Mind the Gap: Does Brain Age Improve Alzheimer's Disease Prediction?

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Brain age models predict chronological age and serve as a marker for brain health.
• Machine learning models trained directly on outcomes (direct models) often outperform those trained on surrogate objectives.
• The efficacy of brain age models versus direct models for specific health outcome prediction remains unclear.

Purpose of the Study:

• To compare the predictive performance of large-scale brain age models against direct models for Alzheimer's disease (AD) dementia prediction.
• To investigate the utility of different representations from brain age models (scalar vs. high-dimensional intermediate representations).
• To assess the impact of fine-tuning brain age models on specific health outcome prediction.

Main Methods:

• Utilized anatomical T1 scans from 1,848 participants across three continents.
• Compared pretrained large-scale brain age models (N=53,542) with direct prediction models.
• Evaluated performance using scalar brain age gap and higher-dimensional intermediate representations, with and without fine-tuning.

Main Results:

• A single scalar brain age gap showed poor predictive performance for AD dementia.
• Higher-dimensional intermediate representations from brain age models improved prediction but were still outperformed by direct models.
• Fine-tuning brain age models using intermediate representations was necessary to match the performance of direct models.

Conclusions:

• Brain age is a valuable marker for general brain health.
• Predicting specific health outcomes like AD dementia may be suboptimal using chronological age as a pretraining target for brain age models.
• Direct prediction models or fine-tuned brain age models are more effective for specific health outcome prediction.