BrainAGE: Revisited and reframed machine learning workflow.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomarkers of Brain Health

Background:

• Brain age prediction from magnetic resonance images is a valuable tool for assessing brain health.
• Previous methods like BrainAGE using relevance vector regression (RVR) have limitations with large datasets.
• The development of advanced machine learning techniques is crucial for improving brain age estimation.

Purpose of the Study:

• To revise and improve the BrainAGE method for more stable and accurate brain age prediction.
• To extend the global approach to a regional brain age estimation for spatially specific insights.
• To validate the performance of the new algorithms on diverse datasets, including clinical samples.

Main Methods:

• Replaced relevance vector regression (RVR) with Gaussian process regression (GPR) in the BrainAGE approach.
• Developed a regional BrainAGE model to provide brain age scores for specific brain lobes.
• Validated the models on the UK Biobank (UKB), ADNI, and schizophrenia datasets, as well as synthetic data.

Main Results:

• The reframed global BrainAGE model achieved a mean absolute error (MAE) below 2 years on the UKB sample.
• Significant differences in brain age were observed between healthy individuals and patients with Alzheimer's disease and schizophrenia.
• The regional BrainAGE model demonstrated disease-specific patterns in clinical samples, indicating its potential for diagnostic applications.

Conclusions:

• The improved BrainAGE algorithms, utilizing GPR, offer reliable and valid brain age estimations.
• Regional brain age analysis provides valuable, spatially specific information for understanding brain health and disease.
• These advancements hold promise for the use of brain age as a biomarker in clinical neuroscience.