Population-neuroscience Tokyo TEEN Cohort (pn-TTC)

Image acquisition and processing

For the four waves of data collection (from September 2013 to June 2023), two 3-T scanners and three acquisition procedures were used (S1 Fig). Only T1-weighted images were acquired for procedures 1 and 2, whereas T1-weighted and T2-weighted images were acquired for procedure 3 (S2 Methods and S1 Table). When only T1-weighted images were available, the legacy style of the Human Connectome Project pipeline was used for image preprocessing. In cases in which both T1-weighted and T2-weighted images were available, the standard style of the Human Connectome Project pipeline was used for image preprocessing [24]. Then, we extracted the 75 features of cortical thickness, cortical surface area, and subcortical volume using the Desikan–Killiany atlas. Traveling subject harmonization was performed to diminish the procedural difference for each brain feature (S2 Methods) [10,25,26]. Quality controls were performed in each preprocessing step in a standardized manner (S2 Methods) [10,25].

The first COVID-19 case, the declarations of a state of emergency, and the definition of the four periods

Wuhan Municipal Health Commission, China, reported a cluster of pneumonia cases in Wuhan, Hubei Province, China, to the World Health Organization (WHO) on December 31, 2019 [27]. The first COVID-19 case in Japan was reported to WHO on January 16, 2020 [28]. The first declaration of a state of emergency in Tokyo, Japan, was from April 7, 2020, to May 25, 2020 [29]. The second was from January 8, 2021, to March 21, 2021; the third was from April 25, 2021, to June 20, 2021; and the fourth was from July 12, 2021, to September 30, 2021.

In this study, the four periods were defined as follows: (1) Period 1: July 2019 to December 2019. This is the period before the COVID-19 pandemic. (2) Period 2: January 2020 to May 2020. This period includes the first COVID-19 case in Japan and the first declaration of a state of emergency. (3) Period 3: June 2020 to December 2020. This is the period during which no state of emergency was declared. (4) Period 4: January 2021 to September 2021. This period includes the second, third, and fourth declarations of a state of emergency.

Preprocessing of time-series K6 scores

First, the acquisition time of the K6 score data was rounded to the nearest day. Second, because the questionnaire asked about participants’ status in the past 30 days, we extended the same data to the previous 30 days, assuming that the same status continued during this period. If the extension resulted in more than one data point on the same day, the data were averaged. Third, using gaussian process regression, the data were further interpolated to monthly data. The kernel function was defined using a combination of the sklearn.gaussian_process.kernels class in the python scikit-learn (https://scikit-learn.org), RBF() + ConstantKernel() + WhiteKernel(). Fourth, the data were binarized to perform the energy landscape analysis. The average was computed for each item per participant over the entire period. Using this average, the time series was binarized and assigned a value of 1 if above average and 0 otherwise. Values were set to 0 for items for which all values were equal. We express an individual response as an N-dimensional vector (hereafter referred to as a “response vector”) whose components consist of binarized responses to N questionnaire items. This preprocessing workflow is illustrated in S2 Fig.

Stratifying participants using time-series K6 scores

Participants were stratified using k-means clustering based on preprocessed time-series K6 scores (Python scikit-learn). We also computed the within-cluster sum of squared errors (SSE) to assess the accuracy of the clustering. To see the difference in depressive symptoms at the times of the MRI scans at Waves 3 and 4 between the clustering groups, we used a logistic regression model (S2).

Energy landscape analysis

The programming code we used to calculate the energy landscape is a Python version of the MATLAB code available from [30] along with a tutorial. The Python code is available on GitHub (https://github.com/ttttmmttddiikk/Energy_Landscape_Analysis) and Zenodo (https://doi.org/10.5281/zenodo.17585043).

Pairwise maximum entropy model.

