Large-scale harmonization and integration of multi-modal ADNI data for the early detection of Alzheimer's disease and related dementias

点击翻译按钮获取中文摘要

Alzheimer’s disease (AD) and Alzheimer’s Disease Related Dementia (ADRD) are highly heterogeneous in pathology with mixed signatures on clinical biomarkers, making the early diagnosis challenging. Over the past few decades, large cohorts of multi-modal data have been collected to identify the interactions between these key pathologies. However, the utility of such cohorts has been compromised by the heterogeneity of the data collected from multiple sites and scanners, creating technical variability that can introduce noise and bias. Without comprehensive data harmonization and aggregation, these non-biological sources of variability can systematically bias the results of data-driven efforts in biomarker development. Our long-term goal is to identify specific AD and ADRD disease pathology markers and how they evolve. This project aims to improve the early detection of AD and ADRD so that future disease-modifying therapy can be allocated more efficiently to patients. To achieve this objective, we aim to harmonize trans-national cohorts of the Alzheimer’s Disease Neuroimaging Initiative (ADNI) to improve the diagnostic classification of AD and ADRD. The central hypothesis of our study is that by harmonizing the multi-modal American ADNI (versions 1, 2, 3, and GO) and Japanese ADNI datasets and building state of the art predictive models from each modality integrated into comprehensive ensembles, we can identify novel classifiers and features for early AD diagnosis and differentiation from ADRD. The central hypothesis will be tested by pursuing three specific aims: 1) Harmonization of multi-modal ADNI data, 2) Development of a suite of effective classifiers from diverse, harmonized ADNI data modalities, 3) Integration of multi-modal predictors into an ensemble model for AD/ADRD/healthy control classification, validation of the model in international ADNI cohorts, and sharing of the data and software products. We will pursue these aims by applying innovative computational approaches that combine traditional machine learning and more recent deep learning methods for unstructured neuroimaging and structured clinical data in ADNI. Moreover, we will leverage ensemble learning techniques to effectively combine models built from these diverse data modalities to optimize for robust classifiers of AD, ADRD, and the health status of patients. The results from this proposal will have a significant impact on better understanding the spatial dynamics and other mechanisms of AD and ADRD pathogenesis. Importantly, this project will create publicly available resources for multi-modal data harmonization and predictive modeling that can be used to explore further AD, ADRD, and other neurological disorders in future studies.