Additional Insights into the Relationship Between Brain Network Architecture and Susceptibility and Resilience to the Psychiatric Sequelae of Childhood Maltreatment

Abstract

Childhood maltreatment is associated with increased risk for psychiatric and substance use disorders. However, some maltreated individuals appear resilient to these consequences while manifesting the same array of brain changes as maltreated individuals with psychopathology. Hence, a critical issue has been to identify compensatory brain alterations in these resilient individuals. We recently reported that maltreatment is associated with a more vulnerable structural brain network architecture. Resilient individuals have this same vulnerability but appeared to be able to effectively compensate due to reduced nodal efficiency (ability of a node to influence the global network) in 9 specific brain regions that moderate the relationship between maltreatment and psychopathology. Following up we now report that network vulnerability increases progressively during late adolescence to plateau at about 21 years of age, which may help to explain age of onset of psychopathology. Further, we found that network vulnerability was most significantly affected by parental verbal abuse between 16–18 years of age and number of types of maltreatment during childhood. Asymptomatic individuals with no history of psychopathology had more prominent alterations in nodal efficiency than asymptomatic individuals with prior history, who specifically showed reduced nodal efficiency in right amygdala and right subcallosal gyrus. Experiencing inadequate financial sufficiency during childhood increased risk of susceptibility versus resilience by 2.98-fold (95% CI 1.49–5.97, p = 0.002) after adjusting for differences in exposure to maltreatment. Interestingly, adequate-to-higher financial sufficiency appeared to be protective and was associated with reduced nodal efficiency in the right postcentral gyrus and subcallosal gyrus ‘resilience’ nodes.

Childhood maltreatment is the most important risk factor for psychopathology. Adverse childhood experiences in the form of maltreatment and household dysfunction account for about 33%, 54%, 64% and 67% of the population attributable risk for anxiety disorders (Green et al., 2010), depression (Anda et al., 2002), addiction to illicit drugs (Shanta R Dube et al., 2003) and suicide attempts (S R Dube et al., 2001), respectively. Abuse and neglect are also recognized as major risk factors for personality disorders (Afifi et al., 2011; Bierer et al., 2003; Cutajar et al., 2010; Herman et al., 1989; Lyons-Ruth et al., 2013; Widom et al., 2009; Zanarini et al., 1997) and to substantially increase risk for psychosis (Cutajar et al., 2010; Arseneault et al., 2011; Bebbington et al., 2011; Bendall et al., 2008; Daly, 2011; Fisher et al., 2010; Gaudiano and Zimmerman, 2010). These strong associations justify a rigorous effort to understand how maltreatment gets ‘under the skin’ to increase risk in some individuals and a comparable effort to identify neurobiological and experiential factors associated with better than expected outcomes in other individuals in order to identify youths at greatest risk, to preempt development of psychiatric disorders and to enhance neurobiological substrates of resilience.

We have pioneered and pursued the hypothesis that maltreatment increases risk for substance use disorders (SUDs) and psychopathology by altering trajectories of brain development. A growing body of reproducible findings, from our lab and others, links maltreatment to brain differences. Consistent findings are smaller midsagittal area (Teicher et al., 1997; De Bellis et al., 1999; De Bellis et al., 2002; S. L. Andersen et al., 2008; Teicher et al., 2004) or decreased fractional anisotropy (Jackowski et al., 2008; Teicher et al., 2010) of the corpus callosum and lower hippocampal volume in adults (Bremner et al., 1997; Frodl et al., 2010; Stein et al., 1997; Vermetten et al., 2006; Vythilingam et al., 2002; Weniger et al., 2009; S. L. Andersen et al., 2008; Brambilla et al., 2004; Driessen et al., 2000; Schmahl et al., 2003; Opel et al., 2014; Teicher et al., 2012a) but not necessarily in youths (De Bellis et al., 1999; De Bellis et al., 2002; Carrion et al., 2007; Tupler and De Bellis, 2006). Maltreatment is also associated with attenuated size of the anterior cingulate (Cohen et al., 2006; De Bellis et al., 2000; Tomoda et al., 2009b; Bremner et al., 2003), orbitofrontal (L. M. Shin et al., 1999; Bremner et al., 2003; Hanson et al., 2010) and dorsolateral prefrontal cortex (Sheu et al., 2010; Tomoda et al., 2009b), and enhanced amygdala response to threat (Dannlowski et al., 2013; Dannlowski et al., 2012; Zhu et al., 2019) and blunted striatal response to reward anticipation (Dillon et al., 2009; Mehta et al., 2010; Boecker et al., 2014). We direct interested readers to a number of comprehensive reviews (Teicher and Samson, 2016; Teicher et al., 2016; Nemeroff, 2016; Hart and Rubia, 2012).

