Community detection in the human connectome: Method types, differences and their impact on inference

Area of Science:

• Neuroscience
• Network Science
• Computational Biology

Background:

• Brain network community structure is key to organization, but no standard method exists for its detection.
• The impact of community detection method choice on accuracy, robustness, and biological correlations is poorly understood.

Purpose of the Study:

• To compare three classes of community detection methods: modularity maximization, probabilistic (Bayesian inference within SBM), and geometric.
• To investigate how brain network topology influences method accuracy, reliability, and agreement.
• To assess method-dependent associations between network communities and individual measures.

Main Methods:

• Analysis of large-scale resting-state fMRI data from the Adolescent Brain Cognitive Development (ABCD) study ($$ =\u20095251).
• Analysis of synthetic networks ($$ =\u20095338) with varied topologies.
• Comparison of Newman, Louvain, Bayesian inference within SBM, and graph Ricci flow methods.

Main Results:

• Method accuracy and reliability depend heavily on brain network topology.
• In networks with clear community structure, most methods performed similarly.
• In complex networks, Bayesian inference within SBM showed superior accuracy.
• Some method-dependent associations were found between network communities and individual measures.

Conclusions:

• Probabilistic methods, particularly Bayesian inference within SBM, provide reliable community structure estimates across network topologies.
• Method dependence in findings highlights potential issues in reliability and reproducibility.
• Confirming network communities and their correlates with multiple detection methods is recommended for robust biological inferences.