This is How Machine Learning Can Measures Your Rate Of Aging

Using various aging markers, machine learning can estimate physiological age.

This is How Machine Learning Can Measures Your Rate Of Aging

Looking at the dates on the calendar and observing the Earth's orbit around its axis and the sun makes it appear that we are aging at a constant rate. After all, another day in life is a day of growth.

Thus, our chronological age is measured by the ticking of the clock and the passage of days. But we also have a biological age, which doesn't always exactly match our clocks and calendars.

Of course, we can't reverse age (yet?). But biologically, not everyone ages the same way. You may be younger or older than your chronological age.

Since we cannot rely on clocks and calendars to measure our biological age, we have to look for other aging times. Many recent studies have focused on tracking internal biological signals that correlate with various aging processes in our body.

For example, blood protein levels can be used to quantify biological age or track inflammation. Recent methods of determining biological age include several hours of epigenetic aging. The dynamics of epigenetic tags that can attach to DNA and alter gene activity are thought to reflect various aspects of aging. (These epigenetic clocks have been used, for example, to study embryonic development or to figure out why some animals appear to be resistant to aging.)

There are many different aging watches, so it might be worth considering combining them. This is already  done with a different epigenetic clock, but what if we go up to one level and look at multiple physiological variables like blood markers, cardiovascular function, etc.? (Of course, epigenetic changes can affect all of this.)

Well, the new study begins by claiming:

"…few biological age measurements integrate traits from distinct hierarchical levels (e.g., proteome, metabolome, and phenome) and across the multiple organ systems in which aging is manifest."

If you read something like this in the article preface, you can be sure that the author will try something about it. In this case, the scientists used the data sets SardiNIA (> 6000, ages 14-101, mean  43.7 years, 57 men) and InCHIANTI (> 1400, age 21-102, mean  69 years). , 56 - male), both of which contain a  set of longitudinal meanings for many traits (blood indicators, cardiovascular function, personality tests, etc.). Using much of this data as training data, the researchers applied several machine learning models to isolate traits that were highly correlated with biological aging. We then asked the model to predict the ages of people in the data set. Averaging this over many people will give you the most accurate model. (I am assuming, on average, than the predicted physiological age and the actual age are not significantly different.

For large data sets, this seems to be a reasonable assumption. You can later apply a better model to the individual and see how  they deviate from the mean in terms of  biological age.) The most efficient model is a random forest classifier. This is a technique that relies on creating multiple decision trees for each prediction and then looking at the "forest" of all the decision trees. You can have a good predictive model when many trees look the same. In the end, the researchers found 52 properties for cholesterol, from blood counts  to blood pressure and pulse wave velocity. Overall, cardiovascular disease risk indicators had the greatest impact. These properties allowed researchers to quantify the physiological rate of aging  (PAR).

Here are some interesting correlations:

PAR is associated with mortality (obviously).

PAR correlates quite well with how many hours of epigenetic aging (thus epigenetic change seems to be very well reduced or increased -).

PAR was moderately genetic. Of course, this may be due to the common environment, common habits of parents and children, etc.

However, the scientists also had genetic data to study and found several single-letter DNA alterations (SNPs) associated with  PAR. (Most of them are in the CFI and GAR1 genes, and some are in  the LINC00202 region or in the vicinity of those encoding long non-coding RNAs that affect gene activity.) But despite genetic factors,  the researchers cautiously point out: 

"The majority of PAR is still attributable to environmental factors, including the effect of pathology."


Our current study showed that PAR is an important survival rate and correlates with epigenetic aging rates, providing evidence for an accurate measure of aging.