Brain Network State Transformer: Leveraging State Functional Connectivity for Enhanced Brain Network Analysis

Recent advancements in functional Magnetic Resonance Imaging (fMRI) have highlighted the importance of capturing the dynamic nature of brain activities, prompting a shift from static Functional Connectivity (FC) to dynamic FC (DFC). However, existing DFC approaches often struggle to balance temporal granularity with interpretability, leading to challenges in disentangling meaningful connectivity patterns. In this work, we introduce the Brain Network State Transformer (BNST), a novel framework that leverages State FC to enhance brain network analysis. Our approach integrates three key steps: (1) Deep Clustering to identify recurring brain states from high-dimensional DFC matrices, (2) State-Based Rechunking to reorganize BOLD time series according to these states, and (3) a Transformer-Based Feature Extraction mechanism that models intra-state and inter-state relationships for downstream prediction tasks. We demonstrate the effectiveness of BNST on two publicly available fMRI datasets-ABCD and HCP-across both classification and regression tasks. By capturing structured temporal dynamics, BNST not only boosts prediction performance but also improves interpretability by identifying distinct brain states and their functional significance, providing a structured representation that aligns with meaningful cognitive and neural processes.