Graph Neural Networks for Analyzing Trauma-Related Brain Structure in Children and Adolescents: A Pilot Study

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Provide a justification for the chosen correlation threshold (0.5) used in constructing brain networks. Were sensitivity tests conducted with different thresholds?

Additional References: Include recent works (2019 onwards) that explore GNNs in neuroimaging to strengthen the literature review.

 

Author Response

Comments 1: Provide a justification for the chosen correlation threshold (0.5) used in constructing brain networks. Were sensitivity tests conducted with different thresholds?

Response 1: Thank you for your comment. The threshold of 0.5 was chosen considering the current limited guidelines for applying GNNs in brain structure analysis. This value allows for a focus on moderate-to-strong connections within the 0 to 1 range of Pearson correlation coefficients, ensuring biologically meaningful network construction. Sensitivity tests with different thresholds were not conducted in this study, as the focus was on exploring the feasibility of applying GNNs to trauma-related brain networks using a standard threshold. The revised manuscript now includes this explanation in Section 2.3.1, lines 105–107: "Edges were defined using a threshold of 0.5, as shown in Eq. (1). This threshold was chosen considering the current limited guidelines for applying GNNs in brain structure analysis and the need to focus on moderate-to-strong Pearson correlations."

Comments 2: Additional References: Include recent works (2019 onwards) that explore GNNs in neuroimaging to strengthen the literature review.

Response 2: Thank you for this suggestion. In response, we have incorporated additional references in the Introduction section (lines 53–62). These highlight applications such as disease classification, biomarker discovery, and missing brain graph synthesis, while also addressing their potential in advancing network neuroscience. The revised manuscript now notes: "Recent studies have reviewed GNN-based methods in network neuroscience, emphasizing their potential in advancing diagnosis of neurological disorders  and population graph integration [19, 20]."

Reviewer 2 Report

Comments and Suggestions for Authors

The article explores the use of graph neural networks to analyze brain structure changes in trauma-exposed children and adolescents, revealing differences in neural network patterns between the trauma and control groups. The topic of the article is relevant. The structure of the article follows the format accepted at MDPI for research papers (Introduction, including a review of analogs, Models and Methods, Results, Discussion, Conclusions). The level of English is acceptable. The article is easy to read. The figures in the article are of acceptable quality. The article cites 34 sources, many of which are outdated.

The following remarks and recommendations can be made regarding the article's material:

1. The study does not specify the number of participants or the demographics (e.g., age, gender, socio-economic status) in each group. A small or homogenous sample could limit the generalizability of the findings to broader populations of trauma-exposed children and adolescents.

2. While the study highlights structural differences between the trauma-exposed and control groups, the cross-sectional nature of the analysis makes it difficult to establish causal relationships between trauma exposure and the observed brain network changes. Longitudinal studies would be necessary to draw conclusions about the causal impact of trauma on brain connectivity.

3. The study presents network differences between the trauma-exposed and control groups, but the specific neurobiological mechanisms underlying these differences remain unclear. Further research is needed to explore whether the observed structural changes are directly related to trauma-related psychopathology or represent broader neurodevelopmental changes.

4. The study utilizes Graph Neural Networks (GNNs) to analyze brain networks, but the choice of this method may not fully capture all aspects of brain connectivity. GNNs are powerful tools but might overlook certain complex interactions between regions, especially if they are not directly correlated. Other network-based methods, such as functional connectivity analysis, could complement the findings.

5. The study primarily focuses on MRI-based data, which provides valuable structural information, but additional neuroimaging techniques such as fMRI or PET scans could provide more comprehensive insights into functional connectivity and metabolic changes related to trauma exposure.

6. Although the study controls for trauma exposure, other factors such as comorbid mental health conditions, substance use, or medication use in the participants could influence brain connectivity. These potential confounding variables are not addressed in the analysis, which could impact the interpretation of the results.

