Combining fast and slow fMRI sampling rates can enhance predictive power in resting-state data.

Data collection technology in functional magnetic resonance imaging (fMRI) is rapidly developing, leading to continuous growth of spatio-temporal resolution. The need to understand brain dynamics, as it plays a crucial role in understanding brain function, continues to push innovation in this direction as limits on the frequency of data measurement limit the kinds of questions that may be asked. In parallel, researchers continue to amass large volumes of fMRI data using the highest sampling frequencies available with current technology. A common and plausible assumption is that higher measurement frequencies may lead to more informative data about the brain dynamics and help mitigate physiological noise from neurovascularly coupled signal. This assumption leads to the tendency to discard the older datasets collected with lower temporal resolution in favor of more recent collections. Moreover, as we will show, it leads to under-utilizing the current MRI technology by only collecting at the fastest available rate. A recent theoretical study demonstrated that combining high frequency data with data collected at a deliberately slower sampling rate can, in some conditions, lead to gains in information about the dynamics. We hypothesize that similar effects can be observed in fMRI datasets where data is collected at multiple timescales, as opposed to datasets created by subsampling from a single acquisition rate. A resting state fMRI dataset collected from 10 subjects at a slow (2150 ms) and fast (100 ms) repetition time (TR) is analyzed, demonstrating informative gains in predictive power by combining the two. This gain is in contrast to diminishing returns in the single TR dataset performance, where the data has been manually-undersampled to a slower sampling rate and combined with the original. Performance outcomes were also compared in gender prediction across a multi-rate dataset and single rate dataset, with multi-rate results showing gains in composite features. Our experiments demonstrate agreement with the theoretical results in showing that features formed as a combination of slow and fast sampling rates yield greater predictive power than features from either slow or fast rates alone in some settings.