Dynamic Functional Connectivity Analysis with Sliding-Window Correlations, Dimensionality Reduction and k-means Clustering

Code to accompany Spencer APC & Goodfellow M 2022 "Using Deep Clustering to Improve fMRI Dynamic Functional Connectivity Analysis" NeuroImage. doi: https://doi.org/10.1016/j.neuroimage.2022.119288

This code runs sliding-window correlations (SWC) on preprocessed node-averaged timeseries data, followed by an optional dimensionality reduction step using either PCA, UMAP or autoencoders, then k-means clustering to find dynamic functional connectivity states. Place raw timeseries data into the data/raw_tseries folder, with one .csv file per subject.

Run python3 main.py -h for help, or see run.sh for examples. To see a demonstration, use the create_toysimulation.m script in the SimTB_model_wrapper folder to generate synthetic data, then move this to the data/raw_tseries folder and run ./run.sh.

Python code was originally run with Python 3.6.8, using the following packages (and versions):

• numpy 1.18.5
• matplotlib 3.2.2
• tensorflow 2.2.0
• scipy 1.4.1
• sklearn 0.22.2
• umap 0.4.6
• pyclustering 0.10.1.2
• skggm 0.2.8

The SimTB wrapper runs in MATLAB and requires the SimTB toolbox (https://trendscenter.org/trends/software/simtb/). The code for the wrapper was adapted from that used in Allen et al., "Tracking whole-brain connectivity dynamics in the resting state." Cereb Cortex. 2014 24(3):663-76.

Modifications to the SimTB wrapper include:

• Generating batches of synthetic data, representing data from multiple heterogeneous subjects. Between-subject differences in HRF shape, noise parameters, and underlying state time courses.
• State FC matrices are randomly generated for a given dataset.
• Underlying state time courses are randomly sampled from a hidden Markov model.