Comparing dynamic functional connectivity models in fMRI

1706 

Abstract Submission 

Yiming Wei¹, Stephen Smith¹, Mark Woolrich¹, Rezvan Farahibozorg¹, Stanislaw Adaszewski², Stefan Fraessle²

¹University of Oxford, Oxford, Oxfordshire, ²F. Hoffmann-La Roche AG, Basel, Basel-Stadt

Yiming Wei  
University of Oxford
Oxford, Oxfordshire

Stephen Smith  
University of Oxford
Oxford, Oxfordshire

Mark Woolrich  
University of Oxford
Oxford, Oxfordshire

Seyedeh-Rezvan Farahibozorg  
University of Oxford
Oxford, Oxfordshire

Stanislaw Adaszewski  
F. Hoffmann-La Roche AG
Basel, Basel-Stadt

Stefan Fraessle  
F. Hoffmann-La Roche AG
Basel, Basel-Stadt

Functional connectivity (FC) measures how different brain regions interact with each other. Recent research has increasingly focused on dynamic functional connectivity, which explores how these interactions change over time [1]. Despite various dynamic functional connectivity models in the literature, a comprehensive comparison of these approaches is absent. Here we conducted a systematic comparison of three models: Hidden Markov Modeling (HMM) [2], Dynamic network modeling (DyNeMo) [3,4] and Sliding Window Correlation (SWC) [5]. Our analysis encompassed a range of hyperparameters (number of input channels and number of distinct dynamic states), examining reproducibility and model goodness.

We utilized rfMRI data from the Human Connectome Project (HCP) S1200 release (N = 1003 subjects, four 15-min scans per subject) [6,7]. Processing included ICA-FIX denoising, surface alignment by MSMAll [8] and group-level spatial Independent Component Analysis (ICA) [9]. This process yielded spatial maps and the N_channels associated time series which are the input channels to dynamic modelling. The time series were z-scored for each 15min run. We then fitted HMM [2] and DyNeMo [3,4] (Fig 1) on these data via stochastic variational inference using the osl-dynamics repository [10]. The range of hyperparameters was specified as N_channels = {15, 25, 50, 100} and N_states = {4, 8, 12, 16, 20}. For SWC [5], we first applied a Butterworth filter (order = 16, high-pass cut-off frequency = 0.25 Hz), set the window_length = 143, step_size = 118 to calculate functional connectivity, and finally employed K-means clustering to determine the centroids of these matrices in FC space.

To assess model reproducibility, we divided all 4012 sessions randomly into two halves and trained our models on both splits, computed Fisher-z transformed correlation coefficients to quantify the similarity of states' functional connectivity matrices, and used the Hungarian algorithm to pair the model states obtained from different splits. Finally, the mean diagonal value of pair-reordered Fisher-z transformed correlation served as the measure of reproducibility. To evaluate model goodness, we conducted five separate model training sessions on the entire dataset and computed the average free energy.

As shown in Fig 2A-C, the reproducibility of FC matrices from HMM and SWC is satisfactory, while FCs in DyNeMo exhibit poorer reproducibility across all hyperparameter settings. For example, with N_channels = 50, N_states = 12 (Fig 2B, left), state 0 in the first half of DyNeMo displays a low correlation with all other states from the second half. Comparison across different numbers of channels and states in Fig 2A-C also demonstrates a trend: larger N_channels and smaller N_states yield better reproducibility.

Free energy decreases in the HMM as more states are introduced (Fig 2D). However, in DyNeMo, the free energy fluctuates as N_states increase (Fig 2E). This suggests the possibility of an optimal number of states for the free energy metric. It is also possible that there is instability in DyNeMo inference (for these rfMRI data). Moreover, the varying magnitudes of free energy across different models suggest that free energy may not be a reliable metric to select model type (HMM and DyNeMo) and N_channels. It should be noted that free energy is only available for generative models, and thus is not applicable to the SWC approach.

Our findings reveal that both HMM and SWC tend to produce states that are similar (within-method) to each other and exhibit superior reproducibility. Conversely, DyNeMo yields networks that are more distinct from each other and demonstrate greater representation capacity. As expected, larger numbers of states tend to explain the data better in different models, but this can come at the cost of lowering reproducibility.

Bayesian Modeling

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling ¹

Task-Independent and Resting-State Analysis ²

Computational Neuroscience

Data analysis

FUNCTIONAL MRI

Machine Learning

Modeling