¹Jeonbuk National University Hospital, Psychiatry, Jeonju, Korea, Republic of

Background: Developing a reliable and objective method to distinguish patients from healthy controls using graph machine learning algorithms is crucial for improving diagnostic accuracy and contributing to personalized treatment interventions. However, few studies have focused on Post-Traumatic Stress Disorder (PTSD) to date.

Aims & Objectives: This study aims to develop a diagnostic model to distinguish between individuals with PTSD and healthy controls (HC). Utilizing memory task-based fMRI data, we can seek to identify distinctive patterns in brain activity that can serve as markers for the diagnosis of PTSD and HC individuals.

Method: Twenty-three healthy controls and 13 PTSD diagnosed underwent verbal memory task fMRI scanning. Graph analysis was conducted through the application of the CONN-fMRI FC toolbox (version 21a) in conjunction with SPM12 software. Graph matrices were constructed with 164 nodes as Regions of Interest (ROI), and a connectivity matrix was derived by ROI-to-ROI correlations. After obtaining the connectivity matrix, we applied thresholding to eliminate less relevant information, creating an adjacency matrix as part of the preprocessing. Subsequently, we extracted low-dimensional features (specifically, 10 features) from functional brain connectivity networks using the "graph2vec" graph embedding technique. Our classification approach involved employing RandomForest, XGBoost, Support Vector Classifier (SVC), Gaussian Naïve Bayes (Gaussian NB), and k-Nearest Neighbors (KNN) models to identify PTSD-related brain networks based on the aforementioned low-dimensional features. The obtained classification results were compared against the ground-truth to derive classification accuracy. To assess prediction accuracy robustly, we employed k-fold cross-validation with k=5.

Results: Result of this study illustrates the classification accuracy achieved across the rest, encoding, and retrieval datasets. In the rest dataset, XGBoost exhibited the highest performance with a 73% classification accuracy at the 0.05 threshold level, while Gaussian NB and KNN models achieved a comparable accuracy of 73% at the 0.15 threshold level. For the encoding dataset, XGBoost demonstrated the most effective classification performance, achieving a 75% accuracy at the 0.25 threshold level. Finally, in the retrieval dataset, Gaussian NB showed as the best performance model, attaining a 70% classification accuracy at the 0.45 threshold level.

Discussion & Conclusions: This study distinguishes itself from prior machine learning-based classification studies because we utilize memory tasks in the fMRI dataset for PTSD. Our study is the first to use memory task-based fMRI along with graph machine learning algorithms to identify PTSD. The different brain functional network patterns observed during memory tasks provide valuable insights for clinical understanding and identification of PTSD. In our results, XGBoost demonstrated strong performance in both the rest and encoding datasets, indicating its ability to effectively capture complex patterns associated with PTSD. On the other hand, Gaussian NB was outstanding in the retrieval dataset, highlighting its proficiency, particularly in the memory recall process. The result that different models succeeded in various datasets highlights importance of customizing machine learning methods for specific memory processing in PTSD. In conclusion, our findings suggest that different brain functional network patterns during memory tasks are valuable for clinical understanding and identification of PTSD.