Multi-Link Analysis: Brain Network Comparison via Sparse Connectivity Analysis

Alessandro Crimi, Luca Giancardo, Fabio Sambataro, Alessandro Gozzi, Vittorio Murino, Diego Sona, the Alzheimer’s Disease Neuroimaging InitiativebioRxiv 2018 · DOI

TL;DR

This paper introduces "Multi-Link Analysis" (MLA), a novel sparse multivariate approach for identifying brain sub-networks that differ between two groups (e.g., healthy vs. diseased). MLA excels at discovering subtle, localized connectivity patterns that might be missed by traditional univariate methods, offering a powerful tool for hypothesis generation in connectomics research.

What Problem Does This Paper Address?

Understanding the brain's complex network of connections (connectome) is crucial for identifying biomarkers of disease. However, analyzing this high-dimensional data is challenging. Traditional methods often rely on univariate statistical tests, which can miss multivariate relationships between connections. Furthermore, purely exploratory data analysis can be hindered by a lack of prior knowledge, making it difficult to pinpoint specific network differences. Existing machine learning approaches, while improving classification accuracy, don't always guarantee interpretable results regarding why groups differ. This paper aims to bridge this gap by providing a method that not only identifies differences but also highlights specific, interpretable sub-networks.

Methods

The core of the proposed method, Multi-Link Analysis (MLA), is a two-stage feature selection process based on sparse discriminant analysis (SDA).

  1. Ensemble of Sparse Classifiers: MLA trains multiple sparse linear discriminant models on bootstrapped subsamples of the data. This ensemble approach helps to identify connections that are consistently important for discriminating between the two groups.
  2. Stability Selection: To ensure robustness and overcome the inherent instability of some sparse methods, MLA employs a stability selection criterion. Only connections that are frequently selected across the ensemble of models are retained. This process effectively filters out connections that are only relevant in specific subsamples, leading to a more reliable set of discriminative features.

The method is designed to work with weighted networks (representing connectivity strength) and does not require prior anatomical parcellation, making it flexible for various datasets. It leverages sparsity through L1 regularization to select a parsimonious set of features, aiming for interpretability rather than solely maximizing classification accuracy.

Datasets

The authors evaluated MLA on three distinct datasets:

  • Mice Structural Connectivity Data: A dataset of structural connectomes from BTBR mice (a model for autism) and control C57BL/6J mice. This dataset was used to validate the method's ability to identify known anatomical differences, specifically the absence of the corpus callosum in BTBR mice.
  • Human Alzheimer's Disease Structural Connectivity Data: A publicly available dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI), comparing patients with Alzheimer's disease to healthy elderly controls. This dataset was used to assess MLA's performance on human structural connectomes and compare it with existing methods.
  • Human ADHD Functional Connectivity Data: A publicly available resting-state fMRI dataset comparing children with ADHD to typically developing controls. This dataset was used to demonstrate MLA's utility in identifying functional connectivity differences, particularly in cases where univariate methods struggle.

Key Insights and Ideas

  • Multivariate Relationships Matter: The paper strongly advocates for multivariate approaches, showing that connections often work in concert to define group differences, and analyzing them individually can lead to missed insights.
  • Interpretability Over Pure Accuracy: MLA prioritizes identifying interpretable sub-networks that explain group differences, even if it means slightly lower classification accuracy compared to methods that might find complex, less understandable patterns.
  • Sparsity as a Discovery Tool: By enforcing sparsity, MLA effectively prunes away less relevant connections, highlighting the most discriminative ones and acting as a powerful data-driven hypothesis generator for neuroscientists.
  • Robustness Through Ensemble and Stability: The combination of ensemble learning and stability selection makes MLA more robust to noise and variations in the data, leading to more reliable identification of discriminative features.
  • Flexibility in Data and Parcellation: The method's ability to handle both structural and functional data, and its independence from pre-defined anatomical atlases, makes it broadly applicable.

Key Results

  • Mice Data: MLA successfully identified structural differences in acallosal BTBR mice, including the lack of corpus callosum, aligning with known literature and demonstrating the method's ability to detect expected anatomical alterations.
  • Alzheimer's Data: MLA detected a higher number of discriminative connections compared to NBS and SVM, achieving a classification accuracy of 75%. The identified connections were largely localized to temporal and para-hippocampal regions, consistent with known Alzheimer's pathology.
  • ADHD Data: MLA identified 8 discriminative connections with 70% accuracy, including those involving frontal pole and cingulate gyrus, which are key hubs of the default mode network (DMN) known to be altered in ADHD. Notably, the NBS method failed to find any significantly discriminative connections in this dataset.
  • Computational Efficiency: The analysis of large datasets was relatively fast, with the mice experiment taking less than 40 minutes and the human Alzheimer's dataset analysis taking less than 1 second on a standard machine.

Limitations and Future Work

  • Parameter Tuning: While the method is generally robust, the sparsity parameter ($\alpha$) can be adjusted to control the number of detected features, allowing users to balance sensitivity and specificity. The authors suggest that this parameter can be manually adjusted by neuroscientists to guide the discovery process.
  • "Half the Number of Subjects" Threshold: The paper mentions using "half the number of subjects" as a threshold for feature selection stability. While this is a common practice, the optimal threshold might vary across datasets, and further investigation into its impact could be beneficial.
  • Exploratory Framework: The authors emphasize that MLA is primarily an exploratory tool for generating hypotheses. While it can achieve good classification accuracy, its main strength lies in its ability to reveal underlying network differences for further investigation.

Future work could involve further exploring the impact of different parcellation strategies (or lack thereof) on MLA's performance and extending the method to more complex group comparisons (e.g., more than two groups).

Summary updated

Abstract

ABSTRACT The analysis of the brain from a connectivity perspective is unveiling novel insights into brain structure and function. Discovery is, however, hindered by the lack of prior knowledge used to make hypotheses. On the other hand, exploratory data analysis is made complex by the high dimensionality of data. Indeed, in order to assess the effect of pathological states on brain networks, neuroscientists are often required to evaluate experimental effects in case-control studies, with hundreds of thousand connections. In this paper, we propose an approach to identify the multivariate relationships in brain connections that characterise two distinct groups, hence permitting the investigators to immediately discover sub-networks that contain information about the differences between experimental groups. In particular, we are interested in data discovery related to connectomics, where the connections that characterize differences between two groups of subjects are found. Nevertheless, those connections not necessarily maximize accuracy in classification since this does not guarantee reliable interpretation of specific differences between groups. In practice, our method exploits recent machine learning techniques employing sparsity to deal with weighted networks describing the whole-brain macro connectivity. We evaluated our technique on functional and structural connectomes from human and mice brain data. In our experiments, we automatically identified disease-relevant connections in datasets with supervised and unsupervised anatomy-driven parcellation approaches, and by using high-dimensional datasets.
BibTeX
@misc{crimi2018multi,
  author = {Alessandro Crimi and Luca Giancardo and Fabio Sambataro and Alessandro Gozzi and Vittorio Murino and Diego Sona and the Alzheimer’s Disease Neuroimaging Initiative},
  title = {{Multi-Link Analysis: Brain Network Comparison via Sparse Connectivity Analysis}},
  year = {2018},
  doi = {10.1101/277046},
  howpublished = {bioRxiv},
}
Figure