Discriminative Network Models of Schizophrenia

Guillermo A. Cecchi, Irina Rish, Benjamin Thyreau, Bertrand Thirion, Marion Plaze, Marie-Laure Paillère-Martinot, Catherine Martelli, Jean-Luc Martinot, Jean-Baptiste Poline

Introduction

Schizophrenia is increasingly viewed not as a localized brain lesion, but as a systemic "disconnection" of neural networks. For decades, neuroscientists have struggled to find a specific, local abnormality in brain activity that consistently identifies the disorder. Traditional fMRI analysis often relies on the General Linear Model (GLM), which looks at which individual voxels "light up" during a task.

However, schizophrenia is hypothesized to be a disorder of emergent properties—it’s not about where the brain is active, but how different parts of the brain coordinate. This paper addresses the failure of localized, "mass-univariate" approaches to capture this delocalized pathology, demonstrating that data-driven functional network features are far more predictive of the disorder than traditional brain activation maps.

Methods and Data

The researchers compared two main approaches to analyzing fMRI data:

  1. Model-Driven (ROI) Approach: Examining 10 pre-defined Regions of Interest (ROIs) associated with language and schizophrenia, measuring correlations and interactions between these regions.
  2. Data-Driven (Network) Approach: Constructing voxel-level functional networks where an edge exists between two voxels if their activity correlation exceeds 0.7. From these networks, they extracted Degree Maps (Full Degree, Long-distance Degree, and Inter-hemispheric Degree).

To classify subjects, standard linear methods (Gaussian Naive Bayes and SVM) were compared against a Sparse Gaussian Markov Random Field (MRF), which explicitly models the conditional dependencies between voxels rather than treating them as independent features.

The study utilized fMRI data from 22 subjects (11 schizophrenic patients with persistent auditory hallucinations and 11 age-matched healthy controls) performing a simple auditory task.

Key Insights

The "Hub" Deficit and Network Stability Healthy brains exhibit clear "hubs"—voxels in the auditory cortex with high connectivity. In schizophrenic patients, these hubs are significantly diminished, resulting in a network that is more evenly spread and less efficient. Crucially, while standard activation maps vary wildly between different subsets of data, degree maps are remarkably stable. This suggests that network topology is a much more robust biological marker than task-based activation.

Statistical significance of degree maps vs. activation maps.

The Disconnection Hypothesis The results provide quantitative evidence for the century-old "disconnection hypothesis." Specifically, the schizophrenic group showed a significant bias toward low relative inter-hemispheric connectivity, meaning the two halves of the brain aren't communicating as effectively as they should.

Histogram showing reduced inter-hemispheric connectivity in schizophrenic patients.

Results

By modeling the brain as a complex graph, the data-driven approach achieved remarkable predictive success:

  • High Classification Accuracy: The Sparse MRF using long-distance degree maps achieved 86% accuracy, demonstrating that the relationships between brain regions contain more discriminative information than the activity levels themselves.
  • Network vs. Activation: Standard activation maps performed poorly, often hovering near chance level. Even the best-performing activation map was significantly outperformed by the simplest degree map.
  • ROI Failure: Traditional ROI-to-ROI correlation analysis failed to find any statistically significant differences between the groups after correcting for multiple comparisons, highlighting the necessity of the finer-grained, voxel-level network approach.

Limitations and Future Work

While the results are statistically significant and compelling, the study is based on a small cohort of 22 subjects. Larger datasets are needed to ensure these biomarkers generalize across the highly diverse schizophrenia spectrum. Furthermore, the patients in this study were specifically chosen for prominent auditory hallucinations; it remains unclear if these network disruptions manifest similarly in patients with primarily "negative" symptoms. Future work could also employ effective connectivity measures to investigate if the direction of information flow is similarly disrupted.

Summary updated

Abstract

Schizophrenia is a complex psychiatric disorder that has eluded a characterization in terms of local abnormalities of brain activity, and is hypothesized to affect the collective, ``emergent working of the brain. We propose a novel data-driven approach to capture emergent features using functional brain networks [Eguiluzet al] extracted from fMRI data, and demonstrate its advantage over traditional region-of-interest (ROI) and local, task-specific linear activation analyzes. Our results suggest that schizophrenia is indeed associated with disruption of global, emergent brain properties related to its functioning as a network, which cannot be explained by alteration of local activation patterns. Moreover, further exploitation of interactions by sparse Markov Random Field classifiers shows clear gain over linear methods, such as Gaussian Naive Bayes and SVM, allowing to reach 86% accuracy (over 50% baseline - random guess), which is quite remarkable given that it is based on a single fMRI experiment using a simple auditory task.
BibTeX
@inproceedings{cecchi2009discriminative,
  author = {Guillermo A. Cecchi and Irina Rish and Benjamin Thyreau and Bertrand Thirion and Marion Plaze and Marie-Laure Paillère-Martinot and Catherine Martelli and Jean-Luc Martinot and Jean-Baptiste Poline},
  title = {{Discriminative Network Models of Schizophrenia}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2009},
  url = {https://openreview.net/forum?id=ByV0ztW_WB},
}
Figure