We recruited 215 healthy adult participants from Guangzhou Medical University and the surrounding community between July, 2013 and August, 2019. To ensure eligibility, we administered the Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders–IV Edition (DSM- IV) Non-Patient Edition to all participants, confirming that they had no history of Axis I disorders. Additionally, we excluded anyone with a family history of psychiatric illness among first- to third-degree biological relatives, as well as individuals with a history of seizures, head trauma, significant surgeries or medical conditions, substance abuse or dependence, or contraindications for magnetic resonance imaging (MRI). We excluded nine subjects due to quality control issues of head motion and spatial normalization, leaving 206 subjects eligible for subsequent analysis (Table 1).

All participants were right-handed and provided written informed consent before their involvement in the study. The study protocol received ethical approval from the Institute Research Board of the Affiliated Brain Hospital, Guangzhou Medical University.

Before MRI scanning sessions, we conducted a thorough assessment of CM severity for all participants using the Chinese version of the CTQ Short Form [17, 18]. This self-report scale comprises 28 items, encompassing five distinct dimensions: emotional abuse, physical abuse, sexual abuse, emotional neglect, and physical neglect, and each item is scored from 1 (“never”) to 5 (“very often”). This scale has been widely applied in clinical and non-clinical populations [18], and its Chinese version has shown strong reliability and validity [17]. A detailed assessment of the CTQ scale and subscales for all qualified subjects (N = 206) can be seen in Table 1.

All MRI datasets for all participants were acquired utilizing a 3T Philips Achieva X-series MRI scanner, equipped with an eight-channel phased-array head coil in the Affiliated Brain Hospital, Guangzhou Medical University.

The resting-state fMRI datasets were collected using a gradient-echo echo-planar imaging sequence, with parameters as follows: repetition time (TR) = 2000 ms, echo time (TE) = 30 ms, flip angle = 90°, field of view (FOV) = 220 mm $\times$ 220 mm, and acquisition matrix = 64 $\times$ 64. The scan consisted of 33 transverse interleaved slices, each with a thickness of 4 mm and a gap of 0.6 mm. Throughout the scanning process, participants were instructed to remain at rest with their eyes closed, and none of them reported falling asleep upon being queried immediately after the scan.

A T1-weighted 3D turbo field-echo sequence was employed to acquire the structural datasets, with parameters as follows: TR = 8.2 ms, TE = 3.7 ms, inversion time = 1100 ms, flip angle = 7°, FOV = 256 mm $\times$ 256 mm, acquisition matrix = 256 $\times$ 256, and voxel size = 1 $\times$ 1 $\times$ 1 mm³. The scan encompassed continuous 188 sagittal slices covering the brain.

The resting-state fMRI images were preprocessed by the Data Processing and Analysis for Brain Imaging (DPABI) pipeline (http://rfmri.org/dpabi), as implemented in the Matlab (v 2016a, Mathworks, Natick, MA, USA) platform.

The preprocessing pipeline included the following steps: (1) removal of the first 10 volumes; (2) slice-timing and head-motion correction; (3) co-registration of functional images to structural space; (4) regression of whole-brain, cerebrospinal fluid, and white matter signals, as well as linear trends and Friston-24 head-motion parameters; (5) non-linear normalization to Montreal Neurological Institute (MNI) space; (6) smoothing with a 5-mm full-width-at-half-maximum; and (7) temporal filtering within a frequency range of 0.01 to 0.1 Hz.

After preprocessing, nine participants were excluded due to quality control issues of the head motion and spatial normalization, and 206 participants were kept for further analyses.

For the estimation of individual EC, we applied the Schaefer atlas [19] (100 parcels, Fig. 1a) to cover the cortex and extracted the first eigenvariate for each region from the preprocessed functional images using fslmeants/FSL. This eigenvariate served as the time series data, which was then utilized in the calculation of EC.

Fig. 1.

The Schaefer atlas and effective connectivity (EC) matrix averaged across all participants. (a) shows the utilized Schaefer atlas (100 parcels) used for extracting individual timeseries and the color indicates the different parcellations. (b) illustrates averaged EC across subjects and the color bar indicates the EC magnitude.

As a recently developed variant of DCM, rDCM was utilized to estimate individual whole-brain EC efficiently [14]. This method transforms conventional linear DCM into the frequency domain, making it a special case of the Bayesian linear regression model [14]. Using this approach, we specified a full-connection model to estimate EC, where a 100-by-100 matrix was built for each subject (Fig. 1b). Additionally, to examine the prediction performance from separate networks, we down-resampled the whole-brain EC matrix into seven networks including the visual network (Vis), somatosensory network (SMN), dorsal attentional network (DAN), ventral attentional network (VAN), limbic network (Lim), fronto-parietal network (FPN), and DMN. This down-sampling grouped parcels based on the Schaefer atlas [19, 20], aligning each of the 100 parcels with one of seven networks. We extracted within-network connectivity for each network by selecting matrix entries where nodes shared the same network label. This approach enabled a targeted comparison of whole-brain and network-specific EC predictions.

