Knowledge-informed Deep Learning for Apnea Detection with Limited Annotations

用于具有有限注释的呼吸暂停检测的知识型深度学习

• 批准号: 10509437
• 负责人: Changyue Song
• 金额: $ 16.05万
• 依托单位: UNIVERSITY OF FLORIDA
• 依托单位国家: 美国
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-09-01 至 2025-06-30
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10509437
• 关键词: Address; Adopted; Affect; Air Movements; Algorithms; Apnea; Arousal; Attention; Breathing; Clinical; Clinical Data; Complex; Computer Assisted; Consumption; Data; Data Analytics; Detection; Diagnosis; Engineering; Event; Goals; Grain; Health; Healthcare; Home; Knowledge; Learning; Machine Learning; Manuals; Modeling; Monitor; Noise; Oxygen; Patients; Pattern; Performance; Physicians; Physiological; Polysomnography; Probability; Process; Records; Research; Severities; Signal Transduction; Sleep; Sleep Apnea Syndromes; Sleep Disorders; Support System; Techniques; Time; United States; Woman; base; chronic respiratory disease; clinical decision support; clinical decision-making; clinical practice; cost effective; deep learning; deep learning model; detection method; improved; indexing; insight; large datasets; mHealth; machine learning framework; men; novel; patient oriented; statistical learning; supervised learning; wearable device

PROJECT SUMMARY Sleep apnea is a common chronic respiratory disease characterized by breathing difficulties during sleep. Prevalent clinical practice to diagnose sleep apnea requires manual identification of apnea occurrences, which is expensive and time-consuming. Recently, machine learning has attracted much attention to diagnose apnea based on physiological signals collected via wearable devices. However, most existing studies rely on strongly supervised learning for the detection, and fine-grained annotations are required to achieve a high level of granularity. In practice, it is usually expensive and time-consuming to acquire a large dataset with temporally fine-grained annotations (i.e., detecting apnea within short time epochs). Consequently, the limited availability of fine-grained annotations hinders the wide implementation of machine learning and limits its granularity. The ultimate goal of this research is to create a weakly-supervised machine learning framework that incorporates annotations of different granularity levels and clinical domain knowledge for healthcare data analytics. In particular, this study focuses on deep learning because it has shown superior performance and great potential in aiding the analysis of clinical data. The technical objective of the proposed study is to create new deep learning models that incorporate coarse-grained annotations and clinical knowledge for detecting apnea at a high level of granularity based on multiple physiological signals. The specific aims of this proposal are as follows. Aim 1. Systematically identify and quantify the apnea-related patterns in physiological signals. The proposed study will numerically explore the physiological signals to elucidate the patterns related to apnea and other sleep disorders based on feature engineering and statistical learning techniques. Aim 2. Incorporate coarse-grained annotations and clinical knowledge into deep learning models for apnea detection. We will establish new deep learning models to integrate incomplete fine-grained annotations, coarse-grained annotations, and clinical knowledge for apnea detection. Aim 3. Develop an algorithm to adaptively acquire annotations for performance improvement. To further improve the performance of the deep learning model, we will develop an adaptive algorithm to determine whether and where to acquire more annotations from physicians and the level of granularity. The proposed study will address the challenge of generating fine-grained predictions given incomplete or no fine-grained annotations in computer-aided apnea detection. The proposed model will be an advancement to robust and interpretable deep learning that incorporates coarse-grained annotations and domain knowledge. The expected results of study will provide important insights in addressing similar challenges in other biomedical applications, enabling novel real-world solutions such as clinical decision-making support systems, in-home apnea monitoring, and mobile health.

项目概要 睡眠呼吸暂停是一种常见的慢性呼吸系统疾病，其特征是睡眠期间呼吸困难。 诊断睡眠呼吸暂停的普遍临床实践需要手动识别呼吸暂停的发生，这 既昂贵又耗时。最近，机器学习在诊断呼吸暂停方面引起了广泛关注 基于通过可穿戴设备收集的生理信号。然而，大多数现有研究强烈依赖于 检测的监督学习和细粒度的注释需要达到高水平 粒度。在实践中，获取具有时间特征的大型数据集通常是昂贵且耗时的。 细粒度注释（即在短时间内检测呼吸暂停）。因此，可用的资源有限 细粒度的注释阻碍了机器学习的广泛实施并限制了其粒度。 这项研究的最终目标是创建一个弱监督的机器学习框架，其中包含 用于医疗数据分析的不同粒度级别和临床领域知识的注释。在 特别是，这项研究重点关注深度学习，因为它表现出了优越的性能和巨大的潜力 帮助分析临床数据。拟议研究的技术目标是创建新的深度学习 结合粗粒度注释和临床知识的模型，用于高水平检测呼吸暂停 基于多个生理信号的粒度。该提案的具体目标如下。 目标 1. 系统地识别和量化生理信号中与呼吸暂停相关的模式。这 拟议的研究将对生理信号进行数值探索，以阐明与呼吸暂停相关的模式 以及基于特征工程和统计学习技术的其他睡眠障碍。 目标 2. 将粗粒度注释和临床知识纳入深度学习模型中 呼吸暂停检测。我们将建立新的深度学习模型来整合不完整的细粒度 注释、粗粒度注释和呼吸暂停检测的临床知识。 目标 3. 开发一种算法来自适应获取注释以提高性能。到 为了进一步提高深度学习模型的性能，我们将开发自适应算法 确定是否以及在哪里从医生那里获取更多注释以及粒度级别。 拟议的研究将解决在不完整或没有的情况下生成细粒度预测的挑战 计算机辅助呼吸暂停检测中的细粒度注释。所提出的模型将是一个进步 强大且可解释的深度学习，结合了粗粒度注释和领域知识。 研究的预期结果将为解决其他生物医学领域的类似挑战提供重要见解 应用程序，实现新颖的现实世界解决方案，例如临床决策支持系统、家庭 呼吸暂停监测和移动健康。