Identifying Individualized Functional Brain Networks: An Unsupervised Deep Temporal Attention Based Autoencoder Model

Identifying functional brain networks (FBNs) under naturalistic stimuli is crucial for understanding the brain function. While numerous studies have employed deep learning methods to model neural responses, most of them overlook self-attention mechanism of brain and the temporal dependencies in neural activity. In this study, we proposed an unsupervised deep temporal attention based autoencoder model (DTA-AE) that aimed to model brain function in a compact, standardized latent space, thus representing complex brain activities via dense embedding vectors. Our method employed self-attention mechanisms to establish inter-regional temporal dependencies and simulate brain's attention allocation mechanism. Experimental results demonstrated that reliable and distinctive FBNs can be extracted. Specifically, the fingerprint analysis revealed that the identified FBNs can capture the individualized characteristics, enabling accurate subject recognition. Furthermore, temporal features exhibited good temporal consistency across subjects demonstrated by the inter-subject correlation (ISC) analysis, further confirming the effectiveness of the proposed method.