Deep learning Based Phenotyping and Treatment Optimization for Heart Failure with Preserved Ejection Fraction

射血分数正常的心力衰竭（HFpEF）是一个主要的公共卫生问题 随着人口老龄化和肥胖、糖尿病和糖尿病的流行， 高血压HFpEF约占所有心力衰竭（HF）病例的50%， 在美国至少有300万人患有HFpEF，与高发病率和死亡率相关。HF后 在住院治疗期间，HFpEF的5年存活率为令人沮丧的35%，这比大多数癌症更差。 此外，HFpEF的生活质量与射血分数降低的HF一样差或更差 （HFrEF）。已经进行了一系列大规模的临床试验，但其中大多数仅提供 中性结果，未能证明治疗的有效性。HFpEF的惊人趋势， 病人缺乏有效的治疗方法是一个主要的公共卫生问题。 最近的研究将这一失败归因于HFpEF综合征独特的全身性质 并提出异质性HFpEF综合征中的亚表型， 越来越需要针对特定HFpEF亚型的更好靶向治疗。看似 不同但复杂的相互关联的表型，沿着合并症，生活方式和 环境因素，使多器官综合征从大数据方法中获益最大。 然而，传统研究通常只包括有限的横断面临床症状， 心脏成像的实验室结果和/或大体测量，以研究HFpEF，忽略 电子健康记录（EHR）中丰富的时间信息和详细的空间信息 保留在成像中。 在这个方案中，我们将引入先进的形状分析方法来提取新的图像 特征和生物标志物，并在人群水平上进行验证（目标1）。然后我们将 联合收割机将图像信息与多维时间EHR数据相结合， 使用最先进的机器学习技术进行显著的HFpEF亚类（即表型分析） (Aim 2）。为了达到基于表型的治疗目标，我们将进一步研究最佳的 使用基于深度强化学习（RL）的当前可用代理的治疗策略 在正在进行的试验提供足够的证据之前， 研究经临床证实有效的新药（目标3）。此外，我们将开发一个在线的，开放的- 访问平台，以促进本研究的代码、数据和知识的共享（目标4）。我们 相信这项研究可以提高我们对HFpEF的理解、表型分析和管理， 这可能反过来在美国和世界范围内积极地减轻临床和经济负担。

Heart failure with preserved ejection fraction (HFpEF) is a major public health problem that is rising in prevalence with the aging population and the epidemics of obesity, diabetes, and hypertension. HFpEF accounts for around 50% of all heart failure (HF) cases with a prevalence of at least 3 million in the U.S. HFpEF is associated with high morbidity and mortality. After HF hospitalization, the 5-year survival of HFpEF is a dismal 35%, which is worse than most cancers. In addition, quality of life in HFpEF is as poor or worse than HF with reduced ejection fraction (HFrEF). A series of large-scale clinical trials has been conducted, but most of them only provided neutral result and failed to prove the efficacy of treatments. The alarming trend of HFpEF with lack of effective therapies for patients constitutes a major public health problem. Recent studies have attributed this failure to distinct systemic nature of HFpEF syndrome and proposing sub-phenotypes within the heterogeneous HFpEF syndrome, which highlighted the increasing need for better-targeted therapies to specific HFpEF subtypes. The seemingly disparate but complex interrelated phenotypes, along with comorbidities, lifestyle and environmental factors, make the multi-organ syndrome best beneficial from a big data approach. However, conventional studies usually only included limited cross-sectional clinical symptoms, lab results and/or gross measurements on cardiac imaging to investigate HFpEF, overlooking the rich temporal information from electronic health record (EHR) and detailed spatial information reserved in imaging. In this proposal, we will introduce advance shape analysis method to extract novel image features and biomarker from CMR images and validate at population level (Aim 1). We will then combine image information with multi-dimensional temporal EHR data to jointly identify clinically significant HFpEF subclasses (i.e. phenotyping) using state-of-art machine learning technique (Aim 2). Towards therapeutic goals based on phenotyping, we will further investigate optimal treatment strategies with current available agents using deep reinforcement learning (RL) based on massive EHR data to meet the pressing need before ongoing trials provide sufficient evidence on new drugs with proved clinical efficacy (Aim 3). Furthermore, we will develop an online, open- access platform to facilitating the sharing of code, data and knowledge of this study (Aim 4). We believe this research can improve our understanding, phenotyping and management of HFpEF, which might positively ease the clinical and economic burdens in turn both in U.S. and worldwide.