A novel neural network model with SAGPooling graph decodes changes in cognitive trajectories.

Accurately predicting the progression of neurodegenerative diseases and identifying key brain regions influencing the disease course are critical research topics in neuroimaging. Self-Attention Graph Pooling (SAGPooling) can be used to enhance prediction efficiency and accuracy. The atlas-constrained spatial-informational SAGPooling graph neural network (ACSI-SGNN) model is proposed to enable precise prediction of patients' cognitive resilience and identification of brain regions closely associated with resilience. Considering the complex structure and functionality of brain networks, frequency-domain Granger causality is first utilized to calculate the functional states representing neuronal activity within specific frequency bands of raw fMRI signals and the information flow pathways between brain regions. These were integrated with spatial encoding based on the shortest path to construct a brain network model. The resulting brain network is used for feature aggregation in SGNN, leveraging the ROI selection pooling layer in the model to highlight brain regions. For the cognitive resilience three-class classification task, the proposed ACSI-SGNN model achieved an ACC of 87.50%, an F1 score of 87.68%, a precision of 89.58%, and a recall of 88.33% on the ADNI dataset and an ACC of 88.24%, an F1 score of 86.71%, a precision of 89.87%, and a recall of 83.33% on the HABS dataset, respectively. Results indicate the effectiveness of the proposed method in identifying brain aging rates and neurodegenerative disease progression. Our approach outperformed other methods across three evaluation metrics and identified key brain regions highly consistent with those previously reported as critical to cognitive resilience in neuroimaging studies.