SCH:INT: Collaborative Research: Deep Sense: Interpretable Deep Learning for Zero-effort Phenotype Sensing and Its Application to Sleep Medicine

睡眠占每个人生命的三分之一，影响着每个人的生活质量和健康。传统上，由于缺乏方便的监测设备和自动算法，睡眠模式的长期特征（睡眠表型）在很大程度上是未知的。 最近，活动数据、脑电图、呼吸监测数据和心电图等海量健康感知数据正在诊所和家庭中收集，这为了解诊所外的睡眠表型带来了前所未有的机会。然而，将这些噪声和不可靠的多模态感测数据转化为准确的表型（例如睡眠阶段和呼吸暂停事件）存在巨大的挑战。除了睡眠，许多神经系统疾病，如阿尔茨海默氏症和帕金森氏症，都希望客观跟踪长期疾病进展，这是目前不可能的。该项目将提供计算能力，以在家中进行表型跟踪，重点是睡眠表型。将开发机器学习方法和软件，以在传感器仪器、数据收集和分析上花费最少的努力来进行准确的睡眠表型分析。该项目旨在开发DeepSense，这是一种深度学习工具箱，用于对大量数据流进行建模，包括多导睡眠图等临床监测数据和来自无线传感设备的新型射频信号。研究团队将开发精确的深度学习方法，使用多导睡眠图数据自动进行睡眠监测。他们将发明对抗性深度学习方法来建模射频信号，并利用大量历史多导睡眠图数据来帮助改进射频信号模型。他们将开发可解释的模型，利用和扩展睡眠表型的医学知识。最后，所有提出的模型将通过一项前瞻性研究进行验证，目标是自动化手动睡眠研究，并通过射频信号数据评估睡眠研究的可行性。该研究团队计划发布该项目的开源软件和大型数据集，以造福计算机科学、工程和医学界。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Sleep represents one third of everyone’s life and affects the quality and the health of everyone’s life. Traditionally, long-term characteristics of sleep patterns (sleep phenotypes) are largely unknown due to the lack of convenient monitoring devices and automatic algorithms. Recently, massive health sensing data such as activity data, electroencephalogram, respiratory monitoring data and electrocardiography are being collected in clinics and at home, which brings unprecedented opportunities for understanding sleep phenotypes outside clinics. However, there are tremendous challenges to translate these noisy and unreliable multimodal sensing data into accurate phenotypes such as sleep stages and apnea events. Beyond sleep, many neurological conditions such as Alzheimer’s and Parkinson’s all expect objective tracking of long-term disease progression, which is currently impossible. This project will provide the computational capability to conduct phenotype tracking at home with the focus on sleep phenotypes. Machine learning methods and software will be developed to conduct accurate phenotyping of sleep with minimal effort on sensor instrumentation, data collection and analysis. This project aims at developing DeepSense, a deep learning toolbox to model massive data streams including in-clinic monitoring data such as polysomnography and novel radio frequency signals from a wireless sensing device. The research team will develop accurate deep learning methods to automate sleep monitoring using polysomnography data. They will invent adversarial deep learning methods for modeling radio frequency signals and leverage large historical polysomnography data to help improve models for radio frequency signals. They will develop interpretable models that leverage and expand medical knowledge on sleep phenotypes. Finally, all the proposed models will be validated through a prospective study with a goal of automating manual sleep studies and assessing the feasibility of sleep studies via radio frequency signal data. The research team plans to release the open-source software and large datasets from this project that can benefit computer science, engineering and medical community.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.