PT  - JOURNAL ARTICLE
AU  - Balaji, Vibha
AU  - Song, Tzu-An
AU  - Yang, Fan
AU  - Jacobs, Heidi
AU  - Johnson, Keith
AU  - Dutta, Joyita
TI  - <strong>A graph neural network model for the prediction of longitudinal tau aggregation</strong>
DP  - 2022 Aug 01
TA  - Journal of Nuclear Medicine
PG  - 2233--2233
VI  - 63
IP  - supplement 2
4099  - http://jnm.snmjournals.org/content/63/supplement_2/2233.short
4100  - http://jnm.snmjournals.org/content/63/supplement_2/2233.full
SO  - J Nucl Med2022 Aug 01; 63
AB  - 2233 Introduction: Alzheimer’s disease (AD) is neuropathologically characterized by the presence of tau neurofibrillary tangles and amyloid-β (Aβ) plaques. Tau spread patterns in the brain have been linked to structural connectivity profiles. We present here a machine learning model for making individualized predictions of tau spread in the brain from baseline imaging, connectomics, and demographic data. Our model relies on a graph neural network (GNN) trained at a population level which yields individual-level predictions for the spatial distribution of tau at follow-up. We assess the impact of different graph-domain regularization schemes on the GNN’s predictive accuracy. Specifically, we investigate the relative accuracies of (i) an unregularized GNN (GNN-NoReg), (ii) a total-variation-regularized GNN (GNN-TV), and (iii) a physics-informed GNN (GNN-Phy) for the longitudinal predictive modeling of tau. Methods: Our GNN model learns spatiotemporal evolution patterns of tau from longitudinal training data samples. Accuracies are computed in a separate validation dataset. We used baseline and follow-up tau PET standardized uptake value ratio (SUVR) measures based on the 18F-Flortaucipir radiotracer. Baseline tau PET SUVRs at 66 FreeSurfer regions-of-interest (ROIs) structured as a node signal vector were provided as an input to the GNN. Amyloid positivity status (a binary variable) determined from Pittsburgh Compound-B (PiB) PET imaging and age were provided as additional “global” inputs to the GNN. The graph edge inputs were based on structural connectivity and was derived from diffusion tensor imaging (DTI). The GNN output consists of a node signal vector representing predicted annualized differentials in the tau PET SUVR (hereon referred to as ADSUVR). The data used here are from the Harvard Aging Brain Study (HABS). We used longitudinal (two-timepoint) data from 163 subjects (73.471 ± 8.406, 102 females, 40 PiB+), which were split into 112 training (73.736 +/- 8.821, 112 total, 70 females, 30 PiB+) and 51 validation (72.887 +/- 7.379, 51 total, 32 females, 10 PiB+) samples. We compared a standalone (unregularized) GNN with two types of regularized GNNs: a GNN based on a graph-domain total variation penalty and a physics-informed GNN that uses the Fisher-Kolmogorov-Petrovski-Puskinovand (Fisher-KPP) reaction-diffusion equation to model tau spread along the structural network. Results: To assess the relative performance of these approaches, we computed the root-mean-square error (RMSE) of the ADSUVRs as our figure of merit over 10 selected ROIs where the tau quantitation is considered critical for preclinical AD: amygdala, parahippocampal gyrus, posterior cingulate cortex, temporal pole, hippocampus, superior temporal gyrus, middle temporal gyrus, inferior temporal gyrus, fusiform gyrus, and entorhinal cortex. In the validation dataset, RMSEs observed across 10 ROIs compared as follows: GNN-Phy (0.06358 ± 0.03170) < GNN-TV (0.06381 ± 0.03334) < GNN-NoReg (0.064411 ± 0.03151). Conclusions: Individualized forecasting of the spatiotemporal trajectory of tau is vital for the diagnosis of preclinical AD. Our results demonstrate that tau spread patterns can be predicted from an individual’s baseline tau distribution, amyloid positivity status, and age using a GNN model that relies on structural connectivity. Furthermore, we show that regularization techniques, particularly physics-based regularization, can further enhance GNN accuracy.