However, a perplexing problem has emerged. Although maltreatment has robust effects of brain structure, function and connectivity and is a critically important risk factor for substance use and psychopathology it has not been possible to draw direct connections between brain regions affected and clinical outcomes. What complicates the linkage is the vast number of studies reporting that the maltreatment-associated abnormalities enumerated above are statistically unrelated to presence or absence of psychopathology in mixed samples (Teicher et al., 2004; S. L. Andersen et al., 2008; van Harmelen et al., 2010; Teicher et al., 2012a; Tomoda et al., 2012; Heim et al., 2013; van der Werff et al., 2013; van Harmelen et al., 2013; Chaney et al., 2014; Opel et al., 2014; Van Dam et al., 2014; van Harmelen et al., 2014; Hanson et al., 2015; Ugwu et al., 2014) and discernible even in studies of maltreated individuals with no evidence of psychopathology (Cohen et al., 2006; Paul et al., 2008; Seckfort et al., 2008; Teicher et al., 2010; Edmiston et al., 2011; Carballedo et al., 2012; Dannlowski et al., 2012; Everaerd et al., 2012; Frodl et al., 2012; Gerritsen et al., 2012; Huang et al., 2012; L. M. Baker et al., 2013; Philip et al., 2013; Samplin et al., 2013; Hanson et al., 2015). Briefly, alterations reported in maltreated individuals without apparent psychopathology include: reduced volume of hippocampus, amygdala, anterior and subgenual cingulate, orbitofrontal cortex, dorsolateral and dorsomedial prefrontal cortex, insula and cerebellum as well as decreased integrity of corpus callosum, superior longitudinal fasciculus and cingulum bundle, enhanced amygdala response to threat and blunted striatal response to reward anticipation (Teicher et al., 2016).

These findings led us to conclude that maltreated individuals with better than expected outcomes were not unaffected but were effectively compensated (Teicher and Samson, 2016; Teicher et al., 2016). The critical question is how do they manage, neurobiologically, to maintain mental health despite a host of abnormalities in stress susceptible structures? Our suspicion was that to answer this question we needed to use an approach that would enable us to simultaneously consider the interconnections between the large number of regions that appear to be affected in maltreated individuals. Hence, we conducted the first large scale neuroimaging studies of brain network architecture with hundreds of maltreated participants.

However, to better understand the findings we need to provide some background information for readers unfamiliar with graph theory and network analysis. Interestingly, network analysis had its origins in social psychology and sociology with the study of social networks, which laid out in graphic form the interrelationships between individuals (e.g., who knows whom) (Fienberg, 2012). In such a network each individual is a node (also called a vertex). The lines connecting individual to each other are edges or links, which in an unweighted network are either present or absent (e.g., A and B know each other, A and C do not). The number of edges associated with a node is the degree centrality of the node. Typically, some individuals in a social network will have many more edges (e.g., acquaintances) than other individuals. A hub is a node with a much greater than average number of edges. The presence and distribution of hubs plays an important role in delineating the topology of the network.

In many networks there are a number of hubs that are more connected to each other than to other nodes in the network. This collection of hubs is called the ‘rich club’ and it serves as the communication backbone of the network. One can imagine that vital information or rumors spread quickly throughout the network when the information reaches someone in the ‘rich club’. Nodal efficiency is another measure of the centrality or importance of a node and specifically indicates the ability of a node to spread information to the other nodes in a network.

Social networks, brain networks and computer networks often share an important property known as small worldness, which derived from the idea that all people are six, or fewer, social connections away from each other (i.e., six degrees of separation). Networks with small world properties are organized into clusters of communities (also called modules) that consist of locally interconnected nodes with rich club hubs interconnecting the different communities.

In small world networks there is a balance between integration, which takes place within communities (referred to as local clustering), and separation, which takes place between communities. This organization provides an optimal design for brain networks as connections are metabolically costly. This design enables communication to occur between brain regions (nodes) through only a relatively small number of steps, though many nodes will only have a few local connections. The average of the smallest number of steps that it takes each node to connect to every other node in the network is known as the average path length. The global efficiency of a network is defined as the inverse of the average path length. The degree of small worldness reflects the ratio of local clustering to average path length.

Another property of networks is their vulnerability. One way this is calculated is to assess how much the removal of a node from a network effects the global efficiency of the network by increasing the average path length. The network’s vulnerability is then defined as the consequence of removing the node with the greatest impact.

Applying these concepts to brain networks in maltreated and non-maltreated individuals have been revealing. First, we reported in an examination of cortical network architecture, that regions comprising the highly interconnected communication backbone (‘rich club’) were almost entirely different in maltreated and non-maltreated participants, with a paucity of frontal hubs in the maltreated network (Teicher et al., 2014a). Second, we discovered using diffusion tensor imaging and tractography that maltreated individuals had a sparser fiber stream network architecture. Clustering at the community level did not differ but there were fewer interconnections between communities in the maltreated network. This was associated with reduced global efficiency, increased small worldness and increased vulnerability to disruption (K. Ohashi et al., 2017b). Third, we then showed in an expanded data set (N=342) that asymptomatic maltreated individuals (without substance use or clinically significant symptoms on the Kellner Symptom Questionnaire (Kellner, 1987)) had the same array of abnormalities in global network architecture as maltreated individuals with clinically significant symptoms (K. Ohashi et al., 2019). Fourth, we found that resilient maltreated individuals had reduced nodal efficiency (N_eff – ability to propagate information throughout the network) of their right amygdala relative to susceptible subjects and unexposed controls (K. Ohashi et al., 2019). Fifth, we identified 8 other nodes with reduced N_eff in resilient individuals and found that they moderated susceptibility / resilience to different sets of symptoms and degree of risk for different classes of drug (K. Ohashi et al., 2019). Interestingly, the AI model used N_eff to distinguish symptomatic from asymptomatic maltreated participants. N_eff in the identified nodes associated with resilience could have been increased in some nodes and decreased in others but it turns out that it was reduced in all the identified nodes, which we did not anticipate. For convenience we refer to these 9 nodes associated with susceptibility/resilience as ‘resilience’ nodes.

