Population Specific Biomarkers of Human Aging: A Big Data Study Using South Korean, Canadian, and Eastern European Patient Populations

Abstract

Accurate and physiologically meaningful biomarkers for human aging are key to assessing antiaging therapies. Given ethnic differences in health, diet, lifestyle, behavior, environmental exposures, and even average rate of biological aging, it stands to reason that aging clocks trained on datasets obtained from specific ethnic populations are more likely to account for these potential confounding factors, resulting in an enhanced capacity to predict chronological age and quantify biological age. Here, we present a deep learning-based hematological aging clock modeled using the large combined dataset of Canadian, South Korean, and Eastern European population blood samples that show increased predictive accuracy in individual populations compared to population specific hematologic aging clocks. The performance of models was also evaluated on publicly available samples of the American population from the National Health and Nutrition Examination Survey (NHANES). In addition, we explored the association between age predicted by both population specific and combined hematological clocks and all-cause mortality. Overall, this study suggests (a) the population specificity of aging patterns and (b) hematologic clocks predicts all-cause mortality. The proposed models were added to the freely-available Aging.AI system expanding the range of tools for analysis of human aging.

Figure 1.

Study design. First, blood samples of three populations (Canadian, Korean and Eastern European) with 21 the most relevant features with maximum samples available were used to train three population specific predictors. Afterwards, the resulting dataset consisting of samples from all three populations was used to train and test DNNs for predicting patient age.

Figure 2.

Actual chronological age vs predicted age for Canadian (A), Korean (B), and European (C) populations of patients. The linear regression line is shown in dark grey. Log₂transformed Aging ratio for for Canadian (D), Korean (E), and European (F) population predictions. Log₂Aging ratio of 1 means that sample is predicted twice older than a chronological age and Log₂Aging ratio of −1 means sample is predicted half as old.

Figure 3.

(A) Actual chronological age vs predicted age for the resulting network trained and tested on the all three populations. The linear regression line is shown in dark grey. (B) Log₂ transformed aging ratio. Log₂Aging ratio of 1 means that sample is predicted twice older than a chronological age and Log₂Aging ratio of −1 means sample is predicted half as old.

Figure 4.

Actual chronological age vs predicted age for (A) Canadian, (B) Korean, and (C) European patient populations tested on the network trained on all population samples. Linear regression lines are shown in dark grey. Log_2, transformed aging ratio for (D) Canadian, (E) Korean, and (F) European populations tested on the network trained on all population samples. Log₂Aging ratio of 1 means that sample is predicted twice older than a chronological age and Log₂Aging ratio of −1 means sample is predicted half as old.

Figure 5.

Feature importance plots of the model trained on (A) Canadian population samples, (B) on Korean population samples, and (C) on Eastern European population samples. Permutation feature importance (PFI) method was used to rank blood markers and sex by their importance in age prediction. (D) The top seven most important features across all predictors trained on different populations. Albumin, sex, hemoglobin, and urea are ranked as the most important markers for age prediction in all three models; (E) the most important markers for the network trained on the three populations. Albumin, glucose, and erythrocyte count were ranked as the most markers for age prediction in this model. PFI method was applied to rank blood markers, sex and population by their importance in age prediction.

Figure 6.

Validation of models. Actual chronological age vs predicted age for NHANES dataset using networks trained on Canadian (A), Korean (B), European (C), and (D) all patient population samples. The linear regression lines are shown in dark grey. Networks trained on both E. European and all patient samples demonstrated the higher accuracy of age prediction of NHANES dataset. (E) Hazard ratios for the NHANES and Canada datasets. A Cox proportional hazards regression model was used to relate survival time to the accelerated aging group (delta >5) and slowed aging group (delta <5). Patients predicted younger their chronological age has a lower mortality risk, while patients predicted older has a higher risk. Each row represents a hazard ratio and 95% confidence interval. Note: “∗∗∗” for p-value of .001; “∗∗” for p-value of .01; “∗” for p-value of .05.