DCHO: A Decomposition-Composition Framework for Predicting Higher-Order Brain Connectivity to Enhance Diverse Downstream Applications

TL;DR

DCHO tackles the challenge of dynamic HOBC modeling by decomposing HOBC forecasting into HOBC tensor inference and latent trajectory prediction. It combines a dual-view encoder to capture multiscale local and global topology with a higher-order decoder that models triplet interactions via a latent combinatorial learner, and it adds a latent-space prediction loss to robustly forecast HOBC dynamics. The method is evaluated on synthetic and HCP fMRI datasets, showing superior performance for HOBC prediction, task-state classification, and brain dynamics forecasting, with ablations confirming the value of each component. The work advances HOBC-enabled neurocognitive modeling and offers a scalable framework for downstream neuroscience applications. Theoretical bounds further justify the two-stage error decomposition, underscoring improved controllability over end-to-end approaches.

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.

Figures (4)

• Figure 1: Motivation. (Left): Previous methods analyze HOBC in static windows, overlooking their temporal evolution and limiting predictive applications. (Right): DCHO overcomes this limitation by forecasting the dynamic trajectories of HOBC.
• Figure 2: Overview of DCHO (A) Decomposition–Composition Framework: DCHO framework decomposes prediction into HOBC inference and latent trajectory prediction, using a latent-space prediction loss to model high-level temporal dynamics. (B) Dual-view Encoder: DCHO applies two parallel GNN branches that extract local and global topological features. (C) Higher-order Decoder: DCHO proposes a latent combinatorial learner to capture high-level HOBC information. (D) DCHO leverages the pretrained encoder and predictor to support both non-predictive and predictive tasks.
• Figure 3: (a)-(b): Multi-step HOBC prediction performance of DCHO and LSTM on the Emotion dataset. (c)-(d): Multi-step raw fMRI signal prediction performance of DCHO and AMAG on the Emotion dataset. (Metrics: MAE (left) and RMSE (right), Prediction lengths: 10 to 50 steps)
• Figure 4: Comparison of task state classification performance across raw fMRI signals, FC, HOBC, and DCHO representations: DCHO outperforms raw fMRI signals, FC, and HOBC across seven cognitive task datasets and five metrics, as shown in (a–g), with average results in (h).

Theorems & Definitions (2)

• Theorem 1
• Remark 1