Mathematical model of an intelligent stroke prediction system and its implementation based on a hybrid neural network algorithm of recurrent type

Maslennikov Vladimir

The task of binary classification based on the use of a connected complex of recurrent neural networks with a modified architecture of LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Units) is considered in order to predict the possibility of a stroke in the human brain. The aim of the work is to obtain mathematical algorithms that determine the principles of functioning of neural network modules to search for patterns in data containing information about patients who have been diagnosed for stroke, as well as the software implementation of neural network modules as a single intelligent system. Obtaining mathematical algorithms is carried out on the basis of a conceptual analysis of experimental studies on the development of binary classifiers using recurrent algorithms for intelligent prediction, and individual structural and parametric aspects of artificial neural networks of a recurrent type. The software implementation is performed using the TensorFlow machine learning library. Mathematical algorithms of neural network modules are obtained in the form of systems of equations that model the concept of internal short-term memory to find the correlation in data between the values ​​of quantitative and categorical variables. The results of the software implementation of an intelligent system using training, validation and test samples, formed on the basis of a set from a machine learning data repository, are presented.

Keywords:artificial intelligence, mathematical modeling, deep learning, recurrent neural networks, stroke prediction, binary classifier.