A novel feature extractor based on Constrained Cross Network for Detecting Sleep State

Abstract With increasing awareness of healthy living and social pressure, more and more people have begun to pay attention to their sleep state. Therefore, detecting sleep state is important for personal physical and mental health. Usually, when we talk about sleep detection, the main data source is polysomnography (PSG). Sleep detection based on wrist-worn device data is an excellent alternative to PSG data. However, the application of machine learning and deep learning techniques to wrist-worn device data is still in its early stages. Most existing methods that utilize this type of data for detection rely on heuristic algorithms or traditional machine learning, which suffer from low classification efficiency and insufficient accuracy. To further increase the accuracy, this paper improves the weight matrix of Cross Network in Improved Deep&Cross Network and adds it to the traditional feature extractor. First, we verify the theoretical validity of this method. Then, in the experimental phase, the effectiveness of different feature sets and parameter scales of the feature extractor is carried out. The experimental results show that the accuracy increases from 90.0% to 95.7% when using fully connected layers as the classifier. The ablation experiment shows that at a baseline of 90.00%, the accuracy of the feature derivation module has increased by 3.72%, and the accuracy of the feature crossover module has increased by 2.90%. Using the Constrained Cross Network, the average accuracy increases from 4.45% to 5.67%, which can accurately detect sleep state and then be used to evaluate sleep quality.