Reconsidering Brain Age: Why Age-Prediction Models Fail as Measures of Brain Aging

This paper argues that current brain age models are fundamentally flawed as biomarkers for accelerated aging because they are trained to prioritize shared chronological patterns while ignoring individual trajectories, leading to misleading conclusions that stable anatomical differences are signs of neurodegeneration.

The Gist

The Big Idea: The "Brain Age" Clock is Broken

Imagine you have a machine that looks at a photo of your brain and guesses your age. If it guesses you are 60 but you are actually 50, people say, "Wow, your brain is aging fast!" If it guesses 40, they say, "Great, your brain is young!"

This paper argues that this machine is fundamentally broken. It isn't actually measuring how fast your brain is aging; it is mostly measuring how big your brain was when you were born.

The authors say that using this "Brain Age Gap" to tell if someone is aging too fast is like trying to measure how fast a car is speeding by looking at how big the car is. It just doesn't work.

The Core Problem: The "Average" Trap

To understand why the machine is broken, imagine a classroom of students taking a test every year.

1. The Goal: The teacher wants to build a computer program that can look at a student's test score and guess their grade level (e.g., 5th grade, 6th grade).
2. How the Computer Learns: To be good at guessing the grade, the computer looks for patterns that are the same for everyone. It learns that "5th graders usually get 80% right" and "6th graders usually get 90% right."
3. The Mistake: The computer is trained to ignore the students who are different.
   • If a student is naturally very smart and always scores high (a stable trait), the computer thinks, "Ah, this student must be in a higher grade!"
   • If a student is naturally less skilled and always scores low, the computer thinks, "This student must be in a lower grade!"

The Catch: The computer is so focused on guessing the average grade that it ignores how much a student is actually improving or slowing down over time. It confuses being naturally big with growing fast.

The Two Experiments: Proving the Machine is Lying

The authors tested this idea with two real-world scenarios to show how the "Brain Age" machine fails.

1. The "False Alarm" Test (Birth Weight)

• The Setup: They looked at people's birth weight. We know from science that babies born heavier tend to have bigger brains their whole lives. However, we also know that birth weight does not change how fast a brain shrinks or ages later in life. A heavy baby doesn't age faster or slower than a light baby; they just start with a bigger brain.
• The Result: The "Brain Age" machine looked at the heavy babies and said, "Your brain is older than it should be!" It thought the big brain meant "accelerated aging."
• The Reality: The machine was just confused. It saw a big brain (a stable trait) and mistakenly called it "fast aging." This is a False Positive.

2. The "Missed Signal" Test (Tau Proteins)

• The Setup: They looked at Tau proteins, which are a sign of Alzheimer's disease. When these proteins build up, they cause the brain to actually shrink and change rapidly over time. This is real, active "aging" or damage.
• The Result: The "Brain Age" machine barely noticed this. It was not very good at detecting that these people were losing brain tissue.
• The Reality: Because the machine was trained to ignore "differences between people" (to get a perfect age guess), it actually ignored the very people who were changing the most. It missed the real danger. This is a False Negative.

The "Reliability Paradox"

The paper points out a funny irony: The better the machine gets at guessing your age, the worse it gets at telling you if you are aging strangely.

• If the machine is perfect at guessing age, it has learned to ignore all the "noise" (the differences between people).
• But the "noise" is exactly where the real story of aging is hiding.
• So, a "perfect" brain age model is actually blind to the individual differences it is supposed to find.

The Conclusion: Stop Using the "Brain Age" Score

The authors conclude that we cannot use the "Brain Age Gap" (the difference between predicted age and real age) as a measure of brain health or aging speed.

• What it actually measures: It mostly measures stable, lifelong differences in brain size and shape (like how tall you are).
• What it fails to measure: It fails to measure the actual rate at which your brain is changing or deteriorating.

The Solution: Instead of asking, "How old does this brain look?", we should ask, "How much has this brain changed compared to where it started?" The paper suggests we need new models that look for change, not just a static guess of age. Until then, the "Brain Age" number is misleading and should not be used to diagnose accelerated aging or brain health issues.

Technical Summary

Technical Summary: Reconsidering Brain Age: Why Age-Prediction Models Fail as Measures of Brain Aging

Problem Statement
Brain age models, which use machine learning to predict chronological age from neuroimaging data, are widely interpreted as biomarkers for accelerated or decelerated brain aging. The "brain age gap" (the difference between predicted and chronological age) is commonly used to infer individual differences in aging rates across various domains, including psychiatric disorders, genetic risks, and lifestyle factors. However, this interpretation relies on the unexamined assumption that models trained to predict chronological age are sensitive to individual differences in the rates of brain change. The authors argue that this assumption is fundamentally flawed: the learning objective of these models (minimizing error in predicting chronological age) systematically prioritizes features that change consistently across individuals while suppressing features that capture individual variability in trajectories. Consequently, brain age models may misclassify stable, lifelong anatomical differences as "accelerated aging" while failing to detect genuine longitudinal neurodegenerative changes.

