Source-Decomposed Multimodal Graph Connectivity: A Temporal Mixture-Manifold Pipeline for Tracking Ischemic Injury and Treatment

To capture the complex, dynamic network reorganization following acute focal brain injuries that standard static or unimodal neuroimaging frameworks routinely miss, we present an open, reproducible computational multimodal pipeline that integrates resting-state functional MRI and diffusion tensor imaging (DTI) into dynamic, source-decomposed graph networks. This framework preserves sub-voxel, multi-peak distributional dynamics over time by modeling regional blood-oxygen-level-dependent signals within sliding windows using a Gaussian mixture model, with model complexity dynamically optimized. The resulting pairwise similarities define time-varying edge weights across local, nonlocal, and DTI-supported structural sources, unifying them into a multimodal connectivity manifold that precisely quantifies node strength, topological turnover, and boundary modifications. Applied to a non-human primate model of photothrombotic ischemia, the methodology revealed severe post-injury disruptions, with 66 of 144 nodes (45.8%) fundamentally altering their connectivity across localized sources. Furthermore, coupling this architecture with an Ensemble Kalman Filter (EnKF) allowed for the continuous tracking of a tissue “reversibility” coordinate, yielding earlier detection of metabolic state collapse than unimodal baselines and generating a forecast-divergence index as a rigorous early-warning signal. Extended across post-injury treatment boundaries to evaluate network recovery toward baseline, this interpretable, modality-aware framework establishes a standardized, shareable methodological foundation for tracking complex pathophysiological dynamics in neuroinformatics research.