Background: Investigation of structural connectivity in autism using diffusion MRI (dMRI) based structural connectomes has not produced consistent findings, primarily due to small sample sizes of individual studies and high heterogeneity in autism. To maximize information gleaned from existing studies, it is essential to integrate data retrospectively while overcoming extensive site-related differences. Solving this need requires specialized harmonization paradigms to address statistical complexity of structural connectomes while preserving biological variability. We anticipate that harmonized data will facilitate robust detection of autism-related patterns by increasing pooled sample size.

Objectives: To propose a harmonization paradigm for pooling structural connectomes from retrospective studies, and to investigate divergence in structural connectivity associated with autism in a large, pooled cohort.

Methods: We analyzed dMRI of 1503 participants (1194 neurotypicals and 309 autism, aged 6-21 years) from 6 sites (Table1). Structural connectomes were created by parcellating the brain into 86 regions defined by Desikan-Killiany atlas and performing probabilistic tractography, with streamline counts used as connectivity weights, yielding 86x86 connectivity matrices. Since connectivity weights were non-gaussian distributed, we modeled edgewise weights using a gamma-distributed generalized linear model (gamma-GLM) with a log link, incorporating site, sex, age, and age² as covariates. Site-related coefficients were estimated edgewise from neurotypicals alone and applied to harmonize structural connectomes for autism and neurotypical groups. Removal of site effects was validated using ANOVA tests on harmonized edgewise weights and 6 graph topological measures, controlling for sex, age, and age². We assessed group-level differences and group-by-age interactions in graph measures in harmonized data between autistic and neurotypicals. Only males (n=890) were included in this analysis. We further tested Spearman correlations between graph measures and autism clinical assessments, controlling for age. We compared harmonized structural connectomes to unharmonized data where site covariate was regressed out at the level of graph measures.

Results: Among 1480 edges after consistency-based thresholding, 1377 (93%) exhibited significant site effects pre-harmonization (p<0.05). Site effects at all edges were non-significant following gamma-GLM harmonization (Fig.1A). Additionally, all derived graph measures showed significant site effects before harmonization (p<0.05), while none were significant after harmonization (Fig.1B).

Unharmonized data revealed no group-level differences in graph measures. A weak group-by-age interaction in clustering coefficients (standardized β=0.16, p=0.043) and a marginally significant interaction for global efficiency (standardized β=0.16, p=0.052) were noted (Fig.1C top), but no measures were linked to clinical scores.

Post harmonization, males with autism showed significantly higher intra-hemispheric strength compared to neurotypicals (Cohen's d=0.37, p=0.041). Significant group-by-age interactions were observed for clustering coefficient (standardized β=0.27, p<0.001) and global efficiency (standardized β=0.26, p=0.002), with both measures increasing with age in autism but decreasing in neurotypicals (Fig.1C bottom). Intra-hemispheric strength positively correlated with SRS-2 t-score (R=0.12, p=0.012), while clustering coefficient (R=0.15, p=0.006) and global efficiency (R=0.14, p=0.013) correlated positively with VABS communication subscale.

Conclusions: Our gamma-GLM harmonization can effectively remove site-related differences in structural connectomes for retrospective data integration. With gamma-GLM harmonized structural connectomes from 6 sites, we observed hyper intra-hemispheric connectivity and divergent age-related patterns in clustering coefficient and global efficiency in autistic children compared to neurotypicals.