Transformer-based Functional Time Series Modeling to Unveil Dynamic Brain State Transitions

Brain state is dynamic in nature, and identifying the transitions between different brain states is crucial for understanding ongoing cognitive functions. Here, we opted for a transformer-based blood-oxygen-level-dependent transformer model (BolT) to investigate dynamic transitions of brain states across time. Using the given time series data, the model predicted subsequent time series. At the same time, it provided information about the contribution of each time point for the prediction with the importance weights. We extracted the time series corresponding to the top 10% and bottom 10% contribution of the importance weights. For each time series data, we constructed the connectivity matrix and calculated the degree centrality values. We compared the degree values between the top 10% and bottom 10% contribution data and found significant differences in the regions involved in the default mode network. Specifically, decreases in the degree values were observed at the highly contributed time points. Our findings indicate that the reorganization of connectome organization occurs in the default mode network during transitions in brain states.