It may seem confusing that reduced N_eff was associated with resilience as a normal inclination would be to believe that resilience should stem from enhanced connectivity. However, when it comes to the brain, bigger may not be better and less may be more. Postnatal brain development goes through both an overproduction and a pruning phase, with marked increments in efficiency and performance resulting from the strategic pruning of connections. In network design it is not beneficial for nodes to have as many connections as possible, they need to have an optimal number and configuration of connections. In a sparse network with fewer frontal hubs it may well be optimal for certain nodes to have reduced connectivity or nodal efficiency.

These finding are easily understood if the identified nodes were functioning in a problematic manner even in asymptomatic maltreated individuals. A priori we focused on the right amygdala as studies have shown that this region has a hyperactive response to negative stimuli in maltreated individuals whether or not they are symptomatic (van Harmelen et al., 2013; Dannlowski et al., 2013; Dannlowski et al., 2012). As another interesting example we found that the next most important node in the model, after the right amygdala, was the left frontal inferior gyrus pars triangularis. This region contains part of Broca’s area and we postulated that it may be responsible for the hypercritical ‘inner voices’ or thoughts that maltreated individuals often report experiencing. In individuals with N_eff values in the left pars triangularis comparable to non-maltreated controls there was a strong association between degree of exposure to maltreatment and severity of depression and anxiety. In individuals with low N_eff in this region there was a much weaker relationship resulting in no significant increase in symptom scores. In asymptomatic maltreated individuals low N_eff in the pars triangularis appeared to be specifically associated with reduced connectivity of this region to the right limbic system and right temporal lobe (K. Ohashi et al., 2019).

These findings led us to propose that psychopathology and substance use emerge in maltreated individuals when a sparse, vulnerable network can no longer effectively compensate for abnormalities in one or more nodes. Resilience results from effective compensation which can occur if the impact of certain problematic nodes are diminished.

We then evaluated the strength of this model by determining how well 5 measures of global network architecture and 9 measures of N_eff could predict whether individuals were unexposed controls, maltreated-resilient or maltreated-susceptible and found using multiclass AI predictive analytics that it could do so with 75%, 80% and 82% balanced, cross-validated accuracy, respectively versus 33% chance (p<10⁻⁷) (K. Ohashi et al., 2019). This finding was further confirmed using a balanced split into training and test sets. Together they suggest that this is a promising approach to understanding the neurobiology of susceptibility/resilience to the psychiatric consequences of maltreatment.

However, a number of important questions were left unanswered in our initial report. A key concern is how does network vulnerability changes over the course of development in maltreated individuals. Our hypothesis is that network vulnerability increases progressively in maltreated individuals during adolescents and early adulthood as a consequence of pruning processes and failure to develop frontal hubs that typically emerge between 15–19 years of age (S. T. Baker et al., 2015). A second important issue is to more specifically delineate experiential factors that underlie these maltreatment-associated alterations in brain network architecture. In particular, can we identify developmentally sensitive periods when exposure to specific types of maltreatment have maximal influence on measures of network vulnerability and small world organization. A third developmental concern is whether differences in N_eff in these ‘resilience’ nodes is a preexisting protective factor or whether these are compensatory changes that emerge following exposure to maltreatment or in response to the onset of psychiatric symptoms. We suspect that reduced levels of N_eff in specific nodes may be preexisting alterations in some but emerge as compensatory changes in others. This should then result in different trajectories of response to maltreatment. Briefly, Bonanno and colleagues (Bonanno and Diminich, 2013; Bonanno et al., 2012b) have done some elegant work using latent growth mixture models to identify response trajectories. They have identified three key patterns. One is an ‘minimal impact’ pattern in which individuals have a slight increase in symptomatology during exposure to adversity which rapidly abates when the exposure is over. Another is an ‘emergent’ or ‘recovery’ pattern in which there is a substantial increase in symptoms during exposure which returns to preexposure levels after the exposure ends during the observation period. The third pattern is ‘chronic’ as there is a substantial increase in symptoms that persist after the exposure ends during the observation period. We propose that individuals with preexisting low N_eff values in certain nodes would follow a minimal impact pattern whereas individuals with a gradual compensatory reduction in N_eff would follow an emergent or recovery resilience pattern. A final concern is to ascertain whether there are additional experiential factors, such as poverty or parental education, that tip the balance between susceptibility / resilience to psychopathology in maltreated individuals and if we can delineate network alterations associated with these moderating factors.

Methods and Materials

Subject Recruitment

This article consists of additional analyses conducted on a previously published study (K. Ohashi et al., 2019). The Partners Healthcare institutional review board approved this study and all subjects provided informed consent. Recruitment followed previously reported methods (K. Ohashi et al., 2019; Kyoko Ohashi et al., 2017a; Teicher et al., 2014b; Teicher et al., 2012b). Briefly, subjects were recruited by advertisements. Potentially interested participants were phone screened to be right-handed, medically healthy, unmedicated (except for hormone replacement, contraceptives and occasional use of albuterol inhalers or non-sedating antihistamines) and between 18–25 years of age. Those who appeared eligible were invited to log onto a HIPAA-compliant enrollment system to provide detailed information on demographics, medical, psychiatric and developmental history, life experiences, psychiatric symptomatology and history of childhood maltreatment. Those that meet criteria were invited to the laboratory for further evaluation. Subjects were selected based on their degree of exposure to maltreatment regardless of psychiatric outcome in order to include a representative distribution of participants with high and low degrees of psychopathology. Participants with exposure to three or more types of maltreatment were oversampled.

