Abstract
Higher-order brain connectivity (HOBC), which captures interactions among three or more brain regions, provides richer organizational information than traditional pairwise functional connectivity (FC). Recent studies have begun to infer latent HOBC from noninvasive imaging data, but they mainly focus on static analyses, limiting their applicability in dynamic prediction tasks. To address this gap, we propose DCHO, a unified approach for modeling and forecasting the temporal evolution of HOBC based on a Decomposition-Composition framework, which is applicable to both non-predictive tasks (state classification) and predictive tasks (brain dynamics forecasting). DCHO adopts a decomposition-composition strategy that reformulates the prediction task into two manageable subproblems: HOBC inference and latent trajectory prediction. In the inference stage, we propose a dual-view encoder to extract multiscale topological features and a latent combinatorial learner to capture high-level HOBC information. In the forecasting stage, we introduce a latent-space prediction loss to enhance the modeling of temporal trajectories. Extensive experiments on multiple neuroimaging datasets demonstrate that DCHO achieves superior performance in both non-predictive tasks (state classification) and predictive tasks (brain dynamics forecasting), significantly outperforming existing methods.
Introduction
In network neuroscience, the standard paradigm for modeling brain function relies on Functional Connectivity (FC), typically represented as pairwise statistical correlations between brain regions. However, the brain's computational architecture involves complex, simultaneous coordination among multiple regions—a phenomenon known as Higher-Order Brain Connectivity (HOBC). While static HOBC inference has seen progress, brain activity is inherently dynamic and non-stationary.
Predicting the temporal evolution of these higher-order networks is challenging due to the "combinatorial explosion" of possible multi-region interactions and the resulting high dimensionality. DCHO addresses these challenges by providing a framework capable of forecasting HOBC dynamics, thereby supporting both state classification and future brain activity prediction.
Methodology: The Decomposition-Composition Strategy
To circumvent the mathematical complexity of high-dimensional HOBC tensors, DCHO employs a Decomposition-Composition strategy. This approach splits the forecasting task into two distinct, manageable phases:
- HOBC Inference (Spatial): Historical fMRI signals and pairwise FC matrices are processed through a Dual-view Encoder. This component utilizes parallel Graph Neural Networks (GNNs) to capture both local and global topological features. A Higher-order Decoder, featuring a latent combinatorial learner, then reconstructs the HOBC tensor from this compressed latent space.
- Latent Trajectory Prediction (Temporal): Instead of predicting the future HOBC tensor directly, DCHO forecasts the evolution of the latent representations. By freezing the encoder and utilizing an LSTM-based predictor guided by a specialized "latent-space prediction loss," the model ensures that temporal dynamics are modeled in a robust, low-dimensional space.
This decoupling of spatial inference and temporal forecasting allows for more controlled error propagation and improved stability compared to end-to-end architectures.
Experimental Evaluation
The DCHO framework was rigorously tested across diverse datasets, including the Human Connectome Project (HCP) (covering seven task-based fMRI datasets and one resting-state dataset) and synthetic non-linear dynamical systems like the Lorenz system and the Hindmarsh-Rose (HR) model.
Key Results
- Dynamic Forecasting: DCHO consistently outperformed standard temporal models (MLP, LSTM, Transformer) in both short-term and long-term HOBC prediction. In raw fMRI dynamics forecasting, DCHO surpassed state-of-the-art models like AMAG and NetFormer across all ten datasets.
- Downstream Classification: The learned representations significantly improved cognitive state classification. On average, DCHO achieved a 6.24% increase in accuracy and a 6.77% improvement in AUROC compared to traditional FC and static HOBC baselines.
- Cognitive Relevance: The most significant performance gains were observed in complex cognitive tasks (e.g., Emotion, Working Memory), suggesting that HOBC effectively captures the integrative computations necessary for advanced human cognition.
Conclusion
DCHO provides a robust solution for modeling the dynamic nature of higher-order brain networks. By leveraging a latent-space decomposition-composition approach, it overcomes the dimensionality hurdles of HOBC. Future work aims to extend this framework to multimodal neuroimaging data, such as combining fMRI with EEG, to further enhance its predictive and diagnostic capabilities.
Summary updated