Echocardiography Marks Forecasting Biological Age

Introduction. It is known that chronological age may differ from biological age. Unlike chronological age, which is quite easy to determine, biological age is rather a cumulative assessment of the morphological, physiological and biochemical states of many systems of the human body. The biological age includes both a genetic component and an ontological one, considering the lifestyle of a person and the amortization of his body. The calculation of biological age is a more complex and, at the same time, a personalized approach for assessing the state of an organism compared to chronological age. There is still no single formula for determining the biological age of a person, however, attempts have been made to create local estimates, usually considering the state of an organ or system in the body. One of such local assessments may be a study of the patient's cardiovascular system, and one of the most informative instrumental diagnostics is Echocardiography. As a rule, during Echocardiography, morphological and functional changes of the heart are evaluated, indicators reflecting wall thicknesses, the volume of chambers, velocity blood flow, etc. are calculated.

The aim of the work was to create and train a predictive model using a neural network algorithm to estimate biological age based on Echocardiography data.

Material & Methods. The study included Echocardiography data from 300 patients of The Russian Clinical Research Center for Gerontology (Pirogov Russian National Research Medical University) who do not have chronic diseases. The age of patients ranged from 23–82 years in women, and 25–72 in men. Statistical analysis was performed in the statistical environment R. Echocardiography marks that have a native estimate (cm, mm, l, ml) were subject to normalization for height (cm). Spearman's correlation analysis was used as a nonparametric test to calculate the strength of the association between variables (Echocardiography marks) and age. The studied dataset was divided into training and testing sets. As a model, the architecture of a fully connected neural network (FCNN) was created using the Keras library. The architecture of the model is a deep network containing 10 hidden layers. The metrics of model quality during training were MAE (Mean Absolute Error), MSE (Mean Square Error), RMSE (Root Mean Square Error), R2 (R-squared coefficient of determination), and ε_acc (ε-accuracy — epsilon accuracy, where ε = 10, i.e. ± 10 years is the spread at accuracy assessment).

Results. The first step was an assessment of the ability of the Echocardiography data to describe an age. For this purpose, a nonparametric correlation analysis was carried out for groups of men and women separately. Out of 48 indicators, the top 15 were determined, which demonstrated a correlation with age of more than 0.55 (in abs value). Of these, the most common study protocol includes 8: E_A, IVS, LV_PW, LV_CO, LV_EDV, RWT. As well as H_RWT and L_E_A, which are calculated from RWT and E_A, respectively. Since the correlation of the selected parameters varies in its strength depending on gender (however, it retains the direction of dependence), two models were built — for men and for women. As a result, based on these parameters, we identified possible combinations and trained models based on them in three modes: general, for men, for women. According to the training results, several models demonstrated the highest accuracy of age assessment. The best models were combined into 2 two final complex models for each sex. As a result, predictive age models provide five Echocardiography marks as input data: LV_CO (cardiac output, l/min), E_A (E/A ratio of maximum flow rates in the 1st and 2nd phases), RWT (wall thickness ratio), IVS (interventricular septum thickness, cm), LV_PW (posterior wall thickness of the left ventricle, cm), and their two derivatives H_RWT (RWT ≥ 0.42) and L_E_A (E_A < 1); height (cm) and sex of the patient. The models intended for age forecasting have the following qualities: MAE = 4.92, MSE = 38.33, RMSE = 6.16, R2 = 0.78, ε_acc = 0.88, in men; MAE = 5.09, MSE = 39.42, RMSE = 6.28, R2 = 0.77, ε_acc = 0.89, in women.

Conclusions. Thus, two models for predicting age in men and women were created and trained, using five Echocardiography marks as predictors.