基于多数据多模型融合的睡眠节律评估与智能干预研究

Sleep assessment and sleep intervention have always been a common concern in scientific and social fields. The rapid development of portable sleep monitoring terminals provides various types of sleep data, the size and collection of these data can not meet the requirements of the artificial annotation sleep state. Considering small data samples with labels, artificial constructed features, and traditional machine learning method, a new learning method and model needs to be introduced. The rapid development of semi-supervised learning methods and deep learning models makes it possible to conduct personalized intelligent sleep assessment from a large amount of data. The project introduce semi-supervised learning methods and deep learning models into sleep assessment and intervention studies based on various sleep data from professional devices and portable devices. The study targets at the following aspects: 1. Sleep phase features based on multiple data; 2. Sleep assessment based on multiple models; 3. Sleep intervention based on neural feedback. This study not only has a deeper understanding of the characteristics of human sleep, but also provides a scientific basis for the theory and practice of personalized intelligent sleep.

睡眠的评估和睡眠问题的干预一直是科学领域和社会领域人们普遍关注的热点。便携式睡眠监测终端的迅速发展提供了大量的多种类睡眠数据，这些数据的规模和收集方式无法满足人工标注睡眠状态的要求，使得以小样本有标签数据、人工构建特征、和传统机器学习方法为主的睡眠研究亟待引入新的学习方法与模型。半监督学习方法与深度学习模型的迅速发展，使得人们从大量数据中进行个性化智能睡眠评估成为可能。本项目针对来自专业设备和便携式设备的多种类睡眠数据，将半监督学习方法和深度学习模型引入睡眠评估与干预研究中，从以下三个方面展开研究：1、基于多数据的睡眠时相特征表达；2、基于多模型的睡眠评估；3、基于神经反馈的睡眠干预。这项研究不仅对人类睡眠的特点有更深层次的认识，同时也为个性化智能睡眠提供了理论和实践的科学依据。

睡眠节律的准确评估不仅有利于人们及时了解自身的身体健康状态，同时也为睡眠障碍相关的心理和生理疾病的治疗与干预提供了必要的依据和手段。本项目针对睡眠相关的生理和行为数据，将半监督学习方法和深度学习模型引入睡眠评估与干预研究中，以解决睡眠大数据带来的人工睡眠状态标注的局限，借助多种类数据的多模型融方法与神经反馈技术，实现睡眠节律的个性化评估与智能干预。取得的主要成果如下：1）提出了基于智能搜索算法的睡眠时相特征提取方法，为构建睡眠干预模式提供依据；2）对半监督稀疏表示的学习方法进行改进，构建了半监督稀疏表示睡眠分期模型并进行了评估；3）对无监督特征学习模型、迁移学习模型、以及数据增强与平衡模型进行改进与应用，实现了针对睡眠数据的深度学习模型自动构建与评估；4）设计并完成了针对睡眠改善的fMRI和EEG神经反馈情绪调节实验。上述研究成果推动了深度学习方法在睡眠大数据应用中的发展，促进了睡眠节律评估与干预的的应用转化。