Human functional connectome fingerprinting using resting-state dynamic functional connectivity

1406 

Abstract Submission 

Amin Ghaffari¹, Yufei Zhao¹, Jason Langley¹, Xu Chen¹, Xiaoping Hu¹

¹University of California, Riverside, Riverside, CA

Amin Ghaffari  
University of California, Riverside
Riverside, CA

Yufei Zhao  
University of California, Riverside
Riverside, CA

Jason Langley  
University of California, Riverside
Riverside, CA

Xu Chen  
University of California, Riverside
Riverside, CA

Xiaoping Hu  
University of California, Riverside
Riverside, CA

Resting-state functional connectivity (FC) networks have distinct, personalized connectivity patterns that could act as a unique fingerprint for individual identification (Kazeminejad, A. (2019)). Prior work on functional connectome fingerprinting relied on static FC patterns and considered a single functional connectivity map across the whole data collection session for subjects for individual identification (Finn, E.S. (2015)). However, the brain is a dynamic system that switches between multiple metastable states rather than staying in a single state (Fox, M.D. (2005)), and each of the metastable states could represent a different FC map that has unique information for each individual. Based on this understanding, incorporating dynamic FC information may improve identification accuracy in FC fingerprinting.

In this paper, we use resting state data from the Human Connectome Project (HCP) with a sample size of 100 subjects to evaluate the performance of our dynamic states-based brain fingerprinting method. Regions of Interest (ROIs) are extracted based on the 268-node functional parcellation map of (Shen, X. (2013)). Using the time series of each ROI and a sliding time window of 20 time points, we develop a framework for the analysis of dynamic functional connectivity of the subjects and corresponding individualization. Initially, at each time window, we calculate the Pearson correlation coefficients between pairs of regions and obtain the functional connectivity maps. By subtracting the mean of all of the FC maps across all subjects and time windows, we obtain the demeaned version of FC networks. We consider each brain to go into 3 dominant connectivity states within the scan session, and utilize the K-Means clustering methods twice to obtain these states. To achieve this, first, we apply a data reduction technique on the FC maps in which we sum all correlation coefficients for each ROI and call that the correlation strengths (Ou, J. (2013)). Then, we perform K-Means (K=30) clustering on correlation strengths and calculate the centers of clusters for each individual. Next, we place all of those clusters' centers from all subjects into a pool of correlation strengths and acquire the main dominant states by computing the 3 main cluster centers. By calculating the correlations between each individual's correlation strength at each time window with the dominant brain states, we obtain the average representation of the 3 states for each subject. Finally, by normalizing and concatenation of those representations, we achieve a final vectorized representation for the individuals, and using the Pearson correlation coefficient between these vectors, we identify subjects' scans from the rest of the group.

Our results demonstrate a notable enhancement in identification accuracy with dynamic states-based fingerprinting. More specifically, incorporating dynamic functional connectivity patterns for identification resulted in an accuracy of 100% when identifying day 2 scans based on day 1 data, and 98% when identifying day 1 scans using day 2 data, surpassing the performances based on the whole-brain static functional connectome, which is below 95% (Finn, E.S. (2015)).

Our results indicate that for more accurate individualization of the human functional brain map, it is beneficial to consider the effects of dynamics in the resting state network. Considering the average functional connectivity map reduces the differentiation power between subjects and is associated with more misidentified samples because it neglects the inherent dynamics of the functional connectivity pattern of the brain. In contrast, looking at representations of different brain states by each subject has a better "fingerprinting" ability and therefore is a better method of identification.

Classification and Predictive Modeling ¹

Connectivity (eg. functional, effective, structural) ²

fMRI Connectivity and Network Modeling

Task-Independent and Resting-State Analysis

Data analysis

FUNCTIONAL MRI

Statistical Methods

Other - Brain Fingerprinting, Functional Connectivity