Characterizing the inter-subject variability of physiological response functions

1689 

Abstract Submission 

Shiyu Wang¹, Roza Bayrak², Jingyuan Chen^3,4, Catie Chang^2,5

¹Department of Biomedical Engineering, Vanderbilt University, Nashville, TN, USA, ²Department of Computer Science, Vanderbilt University, Nashville, TN, USA, ³Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, USA, ⁴Department of Radiology, Harvard Medical School, Boston, MA, USA, ⁵Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, USA

Shiyu Wang  
Department of Biomedical Engineering, Vanderbilt University
Nashville, TN, USA

Roza Bayrak  
Department of Computer Science, Vanderbilt University
Nashville, TN, USA

Jingyuan Chen  
Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital|Department of Radiology, Harvard Medical School
Charlestown, MA, USA|Boston, MA, USA

Catie Chang  
Department of Computer Science, Vanderbilt University|Department of Electrical and Computer Engineering, Vanderbilt University
Nashville, TN, USA|Nashville, TN, USA

Low-frequency fluctuations in respiration volume and heart rate contribute to functional magnetic resonance imaging (fMRI) blood-oxygen level dependent (BOLD) signals across the brain[1]. Understanding the variability of how physiological factors associate with BOLD signals across subjects could help us better model their relationship. People have studied the inter-regional variability of the physiological response functions (PRFs) averaged across young adults [1] and the inter-subject differences of PRFs fitted to the global signal [2], but inter-subject differences in the full spatiotemporal pattern of PRFs remain to be quantified. Here, we characterized the range of whole-brain physiological response patterns that can be obtained across resting-state autonomic activities, using data that includes a variety of spontaneous breathing behaviors and heart rate levels.

We included 1500 3T resting-state fMRI scans (375 subjects, each with 4 scans) from the Human Connectome Project (HCP) under the HCP Minimal Preprocessing pipeline (TR=720ms; 1200 frames; voxel size=2mm isotropic). In addition, we detrended the signal, applied band-pass filtering (0.01–0.15Hz), downsampled by a factor of 2, extracted 497 cortical and subcortical ROIs using 4 atlases [3-6], and z-normalized the time courses.
Respiratory variation (RV) was defined as the standard deviation of a 6 s window of the respiratory belt recording centered at each time point. Heart rate (HR) was 1/(mean inter-beat-interval) of the pulse oximeter signal within the same windows. RV and HR were z-normalized, detrended, each convolved with 5 pre-defined basis functions [1], and fitted to each ROI's time course to obtain PRFs. Specifically, RRFs (respiratory response functions) and CRFs (cardiac response functions) were defined by multiplying the convolved RV or HR with their fitted β.
We used modularity detection to obtain whole-brain PRF patterns across scans. We concatenated the 497×2 PRFs (both RRFs and CRFs) for each scan and correlated the concatenated PRFs, forming a 1500×1500 correlational matrix. Next, we applied modularity detection on the correlational matrix using the brain connectivity toolbox [7]. The goal was to derive modules that maximize the within-cluster similarity while minimizing the across-cluster similarity [8]. A resolution parameter γ was introduced to overcome the resolution limit of modularity [9]. We tested γ values ranging from 0.9 to 1.2 with an increment of 0.01, with γ > 1 encouraging smaller clusters. The optimal γ was decided by selecting the most stable clustering results across all γ, where stability was measured by mean normalized mutual information [10].

The correlational matrix among the 1500 scans' PRFs were shown in Fig 1. The optimal γ value was 1.05, giving 11 clusters in total. There were 3 large clusters, 3 medium-sized clusters and 5 small clusters. Among the 375 subjects, at least 3 out of 4 PRFs from 4 scans were assigned the same cluster for 252 subjects, with only 12 subjects for whom all 4 PRFs were assigned to different clusters. Fig 2 shows the centroids of the whole-brain RRFs from 2 of the biggest clusters. The major difference between them was that cluster #8 had an initial dip before an increase in the RRFs, whereas cluster #10 peaked immediately after the physiological onset, which was more consistent with the previous population-level spatiotemporal dynamics of RRFs [1].

We observed several whole-brain PRF patterns that were shared across groups of subjects. Further, the cluster assignment of different scans within a given subject was relatively consistent. The observed inter-subject differences may arise, to some extent, from differences in subjects' breathing patterns during the scans. Our future steps include comparing PRFs across scans with matched breathing patterns, modeling respiratory and cardiac effects separately, and examining different physiological basis sets.

Exploratory Modeling and Artifact Removal ¹

Cerebral Metabolism and Hemodynamics

Neurophysiology of Imaging Signals ²

Autonomics

Data analysis

FUNCTIONAL MRI

Modeling