Domain Adversarial Convolutional Neural Network Improves the Accuracy and Generalizability of Wearable Sleep...

Area of Science:

• Biomedical Engineering
• Sleep Science
• Artificial Intelligence

Background:

• Wearable accelerometers are crucial for at-home sleep monitoring in research and clinical practice.
• Existing algorithms for accelerometer-based sleep analysis have limitations in accuracy and generalizability.
• Improving sleep assessment tools is vital for understanding and managing sleep disorders.

Purpose of the Study:

• To evaluate a deep learning domain adversarial convolutional neural network (DACNN) model for sleep-wake classification and outcome estimation using wrist-worn accelerometry.
• To compare the performance of the DACNN model against established sleep algorithms.
• To assess the generalizability of the DACNN model across different wearable devices and activity count datasets.

Main Methods:

• Application of a deep learning domain adversarial convolutional neural network (DACNN) model to wrist-worn accelerometer data.
• Classification of sleep-wake states and estimation of sleep outcomes.
• Validation of the model on an independent dataset from different wearable devices.

Main Results:

• The DACNN model demonstrated superior performance compared to existing sleep algorithms in classifying sleep-wake and estimating sleep outcomes.
• Achieved 80.1% accuracy, with 84% sensitivity and 58% specificity, generalizing well to a new dataset.
• Significantly reduced errors in Wake After Sleep Onset (WASO) and Sleep Efficiency estimation compared to Cole-Kripke, Sadeh, and z-angle algorithms.

Conclusions:

• Deep learning DACNN models enhance the accuracy and specificity of sleep-wake classification from wrist accelerometer data.
• Accelerometer-based sleep monitoring remains a valuable, cost-effective, and user-friendly method for sleep assessment.
• The DACNN model represents a scalable and valid advancement for real-life sleep outcome assessment.