Methodology
The study employs a three-pronged approach combining theoretical analysis, controlled simulations, and empirical validation using longitudinal neuroimaging data.

1. Theoretical Analysis: The authors analyze the mathematical properties of brain age models (both linear and optimal non-linear). They decompose brain feature variance into three components: measurement noise, baseline differences (level effects), and individual differences in change (slope effects). They demonstrate that to minimize prediction error for chronological age, models must assign high weights to features with low inter-individual variance and high age-dependency. Conversely, features with high inter-individual variance in change (the very signal required to measure aging rates) are down-weighted because they introduce "noise" relative to the target of chronological age.
2. Simulation Experiments: A synthetic dataset was generated with two features: one with a uniform aging rate across all individuals (F1) and one with accumulating individual differences in aging rates (F2). Models (Linear, Optimal Non-linear, and XGBoost) were trained to predict age. The study quantified how much of the individual trajectory variance (slope effects) was captured by the model outputs versus how much was suppressed.
3. Empirical Validation (Double Dissociation): The authors tested two complementary failure scenarios using data from multiple cohorts (ADNI, DLBS, HABS, PreventAD, LCBC, and UK Biobank, totaling over 42,000 subjects for training and various subsets for testing):
   • False-Positive Test (Level Effects): They examined the association between birth weight (a fixed, lifelong trait) and brain age gap. Since birth weight correlates with brain volume but not with the rate of volume change, any association with brain age gap would represent a misclassification of level effects as aging.
   • False-Negative Test (Slope Effects): They examined the association between tau accumulation (measured via PET, a biomarker of neurodegeneration characterized by longitudinal change) and brain age gap. They compared the sensitivity of brain age models against simple volumetric measures (hippocampal and total brain volume) in predicting tau positivity.

Key Contributions and Results

• Theoretical Suppression of Aging Signals: The analysis confirms that brain age models are optimized to ignore the signal of individual differences in aging. In simulations, as individual differences in change (slope effects) increased the variance of a feature, the model's reliance on that feature dropped to near zero. The model effectively "flattens" individual trajectories to fit the mean age trajectory.
• False Positives (Birth Weight): In the empirical analysis, both Elastic Net and XGBoost brain age models showed a significant association between lower birth weight and "accelerated brain aging" (higher brain age gap). However, longitudinal analyses revealed no significant relationship between birth weight and the actual rate of brain volume change. This confirms that the brain age gap was misclassifying stable, pre-existing anatomical differences (level effects) as differences in aging rates.
• False Negatives (Tau Pathology): Conversely, brain age models showed low sensitivity to tau-related neurodegeneration. While simple measures of longitudinal hippocampal and total brain volume change showed strong associations with tau positivity (Odds Ratios ~2.4–2.9), the brain age gap showed significantly weaker associations (Odds Ratios ~1.2). Even raw, cross-sectional volumes were more sensitive to tau than the brain age models. This demonstrates that the models fail to capture the very biological changes they are often used to detect.
• The Reliability Paradox: The authors note a paradox where more accurate chronological age predictors may be less sensitive to individual differences in aging, as higher accuracy requires stricter suppression of inter-individual variance.

Significance and Claims
The paper claims that the "brain age gap" cannot be interpreted as a measure of individual differences in brain aging or brain health without additional longitudinal evidence. The authors argue that:

1. Misinterpretation Risk: Conventional brain age models risk converting stable between-person anatomical differences (e.g., due to genetics or prenatal factors) into apparent accelerated aging, leading to false-positive inferences in conditions like schizophrenia or developmental disorders.
2. Insensitivity to Pathology: These models are inherently less sensitive to genuine, clinically meaningful neurodegenerative changes (like tau-driven atrophy) than simple volumetric measures because the models penalize the high-variance features that characterize such changes.
3. Reorientation of the Field: The authors propose that the field must reorient its prediction targets. Instead of predicting chronological age, models should be optimized to predict future change directly from baseline features or to capture inter-individual differences in aging-related change (e.g., via canonical correlation analysis or delta expansion). They suggest that cross-sectional data can still inform aging research, but only if the models are explicitly designed to capture aging-relevant variation rather than shared age-related patterns.

The authors conclude that the current utility of brain age models is limited to summarizing anatomical variation, not measuring aging dynamics, and caution against their use as standalone proxies for accelerated aging or brain health.