Background

Posttraumatic stress disorder (PTSD) is a delayed and lasting dysfunctional response to psychological stress. Patients usually have a long illness with recurring symptoms, often complicated by comorbidities such as substance abuse, depression, anxiety disorder, aggressive behavior, self-injury and suicide, as well as medical complications such as chronic pain and infection, cardiovascular disease and increased risk of dementia [1, 2]. The overall burden of disability and premature death caused by PTSD is therefore high [3]. Children are more vulnerable to PTSD than adults, being 12–25% more likely to develop depression, suicidal behavior and cognitive impairment [4]. The etiology and neuropathology of this complex disease are still not clear [5], and accurate prognosis suffers from the lack of reliable biomarkers.

Two decades ago it was hoped that neuroimaging-based biomarkers would prove diagnostically and prognostically effective in a number of neuropsychiatric diseases. This hope has not yet been realized, as research has revealed an increasingly complex picture of subtle, distributed brain changes varying with individual clinical characteristics. Neuroimaging biomarkers capable of distinguishing PTSD from non-PTSD subjects have received attention [6,7,8,9], and two recent studies obtained good results using resting-state fMRI [7, 9]. The brain is a highly interconnected network, and the development of psychiatric illness appears increasingly linked to dysfunctional integration of networks between the cortex and subcortical regions. Many studies have therefore taken whole-brain network metrics and used them as input to a single subject classification [10,11,12]. Recent advances in psychoradiology, allow the direct noninvasive characterization of brain network topology in neuropsychiatric patients [13,https://www.activestate.com/products/python/), where the neural network was implemented in the Pytorch library [1 shows an overview of the classification approach showing the main steps in the pipeline. Figure 2 shows the deep network training model. The Supplementary Figure illustrates this process, using a simple example of a 3-layer network; see [34] for details.

Fig. 1
figure 1

Overview of the employed classification approach showing the main steps of the pipeline. The raw images were preprocessed, and then the whole brain functional connection matrix was calculated to obtain the graphic topological attributes. Finally, the deep learning model was used to classify the groups

Full size image
Fig. 2
figure 2

Deep network training. (a) An unsupervised step is first performed that sequentially trains individual autoencoders (AE). (b) The supervised step stacks the initialized AEs (thus creating the deep network) and then adds one additional layer for the supervised training only (the training label layer) which contains the binary diagnosis label for each binary high-dimension feature vector in the training population [38]

Full size image

Other statistical methods

The statistical significance of between-group differences in demographic and clinical characteristics was tested by the two-tailed two-sample t test (continuous variables) or the two-tailed Pearson Chi-square test (categorical variables).

Results

Demographic and clinical characteristics

There were no significant differences in age, gender, education between pediatric PTSD and HC (p > 0.05; Table 1).

Table 1 Demographic and clinical characteristics of participants a
Full size table

Classification performance

The single-subject classification of pediatric PTSD and HC using graph-based topological metrics was assessed for accuracy, sensitivity and specificity at 10-fold cross-validation. The average accuracy of classification was 71.2 ± 12.9%, the average sensitivity was 59.7 ± 21.9% and the average specificity was 82.7 ± 13.9% in the DL model (p < 0.001).

Regions with greatest contribution to single subject classification

To identify the classification pattern in patient and HC group, we investigated feature contributions to the non-linear dimensionality reduction in patient group in the DL model. The 10 features with the highest contribution values across the DL models are reported in Table 2 and represented graphically in Fig. 3. These regions were mainly located in frontoparietal areas, with some spread to subcortical regions such as the anterior cingulate cortex, median cingulate cortex, and amygdala.

Table 2 Top 10 most relevant topological properties of brain regions for Deep Learning classification analysis a
Full size table
Fig. 3
figure 3

Regions providing the greatest contribution to single subject classification of patients and controls. The nodes (brain regions) were mapped onto the cortical surfaces using the BrainNet Viewer package (http://www.nitrc.org/projects/bnv). For brain regions, red represents the nodal betweenness, blue represents the nodal efficiency, and yellow represents the nodal degree. Abbreviation: DCG, median cingulate and paracingulate gyri; ROL, Rolandic operculum; AMYG, amygdala; ACG, anterior cingulate and paracingulate gyri; MFG, middle frontal gyrus; SMA, Supplementary motor area; SPG, superior parietal gyrus; R, right hemisphere; L, left hemisphere

