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

With increasing awareness of healthy living and social pressure, more and more people have begun to pay attention to their sleep state. Most existing methods that utilize wrist-worn devices data for detection rely on heuristic algorithms or traditional machine learning, which suffer from low classification efficiency and insufficient accuracy. This study explores an improved feature extractor based on the Constrained Cross Network to enhance the accuracy of the sleep-wake binary classification problem. The feature extractor consists of Feature Derivation Module and Feature Interaction Module. Feature Derivation Module leverages dilated convolutions, gated recurrent units, and attention mechanisms to construct new features in batches. The main structure of Feature Interaction Module is composed of Constrained Cross Network. This module is based on the improved Cross Network in the Improved Deep&Cross Network (DCN-v2), aiming to model high-order feature interactions. The dataset consists of 277 individuals, with varying numbers of recorded days. The ratio of sleep to wake states is 3:7. We extracted 80% of the 7,523 subsamples (divided by day) for training and validation, while the remaining 20% was used as the test set. Compared with the CNN-based method, the proposed method improves the F1-score from 75.84% to 91.14%, and the accuracy increases from 90.03% to 95.70%. After adding a non-wear identification mask, the proposed method achieves an F1-score of 94.25%, with the accuracy further improved to 97.38%. When using the same classifier, the Constrained Cross Network contributed approximately 34% to the feature extractor's performance, while the entire feature extractor further improved the effectiveness of feature extraction. Compared to traditional DNNs, the proposed method offers a more efficient approach to feature extraction, resulting in a notable enhancement in model performance, albeit with a moderate increase in computational complexity. Furthermore, given the explicit feature construction characteristics of the Cross Network, this method has the potential to assist in developing more pronounced manual features in future research.