Paper 3: Introduction
Exploring the potential use of individualized network approaches in studies of emotion processing
Since its introduction in the early 1990’s, functional magnetic resonance imaging (fMRI) has become a central pillar of research aimed at understanding the inner workings of the human brain. By examining changes in the blood-oxygen-level-dependent (BOLD) signal, which serves as a proxy for neural activity, researchers are able to identify and map neural substrates of human cognition. The focus of much fMRI work has been to characterize neural correlates of core cognitive capacities in the average human brain. As the field has grown, fMRI approaches have evolved from characterizing average patterns of task-related activation to exploring how these patterns of activation vary between individuals as a function of other factors such as the presence of psychopathology, age, and socioeconomic status (a.k.a., individual differences). While individual differences work has illuminated much about brain-behavior relationships, a series of recent studies have revealed that many common fMRI tasks exhibit poor test-retest reliability (Elliott et al., 2020; Fröhner et al., 2019; Nord et al., 2017; Flournoy et al., 2023). Test-retest reliability—the maintenance of rank order of subjects across time—is essential to studies of individual differences (Revelle & Condon, 2019), as such approaches assume constructs used to study differences across people reflect stable phenomena. The low test-retest reliability of neural activation observed in many canonical fMRI tasks calls into question the legitimacy of using task-based fMRI activity as a construct in individual differences studies (Elliott et al., 2020; Fröhner et al., 2019; Nord et al., 2017; Flournoy et al., 2023). Identifying sources of variability that underlie the poor test-retest reliability of task-based fMRI activity is a necessary first step towards validating individual differences approaches.
Reliability of Neural Correlates of Emotion Processing. One of the tasks at the center of studies revealing poor test-retest reliability of fMRI is a canonical emotion processing task (the task used in Project 2) (Elliott et al., 2020; Fröhner et al., 2019; Nord et al., 2017; Flournoy et al., 2023). The task was originally designed to characterize amygdala function in the average human brain (Hariri et al., 2000, 2002). Using this and other similar tasks, researchers have found robust amygdala activation to threat-relevant emotional faces, including faces exhibiting fear or anger, relative to non-emotional cues (Hariri et al., 2002; J. Morris, 1998; J. S. Morris et al., 1998; Whalen et al., 1998). Though this early work resulted in an emphasis on the role of the amygdala in threat detection, follow-up work has shown that the amygdala also responds to a broad range of expressions of affect, including positively-valenced emotions (Kim et al., 2003; Somerville et al., 2004; Winston et al., 2003; Yang et al., 2002). These findings led to an expanded view of the amygdala as a region that supports general detection of biologically salient stimuli (Whalen et al., 1998; Sander et al., 2003). Given the role of the amygdala in the detection of emotional salience, amygdala reactivity was identified as a possible biomarker of risk for psychopathology (Herman and Cullinan, 1997; Kim et al., 2003; Murty et al., 2010; Pessoa and Ungerleider, 2004; Swartz et al., 2015). By using this canonical task and other similar emotional face processing tasks, researchers discovered heightened amygdala reactivity in individuals exposed to major life stress (Dannlowski et al., 2013; McCrory et al., 2011; van Harmelen et al., 2013; Weissman et al., 2020), as well as in individuals with psychopathology (Yurgekun-Todd et al., 2000; Thomas et al., 2000; Etkin & Wager, 2007; Whalen et al., 2002; Groenewold et al., 2013; Sicorello et al., 2020; Simmons et al., 2011). Furthermore, prospective studies have shown that heightened amygdala reactivity predicts psychopathology onset following severe life stress (McLaughlin et al., 2014; Stevens et al., 2016), as well as typical life stress (Swartz et al., 2015). Given the breadth of work identifying connections between amygdala reactivity and behavioral health, numerous massive data collection efforts aimed at linking brain function to behavior and health outcomes have included this now canonical emotion processing task in their task batteries (eg. Human Connectome Project, the Dunedin Study, and UK Biobank) (Van Essen et al., 2013; Poulton, Moffitt, & Silva, 2015; Bycroft et al., 2018). Though this canonical emotion processing task has become central to a wide body of work aimed at understanding individual differences in amygdala reactivity, the amygdala activity in this task (the task target) exhibits particularly poor test-retest reliability (Elliott et al., 2020; Flournoy et al., 2023). These findings call into question the validity of using the amygdala as a target in studies of individual differences using this task. Identifying alternative targets for examination may provide an opportunity to continue to explore factors that underpin individual differences in emotion processing.
