Individual auto-regressive models for long-term prediction of BOLD fMRI signal
François Paugam, Guillaume Lajoie, Pierre Bellec, Université de Montréal, Canada; Basile Pinsard, Centre de Recherche de l'Institut Gériatrique de Montréal, Canada
Posters 3 Poster
Pacific Ballroom H-O
Sat, 27 Aug, 19:30 - 21:30 Pacific Time (UTC -7)
Auto-regressive models of functional neuroimaging time series capture the statistical structure of intrinsic brain dynamics. Auto-regression plays a key role in statistical mapping of the human brain, as a component of general linear models. Most fMRI auto-regressive models thus far have been proposed at the group level, ignoring individual idiosyncrasies. Densely sampled individual data have recently demonstrated large and systematic inter-individual variations in brain organization. In this work, we use a large naturalistic video watching dataset, featuring 19h of BOLD fMRI data per subject (N=6) to investigate individual-specific auto-regressive models of fMRI time series, benchmarking eleven popular machine learning models. Though all models had comparable performance on short-term predictions (1.49s), there was a clear distinction on longer-term predictions (8.94s) with an advantage to models that account for low frequency interactions through the use of longer input sequences. Non-linearities were found to be beneficial only for graph convolutional networks which were also the only ones to outperform standard linear regression. This work provides a benchmark of fMRI auto-regression at the individual level in ideal conditions, and can serve as a reference for future work on shorter time series and groups of subjects.