chatGPT(5.6paid)
Paper reviewed
A generalized approach for estimating the pace of aging, Sluiskes et al., Journal of Gerontology: Medical Sciences, 2026. DOI: 10.1093/gerona/glag205.
Overall assessment
This is a useful methodological proof of concept. Its main innovation is to replace the equal weighting used in the original Dunedin Pace of Aging with weights learned from mortality.
The approach is intuitively attractive, especially when some candidate biomarkers may be uninformative. However, the paper does not yet demonstrate that the resulting weighted score is substantially better, externally valid, or a literal measure of the rate of biological aging. The strongest interpretation is that it is a mortality-weighted index of longitudinal functional deterioration.
Summary
Research question
The original Dunedin Pace of Aging:
- Estimates each person’s rate of change across multiple biomarkers.
- Standardizes the biomarker slopes and adds them together.
- Consequently gives each biomarker equal importance.
- Was developed in relatively young adults, among whom mortality was rare.
The authors ask whether the biomarker slopes can instead be weighted according to how strongly they predict mortality.
Cohort
The analysis used the Swedish Adoption/Twin Study of Aging:
- 745 adults with sufficient biomarker data.
- Mean age at first assessment: 62.3 years.
- Mean age at last assessment: 77.1 years.
- 61.5 percent women.
- Mean of 4.8 testing occasions over as many as nine waves.
- 572 deaths, representing 76.8 percent of participants.
- Median follow-up from the final assessment to death or censoring: 6.5 years.
The ten measures were:
- Waist-to-hip ratio
- Grip strength
- Mini-Mental State Examination
- Flexibility
- Balance
- FEV1
- FVC
- FEV1/FVC
- Haemoglobin
- Blood glucose
BMI and triglycerides were excluded because their mortality relationships can reverse in old age. Several lipid measurements were excluded because their age trajectories appeared to be affected by lipid-lowering treatment.
Construction of the weighted Pace of Aging
The procedure had two main stages:
-
A separate mixed-effects model was fitted for each biomarker.
These models allowed the average biomarker trajectory to be nonlinear with age and to differ by sex. They then estimated each participant’s random slope, meaning that person’s longitudinal deviation from the average age trajectory.
-
The ten estimated random slopes were entered into a Gompertz accelerated failure-time mortality model.
The mortality coefficients became the weights used to combine the biomarker slopes. The resulting weighted Pace of Aging, or wPoA, was expressed as a multiplier centred near 1.
The authors interpret 1.0 as average aging and 1.1 as aging approximately 10 percent faster than average.
Results
The wPoA had:
- Mean: 1.00
- Standard deviation: 0.04
- Range: 0.89 to 1.16
- No association with chronological age at the last assessment
The outcome models report hazard ratios per one standard deviation higher score:
| Outcome | Weighted PoA | Unweighted PoA | External weighted PoA | Matched unweighted PoA |
|---|---|---|---|---|
| Cardiovascular disease | 1.49 | 1.35 | 1.34 | 1.29 |
| Diabetes | 1.27, not significant | 1.16, not significant | 1.21, not significant | 1.12, not significant |
| Dementia | 1.44 | 1.53 | 1.67 | 1.61 |
| All-cause mortality | 1.50 | 1.42 | 1.41 | 1.40 |
The externally weighted version used mortality associations taken from previous studies and only seven of the ten biomarkers.
Among 425 participants with complete comparison data, wPoA correlated:
- 0.80 with the unweighted PoA
- 0.89 with the externally weighted version
- 0.42 with the Functional Aging Index
- 0.34 with the Frailty Index
- 0.14 with PhenoAge acceleration
- 0.13 with GrimAge acceleration
- 0.12 with physiological age acceleration
- 0.12 with DunedinPACE
Thus, its overlap with molecular aging clocks was statistically significant but quantitatively weak.
Novelty
What is genuinely new
The principal novelty is the first implementation of mortality-derived weights for the individual biomarker slopes underlying a Pace of Aging score. The original Dunedin paper discussed weighting as a possibility but did not implement it.
Other useful refinements include:
- Combining longitudinal mixed models with an accelerated failure-time model.
- Expressing the result as a putative aging-speed multiplier.
- Allowing nonlinear average biomarker trajectories in older adults.
- Accounting for sex-specific age trajectories.
- Including twin-pair clustering in the longitudinal component.
- Beginning mortality follow-up at the final biomarker assessment, thereby avoiding use of future measurements to predict earlier survival.
What is not new
The individual components are established methods:
- Random slopes have already been used in several longitudinal aging indices.
- Nonlinear biomarker trajectories in older adults were incorporated by Balachandran et al. in 2025.
- Outcome-trained biological-age measures are already common.
- Two-stage longitudinal and survival modelling is a conventional statistical strategy.
The novelty is therefore best described as a moderate methodological extension and synthesis, rather than a fundamentally new model of aging.
Strengths
- The cohort has unusually long longitudinal follow-up and a high number of deaths.
- The distinction between nonlinear population aging and individual deviations is sensible.
- Starting survival follow-up at the last measurement avoids an obvious look-ahead bias.
- The authors compare weighted and unweighted versions using the same biomarkers.
- They recognise the danger of internal overfitting and attempt an externally weighted sensitivity analysis.
- The paper openly discusses the limitations of two-stage modelling and the potential value of joint models.
