Biological age estimates vary with the data and the model’s prediction target

Biological age estimates vary with the data and the model’s prediction target

On October 19, NaukaPRO published a 22-minute lecture by researcher Dmitry Kryukov. He explains that a calculator combines measurements of the body into a single number whose meaning depends on the data and the target the algorithm was trained to predict.

Kryukov describes biological age as an “artificial construct.” It is calculated from biomarkers, measurable indicators of the body’s condition, such as blood pressure or blood test results. The developer selects the biomarkers, the calculation method, and the target the algorithm should predict.

A calculator trained to predict chronological age produces an estimate close to chronological age. A model can also be trained to predict the risk of death or disease, in which case its output estimates that risk. In a paper by Kryukov and colleagues, the authors write that each such choice defines its own version of biological age.

During cell reprogramming, clocks sometimes returned ages below zero: the model encountered data that were not represented in its training sample. Kryukov suggests examining the biomarkers themselves to determine whether they reflect and influence the mechanisms of aging. The paper’s authors consider clocks useful for estimating risk when the algorithm performs consistently in an appropriate population.

Kryukov advises asking services about the proxy, the measurable target the model uses in place of a direct measure of aging, and about the biomarkers in the training data:

“Show the proxy and the original biomarkers used to train the model.”

A 2026 systematic review included 81 models based on measurements from routine tests and examinations across 59 studies. For 31 models, both internal validation using data from the original study and external validation in a different population were missing. Interpreting a biological age estimate requires knowing three things: the outcome it predicts, the measurements used to calculate it, and the population in which the algorithm was validated.