Advanced machine learning algorithms that integrate genomewide, longitudinal MRI and demographic data to predict future cognitive decline toward dementia

摘要 阿尔茨海默病(AD)的“临床前”阶段的特点是大脑水平异常 淀粉样蛋白在没有主要症状的情况下堆积，可以持续数十年，并可能持有 成功治疗策略的关键。今天迫切需要定量的生物标记物。 以及可以在个体水平上预测临床进展的基因测试。这个项目将会发展 尖端机器学习算法，将挖掘高维、多模式和 纵向数据，以得出在以下情况下产生个人水平临床预测的模型 痴呆症。开发的预测模型将特别利用无处不在和负担得起的数据 类型：结构性脑MRI扫描，唾液或血液来源的全基因组序列数据，以及 人口统计变量(年龄、教育程度和性别)。先前的研究已经证明，所有这些 变量与临床痴呆症的下降密切相关，然而到目前为止我们还没有模型。 可以获取嵌入在这些高维数据中的所有预测性信息。 机器学习(ML)算法越来越多地被用来计算来自高风险的临床预测 多维生物医学数据，如临床扫描。然而，大多数先前的最大似然方法都是为 ‘’预测‘’任务是关于并发条件的应用程序(例如，区分情况和 控制)；已确定的风险因素(如年龄)、多种模式(如基因和图像)和 纵向数据没有得到充分利用。该应用程序的核心创新将是开发 严谨、灵活和实用的ML方法，可以充分利用多模式、纵向和高 计算预后临床预测的生物医学数据维度。 拟议的项目将建立在PI在计算建模和分析方面的强大背景之上 大规模生物医学数据。我们将采用创新的贝叶斯ML框架，以提供 灵活地处理和利用真实生活中的纵向和多模式数据。我们假设 开发的模型将比替代基准更有用地识别临床前 面临即将出现临床衰退的高风险的个人。我们将使用一个统计上严格的 用于发现、交叉验证和对开发的工具进行基准测试的方法。这个项目将 提供免费分发、记录和验证的软件和模型，用于预测未来的临床 基于全基因组、纵向结构磁共振成像和人口统计学数据的进展。我们相信 我们开发的算法和软件将为临床前AD的分层提供宝贵的工具 受试者在药物试验中，优化未来的治疗方法，并将不良反应的风险降至最低。

ABSTRACT The “preclinical” phase of Alzheimer’s disease (AD) is characterized by abnormal levels of brain amyloid accumulation in the absence of major symptoms, can last decades, and potentially holds the key to successful therapeutic strategies. Today there is an urgent need for quantitative biomarkers and genetic tests that can predict clinical progression at the individual level. This project will develop cutting edge machine learning algorithms that will mine high dimensional, multi-modal, and longitudinal data to derive models that yield individual-level clinical predictions in the context of dementia. The developed prognostic models will specifically utilize ubiquitous and affordable data types: structural brain MRI scans, saliva or blood-derived genome-wide sequence data, and demographic variables (age, education, and sex). Prior research has demonstrated that all these variables are strongly associated with clinical decline to dementia, however to date we have no model that can harvest all the predictive information embedded in these high dimensional data. Machine learning (ML) algorithms are increasingly used to compute clinical predictions from high- dimensional biomedical data such as clinical scans. Yet, most prior ML methods were developed for applications where the ``prediction’’ task was about concurrent condition (e.g., discriminate cases and controls); and established risk factors (e.g., age), multiple modalities (e.g., genotype and images) and longitudinal data were not fully exploited. This application’s core innovation will be to develop rigorous, flexible, and practical ML methods that can fully exploit multi-modal, longitudinal, and high- dimensional biomedical data to compute prognostic clinical predictions. The proposed project will build on the PI’s strong background in computational modeling and analysis of large-scale biomedical data. We will employ an innovative Bayesian ML framework that offers the flexibility to handle and exploit real-life longitudinal and multi-modal data. We hypothesize that the developed models will be more useful than alternative benchmarks for identifying preclinical individuals who are at heightened risk of imminent clinical decline. We will use a statistically rigorous approach for discovery, cross-validation, and benchmarking the developed tools. This project will yield freely distributed, documented, and validated software and models for predicting future clinical progression based on whole-genome, longitudinal structural MRI and demographic data. We believe the algorithms and software we develop will yield invaluable tools for stratifying preclinical AD subjects in drug trials, optimizing future therapies, and minimizing the risk of adverse effects.