Atypical attentional network dynamics in adolescent depression during emotional movie viewing

Abstract

Emotion studies have commonly reported atypical emotional processing in clinically depressed adolescents in the context of short-lasting emotional cues. However, interindividual differences in the moment-to-moment brain network dynamics that underlie this impaired emotional reactivity remain unclear, and the use of poorly matched controls and relatively small sample sizes represents major limitations in most neuroimaging depression studies to date. Here, we address these concerns by using the temporal features of a rich naturalistic paradigm (i.e. a clip from the movie ‘Despicable Me’) to investigate brain network dynamics in 42 clinically depressed and 42 nondepressed adolescents aged 16–21 years, matched for age, gender, and psychiatric comorbidities. Using a dynamics functional connectivity analysis technique called Leading Eigenvector Dynamics Analysis, we found that the clinical group exhibited significantly higher probability of occurrence of the dorsal attention network and lower recruitment of the fronto-parietal, default mode network, ventral attention, and somato-motor networks throughout the task. This brain/behaviour relationship was prominent during less emotional moments of the movie, consistent with previous findings. Our findings demonstrate the key role of continuous affective measures in providing information about how activity in the depressed brain evolves as emotional intensity unfolds throughout the movie. Future studies with a larger sample size are needed in order to corroborate the present findings.

Introduction

Adolescent service use for affective disorders has sharply increased in the last decades (Collishaw 2015). A recent survey found that the prevalence of past-year major depressive episode increased from 8.1% to 15.8% between 2009 and 2019 in the USA (Daly 2022). This is concerning, especially since adolescence is a period where key life transitions and major socio-emotional and cognitive changes are occurring, such as the rapid development of the prefrontal cortex, leading to the consolidation of key executive, reasoning, and self-regulatory skills (Patton et al. 2016). In particular, major depression is a common affective disorder in adolescents (Essau 2008). It is estimated that the 1-year point prevalence rate of major depression is quite low in adolescent children who are younger than 15 years (1%–3%), but then rises rapidly from 14% to 17% between 16 and 21 years of age (Kessler et al. 2001, Costello et al. 2003).

In the last decade, a growing number of studies have revealed that adolescent depression is characterized by impaired executive functioning in the domains of attention, memory, and verbal reasoning (see Goodall et al. 2018 for a systematic review and meta-analysis). These difficulties with sustaining attention have been associated with atypical emotional reactivity, such as a blunted emotional response to positive images and an increased response to negative images (Bloch et al. 2013). Furthermore, there is a general consensus across studies regarding aberrant sustained attention and gaze durations for happy and sad stimuli in depression (see Suslow et al. 2020 for a systematic and meta-analytic review). From a neuroimaging perspective, when asked to ignore threat-related images, adolescents with major depressive disorder have been found to recruit, to a greater degree than control participants, brain regions involved in cognitive control and attention-orientation, such as the fronto-cingulate network (Colich et al. 2017).

Together, these findings are indicative of a strong relationship between atypical emotional reactivity and the way depressed adolescents deploy attention with respect to affective stimuli. However, one of the main limitations to these studies is that they have employed ‘static’ emotional stimuli (i.e. images), and little is known about the links between emotional reactivity and attentional processes in the context of ‘dynamic’ cues such as more ecological naturalistic paradigms (i.e. music or movies). Indeed, the literature suggests that exploring brain function under time-varying, real-life situations may help gain a better understanding of adaptive behaviour (Sonkusare et al. 2019) as the brain progresses from assessing the relevance of the emotional cue to detecting its significance and evaluating how to cope with it (component process model; Scherer 2009).

Because movies represent a rich repertoire for the exploration of a range of cognitive and socio-emotional mechanisms encountered in everyday contexts (e.g. emotional reactivity, social interactions, and attentional processes), they have increasingly been used in emotion research, including in neuroimaging (see Jääskeläinen et al. 2021 for a review). Indeed, watching movies has been shown to robustly induce deep emotional reactions beyond any other mood induction procedure (Westermann et al. 1996), and movie-watching functional Magnetic Response Imaging (fMRI) has been found to outperform resting-state fMRI for predicting trait-like cognitive and emotional behaviours in healthy young adults (Finn and Bandettini 2021) and in children (Vanderwal et al. 2019).

