💻 bioinformatics

Trustworthy ML/AI for Aging Clocks: Preventing Systematic Prediction Bias in Biological Age Estimation

This paper identifies and addresses the critical issue of systematic prediction bias in machine learning-based aging clocks, which can distort downstream association analyses, by proposing a constrained optimization framework to ensure valid biological age estimation and inference.

Original authors: Lee, H., Ye, Z., Yang, Y., Pan, Y., Maron, B., Wang, Z., Kochunov, P., Thompson, P., Hong, L. E., MA, T., Chen, C., Chen, S.

Published 2026-09-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lee, H., Ye, Z., Yang, Y., Pan, Y., Maron, B., Wang, Z., Kochunov, P., Thompson, P., Hong, L. E., MA, T., Chen, C., Chen, S.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine trying to measure how fast a car is aging by looking at its engine, its paint, and the wear on its tires, rather than simply checking the odometer. This is the essence of biological age research. While we all know our chronological age—the number of years since we were born—our bodies do not always age at the same pace. Some people remain physically robust well into their seventies, while others show signs of decline much earlier. Scientists have developed tools called "aging clocks" to estimate this biological age. These tools analyze complex data from our DNA, blood, or brain scans to predict how old we are on the inside. The goal is to identify who is aging too fast and why, so doctors can intervene to prevent disease. However, a new study suggests that the very computers used to build these clocks might be tricking us, leading to conclusions that are not just wrong, but sometimes the exact opposite of the truth.

The problem lies in how these computer programs, known as machine learning models, are trained. To teach a computer to guess biological age, researchers feed it data from thousands of people, using their known chronological age as the answer key. The computer learns to spot patterns in the biological data that match the years on a calendar. But these programs have a hidden flaw: they are naturally biased toward the average. When a model is unsure, it tends to guess the middle ground. For a young person, this means the computer guesses they are older than they really are. For an elderly person, it guesses they are younger. This is not a random mistake; it is a systematic error that shrinks the extremes toward the center. In the world of aging research, this creates a distorted picture where the oldest people appear to be biologically younger than they are, and the youngest appear older.

Researchers Hwiyoung Lee and Shuo Chen, along with their team at the University of Maryland and other institutions, discovered that this bias does more than just miscalculate a number. It fundamentally breaks the science that follows. When scientists use these flawed clocks to study how lifestyle factors like smoking or diet affect aging, the bias can flip the results entirely. In one real-world test involving brain scans, the standard computer models suggested that people with better cognitive skills had "older" brains, implying that being smart made you age faster. This is a nonsensical conclusion. When the researchers corrected the computer's bias, the result flipped to the expected reality: people with better cognitive skills had "younger," more resilient brains. Similarly, in a study of kidney function, the flawed models suggested that worse kidney health was linked to a younger biological age, while the corrected models showed the logical link: poor kidney health is associated with accelerated aging.

To fix this, the team developed a new method called Unbiased Machine Learning Regression. Instead of letting the computer find the path of least resistance, which leads to that average-shrinking error, they added strict rules to the training process. They forced the model to be perfectly accurate for two specific groups of people: those who are younger and those who are older. By anchoring the model to these two points, the computer is compelled to draw a straight, accurate line through the data, ensuring that a prediction of "older" truly means older, and "younger" truly means younger. When they applied this new method to data from the UK Biobank and the Framingham Heart Study, the systematic errors vanished. The new clocks no longer showed the strange, reversed relationships between health and age.

The findings, published in a preprint, highlight a critical need for caution in how we use artificial intelligence to understand human health. The study demonstrates that simply having a computer model that looks accurate on paper is not enough; if the model is biased, it can lead to dangerous misunderstandings about what causes disease and how to treat it. The researchers have made their new, unbiased tools available to other scientists, offering a way to ensure that the next generation of aging clocks tells the truth about our bodies, rather than a story distorted by the limitations of the software used to build them.

Technical Summary: Trustworthy ML/AI for Aging Clocks

Problem Statement
Machine learning (ML) and artificial intelligence (AI) regression models are increasingly used to construct "aging clocks," which estimate biological age from omics and medical imaging data. A fundamental but underappreciated statistical limitation of these models is systematic prediction bias. Standard ML regression methods (e.g., random forests, gradient boosting, neural networks, regularized regressions) optimized for common objective functions like Mean Squared Error (MSE) tend to shrink predictions toward the mean. This results in a systematic pattern where predicted biological ages are underestimated for older individuals and overestimated for younger individuals.

