2026-09-03
MOSAIC-Age from MD Anderson trains eight organ-specific aging clocks on 9,971 anomaly-free chest CTs. Across 35,293 participants in two prospective cohorts, organs that look older on CT track with chronic disease and mortality, with the aortic age gap the most consistent predictor of death.
Biological aging runs at different rates across individuals and across organs within one person. The available clocks are either blood-based proteomic models (molecular, blind to which organ) or brain MRI models (one organ). Chest CT sits oddly unused: a single scan covers lungs, heart, aorta, bone, and muscle, and tens of millions are performed each year for other reasons, archived after one read.
This MD Anderson preprint builds the opportunistic-screening case end to end: from a routine chest CT, can you read which organs are aging fast, and does that predict future disease and death?
MOSAIC-Age works in four steps:
After development the pipeline was locked and applied untouched to two independent prospective cohorts: NLST (25,824 participants) and COPDGene (9,469), totaling 35,293 people with long-term follow-up.
Age prediction is solid: median Pearson r = 0.94 with chronological age across organs, median MAE 5.14 years per organ clock and 3.70 for the composite, comparable to prior MRI clocks. Male-female MAE differs by a median 0.35 years; agreement between diagnostic CT and PET/CT-derived estimates is median r = 0.84.
The clinical associations, all after multiple-testing correction:
| Association | Effect | Detail |
| Lung BAG × prevalent COPD | OR 1.50 | per +1 SD, q=2.4e-95 |
| Coronary BAG × prevalent heart disease | OR 1.98 | q=9.9e-226 |
| Extremely aged aorta × all-cause death | HR 2.09 | BAG>1.5 SD, q=1.5e-34 |
| ≥5 extremely aged organs × death | HR 2.64 | vs. none, q=3.4e-10 |
| Composite BAG top vs. bottom quartile × death | HR 1.77 | P=1.0e-45 |
| Aorta pace +1 SD × death | HR 1.20 | adjusted for baseline BAG, q=1.7e-11 |
Of 180 disease-BAG pairs, 66% were significantly positive for prevalent disease. With all eight organ BAGs in one model, muscle, aorta, lung, and coronary remained independent mortality predictors. Cause-specific analyses line up anatomically: lung BAG leads respiratory death (HR 1.47), coronary BAG leads cardiovascular death (HR 1.45). Against genetics, lung BAG predicts incident COPD (OR 1.41) better than a lung-function polygenic risk score (OR 1.13), and the two correlate at only r=0.051, imaging and inherited risk carry distinct information.
One counterintuitive result: the strongest predictor of incident lung cancer is not lung BAG but aorta BAG (HR 1.17 per SD, the only survivor in multivariable analysis). The authors stay measured: this likely reflects shared pathways, smoking, chronic inflammation, oxidative stress, rather than local aortic aging causing carcinogenesis.
Tens of millions of chest CTs are read once and archived. This work shows the same scan carries a multi-organ aging report at zero additional imaging cost or blood draw. For lung screening populations it adds a risk-stratification dimension current guidelines lack.
Methodologically it is a complete template for opportunistic imaging AI: an open-source segmentation model, under 10,000 training scans, and a locked-pipeline external validation that other radiology groups can copy for other organ systems.
For ML practitioners, the transferable pieces are the per-organ heads, label smoothing on a continuous target, domain bridging via low-dose simulation, and downstream statistics on residuals rather than raw predictions.