Unified Multi-Cohort Harmonisation and Normative Modelling of Neuroimaging Data via Hierarchical GAMLSS

Area of Science:

• Neuroimaging
• Biostatistics
• Data Science

Background:

• Large-scale neuroimaging studies pool data from multiple sources, leading to technical variations.
• Existing harmonization methods like ComBat have limitations, including Gaussian assumptions and focus on mean/variance correction.
• Addressing these technical variations is crucial for accurate biological inference in neuroimaging.

Purpose of the Study:

• To introduce a unified hierarchical Generalised Additive Models for Location, Scale and Shape (GAMLSS) framework for neuroimaging data harmonization and normative modeling.
• To model cohort effects across all distributional parameters, accommodating various parametric families.
• To enable joint harmonization and normative inference, providing harmonized values on the original scale.

Main Methods:

• Developed a hierarchical GAMLSS framework to model cohort effects in location, scale, and shape parameters.
• Utilized centile-based quantile mapping for harmonized values on the original measurement scale.
• Evaluated the framework on a large longitudinal dataset (88,126 observations, 237 features) from six cohorts.
• Compared GAMLSS with ComBat, ComBat-GAM, and ComBat-LS using criteria like data retention and biological signal preservation.

Main Results:

• GAMLSS achieved near-complete removal of residual cohort effects.
• The framework retained almost all valid observations post-harmonization.
• GAMLSS demonstrated superior preservation of age- and sex-related biological signal compared to existing methods.
• It better preserved age trajectories for complex features like white matter hypointensity volume.

Conclusions:

• Hierarchical GAMLSS offers a flexible and practical alternative for large-scale neuroimaging harmonization.
• The framework excels with non-Gaussian residual distributions and complex cohort effects.
• GAMLSS provides a unified approach for harmonization and normative modeling, enhancing biological inference.