We fit the pairwise maximum entropy model to the pooled individual response data in the same manner as in the previous studies [31]. Let be the set of all the possible “states” of -dimensional response vectors, where is the number of questionnaire items ( in the case of K6). has elements in total. For all , we assign the probability such that

(1)

The Shannon entropy is given by

(2)

where log is a natural logarithm. We maximize the entropy (2) under equation (1) and the following constraints (3) and (4):

(3)(4)

where , , , . Here, denotes the mean, and } denotes the th component of the th response vector in the pooled response data consisting of T individual responses. Using the method of Lagrange multipliers, we obtain the following equation about the model probability:

(5)

where

(6)

is the energy, and and are the parameters of the model ( is symmetric: ). Here, interactions higher than second order are ignored, hence the name pairwise maximum entropy model. Equations (5) and (6) match those of the Boltzmann distribution and the Ising model, respectively. The maximum-likelihood solution is expressed as

where is the likelihood given by

and is the th response vector in the pooled response data. We maximize the likelihood by a gradient ascent scheme given by

and

where the superscripts new and old represent the values after and before a single updating step, respectively, and is the learning rate. Updating was stopped when

became less than the permissible error (the first term in the numerator is the square of the 2-norm; the second term is the square of the Frobenius norm). We set and . For the monthly data, we set to reduce computation time. Initial values of and were set to 0.

The model variables and parameters , , , and are particularly important throughout the study. They are listed with explanations in S3 Table.

Accuracy indices.

To quantify the accuracy of the pairwise maximum entropy model fitted to the data, we used the following two accuracy indices [19,32,33]: The first accuracy index is defined by

where is the Shannon entropy of the maximum entropy model incorporating correlations up to the th order. The probability distribution represents the independent maximum entropy model in which we ignore any interactions between items by forcing for all . Distribution represents the pairwise maximum entropy model. Distribution is equivalent to the distribution of the given data, . The second accuracy index is defined by

where is the Kullback–Leibler divergence between two distributions and .

Statistical analysis of brain image features

The analysis of adolescent brain trajectory and its relationship with the clustering was performed using R software version 4.3.2. To explore the adolescent developmental trajectory of each brain feature, we built a GAMM implemented with the mgcv package (version 1.9-1) [36], including the main effect of sex as a linear term, the main effect of age as a smooth term using a penalized cubic regression spline and a basis function of 4, a tensor interaction term between age and sex, intracranial volume as a linear covariate, and a random intercept per participant (S2 Methods). Because the use of a subset of brain images (266 scans) from the 84 participants included in the energy landscape analyses may cause sampling bias in relation to the whole cohort (1,211 scans from 471 participants), we extracted the deviation from the nonlinear model for each scan and feature [37]. We then tested the difference between the groups from the clustering using GLMMs implemented with the lme4 package (version 1.1-35.1), including the main effect of group (G1 or G2), age × group interaction, child IQ, socioeconomic status, and handedness as covariates, and a random intercept per participant. Because we conducted 75 repeated tests, FDR correction was applied. To see whether the GHQ-28 scores at Wave 3 and/or the difference between waves was also associated with adolescent brain features, we further tested the same GLMMs. For the features with significant associations, we tested whether the group (G1 or G2) and/or the GHQ-28 scores could explain the brain difference and trajectory by performing the GLMMs adding the main effects of both variables and the interactions by age.

Ethics statement

This study was approved by the ethics committee of the Faculty of Medicine, The University of Tokyo (No. 10069), and the research ethics committee of The University of Tokyo (No. 22-396). Written informed consent was obtained from participants in the monthly web-based survey. After written informed consent was obtained, the participants’ parents were informed of their participation and had the opportunity to withdraw consent. For brain MRI scans, written informed consent was obtained from participants’ primary caregivers and participants 15 years of age or older prior to participation in this study.

STROBE guideline and TRIPOD statement

This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 Checklist) and the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis Or Diagnosis (TRIPOD) statement (S2 Checklist).

