Investigating systematic bias in brain age estimation with application to post-traumatic stress disorders

Area of Science:

• Neuroimaging
• Machine Learning
• Biomarker Development

Background:

• Brain age prediction is a growing field with potential for biomarkers in typical development and neuropsychiatric disorders.
• A persistent bias exists: predicted brain age is overestimated in younger individuals and underestimated in older ones.
• Existing explanations for this bias are varied, citing methodological and data-related factors.

Purpose of the Study:

• To investigate the origins of the brain age prediction bias.
• To propose and validate a statistical method for correcting this bias.
• To assess the bias adjustment's effectiveness in a large, multi-modal neuroimaging dataset.

Main Methods:

• Analysis of a large neuroanatomical dataset (N=2,026, ages 6-89) from multiple sources.
• Demonstration that the bias is independent of data source and specific machine learning methods, including deep neural networks.
• Application of a general linear model (GLM) for statistical bias adjustment.

Main Results:

• The brain age prediction bias is not dependent on the dataset or the specific machine learning method used.
• A simple GLM-based method effectively adjusts for the observed bias.
• Bias adjustment improved accuracy in a multi-modal neuroimaging dataset (N=804, ages 8-21) including healthy controls and PTSD patients.

Conclusions:

• The observed brain age prediction bias has a statistical origin, not solely data or method-dependent.
• A GLM-based approach provides an efficient and effective solution for correcting brain age prediction bias.
• This bias correction method has implications for improving the reliability of brain age as a biomarker.