Bias-adjustment in neuroimaging-based brain age frameworks: A robust scheme.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Brain age estimation frameworks are crucial for understanding brain health and aging.
• Prediction errors in current brain age models can affect the reliability of statistical inference.
• Existing methods often struggle with bias, impacting clinical applicability.

Purpose of the Study:

• To present an effective and straightforward bias-adjustment scheme for brain age estimation.
• To reduce prediction bias in machine learning-based brain age frameworks.
• To enhance the accuracy and clinical utility of brain age estimation.

Main Methods:

• Developed a bias-adjustment scheme using chronological age as a covariate in the training set.
• Applied the scheme to a machine learning-based brain age framework using fluorodeoxyglucose positron emission tomography (FDG-PET) data from 675 cognitively unimpaired adults.
• Validated the method on independent test sets including cognitively unimpaired adults, mild cognitive impairment (MCI) patients, and Alzheimer's disease (AD) patients.

Main Results:

• Achieved a high R² of 0.81 and a low mean absolute error (MAE) of 2.66 years in cognitively unimpaired adults using the bias-adjusted method.
• Without bias-adjustment, the R² was 0.24 and MAE was 4.71 years on the same independent set.
• Demonstrated significant reduction in prediction error, particularly for clinical populations.

Conclusions:

• The proposed bias-adjustment scheme effectively diminishes prediction error in brain age estimation.
• This method enhances the robustness and accuracy of brain age estimation frameworks for clinical applications.
• Improved brain age estimation holds potential for better diagnosis and monitoring of neurological conditions.