It is worth noting, however, that only a few studies have so far used movie clips to explore adolescent depression, despite the growing popularity of movie fMRI. Gruskin et al. (2020) found a relationship between depression symptom severity and the synchronization of brain activity, as measured with dynamic intersubject correlation, in adolescents (13–21 years old), but not in children (7–12 years old), in the context of emotional movie watching. Adolescents with more similar depressive symptom severity shared more similar neural responses to the movie, and this relationship between behavioural and neural profiles was stronger during less emotional moments of the movie, especially in the medial prefrontal cortex and the right posterior cingulate cortex (Gruskin et al. 2020), two regions of the default mode network (DMN) that have been shown to increase when participants share similar interpretations of the narrative of a story (Yeshurun et al.2017). These findings have several implications. First, they suggest that the relationship between depressive symptom severity and neural activity during movie watching emerges during adolescence, which makes movie-based fMRI very promising for the exploration of adolescent depression. Second, they highlight the importance of exploring the brain/behaviour relationship at the dynamic level instead of treating emotion processing as time-invariant. Third, they reinforce that using continuous measures providing information about the moment-to-moment emotional intensity elicited by the movie can help shed light on how activity in the depressed brain evolves as a function of how emotional intensity unfolds throughout the movie. However, the study by Gruskin et al. (2020) also has an important limitation in that psychiatric comorbidities were not controlled for in the analyses.

In this study, we used the Leading Eigenvector Dynamics Analysis (LEiDA) method to investigate interindividual differences in brain dynamics at play during movie watching in depressed, compared with nondepressed adolescents, matched for psychiatric comorbidities. Using LEiDA, Cahart et al. (2024) recently found increased activity of the dorsal attention network (DAN) in healthy young adults with higher levels of depressive symptoms during an emotional music listening task. More precisely, increased DAN recruitment predicted elevated depressive symptoms, which, in turn, predicted blunted emotional responses (Cahart et al. 2024). Given the key role of the DAN in voluntarily orienting attention towards the external environment, aberrant recruitment of this network during emotional tasks may underlie difficulties in processing and engaging with emotionally relevant cues (Corbetta and Shulman 2002), typically observed in major depressive disorder (Rottenberg et al. 2005). In particular, atypical DAN activity at rest (see Kaiser et al. 2015 for a meta-analysis) and during cognitive tasks (Sambataro et al. 2017) has previously been found in adult depression. However, it is currently unclear whether similar aberrant recruitment of the DAN is evident in an adolescent clinical cohort in the context of emotionally relevant naturalistic paradigms. In particular, LEiDA characterizes the dynamic behaviour of brain networks, revealing changes in functional connectivity over the course of a scan. It identifies distinct patterns of coordinated neural activity that the brain transitions between. Each network, or ‘state’, is defined by synchronization between a specific set of brain regions. Desynchronization of typically observed dynamically synchronized networks may reflect functional pathophysiology contributing to psychiatric conditions. Consequently, LEiDA potentially affords a valuable tool for identifying functional markers of major depression (Alonso et al. 2023).

Materials and methods

Participants

Neuroimaging and behavioural data from the publicly available Child Mind Institute’s Healthy Brain Network (HBN) data portal (Alexander et al. 2017) were used for this study. The HBN project is a large ongoing initiative which has been creating a biobank from a high-risk community sample of children and adolescents. Data from 168 unmedicated participants, aged between 16 and 21 years, with complete fMRI and questionnaire data, were initially downloaded from releases 1–9.

Prior to conducting the research, all participants provided written informed consent. For participants younger than 18 years, written consent was obtained from their parents or legal guardians and written assent was obtained from the participant. The study was approved by the Chesapeake Institutional Review Board.

