A Skewed Loss Function for Correcting Predictive Bias in Brain Age Prediction.

Area of Science:

• Neuroimaging
• Biomarker Discovery
• Machine Learning in Medicine

Background:

• Brain age delta, the difference between predicted and chronological brain age, is a potential biomarker for pathological phenotypes.
• Regression models often exhibit bias in brain age delta estimation, overestimating for younger and underestimating for older individuals.
• This age-related bias complicates the analysis of brain age delta in relation to other age-associated variables.

Purpose of the Study:

• To propose a novel bias correction method for regression models used in brain age estimation.
• To introduce a skewed loss function to address the overestimation/underestimation bias in brain age delta.
• To develop a corrected brain age delta metric that is independent of chronological age and improves predictive accuracy.

Main Methods:

• A novel bias correction method employing a skewed loss function was developed to replace traditional symmetric loss functions in regression models.
• The method was validated using three deep learning architectures (ResNet, VGG, GoogleNet) on public neuroimaging aging datasets.
• The approach requires minimal preprocessing, applicable to various MR image types sensitive to age-related changes.

Main Results:

• The proposed skewed loss function effectively corrected the bias in brain age delta estimation.
• The corrected brain age delta demonstrated no linear correlation with chronological age.
• The method achieved higher predictive accuracy compared to a conventional two-stage bias correction approach.
• The approach showed flexibility across different deep learning models and hyperparameter settings.

Conclusions:

• The novel skewed loss function provides an effective bias correction for brain age delta estimation in neuroimaging.
• The corrected brain age delta offers a more reliable biomarker, free from chronological age-related bias.
• This method enhances the utility of brain age delta for studying age-associated neurological conditions and improving predictive accuracy.