Collection of K6 questionnaire responses from adolescents before and during the COVID-19 pandemic

To analyze changes in adolescents’ emotional states during the COVID-19 pandemic, we obtained monthly responses to the K6 questionnaire [15], which measures nonspecific psychological distress, completed by 16- to 18-year-old high school students in Tokyo who participated in the pn-TTC from July 2019 to October 2021, starting 8 months before the first state of emergency declaration (Figs 1A and S1). Participants included 42 males and 42 females. The questionnaire contains 6 questions about the participants’ psychological distress over the past 30 days, and each item is rated on a 5-point scale (Fig 1B). The results showed Pearson correlation coefficients ranging from 0.53 to 0.81 between items (Fig 1C), suggesting that the items of the K6 questionnaire, whose total score indicates the level of psychological distress, were moderately to strongly correlated, as shown in previous validation studies [38–40].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Collecting K6 scores from high school students in Japan.

(A) The timeline of K6 responses is described for each participant (n = 84 participants and 1,278 total responses). Black dots indicate the date on which the response was received. (B) Questionnaire items and scales. (C) The correlation matrix between the item-level scores. Pearson’s correlation coefficients are shown in pseudo-colors ranging from orange to purple. Each item is arranged based on hierarchical clustering. (D) The time series of the item-level scores after data cleaning (see Methods, S2 and S3 Figs). All items from K6-1 to K6-6 are plotted with time on the x-axis and score on the y-axis for each participant. The thick line represents the mean. Error bars indicate standard deviation (SD).

https://doi.org/10.1371/journal.pmed.1004884.g001

Conventional statistical analyses of K6 responses have mainly focused on the total score and have not been able to fully exploit the multidimensional structure and dynamic features of the questionnaire. In order to analyze the dynamics of K6 responses in an integrated manner, encompassing six dimensions, we employed a method referred to as an “energy landscape analysis”. In this analysis, each questionnaire item is represented by a spin that takes the value of either 0 or 1 (in statistical physics, a spin usually takes the value of +1 or −1, but the mathematical model is the same except for differences in constants [41]). We assign an “energy” to each spin () and to the interaction of two spins (), and the total energy represents the probability of a specific combination of K6 responses (lower energy means higher probability; see Methods and S1 Note (Glossary) for further details).

Thus, the collected and preprocessed questionnaire responses (Figs 1D, S2, and S4) were binarized according to the procedures described in the Methods (illustrated in S3A Fig), and the binarized K6 scores were used in subsequent analyses. Here, binarization cutoffs were set to the individual mean and were therefore allowed to vary between individuals and between questionnaire items. See S2 Note and S5 Fig for a comparison of different binarization methods.

Energy landscapes for K6 scores during the COVID-19 pandemic

We first pooled the binarized K6 scores of all participants and constructed the energy landscape for the entire period, which we visualized as a “disconnectivity graph” (Methods and S3B Fig). We provide an explanation of a disconnectivity graph in Figs 2A and S1 Note. The accuracy of the model fitting is evaluated in S6A Fig.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Energy landscape analysis of time-series K6 questionnaire responses for all participants.

(A) An example of a disconnectivity graph (left) and the equivalent ball-and-cup diagram (right) are shown with explanations. Both display stable states and the energy barrier between them. Additional details on disconnectivity graphs can be found in S1 Note (Glossary). (B) The disconnectivity graph of the energy landscape for the entire period is shown. Each stable state is represented by a six-bit number consisting of the binarized values of K6-1, K6-2, K6-3, K6-4, K6-5, and K6-6 from left to right. The cutoff value of binarization was set to the individual mean; thus, 0 is below or equal to the individual mean and 1 is above the individual mean. (C) The graph of the basin of attraction obtained from the energy landscape analysis for the entire period. The stable states are plotted as squares. Each state is connected to the direction of attraction by an edge and each basin of attraction is shown in a different color. The basin 000000 is shown in red; the basin 111111 is shown in green. The density of the color corresponds to the magnitude of the energy. The darker the color, the lower the energy, and the lighter the color, the higher the energy. The six-bit number under each state shows the binarized values of K6-1, K6-2, K6-3, K6-4, K6-5, and K6-6 from left to right. (D) The definition of the four periods utilized in the analysis is explained, along with the number of school days per week for each participant. The thick line represents the mean. Error bars indicate standard deviation (SD). SoE refers to a state of emergency. (E) The mean of the total K6 scores for the 4 periods is shown. Error bars indicate standard deviation (SD). (F) The disconnectivity graphs of the energy landscape for the 4 periods are shown. (G, H) The simulated times from 000000 to 111111 or from 111111 to 000000 are compared between the 4 periods. 000000 (binarized responses of 0 to all six items) can be considered “healthy,” whereas 111111 (binarized responses of 1 to all six items) represents the “depressive” state. Box plots show the median (center line), interquartile range (box), and range (whiskers). The upper outliers are hidden to improve visibility. Yellow indicates increase compared with Period 1.