Of those 168 participants, 35 were excluded either for scans not passing visual inspection or for maximum head motion exceeding 3.3 mm. Of the 133 remaining participants, the data from all 42 participants with a depression diagnosis were used for subsequent analyses, as well as a further 42 participants without depression who were identified as matching the depression sample in terms of age, gender, and comorbidities. Details about how both groups were matched in terms of age, gender, and comorbidities can be found in the ‘Matching analysis’ section. The comorbidities present in the final sample included Generalized Anxiety Disorder, Attention Deficit Hyperactivity Disorder (ADHD)—Inattentive subtype, and ADHD—Combined subtype, Autism and Trauma. The remaining 49 participants were excluded from the analyses. Therefore, the final number of participants included in our analyses consisted of 42 participants with depression (15 males and 27 females, age = 17.63 ± 1.6 years), which we will refer to as the depression group, and 42 participants without depression (17 males and 25 females, age = 17.58 ± 1.4 years), whom we will refer to as the control group.

MRI data acquisition

Participants were scanned at the Rutgers University Brain Imaging Center, Newark, USA, using a Siemens 3-Tesla Tim Trio scanner. For each participant, fMRI images were acquired using a multiband 6 sequence with the following parameters: repetition time (TR) = 0.8 s, echo time (TE) = 30 ms, flip angle = 31°, slice thickness = 2.4 mm, in-plane voxel resolution 2.4 mm², and field of view = 204 mm. Anatomical T1-weighted scans had the following parameters: TR = 2.5 s, TE = 3.15 ms, flip angle = 8°, slice thickness = 0.8 mm, in-plane voxel resolution 0.8 mm², and field of view = 256 mm. Full details about the scanning parameters used for the HBN project can be found at http://fcon_1000.projects.nitrc.org/indi/cmi_healthy_brain_network/MRI%20Protocol.html.

Resting-state fMRI, diffusion kurtosis imaging, and Peer Eye Estimation Regression were also collected for all participants. However, here, we focus solely on the fMRI data acquired while participants watched a short movie clip.

MRI preprocessing

The data were preprocessed using the CONN toolbox Version 20b (Whitfield-Gabrieli and Nieto-Castanon 2012) and MATLAB R2020a (MathWorks, Natick, MA, USA). The images were first realigned, coregistered, and then spatially normalized into the Montreal Neurological Institute (MNI) standardized space. Because of the clinical nature of this population, we used Diffeomorphic Anatomical Registration Through an Exponentiated Lie algebra [DARTEL registration (Ashburner 2007)] to create a template specific to this study. Each participant’s T1-weighted image was first segmented into grey matter (GM), white matter (WM), and cerebrospinal fluid, which led to the creation of a group-specific template using DARTEL registration. This template was used for normalizing the functional images to MNI space. Smoothing was then carried out on the normalized GM and WM segments using a Gaussian filter of 8-mm spatial full-width at half-maximum value.

Depressive symptom inventory

Each participant was administered a semistructured DSM-5-based clinical interview (i.e. the Kiddie Schedule for Affective Disorders and Schizophrenia - Computerized Version KSADS-COMP; Kaufman et al. 1997) by a qualified clinician, and consensus DSM-5 diagnoses were evaluated following the completion of the clinical interview and other study procedures. As part of a phenotypic assessment, all participants were also required to complete questionnaires that measured their psychiatric, behavioural, and cognitive functioning using standard assessment instruments (Alexander et al. 2017). For this study, we focused on the Moods and Feelings Questionnaire—long version (MFQ; 18), which is made up of 33 statements describing how the participant has been feeling in the past 2 weeks, such as ‘I felt miserable or unhappy’ or ‘I felt lonely’. Participants rated each statement on a 3-point scale (e.g. not true, sometimes, and true).

