Aging clocks have become a routine tool: biological age is computed from DNA methylation, the transcriptome, or brain scans and then compared with chronological age. An upward deviation is read as accelerated aging and linked to disease, while a downward one is read as resilience. However, standard machine learning models (random forest, gradient boosting, neural networks) systematically compress predictions toward the mean: age is underestimated in older people and overestimated in younger ones. This bias is inherited by aging clocks because true biological age is unobservable, and models are trained on chronological age as a surrogate. As a result, associations between accelerated aging and health can be distorted to the point of changing sign, which casts doubt on the conclusions of many studies. The work stands apart: nearby in the archive are papers on cell, battery, and brain aging, but no one has tackled the problem of prediction bias specifically for aging clocks.
The authors first formalize the systematic bias of predictions in general machine learning regression: if the regression of the predicted outcome on the true one has a slope less than one, the bias is linear and proportional to 1 − η. They then carry this over to aging clocks, where true biological age is hidden and chronological age is used as a surrogate, and prove a theorem on the consistency of the bias estimate from the surrogate. To remove the bias, they propose an unbiased machine learning regression (UMLR) method: two constraints on mean predictions in two non-overlapping subgroups split by outcome (for example, by the median age) are added to the objective function. These constraints act as two anchor points that pin the prediction line to the diagonal 'predicted = true', ensuring a slope of 1. The method is tested in simulations and on two real datasets: brain age from fractional anisotropy of white matter from UK Biobank (training) and the Human Connectome Project in Aging (test), as well as epigenetic clocks from CpG sites from the Framingham Heart Study, with training on Minnesota data and testing on Johns Hopkins data.
In simulations, standard models (random forest, XGBoost, neural networks, LASSO) show a systematic bias: the slope of the regression of predicted on true outcome is less than one, and the bias is largest at the edges of the age range. UMLR removes the bias, yielding a slope close to one, although the root mean square error may be higher and R² lower than for biased models. In the brain age case on the HCP test set, biased models give a positive link between the biological age gap and cognitive score (t= −2.62, p=0.066), meaning accelerated aging is linked to better cognitive function, which contradicts expectations. UMLR flips the sign: the link becomes negative (t=4.38, p=1.58×10⁻⁵), meaning accelerated aging is linked to worse cognitive function, as it should be. In the epigenetic clock case, the link between the gap and serum creatinine (a marker of kidney function) also flips sign: biased models give a negative association (t= −3.68, p=2.94×10⁻⁴), while UMLR gives a positive one, though not significant (t=0.79, p=0.43). The bias parameter η for standard models in the test data was about 0.5–0.7, whereas for UMLR it was close to 1.0.
The work shows that conclusions about the link between accelerated aging and diseases and risk factors may be an artifact of prediction bias rather than a biological signal. Unlike neighboring studies that build clocks or look for associations without controlling for bias, this one provides the first formal definition of bias for aging clocks, proves that it carries over to unobservable biological age, and proposes a correction method at the training stage. This changes the standard of analysis: instead of post hoc calibration, which makes predictions too dependent on chronological age, UMLR builds unbiasedness into the model itself. For the epidemiology of aging, this means that many published associations may need to be revised, and future studies should use unbiased methods.
The UMLR method can be extended to other types of aging clocks (transcriptomic, proteomic, multi-omic) and to deep learning models, where bias is also present. The constraints could be adapted for nonlinear bias patterns and for censored outcomes, for example in longitudinal studies. Another direction will be the development of diagnostic tests to check for bias in already published clocks and the recalculation of associations in large biobanks using UMLR. It would also be interesting to study how bias affects causal estimates (for example, in Mendelian randomization) and clinical decisions based on biological age.
The results will affect the epidemiology of aging, gerontology, neuroimaging, and epigenetics, where aging clocks are used as biomarkers. They are also important for clinical trials, where biological age serves as a surrogate endpoint, and for personalized medicine, which assesses an individual's rate of aging.
The immediate steps include applying UMLR to existing large cohorts (UK Biobank, Framingham, HCP) to recalculate known associations and publishing a software package with a user-friendly interface. It is also necessary to validate the method on external datasets with different age distributions and compare it with alternative approaches to removing bias.
In the field of aging, the question of what clocks actually measure remains unresolved: biological age or training artifacts. This work links that problem to the fundamental statistical task of inference from predicted outcomes, showing that without controlling for bias one cannot separate true accelerated aging from regression to the mean. This echoes the long-standing debate about the validity of aging biomarkers and the need for a gold standard for biological age.
In the brain age case, biased models claimed that people with better cognitive function have brains that age faster, while unbiased ones said the opposite. The same dataset led to opposite medical recommendations depending on whether the bias was corrected.
In the series Altered Carbon, biological age can be separated from chronological age with technology, but even there estimates of cerebral cortex age suffer from systematic errors if the algorithm is not calibrated. The work reminds us that before trusting 'aging clocks', we need to make sure they do not lie at the edges of the range.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.