Longitudinal predictive modeling for tau in Alzheimer's disease

PROJECT SUMMARY Alzheimer’s disease, the most common cause of dementia in the elderly, is characterized by a cognitively asymptomatic preclinical stage which is identified and monitored via longitudinal tracking of pathophysiological biomarkers, e.g., tau and amyloid. Since the aggregation of tau protein tangles in the medial temporal lobe is a key driver of memory impairment, accurate image-based longitudinal prediction of tau burden could fill a critical gap in biomarker development for preclinical Alzheimer’s disease. Tau tangles exhibit stereotypical neuroanatomical patterns of spatiotemporal spread that correlate strongly with the progression of neurodegeneration. Studies in animal models have suggested that the characteristic patterns of tau spread associated with Alzheimer’s progression are determined by neural connectivity rather than physical proximity between different brain regions. Graph-theoretic methods that utilize macroscale structural connectivity mapping in humans to predict future tau burden could lead to valuable prognostic tools for Alzheimer’s disease. The overarching research goal of this R01 Research Project Grant is to develop an interpretable machine learning model that uses individual structural connectomics to make personalized predictions of differential measures of tau from multimodal baseline data. Our approach relies on longitudinal 18F-Flortaucipir positron emission tomography (PET) for the imaging of tau tangles, 11C-Pittsburgh Compound B (PiB) for the imaging of amyloid plaques, and high-angular-resolution diffusion magnetic resonance (MR) imaging for individualized structural connectomics in human subjects. We will develop a physics-informed and interpretable graph neural network to predict the annual rate of change of the regional tau burden from multimodal inputs, including baseline tau, Aβ, and an array of structural connectivity metrics. We will also develop novel physics-based analytic models for tau progression, which will be used to effectively guide the machine learning framework. Finally, we will apply the machine learning model to investigate the earliest cortical site of tau aggregation, to examine the connectomic basis of early tau spread, and to leverage our model’s interpretability to discover and validate novel connectomic biomarkers to characterize preclinical Alzheimer’s disease. To validate the machine learning model, we will use serial tau PET data at two and three timepoints from the Harvard Aging Brain Study, one of the largest longitudinal imaging resources for preclinical Alzheimer’s disease. To ensure scientific rigor, secondary validation of the models will be performed using data from the Alzheimer’s Disease Neuroimaging Initiative database. The proposed personalized predictive model could significantly impact preclinical Alzheimer’s prognosis, facilitate ongoing clinical trials, and shed light on the neuroconnectomic and biological underpinnings of Alzheimer’s disease.