Movie clip and subjective ratings

Functional MRI data were collected while the participants watched a 10-min emotionally provocative clip from the movie ‘Despicable Me’ (spanning from 1:02:09 to 01:12:09). This movie clip was chosen because it alternates humorous scenes such as an adoptive caregiver developing a strong emotional connection with children while reading them bedtime stories and sad scenes depicting rejection by the caregiver and the children being taken away from him. Additionally, previous studies have shown that this clip elicits strong emotional responses in adolescents and that adolescents with more similar levels of depressive symptom severity exhibit more similar brain responses while watching the clip (Gruskin et al. 2020). A more detailed description of the movie clip can be found at http://fcon_1000.projects.nitrc.org/indi/cmi_healthy_brain_network_working/mri_protocol.html. For this study, we used continuous subjective ratings previously collected by D.C.G., first author from Gruskin et al. (2020). The subjective ratings were provided by 20 healthy independent adult raters (mean age = 25 ± 3.74 years), each of whom continuously rated the emotional valence of the movie clip on a scale of −4 (unpleasant) to 4 (pleasant) three times, resulting in 60 trials altogether. Ratings were collected every 250 ms.

Matching analysis

We used the MatchIt library (R Foundation for Statistical Computing) to ensure that both groups were matched in terms of age, gender, and comorbidities (i.e. Generalized Anxiety Disorder, ADHD—Inattentive subtype and ADHD—Combined subtype, Autism and Trauma). After carrying out the matching analysis in R, an independent t-test was run as a post hoc analysis for confirming that no significant differences were observed between the groups in terms of age, and chi-square analyses were used to check for differences for the gender and comorbidities variables.

Dynamic analyses

For the analysis of dynamic connectivity, we used LEiDA codes initially developed by [(Cabral et al. 2017) https://github.com/juanitacabral/LEiDA] that we adapted to our data. The analyses were carried out within MATLAB R2020a (MathWorks, Natick, MA, USA).

A total of 105 Regions of Interest (ROIs), defined anatomically based on the Harvard-Oxford cortical atlas, were extracted from the CONN toolbox (Whitfield-Gabrieli and Nieto-Castanon 2012). For each ROI, the BOLD time series were first Hilbert-transformed into an analytic signal, which, unlike the real BOLD signal, has a phase associated with it. The Hilbert transform typically expresses the real signal as the product between the instantaneous amplitude of the signal and the cosine of the instantaneous phase angle of the signal. The dynamic phase-locking matrix [dPL(t)] was then calculated to evaluate the degree of synchrony between pairs of ROI phases at each time point t (Fig. 1a). The leading eigenvector V₁(t) of the matrix dPL(t) was then computed (Fig. 1b–c). V₁(t) typically captures the main orientation of all 105 BOLD phases, with reduced dimensionality, at a given time point t. It contains 105 elements (i.e. brain regions) which can be divided into two communities (i.e. blue or red), depending on the direction of the projection of their phase onto V₁(t). By convention, most elements in V₁(t) have a negative value (Alonso Martínez et al. 2020, Vohryzek et al. 2020). When all elements have a negative sign, then it is indicative of global coherence. In contrast, a positive sign reflects brain regions that detach from global coherence, project onto the opposite direction of V₁(t), and represent meaningful functional networks that dominate V₁(t).

A k-means clustering algorithm was then applied to the concatenated data to divide into clusters the set of leading eigenvectors from all 84 participants and all 750 time points, with k ranging from 5 to 10 clusters. A Dunn index (Dunn 1973) was obtained for each clustering solution between 5 and 10, and the highest Dunn value was selected as it reflects minimal variance within cluster and maximal distance between clusters (Dunn 1973). This led to the identification of an optimal number of clusters k, each corresponding to a distinct BOLD phase-locking state (Fig. 1d).

The probability of occurrence, described as the percentage of time points during which a given state dominates during the scan, was then calculated for each state and each group.

To identify meaningful reference functional networks within each state, we then calculated the spatial similarities shared between each state and each of the seven Yeo networks previously identified in the literature (Yeo et al. 2011): visual, somato-motor, DAN, ventral attention network (VAN), limbic, fronto-parietal, and DMN.

More specifically, we first calculated, for each of the 105 ROIs, the proportion of voxels assigned to each of the seven Yeo networks, thus leading to seven 105 × 1 vectors where each element represents its contribution to the Yeo network. We then calculated Pearson’s correlation coefficients between each of these seven networks and the centroids V_k previously obtained from the k-means clustering analysis, following the methodology described in Vohryzek et al. (2020). Significance was set at P < .01/k.