Exclusion criteria included: history of neurologic disease, concussion or head trauma resulting in loss of consciousness for more than 5 minutes, multiple unrelated forms of adversity including natural disaster, motor vehicle accidents, near drowning, house fire, mugging, witnessing or experiencing war, gang violence or murder, riot, or assault with a weapon or animal attack. Additionally, high levels of current drug or alcohol use, or previous use of heroin or cocaine were grounds for exclusion.

Overall, 2188 subjects provided on-line information. From this group we interviewed 670 subjects, and from this interviewed pool we selected the neuroimaging sample. The large number of subjects screened was not specifically required to provide the required number of participants with moderate-to-high levels of maltreatment. Most online screened subjects were eliminated due to histories of head injury / possible concussion, exposure to multiple types of trauma (e.g., natural disasters, motor vehicle accidents), prematurity or birth complications and binge drinking.

Subject Assessments

Data were collected on race, ethnicity, parental education and perceived financial sufficiency during childhood (Choi et al., 2009), which was rated from 1 (much less than enough money than to meet family’s needs) to 5 (much more than enough money to meet family’s needs). Structured Clinical Interviews for DSM-IV Axis I and II psychiatric disorders (SCID I and II) (First et al., 1997) were used to establish lifetime psychiatric diagnoses.

Maltreatment and Abuse Chronology of Exposure Scale (MACE).

The MACE was created to assess severity of exposure to 10 types of MAL (including peer physical and emotional bullying) across each year of childhood (Teicher and Parigger, 2015). It was developed using item response theory in which each MAL category fits a Rasch model that delineates a latent trait and provides a severity of exposure scale with at least interval scaling properties (Jackson et al., 2014). MACE scores have high test-retest reliability (r=0.91, N=75) and do not show significant negative attribution bias (Teicher and Parigger, 2015). Timing of exposure across years also has high reliability (range 0.722 –.994) with highest test-retest reliabilities for types of MAL that are usually episodic and perhaps most ingrained as specific events (e.g., sexual abuse r=.994; peer physical abuse r=.954; peer verbal abuse r=.884; physical neglect r=.859 and parental physical abuse r=.803). Rating of abuse, in the aggregate, had acceptable test-retest reliability from age 2–3 on (Teicher et al., 2018). This fits with the observation that age of onset for adult memories of childhood typically occurs between 2 and 4 years of age and are earlier for salient events so that hospitalizations and birth of siblings can be accurately recalled if they took place at age 2 and moving homes at age 3 (Usher and Neisser, 1993). Ratings of neglect were reliable from age 1 on (Teicher et al., 2018) with subjects making inferences about what they believed their family was like for ages they were too young to remember (e.g., availability of family members to take care of you and protect you, having enough to eat, etc.).

Kellner Symptom Questionnaire (SQ).

The SQ is a 92-item self-report scale that provides current ratings of depression, anxiety, somatization and anger-hostility as well as measures of well-being (Contented, Relaxed, Physically Well and Friendly) (Kellner, 1987). It has good psychometric properties, was specifically designed to detect change in symptoms over time in treatment trials (Kellner, 1987) and we have used it extensively in our research studies (Teicher et al., 2010; Khan et al., 2015; Polcari et al., 2014; Teicher et al., 2006). The scale is very sensitive as it detects both symptoms (e.g., sad, blue) as well as absence of well-being (e.g., happy) and can even distinguish between euthymic bipolar patients and healthy controls (Kellner, 1987). Scores of 12 or above are taken to connote presence of clinically significant symptoms in each domain.

Criteria for susceptibility or resilience

Subjects’ reporting exposure to two or more types of maltreatment on the MACE were designated as ‘at risk’, as this degree of exposure increased odds of lifetime history of MDD by 4.67-fold (95% CI 2.68–8.39, p < 10⁻⁸). Subjects were further partitioned into those with presence or absence of clinically significant symptoms on Kellner’s SQ, as it resulted in the delineation of symptomatic and asymptomatic at-risk maltreated groups who did not differ in their degree of exposure to any type of maltreatment (K. Ohashi et al., 2019). Further, the asymptomatic at-risk group did not differ from the not at-risk group in Kellner SQ scores or in degree of alcohol consumption or risk for polysubstance abuse.

Criteria for Minimal Impact versus Recovery Groups

For these analyses we included our previously defined not at-risk controls (N=118) who were all currently asymptomatic (but could have had past psychiatric episodes) and symptomatic maltreated participants (N= 106) (K. Ohashi et al., 2019). Note 24 subjects (17%) with no-to-low exposure to maltreatment were symptomatic and excluded from the control group. We divided the asymptomatic maltreated participants (N=86) into two separate groups. One group was defined as asymptomatic – minimal impact and it consisted of at-risk individuals (based on MACE) who were currently asymptomatic on the SQ and had no lifetime history of Axis I or II psychiatric disorders on SCID I and II (N=31). The second group was defined as asymptomatic –recovered and consisted of at-risk participants who were currently asymptomatic on the SQ but had a previous history of Axis I or II psychiatric disorders on SCID (N=55).

MRI Data Acquisition

MRI scans were acquired using 3T Siemens TIM Trio (Erlangen, Germany) using previously reported methods (Kyoko Ohashi et al., 2017a) (see S3). Briefly, multiple diffusion-weighted images were acquired in 72 directions. Scan parameters were: b=1000 sec/mm2; echo time (TE)/repetition time (TR)=81 msec/6sec; matrix=128×128 on 240mmx240mm field of view (FOV); slices 3.5mm without gap.

