Machine-learning on brain connectomics: Individual prediction of cognitive functioning in health and cerebral small vessel disease

这个项目的总体目标是开发、实施和评估计算模型，以预测健康受试者和脑部小血管疾病患者的脑连接技术对个体认知表现的预测。因此，我们解决了研究中的一个巨大差距，迄今为止，基于连接组的预测模型要么集中在年轻的健康成年人身上，要么集中在痴呆症的分类上，主要是阿尔茨海默病，而对于有风险或有症状前血管脑部变化的老年患者的脑行为关联研究工作很少。针对这一人口统计群体，我们将建立一个改进的机器学习框架，用于从神经成像数据中预测认知表型的样本外。这项工作将基于三个以人群为基础的大型队列(汉堡市健康研究：N=3,000,1000个大脑：N=1,200，英国生物库：N=48000)，它们提供了对认知功能的全面评估。重要的是，我们将单独和结合考虑不同的成像衍生指标(灰质体积、结构和功能连通性、白质高信号)，以得出每个样本和行为表型的最佳模型。这些将以仅使用社会人口信息和其他背景数据的模型为基准，以阐明成像对个人预测的附加价值。在临床环境中使用机器学习方法的关键挑战之一是，临床样本通常太小，无法进行模型训练和验证。这里我们提出了一种新的策略来解决这个问题，即元学习。关键的想法是利用这样一个事实，即大脑与行为的联系生活在一个有限的流形上，即大多数行为都是相关的，并与一组有限的神经生物学变异模式有关。因此，机器学习模型在一个大的队列上训练，从神经成像数据中预测特定的行为，也应该在一个较小的临床相关样本中捕捉到其他相关行为的信息。因此，在该项目的第二部分，我们将开发、应用和评估一种新的元学习策略，将在大人口样本中训练的模型转移到新的数据集，这些数据集反映了临床应用的关键结果衡量标准，但太小，无法对这些方面的预测算法进行稳健的训练。这一策略将在临床使用案例中进行测试，分别涉及三种不同的脑小血管疾病和中风患者样本。通过这一点，当前的提案将基于ConnectCome的个体特征预测的“大数据”战略结合在一起，重点是能够转移到临床使用案例，样本量有限。综上所述，我们的目标是促进脑连接技术在个体患者风险评估、表型预测和潜在临床诊断中的计算开发。

The overarching aim of this project is to develop, implement, and evaluate computational models for the prediction of individual cognitive performance from brain connectomics in healthy subjects and patients with cerebral small vessel disease. Hereby we address a substantial gap in research, connectome-based predictive modeling have to date focused either on young healthy adults, or on classification of dementia, mainly Alzheimer’s disease, while only little work has been done to study brain-behavior association in elderly patients at risk or with pre-symptomatic vascular brain changes. Focusing on this demographic group, we will establish an improved machine-learning framework for out-of-sample prediction of cognitive phenotypes from neuroimaging data. This work will be based on three large, population-based cohorts (Hamburg City Heath Study: N=3000, 1000Brains: N=1200 and UKbiobank: N=48000), which provide comprehensive assessments of cognitive function. Importantly, we will consider different imaging derived measures (grey matter volume, structural and functional connectivity, white-matter hyperintensities) alone and in combination to arrive at optimal models for each sample and behavioral phenotype. These will be benchmarked against models using only socio-demographic information and other background data to elucidate the added value of imaging for individual prediction. One of the key challenges for employing machine-learning approaches in clinical settings, i.e., translational application, is the fact that clinical samples are usually too small for model training and validation. Here we propose a novel strategy to address this problems, namely meta-learning. The key idea is to exploit the fact that brain-behavior associations live on a confined manifold, i.e., most behaviors are correlated and relate to a limited set of neurobiological modes of variation. A machine-learning model trained on a large cohort to predict a particular behavior from neuroimaging data should therefore also capture information about other, related behaviors in a smaller, clinically relevant sample. In the second part of this project, we will thus develop, apply and evaluate a novel strategy of meta-learning to transfer models trained in large population samples to new datasets that reflect critical outcome measures for clinical applications but are too small for robust training of predictive algorithms on these. This strategy will be tested in a clinical use-case involving three different samples of patients with cerebral small vessel disease and stroke, respectively. By thism the current proposal brings together a “big-data” strategy for connectcome-based prediction of individual traits with a focus on enabling a transfer to clinical use-cases with limited sample sizes. Taken together our goal is to foster the computational exploitation of brain connectomics for risk assessment, phenotype prediction and potential clinical diagnostics in individual patients.