To explore the relationship between whole-brain and network-specific EC and CM severity, we employed a ridge-regularized linear regression model within a 10-fold cross-validation scheme, where EC was specified as features and the total scores summing from all CTQ subscales were used for targets. The cross-validation process involved splitting samples into training and testing sets. In the training set, EC features were correlated with CTQ total scores, and the top 5% of features were extracted based on the highest positive and negative correlations, respectively. A ridge-regularized linear regression model was then trained on the training set, which was subsequently used to predict scores in the testing set, calculating the correlations between predicted and actual scores. An average prediction correlation was calculated across all cross-validation folds.

To improve the stability of prediction performance, we repeated this prediction process 100 times and obtained an average prediction correlation across all iterations for each prediction case. To assess the statistical performance of the prediction model with the feature selection threshold of 5%, we applied a label-shuffled permutation (n = 500 times) test, where CTQ total scores were randomly shuffled before each cross-validation loop. The permutation process serves to evaluate the robustness of our model by generating a null distribution of prediction correlations. By comparing the empirical prediction performance against this null distribution, we can determine the statistical significance of our observed results with a threshold of p = 0.05. Given that predictions were conducted for the whole brain and each of the seven networks, we additionally applied a Bonferroni-corrected threshold of p = 0.0063 (0.05/8) for multiple comparisons to the uncorrected p-values obtained from each permutation test. This provides both uncorrected and corrected interpretations, offering a more comprehensive view of our findings.

Considering that feature selection thresholds may impact prediction performance, we repeated our prediction processes with the feature selection conditions of 10% and 20%, which enabled us to confirm the consistency of prediction performance patterns among whole-brain and network-specific predictions. Additionally, to verify our findings, we employed a least absolute shrinkage and selection operator (LASSO)-regularized linear regression model and compared its results with those of the ridge-regularized model at a 5% feature selection threshold.

Table 1 summarizes the demographic characteristics and CTQ scores for the 206 participants included in the study after imaging preprocessing. The sample comprised 91 males and 115 females, with an average age of 25.3 years. Participants had an average of 14 years of education. The mean CTQ total score was 34.7, with subscale averages as follows: emotional abuse, 6.6; physical abuse, 5.8; sexual abuse, 5.3; emotional neglect, 9.6; and physical neglect, 7.3.

In this study, we examined the prediction correlation between CTQ total scores and EC features selected from the whole brain and separated networks using a ridge-regularized linear regression model, where a feature selection threshold of 5% was applied. Then, to improve the robustness and consistency of the above predictions, we also employed 10% and 20% feature selection thresholds to examine whole-brain and network-specific EC predictions. Additionally, a LASSO-regularized model was employed to validate the prediction results.

When considering a feature selection threshold of 5% (Fig. 2 and Table 2), we found that whole-brain EC features marginally predicted CTQ total scores, with a correlation coefficient of r = 0.16. However, when predictions were examined at the network-specific level, only EC features from the DMN successfully predicted CTQ total scores, yielding a stronger correlation of r = 0.25 (p 0.05). In contrast, EC features from other networks, including the Vis, SMN, DAN, VAN, Lim, and FPN, failed to be statistically significant in prediction (p $>$ 0.05). Their prediction correlations ranged from 0.1 to 0.15. Notably, EC features from the DMN outperformed those from all other cases in predicting CTQ total scores, exhibiting the highest prediction correlation compared with both the whole-brain and other network-specific predictions. However, when including a multiple-comparison corrected threshold of p = 0.0063 (Bonferroni corrected), none of them were statistically significant.

Fig. 2.

Violin plot showing the distribution of prediction correlations for whole-brain and network-specific EC features derived from 100 repetitions with a 5% feature selection threshold. The x-axis represents the EC feature sources used to predict the CTQ total scores, while the y-axis indicates the correlation between the empirical and predicted scores. Each dot represents an individual prediction correlation. Abbreviations: WB, whole brain; Vis, visual network; SMN, somatosensory network; DAN, dorsal attentional network; VAN, ventral attentional network; Lim, limbic network; FPN, fronto-parietal network; DMN, default mode network.

To examine the impact of feature selection thresholds on prediction correlations for CTQ total scores, we varied the feature selection thresholds from 5% to 10% and 20% (Table 3). We observed that the significance level of whole-brain EC prediction did not vary across different thresholds, with no significant predictions for CTQ total scores at the 10% or 20% thresholds (p $>$ 0.05). In contrast, EC features from the DMN remained statistically significant (p 0.05) even after applying a multiple-comparison correction (p 0.0063) with these thresholds. Importantly, varying the feature selection thresholds did not affect the statistical significance (p $>$ 0.05) of EC features in the other networks.

