FASTIMAGES: Validating replay detection methods in human neuroimaging using a combined MEG and fMRI dataset
bioRxiv
Abstract
Studies in rodents and humans using invasive electrophysiology have established that neural replay is a ubiquitous brain phenomenon associated with memory, planning, and decision-making. However, invasive recording in humans remains difficult, limiting our understanding of replay in the human brain. To comprehensively study replay in humans, reliable non-invasive detection methods are needed. Several non-invasive approaches have been proposed, but none have been fully validated against known ground-truth signals. In this study, we present FASTIMAGES, a benchmark dataset comprising parallel fMRI recordings from 40 participants and MEG recordings from 30 participants. The dataset contains known neural sequences evoked by fast visual stimulation alongside functional localizer trials. Neural sequences were generated using five visual stimuli presented at onset-to-onset intervals of 132, 164, 228, and 612 milliseconds. Using this dataset, we evaluate two sequence-detection methods: Temporally Delayed Linear Modelling (TDLM) and Slope Order Dynamic Analysis (SODA). TDLM was originally developed for MEG by Liu et al. (2021), whereas SODA was developed for fMRI by Wittkuhn and Schuck (2021). We examine the assumptions underlying each method and assess their strengths and limitations when applied to MEG and fMRI data. Both methods perform best in their native modality, with TDLM excelling for MEG and SODA for fMRI. Under idealized benchmark conditions, the two approaches achieve comparable effect sizes. Transferring these methods across modalities remains challenging. The FASTIMAGES dataset provides clearly defined neural sequences that can serve as ground truth for benchmarking and validating future sequence-detection methods under idealized conditions.
