Warped Bayesian linear regression for normative modelling of big data.

Area of Science:

• Neuroimaging
• Biostatistics
• Computational Neuroscience

Background:

• Normative modeling in neuroimaging predicts individual deviations from typical development.
• Current methods like Gaussian process regression struggle to scale with large neuroimaging datasets.
• Existing models often assume Gaussian distributions, leading to inaccuracies, especially in outer centiles.

Purpose of the Study:

• To develop a scalable normative modeling framework for large neuroimaging cohorts.
• To address the challenge of non-Gaussian predictive distributions in neuroimaging data.
• To improve the accuracy of estimating individual deviations from normative trajectories.

Main Methods:

• Introduced a novel framework using Bayesian linear regression with likelihood warping.
• Applied the warped Bayesian linear regression (BLR) to neuroimaging-derived measures from the UK Biobank.
• Evaluated computational scalability and accuracy for image-derived phenotypes and diffusion tensor imaging measures.

Main Results:

• Demonstrated good computational scaling of the warped BLR method.
• Showed improved accuracy for certain image-derived phenotypes and voxel-wise measures with non-normal residuals.
• Confirmed the framework's ability to model non-Gaussian distributions effectively.

Conclusions:

• The warped Bayesian linear regression offers enhanced computational scalability for neuroimaging normative modeling.
• This framework flexibly incorporates non-linearity and non-Gaussianity, expanding applicability to diverse neuroimaging datasets.
• The method improves the accuracy of normative modeling, particularly for estimating deviations in sparse outer centiles.