Harmonization of Structural Brain Networks for Multi-site Studies

Research on structural networks often suffers from limited sample sizes and inherent selection biases in individual studies, which restrict their ability to address complex questions regarding human brain organization. Pooling data across studies is crucial for achieving a more comprehensive representation of the population and effectively managing individual heterogeneity; however, structural networks acquired from multiple sites are susceptible to significant site-related differences. This necessitates harmonization to mitigate biases and reveal true biological variability in multi-site analyses. Our work marks the first effort to develop and evaluate harmonization frameworks specifically for structural networks. We adapt several statistical approaches for harmonizing structural networks and provide a comprehensive evaluation to rigorously test their effectiveness. Our findings demonstrate that the adaptation of the Gamma Generalized Linear Model (gamma-GLM) outperforms other methods in modeling structural network data, effectively eliminating site-related effects in structural connectivity matrices and downstream graph-based analyses while preserving biological variability. Additionally, we highlight gamma-GLM’s superiority in addressing confounding factors between site and age. Two practical applications further illustrate the utility of our harmonization framework in tackling common challenges in multi-site structural network studies. Specifically, harmonization using gamma-GLM enhances the transferability of brain age predictors to new datasets and facilitates data integration in patient studies, ultimately improving statistical power. Together, our work offers essential guidelines for harmonizing and integrating multi-site structural network studies, paving the way for more robust discoveries through collaborative research in the era of big data.