Project 3: Discussion
In the present study, I examined whether an individualized network approach to characterizing activity during a canonical emotion processing task improves test-retest reliability with the aim of identifying alternative targets for measuring functional recruitment in studies of emotion processing. I compared test-retest reliability of activity within the individualized networks to activity within the amygdala, a set of meta-analytic ROIs, as well as networks defined using a group average parcellation. Following this reliability analysis, I examined whether network-level activation during emotion processing is associated with stress and psychopathology, including depression and anxiety symptoms. I examined these associations using both a between-person (ie. individual differences) approach, as well as within-person approach (ie. within-person fluctuations). Finally, to confirm that activity within the individualized networks is associated with meaningful task-level behavior, I additionally examined associations with task performance, including task accuracy and response time.
Evaluating Test-Retest Reliability of Neural Function During Emotion Processing. Characterizing activity during this canonical emotion processing using a network-level framework did not improve test-retest reliability. Though I observed significant activity during emotion processing within the individualized networks, the test-retest reliability of this activity was poor across all individualized networks, as well as across all group average networks examined. Furthermore, I replicated previous accounts of poor test-retest reliability of activity within the amygdala (Elliott et al., 2020; Flournoy et al., 2023). The left and right fusiform were the only regions that exhibited fair-good reliability. In reliability assessments of tasks that span multiple domains, this emotion processing task exhibited the lowest test-retest reliability (Elliot et al., 2020). Higher test-retest reliability estimates were observed in tasks targeting motor and sensory functions, as well as working memory. This indicates that test-retest reliability varies significantly based on task design and specific domain of information processing. Emotion-relevant paradigms, such as the present task, may be particularly sensitive to within-individual variation in affect (Rocke et al., 2009). Therefore, the poor test-retest reliability observed in the present study likely reflects both high measurement error in the BOLD fMRI signal, in conjunction with high within-person variability in neural responses to this task. Indeed, in a recent study, Flournoy et al., found that roughly half of the variance observed in task-evoked activity in a similar emotion processing task could be attributed to reliable within-person variability and that this variability could be explained in part by within-person fluctuations in mood, sleep, and stressful life events (Flournoy et al., 2023). Other factors such as variability in task completion strategy may additionally increase within-person variability in functional activity during this task. Characteristics of this canonical task, such as the broad contrast employed between emotional faces and shapes, may make it particularly liable to high within-person variability across time. To successfully complete this task, subjects can use a wide range of strategies and engage in numerous forms of information processing —from attending to the low-level features of the face, to evaluating the identity of the faces, to processing the emotions exhibited on the faces. These various processes will result in varied neural responses that are not dissociable given the contrast used (i.e., the comparison of emotional faces to simple geometric shapes). Indeed, in Paper 2 I observed that alongside activity in the visual-central network and dorsal attention network B, there were a number of other networks that exhibited meaningful activity across many subject’s sessions (specifically control network A and default network B). While I did see these patterns across many subject-sessions, they were not ubiquitous even within the same subject, suggesting that subjects may be engaging in different task completion strategies across sessions. This may further contribute to the poor test-retest reliability of functional activation observed. In sum, using individualized networks as the units of analysis does not address issues of poor test-retest reliability of this task, which may have design characteristics that inherently prevent the identification of reliable neural signatures of emotion processing.
Associations with Stressful Life Events and Psychopathology. I did not observe significant associations between neural activity during emotion processing and stressful life events, depression symptoms, or anxiety symptoms within any of the boundaries examined–left and right fusiform, the amygdala, individualized networks, or networks defined using group averaged data–at either the between-person or within-person level. These results are perhaps unsurprising given the poor test-retest reliability observed for neural activity and the small sample used in the present study. Good test-retest reliability of task-evoked activity is essential for studies of individual differences, as it sets an upper limit on our ability to detect valid associations. Furthermore, recent work has suggested that cross-sectional brain-wide associations require thousands of individuals to be reliably identified (Marek et al., 2022). Therefore, a lack of between-person associations observed in the present study is not entirely unexpected. However, I additionally did not observe significant associations between neural activity during emotion processing and within-subject fluctuations in stressful life events, depression symptoms, or anxiety symptoms across the 12 months of the study in any of the boundaries examined. Flournoy et al. (2023), found within-person variability in neural activity during a very similar emotion processing task was associated with within-person fluctuations in mood, sleep, and stressful life events across a similar 12-month time frame sampled at the monthly level, but I did not replicate these findings. Importantly, the task used across these two studies was not identical. Flournoy et al., used a task with a much tighter contrast (Fearful Faces > Neutral Faces) that was able to isolate neural responses to threatening emotions specifically. It is possible that associations of neural activity during emotion processing with fluctuations in mood, sleep, and stressful life events are specific to threat-relevant processing and do not generalize to the broader range of processes captured in the present task. These findings further support the notion that this canonical task may lack the specificity needed to detect meaningful individual differences, thereby limiting its usefulness in studies of emotion processing. Furthermore, these findings are important given that this task continues to be used in some of the largest scale data collection efforts in cognitive neuroscience.
Associations with Task Performance. Although I did not observe meaningful associations between activity in this task with factors such as stressful life events and symptoms of psychopathology , I also examined whether functional activity in the individualized networks is associated with task behavior. The emotion processing task has two behavioral measures that could be evaluated: task accuracy and response time. Neural activity exhibited significant associations with task accuracy only in the right and left fusiform. Activity within both the amygdala and the individualized networks (the visual-central network and dorsal attention network B) was not associated with task accuracy. Importantly, there was a ceiling effect for task accuracy as the task of matching shapes and faces is very simple. The presence of region-level effects in the fusiform, but lack of network-level associations, suggests that the networks may not be as sensitive to small effects as focal regions.
While I did not observe associations of task accuracy with activity within the individualized networks, I did see significant associations with response time. Associations between response time and activity during the Faces > Shapes contrast were present across all the boundaries (amygdala, left and right fusiform, visual-central network, and dorsal attention network B), suggesting that the faster a subject responded in the task in general, the more differentiated the activity between the faces and shapes conditions across each of these approaches for defining the boundaries of neural activity. Taken together these results suggest that the individualized networks are sensitive to behavioral outcomes in this task, though do show less sensitivity relative to localized regions.
Conclusion. In conclusion, using an individualized network approach to characterize activity within a canonical emotion processing task does not improve test-retest reliability. Furthermore, although activity within individualized networks was associated with reaction time during the task, this activity was not associated with factors such as stressful life events, depression symptoms, or anxiety symptoms, either between individuals or within individuals. In sum, our findings add to the growing evidence that calls into question the value of this highly utilized canonical emotion processing task for studying individual differences.