The effect of temporal observation selection on the prediction of visual stimulus from block design functional MRI

Multi-voxel pattern analysis is an approach to investigating brain activity measured by functional Magnetic Resonance Imaging (fMRI) in response to given stimuli. The signal acquired using fMRI is spatiotemporal, and can be used to predict the stimuli causing brain activation. Existing prediction methods suffer from the `curse of dimensionality´ by embedding all time points of the experiment in feature space. Although this problem can be alleviated by feature selection in spatial domain so that informative voxels are selected, feature selection in temporal domain has not been attempted. Henceforth, it is unclear which spatiotemporal combination of fMRI data gives the best prediction. In this study, we investigate the effect of using different combinations of fMRI time points on the prediction accuracy of visual stimuli, using support vector machine and random forest as classification methods. Using a publicly available fMRI dataset, we demonstrate that classification using multiple concatenated time points significantly outperforms a single time point based classification. Our results highlight the necessity of considering both temporal and spatial patterns to achieve better prediction of visual stimuli from fMRI data.