On Time-frequency Feature Selection Method for Neural Imaging Analysis With Small Sample Size

In most functional studies in neuroimages, such as electro-encephalography (EEG) and functional magnetic resonance imaging (fMRI), only time-average characteristics were extracted from the time series of signals in region-of-interest (ROI) or links between ROIs, which implies that temporal sequential information in the images may be lost. Therefore, provided with a small sample size, this sort of methods are incapable for significant statistic detection for a large load of family-wise error rate (FWER) control. In this paper, we propose a novel approach for difference detection of data of time series between groups. By taking the time-frequency features into considerations and employing the Fisher's pooling method, our approach demonstrates a significant enhancement of statistical power, particularly for a small size of data but strict FWER control. The simulation model shows that it can greatly reduce the false positive rate with a minor loss of false negative rate. We employ our approach to two sets of experimental data: EEG of schizophrenia subjects and resting-state fMRI for anxiety subjects. It is shown that our approach performs better to identify statistically significant spatial characteristic, such as ROI and link of pairs of ROIs, between patient and healthy control groups. Moreover, this approach enables to identify the significant frequency-band feature in the group comparison.