Machine Learning Based Comparison of Pearson's and Partial Correlation Measures to Quantify Functional Connectivity in the Human Brain

Abstract

Functional connectivity gives the statistical association or dependence between two or more distinct time series. Quantification of functional connectivity is normally done using Pearson's correlation coefficient, which measures the degree of co-activation of two different brain regions. But the brain does not function merely on pairwise relations. Brain functioning is based on interrelationships between several functional units simultaneously. Partial correlation is one such measure which considers these interrelationships. It quantifies the correlation between two distinct time series, but also removes the confound of the other correlations. This paper compares these two measures using functional magnetic resonance images in a machine-learning framework. Connectivity analysis and classification of autistic individuals from control population was done using these two measures. Classification accuracies were compared, with the conclusion that the measure which results in statistically significant accuracy has better predictive ability, and is better suited for fMRI functional connectivity modelling. It was experimentally found that Pearson's correlation coefficient gave better classification accuracy of around 2% than partial correlation measure.