BrainNet with Connectivity Attention for Individualized Predictions Based on Multi-Facet Connections Extracted from Resting-State fMRI Data

Resting-state functional magnetic resonance imaging (RS-fMRI) has great potential for clinical applications. This study aimed to promote the performance of RS-fMRI-based individualized predictive models by introducing effective feature extraction and utilization strategies and making better use of information hidden in RS-fMRI data. We proposed a novel framework named multi-facet BrainNet with connectivity attention (MFBCA) to fulfill the purpose, and the framework is characterized by the following three strategies. First, in addition to the overwhelmingly popular functional connectivity, we also used distance correlation and weighted directed connectivity as multi-facet inputs for MFBCA. Second, a connectivity attention layer was proposed to force MFBCA to focus more on connections that are important for predictions. Finally, a BrainNet-based architecture with a feature fusion module was introduced to facilitate final predictions. We evaluated the performance of MFBCA with predictions of individuals' age and intelligence quotient as test cases based on three public RS-fMRI datasets. The results indicate that MFBCA can effectively utilize the information hidden in RS-fMRI data and outperform baselines. The predicted-vs-actual correlations for age predictions were 0.876 (7.314 years), 0.873 (8.121 years), and 0.681 (3.865 years), and for IQ predictions was 0.615 (4.287). The connectivity attention layer made it possible for us to determine the connections important for individualized predictions. MFBCA can be widely applied to predictions based on neuroimaging data from which connectivity maps can be extracted. Furthermore, the explicit physiological basis for predictions provided by the connectivity attention layer makes MFBCA a profitable choice for clinical applications.