Advanced machine learning algorithms that integrate multi-modal neuroimaging to quantify the heterogeneity in Alzheimer's Disease

点击翻译按钮获取中文摘要

Abstract Alzheimer's Disease (AD) affects over 5 million Americans posing a significant burden to the community and health care system. Machine learning (ML) methods have been crucial in detecting the disease and characterizing its progression. Due to the lack of an in vivo “ground truth” diagnosis, ML approaches have typically relied on clinically derived labels and a case-control design in their search for a single imaging pattern that optimally distinguishes between the two groups in the case-control design. However, heterogeneity within clinical labels may degrade performance and interpretability. The goal of this project is to address this limitation and accurately characterize heterogeneity in preclinical and symptomatic AD. Given that age is a major risk factor for developing dementia, we will characterize healthy aging using multimodal neuroimaging data and ML in Aim 1. To this end, we propose to develop a novel unsupervised multi-view machine learning tool that can integrate information from multiple imaging modalities (i.e., structural Magnetic Resonance Imaging, and amyloid and tau sensitive Positron Emission Tomography) in a principled way. This will enable us to define the normal trajectory of age- related changes across all modalities, providing the necessary context to understand AD pathology. We will characterize AD pathology using multimodal neuroimaging data and ML in Aim 2. To this end, we propose to develop a novel semi-supervised ML framework that integrates multimodal information and derives data-driven disease dimensions. This is achieved by identifying and quantifying at the individual level imaging patterns that capture neuroanatomical and neuropathological alterations. Our approach builds on our extensive prior work on using an advanced, unsupervised multivariate pattern analysis technique, termed orthonormal projective non-negative matrix factorization, for analyzing neuroimaging data. Importantly, our project leverages two large multimodal datasets, the Knight AD Research Center (ADRC) cohort and AD Neuroimaging Initiative (ADNI), which sample participants across the continuum of AD making them ideal for investigating heterogeneity of AD pathology using advanced ML techniques. If successful, our approaches could be used for studying any brain disorder and could be readily integrated into personalized medicine strategies in the future when rich, multimodal imaging data collection will become a routine diagnostic procedure in hospitals.