🧬 biology

Does AI-predicted biochemical age predict death? A cross-validated re-examination in NHANES 1999–2018 with linked mortality follow-up

This study demonstrates that while complex machine learning models can more accurately predict chronological age from blood chemistry, their resulting "age acceleration" metrics are less effective at predicting mortality than simpler models or biomarkers directly trained on death outcomes, revealing that the training target rather than the model complexity or biomarkers is the critical factor in developing useful biological age predictors.

Original authors: John Feng

Published 2026-08-06

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: John Feng

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to guess how fast a car is wearing out just by looking at the odometer. In the world of aging science, there is a popular idea called "biological age." The theory is simple: two people might be the same age in years (chronological age), but one might feel and function like they are much older or younger due to how their body is actually aging. Scientists have been building "aging clocks" to measure this. Usually, they do this by training a computer to look at a person's blood test results and guess their age. If the computer guesses you are older than you really are, the idea is that you are aging faster than normal. This "extra" age is called "age acceleration," and many researchers hope it can predict who will get sick or die sooner. But here is the big question: Does a computer that is really good at guessing your age actually tell us anything useful about how long you will live?

This paper, written by researcher John Feng, dives right into that question using a massive dataset of blood tests from over 45,000 American adults. The author set out to test a specific claim: that the "mistakes" a computer makes when guessing your age (predicting you are 50 when you are actually 45) are a sign of how quickly you are dying. To do this, the study used a giant pool of data from the National Health and Nutrition Examination Survey (NHANES), tracking blood chemistry and seeing who passed away over the next two decades. The results are a bit of a plot twist for the scientific community. The study found that while fancy AI models are excellent at guessing chronological age, the "age acceleration" they produce is actually a terrible predictor of death. In fact, the better the AI gets at guessing your age, the less useful its prediction becomes for telling you if you are at risk of dying. The real secret to predicting death wasn't in the "mistakes" of the age-guessing models, but in training the computer to look directly for signs of death in the blood.

The Great Age-Guessing Experiment

Think of the blood tests in this study as a 38-piece puzzle. The researchers took these puzzle pieces—things like cholesterol, blood sugar, and white blood cell counts—and asked a computer to solve a specific riddle: "Based on these numbers, how old is this person?" They tried six different types of computer brains, ranging from simple math (like a straight line) to complex, flexible neural networks (like a deep-thinking AI).

The first part of the story is a victory for the complex AI. When the researchers asked, "Who can guess the age best?" the flexible neural networks and gradient boosting models won hands down. They were much better at predicting age than the simple linear models. The best neural network got an accuracy score (called an R²) of 0.557, meaning it could explain about 56% of the variation in age. The simple models only managed about 40%. If you were just looking for a computer that could guess your birthday from a blood test, the complex AI would be your champion.

The Twist: Being Good at Guessing Age is Bad for Predicting Death

Here is where the story takes a sharp turn. The researchers then took the "mistakes" from these age-guessing computers. If the computer said you were 50 but you were actually 45, that +5 difference is your "age acceleration." The big hope was that this number would tell us who is dying sooner.

The results were surprising. The computer models that were best at guessing age turned out to be the worst at predicting death.

The simple models (like ridge regression), which were less accurate at guessing age, produced an "age acceleration" score that improved the prediction of death by a tiny bit (adding about 0.0124 to a statistical score called the C-index).
The super-smart neural networks, which were the best at guessing age, produced a score that added almost nothing to the prediction of death (only 0.0056).

The study found a strong negative link: the more accurate the age prediction, the less useful the "age acceleration" was for predicting mortality. It's as if the complex AI was so busy memorizing the tiny, specific details of how blood changes with age that it accidentally filtered out the very signals that actually signal danger. The paper suggests that by trying to be perfect at guessing age, the AI threw away the information that matters for survival.

