Background & Aims
This study aims to used resting state functional connectivity to derive biomarkers for chronic pain. Recently, graph neural networks (GNNs) have seen a surge in prominence because of their success in illustrating unstructured relational information. However, the recent advancements in GNNs have not yet been fully utilized for the analysis of rs-fMRI data, especially when it especially to its spatiotemporal dynamics. In this study, we introduce a deep neural network structure that seamlessly combines both temporal and graphical convolutional neural networks. This enables us to comprehensively learn from the spatial and temporal aspects of rs-fMRI data in a unified manner. This approach integrates learning temporal dynamics, displaying intra- feature learning while likewise receiving interactions between ROI-wise dynamics, consequently working with inter-feature learning. We hypothesized that this methodology can improve our classification of chronic pain patients using resting state fun
Methods
Resting-state functional MRI (rsfMRI) information, including pain status, were extracted from the 9,000 UK BioBank (UKBB) dataset. The graph attention temporal convolutional networks (GATCN) were developed for predicting sex and chronic pain after splitting 70/30% of the data into training and testing sets. The Receiver Operating Characteristic Area Under the Curve (ROC-AUC) Score was used to compare model performance.
Results
We first tested the utility of our pipeline by classifying male and female participants using resting state functional connectivity. The model achieved an exceptional AUC of 0.92 (95% CI:0.89 – 0.94), indicating exceptional performance when training our GATCN on the time series for the classification of male and female participants. We next used the model to predict the presence or absence of chronic pain, a comparison for which we obtained close to random performance using linear models trained on connectivity matrices derived from the time series. Here, the GATCN model has a decent amount of accuracy when it comes to predicting chronic pain. The AUC was 0.71 (95% CI: 0.68-0.74), illustrating that the model could classify individuals with and without chronic pain considering time series information data.
Conclusions
The model’s robust performance in sex classification is highlighted by the GATCN results in combination. Predicting chronic pain based on GATCN model shows its potential utility in supporting the early diagnosis and understanding of chronic pain conditions. We propose that this model could lay the foundation for future deep learning methodologies centered on utilizing the intrinsically and inseparably spatio- temporal nature of rs-fMRI data.
References
[1] Arslan, S. et al. Graph saliency maps through spectral convolutional networks: application to sex classification with brain connectivity. Lect. Notes Comput. Sci. Springer Int. Publ., pp. 3–13, 2018.
[2] Gadgil, S. et al. Spatio-temporal graph convolution for resting-state fmri analysis. in: Medical image computing and computer assisted intervention,. MICCAI Springer Int. Publ., pp. 528–538, 2020.
[3] Li, X. et al. Braingnn: Interpretable brain graph neural network for fmri analysis. Elsevier, 7(102233), 2021.