Detecting abnormal connectivity in schizophrenia via a joint directed acyclic graph estimation model

Gemeng Zhang, Aiying Zhang, Biao Cai, Zhuozhuo Tu, Vince D. Calhoun, Yu‐Ping WangNeuroImage 2020 · arXiv

Abstract

Functional connectivity (FC) has been widely used to study brain network interactions underlying the emerging cognition and behavior of an individual. FC is usually defined as the correlation or partial correlation between brain regions. Although FC is proved to be a good starting point to understand the brain organization, it fails to tell the causal relationship or the direction of interactions. Many directed acyclic graph (DAG) based methods were applied to study the directed interactions using functional magnetic resonance imaging (fMRI) data but the performance was severely limited by the small sample size and high dimensionality, hindering its applications. To overcome the obstacles, we propose a score based joint directed acyclic graph model to estimate the directed FC in fMRI data. Instead of using a combinatorial optimization framework, the structure of DAG is characterized with an algebra equation and further regularized with sparsity and group similarity terms. The simulation results have demonstrated the improved accuracy of the proposed model in detecting causality as compared to other existing methods. In our case-control study of the MIND Clinical Imaging Consortium (MCIC) data, we have successfully identified decreased functional integration, disrupted hub structures and characteristic edges (CtEs) in schizophrenia (SZ) patients. Further comparison between the results from directed FC and undirected FC illustrated the their different emphasis on selected features. We speculate that combining the features from undirected graphical model and directed graphical model might be a promising way to do FC analysis.

Introduction

Functional connectivity (FC) is a cornerstone of neuroimaging research, typically defined by correlations between brain regions. However, correlation does not imply causation, and standard FC fails to capture the directionality of interactions. Directed Acyclic Graphs (DAGs) offer a solution by mapping the causal flow of information, but they are notoriously difficult to estimate in high-dimensional fMRI datasets where the number of brain regions (variables) often far exceeds the number of available samples. Traditional "greedy search" methods struggle with this dimensionality, often producing unreliable results.

Methodology

To address these challenges, we propose a joint DAG estimation model that shifts from discrete combinatorial searches to a continuous optimization framework. Our approach incorporates three key innovations:

  1. Algebraic Constraint: We utilize the NOTEARS framework to represent the acyclicity requirement as a smooth algebraic equation: $h(W) = Tr(e^{W \circ W}) - d = 0$. This allows the use of standard calculus-based optimization (like L-BFGS) rather than exhaustive graph searches.
  2. Joint Learning: Instead of estimating graphs for healthy and patient groups independently, we use a joint objective function with a group regularization term. This encourages the models to share a similar underlying structure, effectively "borrowing strength" across the entire dataset to filter out noise.
  3. Sparsity: $L_1$ penalties are incorporated to ensure the resulting networks are sparse, highlighting only the most significant causal links.

Experimental Results

Simulation Studies

In synthetic tests using various noise distributions (Gaussian, Exponential, Gumbel), the joint model significantly outperformed existing methods such as PC, FGES, and LiNGAM. As shown in Figure 1, the joint model maintains a high True Positive Rate (TPR) and a low False Discovery Rate (FDR) even in low-sample regimes ($n=50$), which are typical for clinical fMRI studies.

Performance comparison of the Joint model against existing methods (PC, FGES, LiNGAM, NoTears) across varying node sizes with a small sample size (n=50).

Clinical Application: Schizophrenia

We applied the model to fMRI data from the MIND Clinical Imaging Consortium (MCIC), involving 92 schizophrenia (SZ) patients and 116 healthy controls (HC). The analysis revealed that SZ patients exhibit significantly lower global and local efficiency in their directed networks, suggesting a deficit in information integration.

Crucially, the model identified "Characteristic Edges" (CtEs) unique to each group. As visualized in Figure 2, healthy controls displayed a robust "in-hub" structure in the gyrus rectus (prefrontal cortex) that was disrupted in SZ patients. Many of the missing links in the SZ group were related to the olfactory cortex and subcortical regions like the pallidum, aligning with known sensory processing issues in the disorder.

Circular visualization of characteristic edges (CtEs) identified in schizophrenia patients versus healthy controls, highlighting disrupted connectivity patterns across brain lobes.

Discussion and Conclusion

A key finding of this work is the distinction between directed and undirected connectivity. We observed only a 13-15% overlap between the edges identified by our DAG model and those found via standard correlation methods. This suggests that directed graphs provide a unique and complementary "sketch" of brain activity.

By transforming DAG estimation into a continuous optimization problem and leveraging joint group learning, this research provides a robust framework for mapping the causal architecture of the brain. These findings offer new insights into the specific network disruptions underlying schizophrenia, demonstrating that the disorder is characterized not just by reduced connectivity, but by a fundamental reorganization of directed information flow.

Summary updated

Abstract

Functional connectivity (FC) has been widely used to study brain network interactions underlying the emerging cognition and behavior of an individual. FC is usually defined as the correlation or partial correlation between brain regions. Although FC is proved to be a good starting point to understand the brain organization, it fails to tell the causal relationship or the direction of interactions. Many directed acyclic graph (DAG) based methods were applied to study the directed interactions using functional magnetic resonance imaging (fMRI) data but the performance was severely limited by the small sample size and high dimensionality, hindering its applications. To overcome the obstacles, we propose a score based joint directed acyclic graph model to estimate the directed FC in fMRI data. Instead of using a combinatorial optimization framework, the structure of DAG is characterized with an algebra equation and further regularized with sparsity and group similarity terms. The simulation results have demonstrated the improved accuracy of the proposed model in detecting causality as compared to other existing methods. In our case-control study of the MIND Clinical Imaging Consortium (MCIC) data, we have successfully identified decreased functional integration, disrupted hub structures and characteristic edges (CtEs) in schizophrenia (SZ) patients. Further comparison between the results from directed FC and undirected FC illustrated the their different emphasis on selected features. We speculate that combining the features from undirected graphical model and directed graphical model might be a promising way to do FC analysis.
BibTeX
@article{zhang2020detecting,
  author = {Gemeng Zhang and Aiying Zhang and Biao Cai and Zhuozhuo Tu and Vince D. Calhoun and Yu‐Ping Wang},
  title = {{Detecting abnormal connectivity in schizophrenia via a joint directed acyclic graph estimation model}},
  journal = {NeuroImage},
  year = {2020},
}
Figure