DynFS-MoE: Dynamic Functional-Structural Mixture-of-Experts for Post-Traumatic Epilepsy Diagnosis

A research team has proposed a dynamic multimodal framework called DynFS-MoE to improve early identification of post-traumatic epilepsy, a severe complication of traumatic brain injury [1]. Post-traumatic epilepsy (PTE) develops after traumatic brain injury (TBI), but early detection remains difficult because of the complex structural and functional brain changes the injury triggers [1][2]. The new framework, detailed in a paper submitted to arXiv on June 15, 2026, integrates functional MRI and structural MRI through a time-aware functional-structural encoder and a class-conditioned expert routing system [1][2]. The architecture, named DynFS-MoE, has three components: a time-aware functional-structural encoder that refines temporal fMRI and structural sMRI data into modality patches; a mixture-of-experts layer with modality-specific and cross-modal experts that capture both within-modality characteristics and cross-modal interactions; and a modality-class MoE routing module that uses a gating mechanism conditioned on class-aware representations to dynamically assign expert weights [2][3]. Across three binary classification tasks, the framework consistently outperformed static fusion baselines [1][2]. The dynamic routing strategy overcomes limitations of static fusion methods and improves discrimination performance across multiple classification objectives, the authors report [2][3]. Interpretability analyses revealed region-of-interest interaction patterns, including thalamic hyperconnectivity and limbic disruptions, which the researchers describe as clinically relevant insights [2][3]. The model effectively captures class-dependent brain interaction patterns and provides an interpretable approach for PTE diagnosis and risk stratification [1][2]. The work appears on arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 submissions per month and hosts papers across mathematics, physics, computer science, and related fields [9]. Papers on arXiv are moderated but not peer-reviewed [9]. The DynFS-MoE framework joins a growing body of research applying mixture-of-experts architectures to neurological disease identification. A separate study proposed NeuroMoE, a transformer-based MoE framework for classifying Parkinson’s disease, idiopathic REM sleep behavior disorder, and healthy controls using MRI scans, clinical assessments, and serum biomarkers, achieving a validation accuracy of 82.47% [4]. Another effort, BrainNet-MoE, applied brain-inspired MoE learning to differentiate Alzheimer’s disease from Lewy body dementia, designing disease-specific expert groups to process brain sub-networks under different conditions [5].