7. The higher reconstruction loss observed in the trauma-exposed group may reflect the complexity of their brain network, but it is not clear whether this result is due to the GNN model’s limitations or the actual neurobiological differences between groups. A more detailed analysis of the model's performance, including comparisons with other machine learning models, could provide more context.

8. The study focuses on predefined Regions of Interest (ROIs) and their connectivity. While this approach provides valuable insights into specific brain regions, it may miss broader network interactions that could be relevant for understanding trauma’s impact. More holistic, whole-brain analyses could further elucidate the complexity of brain connectivity.

9. The preprocessing steps used to prepare the MRI data for GNN analysis are not detailed in the study. Variations in data quality, such as motion artifacts or partial volume effects, could affect the accuracy of the results and should be addressed in future research.

Author Response

Comments 1: The study does not specify the number of participants or the demographics (e.g., age, gender, socio-economic status) in each group. A small or homogenous sample could limit the generalizability of the findings to broader populations of trauma-exposed children and adolescents.

Response 1: Thank you for this comment. Participant demographics have been clarified in the Data Acquisition section (lines 71–78). Specifically, we now state: "Of these, 9 were female, with a mean age of 15.13 years (SD = 0.83). The control group included 18 participants recruited from the local community, 5 of whom were female, with a mean age of 15.16 years (SD = 1.54)." While socio-economic status and other variables were not collected, we acknowledge this as a limitation, which has been noted in the Discussion section (lines 254–255). We also acknowledged the limitation of current study as having a small sample size, and added this in the discussion section.

Comments 2: While the study highlights structural differences between the trauma-exposed and control groups, the cross-sectional nature of the analysis makes it difficult to establish causal relationships between trauma exposure and the observed brain network changes. Longitudinal studies would be necessary to draw conclusions about the causal impact of trauma on brain connectivity.

Response 2: We appreciate your comment regarding the cross-sectional nature of the analysis. This limitation has been addressed in the Discussion section on line 263, where we note the importance of future longitudinal research to better understand the causal relationships between trauma exposure and brain connectivity changes.

Comments 3: The study presents network differences between the trauma-exposed and control groups, but the specific neurobiological mechanisms underlying these differences remain unclear. Further research is needed to explore whether the observed structural changes are directly related to trauma-related psychopathology or represent broader neurodevelopmental changes.

Response 3: Indeed the cross-sectional study design limited our ability to examine the neurodevelopmental component these populations. This was now added to the limitation section of the discussion as follows: : "While this study highlights structural differences in the brain, the neurodevelopmental changes of these differences could not be examined due to the cross sectional design.  Future research should explore the neuro-developmental trajectory of the observed structural changes."

Comments 4: The study utilizes Graph Neural Networks (GNNs) to analyze brain networks, but the choice of this method may not fully capture all aspects of brain connectivity. GNNs are powerful tools but might overlook certain complex interactions between regions, especially if they are not directly correlated. Other network-based methods, such as functional connectivity analysis, could complement the findings.

Response 4: We acknowledge the potential limitations of GNNs in fully capturing complex interactions between brain regions. This point has been addressed in the Discussion section (lines 267–270), where we state: "While this study used an unsupervised approach to analyze group differences, future studies could validate the effectiveness of GNNs by incorporating supervised learning with defined target variables. Expanding on these directions will further establish GNNs as a robust tool for brain imaging research."

Comments 5: The study primarily focuses on MRI-based data, which provides valuable structural information, but additional neuroimaging techniques such as fMRI or PET scans could provide more comprehensive insights into functional connectivity and metabolic changes related to trauma exposure.

Response 5: We completely agreed with this point, but those methodologies were beyond the scope of our research goal. Indeed it would be very helpful to incorporate our findings with the future fMRI and/or PET findings. 

Comments 6: Although the study controls for trauma exposure, other factors such as comorbid mental health conditions, substance use, or medication use in the participants could influence brain connectivity. These potential confounding variables are not addressed in the analysis, which could impact the interpretation of the results.