Critique
1. Mortality prediction is partly circular
The main wPoA weights were learned from mortality in SATSA and then evaluated for mortality in the same participants. Its stronger mortality association is therefore expected and cannot be treated as independent validation.
The externally weighted analysis reduces this problem, but its results are revealing: the mortality hazard ratio is 1.41 for the weighted score and 1.40 for its matched unweighted counterpart. The apparent advantage almost disappears.
Cross-validation, bootstrap optimism correction or, preferably, an independent cohort is needed.
2. The claimed improvement is small
Weighting produces only modest changes in hazard ratios:
- It improves cardiovascular disease prediction slightly.
- It does not produce significant diabetes prediction.
- Internally estimated weighting makes dementia prediction worse.
- Externally derived weighting adds almost nothing to mortality prediction.
No confidence intervals or formal tests comparing the scores are presented. Nor are discrimination, calibration, prediction error or incremental value over age, sex and baseline health reported. A larger hazard ratio alone does not establish a better predictor.
3. It is not necessarily a general measure of aging
Training on mortality defines aging according to whatever predicts death in this particular population. That may be useful, but it is not outcome-neutral.
The MMSE result demonstrates the problem. Mortality weighting gives cognitive decline little weight, making the score less predictive of dementia. A score trained on dementia, disability or multimorbidity would assign different weights.
The measure is therefore better described as a mortality-weighted deterioration score than as a universal pace of aging.
4. Individual slopes may be noisy
Participants needed only one complete testing occasion to enter the study. A genuine individual slope cannot be estimated from one observation without relying almost entirely on model-based shrinkage.
Even with several observations, slope precision varies according to:
- Number of measurements
- Length of the observation period
- Measurement error
- Missed visits
- Age at observation
The second-stage survival model treats these estimated slopes as if they were measured without error. This can bias weights and understate uncertainty. A joint longitudinal-survival model would handle this more appropriately.
5. Informative dropout and terminal decline are important risks
The last assessment differs between participants. People may stop attending because of illness, cognitive decline, disability or impending death. Consequently, both the estimated slope and the chosen start of survival follow-up may depend on health.
Starting prediction at the last visit removes look-ahead bias, but it does not remove healthy-survivor bias, informative dropout or terminal-decline effects. The score may partly detect established or prodromal disease rather than a stable underlying aging rate.
6. The high-dimensional claim is speculative
The authors suggest that the method would be useful for large omics datasets where marker relevance is uncertain. This was not demonstrated.
An unpenalised mortality model containing many correlated, error-prone slopes would be unstable in high dimensions. Regularisation, dimension reduction, multivariate longitudinal modelling and external validation would be essential.
Moreover, the current application preselected biomarkers using expected age direction and monotonic mortality relationships. It therefore does not actually demonstrate completely agnostic biomarker inclusion.
7. Correlated biomarkers may destabilise weights
FEV1, FVC and FEV1/FVC are mathematically and biologically related, yet all three enter the internally trained model. Separate mixed models also ignore correlations among the ten longitudinal processes.
Multicollinearity can make individual weights unstable or cohort-specific. Weight stability under resampling is not reported.
8. Convergent validity is limited
Correlations with DunedinPACE, GrimAge and PhenoAge are only about 0.12 to 0.14. The stronger correlations with functional and frailty measures may arise partly because they use similar physical-function information.
The authors interpret weak correlations as evidence that different measures capture different dimensions of aging. That is possible, but weak correlation could also reflect measurement error, low reliability or poor construct agreement.
9. The aging-speed interpretation needs clarification
There appears to be a sign inconsistency in the manuscript. The stated regression for log survival time uses a positive linear predictor, which under the conventional accelerated failure-time formulation lengthens survival. The subsequent survival expression treats the same positive predictor as compressing survival and producing faster aging. Either a minus sign or an inverse transformation appears to be missing.
More fundamentally, a mortality time-scaling factor is not automatically equivalent to physiological years aged per chronological year. A value of 1.1 should not be interpreted literally as 10 percent faster biological aging without stronger model validation.
10. Additional statistical issues
- The survival-model descriptions do not state whether twin dependence was handled with robust standard errors or a shared frailty.
- Competing death is particularly important for dementia, diabetes and cardiovascular outcomes in this older cohort.
- Common weights are assumed for men and women despite possible sex differences in prognostic effects.
- The lack of intervention data means responsiveness to geroprotective treatment is unknown.
- The reported lack of correlation with chronological age is partly expected because age and sex trajectories were already modelled out. It is not strong independent validation.
Bottom line
| Question | Assessment |
|---|---|
| Is the weighting idea novel? | Yes, within the Pace of Aging framework |
| Is the overall statistical strategy fundamentally new? | No |
| Does weighting clearly outperform equal weighting? | Not yet demonstrated |
| Is wPoA externally validated? | No |
| Is it a literal biological aging-rate measure? | Not established |
| Is it a useful proof of concept? | Yes |
The most important finding is not that the authors have produced a superior aging clock, but that they have established a flexible framework in which the definition of aging can be made explicit through outcome-based weighting. Its value will depend on independent validation, reliable longitudinal slope estimation and careful choice of the outcome used to define “aging.”
One limitation of this review is that the supplied accepted-manuscript PDF jumps from page 32 to page 68. The cited supplementary notes, outcome definitions and Table S2 containing the individual biomarker weights are therefore absent, preventing scrutiny of the exact weights and their stability.