MRI Analysis and Network Construction

Unweighted 90-node fiber stream networks were constructed using previously published methods (K. Ohashi et al., 2019; Kyoko Ohashi et al., 2017a). Graph theory was used to calculate global network parameters including: global efficiency; degree; vulnerability and small-worldness. These network parameters have been described in detail in previous papers (K. Ohashi et al., 2019; Kyoko Ohashi et al., 2017a; Anderson and Cohen, 2013; Bassett and Bullmore, 2006; Bullmore and Sporns, 2009; Bullmore et al., 2009; Latora and Marchiori, 2001; Rubinov and Sporns, 2010; Stam and Reijneveld, 2007) (R package: igraph, brainGraph).

Statistical Procedures

Changes in Network Vulnerability Across Age in Maltreated Individuals.

Cross-sectional measures of network vulnerability were assessed between 18 and 25 years of age using ANCOVA and were fit to a polynomial model.

Identification of Sensitive Exposure Periods.

A critical question is whether exposure to particular types of maltreatment at specific ages are important risk factor for alterations in global network architecture. Conventional analytic techniques, such as multiple regression, are not suitable because of marked collinearity in the degree of exposure to specific types of maltreatment at adjacent ages. Instead, we identified the most important cross-validated ‘predictors’ associated with global network measures using random forest regression with conditional inference trees (RFR-CIT) (cforest in R package party), an artificial intelligence machine learning strategy that is highly resistant to collinearity, which we have used in an expanding series of sensitive period studies (Khan et al., 2015; Pechtel et al., 2014; Schalinski et al., 2017; Schalinski et al., 2018; Schalinski et al., 2016; Teicher et al., 2018; Tomoda et al., 2009a; Tomoda et al., 2012). Random forest regression predicts outcome by creating a forest of different decision trees with each tree generated from a different subset of the data and constrained in the number of predictors that can be considered at each branch point (Breiman, 2001). This “wisdom of the crowd” strategy provides superior predictions and is highly resistant to collinearity (Liaw and Wiener, 2002). The tree structure can also model interactions and does not assume a linear relationship between exposure and response. We use a variant of Brieman’s approach with conditional inference trees (Strobl et al., 2007) that rectifies a problem in the estimation of importance of predictors with many versus few levels or categories (Strobl et al., 2007).

We selected this strategy following simulation studies using actual exposure data with artificial outcomes specified by type and timing of MAL and diluted with random noise. Data were analyzed using several different AI algorithms. RFR-CIT performed with remarkable accuracy in hundreds of simulations of varying complexity with the embedded signal accounting for 5–10% of variance. Other algorithms were moderately to severely influenced by collinearity and often misidentified type and timing of MAL that were correlated with the actual predictors.

For these analyses the random forest was trained using data from 63.3% of the subjects and evaluated on the ‘out of bag’ test set (36.7%). The variable importance (VI) of each predictor in the model was assessed by permuting the variable, refitting the random forest, and calculating how much permutation of that variable increased the mean square error of the fit. Permuting important regressors produces a large increase in mean square error while permuting unimportant regressors has negligible impact. This process was repeated 50 time to calculate average measures of VI for each variable. To estimate significance the overall process was then repeated 1000 times using reshuffled network values to calculate chance mean and SD importance levels for each variable. The significance of Z-test difference between observed and chance VI measures were calculated for each variable and adjusted using Bonferroni correction to control for multiple comparisons.

Results

The sample consisted of N=342 participants (132M/210F) 21.7 ± 2.5 years of age. Fifty-six percent of the participants were at-risk with moderate-to-high-exposure to maltreatment (mean 4.1 ± 2.1 types) and the remainder were not at-risk with no-to-low exposure (0.4 ± 0.5 types of maltreatment).

Changes in Network Vulnerability Across Age in At-Risk Participants.

Overall, there was a significant linear effect of age on network vulnerability (F_1,188 = 10.12, p < 0.002) in at-risk individuals after covarying for sex and parental education which were also significant predictors. There was no significant age x sex interaction (F_1,187 = 0.21, p > 0.64). As seen in Figure 1, network vulnerability increased from age 18 (earliest age studied) to plateau at about age 21. Small worldness had a very similar linear relationship with age (F_1,188 = 9.88, p < 0.002) and a parallel time course reaching plateau value at about 22 years of age.

Fig 1.

Change in network vulnerability across age in maltreated participants with confidence limits based on quadratic polynomial fit

Sensitive Exposure Periods for Network Vulnerability and Small Worldness.

As seen in Figure 2A significant ‘predictors’ of network vulnerability were parental verbal abuse at ages 16 (variable importance (VI) = 0.66, p < .002), 17 (VI = 0.56, p < .01) and 18 years (VI = 0.88, p < 10⁻⁵), number of different types of maltreatment during childhood (VI = 0.91, p < 10⁻¹⁸), current age (VI = 1.47, p < 10⁻⁵) and parental education (VI = 0.66, p < .03). Results for small worldness (Fig 2b) were similar with the most important ‘predictors’ being parental verbal abuse at age 18 (VI =0.58, p < .04), number of types of maltreatment (VI = 0.93, p < 10⁻²³), severity of maltreatment across childhood (VI = 0.53, p < 10⁻⁵) as well as age (VI = 2.28, p < 10⁻²³), sex (VI = 0.84, p < .03) and parental education (VI = 0.89, p < .0001).

