Description:

Speaker: Haochang Shou, PhD

Statistical Challenges for Data Integration in Large Neuroimaging Cohort Studies

With the increasing needs for big data analytics in medical imaging, integrating data from multiple study sites and various biological domains have become critical to better understand the complex human diseases. For example, large-scale observational studies often collected multiple modalities of measurements such as imaging, mobile health and survey data from samples recruited over several clinical centers. Analysis of such data have become particularly challenging due to the existence of site differences caused by unwanted technical variations that could mask the biological associations of interest, and that different data property of various modalities further poses difficulties in conducting integrative analyses to evaluate the joint relationship among modalities. In this talk, we will discuss several most recent developments of statistical harmonization methods in large neuroimaging studies under various data modalities.

Our approaches are designed to mitigate site differences that exist in mean, variance, and covariance structures in structural and functional imaging outcomes. We will then discuss a novel distance-based regression model, which we refer to as Similarity-based Multimodal Regression (SiMMR), that enables simultaneous regression of multiple modalities through their distance profiles. Our proposed method can detect associations in multimodal data of differing properties and dimensionalities, even with modest sample sizes. Our methods are motivated by iSTAGING (Imaging-based coordinate SysTem for AGing and NeurodeGenerative diseases) consortium, which brings together many multisite neuroimaging studies to understand the complex and heterogeneous process of normal aging as well as in Alzheimer’s disease.