Automated brain-age prediction and its interpretation

点击翻译按钮获取中文摘要

The biological age of a brain (brain-age) is a valuable summary of its structural or functional brain state. Importantly, elevated brain-age – the BrainAGE score – can serve as a biomarker to identify people with a higher risk of developing age-related diseases. However, brain-age cannot be measured and must be estimated, which calls for accurate methods based on non-invasive brain-imaging data. Although various machine learning (ML) based methods have been proposed for brain-age prediction, two crucial aspects remain notably understudied. Firstly, the impact of the manifold choices regarding data selection and representation or ML algorithms and their interaction with contextual aspects (e.g. restricted age range or a specific site) on prediction accuracy has not been systematically evaluated. As no single method can work well in all scenarios, there is a clear need to identify workflows (combinations of representation, ML algorithm and context) that may improve prediction accuracy. Secondly, localization of age-sensitive brain regions responsible for the predicted brain-age and in turn its deviation from the chronological age has only been coarsely explored. In addition, the impact of individual-specific factors (e.g. atypical aging in neurodegenerative diseases) remains poorly understood. In this work, we will address these issues by: (1) providing workflow design guidelines based on systematically evaluating a large number of workflows and contexts in a big-data framework, and (2) gaining neurobiological insights by testing and devising interpretation methods to localize age-sensitive brain regions and explain individual predictions.