Response 6: We completely agreed with this. We confirmed that there was no previous or current psychiatric medication use in each group. It is extremely rare to use psychiatric medication ofr this population as a routine practice in the country the study was conducted. In addition, although there were comorbidities of mental health conditions in the trauma-exposed group, there was no mental health conditions in the control group. There was no substance use issue in neither group. As for the comorbid diagnoses, we believe that they are highly correlated with trauma exposure, since it is common to who disruptive mood and behavior symptoms in the children with past and current trauma exposure history, and the majority of the diagnoses were disruptive mood and behavior disorders, including Attention-Deficit/Hyperactivity Disorder (ADHD), Oppositional Defiant Disorder, Conduct Disorder, and Depressive Disorder. 

Comments 7: The higher reconstruction loss observed in the trauma-exposed group may reflect the complexity of their brain network, but it is not clear whether this result is due to the GNN model’s limitations or the actual neurobiological differences between groups. A more detailed analysis of the model's performance, including comparisons with other machine learning models, could provide more context.

Response 7: Thank you for the comment. The potential reasons behind the higher reconstruction loss in the trauma-exposed group, including both GNN model limitations and neurobiological differences, have been partially addressed in the Discussion section (lines 264–270). Specifically, we state: "Future research should explore more tailored GNN architectures and training methods that align with the unique characteristics of brain image data. Additionally, combining GNNs with complementary methods, such as functional connectivity analysis, may provide a more comprehensive understanding of brain networks. While this study used an unsupervised approach to analyze group differences, future studies could validate the effectiveness of GNNs by incorporating supervised learning with defined target variables."

Comments 8: The study focuses on predefined Regions of Interest (ROIs) and their connectivity. While this approach provides valuable insights into specific brain regions, it may miss broader network interactions that could be relevant for understanding trauma’s impact. More holistic, whole-brain analyses could further elucidate the complexity of brain connectivity.

Response 8: We acknowledge the importance of whole-brain analyses in understanding the broader network interactions relevant to trauma’s impact. The potential limitations of focusing on predefined Regions of Interest (ROIs) have been addressed in the Discussion section (lines 264–270). Specifically, we note: "Future research should explore more tailored GNN architectures and training methods that align with the unique characteristics of brain image data. Additionally, combining GNNs with complementary methods, such as functional connectivity analysis, may provide a more comprehensive understanding of brain networks." This highlights the need for more holistic approaches to further elucidate the complexity of brain connectivity.

Comments 9: The preprocessing steps used to prepare the MRI data for GNN analysis are not detailed in the study. Variations in data quality, such as motion artifacts or partial volume effects, could affect the accuracy of the results and should be addressed in future research.

Response 9: Thank you for your comment. The preprocessing steps, including skull stripping, normalization, and ROI-based signal intensity analysis, are detailed in the Data Preprocessing section (lines 87–92). Specifically, we state: "Preprocessing steps were performed using the AFNI software suite. First, skull and non-brain tissues were removed to isolate brain tissues. The images were then normalized to a standard space using the Automated Anatomical Labeling (AAL) template." As this study utilized only T1-weighted data, additional processing for motion artifacts was not applied.

Reviewer 3 Report

Comments and Suggestions for Authors

The topic of the research described in the article is consistent with the current trend in neurology concentrating on the use of modern technologies for the diagnosis and analysis of brain networks in patients exposed to traumatic events. This is a very important issue from a social and medical point of view, especially if we take into account the increasing number of traumatic events in the population of children and adolescents around the world. The authors’ contribution to science is that the technology of graph neural networks (GNNs) they created and used provides deeper insight into structural brain changes in children with trauma exposure. From a methodological point of view, the method of conducting the study and the analysis of the obtained results should be considered correct. Nevertheless, the text of the article (especially section 2.1.) lacks important information regarding the description of the studied group (e.g. what criteria were used to qualify the participants, whether the patients had specific neurological symptoms, etc.). Moreover, considering the small size of the group, the authors should indicate that their study is a pilot study. I propose that my comments be included in the revised version of the text. After making appropriate corrections, I support the publication of the reviewed article.