This bias is particularly problematic in biological age research because the true biological age is a latent (unobserved) variable. Models are typically trained using chronological age as a surrogate outcome. While standard post-hoc calibration strategies exist, they often make predicted biological age overly dependent on chronological age, limiting the ability to capture genuine biological variation. More critically, this systematic bias propagates into downstream association analyses. When aging clocks are used as outcomes to assess associations with health factors (e.g., smoking, kidney function), the bias can distort or even reverse the sign of the estimated associations, leading to spurious conclusions (e.g., suggesting that older predicted brain age correlates with better cognitive performance).

Methodology
The authors propose a principled framework called Unbiased Machine Learning Regression (UMLR) to prevent systematic prediction bias during the model construction phase, rather than correcting it post-hoc.

Theoretical Characterization: The paper defines systematic prediction bias mathematically. For a true outcome $Y$ and a prediction $\hat{Y}$ , bias is characterized by a linear shrinkage toward the mean: $E(\hat{Y} | Y = y^*) - y^* = (\eta - 1)y^*$ , where $\eta \in [0, 1]$ . A value of $\eta < 1$ indicates bias.
Latent Outcome Handling: In aging clock applications, since true biological age $Y$ is unobserved, the authors prove (Theorem 1) that the systematic bias estimated by regressing the predicted biological age ( $\hat{Y}$ ) against chronological age ( $A$ ) is a consistent estimator of the bias relative to the latent biological age. This allows the bias to be quantified and corrected using the observable surrogate outcome.
Constrained Optimization: UMLR modifies the standard MLR objective function by adding constraints. The model minimizes the loss function (e.g., MSE) subject to two mean-matching constraints on two disjoint subsets of the training data ( $R_1$ $R_{1}$ and $R_2$ $R_{2}$ ), typically defined by a cutoff (e.g., the sample mean):
- $\sum_{i \in R_1} (f(X_i) - Y_i) = 0$
- $\sum_{i \in R_2} (f(X_i) - Y_i) = 0$
  Geometrically, these constraints force the fitted regression line to pass through two distinct anchor points on the unbiased 45-degree line, ensuring the slope equals 1 ( $\eta = 1$ ) and eliminating systematic bias.

Key Results
The paper validates the UMLR framework through synthetic experiments and two real-world case studies:

Synthetic Experiments: Demonstrated that standard MLR models consistently exhibit $\eta < 1$ (bias), leading to biased downstream association estimates. In contrast, UMLR achieved $\eta \approx 1$ , yielding unbiased estimates of the biological age gap and its associations with covariates.
Case Study 1 (Brain Age): Using neuroimaging data from the UK Biobank (training) and Human Connectome Project (testing), the authors compared standard MLR and UMLR.
- Standard MLR showed significant bias ( $\eta \approx 0.30$ in training, $\approx 0.37$ in testing).
- Downstream Impact: Standard MLR suggested a positive association between brain age acceleration and fluid cognition (implying older brains have better cognition), whereas UMLR showed the expected negative association (accelerated aging linked to worse cognition).
Case Study 2 (Epigenetic Age): Using DNA methylation data from the Framingham Heart Study (UMN training, JHU testing) to predict kidney function (serum creatinine).
- Standard MLR exhibited bias ( $\eta \approx 0.33$ in training, $\approx 0.37$ in testing).
- Downstream Impact: Standard MLR suggested a negative association between epigenetic age gap and creatinine (implying worse kidney function links to younger epigenetic age). UMLR corrected this, showing a positive (though not statistically significant in this specific test) association consistent with biological expectations (worse kidney function links to accelerated aging).

Notably, while UMLR often results in a lower $R^2$ and higher RMSE relative to chronological age compared to standard MLR, the authors argue this is a desirable trade-off. A lower $R^2$ with the surrogate outcome indicates the model is less constrained to simply mirror chronological age and is better at capturing biologically meaningful deviations.

Significance and Claims
The paper claims that systematic prediction bias is a pervasive issue in ML/AI-based aging clocks that can fundamentally alter biomedical inferences. The significance of this work lies in:

Identifying a Critical Flaw: Highlighting that standard ML optimization criteria favor biased predictions that distort downstream causal or associative inferences in biological age research.
Providing a Solution: Introducing UMLR as a broadly applicable, principled framework that prevents bias at the source through constrained optimization, ensuring valid downstream inference without requiring the true biological age to be observed.
Correcting Misleading Conclusions: Demonstrating through case studies that ignoring this bias can lead to conclusions that are not only incorrect but diametrically opposite to biological reality (e.g., reversing the direction of risk associations).

The authors emphasize that for studies using aging clocks as primary endpoints or biomarkers to inform public health and clinical decisions, adopting unbiased frameworks like UMLR is essential to ensure the validity of the inferred relationships between aging and health outcomes.

Technical Summary: Trustworthy ML/AI for Aging Clocks

More like this