Changes in large-scale neural networks under stress are linked to affective reactivity to stress in real life

Reviewer #1 (Public review):

Summary:

In their paper, Tutunji et al aim to investigate the dynamic effects of stress on activity of different brain networks (salience network, executive network, and default mode network). Crucially they differentiate between rapid (<1 h) and late (>1) effects of stress. Lastly, they connect acute changes in brain activity with inter-individual differences in stress reactivity in real-life assessed using EMA.

They first show the expected dynamics in stress-induced brain activity with a transient increase in salience network activity and a decrease in default mode network activity although in contrast to expectations, this did not disappear in the late phase. Notably, the increase in salience network activity was associated with a 'resilience index' derived from EMA that captures whether an individual responds with more or less reduction in positive effect than expected based on the number of above average stress events.

Linking acute stress to long-term affective stress reactivity is a crucial step to better understand how adaptive or maladaptive stress responses play out in the long term and how they might be related to mental health problems.

Strengths:

The link of the acute stress response to stress reactivity in daily life is highly relevant and a major strength of the paper. Moreover, the design of the EMA component assessing a week with low stress and one with high stress (exam week) in all participants and thus including a naturalistic manipulation enables a quantification of stress reactivity that captures 'real life'.

The authors do not only quantify the magnitude of the acute stress response but take into account an early as well as late response to disentangle the dynamic nature of the stress response. In that way, it is possible to establish which parts of the stress response are relevant for the affective response.

In addition to reporting changes in network activation, the authors also report behavioral outcomes of the tasks which is crucial to evaluate the meaning and relevance of the neural outcomes.

Weaknesses:

Although the authors assess multiple physiological outcomes to the stress task, only the cortisol response is analyzed with regard to its association with the stress-induced changes in network activity. Considering that it is mainly the salience network that shows an increase and this in the early phase that is characterized by the noradrenaline and not so much the cortisol response, an association with a marker of the NA response would be interesting.

To evaluate the association of the acute stress response with stress reactivity in real life more conclusively it would be interesting to see whether and how the affective response to the acute stress is related to stress reactivity in real life.

In the introduction, the authors hypothesize that all networks show distinct activation patterns during the stress response and expect all of them to be associated with the stress reactivity during EMA. However, no correction for multiple comparisons across the many tests (each network at two phases) is reported.

All stress-induced changes in activity are assessed by using other tasks since it is not trivial to measure changes in activation of specific regions without comparing different conditions of a task. Nonetheless, with the chosen approach it is not completely clear whether stress only modulates brain responses to other tasks or changes activation within those networks independently of any other tasks. Moreover, one of the tasks did not elicit the expected activation contrast and it is unclear whether this affects stress-effects.

Some of the less central results that are discussed in the paper such as the association of the real-life stress reactivity measure with neuroticism, the sex-effect of the cortisol response or the mediation and moderation models of the stress-induced changes in network activity and performance in the tasks seem slightly overinterpreted considering that they are either not quite significant or not hypothesized and thus it is not clear why for example once a mediation and in another outcome a moderation model was chosen.

Reviewer #2 (Public review):

Summary:

This study aimed to investigate changes in neural responses over time after acute stress and their association with real-life stress. To this end, functional MRI data was collected from 3 tasks (Oddball, 2-back, Associative retrieval) early and late following stress and control conditions. Emotional ratings during a stressful week before an exam and a non-stressful week without an exam were used to index real-world stress. In total, data from 70 individuals were used for the analyses in the paper. Results showed increased oddball related activation early after stress whereas activation to the associative retrieval was reduced across early and late trials following stress compared with control. Brain activation during the oddball task after stress contrasted against control correlated with the index used to measure stress in the real-world. This is a very ambitious study and the findings that stress has opposite effects on the oddball and the associative retrieval tasks is new. However, I am not convinced that brain responses are correlated with real-world stress from the results presented in the paper. I also have several other concerns listed below.

Strengths:

The study uses a unique design based on hypothesis firmly grounded in theories of stress related brain function. Large amounts of data are collected for all of the 70 participants included in the analyses and the hypotheses tested using paired tests have strong statistical power. Data collection methods are sound aiming to reduce stress induced by being in the scanner environment for the first time and reducing variation in cortisol due to circadian rhythm.

