Abstract
Fusing complementary information from different modalities can lead to the discovery of more accurate diagnostic biomarkers for psychiatric disorders. However, biomarker discovery through data fusion is challenging as it requires extracting interpretable and reproducible patterns from datasets of different orders. This paper presents a structure-revealing fusion framework using Advanced Coupled Matrix and Tensor Factorization (ACMTF) to analyze EEG, fMRI, and sMRI data. By coupling these modalities along a shared subject mode, the researchers successfully identified significant biomarkers for schizophrenia, demonstrating improved clustering performance and spatial resolution compared to single-modality analysis.
Introduction: The Challenge of Multi-Modal Data Fusion
Neuroscience relies on various imaging modalities to study psychiatric disorders. Electroencephalography (EEG) provides excellent temporal resolution but is spatially coarse. Functional MRI (fMRI) offers high spatial resolution but is slow, while structural MRI (sMRI) provides anatomical maps. The core challenge is data fusion: integrating a 3D EEG tensor (Subject × Time × Electrode) with 2D fMRI or sMRI matrices (Subject × Voxel).
Traditional methods often flatten the EEG tensor into a matrix, which destroys its inherent multi-way structure. Furthermore, existing techniques struggle to distinguish between neural patterns that are shared across modalities and those that are unique. This study addresses these issues to find more reproducible and interpretable diagnostic biomarkers.
Methodology: Advanced Coupled Matrix and Tensor Factorization (ACMTF)
The researchers utilized Advanced Coupled Matrix and Tensor Factorization (ACMTF). Unlike standard fusion, ACMTF treats EEG data as a third-order tensor, preserving its multi-way structure. The model factorizes the EEG tensor and the fMRI/sMRI matrices simultaneously, coupling them through a shared subject mode. This assumes that the underlying weight of a specific neural component for a given individual remains consistent across different imaging modalities.
The "Advanced" component of ACMTF incorporates sparsity penalties ($\ell_1$-norm regularization). This allows the model to identify which biomarkers are shared and which are modality-specific by effectively "turning off" components in modalities where they are absent. This framework provides a more flexible and interpretable alternative to joint Independent Component Analysis (jICA).
Experimental Setup and Datasets
The study utilized data from 32 participants (21 healthy controls and 11 patients with schizophrenia) performing an Auditory Oddball Task (AOD). In this task, subjects listen to tones and respond to infrequent "target" sounds. The multi-modal dataset included:
- EEG: 62 electrodes, 451 time points.
- fMRI: ~60,000 voxels.
- sMRI: ~300,000 voxels (gray matter maps).
Key Results and Insights
- Enhanced Diagnostic Accuracy: Fusing EEG and fMRI using ACMTF resulted in a clustering accuracy of 88–91% for distinguishing patients from controls, a significant improvement over the 81% accuracy achieved using EEG data alone.
- Biomarker Identification: The model identified a critical component related to the P3 peak (an event-related potential associated with attention). Patients with schizophrenia showed decreased activation in the sensorimotor cortex and parietal lobe within this component.
- Spatial Resolution Gain: Joint analysis effectively "enhanced" the spatial resolution of EEG biomarkers by leveraging fMRI data, pinpointing specific regions like the anterior sensorimotor cortex.
- Reproducibility: Sensitivity analysis confirmed that the neural patterns extracted by ACMTF are stable and reproducible across different subject subsets.
- Preprocessing Sensitivity: The inclusion of sMRI data revealed that results are highly sensitive to preprocessing. Without centering data across subjects, the model tended to capture average brain structure rather than the subtle differences indicative of schizophrenia.
Conclusion and Future Directions
By preserving the multi-way structure of neuroimaging data, ACMTF offers a powerful framework for discovering interpretable biomarkers. While the small sample size is a limitation, the results demonstrate that structure-revealing fusion can significantly improve the identification of psychiatric disorders. Future work will focus on automating parameter selection and applying this framework to larger cohorts and other neurological conditions, potentially incorporating additional modalities like Diffusion Tensor Imaging (DTI).
Summary updated