Unraveling Diagnostic Biomarkers of Schizophrenia Through Structure-Revealing Fusion of Multi-Modal Neuroimaging Data

Evrim Acar, Carla Schenker, Yuri Levin‐Schwartz, Vince D. Calhoun, Tülay Adalı

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.

Figure 1: Coupled data sets including fMRI matrix (Y), EEG tensor (X), and sMRI matrix (Z), all coupled along the subject mode.

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).

Figure 2: Modeling of the EEG tensor coupled with fMRI and sMRI matrices using the ACMTF framework, showing the decomposition into shared and unique components.

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

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 since it requires extracting interpretable and reproducible patterns from data sets, consisting of shared/unshared patterns and of different orders. For example, multi-channel electroencephalography (EEG) signals from multiple subjects can be represented as a third-order tensor with modes: subject, time, and channel, while functional magnetic resonance imaging (fMRI) data may be in the form of subject by voxel matrices. Traditional data fusion methods rearrange higher-order tensors, such as EEG, as matrices to use matrix factorization-based approaches. In contrast, fusion methods based on coupled matrix and tensor factorizations (CMTF) exploit the potential multi-way structure of higher-order tensors. The CMTF approach has been shown to capture underlying patterns more accurately without imposing strong constraints on the latent neural patterns, i.e., biomarkers. In this paper, EEG, fMRI, and structural MRI (sMRI) data collected during an auditory oddball task (AOD) from a group of subjects consisting of patients with schizophrenia and healthy controls, are arranged as matrices and higher-order tensors coupled along the subject mode, and jointly analyzed using structure-revealing CMTF methods [also known as advanced CMTF (ACMTF)] focusing on unique identification of underlying patterns in the presence of shared/unshared patterns. We demonstrate that joint analysis of the EEG tensor and fMRI matrix using ACMTF reveals significant and biologically meaningful components in terms of differentiating between patients with schizophrenia and healthy controls while also providing spatial patterns with high resolution and improving the clustering performance compared to the analysis of only the EEG tensor. We also show that these patterns are reproducible, and study reproducibility for different model parameters. In comparison to the joint independent component analysis (jICA) data fusion approach, ACMTF provides easier interpretation of EEG data by revealing a single summary map of the topography for each component. Furthermore, fusion of sMRI data with EEG and fMRI through an ACMTF model provides structural patterns; however, we also show that when fusing data sets from multiple modalities, hence of very different nature, preprocessing plays a crucial role.
BibTeX
@inproceedings{acar2019unraveling,
  author = {Evrim Acar and Carla Schenker and Yuri Levin‐Schwartz and Vince D. Calhoun and Tülay Adalı},
  title = {{Unraveling Diagnostic Biomarkers of Schizophrenia Through Structure-Revealing Fusion of Multi-Modal Neuroimaging Data}},
  booktitle = {Frontiers in Neuroscience},
  year = {2019},
  doi = {10.3389/fnins.2019.00416},
}
Figure