Trustworthy Disease Discovery: A Case Study in Alzheimer’s

Abstract

A small number of high-dimensional training samples is a challenging problem for understanding the dynamics of multivariate spatiotemporal neuroimaging. Often, reducing these dynamics to a small number of handcrafted features helps. For example, the matrix of Pearson's correlation coefficients is highly predictive. Nevertheless, it is hard to perceive the dynamics and the disorder from these compressed proxy representations. In this paper, we propose a hierarchical recurrent model with attention that learns dynamics directly from temporal signals and captures stable interpretations of abnormal conditions predictive of disorder under consideration. We study these abnormalities in dynamics through feature importance estimation in the resting-state functional MRI data using different interpretability methods. We validate this feature estimation by introducing the Retain and Retrain (RAR) process and demonstrate its utility on an Alzheimer's disease dataset. Furthermore, we show that the proposed model is adaptable to small sample case by offering a self-supervised pretraining scheme of the same model. With this scheme, we demonstrate that the model can leverage a large unrelated but publicly available dataset to learn improved representation to maintain adequate predictive capacity and extract useful disorder-specific information.

Tasks

Reproductions