Do Aging Organs Compound Each Other's Risk? A Measurement-Corrected Test of Multi-Organ Synergy

Aging clocks are models that estimate how "old" a person's body looks from blood proteins, metabolites, or scans, and then report the gap between that estimate and calendar age. They now exist for individual organs: brain, heart, kidney, liver, pancreas and others. A popular theory of aging holds that death accelerates with age because failing parts of the body strain one another, so one weak organ makes a second weak organ more dangerous. If that is true, a person with both an old heart and an old kidney should carry more risk than the two gaps would predict separately. No one has tested this properly. Existing multi-organ studies either add the organ gaps together or sort people into a few coarse groups. Both choices rule out, by construction, the compounding effect the theory predicts.

We propose to test for that compounding directly by adding organ-by-organ product terms in a final block of a nested regression, after chronological age, clinical labs, and each organ's own effect have been entered. The naive version of this test is easy to run and easy to get wrong in two opposite directions. A product term can look significant only because risk curves upward in a single overall "aging" factor, which is a false positive. It can also be shrunk toward zero because clock gaps are noisy, which is a false negative. The contribution here is a test that closes both holes. The null model includes curvature and a shared aging factor. The interaction estimate is corrected for noise by using a second, independent measurement of the same organ from a different data type. The result should settle, in either direction, whether multi-organ aging compounds risk.

Core Claim

Pairwise synergy between organ age gaps, if it exists, is detectable only when three conditions hold. First, the null model already includes a common aging factor and each gap's own nonlinear effect. Second, the interaction is judged on the scale the theory actually concerns. Third, the interaction estimate is corrected for the attenuation that clock noise causes. Under these conditions we predict positive interactions for the organ pairs that structural equation modeling in Imaging-based organ-specific aging clock predicts human diseases and mortality identified as linked, namely heart with kidney and heart with pancreas. We predict null interactions for pairs with no such link. A clean null across the board would also be informative. It would mean that additive multi-organ risk scores lose nothing, and that the interdependence theory does not need pairwise organ synergy at the resolution clocks can measure.

Why the Scale and the Null Model Matter

The motivating theory comes from The Gompertz Law emerges naturally from the inter-dependencies between sub-components in complex organisms. That paper shows that exponential growth of mortality with age can arise from cascading failures among interdependent subsystems. It fits the Danish mortality registry with roughly $10^6$ subsystems. That scale is far finer than seven or eleven organs, and this constrains our proposal. Pairwise synergy at the level of organs is a hypothesis the theory makes plausible, not one it requires. A proportional-hazards model with additive organ gaps already makes risks multiply, because the gaps add on the log-hazard scale. "Synergy" must therefore mean something beyond that. We define it as a positive coefficient on the product term in

$$\log h_i(t) = \log h_0(t) + f(\text{age}_i) + \sum_k \big(\beta_k g_{ik} + \gamma_k g_{ik}^2\big) + \delta\, G_i + \theta_{jk}\, g_{ij} g_{ik},$$

where $h_i(t)$ is person $i$'s hazard at time $t$ and $h_0(t)$ is the baseline hazard. The function $f$ is a spline in chronological age. The term $g_{ik}$ is the standardized, age-residualized age gap of organ $k$. The quantity $G_i$ is the first principal component of all gaps, which we call the "general aging factor". The coefficient $\theta_{jk}$ is the synergy coefficient for organs $j$ and $k$.

We report $\theta_{jk}$ alongside the additive-scale interaction, meaning the excess absolute risk when both organs are old. That second measure is the clinically meaningful one. It can be positive even when $\theta_{jk}=0$.

The $\gamma_k$ and $\delta$ terms are the part most naive analyses omit. If organ gaps are positively correlated and risk curves upward in overall frailty, then $g_{ij} g_{ik}$ partly stands in for $G_i^2$. That stand-in would be read as synergy.

Correcting for Clock Noise

Measurement error attenuates interaction effects more severely than main effects, a limitation the research file notes for hierarchical regression in general. For centered, roughly independent errors, the reliability of a product is approximately the product of the two reliabilities. Clocks with reliability near 0.5 therefore give a product term with reliability near 0.25. The metabolome clocks in Multi-organ metabolome biological age implicates cardiometabolic conditions and mortality risk correlate with chronological age at only r = 0.25 to 0.42 on test data. Their gaps are therefore dominated by variation of uncertain origin.

Our correction uses a cross-modality replicate. In UK Biobank participants who have both organ MRI and Olink plasma proteomics, one organ is measured twice by independent instruments. The MRI-based gaps come from Multi-organ MRI digitizes biological aging clocks across proteomics, metabolomics, and genetics. The proteomic organ gaps come from the Oh et al. models that were applied in Biological organ ages associate with risk of chronic diseases in a community-based population. The two measurements share the organ's true aging state, but not the scanner noise or the assay noise. This lets us estimate each clock's reliability. We then apply regression calibration to the main effects and product terms, with a simulation-extrapolation estimator (which adds extra noise and extrapolates back to zero noise) as a check.

This mechanism is the key new ingredient. Without it, a null result cannot be interpreted. With it, a null result means something.

