Meeting Banner
Abstract #3499

Unsupervised Clustering of FMRI Time Series with the Granger Causality Metric

Santosh B. Katwal1,2, John C. Gore2,3, Baxter P. Rogers2,3

1Electrical Engineering and Computer Science, Vanderbilt University, Nashville, TN, United States; 2VUIIS, Nashville, TN, United States; 3Biomedical Engineering, Vanderbilt University, Nashville, TN, United States

Unsupervised clustering methods such as Self-Organizing Map (SOM) or Hierarchical Clustering (HC) use the conventional Euclidean distance or correlation as the similarity metric to cluster data. The Euclidean distance cannot fully represent the noise points and correlation metric cannot efficiently detect small timing variability in fMRI time-series data. High field fMRI provides high signal-to-noise ratio (SNR) measurements. With high TR during acquisition, small temporal differences, down to 100 ms, can be resolved using the directed influence measure from the Granger causality approach. We use the Granger causality as a similarity metric in SOM or HC to cluster fMRI data with small timing variability.