Brain-age models with lower age prediction accuracy have higher sensitivity for disease detection

Area of Science:

• Neuroimaging
• Machine Learning
• Psychiatry

Background:

• Brain-age models estimate brain age from MRI scans.
• Current models prioritize chronological age prediction accuracy.
• This accuracy focus may not align with detecting brain disorders.

Purpose of the Study:

• Reevaluate brain-age models for neurological and psychiatric disorder detection.
• Challenge the emphasis on chronological age prediction accuracy.
• Investigate simpler models for disease-relevant changes.

Main Methods:

• Analyzed T1 MRI data from 46,381 UK Biobank participants.
• Compared simpler and complex machine learning models.
• Assessed model sensitivity to disease-relevant changes versus age prediction accuracy.

Main Results:

• Simpler models, especially with regularization, showed higher sensitivity to disease.
• These models were less accurate in chronological age prediction.
• Accuracy-optimized models may miss disease-informative variations.

Conclusions:

• Models optimized for age accuracy are not ideal normative brain aging models.
• Simpler models offer better brain health biomarkers by capturing disease variations.
• Focusing on models with larger patient-control effect sizes aids early disease detection.