Finally, Analyses of Covariance (ANCOVAs) were carried out to compare the groups with regard to the probability of occurrence metric. For any states where the groups differed, Pearson’s correlation coefficients were then calculated between MFQ scores and the probability of that state, so as to further explore to what extent the metric of that state was related to the severity of depressive symptoms.

For all analyses, age, gender, and comorbidities were used as covariates of no interest (Miller and Chapman 2001) and correction for multiple comparisons was implemented using false discovery rate (FDR) correction (Benjamini and Hochberg 1995).

Relationship between depressive scores, network behaviour, and perceived emotional intensity

To explore whether emotional intensity was related to the brain/behaviour association, we first calculated the averaged subjective ratings across the independent raters at each time point. To facilitate the analysis, we defined ‘neutral’ moments (i.e. less emotional) as time periods when the values of the averaged ratings ranged between −1 and 1 for at least 40 s (i.e. 50 consecutive time points). We chose this 40-s threshold because movie clips of that duration have previously been shown to elicit strong emotional responses (Carvalho et al. 2012). ‘Sad’ periods were defined by scores below −2, for at least 50 consecutive time points. For each emotion type (i.e. ‘neutral’ and ‘sad’), we then extracted the probability of occurrence of the LEiDA states of interest. We next calculated one-way ANCOVAs to compare the groups directly (i.e. depressed vs nondepressed). Following this, we calculated Pearson’s correlation coefficients between the MFQ scores and the probability of occurrence of the LEiDA state. Again, age, gender, and comorbidities were used as covariates of no interest, and correction for multiple comparisons was implemented using FDR correction in all analyses (Benjamini and Hochberg 1995).

Results

Matching analysis

After matching analysis, the independent t-test revealed no significant differences between groups in terms of age [t₍₈₂₎ = −0.148, P = .883]. The chi-square tests revealed no significant differences between groups in terms of the prevalence of gender and comorbidities (Table 1).

Dynamic analyses

The Dunn score revealed that the optimal number of states (k) was eight.

Correlations between LEiDA states and each Yeo network

State 1 did not significantly correlate with any of the Yeo networks, which is indicative of global coherence. As displayed in Fig. 2, State 2 significantly correlated with the fronto-parietal and DMN networks; State 3, with the DMN and the DAN; State 4, with the DAN; State 5, with the visual network; State 6, with the somato-motor and DMN networks; State 7, with the visual network; and State 8, with the ventral attention (VAN) and somato-motor networks. None of the states correlated with the limbic network.

For the rest of the paper, we will refer to each state by the underlying Yeo functional network(s) they significantly correlated with.

Figure 3 presents each state in cortical space, with coloured areas reflecting brain regions with a positive sign that detach from the global coherence mode.

Comparing LEiDA’s probability of occurrence between groups

One-way ANCOVAs revealed that the probability of occurrence was significantly higher in the control group compared to the depression group for the fronto-parietal and DMN networks [State 2; 14.2% vs 10.7%; F(1,75) = 11.125, pFDR = .0027] and for the VAN and the somato-motor network [State 8; 12.5% vs 9.5%; F(1,75) = 12.585, pFDR = .0027] (Fig. 4). In contrast, the probability of occurrence was significantly lower in the control group compared to the depression group for the DAN ([State 4; 9% vs 13.5%; F(1,75) = 22.861, P < .001]. There was no significant difference between groups for the other five states (pFDR > .05).

To further explore the robustness of our findings, we further investigated these one-way ANCOVA results across different clustering solutions using the cluster-number (k) set between k = 5 and k = 10. F-tests and P-values after FDR correction can be found in Supplementary Table S1. We found that the difference between groups remained highly significant for the DAN across all clustering solutions (pFDR < .001). For the VAN and somato-motor network, P-values were significant across all clustering solutions, with smaller values for k > 6 and larger values for k = 5 and k = 6. For the fronto-parietal and DMN networks, P-values were also significant across all solutions except for k = 10, which might be explained by the finer granularity in this specific clustering solution. Altogether, results are robust across different numbers of states between 5 and 9.

