Search Papers | Poster Sessions | All Posters
Poster C56 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
A scaling study of self-supervised auto-regressive modelling of fMRI time series and performance on downstream sex prediction in the UK biobank sample
Hao-Ting Wang1 (), Francois Paugam2, Nicolas Farrugia3, Pierre Bellec1,2,4; 1CRIUGM, Montreal, QC, Canada, 2Universitaire de Montreal, Montreal, QC, Canada, 3IMT Atlantique, Brest, France, 4Mila, University of Montreal, Montreal, QC, Canada
Limited data availability restricts how deep learning techniques can model patterns of brain activity. This study investigates the impact of training set size on the capacity of autoregressive graph convolutional networks (GCNs) to learn predictive features from fMRI data. GCN's graph structure, few parameters and interpretable features make it a good candidate to model complex brain dynamics. Using a large fMRI dataset, we assessed how dataset size impacts (1) autoregressive GCN's ability to predict future fMRI time points from past time points, and (2) the predictive value of GCN's learned features on a downstream sex prediction task. Our findings show performance saturation at a sample size of 8000 subjects for both the autoregressive and the downstream task, highlighting the model's ability to capture relevant brain signals. Standard deviation pooling from GCN layer weights emerged as the most predictive feature on the downstream task. These results motivate further exploration into more complex model architectures to achieve gains in performance.
Keywords: fMRI auto-regressive model functional connectivity foundation model