https://doi.org/10.1371/journal.pmed.1004884.g002

The disconnectivity graph was drawn with the stable states shown in Fig 2B. Here, each state is represented by a six-bit number, each bit of which represents K6-1, K6-2, K6-3, K6-4, K6-5, and K6-6 from left to right. The stable states, defined as local minima in the energy landscape (i.e., state(s) with lower energy than all adjacent states), were 000000 (i.e., binarized responses to all 6 items were 0) and 111111 (i.e., binarized responses to all 6 items were 1), corresponding to the “healthy” and the “depressive” states (states of psychological distress). The fact that the two stable states had the same digits across all items is consistent with the relatively strong correlation of the K6 questionnaire items. Indeed, all the estimated coefficients of interaction () were positive (S4 Table), which means that states become more stable when different digits take the same values. Furthermore, to quantify the uncertainty of the estimated parameters, we computed confidence intervals for both and and confirmed that these intervals were notably narrow, indicating stable estimates (S7 Fig). We also confirmed stable parameter estimation based on the convergence of estimation errors and the change in parameter estimates across iterations (S8 Fig). The energy difference between 000000 and 111111 was , meaning that the probability of 000000 was times greater than that of 111111 (i.e., lower energy means higher probability).

To see how all states are related in the energy landscape, we also plotted a forest structure (called the graph of basins of attraction) in which each state is connected to an adjacent state with the lowest energy (Figs 2C and S3B). Here, each state is represented by a six-bit number; an adjacent state of state A differs from A by only one bit. The structure consisted of two disjoint trees (called basins of attraction, hereafter “basins”) with root 000000 or 111111. Each basin consists of all states that lead to its root (stable state) by the steepest descent method, i.e., by moving downhill in the energy landscape. In other words, a basin is a set of states that are expected to be attracted to the same stable state. Therefore, the 000000 basin can be interpreted as the relatively healthier states (attracted to the healthy stable state 000000), and the 111111 basin as the relatively more depressive states (attracted to the depressive stable state 111111). The basin sizes of the 000000 and the 111111 basins were 40 and 24, respectively.

To explore the connection between students’ depressive states and their social context, we examined differences across the four periods outlined by the declarations of states of emergency (Fig 2D), considering that social interactions at high schools primarily occur on school days. The first state of emergency was declared in Period 2 (January to May 2020), whereas the second, third, and fourth declarations took place in Period 4 (January to September 2021). Notably, the number of school days during the states of emergency, especially in the first one, was significantly small. When we compared the mean of the total K6 score across the four periods, there was a significant trend indicating a decrease in the K6 score in Period 2 and then a slight increase in Periods 3 and 4 (Jonckheere–Terpstra trend test, ) (Fig 2E; see also S9 Fig). We speculate that an increase in time spent at home during the pandemic, especially in the first state of emergency, led to a reduction in social interactions. This reduction in social interactions may have helped to alleviate psychological distress for high school students, as previously mentioned in [42]. In addition, as demonstrated by our different questionnaire data presented in S10 Fig, we revealed a correlation between fewer school days per week and lower K6 scores.

