Deep Learning for Cognitive Task Presence Prediction from Dynamic Functional Connectivity

Dynamic functional connectivity (dFC) studies the time-varying coordination between brain regions measured with fMRI and is a potential biomarker for understanding cognitive dynamics and tracking the development of neurological disorders. However, a critical methodological challenge lies in the variability of dFC estimates across different dFC assessment methods, raising concerns about the reliability and interpretation of downstream findings. While deep learning (DL) models have demonstrated the ability to capture traditionally inaccessible data patterns in many disciplines, they encounter challenges when applied to neuroimaging data. For instance, the high dimensionality, noise, and temporal complexity inherent in dFC makes it challenging for DL models to extract meaningful and interpretable insights. In this study, we investigated how DL architectures can be developed and adapted to predict task presence over time from task-based dFC data, and additionally, how the choice of dFC assessment method influences the predictive performance of DL models. We developed and compared a convolutional neural network (CNN), a node-level classification graph convolutional network (GCN), and a graph-level classification GCN based on their ability to predict time points at which subjects were performing a cognitive task or at rest. In our study, the results indicate that both DL model architecture and dFC estimation methodology significantly impact task presence prediction capacity, while the specific task paradigm had minimal influence out of the limited types that were explored. This work offers a powerful benchmark for understanding the dynamics of underlying task-driven cognitive state transitions and the analytical flexibility limitations of dFC estimation methods and DL architectures.