Opportunistic Atherosclerotic Cardiovascular Disease Risk Estimation at Abdominal CTs with Robust and Unbiased Deep Learning

项目总结 动脉粥样硬化性心血管疾病(ASCVD)是全球发病率和死亡率的主要原因， 全国有1800多万成年人受到影响。然而，80%的ASCVD死亡可以通过及时的干预来预防 在对ASCVD风险进行早期筛查后--准确的亚临床需求未得到满足的有力理由 ASCVD诊断。因此，在这项研究中，我们评估基于深度学习(DL)的分析是否预先存在 腹部计算机断层扫描(CT)与电子病历(EMR)相结合可以提高预测能力 在一个大型多地点一级预防人群中，心血管死亡、心肌梗死和中风的风险。我们 将在大量不同的现实世界人群中进行这项研究，并进行外部验证，以确定 我们可以改进临床上使用的集合队列方程(PCE)，它有许多缺点。 在美国，每年有2000多万例腹部CT扫描。当这些扫描回答特定的临床问题时 问题，与心脏代谢风险相关的组织表型的定量信息很简单 未评估。DL算法可用于量化脂肪组织、肌肉、骨骼、 肝脏和血管钙化，这些都可以用来改善PCE以确定心血管疾病 事件。在我们提案的目标1中，我们将构建具有内置质量控制的自动分割算法 在125,000多个不同的受试者中提取这些身体组成指标的机制，以确定人口水平 组织大小和放射密度的标准值。在目标2中，我们将使用这些主体来增加PCE协变量 使用高级DL模型预测ASCVD风险的合成值和其他EMR功能。此外， 我们将设计新的算法方法，通过确保类似的模型性能来改善健康公平 跨PCE资格、种族/民族、保险类型和CT扫描仪制造商/型号的患者子组。在AIM 3、我们将建立一个新的ASCVD风险估计器，直接使用3D成像数据。我们将增强这一端到端功能 通过集成利用成像数据和电子病历数据的多模式模型来进行预测。实现 为了提高数字图书馆解决方案的可解释性，我们将构建每个主题的数字双胞胎来描述原因 目前正在进行模型预测，以及患者可以做出哪些改变来降低ASCVD风险。 我们将根据斯坦福大学的数据(25000名患者)训练所有模型，测试斯坦福大学的数据(8000名患者)，以及 使用三个Mayo诊所站点(20K以上患者)的数据对模型进行外部验证，以评估模型的泛化能力 我们的工具。我们已经组建了一个由DL专家、心脏病专家和腹部专家组成的跨学科MPI团队 放射科医生建立这样的ASCVD风险模型。我们开发创新的工具来提高精确度、通用性、 基于DL的ASCVD风险模型的偏倚和可解释性。我们的长期目标是能够及早发现静音 动脉粥样硬化和触发干预措施，最终可能防止超过80万人死亡，心肌梗死， 每年在不同的美国人中发生中风事件。

PROJECT SUMMARY Atherosclerotic cardiovascular disease (ASCVD) is the main cause of morbidity and mortality worldwide, and affects 18+ million adults nationally. However, 80% of ASCVD deaths may be prevented with prompt intervention following early screening for ASCVD risk – a powerful rationale for the unmet need of accurate subclinical ASCVD diagnoses. Thus, in this study we assess whether a deep learning (DL)-based analysis of pre-existing abdominal computed tomography (CT) scans paired with electronic medical records (EMR) improves prediction of cardiovascular death, myocardial infarction, and stroke in a large multi-site primary prevention population. We will conduct this study in a large, diverse, real-world population with an external validation to ascertain whether we can improve upon the clinically-utilized pooled cohort equations (PCE) that have numerous shortcomings. 20+ million abdominal CT scans performed annually in the US. While these scans answer specific clinical questions, quantitative information related to tissue phenotypes associated with cardiometabolic risk is simply not evaluated. DL algorithms can be used to quantify body composition metrics for adipose tissue, muscle, bone, liver, and vascular calcifications, which can all be used to improve upon the PCE for determining cardiovascular events. In aim 1 of our proposal, we will build automated segmentation algorithms with a built-in quality control mechanism to extract these body composition metrics in 125,000+ diverse subjects to ascertain population-level normative values of tissue size and radiodensity. In aim 2, we will augment the PCE covariates with these body composition values and additional EMR features for predicting ASCVD risk with advanced DL models. Moreover, we will devise new algorithmic approaches for improving health equity by ensuring similar model performance across patient sub-groups of PCE eligibility, race/ethnicity, insurance type, and CT scanner make/model. In aim 3, we will build a new ASCVD risk estimator that directly uses 3D imaging data. We will augment this end-to-end prediction approach by integrating multi-modal models that leverage both imaging data and EMR data. Realizing the need for improved explainability of DL solutions, we will build digital twins of each subject to describe why model predictions are being made and what changes a patient could make to lower ASCVD risk. We will train all models on data from Stanford (25k patients), test on data from Stanford (8k patients), and externally validate the models on data from three Mayo Clinic sites (20k+ patients) to assess the generalizability of our tools. We have assembled an inter-disciplinary MPI team of DL experts, cardiologists, and abdominal radiologists to build such ASCVD risk models. We develop innovative tools to improve accuracy, generalizability, bias, and explainability of DL-based ASCVD risk models. Our long-term goal is to enable early detection of silent atherosclerosis and trigger interventions that may ultimately prevent over 800,000 death, myocardial infarction, and stroke events in diverse Americans annually.