A Dunn index curve illustrating the Dunn score for each clustering solution can be found in Supplementary Figure S1.

Relationships between LEiDA’s probability of occurrence and the severity of depressive symptoms

There were a significant positive correlation between the MFQ scores and the probability of occurrence of the DAN (State 4; r = 0.418, pFDR < .001) and a significant negative correlation between the MFQ scores and the probability of occurrence of the VAN and the somato-motor network (State 8; r = −0.338, pFDR = .003). There was no significant correlation between the MFQ scores and the probability of occurrence of the fronto-parietal and DMN networks (i.e. State 2).

Relationship between LEiDA, depressive scores, and emotional intensity

Based on the time series of the subjective ratings averaged across independent raters at each time point, we identified one ‘sad’ period (i.e. emotional), as well as two ‘neutral’ moments (i.e. less emotional), one of which was before the sad period (i.e. ‘neutral_before’), while the other one was just after (i.e. ‘neutral_after’). Figure 5 illustrates the averaged subjective ratings over time, as well as the sad and neutral moments. There was no period of time where averaged ratings were above 2 over at least 50 consecutive time points, and hence why no ‘happy’ moment is displayed in Fig. 5.

As illustrated in Fig. 6, for the fronto-parietal and DMN networks (i.e. State 2), the one-way ANCOVAs revealed a significant difference between groups for the probability of occurrence during ‘neutral_before’ [F(1,75) = 12.830, pFDR < .001]. There was no significant difference between groups during the ‘sad’ and ‘neutral_after’ moments (pFDR > .05).

There were a significant negative correlation between the MFQ and the probability of occurrence of the fronto-parietal and DMN networks during ‘neutral_before’ (r = −0.312, pFDR = .006) and no significant correlation during ‘sad’ and ‘neutral_after’ (pFDR > .05).

For the DAN (i.e. State 4), there was a significant difference between groups for the probability of occurrence during both ‘neutral_before’ [F(1,75) = 8.549, pFDR = .014) and ‘neutral_after’ moments [F(1,75) = 7.149, pFDR = .014]. There was no significant difference between groups during the ‘sad’ moment (pFDR > .05).

There were a significant positive correlation between the MFQ and the probability of occurrence of the DAN during ‘neutral_before’ (r = 0.227; pFDR = .047) and ‘neutral_after’ (r = 0.273, pFDR = .016) and no significant correlation during ‘sad’ (pFDR > .05).

For the ventral attention and somato-motor networks (i.e. State 8), there was a significant difference between groups for the probability of occurrence during ‘neutral_after’ [F(1,75) = 9521, pFDR = .003]. There was no significant difference between groups during the ‘sad’ and ‘neutral_before’ moments (pFDR > .05).

There were a significant negative correlation between the MFQ and the probability of occurrence of the ventral attention and somato-motor networks during ‘neutral_after’ (r = −0.308; pFDR = .006), and no significant correlation during ‘sad’ and ‘neutral_before’ (pFDR > .05).

To check for the robustness of our findings and verify that our findings are not the result of relative undersampling of the probability distributions within the target window, given eight observed network states, we also calculated ANCOVAs for the probability of occurrence during ‘sad’, ‘neutral_before’, and ‘neutral_after’ for clustering solutions k = 5 to k = 7. The findings were similar to those obtained for k = 8.

More specifically, for the fronto-parietal and DMN networks, there was a significant difference between groups during ‘neutral_before’ for k = 5 [F(1,75) = 7.845, pFDR = .03], k = 6 [F(1,75) = 12.602, pFDR < .001], and k = 7 [F(1,75) = 10.595, pFDR = .014].

For the DAN, there was a significant difference between groups during ‘neutral_before’ for k = 5 [F(1,75) = 10.214, pFDR = .01], k = 6 [F(1,75) = 10.120, pFDR = .012], and k = 7 [F(1,75) = 10.862, pFDR = .014]. There was also a significant difference between groups during ‘neutral_after’ for k = 5 [F(1,75) = 4.807, pFDR = .048], k = 6 [F(1,75) = 5.557, pFDR = .049], and k = 7 [F(1,75) = 4.944, pFDR = .048].