However, the total K6 score alone does not reveal which combination of items contributed to the healthy and depressive states, nor how these states changed over time. To see this, we looked at individual K6 items and performed an energy landscape analysis on the pooled data for each of the 4 periods; the disconnectivity graphs were drawn similarly (Fig 2F; the accuracy is evaluated in S6B Fig). The stable states included 000000 and 111111 for all periods. There were also other stable states with higher energy (lower probability), with an energy difference of >3 from 000000 (i.e., more than 20 times lower probability). The energy of 000000 is 0 from the definition of the model. The energy of 111111 was lower in Period 1 (2.40) than in the other periods (2.90, 2.92, and 2.71), meaning the probability of 111111 was higher before the pandemic and became lower after the onset of the pandemic. Note that, when the K6 scores were analyzed item by item, use of traditional statistical methods such as the Jonckheere–Terpstra trend test, Tukey’s test, or Mood’s median test did not reveal significant differences among the different periods (S5 Table), suggesting that these methods were not sensitive enough to detect item-level differences. To further validate the findings from the energy landscape analysis, we performed a latent class analysis based on total K6 scores, which independently reproduced the same temporal pattern (see S3 Note and S11 Fig).

To examine how participants moved between different states over time, we simulated state transitions constrained by the energy landscape using Gibbs sampling (Methods), based on the simplifying assumption of detailed balance [43]. The logarithm of the average dwell time of each state was a linear function of its energy with a negative slope, consistent with the theoretical prediction (the probability of states follows a Boltzmann distribution in the model) (S12 Fig). Notably, when we started the simulation at 000000 and continued until the state reached 111111, the mean time from 000000 to 111111 became significantly longer in Periods 2–4 than in Period 1 (Fig 2G). In contrast, when we started the simulation at 111111 and continued until the state reached 000000, the mean time from 111111 to 000000 was not significantly different among the four periods (Fig 2H). These results suggest that it became more difficult to move from 000000 to 111111 after the pandemic started.

The above trend in the simulations was confirmed by a statistical analysis of the two basins (000000 and 111111) (S4 Note and S13AB Fig). The probability of staying in the healthy 000000 basin increased in Periods 2–4 compared to Period 1 (S13A Fig), and the monthly transition probability from the depressive 111111 basin to the healthy 000000 basin increased in Periods 2–4 compared to Period 1 (S13B Fig), suggesting that students were more likely to be in the healthy 000000 basin after the pandemic.

We further constructed energy landscapes on the pooled data for each month (from July 2019 to September 2021, S14A Fig), although we note that the accuracy of landscape construction may have been lower due to a small sample size (see also S15 Fig for confidence intervals of energy estimates). Nevertheless, the energy difference between 000000 (healthy state) and 111111 (depressive state) was negatively correlated with the mean number of school days per week (r = −0.35, p = 0.071) (S14BC Fig), again suggesting that students were less depressed when school days per week were fewer.

Energy landscapes for diverse psychological distress to the COVID-19 pandemic

The patterns of psychological responses to crises, including to the COVID-19 pandemic, vary greatly among individuals [44]. To gain deeper insights into such diverse responses, we next stratified participants using k-means clustering, with a focus on the time series of total K6 scores (Figs 3A and S16). We identified two distinct groups: G1 consisted of 61 participants whose total K6 score was relatively low (less than 5) and stable over time, and G2 consisted of 23 participants whose total K6 score was higher (with most being higher than 5) and less stable (Figs 3B and S17) (Welch t test with Benjamini–Hochberg correction; , , , for Periods 1–4, respectively). Some participants in G2 maintained a total K6 score above 5 across all periods, while others had a total K6 score below 5 in at least one period (S9 Fig). The number of participants in the four periods, categorized by the clinically relevant cutoffs (5, 8, 13) of the total K6 score (the cutoff values were chosen based on [22]), is shown in S6 Table. G1 included 36 males and 25 females, while G2 included 6 males and 17 females. Similar to the findings for the K6, G2 showed higher scores on the 28-item version of the GHQ (GHQ-28) [16] (S1 Methods and S7 Table), which assesses more subjective affective symptoms than does the K6. G2 also showed a greater increase than G1 in the GHQ-28 score between the two brain MRI scans. No significant differences were observed between G1 and G2 in IQ or parental socioeconomic status (S7 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Energy landscape analysis in the stratified groups.

