Computational Investigation Into Relationships Between Human Brain Signals and Behavioral States and Traits

Abstract

Functional Magnetic Resonance Imaging (fMRI) is a powerful technology for studying human brain activity in both healthy individuals and in patient populations. To most effectively use fMRI for neuroscience or healthcare, it is important to build a strong understanding of how fMRI measurements relate to an individual’s cognitive or behavioral traits as well as to their constantly shifting internal states, such as levels of wakefulness and drowsiness (“vigilance”). In one study, we investigate the temporal properties of fMRI signals, examining whether fMRI temporal complexity features exhibit stable markers of inter-individual differences and whether they relate to behavioral traits. We draw upon two widely used time-series complexity measures – a nonlinear complexity measure (Sample Entropy) and a spectral measure of low-frequency content – to capture dynamic properties of fMRI data. Our results indicate that these two measures are closely related, and that both generate reproducible patterns across brain regions across repeated scans. Using Canonical Correlation Analysis, we find a significant relationship between brain complexity features and behavioral/cognitive trait measures. Next, we mine the low-dimensional representations of brain signals to detect hidden vigilance states in unsupervised manner. Linear and nonlinear dimensionality reduction methods are applied to fMRI signals, followed by a clustering step using Gaussian Mixture Modeling. Our results show that the detected fMRI states align well with discrete vigilance states (high and low alertness) measured from simultaneously collected electroencephalography (EEG; a gold-standard vigilance measure) and behavioral reaction times. We also characterize the brain activation patterns underlying alert versus drowsy vigilance states. In addition, we move beyond discrete vigilance states to estimate continuous, slowly shifting EEG vigilance information from corresponding fMRI data by designing a multi-path variational autoencoder with regularization modules (DR-VAE). Our results indicate that the proposed method can successfully estimate the slowly shifting vigilance trends. Overall, this work makes key developments toward identifying and building the relationships between human brain signals and behavioral states and traits using complementary approaches.