Seeking Alternative Targets. The primary aim of the current work is to evaluate whether current analytical approaches used to examine activity in this canonical emotion processing task might be contributing to poor reliability estimates due to sources of noise that could be reduced with other analytic choices. A major source of variability in the fMRI signal is noise and given its position above the sinus cavities, the amygdala is one region that is particularly affected by a poor signal to noise ratio (LaBar et al., 2001). As the amygdala has been the primary focus of studies of emotion processing, evaluations of test-retest reliability have focused on task-evoked activity within the amygdala (Elliott et al., 2020; Fröhner et al., 2019; Nord et al., 2017). The broad contrast used in this task (threat-relevant faces versus geometric shapes) only eliminates signal due to basic visual processing, capturing a wide range of higher-order processing involved in perceiving emotional faces, including face processing, identity processing, emotion processing, as well as detection of emotional salience. Perhaps unsurprisingly, this task elicits wide-spread cortical activity given the broad contrast used (Elliott et al., 2020; Hariri et al., 2002; Somerville et al., 2018). Expanding the characterization of brain activity beyond the amygdala to include a broader range of cortical regions that are recruited by the task, but are less impacted by poor signal to noise ratios, may provide the opportunity to identity alternative targets that are better situated than the amygdala for use in studies of individual differences in emotion processing.
While shifting to examining cortical responses may serve to improve reliability estimates for this task, standard methods used to isolate cortical activity such as using meta-analytically defined peak activations or average task activation maps, to define a priori boundaries at a group level, suffer from a series of major limitations due to the group-based nature of these approaches (Fedorenko, 2021). A growing body of work is revealing that the precise location of functional activity in response to task demand varies substantially across individuals (Frost & Goebel, 2011; Tahmasebi et al., 2012; Vazquez-Rodriguez et al., 2019; Tavor et al., 2016). For example, the location of the fusiform face area (FFA), a region that shows notable activity in response to the canonical emotion face processing task, varies notably along the length of the fusiform gyrus between subjects, despite the fact that the gyri themselves are consistently aligned across different individuals (Frost & Goebel, 2011). These idiosyncrasies are overlooked when group-averaging methods are applied, and result in missed or diminished effects given poor overlap in signal across idiosyncratic functional activity between subjects (Nieto-Castañón & Fedorenko, 2012; Chen et al., 2017). So while shifting the focus of this task away from the amygdala to cortical activity may serve to improve the reliability of the signal, careful consideration should be made to how the task-based activity can be quantified while limiting potential sources of measurement error.
Individualized Networks. Advances in systems neuroscience, aimed at examining the functional organization of the human brain, have revealed that the human cortex is organized into a series of networks (Gordon et al., 2016; Power et al., 2011; Yeo, et al., 2011) that are functionally discrete (Braga et al., 2020; DiNicola et al., 2020) and highly idiosyncratic in their topography across individuals (Braga et al., 2019; Braga & Buckner, 2017; DiNicola et al., 2020; Gordon et al., 2017; Kong et al., 2019, 2021; Laumann et al., 2015; Tavor et al., 2016). Work examining the idiosyncratic nature of these networks has resulted in analytical methods that produce individualized parcellations that have subsequently been used as a priori boundaries to examine task-evoked activity at the level of the individual (Braga et al., 2020; DiNicola et al., 2020; Gordon et al., 2017). This work has shown improved correspondence between network boundaries and task activity at the level of the individual in a wide range of tasks including a basic finger tapping task (Gordon et al., 2017), the emotion processing task currently in question (Chong et al., 2017), in language tasks (Braga et al., 2020), as well as high-order episodic projection and theory of mind tasks (DiNicola et al., 2020). Many of the areas that exhibit activity during the canonical emotion processing task fall within regions that comprise discrete higher-order association networks, including the medial prefrontal cortex, superior temporal sulcus, and posterior cingulate—primary nodes in the default network, and the insula—a major node in the salience network (Fusar-Poli et al., 2009; Pozzi et al., 2021; Sabatinelli et al., 2011). Given the seeming overlap of task-active cortical regions during the canonical emotion processing task with cortical networks, examining network-level task-evoked activity by employing individualized networks may be an alternative approach to examining emotion processing activity in the cortex. Employing individualized parcellations, as opposed to standard meta-analytic ROIs, will shift the unit of analysis from a few small, isolated group-based regions to distributed idiosyncratic networks encompassing large extents of the cortex. While signal from single small group-based regions may exhibit low reliability, larger swatches of activation aggregated across distributed regions of the cortex may prove more reliable. So, shifting to an individualized network-level approach, and capturing a greater extent of the task-evoked signal by using network boundaries that have been shown to have strong correspondence to task activity, may serve to increase reliability of this signal.
