Premium
Network diffusion model enhances predictions of future tau‐PET burden in Alzheimer’s patients
Author(s) -
Damasceno Pablo F.,
La Joie Renaud,
Shcherbinin Sergey,
Southekal Sudeepti,
Kotari Vikas,
Higgins Ixavier A.,
Collins Emily C.,
Rabinovici Gil D.,
Mintun Mark A.,
Raj Ashish
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.039480
Subject(s) - nuclear medicine , white matter , diffusion mri , mathematics , psychology , medicine , statistics , magnetic resonance imaging , radiology
Background Precisely measuring longitudinal tau PET is important for clinical trial design. Here we use a group of patients diagnosed with AD dementia (N = 28) and mild cognitive impairment (MCI) (N = 60) to investigate whether a network‐based pathological spread model could predict future tau accumulation based on regional standardized uptake value ratios (tau‐SUVR) and amyloid values determined at baseline. Method Baseline and 18‐month follow‐up flortaucipir and MRI images were analyzed from the Avid A05 phase 2 study (NCT 02016560, (Pontecorvo et al. 2019)). PET images were aligned to their respective Freesurfer‐conformed MRI images. Regional mean SUVR were extracted for 68 cortical regions using a white matter reference region (Southekal et al. 2018). The Network Diffusion Model (NDM) (Raj et al. 2012; 2015) was used to estimate longitudinal tau following a simple regression Δτ r = ‐mLτ r , where Δτ r is a vector of regional change in tau‐SUVR, the difference between tau‐SUVR at month 18 and baseline (Figure 2), L is the Laplacian matrix of the healthy‐based structural connectome, τ r is the regional tau‐SUVR vector at baseline (Figure 2), m is a linear model proportional to baseline features (see Figure 1B). For each model m, the parameters were fit via Ordinary Least Regression and the resulting R 2 was computed via bootstrapping. When fitting those parameters, predictions of Δτ r were made either per region (Figure 3A), averaging over all cortical regions (Figure 3B); or averaging over selected ROIs – estimated via the NDM to have Pareto optimal Δτ r (Figure 3C‐3D). Result For patients diagnosed with AD dementia, the explained variance increases as the number of regions decreases, independently of the model used for m (columns in Figure 3). The highest explained variance (R 2 = 0.69) was found when predicting mean Δτ over four regions: (L/R) Superior Temporal, (L/R) Lateral Occipital, and when m is a linear combination of 3 baseline features: total cortical amyloid, total cortical tau, and ratio between these two quantities. Conclusion Structural brain connectivity may contribute to future tau accumulation, indicating its potential impact in monitoring patient’s progression and the design of future clinical trials. Further validation using independent datasets is warranted.