(A) Clustering of time-series total K6 scores. Total K6 scores are plotted with time for each participant. The thick line represents the mean. Error bars indicate standard deviation (SD). The colors correspond to the stratified groups G1 and G2. G1 (plotted in pink) tended to have lower and more stable total K6 scores, whereas G2 (plotted in blue) tended to have higher and more variable scores. (B) The mean of the total K6 score for G1 and G2 is shown over the 4 periods. Periods 1–4, defined in Fig 2D, are as follows: Period 1 (July–December 2019), before the COVID-19 pandemic; Period 2 (January–May 2020), including the first COVID-19 case in Japan and the first state of emergency; Period 3 (June–December 2020), when no state of emergency was declared; and Period 4 (January–September 2021), including the second to fourth state of emergency declarations. Error bars indicate standard deviation (SD). (C, D) The graphs of the basin of attraction are drawn separately for G1 and G2 participants. (E, F) The disconnectivity graphs of the energy landscape for G1 and G2 are shown for the entire period. (G, H) The disconnectivity graphs of the energy landscape for G1 and G2 are shown for the 4 periods. (I–L) The simulated times from 000000 to 111111 or from 111111 to 000000 for G1 and G2 are compared between the 4 periods, respectively. 000000 (binarized responses of 0 to all six items) can be considered “healthy,” whereas 111111 (binarized responses of 1 to all six items) represents the “depressive” state. Box plots show the median (center line), interquartile range (box), and range (whiskers). The upper outliers are hidden to improve visibility. Yellow and blue indicate increase and decrease, respectively, compared with Period 1.

https://doi.org/10.1371/journal.pmed.1004884.g003

We constructed separate energy landscapes for G1 and G2 and compared the basins derived from the energy landscapes (Fig 3C and 3D). The stable states for both groups were 000000 (i.e., healthy state) and 111110/111111 (i.e., depressive(-like) state). The sizes for the 000000 and 111110/111111 basins were 47 and 17 for G1, and 33 and 31 for G2, respectively.

We compared the disconnectivity graphs of G1 and G2 drawn from the energy landscape (Fig 3E and 3F: the accuracy indices are shown in S6C–S6E). The energy of 111111 or 111110 (depressive(-like) state) was higher in G1 (G1: 4.10, G2: 0.54), consistent with G1 participants being less likely to be in the depressive state. When energy landscapes were constructed for the 4 periods, the stable states included 000000 and 010101/010110/011110/111110 (depressive-like states) for G1 participants and 000000, 010101, and 111111 for G2 participants (Figs 3G, 3H, and S6D–S6F). Note that G1 participants had a very small probability of being in state 111111 (fully depressive), and thus 111111 was not a stable state; either K6-1 or K6-6 was 0 in the stable state (i.e., their emotional state lacked nervousness or feelings of worthlessness). The energy of 000000 is 0 as defined in the model. The energies of the depressive-like states in G1 peaked in Period 2 and then decreased (3.57, 4.57, 3.60, and 3.76), but in G2 were at about the same level as 000000 (−0.39, 0.70, 0.50, and 0.72), indicating that G1 participants were least likely to exhibit depressive symptoms in Period 2, whereas G2 participants were equally likely to be healthy or depressed over the 4 periods.

To examine how G1 and G2 participants moved between different states over time, we performed simulations of state transitions constrained by the energy landscape of G1 or G2. Two simulations showed an interesting difference: In the G1 landscape, the mean time from 000000 to 111111 increased more than 2-fold in Periods 2–4 compared with Period 1 (Fig 3I), but the mean time from 111111 to 000000 did not differ significantly across the four periods (Fig 3J). In contrast, in the G2 landscape, the mean time from 000000 to 111111 did not significantly differ across the four periods (Fig 3K), whereas the mean time from 111111 to 000000 decreased more than 2-fold in Periods 2–4 compared with Period 1 (Fig 3L). These results suggest that the pandemic states of emergency prevented the G1 participants from becoming depressed, and at the same time triggered the G2 participants to move into a healthy state.

