Assessment of the bone biological age using multiple regression and neural network analysis

Abstract

Background. Nowadays, many studies confirm the increased risk of osteoporosis in the elderly. Strategies for optimizing diagnosis include a combined use of various methods, so calculating bone biological age (BA) can also be important for disease prediction. Recently, some new methodological approaches to BA calculation with the mathematical methods implementation were proposed. However, studies utilizing advanced approaches, particularly neural network (NN) in Ukraine, are limited. The purpose of this study was to develop a bone BA model and compare its accuracy using stepwise multiple regression (MLR) and NN analysis methods. Materials and methods. In a single-center cohort study, we retrospectively analyzed the data of 3,071 healthy women aged 40–90 years. The analysis of the study indices included the anthropometric parameters, dual-energy X-ray absorptiometry indices (DXA), and the parameters of the 10-year probability of major osteoporotic fractures (FRAX). For the development of bone BA models, MLR methods were used and the construction of the NN model was performed using a multilayer perceptron model. Results. As a result of the calculation, we received the MLR formula for bone BA determining bone mineral density (BMD) of lumbar spine and radius, minimal femoral and hip BMD as well as Trabecular Bone Score, and FRAX. The MLR equation allowed to calculation of the bone BA with an error of less than 4.9 years for study and control groups and demonstrated the high connection between calculated and chronological (ChrA) ages (R = 0.77; p < 0.00001). The use of NN analysis showed the best results using 6 input variables and 1 internal layer of 7 neurons. The assessment of the connection between BA and ChrA demonstrated a high coefficient of correlation (R = 0.88; p < 0.000001) with an average error of age calculation of less than 3.7 years for the study and control groups. Conclusions. A comparison of the accuracy of both models in bone BA estimation revealed a significant advantage of the deep learning NN, however, the use trained NN model requires specialized software, whereas the MLR formula can be used without additional costs.