Machine Learning and Large-scale Imaging analytics for dimensional representations of brain trajectories in aging and preclinical Alzheimer's Disease: The brain aging chart and the iSTAGING consortium

摘要 大脑老化通常伴随着许多神经病理过程，通常是同时发生的 可能会导致认知能力下降和痴呆症。血管对认知损害和痴呆的贡献 (VCID)也非常常见，由于与心血管危险因素有关，可能 目前的治疗方法缓解了这种症状。很明显，AD相关痴呆(ADRD)的有效治疗将 要求在前驱和认知正常阶段及早发现脑部病理性改变。成象 方法提供了机会来研究在衰老和前驱阿尔茨海默病中出现的不同的大脑变化 这在以前是不可能的。描述这些多方面的大脑结构、功能和 病理学不仅提供了对潜在的病理生理过程的洞察，而且还具有新颖性 体内可预测的生物标志物。各种研究表明，神经退行性变的相对早期迹象 可以通过基于人工智能的模式分析和机器学习(PAML)方法来检测进程，并且 这些工具可以提供强大的预测性个性化预测面板。我们小组一直在 发展PAML方法的前沿，并将其应用于新的“基于成像的坐标系” 老龄化和神经退行性疾病“(ISTAGING)联盟，这是一项在 我们目前阶段的赠款，成功地汇集和协调了51,000多个磁共振成像设备和 来自11项研究和约34,000名个人的临床数据。我们的目标是捕捉大脑的异质性 通过应用我们的异质性分析PAML深度学习(DL)，随年龄和前驱AD变化 方法，这些方法有助于构建与不同大脑老化轨迹相关的成像模式。我们的 目标是丰富iSTAGING的不同维度，反映大脑变化的各种模式， 从而在可量化和可复制的指标中捕捉潜在的异质性。尽管我们会 将我们之前导出的rsfmri网络和淀粉样蛋白负荷的测量包括在拟议的工作中。 我们将专注于进一步剖析神经解剖的异质性，即提炼AT(N)中的“N”。 测量阿尔茨海默病神经变性变异性的框架和共同病理学的贡献，等等 使用这些中间的神经成像表型来预测认知能力下降和临床进展。 这将允许我们将每个人放入iSTAGING脑图，并绘制他/她的轨迹 以确定脑变化和认知功能衰退的预测指标。当前项目建立在 前一资金阶段的基础工作，并将这一独特的资源扩展到包括 研究重点是纵向数据、未得到充分代表的社会经济地位群体以及 各种并存疾病，包括高血压、糖尿病、肥胖、吸烟和睡眠障碍。这个 拟议的工作还将利用深度学习的最新发展，并将提供先进的方法 用于协调、异质性分析和预测性建模。

Abstract Brain aging is commonly accompanied by a number of neuropathologic processes, often co-occurring, that may lead to cognitive decline and dementia. Vascular contributions to cognitive impairment and dementia (VCID) are also extremely common and, due to associations with cardiovascular risk factors, may be mitigated with current therapies. It is clear that effective treatments for AD-related dementias (ADRD) will require early detection of pathologic brain change at prodromal and cognitively normal stages. Imaging methods offer the opportunity to study diverse brain changes present in aging and prodromal AD in ways that were previously impossible. Characterizing these multi-faceted aspects of brain structure, function and pathology not only provides insights into the underlying pathophysiological processes, but also novel predictive in vivo biomarkers. Various studies have shown that relatively early signs of neurodegenerative processes can be detected via AI-based pattern analysis and machine learning (PAML) methods, and that these tools can provide powerful predictive individualized panels of predictors. Our group has been on the frontier of developing PAML methods, and applying them to the new “Imaging-based coordinate SysTem for AGing and NeurodeGenerative diseases” (iSTAGING) consortium, a large-scale effort pursued in the current phase of our grant, which successfully brought together and harmonized over 51,000 MRIs and clinical data from 11 studies and ~34,000 individuals. We aim to capture the heterogeneity of brain change with aging and prodromal AD, by applying our heterogeneity analysis PAML deep learning (DL) methods, which help structuring imaging patterns associated with different brain aging trajectories. Our goal is to enrich the different dimensions of iSTAGING which will reflect various patterns of brain change, hence capturing the underlying heterogeneity in quantifiable and replicable metrics. Although we will include our previously derived measures of rsfMRI networks and of amyloid burden, in the proposed work we will focus on further dissecting neuroanatomical heterogeneity, i.e. on refining the `N' in the AT(N) framework to measure variability in AD neurodegeneration and the contributions of copathologies, and on using these intermediate neuroimaging phenotypes to predict cognitive decline and clinical progression. This will allow us to place each individual into the iSTAGING brain chart and map his/her trajectory, as well to determine predictive indices of brain change and cognitive decline. The current project builds on the foundational work of the previous funding phase, and expands this unique resource to include several studies focusing on longitudinal data, on groups of under-represented socio-economic status, as well as on various co-morbidities including hypertension, diabetes, obesity, smoking and sleep disturbances. The proposed work will also leverage recent developments in deep learning, and will offer advanced methods for harmonization, heterogeneity analysis, and predictive modeling.