Predicting efficacy of electroconvulsive therapy for adolescent major depressive disorder using a dual-branch graph attention network fusing multi-modal MRI

Major Depressive Disorder (MDD) significantly contributes to global disease burden, and Electroconvulsive Therapy (ECT) is an effective yet variable treatment. This study aims to develop an individualized prediction framework for ECT treatment response in adolescent MDD patients using multi-modal magnetic resonance imaging (MRI) and advanced deep learning. We recruited 27 adolescent MDD patients undergoing ECT, acquiring structural MRI (sMRI) and functional MRI (fMRI) before and after treatment. Individual morphological similarity networks and functional connectivity networks were created by utilizing sMRI and fMRI data, respectively. We introduced a novel Dual-Branch Graph Attention Network (DBGAN) which integrates two parallel graph attention networks for utilizing similarity and connectivity networks. The proposed deep model dynamically fuses information from sMRI and fMRI via a cross-attention mechanism, improving the prediction performance on ECT treatment response. Among 27 participants, 21 responded positively to ECT. According to experimental results, our DBGAN outperformed traditional machine learning models and deep learning models, achieving a mean accuracy of 0.853, precision of 0.920, recall of 0.910, and an F1-score of 0.905. Interpretability analyses indicated that predictive decisions were influenced by fMRI signals primarily in the right posterior insula and right dorsal cingulate gyrus, and sMRI signals predominantly from limbic areas, including the left amygdala and right hippocampus. Our DBGAN model effectively predicts ECT responses in adolescent MDD patients using multi-modal MRI. Our method provides a potential application for personalized treatment of adolescent MDD. Graphical Abstract: The flowchart illustrates the overall framework of this study for predicting the efficacy of ECT in adolescent patients with MDD. Patients underwent sMRI and fMRI scans before and after ECT and were classified as responders or non-responders. The imaging data were used to construct individual structural and functional brain networks. Finally, topological features from these networks were fed into models to predict individual treatment outcomes. • A novel deep learning model, a Dual-Branch Graph Attention Network (DBGAN), was proposed, which dynamically fuses information from sMRI and fMRI through a cross-attention mechanism. • The DBGAN model outperformed traditional machine learning models and a standard CNN model in predicting ECT efficacy, achieving a mean accuracy of 0.853 and an F1-score of 0.905. • The model's interpretability analysis indicated that the right posterior insula and right dorsal cingulate gyrus from fMRI signals, and limbic areas such as the left amygdala and right hippocampus from sMRI signals, were most influential in predicting ECT response.