A0641
Title: Longitudinal prediction of brain markers using Gaussian process deep kernel learning
Authors: Haochang Shou - University of Pennsylvania (United States) [presenting]
Abstract: Longitudinal prediction of brain markers is of high importance for early diagnosis and prognosis for better understanding and differentiating complex human diseases such as ageing and Alzheimer's disease (AD). The aim is to propose to predict the longitudinal changes of the structural MRI markers by leveraging multimodal features and spatial dependency across brain regions. Using the expressivity of deep kernel learning with Gaussian processes (GP), a personalized and reliable prediction model is presented for noisy, longitudinal data, which can provide individually tailored predictions of longitudinal biomarkers. The model includes the population component that captures the global trend across a set of diverse subjects and the personalized component that adapts the predictions using the history of each subject. The proposed model utilizes a deep neural network to learn complex global trends from a large number of patients and Gaussian processes (GP) to probabilistically quantify the uncertainty of the predictions and model the individual trends of each subject. The model is evaluated on multiple diverse and heterogeneous longitudinal imaging cohorts, and volumetric data for individual regions of interest (ROIs) is predicted, as well as machine learning-based composite scores for brain atrophy. The predicted trajectories are further demonstrated to differentiate between subjects with different disease progression statuses.