Again, this trend in the simulations was also confirmed by a statistical analysis of the two basins (the healthy 000000 basin and the depressive 111110(G1)/111111(G2) basin) (S13C-F Fig). The probability of staying in the healthy basin increased in Periods 2–4 compared to Period 1 in both G1 and G2 (S13C-D Fig), and the monthly transition probability from the healthy basin to the depressive basin decreased in G1 (S13E Fig), whereas the monthly transition probability from the depressive basin to the healthy basin increased in G2 (S13F Fig).

Biological characterization of diverse psychological responses to the COVID-19 pandemic

We further tested whether brain structural features and adolescent developmental trajectories differed between the G1 and G2 groups. In the pn-TTC, a total of 479 participants underwent brain MRI scans, of which a subset (84 participants) participated in a monthly web-based questionnaire survey (Figs 4A and S1 and S2 Methods). After modeling nonlinear developmental trajectories of each brain feature using GAMMs with participant-level random effects and extracting individual deviations from these models, GLMM analyses revealed a decreasing trend in the cortical thickness of the caudal part of the middle frontal gyrus (cMFG) during adolescence, and the degree of decrease was greater in G2 than in G1 after false discovery rate (FDR) correction (; Fig 4C). On the other hand, the increase in GHQ-28 scores between Waves 3 and 4 was associated with an increase in cortical thickness in the temporal pole (TP) and entorhinal cortex (; (FDR correction); Fig 4H) and a decrease in surface area in the same regions (; (FDR correction)). There was no significant association in the GHQ-28 score at Wave 3 (Fig 4E and 4F). For the three psychological features (energy-landscape-related G1/G2 clustering based on the K6 items, GHQ-28 scores at Wave 3, and difference in the GHQ-28 scores between Waves 3 and 4), there was no significant model for the main effect for any brain features.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Developmental trajectories of brain features in stratified groups.

(A) Summary of data collection schedule (see also Methods, Figs 2D and S1 and S1 Methods). IQ, intelligence quotient; SES, socioeconomic status; SoE, state of emergency. (B) cMFG and TP in the brain cortex are visualized along with t-values in the GLMM analysis (S8 and S9 Tables). cMFG, caudal part of the middle frontal gyrus; TP, temporal pole. (C–H) Developmental trajectories of brain features for each group are plotted with age at MRI (y) on the x-axis and cortical thickness on the y-axis for each participant. The thick line represents the estimated value by general linear mixed models (GLMMs). (C) A greater decrease in the cortical thickness of cMFG is observed in G2 than in G1. G1 and G2 represent the clustering groups identified in Fig 3; G1 showed lower and more stable total K6 scores, whereas G2 showed higher and more variable scores. (D) A greater increase in the cortical thickness of TP is observed in G2 than in G1. (E, F) Cortical thickness of cMFG and TP in 479 participants, colored by GHQ-28 score (low/high) at Wave 3. (G, H) Cortical thickness of cMFG and TP in 479 participants, colored by the difference in GHQ-28 score (decrease/increase) between Waves 3 and 4. The illustration of the human brain in Fig 4 was drawn using BrainPainter (https://github.com/razvanmarinescu/brain-coloring).

https://doi.org/10.1371/journal.pmed.1004884.g004

We next tested whether the G1 or G2 group or the difference in GHQ-28 scores was more significantly associated with change in five brain features (cortical thickness of cMFG, cortical thickness and surface area of entorhinal cortex, and cortical thickness and surface area of TP) by simultaneously adding the psychological features to the GLMMs. The models showed significant age × group interactions, but not age × GHQ-28 score difference interactions, on the cortical thickness of cMFG (Figs 4B and S18 and S8 Table and S5 Note; (FDR correction); ) and TP (Figs 4B and S18 and S9 Table and S5 Note; (FDR correction); ), indicating the direction of accelerated brain maturation in G2 compared with G1.

Taken together, the energy-landscape-related group was associated with the change in the cortical thickness in the cMFG and TP, with G2 (the group with higher and unstable K6 scores) showing a greater change. This suggests that morphological changes in these brain regions may be associated with increased sensitivity to psychological fluctuations.