Biological Age Modeling Using Electronic Health Records

Background Chronological age poorly captures individual differences in physiologic health across the lifespan. Although biological age has shown promise as a more informative measure of health status, most studies have relied on curated research cohorts rather than routinely collected electronic health record (EHR) data. Among clinical biological measures, Levine PhenoAge and Klemera-Doubal Method (KDM), which estimate biological age from clinical biomarkers, are among the most widely used. However, KDM, in particular, has been rarely applied outside large biobanks or narrow disease-specific hospital contexts, and direct comparisons of KDM with PhenoAge using identical biomarkers in the same real-world population remain scarce. Objective To evaluate the feasibility and clinical relevance of estimating biological age from EHR data by applying Levine PhenoAge and the Klemera-Doubal Method (KDM) to a real-world, multi-disease patient population, including a head-to-head comparison of the two measures, and to evaluate their associations with mortality, morbidity, frailty, and hospital readmission. Methods We conducted a retrospective study of 93,200 encounters from 2,434 unique patients (ages 30–75) at a large health system in western Pennsylvania, with a median follow-up of 10.2 years. Levine PhenoAge and KDM biological age were calculated using a nine-biomarker panel, and age acceleration was derived as the residual of biological age regressed on chronological age. Associations between biological age acceleration and clinical outcomes, including mortality, comorbidities, frailty, and 30-day hospital readmission, were evaluated using Cox proportional hazards, logistic, and negative binomial regression models. Results The cohort had a mean age of 50.6 years (59.6% female), with a mean PhenoAge of 54.3 years (SD 17.8 years) and a mean KDM age of 50.8 years (SD 21.45 years), both of which were strongly correlated with chronological age (r = 0.75 and 0.638, respectively) and with each other (r = 0.78). Chronological age alone was not significantly associated with 30-day readmission rate (p = 0.78), whereas both PhenoAge and KDM acceleration were associated with increased readmission rate(17.6% and 5.8% per 5-year increase, respectively), indicating that biological age acceleration captures readmission risk that chronological age misses entirely. PhenoAge and KDM acceleration were each independently associated with mortality beyond chronological age (HR = 1.17 and 1.08 per 5-year increase), significantly improving the model fit over age and sex alone (LRT χ² =62.99 and 30.90, respectively, p < 0.001). Both PhenoAge and KDM acceleration were independently associated with frailty beyond chronological age, with significant improvement in model fit over an age- and sex-only model (LRT χ² =25.80 and 20.00, respectively, p < 0.005). PhenoAge and KDM age acceleration were also strongly associated with numerous comorbidities, most prominently renal conditions (chronic kidney disease, hyperkalemia), after correction for multiple comparisons. Longitudinal tertile transition analysis showed that male sex predicted both a greater likelihood of transitioning into the highest acceleration tertile from the lowest acceleration tertile (OR = 1.86) and a lower likelihood of transitioning out of the highest acceleration tertile (OR = 0.56), indicating that males tend toward, and persist within, faster-aging trajectories over time. Conclusions This study demonstrates that KDM can be reliably applied to heterogeneous real-world EHR data and directly compared with PhenoAge using an identical biomarker panel within the same clinical population. To our knowledge, KDM has previously been applied primarily within large biobanks or narrow disease-specific cohorts. This study extends KDM to a heterogeneous, multi-disease clinical population and is the first to link biological age acceleration to 30-day hospital readmission across a broad patient population. These findings support the potential utility of EHR-derived biological age as a clinical biomarker of risk stratification.