To validate our prediction results obtained from the ridge-regularized linear regression model, we then examined the prediction correlations using a LASSO-regularized model (Fig. 3). Our findings were consistent with those of the ridge-regularized model, and EC features within the DMN demonstrated the highest prediction correlation compared with other networks.

Fig. 3.

Violin plot displaying the prediction distributions for whole-brain and network-specific EC features, derived from 100 repetitions at a 5% feature selection threshold using the LASSO-regularized linear regression model. The x-axis represents the EC feature sources used to predict CTQ total scores, while the y-axis shows the correlation between empirical and predicted CTQ total scores. Abbreviations: LASSO, least absolute shrinkage and selection operator.

The present study has shown that EC was predictive of individual differences in CM severity. This aligns with previous literature reporting that the EC feature can classify clinical patients from healthy controls well and predict treatment outcomes and behavioral domains [21, 22, 23]. This supports the utility of EC as a promising feature in understanding and predicting clinical and behavioral variations.

Our study further demonstrated that EC features within the DMN were superior to those within other networks. The DMN structure primarily includes regions of the medial prefrontal cortex, posterior cingulate cortex, precuneus, and lateral parietal cortex [24]. This network is known to be engaged in internally-focused mental processes such as self-reflection, mind wandering, daydreaming, autobiographical memory retrieval, and future planning [25, 26, 27]. Aberrant activities in the DMN may contribute to multiple dysfunctional psychological processes in self-referential thinking and working memory, and attentional impairments [28, 29]. Given that the DMN is crucial in self-referential cognitive functions, it is of great importance in the study of CM [30]. Altered connectivity of the DMN is linked to various psychopathological symptoms in transdiagnostic samples including post-trauma stress disorder (PTSD), major depressive disorder (MDD), and schizophrenia [30, 31, 32]. In schizophrenia, the experiences of CM were associated with connectivity variations in the DMN [33]. In MDD, the altered functional coupling within the DMN may contribute to developing depression in individuals with CM [34]. In PTSD, reduced FC within the DMN may lead to self-perception disturbance and link traumatic experiences to the sense of self [35]. Accumulating evidence shows that traumatic exposure adversely impacts DMN connectivity, which in turn leads to psychopathological symptoms associated with maladaptive self-referential processes [30, 36]. Additionally, previous reviews have highlighted the DMN as a network of particular interest in non-clinical individuals exposed to CM [7, 36, 37]. Numerous studies [38, 39, 40, 41, 42] have documented altered connectivity within the DMN, as well as between the DMN and other networks, in individuals with CM compared with those without. For instance, Lu et al. [38] observed abnormal FC within the DMN and in its connections with the cerebellum and insula in individuals with CM, while Zhao et al. [39] reported increased connectivity within the anterior DMN and reduced connectivity between the posterior DMN and other networks in non-clinical adults with CM. Moreover, DMN connectivity has demonstrated strong predictive contributions for CTQ scores in non-clinical individuals [43], underscoring the DMN’s role as a potential biomarker of trauma. Collectively, these findings, together with existing literature on early-life stress [36, 44, 45], suggested that alterations in DMN connectivity may serve as important neural correlates of early-life adversity. Although research regarding EC and CM is sparse, several studies utilized prior-defined regions of interest to explore EC between the amygdala and medial prefrontal cortex, reporting that traumatic exposures impact the inhibition of the amygdala by the medial prefrontal cortex [13, 46]. This may potentially reflect dysfunction in the DMN, contributing to maladaptive emotional responses impacted by traumatic exposures [46]. Consistent with these studies, our findings demonstrated that EC features within the DMN better predicted CM scores than those from other networks, suggesting that the DMN may play a crucial role in influencing CM effects on self-referential and emotional regulation.

This study further evaluated the stability of our prediction findings by varying the feature selection thresholds. Although the p-values for the statistical significance of EC prediction were impacted by different feature selection thresholds, the significance of most prediction cases was little influenced. This aligns with previous literature [47, 48] suggesting that varying feature selection thresholds (e.g., stricter p values) in selecting correlation-based features may not significantly impact prediction. While those studies primarily used p-values rather than a fixed percentage of connectivity features, our validation analysis across different thresholds demonstrated relatively consistent and robust prediction results, particularly for DMN EC features, where the predictions remained stable across different thresholds. Furthermore, the prediction results obtained from the ridge-regularized model were comparable to those from the LASSO model, showing similar stability of EC features within the DMN.