For the ventral attention and somato-motor networks, there was a significant difference between groups during ‘neutral_after’ for k = 5 [F(1,75) = 6.768, pFDR = .045], for k = 6 [F(1,75) = 4.735, pFDR = .045], and k = 7 [F(1,75) = 20.438, pFDR = .014].

There was no significant difference between groups during ‘sad’ for any of the networks.

Discussion

In line with our hypothesis, we observed an increased probability of occurrence of the DAN in the depressed group compared to the control group in the context of LEiDA analyses. We also observed a lower probability of occurrence of the fronto-parietal and DMN networks and of the VAN and somato-motor networks in the depression group. Together, these findings could reflect an unusual balance between being immersed in self-reflective processes and engaging with the external environment (Sambataro et al. 2017). In particular, this over-recruitment of the DAN and under-recruitment of the VAN are in line with the literature showing attentional difficulties [see Rock et al. (2014) for a meta-analysis] and an atypical engagement with external stimuli [Constructionism; Barrett (2016)] in clinical depression. Aberrant behaviours of the DAN have previously been observed at rest [see Kaiser et al. (2015) for a meta-analysis] and during task performance (Sambataro et al. 2017) in adult depression, and atypical effective connectivity within the VAN has formerly been observed in subclinical adolescent depression (Liu et al. 2019). The VAN is traditionally seen as a bottom-up attentional network that involuntarily responds to unexpected and unattended, yet behaviourally salient, stimuli, while the DAN is thought to be part of a top-down controlled attentional system that voluntarily allocates attention to external stimuli (Corbetta and Shulman 2002). Previous studies have suggested that a dynamic interaction between both networks is critical for flexible attentional control (Vossel et al. 2014) and that aberrant coordination between such systems could underlie the attentional and emotional deficits typically observed in depression (Kaiser et al. 2015).

Our findings showing less recruitment of the fronto-parietal and DMN networks in the depressed group are in line with Figueroa et al. (2019) who found a significantly lower probability of occurrence of the frontal, default mode, and salience networks in participants in stable remission from major depression compared to healthy controls. These networks are involved in cognitive control and flexible switching between internal thoughts and the external environment. Our findings, however, do not conform to those from Alonso Martínez et al. (2020) who found an increased engagement of these networks in participants with higher levels of depressive symptoms. It is worth noting, however, that their study differed from this present research in two ways. First, their sample was made up of healthy adults with subclinical depression, while our participants were clinically depressed adolescents. Second, Alonso Martínez et al. (2020) investigated neural processes in the context of unconstrained resting-state fMRI where participants are simply instructed to lie still and think of nothing in particular, while we explored brain dynamics in response to a standardized emotionally provocative movie where participants were required to actively engage in an external event. In fact, resting-state fMRI measures brain activity in the absence of any specific task, which may lead to brain data being difficult to link to specific behaviours and often being interpreted with reversed inference and assumed mental processes (Finn 2021). In light of this, it is not surprising to observe opposite behaviours of the same neural networks in resting-state fMRI compared to naturalistic paradigms.

We also found a more prominent brain/behaviour relationship during less emotional moments of the movie. We did observe a significant difference between groups in the probability of occurrence for the DAN (i.e. State 4) during both neutral moments, for the fronto-parietal and DMN networks (i.e. State 2) during ‘neutral_before’, and for the ventral attention and somato-motor networks (i.e. State 8) during ‘neutral_after’. In contrast, there was no significant association between depressive symptoms and brain dynamics during the sad moment for any of the three states. These findings are in line with Gruskin et al. (2020) who found lower intersubject correlation values at the group level and a stronger relationship between depression status and neural dynamics during less emotional moments of the same movie clip (i.e. Despicable Me) in clinically depressed adolescents. It could be that participants may have interpreted the movie in less similar ways when it was less emotional and perhaps also more ambiguous and less engaging (Gruskin et al. 2020). Figueroa et al. (2019) also found that patients displayed reduced recruitment of the fronto-parietal, DMN, and VANs during neutral mood, while there was no significant difference between groups during sad mood. Our findings also align with Cahart et al. (2024) who found an elevated recruitment of the DAN during neutral moments of a music listening task in participants with greater depressive symptoms, while there was no association between depressive symptoms and DAN activity during sad moments. Further work is needed to better understand this enhanced brain–behaviour relationship during neutral periods of time.

