BrainFTFCN: Synergistic feature fusion of temporal dynamics and network connectivity for brain age prediction

Using neuroimaging-derived data for age estimation serves as a prominent approach in comprehending the normal pace of brain development and mechanisms underlying cognitive declines due to aging and neurological diseases. Despite the promise of resting-state functional magnetic resonance imaging (rs-fMRI) for brain age prediction, previous deep learning models have prioritized capturing either the temporal dynamics via time courses (TCs) or the inherent network topology revealed by functional network connectivity (FNC). These fragmented models neglect the complementary information available by synergistically integrating both. To address this, we introduced BrainFTFCN, a novel feature fusion network that synergistically integrates TCs and FNC for enhanced brain age prediction and model interpretability. BrainFTFCN uniquely combines a Temporal Attention Autoencoder (TAAE) to model evolving activity patterns within TCs and a Functional Connectivity Graph Attention Network (FCGAT) to capture spatial relationships embedded within FNC. The fused features were then fed into a support vector regression model for final age prediction. BrainFTFCN's efficacy shone on Cam-CAN dataset, outperforming state-of-the-art models by 28.21% in mean absolute error (MAE) and demonstrating consistent improvement across other metrics. Ablation studies solidified the critical role of multi-feature integration in boosting prediction. Notably, the most crucial brain regions and discriminative FNC can be easily unveiled via LASSO regression and GNNExplainer respectively, together unlocking biological interpretability and highlighting the model's potential for uncovering valid aging biomarkers.