An aging atlas of 189,000 Chinese residents built from routine health checkups: the faster a person's organ systems diverge between exams, the higher the risk of death
An aging atlas of 189,000 Chinese residents built from routine health checkups: the faster a person's organ systems diverge between exams, the higher the risk of death
Bioinformaticians at Zhejiang University collected 94,055,326 measurements of 99 biomarkers from 189,095 people who underwent repeated health checkups, and mapped these data across ten organ systems. For 147,492 of these individuals, mortality records were also available, making it possible to test whether the atlas is associated with the risk of death.
Once a year, a person gets blood work done and checks their heart and kidneys, then receives results they rarely compare with the previous year's. Calculating "biological age" from blood tests or DNA methylation is already feasible, and it is well established that the kidneys, liver, immune system, and vasculature within the same person age at different rates. These methods, however, are expensive and limited in scale: this spring, Digital Aging Twin decomposed biological age by organ using MRI and methylation in just two thousand people.
On September 21, a study was published showing that data from ordinary annual checkups are sufficient for such a map. The team mapped the data across ten organ systems (cardiovascular, renal, metabolic, and others) and for each of the 99 biomarkers calculated how it typically changes with age in men and women and how far a given individual deviates from that norm. No single direction of change emerged: some biomarkers rose with age, some fell, and some moved in opposing directions or diverged between men and women. Renal and metabolic markers deteriorated notably over the years, while the immune system restructured across multiple blood cell types.
Because organ systems age at different rates, the natural hypothesis was that greater divergence between them would signal higher mortality. Testing on 147,492 people and 4,599 deaths did not confirm this: the spread between systems at a single exam, taken on its own, showed no independent association with mortality. What was associated with mortality was the average burden of deviations across all systems combined: the higher this burden, the higher the risk of death, by 5.8% per unit increase. A second finding, which the authors themselves describe as preliminary, emerged when they compared consecutive exams for the same individual. Mortality risk increased by 19% per unit of this between-exam instability, and for cardiovascular deaths the effect was nearly one and a half times stronger, although it weakened under a more stringent statistical correction.
Because the atlas explains deaths that have already occurred, the authors tested whether it could also predict health problems that have not yet happened. They trained a neural network called DynLifeNet on the same data. Given a person's history of checkups, the model predicts seven conditions one to three years in advance: decline in kidney function, excess uric acid, hypertension, glucose and cholesterol abnormalities, anemia, and elevated liver enzymes. The model's accuracy was 0.80 out of 1, where random guessing would yield 0.5. This is comparable to conventional machine learning algorithms that do not separate biomarkers by organ system, though the organ-system decomposition does add a statistically significant, if modest, improvement in accuracy on top of the raw measurement history.
Validation on an independent sample of 16,478 people from a separate longitudinal Chinese study confirmed the ranking of system importance but not the absolute values: a scale calibrated in Zhejiang does not transfer directly to a different population. The authors themselves describe their atlas as measurement infrastructure, a tool that turns routine clinic data into continuous monitoring of how differently the body ages.