A new stochastic graph embedding method for Alzheimer’s disease early‐stage prediction and intervention evaluation

• Published in: Alzheimer's & dementia Vol. 16
• Main Authors: Xu, Mengjia, Wang, Zhijiang, Zhang, Haifeng, Sanz, David Lopez, Garces, Pilar, Maestú, Fernando, Wang, Huali, Li, Quanzheng, Pantazis, Dimitrios
• Format: Journal Article
• Language: English
• Published: 01.12.2020
• DOI: 10.1002/alz.047329
• ISSN: 1552-5260

Background Subtle alterations of functional brain networks associated with Alzheimer’s disease (AD) are important for a quantitative characterization of neurodegeneration, however prior studies primarily focused on handcrafted, domain‐specific (ad‐hoc) graph features (). Here we developed a novel deep learning model that learns unsupervised brain network embeddings and automatically extracts AD‐related neural signatures. We assessed the robustness of the model in two downstream tasks, the prediction of conversion of mild cognitively impaired (MCI) patients to AD, and the evaluation of a multi‐domain cognitive intervention to amnestic MCI patients. Method We developed a graph Gaussian embedding method (MG2G) that uses a 3D encoder to learn intermediate representations through a sequence of hidden layers and outputs node‐wise low‐dimensional multivariate Gaussian distributions. Advantages of our model are that i) it discovers the intrinsic dimensionality of brain networks, ii) remaps brain data into a latent space amenable for supervised tasks such as AD classification, and iii) allows the use of the Wasserstein distance (W2) to define a metric space, and pinpoint subtle brain network alterations to specific brain regions. Result We used the MG2G model to embed MEG alpha band resting‐state network data from 48 stable MCI (S) patients, 28 progressive MCI (P) patients, and 53 age‐matched healthy elderly (N) subjects from the Madrid cohort. The obtained latent MEG brain network embeddings predicted AD progression as in Fig. 1. MG2G achieved higher performance than a completing model (node2vec) with 82% 3‐class classification, 93% 2‐class N/S classification, and 87% 2‐class S/P classification. We also computed functional brain networks from resting state fMRI data recorded from 12 aMCI patients before and after 12‐month of multi‐domain behavioral interventions. Using the W2 distance to quantify the probabilistic node‐wise embeddings in the latent space, we identified brain regions with intervention‐related functional alterations (Figs. 2 and 3). Conclusion MG2G provided a novel quantitative approach to assess complex functional connectivity patterns and learn highly‐informative network representations for different stages of AD while quantifying the uncertainty in the predicted outcomes. Acknowledgements: J‐Clinic for Machine Learning in Health at MIT; PSI2009‐14415C03‐01; PSI2012‐38375‐C03‐01; B2017/BMD‐3760; FJC2018‐037401‐I; Z161100000516001; D171100008217007.