Predicting the Longitudinal Spread of Atrophy in Neurodegenerative Disorders

Post by Shireen Parimoo

What's the science?

Progressive neurodegenerative diseases like Alzheimer’s disease and frontotemporal dementia (FTD) are thought to result from the spread of misfolded proteins throughout the brain, eventually leading to neuronal loss and atrophy. The spread of brain atrophy typically follows a distinct pattern over the course of each disease, giving rise to a variety of behavioral symptoms. For example, AD pathology is characterized by the spread of misfolded tau protein that begins in the entorhinal cortex, a region of the brain linked to memory function. The pathology then spreads from this “epicenter” to other anatomically and functionally connected areas (i.e. its network). However, few longitudinal studies have investigated individual differences in the spread of atrophy from an epicenter to its network in neurodegenerative diseases. This week in Neuron, Brown and colleagues used network-based modeling of structural and functional magnetic resonance imaging (MRI) scans to predict longitudinal atrophy in patients with progressive neurodegenerative diseases.

How did they do it?

Structural MRI scans were obtained from 72 patients diagnosed with the behavioral variant of frontotemporal dementia (bvFTD) and the semantic variant of progressive primary aphasia (svPPA), as well as from 288 age-matched controls. Patients were scanned twice: once at baseline and again about a year later. Gray matter volume (GMV) was estimated in each voxel of the control participants’ scans and compared with patient scans, producing a GMV atrophy map for each patient that identified regions of relative atrophy. A subset of the control participants also completed a task-free functional MRI scan that was used to generate functional connectivity (FC) maps, which contained brain regions that were co-activated. To do this, the authors specified 192 cortical areas as seed regions and identified other co-activated brain regions, resulting in 192 FC maps for each participant. These maps were then averaged to produce a group FC map for each cortical seed region. For each seed region’s FC map, the authors correlated the FC values in each voxel with the atrophy in each patient’s GMV map. The cortical seed region whose FC map was most highly correlated with GMV atrophy was chosen as the epicenter for that patient. For example, if the anterior temporal lobe’s FC map was most highly correlated with a patient’s GMV atrophy map, then the anterior temporal lobe was chosen as that patient’s epicenter.

To identify the factors underlying the spread of atrophy from the epicenter, the authors specified a generalized additive model with baseline atrophy, the shortest path length to epicenter, and nodal hazard of each functionally connected region in a network as predictors. The shortest path length is the shortest distance between the epicenter and its functionally connected nodes. The nodal hazard is a measure of how much a node is at risk of atrophy based on the degree of atrophy present in its 5 functionally connected neighbours, with higher values suggesting a greater risk of atrophy. To determine model accuracy, the authors correlated patients’ actual atrophy in the follow-up structural scans with their predicted atrophy values. A cut-off correlation value of r = 0.23 was used to sort accurate (r >= 0.23) and inaccurate (r < 0.23) predictions.

What did they find?

Distinct epicenters of atrophy were observed across the two patient groups, including the anterior cingulate cortex and the frontoinsular cortex among those with bvFTD, and primarily the anterior temporal lobe in patients with svPPA. The spread of atrophy throughout the brain was also unique in each patient group. In bvFTD patients, atrophy progressed to the posterior cingulate cortex, precuneus, inferior parietal lobule, posterior inferior temporal cortex, and the dorsolateral prefrontal cortex. On the other hand, atrophy in svPPA patients spread to the orbitofrontal cortex, posterior temporal lobe, the anterior cingulate cortex, and the mid-cingulate cortex.

The spread of atrophy was predicted by the shortest path length to epicenter, nodal hazard, and baseline atrophy of brain regions. In particular, regions closer to the epicenter had the greatest atrophy over time, whereas regions farther away from the epicenter did not show much change in atrophy. Similarly, regions with higher nodal hazard values (i.e. more atrophied neighbouring regions) had greater longitudinal atrophy than regions with low nodal hazard values (i.e. fewer atrophied neighbours). Interestingly, the relationship between baseline and longitudinal atrophy showed an inverted-U pattern, whereby regions with intermediate levels of atrophy showed the greatest atrophy over time compared to regions with low or high levels of baseline atrophy. The predicted spread of atrophy correlated with the actual spread of atrophy over time (r = 0.64) and the model accurately predicted longitudinal atrophy for 59 out of 72 patients in the study. Thus, there was high spatial overlap in the model’s predictions of atrophy and the actual atrophy observed in the patient scans one year later.

What's the impact?

This study is the first to identify patient-specific epicenters of gray matter atrophy in bvFTD and svPPA and predict their longitudinal spread of atrophy. There is often considerable heterogeneity in both the behavior and the neuropathology associated with neurodegenerative diseases, and the network-based approach used to characterize the spread of pathology in this study has important implications for accurately predicting individual trajectories of disease progression.

Brown et al. Patient-tailored, connectivity-based forecasts of spreading brain atrophy. Neuron (2019). Access the original scientific publication here.