Innovative Machine Learning for Medical Data Analytics

背景：现代临床医生被数据所淹没：放射影像、血液化学、临床记录等等。在这些不同的数据来源中，特别是在这些数据来源之间，存在着对患者健康的关键见解。然而，由于灾难性的信息超载，这些见解在当代临床实践中大多没有被观察到。动机：医学成像中使用的大多数机器学习方法最初都是为计算机视觉设计的，其应用领域与医疗数据非常不同。这些方法不能直接处理医疗数据的异质性（不同的图像模式、不同的采集协议、不同的文档格式、不同的时间序列）和非标准化采集时间的挑战。此外，它们通常被设计为使用非常大的训练数据集（例如，带有标记的地面真值的一百万个训练样本），这在医疗数据应用中通常是不可能的。目的：我们的目标是形成一个多学科的研究计划，从根本上研究创新的机器学习处理医疗数据分析的异质性。具体来说，在5年的时间里，我们将遵循以下目标，探索：1)首次从图像和文本中统一提取特征；2)第一次异构域适应；3)第一代计算机辅助诊断系统；4)纵向风险预测系统。方法：我们将建立在我们之前在基本回归学习和直接分析系统方面的成功。我们新提出的方法将基于最先进的机器学习（例如生成对抗网络）框架和我们自己提出的优化函数。它们一起将为从异构来源获得的所有数据提供一个无缝的分析框架。然后将这些新方法嵌入到多计算机辅助诊断和预测系统中，以实现新的临床应用。新颖性和预期意义：本研究将进一步丰富快速发展的机器学习和数据科学领域，并将其从理论和基础科学转向实际工程。它将能够预测疾病的发生或进展和预后。这将使医生能够更好地治疗患者，并改善患者的健康和生活质量；尽可能多地了解一个病人，在他们生命的早期。这是至关重要的，因为在早期阶段识别严重疾病的警告信号，可以使预防和治疗比疾病后期所需的预防和治疗更简单和更便宜。这项研究将在没有额外成本的情况下为现有临床数据提供附加价值，并带来更有效和高效的医疗保健，这将使患者、医生和更广泛的医疗保健系统受益。

BACKGROUND: The modern clinician is buried in data: Radiographic images, blood chemistry, clinical notes and more. Within, and especially between, these different sources of data lie critical insights into the health of patients. These insights, however, go mostly unobserved in contemporary clinical practice due to catastrophic information overload. MOTIVATION: Most machine learning methods employed in medical imaging were originally designed for computer vision with applications in very different domains than medical data. These methods cannot directly handle the challenge of heterogeneity (different image modalities, different acquisition protocols, different documentation formats, different time series) and non-standardized acquisition time in medical data. Moreover, they are usually designed to use very large training dataset (e.g. one million training samples with labelled ground truth), which is usually not possible in medical data applications. OBJECTIVES: Our aim is to form a multidisciplinary research program to fundamentally investigate the innovative machine learning handling heterogeneity for medical data analytics. Specifically, over 5 years, we will pursue our aim following these objectives, which explore: 1) the first unified feature extraction from both images and texts; 2) the first heterogeneous domains adaption; 3) the first next-generation computer-aided diagnosis system; 4) the longitudinal risk prediction systems. APPROACH: We will build on our previous success in fundamental regression learning and direct analytics system. Our newly proposed approaches will be based on the state-of-art machine learning (e.g. generative adversarial network) framework and our self-proposed optimization functions. Together they will provide a seamless analytic framework for all the data acquired ned from heterogeneous sources. And then these new approaches will be embedded in the multiple computer-aided diagnosis and prediction system to enable the new clinical applications. NOVELTY AND EXPECTED SIGNIFICANCE: This research will further enrich the fast-growing fields of machine learning and data science and move them from theoretical and fundamental science to practical engineering. It will enable prediction of diseases onset or progression and prognosis. This will allow physicians to better treat their patients and lead to improved health and quality of life for patients; to understand as much about a patient as possible, as early in their life as possible. This is crucial since identifying warning signs of serious illness at an early stage enables prevention and treatments that are far simpler and less expensive than those needed at later stages of the disease. This research will provide added value with no extra cost to existing clinical data and lead to more effective and efficient healthcare, which will benefit patients, physicians, and the broader healthcare system.