A novel feature extractor based on constrained cross network for detecting sleep state.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Growing interest in healthy living necessitates accurate sleep monitoring.
• Existing wrist-worn sleep detection methods lack efficiency and accuracy.
• Traditional machine learning and heuristic algorithms struggle with complex sleep-wake classification.

Purpose of the Study:

• To enhance sleep-wake binary classification accuracy using wrist-worn device data.
• To develop an improved feature extractor based on the Constrained Cross Network.
• To address limitations in current sleep detection technologies.

Main Methods:

• Developed a novel feature extractor comprising Feature Derivation and Interaction Modules.
• Feature Derivation Module utilizes dilated convolutions, gated recurrent units, and attention mechanisms.
• Feature Interaction Module employs a Constrained Cross Network, based on DCN-v2, for high-order feature interactions.

Main Results:

• Achieved an F1-score of 91.14% and accuracy of 95.70%, surpassing CNN-based methods.
• Incorporating a non-wear identification mask further improved performance to 94.25% F1-score and 97.38% accuracy.
• The Constrained Cross Network contributed significantly to the feature extractor's performance.

Conclusions:

• The proposed Constrained Cross Network-based feature extractor significantly improves sleep-wake classification accuracy.
• This method offers a more efficient and accurate alternative to traditional deep neural networks for sleep analysis.
• The explicit feature construction has potential for future manual feature development in sleep research.