Full size image

Discussion

We set out to classify between pediatric PTSD and HC using the DL model applied to graphic topological measures, and then explored the regions making the greatest contribution to classification performance. Consistent with our first hypothesis, we found that using topological properties in DL we could distinguish PTSD from HC at the individual level with significant accuracy. This supports the emerging notion that graphic topological properties based on resting-state functional neuroimaging data can be a powerful tool for characterizing brain disorders at the level of the individual [21]. Such methods have achieved 86% accuracy in distinguishing patients with amnestic mild cognitive impairment from healthy controls [9], and 70–80% accuracy in distinguishing schizophrenic patients from non-psychiatric controls [10]. In PTSD, Zilcha-Mano et al. [9] were able to discriminate between 51 PTSD individuals and 76 trauma-exposed healthy control subjects with an accuracy of up to 70.6% by using a whole-brain data-driven definition of functional connectivity biomarkers and regularized partial correlations which revealed differences in functional connectivity within executive control network and salience network between the two groups.

DL methods can automatically identify the optimal representation from the raw data without the need for specialized feature engineering. This is achieved by using a hierarchical structure with varying levels of complexity, including the application of consecutive nonlinear transformations to the raw data. An essential aspect of DL that differentiates it from other machine learning methods is that the features are not manually engineered; instead, they are learned from the data, resulting in a more objective and less bias-prone process. Compared with other machine learning methods such as SVM, DL can achieve higher orders of abstraction, complexity and higher classifier accuracy [29, 30], which makes DL more suitable for detecting complex, scattered and subtle patterns in the data [59].

Consistent with our second hypothesis, a few regions make the largest contribution to classification performance: frontoparietal regions (central executive, CEN) and subcortical areas like (median and anterior) cingulate cortex and amygdala. The CEN is associated with the progress of goal-directed behaviors, such as working memory and attention control [60], and it has also been reported as impaired in PTSD [21, 38]. Specifically, CEN functional disruptions are associated with PTSD symptoms of decreased cognitive functioning across multiple domains, as well as emotion under-modulation associated with impaired regulation of limbic structures [61,62,63]. For instance, a recent study found that in the resting-state, subtype non-differentiated PTSD patients demonstrate reduced CEN convergence, which was associated with decrease orbitofrontal-amygdala connectivity in PTSD, an indicator of reduced prefrontal regulation acting on the resting limbic system [63]. The amygdala is a core area in current neurocircuit models of stress and PTSD [64, 65]. Among its multiple functions, the best known is to encode and extinguish the memory of fearful stimuli [65, 66] so as to direct physiological and behavioral responses to such stimuli. In addition, the amygdala plays an essential role in fear generalization [67], arousal [68] and processing of rewards [69], all of which may be disrupted in PTSD. Exaggerated amygdala activity in response to trauma-related and more generic stimuli is a frequent finding in fMRI studies of PTSD [70, 71]. Recent research has enlarged the functions traditionally ascribed to the cingulate to include emotion [72], pain management [73] and cognitive control [74, 75]. A recent meta-analysis concludes that cingulate plays an important role in emotion and cognitive processing in patients with PTSD [76].

Neuroimaging is still far from becoming a routine tool in clinical psychiatry, mainly because there is still insufficient evidence of diagnostic and prognostic effectiveness. We followed recent recommendations on avoiding methodological issues that may in the past led to overoptimistic results [77,78,79]. A major challenge in applying machine learning to high-dimensional neuroimaging data is the risk of overfitting, i.e., the learning of irrelevant fluctuations within a dataset that limits generalizability to other datasets. To avoid that, we applied DL technology to conduct a dimension reduction and mitigate the effect of spurious signals. We also tried to minimize the risk of overfitting through the use of region-level features rather than voxel-level data (which are associated with more noise and a higher risk of overfitting) [30]. One limitation is that we only explored topological properties based on the AAL brain atlas. Although AAL is widely accepted in neuroimaging studies, it has drawbacks. Future studies should verify our results using the new brain atlases that are now being used in neuroimaging and machine learning studies, such as the Power 264-region atlas [80] and the Dosenbach’s 160 functional atlas [81]. Another limitation is that our PTSD participants were exposed to a specific traumatic event (an earthquake), which might limit the generalizability of our results. This can be tested by replication using subjects exposed to other traumatic events.

Conclusion

Despite these limitations, the present study demonstrates DL as an objective and useful classifier which could differentiate pediatric PTSD and HC based on graphic topological measures using resting-state MRI data with promising accuracy. Further, the CEN, parietal gyrus, cingulate cortex, and amygdala provide the greatest contribution to classification performance in DL model, suggesting that investigating these core nodes may give insight into the heterogeneous clinical profiles of individuals with PTSD. Further studies will be needed to assess the clinical applicability of our method.