Weaknesses:

An important argument in the paper is that neural responses associated with stress in the lab correspond to stress in real life. This conclusion is based on a single correlation analysis. This is weak evidence because the correlation is based on 70 individuals and may be driven by outliers. In fact, the correlation between the difference in stress-related SN activation (Stress-Control) and real life stress residual is likely to be driven by outliers. In fig 5b, there are 3 persons with SN values of around 2, which is twice as much as the fourth highest value. There is also 1 person with a Real life stress residual of -3 or -4, which is three to four times as much as the person with the second lowest value. These 4 outliers should be removed before calculating the correlation coefficient. Also, no power analysis is presented in the paper showing what effect size is needed for significant results given a sample size of 70.

It is not clear why the activation maps from the tasks performed in the scanner are referred to as the SN, ECN, and DMN. They are discussed as if they were resting state networks. They are however not resting state networks because they are the results of contrasting two task conditions to each other and not the results from correlating BOLD time-series data from different regions within subjects. Even though masks corresponding to SN, ECN, and DMN are used to calculate means of all voxels, I think these contrasts should be referred to as the tasks that were used to evoke them. It becomes misleading to call them networks which usually refers to nodes and edges in fMRI studies. The first scan was a resting state scan, but these data are not presented in the paper.

Introduction
In the introduction it is said that there are genomically driven effects of cortisol 1 to 2 hours after stress. This is repeated in the discussion: "[the late stress phase] is thought to be dominated by genomically driven effects of glucocorticoids". (There is no reference to this statement however.) This idea, that gene expression should only be regulated by corticosteroids following stress seems unrealistic. The increase in cortisol was only around 60% from baseline in the current study which seems to be similar to other studies. This means that the baseline cortisol level is far from zero. Therefore, effects of cortisol on gene expression must occur all the time and be tightly regulated by circadian clocks. To propose that genomically driven effects of cortisol only exist 1 to 2 hours following stress is therefore too simplistic.

In the last paragraph, it says that n=83. However, the final sample consists of 70 people. Correct this number.

Methods
The EMA data analysis is difficult to understand. Why are the residuals used instead of means for example? I could not understand how the residual values used in the analysis should be interpreted from the way this section was written. Therefore, I cannot judge whether the index is valid or reliable. Using mean values is more common than using residuals when investigating individual differences in stress responses. The use of residuals needs justification and clarification. The results from an analysis using mean values should also be reported.

How was AUCi calculated? What software was used to calculate AUCi?

How was the mediation analysis performed? The only information I found was: "We additionally ran separate models with an interaction term modelled for neural activity in the targeted ROI's to examine the relationship between task performance and neural responses, with random slopes and intercepts also modelled for ROI activity." This is not how mediation analyses are done conventionally. It is common to use structural equation modelling or a series of regression analyses. What is meant by separate models? Was a reduced model compared to a full model with an interaction term? In this case, this is not a mediation analysis. I think the term moderation is better to describe this analysis.

Reviewer #3 (Public review):

This is a very interesting study that aims to examine the effect of stress induction across about two hours on physiological, behavioral, and neural measures in several brain areas. This aim is of importance for the study of stress response and recovery and their neural bases. There are several strengths to the design, including a within-subject design, adequate sample size, and multiple levels of assessment (including lab-based and real-life), and the authors should really be commended for that. The results indicate an acute cortisol response following stress induction, although HR data show that the manipulation may have been effective only among those who did the stress scan first. Behaviorally, stress induction resulted in effects on one of the tasks. Neurally, temporal changes in response were observed in what is referred to as SN and DMN networks, and associations with real-life stress were evident for SN during early stress response. Together, evidence emerged for some temporal changes in stress response on neural function and its associations with behavior and real-life stress response as indicated by self-report EMA.

These findings, both positive and null, provide important insight to the field, and the authors should be praised for that. At the same time, it is important to emphasize that some aspects or findings complicate interpretation and limit the extent of inference, that many places in the manuscript could benefit from clarification, and that more discussion should be given to the null findings.

All in all, given the importance of the questions and the strengths of the design, this study could provide a major contribution to future research. But, to accurately and optimally guide research, it is important to accurately describe and interpret both what was tested and found, and what was not found. Some more specific points are noted below, where improvements could be made to facilitate extraction of insight by the reader, and thus increase the impact of the study on the field.