Multi-modal Prediction of Future Clinical Dementia

 描述（由申请人提供）：阿尔茨海默病（AD）是最常见的痴呆形式，其过程被认为在症状出现前开始多年。这个“临床前”阶段的特征是与AD一致的脑淀粉样蛋白积累的异常水平，这是确定病因和开发治疗策略的关键。在缺乏敏感和特异性的行为/认知测试的情况下，定量生物标志物和基因测试将是至关重要的， 临床前AD的分层用药该项目将研究两种高维数据模式，即结构性脑MRI扫描和全基因组SNP数据，以获得计算未来痴呆发作的个人水平预测的工具。AD在大脑上留下了一个独特的萎缩特征，在结构MRI扫描中可以识别。汇聚的数据表明，AD相关的萎缩是可检测的临床症状前几年。我们实验室在神经影像分析中提倡的机器学习（或模式分析）方法提供了高度敏感和特异性的萎缩检测器。我们假设这些工具将是非常宝贵的识别临床前AD受试者谁是在痴呆发作的风险增加。迟发性AD（LOAD）占所有AD病例的95%以上，高达70%是遗传性的。除了主要的遗传风险因子APOE 4之外，最近的全基因组关联研究（GWAS）已经确定了越来越多的与LOAD相关的其他常见遗传变异。LOAD基因基础的复杂性表明，将整个基因组的数据聚合起来的复杂模型可能有助于我们解释疾病进展中的一些变异性。使用最先进的机器学习技术开发这样的模型，并利用已经收集的大规模数据集，是我们这项提议的主要目标之一。拟议的项目将建立在首席研究员（Sabuncu）在计算建模和机器学习方面的强大背景基础上，使用尖端方法对大规模数据进行分析。我们将使用多模态数据，包括神经成像和GWAS数据，来开发和验证预测临床前负荷未来下降的模型。我们选择的方法将是一种新的贝叶斯ML算法，专门为纵向数据设计。我们假设，开发的模型将是更有用的替代品（通过区分病例和对照组构建），以确定“淀粉样蛋白阳性”的个人谁是在即将到来的临床下降的风险增加。我们将使用多层次的方法来发现和验证，并使用多模式策略来测试我们的假设。

 DESCRIPTION (provided by applicant): The process of Alzheimer's disease (AD), the most common form of dementia, is thought to begin years before symptoms. This "preclinical" phase, characterized by abnormal levels of brain amyloid accumulation consistent with AD, holds the key to identifying causes and developing therapeutic strategies. In the absence of sensitive and specific behavioral/cognitive tests, quantitative biomarkers and genetic tests will be critical for stratified medicine in preclinical AD. This project will examine two high-dimensional data modalities, namely structural brain MRI scans and genome-wide SNP data, in order to derive tools to compute individual-level predictions about future dementia onset. AD imprints a unique atrophy signature on the brain discernable in structural MRI scans. Converging data suggest that AD-associated atrophy is detectable years before clinical symptoms. The machine learning (or pattern analysis) approach, which our laboratory has advocated in neuroimage analysis, offers highly sensitive and specific atrophy detectors. We hypothesize these tools will be invaluable for identifying preclinical AD subjects who are at increased risk of dementia onset. Late-onset AD (LOAD), which represents >95% of all AD cases, is up to 70% heritable. In addition to APOE4, the major genetic risk factor, recent genome-wide association studies (GWAS) have identified a growing list of other common genetic variants associated with LOAD. The complexity of LOAD's genetic underpinnings suggests that sophisticated models that aggregate data across the genome might help us explain some of the variability in disease progression. Developing such models using state-of-the-art machine learning technology and leveraging already-collected large-scale datasets is one of our main aims in this proposal. The proposed project will build on the principal investigator's (Sabuncu) strong background in computational modeling and machine learning to conduct analyses using cutting-edge methods on large-scale data. We will use multi-modal data, including neuroimaging and GWAS data, to develop and validate models that predict future decline in preclinical LOAD. Our method of choice will be a novel Bayesian ML algorithm, specifically designed for longitudinal data. We hypothesize that the developed models will be more useful than alternatives (constructed by discriminating cases and controls) for identifying "amyloid positive" individuals who are at heightened risk of imminent clinical decline. We will use a multi-level approach for discovery and validation and a multi-modal strategy to test our hypothesis.