Abstract
Brain age is widely regarded as a powerful marker of general brain health. Brain age models are typically trained on large datasets to predict chronological age, which may offer advantages in predicting specific health outcomes, much like the success of finetuning large language models for specific applications. However, it is also well-accepted that machine learning models trained to directly predict specific outcomes (i.e., direct models) often outperform those trained on surrogate objectives. Therefore, despite their much larger training data, it is unclear whether brain age models outperform direct models in predicting specific brain health outcomes. Here, we compare large-scale brain age models (pretrained on 53,542 participants) and direct models for predicting specific health outcomes related to Alzheimer's Disease (AD) dementia. Using anatomical T1 scans from three continents (N = 1,848), we find that summarizing brain age with a single scalar (i.e., brain age gap) led to poor prediction performance. Using higher-dimensional intermediate representations of brain age models led to better prediction, but was still worse than direct models without finetuning. Using intermediate representations of finetuned brain age models was necessary to achieve similar performance as direct models. Overall, our results do not discount brain age as a useful marker of general brain health, but suggest that using chronological age as a pretraining target might be suboptimal for predicting specific health outcomes.
Introduction
In neuroimaging, "Brain Age" has emerged as a popular metric for assessing general brain health. By training machine learning models to predict chronological age from MRI scans, researchers calculate the "Brain Age Gap" (BAG)—the discrepancy between predicted and actual age. A "older" brain is typically interpreted as a sign of accelerated aging. However, a fundamental question remains: is it more effective to use a "foundation model" approach (pretraining on massive datasets for a general task like age prediction) or a "direct model" approach (training on smaller datasets for a specific clinical task)? This study investigates whether the massive data advantage of brain age models (trained on ~53,000 scans) translates into superior performance for Alzheimer's Disease (AD) prediction compared to models trained directly on disease labels.
Methods
The researchers utilized a 3D Convolutional Neural Network (SFCN) architecture to compare five distinct modeling approaches for AD classification and Mild Cognitive Impairment (MCI) progression prediction.
The models evaluated include:
- Direct Model: Trained from scratch to classify AD vs. Cognitively Normal (CN).
- BAG (Brain Age Gap): Uses a pretrained age-prediction model to generate a single scalar (the gap) as the predictor.
- BAG-finetune: The age-prediction model is tuned on local healthy controls before calculating the gap.
- Brainage64D: Extracts 64 high-dimensional features from the pretrained model for classification.
- Brainage64D-finetune: The high-dimensional model is fully finetuned on the AD classification task.
Evaluation was performed on three international datasets (ADNI, AIBL, and MACC; Total N = 1,848). While the brain age models were pretrained on 53,542 participants, the direct models were tested even in low-data regimes (N=50).
Results
The study revealed a significant performance gap between simple brain age metrics and direct classification models.
Key findings include:
- The Scalar Limitation: Summarizing a complex brain scan into a single scalar (BAG) resulted in poor diagnostic performance (AUC ~0.70).
- Task Alignment vs. Data Volume: The Direct Model performed comparably to the finetuned brain age model, even when trained on as few as 50 participants. This suggests that task alignment is often more critical than raw data volume.
- Performance Parity: The Direct Model and the Finetuned 64D model achieved the highest performance, with an AUC of approximately 0.92–0.93.
- MCI Progression: For predicting future decline in MCI patients, features from an AD-specific model were more effective than those from a general age-prediction model.
Discussion
The results suggest that while brain age is a useful general marker, chronological age may be a suboptimal surrogate for specific pathological outcomes like Alzheimer's. The features learned to predict chronological age do not perfectly overlap with those indicating neurodegeneration. Furthermore, using "off-the-shelf" brain age features is insufficient; finetuning on the target task is mandatory to match the performance of direct models.
Future work should explore pretraining on surrogate targets more closely aligned with brain health, such as cognitive scores or mortality, and investigate whether even larger "Foundation Models" might eventually provide a clear advantage over direct task-specific training.
Summary updated