Blockmodels for connectome analysis

Daniel Moyer, Boris A. Gutman, Gautam Prasad, Joshua Faskowitz, Greg Ver Steeg, Paul M. ThompsonProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2015 · DOI

TL;DR

This paper explores the use of stochastic blockmodels (SBMs) to represent human brain connectomes, demonstrating that these models can capture essential network structure for disease classification. They introduce a novel variant of the Mixed Membership Stochastic Blockmodel (MMSB) and show that SBM-based representations, while significantly reducing dimensionality, achieve comparable classification accuracy to raw data on two neuroimaging datasets.

What Problem Does This Paper Address?

Analyzing the human connectome, which represents brain connectivity as a network, is a complex task. Traditional methods often involve high-dimensional data, making it challenging to extract meaningful insights and build robust predictive models, especially for disease classification. The authors aim to address this by leveraging stochastic blockmodels, a class of generative network models, to create more efficient and interpretable representations of connectomes. They hypothesize that these models, by capturing underlying community structures, can effectively summarize the network's properties for downstream tasks like disease prediction.

Methods

The core of this paper lies in applying and extending stochastic blockmodels (SBMs) to brain connectome data. They utilize three main SBM variants:

  • Basic SBM: A foundational model that assumes nodes belong to distinct communities and edge probabilities are determined by community interactions.
  • Mixed Membership Stochastic Blockmodel (MMSB): A novel variant introduced by the authors. Unlike the basic SBM, MMSB allows nodes to have a distribution of memberships across multiple communities. This is achieved by assigning a probability distribution over communities for each node, and then sampling community assignments for each edge. The paper extends the original MMSB (defined for Bernoulli edges) to handle Poisson distributed edge counts, which is more appropriate for tractography data.
  • Degree Corrected Stochastic Blockmodel (DCSB): An extension that accounts for the tendency of nodes with higher degrees to form more connections, by introducing a node-specific parameter.

For model fitting, they employ a collapsed Gibbs Sampler, a common Bayesian approach that is efficient and extensible to multiple networks. To evaluate the effectiveness of these blockmodel representations, they use two classification strategies:

  1. Simple Bayesian Classifier: A generative classifier that directly uses the blockmodel's probability distribution to predict class labels.
  2. Linear SVM: A discriminative classifier trained on the empirical interaction matrices derived from the blockmodel's posterior distributions. This is compared against an SVM trained on the raw, flattened connectome matrices.

Datasets

The study utilizes two publicly available neuroimaging datasets:

  • ADNI Dataset: This dataset consists of diffusion MRI scans from 96 subjects, including 50 cognitively normal controls and 46 individuals with Alzheimer's Disease. The data was collected as part of the ADNI-2 project.
  • Bipolar Dataset: This dataset comprises data from 92 subjects involved in the Searching for Endophenotypes of Bipolar Disorders Study. The data was collected across three sites with varying scanner strengths.

Key Insights and Ideas

  • Dimensionality Reduction with Information Preservation: Blockmodels offer a powerful way to reduce the high dimensionality of connectome data while retaining crucial structural information. The empirical interaction matrices derived from blockmodels are essentially lower-rank approximations of the original adjacency matrices.
  • Community Structure as a Disease Biomarker: The paper provides evidence that the community structure captured by blockmodels is sensitive to neurological conditions. The Bayesian classifier, which directly leverages the blockmodel's generative assumptions, shows promising results, particularly for the ADNI dataset.
  • MMSB's Flexibility: The proposed MMSB variant, allowing for mixed memberships, offers greater flexibility in modeling complex brain networks where nodes might participate in multiple functional communities.
  • Qualitative Validation: The authors emphasize that their primary goal is not to achieve state-of-the-art classification accuracy but to validate blockmodels as sensitive network descriptors. The comparable performance of blockmodel-based classifiers to baseline methods serves as a qualitative endorsement of their representational power.

Key Results

  • ADNI Dataset: For the ADNI dataset, SVM classifiers trained on the blockmodel's empirical interaction matrices achieved accuracy comparable to a baseline SVM trained on the raw flattened connectome data. For instance, the MMSB-SVM with K=20 achieved an accuracy of 0.82, very close to the baseline SVM's 0.80. The Bayesian classifiers also performed reasonably well, with SBM-Bayes (K=6) achieving 0.71 accuracy.
  • Bipolar Dataset: On the Bipolar dataset, which is noted as a more challenging classification task, blockmodel-based SVMs also showed comparable performance to the baseline SVM, though overall accuracies were lower. For example, MMSB-SVM (K=16) achieved 0.61 accuracy, similar to the baseline SVM's 0.58. The Bayesian classifiers also showed competitive results, with SBM-Bayes (K=16) reaching 0.60 accuracy.
  • Bayesian Classifier Trend: A notable trend observed across both datasets is that the Bayesian classifier often performed best with fewer blocks (smaller K values), especially for the ADNI dataset.

Limitations and Future Work

  • Number of Blocks (K) Selection: The paper acknowledges that selecting the optimal number of blocks (K) remains an open problem. While information criteria can be used, their effectiveness with degree correction terms is unclear.
  • Interpretation of Communities: The authors suggest that further exploration of the discovered communities could provide significant scientific insight into neurological organization.
  • Extensibility to Directed Graphs: While the current work focuses on undirected graphs, the estimation method for MMSB is noted as being easily extensible to directed cases.
  • Larger Datasets: For larger datasets, the authors propose exploring block groupings as neurological regions to gain deeper insights.

Summary updated

Abstract

In the present work we study a family of generative network model and its applications for modeling the human connectome. We introduce a minor but novel variant of the Mixed Membership Stochastic Blockmodel and apply it and two other related model to two human connectome datasets (ADNI and a Bipolar Disorder dataset) with both control and diseased subjects. We further provide a simple generative classifier that, alongside more discriminating methods, provides evidence that blockmodels accurately summarize tractography count networks with respect to a disease classification task.
BibTeX
@inproceedings{moyer2015blockmodels,
  author = {Daniel Moyer and Boris A. Gutman and Gautam Prasad and Joshua Faskowitz and Greg Ver Steeg and Paul M. Thompson},
  title = {{Blockmodels for connectome analysis}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2015},
  doi = {10.1117/12.2211519},
}
Figure