In sum, current analytical approaches used to examine activity during fMRI tasks, including a canonical emotion processing task, suffer from a series of limitations that may serve to increase noise in the task-evoked signal resulting in reduced reliability. Limitations of the present task in question include emphasis on the amygdala, a region that suffers from high signal to noise ratio, reliance on group-defined ROIs that have low sensitivity to individual differences in task-evoked activity, and finally reliance on a very small portion of the overall signal.
Examining Test-Retest Reliability of Individualized Networks. The goal of the present study is to determine whether activity within individualized network boundaries is a more reliable marker of task-evoked activity during emotion processing. In other words, do the individualized networks produce activity values that maintain more consistent rank order across subjects over time, when compared to the activity within the amygdala or group-based regions of interest on the cortex? To answer this question I examine the test-retest reliability of emotion processing activity within the amygdala and a set of meta-analytically defined ROIs (Sabatinelli et al., 2011), networks in a group-based network parcellation (Yeo et al., 2011), in comparison to individualized network parcellations (Kong et al., 2018). To assess the test-retest reliability across these boundary approaches I will calculate intraclass correlations (ICCs) values of the BOLD signal for the contrast of Faces > Shapes in each of the boundaries in a dataset where subjects completed the canonical emotion processing task monthly over the course of a year, resulting in 12 task completions (Chen et al., 2021, Koo and Li, 2016; Shrout and Fleiss, 1979). Similar between-subject reliability measures have been used to examine test-retest reliability of task-evoked activity in this task previously (Elliot et al., 2020; Flournoy et al., 2023). By examining the test-retest reliably of task-evoked activity across different approaches (I.e., ROIs versus individually identified networks), Ii will be able to determine whether the individual networks provide any improvement to the test-retest reliability in this canonical task.
Associations of Neural Correlates of Emotion Processing with Stress and Psychopathology. Following the test-retest reliability analysis I will examine whether experiences of life stress and symptoms of psychopathology (depression and anxiety) are meaningfully associated with neural activity during emotion processing in networks defined within the individual. As before, I will also examine activity in the amygdala, as well as the left and right fusiform (the meta-analytic ROIs that exhibited significant activity to Faces > Shapes). This analysis will provide the opportunity to explore whether individualized network-level activity is associated with key factors that contribute to individual differences in emotion processing.
Within-Person Fluctuations. In a recent study, Flournoy et al., 2023 illustrated that poor test-retest reliability observed in emotion processing tasks may not be the result of measurement error, but may reflect meaningful within-person fluctuations in activity over time. They found that roughly half of the variance in the BOLD signal could be accounted for by within-person variability—reliable fluctuations of the BOLD signal over time within individuals (Flournoy et al., 2023). These findings suggest that the emotion processing activity observed in this task could reflect a more state-level, as opposed to trait-level phenomenon. Flournoy et al., found that within-person fluctuations in activity evoked by an emotion processing task were significantly associated with fluctuations in sleep, mood and experiences of stress. These results suggest that activity in the individualized networks could be expected to correlate with fluctuations in stress and psychopathology — not just across individuals but also within individuals over time. We will therefore examine whether within-individual fluctuations in stress and psychopathology are associated with network-level activity during emotion processing.
In sum, I aim to explore whether the use of individualized networks to examine activity in a canonical emotion processing task improves test-retest reliability. I will then explore whether network-level activation during emotion processing is associated with stress and psychopathology both between and within individuals . If I find improved test-retest reliability using the individualized networks and further observe associations with factors such as stress and psychopathology (both between and within individuals), then such approaches could offer a new framework for studying emotion processing and its relationship to mental health and behavior.
UP NEXT: Test-Retest Reliability
Primary Question: What is the test-retest reliability of emotion processing activity within the individualized networks?