Blood-Based Aging Scores Predict Surgical Death Risk but Fail to Travel Between Hospitals

Two laboratory-derived measures of biological aging and frailty—Phenotypic Age Acceleration (PAA) and the Laboratory Frailty Index (FI-Lab)—were associated with postoperative mortality in a dual-center cohort study of 1,223 adults undergoing surgery in eastern China. Yet when researchers tested whether these blood-based markers could genuinely improve mortality prediction in a new hospital, the answer was largely no. The study, published in Frontiers in Endocrinology, offers a sobering lesson about the gap between statistical association and clinically deployable prediction.

The research team, led by anesthesiologists at hospitals affiliated with Qingdao University, Shandong First Medical University, and Nanjing Medical University, analyzed harmonized data collected during 2024 and 2025 from the health examination centers of the Second Affiliated Hospital of Shandong First Medical University and the Affiliated Taizhou People’s Hospital of Nanjing Medical University. From initial candidate pools of 3,527 and 3,411 patients, sequential screening left 671 participants in Shandong and 552 in Jiangsu—just 19.0 percent and 16.2 percent of the original pools. Eligibility required a complete preoperative laboratory panel drawn within 24 hours before the index operation, along with full perioperative covariates including American Society of Anesthesiologists physical status, Charlson Comorbidity Index, FRAIL questionnaire score, surgical stress grade, and cancer-related surgery status.

Phenotypic Age is a mortality-derived measure of biological age calculated from chronological age plus nine circulating biomarkers spanning nutrition, inflammation, hematology, renal function, hepatic function, and metabolism: serum albumin, creatinine, fasting glucose, C-reactive protein, lymphocyte percentage, mean corpuscular volume, red cell distribution width, alkaline phosphatase, and white blood cell count. PAA captures how much older—or younger—a person’s Phenotypic Age is than expected for their chronological age. FI-Lab takes a different route to the same territory, counting each abnormal result among 17 prespecified laboratory deficits and expressing the total as a proportion from zero to one. Nine of the 17 deficits overlap with the Phenotypic Age biomarkers, so the two measures are biologically related but weight abnormalities differently.

Over 1,694.66 person-years of follow-up, 74 participants died—39 in Shandong and 35 in Jiangsu—with 57 of those deaths occurring within 365 days. In the fully adjusted, center-stratified Cox model, each one-standard-deviation increase in PAA was associated with a 34 percent higher hazard of death through 540 days (hazard ratio 1.34; 95 percent confidence interval 1.03 to 1.75). The FI-Lab estimate pointed in the same direction but was borderline (hazard ratio 1.28; 95 percent confidence interval 1.00 to 1.65). Restricted cubic splines found risk gradients compatible with roughly linear dose-response relationships for both indices, and participants with both positive PAA and at least three laboratory deficits had the lowest survival in Kaplan-Meier analyses.

Beneath those headline numbers, however, lay a methodological trap. The researchers used Shandong’s age-to-Phenotypic-Age relationship as a fixed reference for both centers, and that choice backfired in Jiangsu. The regression slope of Phenotypic Age on chronological age was 1.66 in Shandong but only 0.98 in Jiangsu. As a result, fixed-reference PAA was essentially uncorrelated with age in Shandong (correlation near zero) but strongly inversely correlated with age in Jiangsu (r = −0.74). In other words, the absolute PAA values were not directly transportable between hospitals, because the underlying age reference relations differed. When the team recalculated PAA using each center’s own age residuals and standardized within center, the fully adjusted pooled association held firm at a hazard ratio of 1.39 (95 percent confidence interval 1.09 to 1.77)—robust in direction, but not interchangeable as an absolute scale.

The prediction analyses were the study’s most consequential test. A ridge logistic model incorporating standard clinical variables—age, sex, ASA physical status, Charlson Comorbidity Index, FRAIL score, surgical stress, and cancer surgery—was trained in Shandong and applied, without refitting, to Jiangsu for 365-day mortality. That locked clinical model achieved an area under the receiver operating characteristic curve of just 0.60 (95 percent confidence interval 0.45 to 0.73). Adding PAA changed the AUC by essentially nothing (0.60). Adding FI-Lab, or both markers together, nudged the AUC to 0.63, but the confidence intervals for those differences spanned zero, and decision-curve analysis showed no consistent gain in net benefit across clinically plausible thresholds.

