Toward Robust Neuroanatomical Normative Models: Influence of Sample Size and Covariates Distributions.

Area of Science:

• Neuroimaging
• Computational neuroscience
• Biostatistics

Background:

• Normative models in neuroimaging are vital for detecting individual brain deviations.
• Model performance is sensitive to the reference sample's size and demographics.
• Understanding these influences is key for reliable Alzheimer's disease (AD) research.

Purpose of the Study:

• To investigate the impact of reference sample size and covariate composition on normative model performance in AD.
• To evaluate the effectiveness of adaptive transfer learning for improving model accuracy.
• To determine optimal cohort characteristics for robust individual-level brain deviation estimation.

Main Methods:

• Trained normative models on subsamples of healthy controls (HCs) with varying sizes (5-600) and demographics.
• Assessed model fit, deviation estimates, and clinical readouts on test sets and AD cohorts.
• Utilized adaptive transfer learning by pre-training on large-scale data (UK Biobank) and adapting to clinical datasets.
• Validated findings in an independent external sample (AIBL).

Main Results:

• Model performance consistently improved with larger reference sample sizes.
• Demographic matching, especially for age, was critical for accurate deviation estimates.
• Directly trained models stabilized around 200 HCs; adapted models achieved similar performance with as few as 50 HCs.
• Transfer learning significantly reduced the required sample size for reliable modeling.

Conclusions:

• Robust individual-level brain deviation modeling is achievable with moderately sized, demographically matched cohorts.
• Adaptive transfer learning enhances efficiency, enabling reliable estimates with smaller reference groups.
• Findings support the broader application of normative modeling in aging and neurodegeneration research.