 

Author Response

Comments 1: the text of the article (especially section 2.1.) lacks important information regarding the description of the studied group (e.g. what criteria were used to qualify the participants, whether the patients had specific neurological symptoms, etc.). 

Response 1: Thank you for your comment regarding the description of the studied group. To address this, we have included additional details about the exclusion criteria in the Data Acquisition section (lines 80–84). Specifically, we now state: "Participants were excluded if they met any of the following criteria: (1) comorbid psychotic, tic, or pervasive developmental disorders; (2) diagnosed central nervous system (CNS) disorders; (3) active use of psychiatric medications at the time of participation; (4) positive results on urine toxicology or pregnancy tests; (5) Wechsler Abbreviated Scale of Intelligence (WASI) scores below 70; or (6) medical conditions or physical limitations precluding MRI scanning (e.g., presence of metal implants, claustrophobia)." 

Comments 2: considering the small size of the group, the authors should indicate that their study is a pilot study. 

Response 2: Thank you for the suggestion. To address this, we have updated the title to reflect the pilot nature of the study. The revised title now reads: "Graph Neural Networks for Analyzing Trauma-Related Brain Structure in Children and Adolescents: A Pilot Study." This change highlights the exploratory nature of the research and aligns with the study's scope.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

I have formulated the following comments on the previous version of the article:

1. The study does not specify the number of participants or the demographics (e.g., age, gender, socio-economic status) in each group. A small or homogenous sample could limit the generalizability of the findings to broader populations of trauma-exposed children and adolescents.

2. While the study highlights structural differences between the trauma-exposed and control groups, the cross-sectional nature of the analysis makes it difficult to establish causal relationships between trauma exposure and the observed brain network changes. Longitudinal studies would be necessary to draw conclusions about the causal impact of trauma on brain connectivity.

3. The study presents network differences between the trauma-exposed and control groups, but the specific neurobiological mechanisms underlying these differences remain unclear. Further research is needed to explore whether the observed structural changes are directly related to trauma-related psychopathology or represent broader neurodevelopmental changes.

4. The study utilizes Graph Neural Networks (GNNs) to analyze brain networks, but the choice of this method may not fully capture all aspects of brain connectivity. GNNs are powerful tools but might overlook certain complex interactions between regions, especially if they are not directly correlated. Other network-based methods, such as functional connectivity analysis, could complement the findings.

5. The study primarily focuses on MRI-based data, which provides valuable structural information, but additional neuroimaging techniques such as fMRI or PET scans could provide more comprehensive insights into functional connectivity and metabolic changes related to trauma exposure.

6. Although the study controls for trauma exposure, other factors such as comorbid mental health conditions, substance use, or medication use in the participants could influence brain connectivity. These potential confounding variables are not addressed in the analysis, which could impact the interpretation of the results.

7. The higher reconstruction loss observed in the trauma-exposed group may reflect the complexity of their brain network, but it is not clear whether this result is due to the GNN model’s limitations or the actual neurobiological differences between groups. A more detailed analysis of the model's performance, including comparisons with other machine learning models, could provide more context.

8. The study focuses on predefined Regions of Interest (ROIs) and their connectivity. While this approach provides valuable insights into specific brain regions, it may miss broader network interactions that could be relevant for understanding trauma’s impact. More holistic, whole-brain analyses could further elucidate the complexity of brain connectivity.

9. The preprocessing steps used to prepare the MRI data for GNN analysis are not detailed in the study. Variations in data quality, such as motion artifacts or partial volume effects, could affect the accuracy of the results and should be addressed in future research.

The authors have addressed all my comments. I found their responses quite convincing. I support the publication of the current version of the article. I wish the authors creative success.