RT Journal Article SR Electronic T1 Predictive modeling of tau aggregation and propagation using PET and DTI JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 603 OP 603 VO 61 IS supplement 1 A1 Yang, Fan A1 Chowdhury, Samadrita A1 Jacobs, Heidi A1 Johnson, Keith A1 Dutta, Joyita YR 2020 UL http://jnm.snmjournals.org/content/61/supplement_1/603.abstract AB 603Objectives: The aggregation of misfolded tau protein is a neuropathological hallmark of Alzheimer’s disease (AD) and other tauopathies. Tau accumulation follows stereotypical anatomical stages which may be considered as spreading. Studies in animal models have shown that the pattern of tau spatial spread (which is associated with AD progression) is determined by neural connectivity rather than physical proximity between different brain regions. Our objectives are 1) to develop a mathematical model for tau aggregation and propagation along the structural network of the brain and 2) to apply it for longitudinal predictive modeling of tau. Methods: Underlying our model is the mathematical principle of diffusion on graphs, generally referred to as network diffusion. Network diffusion models have been previously used to model the spread of atrophy in neurodegenerative diseases. Most prior works use source-free models that rely on passive diffusion, a mathematical framework that is unable to model the simultaneous accretion and propagation of tau. Our framework features a closed-form solution to a partial differential equation with an exponential source term. In this work, we use longitudinal standardized uptake value ratio (SUVR) measures for tau from positron emission tomography (PET) using the 18F-Flortaucipir radiotracer. We also use structural connectivity based graph Laplacians computed from diffusion tensor imaging (DTI) by means of tractography and streamline counting. We assessed model fitting on 62 subjects (75.85 ± 6.18 years, 37 females) from the Harvard Aging Brain Study with serial, two-time-point Flortaucipir-PET scans and corresponding individual diffusion MRI scans for graph Laplacian computation. For independent validation of the predictive capability of this model, we subsequently applied the model to 10 additional HABS participants (73.67 ± 4.66 years, 2 females) with three-time-point Flortaucipir-PET scans and accompanying DTI data. For these three-time-point datasets, we used the first two time points for model fitting and the third time point for validation. Results: To assess goodness-of-fit and predictive accuracy, we compared the fitted/predicted tau with the observed tau in 10 critical anatomical regions-of-interest (ROIs) obtained via FreeSurfer parcellation. Since tau aggregation is a slow process, ROI PET SUVR averages across different time points tend to be highly correlated. We therefore also examined the goodness-of-fit and predictive accuracy for differential tau across time points. Our model substantially outperformed previously published models featuring standalone diffusion and diffusion with localized impulse sources in terms of coefficient of determination between observed vs. estimated differential tau. Conclusion: Combination of DTI and PET can offer a macroscopic perspective on tau propagation along neural pathways. Differential tau measures are highly susceptible to noise and are difficult to fit or predict. Longitudinal tracking of tau aggregation in several key ROIs is vital for understanding and predicting AD progression. Our model’s ability to predict tau aggregation in these critical ROIs is therefore of great significance in AD prognosis. Alongside the development of a novel mathematical framework, other key novelties of this work include personalized prediction of regional tau burden based on individual (rather than group-level) connectivity graphs, validation using longitudinal tau PET data, and accurate prediction of tau SUVR differentials.