The Real Winner: Training for the Right Goal

To prove that the blood data did contain life-or-death information, the researchers ran a second experiment. Instead of asking the computer, "How old is this person?" they asked, "Who is likely to die?" They trained a new model using the exact same 38 blood tests, but this time the computer learned to predict death directly.

The result? This "death-trained" model was a massive improvement. It boosted the prediction of death by 0.0367—more than three times better than the best age-trained model. Even more telling, when they added this death-trained score to a model that already included the "age acceleration" from the AI, the death-trained score still added a huge amount of new information. This proved that the age-trained models hadn't just missed the signal; they had actively discarded it.

The paper also looked at a famous existing clock called "PhenoAge," which was built to predict death. It performed better than the age-trained models but still wasn't as good as the new death-trained model built from the same blood tests. This suggests that even the best current tools are leaving a lot of valuable information on the table because they are trained on the wrong target.

Why the Numbers Can Be Tricky

The study also solved a mystery about why some previous papers claimed their age-guessing models were incredibly accurate (with scores as high as 0.78), while this study's best was only 0.557. The answer wasn't that the new AI was weaker; it was about the crowd of people being tested.

The researchers found that the accuracy of age-guessing depends heavily on the age range of the group. If you include teenagers, whose bodies are changing rapidly, the computer gets much better at guessing age because the blood changes are so dramatic. If you only look at adults between 20 and 79, the changes are slower and harder to spot. The study showed that the "accuracy score" is mostly a reflection of how wide the age range is in the group, not how good the computer is. This means you can't just compare the accuracy scores of different studies to see which clock is better; you have to look at what they are actually trying to predict.

The Bottom Line

The main takeaway from this paper is a lesson in focus. If you want to build a tool to predict who might die sooner, don't train a computer to guess their age and hope the "mistakes" tell you something. That approach throws away most of the useful information. Instead, train the computer directly on the outcome you care about: death. The paper concludes that the bottleneck isn't the blood tests or the AI technology; it's the goal we set for the AI. By chasing the wrong target (chronological age), the field has been missing the most powerful signals for predicting mortality that are hiding right there in the blood.

Technical Summary: Does AI-predicted biochemical age predict death?

Problem Statement
The field of biological aging often relies on "aging clocks" trained to predict chronological age from biological substrates (e.g., blood chemistry, DNA methylation). The underlying assumption is that the residual error of such a model—termed "age acceleration" (predicted age minus actual age)—serves as a proxy for biological aging and predicts mortality risk beyond chronological age. However, many studies validate these clocks solely by reporting the coefficient of determination ( $R^2$ ) against chronological age, without testing whether the residuals actually predict death. This paper investigates whether the pursuit of higher chronological age prediction accuracy inadvertently discards the very signal required for mortality prediction.

Methodology
The study utilized data from ten cycles of the National Health and Nutrition Examination Survey (NHANES, 1999–2018), linked to the NCHS public-use mortality file. The analytic sample comprised 45,033 adults (aged 20–79) with 5,213 deaths over 445,961 person-years.

Inputs: 38 analytes (23 serum chemistry, 15 hematology) and sex.
Model Families: Six families of models were fitted to predict chronological age under nested cross-validation (5-fold outer, 3-fold inner):
1. Ordinary Least Squares (OLS)
2. Ridge Regression
3. Principal Component Regression (PCA + Ridge)
4. k-Nearest Neighbors (k-NN)
5. Histogram Gradient Boosting
6. Multilayer Perceptron (Neural Network)
Architecture Sweep: To ensure neural networks were not underpowered, a random sample of 360 networks was trained across a 1,296-point factorial grid of hyperparameters (depth, activation, penalties, learning rates).
Age Acceleration Definition: Following neuroimaging conventions, age acceleration was defined as the residual from a linear regression of the out-of-fold predicted age on chronological age (to remove correlation with age).
Evaluation: Models were evaluated on all-cause mortality using Cox proportional hazards models adjusted for chronological age and sex. Performance was measured by Harrell's C-index (discrimination) and Hazard Ratios (HR).
Controls:
- PhenoAge: A published mortality-trained clock (using 9 markers) was computed as a benchmark.
- Mortality-Trained Score: A ridge-penalized Cox model was fitted directly to the same 38 analytes to predict death, serving as an upper bound for the information available in the data.

