Abstract
BACKGROUND: Osteoporosis is a common age-related disease with disabling consequences, the early diagnosis of which is difficult due to its long and hidden course, which often leads to diagnosis only after a fracture. In this regard, great expectations are placed on advanced developments in machine learning technologies aimed at predicting osteoporosis at an early stage of development, including the use of large data sets containing information on genetic and clinical predictors of the disease. Nevertheless, the inclusion of DNA markers in prediction models is fraught with a number of difficulties due to the complex polygenic and heterogeneous nature of the disease. Currently, the predictive power of neural network models is insufficient for their incorporation into modern osteoporosis diagnostic protocols. Studies in this area are sporadic, but are widely demanded, as their results are of great importance for preventive medicine. This leads to the need to search for the most effective machine learning approaches and optimise the selection of genetic markers as input parameters to neural network models. AIM: to evaluate the effectiveness of machine learning and neural network analysis to develop predictive risk models for osteoporosis based on clinical predictors and genetic markers of osteoporetic fractures. MATERIALS AND METHODS: The predictive models were trained using a database of genotyping and clinical characteristics of 701 women and 501 men living in the Volga-Ural region of Russia. Anthropometric parameters, data on gender, bone mineral density level, and the results of genotyping of 152 polymorphic loci of candidate genes and replication loci of the GEFOS consortium's full genome-wide association search were included as input parameters. RESULTS: It was found that the model for predicting low bone mineral density, including 6 polymorphic variants of the OPG gene (rs2073618, rs2073617, rs7844539, rs3102735, rs3134069) and 5 polymorphic variants of microRNA binding sites in the mRNA of genes involved in bone metabolism (COL11A1 - rs1031820, FGF2 - rs6854081, miR-146 - rs2910164, ZNF239 - rs10793442, SPARC - rs1054204 and VDR - rs11540149) (AUC=0.81 for men and AUC=0.82 for women). CONCLUSION: The results confirm the promising application of machine learning to predict the risk of osteoporosis at the preclinical stage of the disease based on the analysis of clinical and genetic factors.