Maximizing Power in AD Clinical Trials via Multimodal Machine Learning

描述（由申请人提供）：正在进行的阿尔茨海默病（AD）研究的重点是确定在疾病进展的各个阶段最能预测未来认知能力下降的生物标志物。然后，这些生物标记物可以作为诊断的早期标记物，并用于选择临床试验的受试者。最近的结果表明，通过采用机器学习方法来识别这种歧视性生物标志物是可能的，但到目前为止，研究主要是孤立地使用模式。这些方法提供的敏感性/特异性在更多临床相关问题上并不令人满意：哪些MCI患者会转化为AD？回答这些问题需要综合利用所有数据源（如成像模式、CSF测量）的新方法。该项目的重点是如何从多个生物标志物的数据应该优化汇总，以最好地预测未来的认知能力下降，以及如何这些模型可以改善阿尔茨海默病的临床试验。假设：通过同时使用多种模式（以及纵向数据），可以显著提高个体受试者水平上区分AD、MCI和健康对照的敏感性和特异性。此外，这些方法将显著提高临床试验的样本量估计，并有助于获得评估新治疗程序的定制结果。具体目标：(1)开发新的基于图像的机器学习算法，该算法可以在统一框架内同时利用多种模式。(2)提供软件并在ADNI和BLSA数据集上广泛评估这些方法，以评估真正的多模态分析方法所能达到的敏感性/特异性。(3)将多模式分类方法与阿尔茨海默病临床试验相结合：(a)通过开发观察特定结果所需的综合样本量估计，并使用这些方法为威斯康星ADRC正在进行的r01资助的观察性/前瞻性研究得出定制结果。方法：我们将开发新的多模态机器学习方法，以最佳方式同时利用所有数据源。我们的模型还将纳入纵向数据，并利用疾病不同阶段的模式之间的相互作用。这将用于衍生多模态疾病标志物（MMDM）（目的1）。这些算法将在大规模特征良好的数据集上进行评估，并作为软件工具提供（目标2）。我们将通过两种方式使用这些模型来改进阿尔茨海默病临床试验：通过样品浓缩和定制结果，提供最大的统计能力来检测治疗效果（目标3）。意义：该项目利用了威斯康星ADRC在机器学习、统计临床试验设计、成像和阿尔茨海默病和阿尔茨海默病前期临床诊断方面的专业知识。这个项目将是第一个

DESCRIPTION (provided by applicant): An emphasis in ongoing Alzheimer's disease (AD) research is identifying those biomarkers which best predict future cognitive decline at the various stages of disease progression. These biomarkers can then serve as early markers for diagnosis, and for selection of subjects into clinical trials. Recent results suggest that the identification of such discriminative biomarkers is possible by adapting machine learning methods for this problem: but studies have primarily used modalities in isolation so far. The sensitivity/specificity offered by these methods is unsatisfactory for more clinically relevant questions: which MCI patients will convert to AD? Answering such questions requires new methods that leverage all data sources (e.g., imaging modalities, CSF measures) in conjunction. This project focuses on how data from multiple biomarkers should be optimally aggregated to best predict future cognitive decline, and how these models can improve clinical trials for AD. Hypothesis: Significant improvements in sensitivity and specificity for discriminating AD, MCI, and healthy controls at the level of individual subjects are possible by making use of multiple modalities (together with longitudinal data) simultaneously. Further, these methods will significantly improve sample size estimates in clinical trials, and help derive customized outcomes for evaluating new treatment procedures. Specific Aims: (1) To develop new image-based machine learning algorithms that can take advantage of multiple modalities simultaneously within a unified framework. (2) To provide a software and extensively evaluate these methods on the ADNI and BLSA datasets, to assess the sensitivity/specificity attainable by truly multi-modal analysis methods. (3) To interface multi-modal classification methods with AD clinical trials: (a) by developing comprehensive sample size estimates needed to observe specific outcomes, and using these methods to derive customized outcomes for an ongoing R01-funded observational/prospective study here at the Wisconsin ADRC. Methods: We will develop new multi-modal machine learning methods that will optimally exploit all data sources simultaneously. Our models will also incorporate longitudinal data, and exploit interaction between modalities at different stages of the disease. This will be used to derive a Multi-Modal Disease Marker (MMDM) (Aim 1). The algorithms will be evaluated on large-scale well-characterized datasets and provided as software tools (Aim 2). We will use these models to improve AD clinical trials in two ways: by sample enrichment and customized outcomes that provide maximum statistical power to detect treatment effects (Aim 3). Significance: This project capitalizes on the Wisconsin ADRC's expertise in machine learning, statistical clinical trial design, imaging, and clinical diagnosis of AD and pre-AD conditions. This project will be the first to implement a multi-modal machine learning metric specifically designed to speed up clinical trials so that potential therapies can be evaluated and an effective treatment arrived at as quickly as possible.