Accommodating site variation in neuroimaging data using normative and hierarchical Bayesian models.

Area of Science:

• Neuroimaging
• Statistical modeling
• Biostatistics

Background:

• Normative modeling enables individualized predictions from neuroimaging data, surpassing case-control methods.
• Site effects in multi-site neuroimaging datasets can confound analyses and bias normative model estimates.
• Existing harmonization techniques may not adequately preserve biological variation or clinical utility.

Purpose of the Study:

• To propose and evaluate a hierarchical Bayesian model for accommodating site effects in neuroimaging data.
• To compare the performance of linear and non-linear hierarchical Bayesian models against traditional harmonization methods.
• To assess the models' ability to retain clinically relevant information while mitigating site-specific biases.

Main Methods:

• A hierarchical Bayesian model incorporating site effects as random effects was developed.
• Linear and non-linear models were applied to cortical thickness data from 570 healthy individuals (ABIDE dataset).
• Performance was compared against regression, ComBat (linear and non-linear), and raw data predictions.

Main Results:

• The proposed hierarchical Bayesian method demonstrated superior predictive performance across multiple metrics.
• Resulting z-scores exhibited minimal residual site effects and retained clinically useful information.
• Two-stage harmonization models, particularly those not preserving age/sex variance, showed poor performance and significant variance loss (>90%).

Conclusions:

• Hierarchical Bayesian regression offers a valuable alternative to harmonization techniques for managing site variation in neuroimaging.
• This approach is particularly advantageous for normative modeling, ensuring accurate inter-subject variation modeling and deviation quantification.
• The method effectively removes site effects while preserving clinically relevant information for individualized predictions.