Sleep quality significantly impacts mental and physical well-being.
• Accurate sleep forecasting is crucial for proactive health management.
• Existing models often struggle with personalized predictions from wearable data.

Purpose of the Study:

• To develop an adaptive spatial and temporal model (AdaST-Sleep) for predicting sleep scores.
• To improve personalized sleep forecasting using data from commercial wearable devices.
• To provide a tool for individuals and healthcare providers to manage sleep patterns.

Main Methods:

• Utilized a hybrid model combining convolutional neural networks (CNNs) for spatial features and recurrent neural networks (RNNs) for temporal data.
• Integrated a domain classifier for subject generalization.
• Experimented with various input (3-11 days) and prediction (1-9 days) window sizes.

Main Results:

• AdaST-Sleep outperformed four baseline models across tested window sizes.
• Achieved a Root Mean Square Error (RMSE) of 0.282 with a 7-day input and 1-day prediction window.
• Demonstrated robust performance in multi-day forecasting and accurate tracking of sleep score fluctuations.

Conclusions:

• The AdaST-Sleep framework offers a robust and adaptable solution for personalized sleep forecasting.
• The model effectively utilizes sparse data from wearable devices and domain adaptation.
• Clinical relevance lies in informing lifestyle interventions and tracking participant progress in improving sleep.