Fig 2.

Sensitive period analysis using random forest regression with conditional inference trees indicating importance of type and timing of maltreatment on measures of structural brain network architecture. A Network vulnerability to disruption. B Small world organization. Importance is defined as the mean increase in mean square error of the overall fit following permutation of each independent variable. Solid horizontal line indicates importance of duration of maltreatment, dotted horizontal line indicates importance of overall severity of maltreatment and dashed horizontal line indicates importance of number of different types of maltreatment across childhood. Abbreviations: EN - emotional neglect; NVEA - nonverbal emotional abuse; Peer_E - peer emotional bullying; Peer_P - peer physical bullying; Phys - parental physical abuse; PN - physical neglect; PVA - parental verbal abuse; SexA - sexual abuse; WIPV - witnessing interparental violence; and Wsib - witnessing violence toward siblings.

Nodal Network Architecture and Time Course of Resilience.

Demographic measures for the four groups are included in Table 1. Overall, there were no differences between the groups in gender distribution, race or ethnicity. Unexposed controls were, on average, about 1.1 years older than symptomatic maltreated participants. Symptomatic maltreated individuals also reported lower levels of perceived financial sufficiency during childhood than controls or the recovered groups. All at risk groups reported lower levels of parental education than controls. These variables were included as covariates. To minimize multiple comparisons, groups were initially compared using MANCOVA with separate multidimensional analyses from measures of global network architecture (i.e., vulnerability, small worldness, degree and efficiency) and measures of N_eff in 9 specific ‘resilience’ nodes (K. Ohashi et al., 2019). We previously reported that measures of global network architecture varied between maltreated and non-maltreated groups but not between symptomatic and asymptomatic maltreated participants (K. Ohashi et al., 2019). Conversely, we reported that measures of N_eff in these 9 ‘resilience’ nodes were reduced in asymptomatic maltreated individuals but did not differ between symptomatic maltreated individuals and non-maltreated controls (K. Ohashi et al., 2019). The key question is whether these parameters differed between asymptomatic maltreated individuals with a minimal impact versus emergent/recovery patterns.

Table 1.

Demographic measures by group.

Measures	Unexposed	Symptomatic	Minimal Impact	Recovered	F- or ChiSq	p values
Gender	52M/66F	34M/72F	14M/17F	22M/33F	χ2=3.9	0.2
Age	22.0±2.4	20.9±2.4§	22.1±2.6	21.5±2.6	4.2	0.007
Parental education	16.6±2.9	14.8±3.1¥	14.7±2.8*	14.8±3.7†	8.1	0.00004
Financial sufficiency	3.4±0.7	2.6±1.0¥	3.1±0.8	3.0±0.8	14.5	10−8
Race					χ2=5.8	0.7
White	70%	66%	62%	63%
Black	9%	12%	7%	12%
Asian	16%	13%	14%	15%
Other	5%	9%	17%	10%
Ethnicity (Hispanic)	15%	20%	21%	13%	χ2=1.73	0.6

^*p<.05

^§p<.01

^†p<.005

^¥p<.001 versus not at-risk participants

Overall, there were significant multidimensional differences between the four groups on measures of global network architecture (F_{4, 303} = 3.9, p < 0.004) and nodal efficiency (F_27,888 = 1.49, p = 0.05) (Table 1). In terms of global network measures, all of the at-risk groups: symptomatic (F_4,218 = 3.8, p <0.006); minimal impact (F_4,143 = 3.9, p < 0.005) and recovered (F_4,167 = 3.0, p < 0.02), differed from the not at-risk group, but did not significantly differ between each other.

In terms of nodal efficiency, the asymptomatic minimal impact group differed significantly from the not at-risk (F_9,138 = 2.071, p < .04) and symptomatic groups (F_9,126 = 2.12, p < .04). As indicated in Table 1 all of the nodes, except the right amygdala in the asymptomatic minimal impact group differed significantly from the not at-risk group in N_eff. In contrast, there were no significant differences between the at-risk symptomatic and not at-risk groups and only N_eff measures in the right amygdala and right subcallosal gyrus differed between the asymptomatic recovered group and the not at-risk group.

Socioeconomic Factors Associated with Resilience.

At risk participants reporting that they grew up in families with the lowest levels of perceived financial sufficiency (i.e., much less than enough money to meet their needs) had a 5.30-fold (95% CI = 1.36–30.44, p < 0.008) greater unadjusted odds of being symptomatic versus asymptomatic on assessment than participants growing up in families with enough, more than enough or much more than enough money to meet their needs. Likewise, participants reporting growing up in families with the second lowest level (i.e., less than enough money to meet their needs) had 3.11-fold (95% CI 1.49–6.76, p = .001) greater unadjusted odds of being symptomatic than participants growing up in families with greater degrees of financial sufficiency. These findings were not simply a consequence of higher exposure to maltreatment in the low financial sufficiency groups, as participants with the two lowest levels of financial sufficiency (less or much less money than necessary) had a 2.98-fold (95% CI 1.49–5.97, p = 0.002) increase in odds of being symptomatic compared to participants from families with greater financial sufficiency after adjusting for number of types of maltreatment, age and sex. Put another way only 52% of participants in the symptomatic group reported being raised in families with at least enough money versus 91% in the controls, 81% in the minimal impact group and 78% in the recovered group. Interestingly, having enough, more than enough, or much more than enough money during childhood was associated with reduced N_eff in two of the ‘resilience’ nodes – the right postcental gyrus (F_1,189 = 8.4, p = 0.004) and the right subcallosal gyrus (F_1,189 = 4.59, p < 0.04) after controlling for multiplicity of maltreatment in at risk participants. While financial sufficiency was associated with reduced nodal efficiency in these regions in maltreated individuals, financial insufficiency was not associated with any differences in nodal sufficiency. This suggests that higher levels of financial sufficiency during childhood may have acted as a protective factor in at-risk participants.

