Comparative Evaluation of Assumption Lean Community Detection Methods for Human Connectome Networks

Area of Science:

• Neuroscience
• Network Science
• Computational Biology

Background:

• Community detection in functional brain networks is crucial for understanding mesoscale organization.
• Existing methods often rely on assumptions like assortative structure and arbitrary thresholding, complicating community count selection (K).
• There is a need for assumption-lean, threshold-free methods for robust community detection in brain networks.

Purpose of the Study:

• To systematically benchmark assumption-lean, threshold-free methods for community detection in weighted functional brain connectivity matrices.
• To compare different strategies for selecting the number of communities (K), including post hoc indices and a novel likelihood-based criterion.
• To assess the developmental trajectory of mesoscale brain architecture using these methods.

Main Methods:

• Benchmarked Weighted Stochastic Block Model, Spectral Clustering, and K-means on synthetic and neuroimaging datasets (Human Connectome Project, Washington University 120, Baby Connectome Project).
• Compared various post hoc indices (e.g., silhouette, modularity, NMI) against a likelihood-based criterion for selecting K.
• Employed a consensus relabeling scheme (Hungarian matching with Hamming distance) to stabilize solutions.

Main Results:

• In simulations, all methods recovered stable partitions, but post hoc indices failed under weak signals and nonassortative mixing.
• The likelihood-based criterion identified a parsimonious K=11 in adults, aligning with known brain systems.
• In infants, the same criterion suggested K≈15, revealing distinct developmental mesoscale architecture, including subdivisions in default mode and frontoparietal systems.

Conclusions:

• Threshold-free weighted methods effectively mitigate assortative bias in brain network community detection.
• The likelihood-based criterion offers a reproducible and reliable approach for selecting the number of communities (K).
• This framework reveals significant developmental differences in the mesoscale architecture of the human brain.