Functional Brain Connectivity Research Using Bayesian Multiplex Graph Classifier

Research Background and Problem Statement

In recent years, research on cognitive control in the elderly has garnered increasing attention, especially against the backdrop of accelerating population aging. Understanding cognitive functions in the elderly becomes particularly important, not just due to the medical costs involved, but also because of the significant economic and social impact brought about by an aging society. Investigating changes in brain functional connectivity during cognitive control tasks in the elderly can provide valuable insights into the field of cognitive neuroscience. This study aims to explore the linkage between cognitive decline and interactions among various brain regions using functional magnetic resonance imaging (fMRI) data.

Paper Source and Author Information

This paper was authored by Sharmistha Guha, Jose Rodriguez-Acosta, and Ivo D. Dinov from Texas A&M University and the University of Michigan. The paper was accepted for publication on May 22, 2024, in the journal Neuroinformatics. Detailed link: https://doi.org/10.1007/s12021-024-09670-w.

Research Process

Data Acquisition and Processing

This study uses fMRI data to generate functional connectivity maps. The data comes from 144 healthy elderly participants, aged 20 to 86, who participated in inhibitory and initiation tasks. These tasks were used to measure functional activity in different brain regions during cognitive control. The specific process is as follows:

1. Data Collection: Functional activity of participants’ brains during inhibitory and initiation tasks was recorded using a Siemens 3T MRI scanner.
2. Task Execution: During the experiment, participants reclined on the scanning bed, observed experimental stimuli on a screen via a mirror, and responded to stimuli using their right hand with a response box, pressing buttons with their index and middle fingers.
3. Data Preprocessing: Preprocessing included rigid-body alignment, removal of outlier volumes, correction for cardiac and respiratory noise, temporal calibration, spatial smoothing, temporal trend removal, motion parameter regression, regression of signals from non-interest regions, and registration of data to standard space (MNI space).

Model Construction and Algorithm Design

Given the need to classify binary age outcomes (normal or aged) using multiplex graphs as predictors, this study proposed a Bayesian Multiplex Graph Classifier (BMGC):

1. Regression Framework Construction: A high-dimensional generalized linear model was constructed, modeling the edge coefficients of functional connectivity maps as double-linear interactions of two-point latent effects, which can reasonably handle the topology of multiplex graphs.
2. Variable Selection: A variable selection framework was applied to node-specific latent effects at all layers to identify nodes significantly related to the outcome.
3. Computational Methods: Bayesian methods with high computational efficiency were employed to quantitatively assess the uncertainty in node identification, coefficient estimation, and binary outcome prediction.

Algorithm Validation and Performance Evaluation

The study used simulated data and actual fMRI data for model validation:

1. Simulation Data Generation: Simulated data with different numbers of layers and node sparsity levels were generated to test the model’s performance. The simulation data’s true labels were validated using low-rank coefficient matrices.
2. Performance Comparison: The model’s performance in coefficient estimation and predictive inference was compared with various algorithms, including Lasso, Bayesian Lasso, Bayesian Horseshoe, Tensor Regression, and neural networks.

Main Research Findings

Simulation Data Results

Simulation experiments showed that BMGC performed exceptionally well in node identification, coefficient estimation, and prediction accuracy:

1. Node Identification Accuracy: The model accurately identified important nodes under high node sparsity conditions, with low uncertainty.
2. Coefficient Estimation: BMGC exhibited lower mean squared error (MSE) than competitors in all simulated scenarios, especially with high node density.
3. Predictive Performance: The area under the receiver operating characteristic (ROC) curve (AUC) and F1 score on holdout test samples showed that BMGC’s classification accuracy was significantly superior to other algorithms.

Actual Data Results

In the analysis of real fMRI data, BMGC successfully identified brain regions related to early aging:

1. Brain Connectivity Symmetry and Asymmetry: Symmetric connections were found in sensorimotor regions, while significant asymmetric connections were observed in the default mode network.
2. Classification Performance: Similar to simulation data, BMGC showed superior classification performance, with AUC and F1 scores significantly higher than other methods.

Research Conclusions and Significance

Scientific and Applied Value

The proposed Bayesian Multiplex Graph Classifier (BMGC) exhibits excellent performance in handling multiplex graph structured data. The model can accurately identify nodes related to the outcome, especially suitable for moderate sample sizes with complex interactions among nodes. Furthermore, BMGC’s outstanding performance on various datasets demonstrates its broad application potential, including but not limited to neuroscience and genomics data analysis.

Research Highlights

1. Novel Approach: A Bayesian multiplex graph classifier framework was proposed to handle multiplex graph data, featuring model simplicity and classification accuracy.
2. Utilization of Multiplex Graph Structure: Effective capture of complex inter-layer associations through double-linear latent effects significantly improved predictive performance.
3. Uncertainty Quantification: A natural advantage of Bayesian methods is providing uncertainty quantification for node identification, a significant advantage over other methods.

Future Research Directions

Future research could further explore the impact of nonlinear edge effects on classification results, combining semiparametric models to capture more complex associations. Additionally, further analysis of posterior probability distributions of left and right brain hemisphere differences can better understand brain symmetry and asymmetry under different pathological states.

Appendix and Additional Information

Complete fMRI data, computational tools, protocols, and supplementary materials can be accessed at the following URLs: https://socr.umich.edu/docs/uploads/2024/fmri_corr_pilot.html and https://github.com/jeroda7105/classification-with-multi-layer-graphs.

Conclusion and Future Work

This paper provides an innovative Bayesian framework for addressing the problem of classification using multiplex graphs, having significant implications for studying brain functional connectivity during cognitive control tasks in the elderly. Future research will expand this model to tackle more complex data structures and real-world application scenarios.