Discriminative analysis of non-linear brain connectivity in schizophrenia: an fMRI Study

Longfei Su, Lubin Wang, Hui Shen, Guiyu Feng, Dewen HuFrontiers in Human Neuroscience 2013 · DOI

Abstract

The classification results suggest that the non-linear functional connectivity provided useful discriminative power in diagnosis of schizophrenia, and the inverse but similar spatial distributed changes between the non-linear and linear measure may indicate the underlying compensatory mechanism and the complex neuronal synchronization underlying the symptom of schizophrenia.

Introduction

Schizophrenia is widely understood as a "disconnection syndrome"—a failure of distributed brain networks to integrate properly. To study this, researchers typically use functional MRI (fMRI) to measure functional connectivity, which is the statistical association between activity in different brain regions. However, the vast majority of these studies rely on the Pearson Correlation Coefficient (PCC), which only captures linear relationships. Given that the human brain is a complex, non-linear system, linear models likely miss subtle but critical pathological signatures. This paper evaluates whether non-linear connectivity can better discriminate between schizophrenic patients and healthy controls.

Methodology

The authors introduced a novel measure called the extended maximal information coefficient (eMIC) to isolate non-linear signals:

  1. Measuring Total Association: They used the Maximal Information Coefficient (MIC) to detect both linear and non-linear associations.
  2. Isolating Non-Linearity: They defined eMIC as the difference between the total association (MIC) and the squared linear correlation ($PCC^2$).
  3. Network Construction: The brain was divided into 116 regions (AAL atlas), resulting in 6,670 unique connection features per subject.
  4. Classification: Using a Support Vector Machine (SVM) and leave-one-out cross-validation, they tested the discriminative power of PCC, MIC, and eMIC across a dataset of 64 participants (32 patients and 32 healthy controls).

Results: Superior Discriminative Power

The eMIC metric consistently outperformed both PCC and MIC across various feature set sizes. When using the full feature set, the eMIC achieved a 82.8% generalization rate (accuracy), compared to 81.2% for PCC and 76.6% for MIC.

Classification performance (GR: Generalization Rate, SS: Sensitivity, SC: Specificity) for PCC (blue), MIC (green), and eMIC (red) across varying feature counts.

Furthermore, the eMIC demonstrated high sensitivity (84.4%), suggesting it captures pathological markers that linear models overlook. The study also identified 122 "consensus" non-linear connections that were highly stable across subjects, providing potential targets for future biomarkers.

Discussion: The Compensatory Mechanism

A striking finding was the "inverse phenomenon" in connectivity strength. While linear connectivity (PCC) generally decreased in schizophrenia patients, non-linear connectivity (eMIC) actually increased.

Connectivity matrices and mean connection strengths showing the inverse relationship between linear (PCC) and non-linear (eMIC) connectivity in patients vs. controls.

The authors hypothesize that this increase in non-linear strength represents a "compensatory mechanism." As the brain's standard linear communication channels break down, it may attempt to reconcile the loss by strengthening more complex, non-linear neuronal synchronizations. These changes occurred in similar hubs—specifically the Default Mode Network (DMN), the cerebellum, and the visual network—though eMIC identified unique connections, such as those between the cerebellum and temporal cortex, that PCC missed entirely.

Conclusion and Future Work

While the results suggest that non-linear connectivity provides useful discriminative power, several hurdles remain. The sample size of 64 is relatively small for machine learning, and the low-resolution fMRI data may have missed finer anatomical details. Future work should validate these findings in larger, multi-site datasets and attempt to link these non-linear changes to specific clinical symptoms to improve personalized medicine.

Summary updated

Abstract

The classification results suggest that the non-linear functional connectivity provided useful discriminative power in diagnosis of schizophrenia, and the inverse but similar spatial distributed changes between the non-linear and linear measure may indicate the underlying compensatory mechanism and the complex neuronal synchronization underlying the symptom of schizophrenia.
BibTeX
@inproceedings{su2013discriminative,
  author = {Longfei Su and Lubin Wang and Hui Shen and Guiyu Feng and Dewen Hu},
  title = {{Discriminative analysis of non-linear brain connectivity in schizophrenia: an fMRI Study}},
  booktitle = {Frontiers in Human Neuroscience},
  year = {2013},
  doi = {10.3389/fnhum.2013.00702},
}
Figure