The current best candidates for a sufficiently robust biomarker of aging are based on patterns of DNA methylation, epigenetic markers that control the pace at which specific proteins are produced, and which are constantly shifting in response to circumstances. The best of these epigenetic clocks have degrees of error in assessed age that are five years or less, depending on implementation. The researchers here have chosen to investigate patterns of protein levels in blood rather than epigenetic markers, in part driven by economic considerations, as the needed tools of biotechnology are more mature and less expensive. They use modern computational techniques to try to build useful biomarker algorithms through an analysis of raw data obtained from large numbers of people at various ages. Their efforts result in a degree of error of around six years, which might be taken as encouraging; it may well be possible to do better via this method.
To perform this study, we trained a series of deep neural networks on anonymized blood tests for patients from three distinct ethnic populations: Korean, Canadian, and Eastern European. We compared the predictive accuracy of our deep learning models first when trained using population-specific data, and then when using a combined and ethnically-diverse dataset that includes patients from all three patient populations. We used the same feature space of 20 blood biochemistry markers, cell counts, and sex to train three separate deep networks on three specific ethnic populations.
We present several novel hematological aging clocks. The best-performing predictor achieved a mean absolute error (MAE) of 5.94 years having greater predictive accuracy than the best-performing predictor of our previously-reported aging clock (which achieved an MAE of 6.07 years), despite being trained on a narrower feature space (21 compared to 41 features). These results are in line with the hypothesis that ethnically-diverse aging clocks have the potential to predict chronological age and quantify biological age with greater accuracy than generic aging clocks. Furthermore, they have a greater capacity to account for the confounding effect of ethnic, geographic, behavioral and environmental factors upon the prediction of chronological age and the measurement of biological age.
Albumin, glucose, urea, and hemoglobin were among the most important blood biochemistry parameters for all three population-specific predictors. Albumin is the most prevalent protein in blood and its primary function is the regulation of oncotic pressure, which is critical for transcapillary fluid dynamics, and hypoalbuminemia is often associated with malnutrition, liver disease, injury, chronic inflammation and the aging process. Blood glucose levels, on the contrary, tend to increase with age, and glucose is able to modify proteins via irreversible glycosylation, a feature that is directly associated with the aging process. Levels of serum urea also increase with age, which is associated with age-related decrease in muscle mass. Age-related decreases in hemoglobin is common in the elderly, a condition that increases the risk of cardiovascular disease, cognitive decline and an overall decline in quality of life.
Our hematological clock is consistent with what is already known about the biology and pathophysiology of aging. While the blood parameters are not accurate biomarkers of aging by themselves, when analyzed in combination they can be used to reasonably accurately predict chronological and biological age. Deep learning based hematological aging clocks, even when trained on a limited feature space, demonstrate reasonably high accuracy in predicting chronological age.