The human brain can be considered a complex network, consisting of different areas that continuously exchange their information with each other and form a complex graph called brain graph. The nodes of this graph are different regions of the brain, and its edges are made based on time series extracted from fMRI data. Therefore, this graph can show the interaction between different areas of the brain, as well as express the abnormal patterns that have arisen under the influence of brain disorders. Recent research has shown that the statistical dependence between the time series of different brain regions changes over time. Therefore, modeling the brain with a fixed graph can cause a lot of information to be lost. On the other hand, researches have shown that considering the relationship between samples and using their phenotypic information, such as age and gender, can be of great help in identifying brain disorders. Therefore, in this research, a hierarchical model; Consisting of two parts, it is introduced to find the appropriate representation (embedding) of dynamic brain graphs, followed by the classification of these graphs. The first part of this model, after creating dynamic brain graphs from fMRI data, proceeds to learn the spatial and temporal characteristics of these data and produces the appropriate display for each of them. Also, by using the attention-based pooling layer, this section can reduce the size and modify the graph structure during training. After that, the second part of the model, by placing the representation obtained from the first part as the features of the nodes and calculating the weight of the edges based on the phenotypic data and the degree of similarity between the samples, builds the population graph and analyzes it by the attention-based graph neural network. Finally, the representation obtained for nodes is used to classify samples with specific brain disorders from healthy samples. Also, based on whether the first and second parts of the model are trained simultaneously or separately, various training strategies have been proposed, some of which require a lot of memory during execution. For this purpose, other training strategies have been suggested and reviewed to reduce the required memory. The performance of the proposed model has been eva‎luated on the ABIDE and ADHD_200 datasets to classify people with brain disorder from healthy people, and the results compared to the results of competing models indicate the superiority of this model in terms of various eva‎luation metrics.

مهران صفاياني

عبدالرضا ميرزايي , فرزانه شايق بروجني