Abstract
803
Objectives: The goal of this study is to infer the causality relationship between the accumulation of two hallmarks of Alzheimer’s disease (AD) pathology -A-beta and Tau- in different brain regions. A statistical technique is developed to infer the causality relationship between these quantities in different cortical and sub-cortical brain regions based on in vivo positron emission tomography (PET) data.
Methods: Eighty-seven cognitively normal elderly participants (age: 76.2 (6.2), M/F: 38/49) underwent PiB (A-beta deposits) and 18F-T807/AV1451 (Tau deposits) imaging acquisitions. PiB PET data was expressed as DVR and 18F-T807 as SUVR, and each individual brain image was partitioned into 594 isotropic voxels of 12mm. Using the PiB and 18F-T807 brain spatial variability of our cross-sectional sample, we apply a recently developed causality test [1] to infer the causal inference from one region to the rest of the regions. The causality test, known as the nonlinear causal discovery with additive noise models, contains two steps: Gaussian process regression and the contingency table independence test. Consider regions A and B, first, we perform Gaussian regression (based on the collected data from 40 subjects that are in the training set) to infer accumulated A-beta (or Tau) at region A as a (nonlinear) function of the accumulated A-beta (or Tau) at region B. Next, we apply contingency table independence test (based on the collected data from 31 subjects that are in the test set) to check whether the residual estimation error is independent of the accumulated A-beta at region B. For each region A, the strength of the arrow from region A to any other region B is log(p1/p2), where p1 and p2 denote the p-values of the tests corresponding to the causal connection from A to B and the one from B to A, respectively. A certain threshold is set to pick the arrows with higher strength.
Results: We found that Tau and A-beta data both display high resolution causality networks in the cognitively normal elderly brain. We report causality networks that are consistent with previously known histological data of Alzheimer propagation, as well as novel putative pathways of pathological spread. Representative inferred causality networks for A-beta and Tau data are shown in Figure 1. The causal connectionsfor the accomulation of Tau in sub-cortical regions is shown in Table 1. There is a causal relation from regions in the first column to the regions in the second column.
Conclusion: This study shows that nonlinear causal discovery is a powerful technique for inferring the causality network of A-beta and Tau, which in turn can be useful to investigate and monitor early pathological stages toward the progression of Alzheimer’s disease.