Calibration told an equally cautious story. Only 23 deaths occurred in the Jiangsu validation cohort by 365 days, making calibration estimates imprecise. Calibration slope point estimates of 2.61, 4.38, and 2.01 for the clinical model, the model plus PAA, and the combined model suggested that predicted risks may have been too compressed, but the wide confidence intervals all included one. The authors explicitly declined to interpret nonsignificant calibration tests as evidence of good calibration—a methodologically disciplined stance that many prediction studies skip. Supplementary reclassification metrics, continuous net reclassification improvement and integrated discrimination improvement, did favor FI-Lab, but the researchers treated them as supplementary rather than proof of clinical utility, since they were not accompanied by gains in discrimination, calibration, or net benefit.

Secondary analyses broadened the picture. FI-Lab, but not PAA, was associated with a 12-month composite of death or incident diabetes, chronic kidney disease, or cardiovascular disease (relative risk 1.36; 95 percent confidence interval 1.08 to 1.71). In disease-specific analyses corrected for multiple testing, both measures were associated with incident diabetes, and PAA was also associated with incident chronic kidney disease. An inverse PAA-cardiovascular disease estimate, based on only 29 events, was flagged as exploratory rather than evidence of protection. Exploratory longitudinal analyses in 820 participants with repeat measurements showed mean PAA increases of about 2.90 years and small FI-Lab increases of 0.02 at approximately 12 months—but the authors stressed these should not be read as annualized biological aging rates, because exact reassessment intervals were unavailable and follow-up chronological age had to be approximated as baseline age plus one year.

The study’s limitations are substantial and honestly catalogued. Only 74 deaths occurred by 540 days, and the primary validation cohort had just 23 deaths within a year, limiting precision for calibration, interaction, and incremental-performance estimates. The analytic cohorts represented less than a fifth of the initial candidate pools, raising the possibility of selection bias related to complete-data eligibility. Analyzer manufacturer and reagent metadata were unavailable, so assay or calibration differences between hospitals cannot be excluded as contributors to the higher FI-Lab distribution in Shandong or to center-specific effect estimates. The health-examination-based cohorts may not represent the broader surgical population, and residual confounding remains possible despite adjustment for comorbidity, frailty, and surgical stress.

The takeaway is nuanced but clear. Laboratory-derived aging and frailty measures behaved as risk markers—reproducibly associated with postoperative mortality across two hospitals, with the PAA association surviving center-specific residualization. But association is not prediction. Neither marker consistently improved externally validated mortality prediction beyond a clinical model whose own discrimination was modest, and absolute fixed-reference PAA values proved dependent on the reference population’s age relationship. The authors conclude that routine perioperative use of these measures should await larger, more diverse multicenter studies with rigorous cross-center laboratory harmonization and demonstrated added clinical net benefit. Until then, a blood test that seems to measure how fast you are aging may tell you something real about surgical risk—just not yet in a form that travels reliably from one hospital to the next.

Subject of Research: External validation of phenotypic age acceleration and laboratory frailty indices for postoperative mortality prediction across two Chinese hospitals

Article Title: External validation and cross-center transportability of phenotypic age acceleration and laboratory frailty for postoperative mortality prediction: a dual-center cohort study

Article References: Tang, R., Yu, J., & Ji, X. (2026). External validation and cross-center transportability of phenotypic age acceleration and laboratory frailty for postoperative mortality prediction: a dual-center cohort study. Frontiers in Endocrinology, 17, Article 1970011. https://doi.org/10.3389/fendo.2026.1970011

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1970011

Keywords: biological aging, phenotypic age acceleration, laboratory frailty index, postoperative mortality, external validation, perioperative risk, frailty, prediction models, calibration, transportability, biomarkers, cohort study