Research Program

We will build three things. The first is a harmonized set of age-residualized organ gaps in UK Biobank from MRI and proteomics, with within-person reliability estimated for each organ that both modalities cover. The second is a pre-registered nested analysis, run separately for incident heart failure, end-stage kidney disease, dementia, and all-cause mortality. The third is an external replication in the ARIC cohort. ARIC has 11 proteomic organ clocks, 11,757 participants, and 26 years of follow-up. It is a different population and a different follow-up horizon, and it tests whether any synergy survives a change of cohort.

The blocks are as follows. Block 1 contains chronological age as a spline, sex, and routine clinical measures (eGFR, lipids, blood pressure, HbA1c, hs-CRP). Block 2 adds each organ gap with its square, plus the general aging factor. Block 3 adds the pre-registered product terms, first for Ren's hub pairs and then for all remaining pairs with false-discovery-rate control.

Blockwise entry here is simply a disciplined way to run nested likelihood-ratio tests. We claim no novelty for the procedure itself. The novelty lies in what goes into Block 2 and in the measurement correction.

Baselines

The do-nothing reference is Block 1 alone: chronological age and routine clinical labs. It answers the question raised by Is age an independent determinant of mortality in cardiac surgery as suggested by the EuroSCORE?. That paper found that calendar age carries little independent signal once organ decline and comorbidity are captured. Every clock-based model must first add to this reference before interactions are worth discussing.

The strongest existing competitor is the additive multi-organ model. It enters all organ gaps as main effects in a Cox model, as done by Ren et al. and by Zheng et al. For mortality, Zheng et al. report an average concordance index of 0.72 for MetBAGs combined with polygenic risk scores. Ren et al. report an AUC of 0.82 for incident dementia from the gray-matter clock. We will refit the additive model on exactly the same participants, gaps, and covariates as our Block 3 model, with the same cross-validation folds.

We do not expect interaction terms to move the concordance index much. Beating the additive model therefore means three things. The Block 3 likelihood-ratio test must be significant after correction. The corrected $\theta$ for the hub pairs must exclude zero. Net reclassification among people with two old organs must improve. A gain in C-index of about 0.005 or more would be a bonus, not the success criterion.

The categorical competitor is the three-cluster phenotype of Liu et al., who report a death hazard ratio of 1.75 for the "older organs" cluster versus the "younger organs" cluster. We will reproduce the clustering in our data and compare fit with the likelihood-based information criterion.

The flexible competitor is gradient-boosted survival trees on the same gaps, with interaction strength read from SHAP interaction values. It receives an equal tuning budget of 100 random-search configurations per outcome. No reported result was found in the file for boosted models on organ gaps.

Two ablations isolate the key ingredients. One drops the measurement correction. The other drops the $\gamma_k$ and $\delta$ terms. They show how much of any naive "synergy" is attenuation or curvature in disguise.

Why not just…?

Why not just add a single whole-body age gap? This is the most serious threat. If organ gaps mostly reflect one shared frailty factor, a single score captures nearly everything, and pairwise terms are a distraction. The design tests this directly, because the general factor and its square sit in Block 2. If Block 3 adds nothing beyond them, we will report that the single-score view wins. That outcome would be useful for clinicians who want one number.

Why not just use clusters, as Liu et al. did? Clustering can reveal that people with many old organs fare badly, and Liu et al. report an ESKD hazard ratio of 6.12 for the oldest cluster. It cannot say whether that risk is additive or compounding, because the cluster label mixes both. The information-criterion comparison against continuous models will show whether discretizing costs fit. If clusters fit as well, we will say so.

Why not let a boosted tree or neural network find interactions automatically? Flexible learners do detect interactions, but they detect curvature too. They offer no calibrated test that separates the two, and they do not correct for noise in their inputs. Ludwig et al., in A multimodal retinal aging clock for biological age prediction and systemic health assessment via OCT and fundus imaging, show that flexible clocks can fail badly outside their training distribution: R² of −1.74 on eyes with macular degeneration. That result concerns clock construction rather than downstream risk modeling, so it is only indirect evidence. The honest test is empirical. If boosting's SHAP interactions point to the same organ pairs, and it outperforms our model on held-out likelihood, the flexible route is preferable. We will run exactly that comparison.

Why not just add product terms to a standard Cox model? That is literally what Block 3 does. Without the Block 2 curvature terms and the correction, however, the coefficient means little. The two ablations will show how far naive estimates drift from the corrected ones.

Proposed Experiments and Decisive Checks

The first decisive check is feasibility. We will count the UK Biobank participants who have both organ MRI and proteomics, and estimate cross-modality reliability for the heart, kidney, liver, and pancreas. If any organ's reliability is below about 0.3, its corrected interactions will be too unstable to report, and we will restrict the analysis to organs above that threshold.

The second check is a simulation. We will build synthetic cohorts with a known $\theta$, a known general factor, and known reliabilities, and confirm that the corrected estimator recovers $\theta$ while the naive one does not.

The third check is the main test. The corrected hub-pair interactions (heart with kidney for kidney failure, heart with pancreas for mortality) must be positive and must replicate in the direction of effect in ARIC. Pre-registered non-hub pairs must stay null. Synergy that appears everywhere would suggest residual confounding by overall frailty rather than organ-specific interdependence.

The fourth check is a negative control. We will shuffle one organ's gap within strata of chronological age and the general factor. This breaks true pairing while keeping each gap's distribution, so any significant interaction that survives the shuffle marks a flaw in the method.