Related Experiment Videos
NeuroMoE++: Patient-Adaptive Multi-Level Multimodal Fusion With Mixture-of-Experts for Neurological Disorder
IEEE Transactions on Bio-Medical Engineering
|May 13, 2026
Summary
NeuroMoE++ enhances neurological disorder diagnosis by integrating multiple data types early. This novel approach significantly improves accuracy by capturing subtle cross-modality patterns for better patient-specific decisions.
Area of Science:
- Neuroscience
- Medical Imaging
- Machine Learning
Background:
- Early diagnosis of neurological disorders (NDs) relies on diverse data like MRI and biomarkers.
- Current multimodal learning often uses late fusion, missing crucial cross-modality interactions.
- Subtle patterns in early disease stages are underutilized by existing methods.
Purpose of the Study:
- To introduce NeuroMoE++, a hierarchical framework for explicit cross-modality interaction in neurological diagnosis.
- To improve the utilization of complementary information from different diagnostic signals.
- To enable patient-specific diagnostic decisions through adaptive integration.
Main Methods:
- Developed NeuroMoE++, a hierarchical end-to-end framework for multimodal learning.
- Enabled explicit interaction across modalities during feature extraction.
- Implemented subject-driven adaptive integration for patient-specific decisions.
Main Results:
- NeuroMoE++ achieved 84.94% accuracy on real clinical datasets.
- The model significantly outperformed unimodal and late-fusion baseline methods.
- Demonstrated the effectiveness of explicit cross-modality reasoning.
Conclusions:
- NeuroMoE++ offers a more effective approach to neurological disorder diagnosis.
- Explicit cross-modality reasoning is valuable for capturing subtle diagnostic patterns.
- The framework provides interpretable, patient-specific decisions for clinical application.
