PhysioRisk: an interpretable survival- modeling framework for separating demographic and physiological mortality risk

Every human life follows a predictable statistical curve: the older we get, the more likely we are to die. This is the fundamental rhythm of mortality, driven by the simple passage of time. Because age is such a powerful predictor, doctors and scientists often build computer models that mix a person's age with their blood test results to guess how long they might live. These models are good at sorting people into groups of high and low risk, but they have a blind spot. When a model combines age and biology into a single number, it becomes impossible to tell if a high risk score comes from the person simply being old, or from their body aging faster than expected for their age. This confusion matters because researchers want to understand biological aging—the wear and tear on our systems—separate from the inevitable march of chronological time. If a score is too tangled with age, it cannot tell us if a treatment is actually helping a person's physiology or just reflecting the fact that they are getting older.

A researcher named Gregory Allen set out to untangle these two threads. He developed a new way to look at mortality risk that separates the expected danger of getting older from the unexpected danger of physiological dysfunction. Using a massive collection of health data from thousands of Americans gathered over nearly two decades, he built a system that first calculates the risk a person faces simply because of their age and sex. He then looked at the remaining risk, which comes from how a person's blood markers—like those measuring kidney function or inflammation—differ from what is typical for someone of their same age and sex. By keeping these two sources of risk separate, he created a framework that can show both the total picture of a person's mortality risk and the specific contribution of their biological state, without letting one obscure the other.

The study relied on data from the National Health and Nutrition Examination Survey, a long-running project that tests and interviews a representative sample of the U.S. population. Allen divided this data into two groups: an earlier group used to build the model, and a later group used to test it, ensuring the findings held up over time. He started by creating a baseline score, which he called DemoRisk, based entirely on age and sex. This score captured the vast majority of the ability to predict who would die, simply because older people are statistically more likely to die than younger ones. Next, he took the blood test results and stripped away the patterns that were already explained by age and sex. For example, if kidney function naturally declines as people get older, the model removed that expected decline. What was left were the deviations: the people whose blood work was worse or better than the average for their specific age and sex.

These remaining deviations were organized into two main categories. One category, which the researcher named HematologicStress, grouped together markers related to blood cells and the immune system. The other, called RenalHepaticTurnover, focused on markers related to kidney and liver function. By combining these two categories, he created a second score, PhysioRisk. This score was designed to be almost completely independent of age; in the study, it showed almost no correlation with how many years a person had lived. Despite being stripped of age, this physiological score still carried important information. People with higher PhysioRisk scores were more likely to die during the study period than those with lower scores, even when the researchers compared people of the same age and sex. This proved that biological dysfunction contains its own signal for mortality, distinct from the signal of simply growing older.

The most significant finding came when the researcher recombined these two separate scores. He took the age-based risk and the physiological risk and added them together using specific weights determined from the initial data. The resulting combined score, which he called CalibratedTotalRisk, performed just as well as the best existing models for predicting death. In fact, it was slightly better than using age alone. This demonstrated that the physiological component added real value to the prediction, improving the ability to rank people by risk without losing the power of the age-based component. The combined model was also compared to a well-known existing measure called PhenoAge, which attempts to estimate biological age. The new combined score performed similarly to PhenoAge, suggesting that separating the demographic and physiological parts did not hurt the overall accuracy of the prediction.

However, the study also revealed a subtle complexity in how these risks behave over time. While the age-based part of the model behaved consistently, the combined score showed that the link between the risk factors and death became stronger the longer people were followed. This means that the difference in risk between a high-scoring and low-scoring person grew wider as time went on, rather than staying constant. This finding suggests that the body's physiological state might have a delayed effect on mortality that becomes more apparent over many years. The researchers noted that this behavior did not fit a standard statistical assumption that risk ratios stay the same over time, indicating that the relationship between biology and death is more dynamic than simple models often assume.

The work does not claim to have found a perfect measure of biological age or a magic bullet for extending life. The physiological score is not a direct measure of how "old" a person's body is in a biological sense, nor does it prove that the specific blood markers cause death. Instead, it offers a clearer way to look at the data. By separating the risk of being old from the risk of being sick, the model allows researchers to ask different questions. They can now see if a person's physiology is deteriorating faster than their chronological age would suggest, independent of the general trend of aging. This separation is valuable because it prevents the overwhelming signal of age from drowning out the more subtle signals of biological health.

The study was conducted using data from a single country and relied on a specific set of blood tests, so the results need to be tested in other populations before they can be applied broadly. The researchers also noted that their model was built on people who had complete blood test data, which might exclude the sickest or most vulnerable individuals who often miss medical appointments. Despite these limitations, the approach provides a new tool for understanding mortality. It shows that it is possible to build a model that respects the power of age while still isolating the unique contribution of physiological health. By keeping the two sources of risk visible and separate, the framework offers a more nuanced view of human aging, one that acknowledges that while we all grow older, our bodies do not necessarily age at the same rate or in the same way.