Parental education, on the other hand, did not have a significant association with symptomatic versus asymptomatic status in at-risk participants. Having parents with less that 12 years of education on average, increased adjusted odds of being symptomatic by 1.68-fold (95% CI 0.69 – 3.99), which fell short of significance (p > 0.25), as did having parents with less than 16 years of education (odds ratio 0.90, 95%CI 0.50 – 1.62, p > 0.71). Likewise, gender did not have a significant influence on susceptibility / resilience after controlling for degree of exposure to maltreatment. At risk males were 1.53-fold (95%CI 0.83 – 2.81) more likely to be symptomatic rather than asymptomatic but that difference was not statistically significant (p > .17).

Discussion

These findings expand on our recently proposed network model for susceptibility / resilience to the adverse psychiatric of childhood maltreatment. The first new finding shows that network vulnerability progressively increases in maltreated individuals to reach a plateau at about 21 years of age. Unfortunately, we do not have comparable network data on participants younger than age 18 but suspect that network vulnerability increases from at least age 15 as the frontal hubs that appear to be missing in the maltreated network (Teicher et al., 2014a) typically develop between 15–19 years of age (S. T. Baker et al., 2015). Increasing network vulnerability in maltreated individuals during adolescence and early adulthood, then likely results from the combined consequences of pruning of general fiber pathways (Cressman et al., 2010) and failure to adequately develop frontal association pathways underlying these frontal network hubs (S. T. Baker et al., 2015).

This leads us to propose that increasing network vulnerability during adolescence and early adulthood in maltreated individuals leads to emergence of psychiatric symptoms and problems with substance use that manifests when network vulnerability reaches a stage when the network can no longer effectively compensate for abnormalities in one of more stress susceptible brain regions. This hypothesis fits with evidence that prevalence of depression and PTSD markedly increase in maltreated individuals between 15–19 years of age (Teicher et al., 2009) and that degree of substance use and binge drinking peak in the early adulthood (S. H. Shin et al., 2013).

The next finding suggests that two experiential factors most directly impact network vulnerability. The first is number of different types of maltreatment experienced during childhood (multiplicity of exposure) and the second is degree of parental verbal abuse between ages 16–18. We suspect that multiplicity of exposure, as a form of cumulative burden, may be acting to either curtail the initial overproduction of neuronal processes or enhance their subsequent pruning. We further suggest that parental verbal abuse between ages 16–18 may more directly target development of frontal hubs that largely occurs between these ages (S. T. Baker et al., 2015).

It is unclear whether reduced levels of N_eff in these ‘resilience nodes’ in asymptomatic maltreated individuals are preexisting protective factors or if this reduction in N_eff occurs gradually as part of a compensatory process. We expect that both are true and that preexisting alterations in certain ‘resilience’ nodes are present in some individuals and that other individuals show compensatory changes in N_eff over time. Further, we propose that individuals with these preexisting protective alterations in N_eff would show a ‘minimal impact’ resilience pattern (Bonanno and Diminich, 2013; Bonanno et al., 2012a; Bonanno et al., 2004) when confronted with childhood abuse, characterized by a modest increase in symptoms during exposure but with rapid return to preexposure baseline. In contrast, individuals who develop compensatory changes in N_eff after exposure would be more likely to show an emergent resilient or recovery profile, characterized by development of substantial symptoms during the period of exposure with gradual return to more baseline levels (Bonanno and Diminich, 2013; Bonanno et al., 2012a; Bonanno et al., 2004).

Interestingly, Bonanno et al (Bonanno et al., 2015; Bonanno, 2012; Bonanno and Diminich, 2013; Bonanno et al., 2012b) classify aversive circumstances as acute (isolated events demanding or resulting in loss of resources for generally less than 1 mo – such as natural disasters, transportation accidents) or chronic (event or series of events that exert a recurrent and cumulative impact over many months or years such as abuse or neglect) and propose that this is the key determinant of the likely trajectory of resilience, with minimal impact following acute events and emergent/recovery following chronic exposure. Our neurobiological model expands upon this frame so that the duration of the aversive event serves as one element determining trajectory but postulates that the primary determinant is the baseline status and rate of change of nodal efficiency in relevant ‘resilience nodes’. This modification fits with our network discoveries and we believe that it improves the model as the idea that duration of exposure determines resilience trajectory strains credulity in situations where adults typically show minimal impact resilience patterns even when confronted by chronic traumatic events (e.g., military deployment (S. B. Andersen et al., 2014; Berntsen et al., 2012; Bonanno et al., 2012b; Porter et al., 2017), cancer (Burton et al., 2015; Deshields et al., 2006; Helgeson et al., 2004; Lam et al., 2012)). We suspect that developmental changes in nodal efficiency of certain resilience nodes occurs in individuals who have successfully dealt with challenges throughout life and that this provides a better understanding why adults are more likely to show a minimal impact resilience pattern than children even when confronted with chronic adversity.

