Longevity & AgingResearch PaperOpen Access

LifeClock uses routine blood tests to track biological age from infancy to old age

A transformer model trained on 24.6 million clinic visits estimates biological age across the lifespan and flags future disease risk in children and adults.

Sunday, November 15, 2026 0 views
Published in Nat Med
Glowing digital timeline from infant to elder silhouettes, overlaid with blood-test vials and flowing data streams in blue teal

Summary

LifeClock is an AI biological clock built from routine electronic health records and lab tests rather than DNA methylation or omics data. Researchers trained a transformer model, EHRFormer, on 24.6 million visits from 9.7 million people in Chinese hospital cohorts, then tested it on external cohorts including UK Biobank. Pediatric and adult physiology differ so much that the team built separate clocks: a 'physiological maturity' clock for under-18s and an aging clock for adults. The gap between predicted and chronological age tracked disease. Children with accelerated or delayed maturity showed the expected growth disorders, and over-aged adults had more disease. Fine-tuned versions predicted diseases such as diabetes, stroke and coronary artery disease with AUCs up to 0.98 for current diagnosis and at least 0.8 for future risk. The work is promising but retrospective and hospital-based.

Detailed Summary

Most aging clocks are built for adults and rely on DNA methylation or other specialized assays. Childhood is largely missing, even though rapid development there reflects scripted maturation rather than accumulated damage. A clock that works across the whole lifespan and uses cheap, routine data could make biological-age assessment practical in everyday care.

The researchers developed EHRFormer, a time-series transformer trained on longitudinal electronic health records and lab tests. Training data came from four hospital cohorts in the China Healthy Aging Investigation: 9,680,764 people and 24,633,025 visits, starting from 184 clinical indicators. The model used random masking of inputs and outputs to handle missing values, and adversarial training to remove differences between hospitals. It was also trained autoregressively, so each visit's representation reflects the person's health trajectory. A regression model trained on healthy individuals predicted chronological age, and that prediction served as biological age. External validation used a fifth cohort (86,257 people) and UK Biobank (146,087 people).

The data showed two distinct lab-marker regimes, a pediatric phase (birth to 18 years) and an adult phase, so the team built separate clocks, which improved accuracy. SHAP analysis showed that the top contributors differed almost completely between them. Low AST, high creatinine and high total protein drove the pediatric clock, while high urea, low albumin and high red cell distribution width drove the adult clock. In UK Biobank, the adult model reached a mean absolute error of 4.14 years, with urea, albumin and RDW again among the top contributors.

Clustering of EHR representations produced 64 clusters, 14 pediatric and 50 adult, with distinct disease trajectories. Over-aged adults (age gap above 3 s.d.) had higher disease prevalence than average-aged peers in the same cluster, and the difference grew over time. In children, accelerated maturity appeared in precocious puberty and overgrowth syndromes, and deceleration appeared in delayed puberty, growth hormone deficiency and developmental delay. Some clusters carried large hazard ratios, such as 37.7 for renal failure in adult cluster 20 and 15.4 for pituitary hyperfunction in pediatric cluster 14. Fine-tuned models diagnosed current disease with AUCs of 0.94 to 0.98, including atrial fibrillation, coronary artery disease, ischemic stroke, diabetes and Parkinson's disease. They predicted future disease with AUCs of at least 0.8 and 5- and 10-year incidence AUCs of 0.80 to 0.91. They outperformed RNN and XGBoost baselines. Pediatric risk using data before age 12 reached AUCs of 0.70 to 0.96, and high-risk groups separated clearly in cumulative incidence curves.

These results suggest that routine labs, including values within normal ranges, contain information about maturity, aging and disease risk that single-visit tools may miss. Caveats are important. The data are retrospective hospital records, so patients are not representative of the general population, and the clock is trained on healthy individuals defined within a clinical setting. Chronological age is the training target, which can limit how well the clock captures true biological aging. High AUCs may partly reflect diagnostic information already in the records, and prospective validation and mortality outcomes would strengthen the claims.

Key Findings

  • LifeClock was trained on 24.6 million visits from 9.7 million people and validated externally in a fifth cohort and in UK Biobank.
  • Pediatric and adult clocks rely on almost entirely different lab markers; urea, albumin and RDW drive the adult clock.
  • In UK Biobank, the adult clock reached a mean absolute error of 4.14 years, with the same top biomarkers.
  • Fine-tuned models diagnosed diabetes, stroke and coronary artery disease with AUCs of 0.97–0.98 and predicted future risk with AUCs of at least 0.8.
  • Children with growth-promoting or growth-inhibiting disorders showed accelerated or decelerated maturity relative to healthy controls.

Methodology

Retrospective multi-cohort study using a transformer foundation model (EHRFormer) with masked reconstruction, adversarial cohort-debiasing and autoregressive next-visit training on 184 lab and clinical indicators. Age regression models trained on healthy individuals defined biological age. Disease models were fine-tuned and validated internally, in an independent cohort and in UK Biobank, using AUC, Cox models and SHAP.

Study Limitations

Data are retrospective hospital records, mostly from Chinese cohorts, so selection bias and limited generalizability beyond hospital populations are likely. The clock is trained to predict chronological age, and it is unclear how closely this captures true biological aging, and the full text provided ends before the discussion, so the authors' stated limitations are not covered. Reported AUCs for current diagnosis may be inflated by diagnostic information already present in the records, and no intervention or mortality outcomes were shown in the available text.

Enjoyed this summary?

Get the latest longevity research delivered to your inbox every week.

Enter your email to subscribe: