Premium
A time‐embedding network model captures dynamic longitudinal pathology changes in a dominantly inherited Alzheimer disease population
Author(s) -
McCullough Austin A,
Gordon Brian A,
McDade Eric,
Bateman Randall J,
Morris John C,
Benzinger Tammie LS
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.046335
Subject(s) - psen1 , presenilin , population , neuroimaging , alzheimer's disease , disease , neuroscience , computer science , artificial intelligence , psychology , medicine , pathology , environmental health
Background Population‐level analyses of autosomal dominant Alzheimer disease (ADAD) cohorts have revealed a roadmap of neuropathological and volumetric changes that increase over disease‐time and culminate in severe cognitive impairment. Traditional statistical analysis approaches struggle to integrate these cohort‐level dynamics with patient data to inform disease trajectory predictions at the individual level. Current machine learning approaches excel at this integration of population and individual level information, but require much larger sample sizes than current Alzheimer disease research cohorts can provide. The current work introduces a time‐embedding neural network in order to accurately represent known disease‐time dependent patterns among neuroimaging variables. The network is trained to predict ADAD carrier status to demonstrate the ability of this framework to make actionable predictions within ADAD populations. Method 468 participants from families carrying gene mutations in presenilin 1 (PSEN1), presenilin 2 (PSEN2), or amyloid precursor protein (APP) enrolled in the Dominantly Inherited Alzheimer Network (DIAN) were selected for analysis. Structural MRI, [ 11 C] Pittsburgh Compound B (PiB) PET, and [ 18 F] Fluorodeoxyglucose (FDG) PET region of interest imaging data was utilized (Table 1). DIAN estimated years to symptom onset (EYO) was used as disease‐time. The time‐embedding network learns EYO‐dependent representations of all imaging data in both mutation carrier and non‐carrier groups. These representations are combined with control variables and passed to a decoder network trained to output carrier status predictions (Figure 1). Result The time‐embedding network captures three main longitudinal pathology dynamics that correspond to known longitudinal PiB PET, FDG PET, and volumetric pathology changes in ADAD populations (Figure 2A). The network was successfully trained to predict ADAD carrier status in individual participants (Figure 2B). Prediction accuracy increases as the time of prediction moves closer to onset of clinical symptoms (Figure 2C). Conclusion The time‐embedding network model is able to accurately represent nonlinear time‐dependent processes in a low‐dimensional, sparsely sampled embedding. This embedding provides an interpretable framework that is consistent with known longitudinal pathology changes when applied to an ADAD population, and can be trained to make accurate longitudinal predictions with currently available data sets.