In the present study we found that maltreated individuals who have had a resilience pattern more consistent with minimal impact (currently asymptomatic with no lifetime psychiatric history) had low levels of nodal efficiency in all of the 9 ‘resilience’ nodes except the right amygdala. In contrast, maltreated individuals whose resilience pattern was more consistent with a recovery pattern (currently asymptomatic but with lifetime history of Axis I or II disorders) had reduced nodal efficiency in only two nodes. Interestingly, these findings raise the possibility that targeting these two specific nodes – the right amygdala and right subgenual cingulate – may help to facilitate recovery in symptomatic maltreated individuals.

Finally, we found that perceived financial insufficiency during childhood was associated with a nearly 3-fold increase in risk in maltreated individual of having clinically significant psychiatric symptoms after controlling for differences in degree of exposure to maltreatment. We have used this simple rating of perceived financial sufficiency during childhood as a measure of sociodemographic risk, rather than family income, as this is something that 18–25-year-old participants are more confident of than childhood family income, and that measures of income alone, even after adjusting for differences in family size and geographic area can mean very different things depending on family spending patterns.

Although we typically think of financial insufficiency or poverty as important risk factors (Palacios-Barrios and Hanson, 2019) our results actually suggest that having enough, more than enough, or much more than enough money to meet the family’s needs was a protective factor associated with reduced nodal efficiency in right post central gyrus and right subgenual cingulate nodes of maltreated individuals. Understanding how early experience may alter measures of nodal efficiency in these specific resilience nodes may provide new insight into the nature of potential protective factors.

Key strength of the study include a relatively large single site neuroimaging sample, graph theory analyses and use of novel AI means for assessing sensitive exposure periods. The key limitation is the use of retrospective reports of maltreatment which theoretically could be affected by psychopathology-associated memory impairments and potential memory biases. There is clearly a pressing need to verify if these network and nodal changes are predictive of psychiatric status in a prospective longitudinal study.

Another limitation is the use of a cross-sectional design to infer developmental differences in network architecture. The critical problem with cross-sectional designs is that they can be affected by cohort effects unrelated to the participant’s age such as age group differences in diet, parental rearing, educational or medical practices, exposure to second hand smoke, etc. This is of particular concern in assessing individuals at the same time across different generations, However, we were only assessing a seven year difference and the data were not collected at one time but were collected over about seven years so that subjects studied at different ages may actually be member of the same age cohort and each age group consisted of member from multiple different age cohorts, so that cohort related effects should be strongly attenuated.

A further limitation is the use of a categorical approach. The main problem is that categorical data types contain substantially less information than continuous data types, this leads to reduced signal-to-noise ratio and increases the risk of false negative findings. Hence, we have also shown that the nodal efficiency findings are associated with strong dimensional effects as well (K. Ohashi et al., 2019) which we discussed in regard to the left inferior frontal pars triangularis findings. Failure to detect a signal does not appear to be an issue in these analyses. An advantage of the categorical approach is that the results may be much easier to understand and may be more compelling, which is why we emphasized the categorical findings here.

Overall, these findings provide additional support for a structural brain network model of domain-specific susceptibility / resilience to the psychiatric consequences of childhood maltreatment. We need to ascertain whether functional measures of brain network architecture provide a similar perspective, as functional connectivity may be more malleable than structural connectivity. Future studies will also need to assess the degree to which effective treatments either reverse the neurobiological effects of maltreatment or serve to move the nodal efficiency measures of susceptible maltreated individuals more into line with the nodal efficiency of more resilient individuals.

Table 2.

Measures of global network architecture and nodal efficiency by group.

Measures	Not At-Risk	Symptomatic	Asymptomatic
			Minimal Impact	Recovered
Small Worldness	2.79±0.145	2.850±0.145†	2.900±0.138†	2.850±0.138†
Vulnerability (x 1000)	0.856±0.116	0.902±0.116†	0.927±0.111§	0.888±0.110*
Global Efficiency	0.483±0.015	0.479±0.015*	0.474±0.0143§	0.478±0.0143*
Degree	12.60±0.93	12.20±0.93§	12.00±0.89†	12.30±0.88§
Neff Amygdala (R)	0.434±0.026	0.434±0.026	0.429±0.024	0.425±0.024*
Neff Frontal Inferior pars triangularis (L)	0.443±0.032	0.444±0.032	0.429±0.031†	0.437±0.031
Neff Supplementary Motor Area (L)	0.438±0.030	0.435±0.030	0.422±0.028†	0.438±0.028
Neff Subcallosal Gyrus (L)	0.485±0.048	0.476±0.048	0.463±0.046*	0.470±0.046
Neff Subcallosal Gyrus (R)	0.494±0.046	0.479±0.046	0.470±0.044*	0.475±0.044*
Neff Middle Cingulum (R)	0.514±0.033	0.512±0.033	0.497±0.032†	0.509±0.032
Neff Paracentral Lobule (R)	0.425±0.018	0.425±0.018	0.417±0.017*	0.423±0.017
Neff Post Central Gyrus (L)	0.499±0.026	0.501±0.026	0.490±0.025*	0.499±0.025
Neff Post Central Gyrus (R)	0.507±0.029	0.499±0.029	0.487±0.028¥	0.500±0.028

^*p<.05

^§p<.01

^†p<.005

^¥p<.001 versus not at-risk participants