CRII: SCH: III: Novel Data-Driven Methods to Analyze Heterogeneous Healthcare Data

该项目的长期目标是提高患者的护理标准，并建立临床医生对利用先进的机器学习和人工智能工具进行计算医疗保健的信任。通过2009年《健康信息技术促进经济和临床健康法案》(HITECH)在全国范围内推动电子健康记录，以及可穿戴传感器技术的最新进展，导致数字健康数据的数量、细节和可用性呈指数级激增。这为研究人员、医疗保健专业人员和患者提供了一个令人兴奋的机会，以推断更丰富的、数据驱动的对健康和疾病的理解。然而，与其他数据类型不同的是，医疗保健数据本质上是嘈杂的，缺少值，并且来自多种不同的来源，如实验室测试、医生病历、医学图像和监护仪读数。这些数据特性使大多数现有的机器学习方法和统计模型在发现有意义的疾病模式或做出可靠的预测方面具有很大的挑战性。为了应对这些挑战，该项目将基于强大的深度学习技术开发和验证新的数据驱动方法，以对医疗数据中存在的复杂相关性和模式进行建模。特别是，建议的数据驱动方法将从异质和有限的医疗数据中学习特定于疾病和特定于患者的特征模式。拟议的数据驱动方法的有效性将在具有挑战性和重要的医疗预测任务中得到展示，例如脓毒症的早期预测和预测重症监护病房患者的结局。该项目将倡导基于模型的数据驱动的计算医疗保健范式转换，并将重点通过开发新的数据驱动方法来解决医疗保健数据分析的主要挑战，即异构性和有限的数据集大小。拟议的数据驱动方法将在以下几个方面加速医学发现和帮助临床决策：(A)连接并从互不相连的不同种类的医疗数据堆中学习；(B)产生新的疾病/疾病表示，以及(C)建立临床医生对数据驱动模型的信任。该项目的技术目标分为三个方面。第一个重点将集中在开发一个新的深度学习框架，以从不同类型的医疗数据中学习共享特征表示。具体地说，研究人员将使用机器学习方法，如多视图学习和相关性分析，以利用存在于不同医疗保健数据源内和跨不同医疗数据源的相关性结构。此外，基于对抗性训练的领域自适应技术将被用于从多队列患者群体中学习联合特征表示。第二个重点将集中在通过利用患者或任务相似性网络和几次学习框架从有限的医疗数据中学习功能。特别是，多任务学习和嵌入技术将用于从可用于特定患者队列或医疗保健任务的有限数据中学习特征表示。在第三个推力中，将使用集成和正则化技术来研究所提出的数据驱动方法的模型不确定性。建议的数据驱动解决方案的充分性将在多个临床相关预测任务的真实医疗数据集上进行验证。该奖项反映了NSF的法定使命，并已通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The long term goal of this project is to improve the standard-of-care of patients and build clinicians’ trust in utilizing advanced machine learning and artificial intelligence tools for computational healthcare. The national push for Electronic Health Records through the 2009 Health Information Technology for Economic and Clinical Health (HITECH) Act and the recent advances of wearable sensor technologies has resulted in an exponential surge in volume, detail, and availability of digital health data. This provides an exciting opportunity for researchers, healthcare professionals, and the patients alike to infer richer, data-driven understanding of health and illness. However, unlike other data types, healthcare data is inherently noisy, has missing values, and comes from multiple heterogeneous sources such as lab tests, doctor notes, medical images, and monitor readings. These data properties make it very challenging for most existing machine learning approaches and statistical models to discover meaningful patterns of diseases or to make robust predictions. To address these challenges, this project will develop and validate novel data-driven methods based on powerful deep learning techniques to model the complex correlations and patterns present in the healthcare data. In particular, the proposed data-driven methods will learn disease-specific and patient-specific feature patterns from heterogeneous and limited healthcare data. The effectiveness of the proposed data-driven methods will be showcased on challenging and important healthcare prediction tasks such as early prediction of sepsis and predicting the outcome of Intensive Care Units patients. This project will advocate a model-based data-driven paradigm shift for computational healthcare, and it will focus on addressing the main challenges of analyzing healthcare data, i.e., heterogeneity and limited dataset size, by developing novel data-driven methods. The proposed data-driven methods will accelerate medical discovery and aid in clinical decision making in several ways: (a) connect and learn from the disconnected heterogeneous piles of healthcare data; (b) yield new representations of illness/diseases, and (c) build clinicians’ trust in the data-driven models. The technical aims of the project are divided into three thrusts. The first thrust will focus on developing a novel deep learning framework to learn shared feature representations from heterogeneous healthcare data. Specifically, the researchers will employ machine learning approaches, such as multi-view learning and correlation analysis, to exploit the correlation structures present within and across different healthcare data sources. In addition, adversarial training based domain adaptation techniques will be used to learn joint feature representations from multi-cohort patient populations. The second thrust will focus on feature learning from limited healthcare data by utilizing patient or task similarity networks and a few-shot learning framework. In particular, multi-task learning and embedding techniques will be used to learn feature representations from limited data available for a specific patient cohort or healthcare task. In the third thrust, model uncertainty of the proposed data-driven methods will be studied using ensembles and regularization techniques. The adequacy of the proposed data-driven solutions will be validated on real-world healthcare datasets for multiple clinically relevant prediction tasks.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.