A CT-derived age gap tracked disease risk in one screening cohort

What is newly reported

First posted September 21, the study trained a model to estimate chronological age from routine body-composition features on CT scans, then asked whether the difference between CT-predicted and chronological age tracked disease. It reports a risk marker from existing images—not an intervention and not a direct measurement of how fast someone is aging.

How the study was organized

The researchers divided 7,837 adults into a 2,247-person development dataset, a 2,416-person prevalent-disease test dataset and a 3,174-person longitudinal test dataset. The final ElasticNet model used 37 CT features. All datasets came from the same University of Tokyo screening program and scanner protocol.

That split helps separate model fitting from testing inside the cohort. It does not provide the independent external validation needed to know whether performance transfers to other hospitals, scanner vendors, countries or populations.

How well did it estimate age?

In validation, the model explained 65.1% of the variation in chronological age (R²=0.651), with a mean absolute error of 4.63 years. Predicting chronological age is a technical benchmark; it does not by itself establish that the residual “age gap” is a biological-age measurement.

The age gap was associated with disease

Each 1-standard-deviation higher age gap was associated with 35% higher odds of prevalent disease (OR 1.35, 95% CI 1.26–1.45). In the longitudinal test set, 943 people developed a disease during follow-up; the corresponding hazard ratio was 1.24 (95% CI 1.17–1.32).

These are relative associations across a one-standard-deviation contrast. They do not tell an individual their absolute probability of disease, diagnose a condition, or show that changing the age gap would change health.

The added prediction was small

For five-year incident disease, the time-dependent area under the curve was 0.700 for the CT-age model versus 0.688 for chronological age. That difference may justify further validation, but it does not demonstrate that using the model improves clinical decisions or patient outcomes.

Why routine screening is not ready

The cohort was selected, mostly male and drawn from one Japanese health-screening setting. The paper reports age bias, uncertain disease labels and sparse events for individual diseases. It lacks external and prospective validation. Data are private, and the manuscript says code and the model will be released on publication; code availability was not treated as verified while the paper remains Article in Press.

CT also exposes people to ionizing radiation. The authors explicitly frame possible use as opportunistic analysis of scans already acquired for another clinical reason, not a stand-alone biological-age test. This study is not a reason to seek a CT scan.

What should readers watch next?

External validation across hospitals, scanners and populations should come first. Prospective studies then need to show calibration, decision value and whether model use improves outcomes without unnecessary imaging or follow-up. Even strong prediction would remain different from proving that an intervention slowed aging.