Optimizing Artificial Neural Network Models to Predict Brain-Age from Functional MRI

Abstract

Large MRI datasets combined with deep learning methods have realized a new state of the art for brain-age prediction. Age prediction may serve as a valuable biomarker for brain health and disease given that over-estimated age based on MRI (usually as predicted by a machine learning model; sometimes called a "brain-age gap") has been associated with neurological and psychiatric disorders. However, most of these results have been achieved via the use of high-resolution structural (T1w) MRI scans. Brain-age prediction via deep learning over large volumes of functional MRI (fMRI) data is less well studied, but could help form a bridge between neural health biomarkers observed in MRI and more portable platforms like functional near infrared spectroscopy (fNIRS), which measure a hemodynamic signal similar to fMRI. In this work, we studied how to optimize deep learning model architectures and training pipelines to predict brain-age from resting state fMRI connectivity data. A wide set of pre-processing and model hyperparameters was explored that included varying the number of nodes and the composition of the input functional connectivity matrices, the size, depth and objective functions of the neural network models, and a time series sub-sampling method as a data augmentation strategy. Model performance was evaluated on both an internal validation set of held-out participants (from the multi-study corpus compiled for training), as well as numerous external corpora not seen during training, which comprised healthy controls and clinical participants. Neural network models with a variety of hyperparameter configurations supported accurate brain-age prediction using fMRI and many models generalized effectively to predict the age of healthy individuals among data sets not seen during training (< 8 years mean absolute error on the external validation dataset). However, we report mixed results regarding a brain-age gap for held out clinical populations using these methods, with a gap observed only among neurodegenerative disorders (here, Alzheimers disease), and not among psychiatric disorders or patients with traumatic brain injury. This work constitutes a valuable step towards scalable, portable brain-age prediction but highlights a number of areas where additional work and improvements are needed.