Key Results

Accuracy vs. Utility Inversion: Flexible models (Gradient Boosting, Neural Networks) significantly outperformed linear models in predicting chronological age. The out-of-fold $R^2$ increased from 0.398 (Ridge) to 0.557 (Neural Network). However, the utility of the resulting age acceleration for mortality prediction moved in the opposite direction.
- Ridge regression (lowest accuracy) yielded the highest mortality discrimination gain: $\Delta C = +0.0124$ .
- Neural networks (highest accuracy) yielded the lowest gain: $\Delta C = +0.0056$ .
- The correlation between age-prediction $R^2$ and C-index gain across model families was $-0.94 $($ p=0.005$).
The Ceiling of Training Targets: The same 38 analytes, when fitted directly against mortality rather than age, produced a score with a C-index gain of $+0.0367$ (3.0 times the best age-trained residual). This mortality-trained score retained 85% of its predictive power even after age acceleration was already included in the model, indicating that age-trained residuals fail to capture the majority of the mortality signal present in the blood panel.
Divergent Signals: The analytes most strongly correlated with chronological age were largely orthogonal to those predicting mortality. For example, Blood Urea Nitrogen was the top age-correlated marker but ranked 30th for mortality; Albumin was the top mortality marker but only 15th for age. The rank correlation between age-correlation and mortality hazard magnitude was near zero ( $+0.02$ ).
Sample Composition Artifact: The reported $R^2$ values for age prediction were found to be heavily dependent on the age window of the sample. Varying the lower age bound from 20 to 12 (including adolescents) raised the $R^2$ from 0.557 to 0.683, while narrowing the window to 20–45 dropped it to 0.247. This suggests that high $R^2$ values in literature often reflect sample heterogeneity rather than model superiority.

Key Contributions

Empirical Validation of Theoretical Concerns: The study provides direct empirical evidence that maximizing chronological age prediction accuracy is counterproductive for generating mortality-predictive biomarkers.
Decoupling Accuracy and Utility: It demonstrates that within a single model class (e.g., linear models tuned by regularization), accuracy and utility rise together, but across model classes, the move to flexible learners captures age-specific structures that are prognostically inert, thereby removing signal from the residual.
Training Target Dominance: It establishes that the training target (age vs. mortality) is the primary determinant of a biomarker's utility, not the data quality or model class.
Re-evaluation of Metrics: It argues that $R^2$ against chronological age is a property of the sample's age distribution, not a valid metric for comparing aging clocks across studies or assessing their clinical value.

Significance and Claims
The paper concludes that fitting blood chemistry to chronological age is a valid regression problem for describing how biochemistry changes with age, but it is a "poor way to build a mortality biomarker." The residual from such a fit discards the majority of the mortality-relevant signal.

The author asserts that:

Prognostic clocks must be trained on prognosis: Models like PhenoAge and GrimAge perform better because they are trained directly on mortality outcomes, not chronological age.
Current evaluation metrics are misleading: Reporting $R^2$ against chronological age hides the fact that the most accurate models yield the least useful residuals.
Clinical value is low for age-trained residuals: The incremental discrimination provided by age acceleration from modern AI models is negligible ( $\Delta C \approx 0.005$ ) compared to the potential available in the same data ( $\Delta C \approx 0.037$ ).

The study does not claim that biological age is an empty concept, nor that chronological-age models are useless for describing physiological changes. Rather, it argues that for the specific purpose of predicting death, the "age-trained residual" approach is systematically suboptimal because the objective function (minimizing age error) is orthogonal to the desired outcome (predicting mortality).