Experience in Predicting the Risks Associated with Mechanical Damage to Human Bone Tissue Using Recurrent Neural Networks

Introduction. There are new, in addition to probabilistic and statistical, methods of risk assessment, which require their own methods of analysis and numerical estimation. One of these methods, recognized in a variety of application fields, is associated with the use of direct propagation neural networks. This approach makes it possible to expand the range of tasks that are solved in the field of risk analysis. There are quite a number of systems that require assessment in terms of risk formation, but which are associated with a large number of random factors related to the risk-forming events of the system and its states. Such systems are difficult to model with the help of well-known neural networks. Within the framework of the work, it is proposed to use the capabilities of deep recurrent neural networks with feedback as stabilizing factors with minimization of operational information that needs to be remembered in the process of calculating and operating such a network. Such a model for mechanical damage to human bone tissue depending on a large number of random or indeterminate input signals is proposed to be used in this work. The purpose of the paper is to develop a technique for the use of deep recurrent neural networks and to create a model for predicting event risks associated with the impact of input signals with a high degree of uncertainty or random signals on the system. To provide opportunities for predicting such risks using examples related to injuries of the human skeletal system for its various conditions and conditions. Results. A technique for using recurrent neural networks to predict the risks associated with the violation of the integrity of the human skeletal system was developed. A model of a recurrent neural network was created to predict random events associated with a violation of the integrity of the human skeletal system. Double calculation, aimed at a variety of results, is a confirmation of the performance of the proposed model. It is shown that, depending on the scope of the task set in the analysis, its result is a three-dimensional matrix in coordinates (Xij ∧ Yp;T; φ). At each subsequent step of the iteration in the matrix (φ Xij ∧ Yp;T), by cutting off that part of the potential risk-generating events that, in the opinion of the neural network, are less predicted for each age composition, in favor of other events, real risk-forming events are filtered out, which have predominant values for the system. Conclusions. On the example of random events that accompany mechanical damage to human bone tissue, the ability of models created on the basis of RNN with feedbacks to avoid the uncertainty of risks accompanying human life in four specified ranges of life time and to determine the most effective ones for each of them for a modern person is shown. The ability of activation functions of the bifurcation nature of one of the synapse layers to qualitatively filter random signals in systems of recurrent neural networks with DT-RNN (Deep Transition RNN) feedbacks is shown. The use of deep recurrent neural networks in the formalized version provides new opportunities for taking into account groups of random but real events in the analysis of event risk by clarifying the feedback, and at each subsequent step of their iteration to obtain more accurate data to predict such risk, avoiding the uncertainty of the system state. The formalization of this process provides opportunities to predict random risks for certain groups of the population as a priority, and to use them in the preventive work of medical institutions of the first group of care. Keywords: risks, random events, recurrent neural network, human skeletal system.