Furthermore, the Yeo limbic network did not significantly correlate with any of the eight states identified by k-means clustering. This is in line with previous studies which found that the Yeo limbic network did not correlate with any LEiDA state for multiband sequences (i.e. factors 4 and 6; Cahart et al. 2022). In fact, Vohryzek et al. (2020) also found that LEiDA did not detect the limbic network unless the number of states k was equal to, or higher than, 9, when using multiband sequences. These findings concur with a growing number of studies showing reduced signal-to-noise ratio (Risk et al. 2021, Cahart et al. 2023) and lower test–retest reliability of functional connectivity metrics (Cahart et al. 2023) in subcortical regions in the context of multiband.

In terms of clinical relevance, LEiDA enables a detailed analysis of brain spatiotemporal dynamics, potentially supporting the development of individualized brain models for personalized psychiatric treatment. In particular, systematic stimulations can be applied through dynamic sensitivity analysis to rebalance brain activity from aberrant to optimal brain dynamics (Vohryzek et al. 2023, 2024). By modelling these dynamics, this framework allows for the identification of key regions where targeted interventions could shift brain states towards healthier patterns, paving the way for powerful therapeutic intervention in.

Limitations

Our study does suffer from some limitations. Firstly, the subjective ratings were provided by healthy adult raters from an independent sample. Even though they shed light on the reliable emotional content of the movie as the story unfolded, the raters’ affective experiences might have slightly differed from those of our adolescent participants. Indeed, a recent review suggests that adolescents typically experience emotions more intensely than adults and emotions tend to become increasingly unstable during adolescence (Bailen et al. 2019). This means that the delineation of the ‘neutral’ and ‘sad’ moments might have slightly differed had we been able to use subjective ratings provided by the participants or others of the same age.

Second, the HBN dataset did not include information regarding how familiar the participants were with the ‘Despicable Me’ movie. In fact, previous studies have shown that familiarity levels may alter affective experience during movie watching (Maffei et al. 2019), such as increasing the arousal regardless of whether the content is positive or negative. Future studies would benefit from collecting information about participants’ familiarity with the movie clips, so this could be controlled for.

Third, we acknowledge that, despite being an improvement compared to earlier research, our sample size was relatively small and our results should therefore be interpreted with caution. It is worth noting that, despite a small sample size, we are confident that our findings are robust given that we found similar results across different clustering solutions. However, future studies with a larger sample size are needed in order to corroborate the present findings.

Conclusion

In this study, we identified distinct neural dynamics in depressed compared to nondepressed adolescents matched for age, gender, and psychiatric comorbidities. We found an over-recruitment of the DAN and an under-recruitment of the VAN, the somato-motor network, the fronto-parietal network, and the DMN in depressed adolescents while watching a naturalistic, emotionally salient movie clip. This pattern was prominent during less emotionally intense moments, consistent with recent findings. Our study’s novel contribution lies in investigating these brain network dynamics in depressed and nondepressed samples, carefully matched for psychiatric comorbidities, throughout unfolding emotional contexts. We hope that these findings will inform future depression studies into further exploring the dynamics of atypical attentional mechanisms during emotionally provocative naturalistic paradigms. Such an approach may enhance the ecological validity of neuroimaging studies in clinical populations and potentially inform the development of more targeted interventions.

Acknowledgements

The analyses in this paper were made possible by the data provided by the Child Mind Institute as part of the HBN biobank. We would also like to thank the National Institute for Health Research Maudsley Biomedical Research Centre and King’s College London for their funding and ongoing support of our neuroimaging endeavours.

Supplementary data

Supplementary data is available at SCAN online.

Conflict of interest:

None declared.

Funding

None declared.

Data availability

The datasets generated in the current study are available from the corresponding author on reasonable request.

References