Confound regression models for intersubject correlation analysis with naturalistic stimuli

843 

Abstract Submission 

Samuel Nastase¹, Uri Hasson¹

¹Princeton University, Princeton, NJ

Samuel Nastase  
Princeton University
Princeton, NJ

Uri Hasson  
Princeton University
Princeton, NJ

Intersubject correlation (ISC) analysis has become a workhorse method for measuring shared, stimulus-evoked neural activity in naturalistic paradigms (Hasson et al., 2004; Nastase et al., 2019). In ISC analysis, we use one subject's (or the average of many subjects') brain activity to model another subject's brain activity. This approach effectively isolates synchronized, stimulus-evoked responses and filters out idiosyncratic signals like head motion. How are ISC analyses affected by nuisance signals like head motion and physiological noise? Inspired by related efforts in the resting-state functional connectivity literature (e.g. Ciric et al., 2017; Parkes et al., 2018), we evaluate a variety of confound regression models for mitigating head motion and physiological noise.

To robustly evaluate different confound models, we used a large sample from the "Narratives" dataset (Nastase et al., 2021): fMRI data for 284 subjects listening to subsets of seven different naturalistic, spoken stories. We used fMRIPrep to minimally preprocess the data and extract a variety of different confound variables (Esteban et al., 2019). For simplicity, we limit our initial analyses to the average time series in early auditory cortex (EAC), which exhibits strong ISC in subjects listening to natural spoken language. We evaluated 20 different confound models ranging from no confound variables (model 0) to models comprising different combinations of head motion (HM), global signal (GS), white matter (WM) and cerebrospinal fluid (CSF) signals, spike censoring, as well as aCompCor and tCompCor components. We evaluated the confound models in two ways: (1) how does confound regression affect ISC in EAC? and (2) how does confound regression affect correlation between ISC and subject-level framewise displacement (FD)?

Overall, ISCs were fairly robust across confound models. No confound models reduced ISC, whereas several models increased ISC. Models containing HM, signals from WM/CSF, and GS performed comparably well. We observed that ISCs were negatively correlated with head motion (r = -.183 for model 0): that is, subjects with larger overall head motion had lower ISCs. Several models provide a good compromise: for example, the relatively simple model 13, comprising 6 head motion parameters and 5 aCompCor components from WM and CSF, both significantly improved ISC (from r = .356 to r = .474; t = 22.705, p < .001, FDR) and reduced the negative correlation between ISC and FD (from r = -.258 to r = -.146). We observed qualitatively similar results for other language ROIs (e.g. IFG).

Overall, ISC is fairly robust to head motion. While head motion can artificially inflate e.g. estimates of functional connectivity, ISC analysis is negatively correlated with head motion (more head motion yields lower ISCs). Several relatively simple confound models both improve ISC and mitigate the negative correlation between ISC and head motion.

Social Neuroscience Other ¹

Motion Correction and Preprocessing ²

Data analysis

FUNCTIONAL MRI

Language

Open Data

Social Interactions

Other - naturalistic stimuli;head motion;intersubject correlation;ISC;confounds