Application of deep neural network and Klemera-Doubal method in estimating biological age of middle-aged and elderly residents in China

Abstract

ObjectiveTo investigate the performance of deep neural network (DNN) and Klemera-Doubal method (KDM) in assessment of biological age (BA) among middle-aged and elderly Chinese population and to select the optimal scheme for evaluating the accuracy of two methods in BA estimation. MethodsThe study data on people aged 45 – 85 years were extracted from the baseline survey of the China Health and Retirement Longitudinal Study (CHARLS) conducted in 2011/2012. The extracted data (including blood marker samples from 11 513 people, physical measurements from 13 603 people, and both blood marker samples and physical measurements from 9 904 people) were used to estimate BA by DNN and KDM. Pearson correlation coefficient, mean absolute error (MAE) and root mean square error (RMSE) were adopted in the evaluation on accuracy of BA estimation and the selection of optimal scheme. Complex sampling logistic regression model and linear regression model were constructed using calculated residual (ΔBA) of regression analysis of estimated BA and chronological age (CA), mortality data, and cognitive function assessment of the people. Receiver operating characteristics (ROC) area under curve (AUC) was used to evaluate the performance of ΔBA in predicting mortality risk. ResultsCompared with the estimations based only on blood marker samples or physical measurements, the DNN- and KDM-derived BA estimations based on blood marker sample and physical measurement showed the highest correlation with CA (rDNN = 0.91, rKDM = 0.48), but the lowest MAE (MAEDNN = 3.11, MAEKDM = 13.74) and RMSE (RMSEDNN = 4.15, RMSEKDM = 17.76). The accuracy of DNN-derived BA estimation was better than that of KDM-derived estimation. The mortality risk increased but the scores of overall cognition, memory and mental status decreased with the increment of DNN-derived ΔBA and KDM-derived ΔBA based on blood marker samples and physical measurements and of CA. KDM-derived ΔBA showed a slightly higher accuracy in mortality risk prediction compared to DNN-derived ΔBA (AUCKDM-ΔBA = 0.717; AUCDNN-ΔBA = 0.699). Compared with those having normal aging (with the ΔBA of ≥ – 5 and ≤ 5), the people having accelerated aging (with the ΔBA of > 5) were at an increased risk of mortality (odds ratio [OR]DNN = 1.760, 95% confidence interval [95%CI]: 1.207 – 2.567; ORKDM = 1.872, 95%CI: 1.308 – 2.679) and having a low score of overall cognition (βDNN = – 0.631, 95%CI: – 1.124 – – 0.139; βKDM = – 0.374, 95%CI: – 0.680 – – 0.068). ConclusionThe predictive BA could be more accurate when derived from blood marker sample and physical measurement than that from either one of the two samples. The BA estimated by DNN modeling method is of higher accuracy than that by KDM but the performance in BA prediction of the two methods is similar. The results suggest that DNN modeling method could be used in evaluating individual aging among middle-aged and elderly Chinese populations.

Keywords

• biological age
• deep neural network
• klemera-doubal method
• aging