Measuring shared responses across subjects using intersubject correlation.

Nastase SA, Gazzola V, Hasson U, Keysers C. · 2019

Read Measuring shared responses across subjects using intersubject correlation. by Nastase SA, Gazzola V, Hasson U, Keysers C. (2019) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6688448.txt.

Introduction

Nastase and co-authors give a tutorial on intersubject correlation for shared brain responses. Their article appears in Social Cognitive and Affective Neuroscience. One person's brain activity can model another's. That trick measures shared information across brains. It works during naturalistic, dynamic experiences. No set response model is required. Speaker-listener coupling becomes measurable this way. Perception and recall can be compared head-on. Standard fMRI designs struggle with such richness.

The tutorial develops the method's logic. Questions the approach can answer are mapped. Spatial patterns and network estimation extend it. Pitfalls in method receive thorough treatment. Best practices are spelled out in full. The authors aim at access for newcomers. Social neuroscience is the target field.

Catalog record

CREATORNastase SA, Gazzola V, Hasson U, Keysers C.
YEAR2019
COLLECTIONPAPERS
FORMATTXT
FILE SIZE103 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC6688448

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Measuring shared responses across subjects using intersubject correlation By Samuel A Nastase, Valeria Gazzola, Uri Hasson, Christian Keysers Published in: Social Cognitive and Affective Neuroscience DOI: 10.1093/scan/nsz037 ABSTRACT Abstract Our capacity to jointly represent information about the world underpins our social experience. By leveraging one individual’s brain activity to model another’s, we can measure shared information across brains—even in dynamic, naturalistic scenarios where an explicit response model may be unobtainable. Introducing experimental manipulations allows us to measure, for example, shared responses between speakers and listeners or between perception and recall. In this tutorial, we develop the logic of intersubject correlation (ISC) analysis and discuss the family of neuroscientific questions that stem from this approach. We also extend this logic to spatially distributed response patterns and functional network estimation. We provide a thorough and accessible treatment of methodological considerations specific to ISC analysis and outline best practices. INTRODUCTION Traditional methods for fMRI data analysis are not conducive to studying the multidimensional dynamics that characterize social interaction in real-life contexts. Methodological constraints require relatively brief, isolated stimulus events or tasks, accompanied by a pre-defined model of the expected neural response. Brain areas involved in a particular function are localized by contrasting neural responses to tightly controlled stimuli varying along a few isolated parameters of experimental interest. As a result, many of the core questions of social and affective neuroscience have proven difficult to study (Zaki and Ochsner, 2009; Hasson and Honey, 2012; Adolphs et al., 2016). For instance, narrative comprehension is triggered by complex situations that unfold over minutes and cannot be captured in brief epochs, while face-to-face social interactions additionally involve a multitude of communication channels such as words, sentences, intonation, facial expressions and gestures (Hasson et al., 2012). Predicting fluctuations in brain activity during these dynamic, continuous episodes is difficult. Finally, the social and affective symptoms of patients with psychiatric disorders may only be revealed in open, complex situations that cannot be boiled down to experimental paradigms with brief, disjoint events (Klin et al., 2002). Intersubject correlation (ISC) analysis provides complementary insights to traditional analyses by circumventing the need for a pre-defined response model and allowing experimenters to measure the consistency of neural responses to complex, naturalistic stimuli across individuals (Hasson et al., 2004, 2010). Beyond simply measuring response reliability, ISC analyses allow us to measure shared content across experimental conditions. By capitalizing on the richness of naturalistic experimental paradigms, ISC has the potential to empower the investigation of social interactions. This tutorial situates ISC among related methods and extends the logic of ISC to spatially distributed response patterns and functional connectivity. We provide illustrative examples of how ISC analyses can be used to address novel questions and put special emphasis on methodological and interpretational considerations. SITUATING ISC AMONG TRADITIONAL METHODS Traditional analyses of functional magnetic resonance imaging (fMRI) data follow a simple conceptual framework. During experimental design, we generate at least two conditions that differ according to some variable of experimental interest: one may be thought to trigger a particular function while the other is as similar as possible without triggering that function, or the conditions may vary parametrically along a single variable such as retinotopic eccentricity. In fMRI, noise dominates signals at frequencies lower than 0.04 Hz, so when presenting stimuli intended to evoke a particular function, instances of each condition typically range from brief ‘events’ (tens of milliseconds to several seconds) to ‘blocks’ about 20 s in duration (Boynton et al., 1996; Chen and Tyler, 2008). We assume that neural activity is roughly constant within each instance of a condition and that a brain region is involved in a function, or tuned to an experimental variable, if its activity increases in response to the condition where the variable of interest is present or increased in magnitude, relative to a condition where the variable of interest is not present (i.e. the control condition) or lesser in magnitude. These are typically referred to as subtraction (Friston et al., 1996) and parametric (Büchel et al., 1998) designs, respectively. These designs lend themselves to generating predictions about the hypothesized time courses of neural activation. The hypothesized time courses serve as predictors in a general linear model (GLM) that quantifies how well the expected time course predicts activity observed in each voxel, thereby localizing the function of interest (Friston et al., 1994). Fluctuations in brain activity over time within a condition or across repetition are considered noise, while the difference across conditions is the signal. Because each predictor typically comprises multiple trials of the same condition and we assume that neural activity is identical across trials of the same condition, this approach effectively collapses across trials (i.e. trial averaging; Dale and Buckner, 1997). This approach is powerful whenever (i) the function can be recruited in short epochs, (ii) tightly controlled stimuli can be generated to isolate and manipulate the parameter of experimental interest and (iii) we have detailed and exhaustive hypotheses about the time course of relevant brain activity. Cutting edge modeling efforts (e.g. Huth et al., 2016) suffer from similar constraints. For example, when using word embeddings to predict brain activity during narrative comprehension, each occurrence of a word receives the same embedding regardless of the overarching narrative. In real-life scenarios, where the response to each token changes as a function of an evolving narrative context, such trial-averaging methods will fall short (Ben-Yakov et al., 2012). ISC analyses provide a complementary, data-driven alternative for identifying brain regions with activity driven by the stimuli or paradigm. The core idea is best illustrated for subjects listening to a spoken story. If multiple subjects listen to the same story, brain regions that are systematically driven by the story will fluctuate synchronously across viewers, while brain regions that do not process the story in the same way across subjects, or are not responsive to the story at all, will not. For example, a voxel in early auditory cortex will consistently track the low-level auditory features of the spoken words across all viewers. The response time course of this voxel will be highly correlated across subjects. On the other hand, a region of the brain that is not entrained by the story (e.g. one involved in low-level visual or motor processing) will not yield a consistent response time course across subjects. Finally, regions that respond to the story in a way that varies temporally to some extent across subjects, for instance because they are involved in emotional reactions to the story that evolve somewhat idiosyncratically from subject to subject, will show intermediate correlations, particularly in the lower frequency range (see Box 1). In summary, correlating brain activity across subjects while they are exposed to a complex stimulus reveals brain areas that process the stimulus in a consistent, time-locked manner. Correlations approaching 1 indicate that the region encodes information about the stimulus and that this information is processed in a stereotyped way across individuals, while correlations approaching 0 reflect regions with idiosyncratic processing or encoding little information about the stimulus. This logic can be meaningfully applied to specific frequency bands of the signal (Box 1). If we study the processing of features of the soundtrack that fluctuate rapidly, we would look for correlation across viewers in higher

Read the complete text