A Method for Determining the Type of Human Breathing Based on Machine Learning

Abstract

Listen

Translate article

The purpose of research is to increase the effectiveness of respiratory rehabilitation through the development of automated methods for determining the type of breathing based on machine learning. Methods. After the COVID-19 pandemic, respiratory rehabilitation became particularly important, as well as methods of home (remote) rehabilitation using the means provided by modern technologies, for which new methods and means began to be developed, including using wireless sensors or motion capture systems. Special attention during respiratory rehabilitation is paid to the type of human breathing, as well as automated methods for analyzing breathing. At the moment, the problem arises that most of the developed methods for analyzing breathing do not work with types of breathing: they either determine only one type, for example, diaphragmatic, or simply analyze the condition of the lungs. In this regard, there is a need to develop a method for analyzing and determining directly the types of human respiration. This article discusses three methods for solving the problem of determining the type of human breathing using a motion capture system and machine learning. The first method is based on static characteristics, for which the Random Forest model was used. The second method, which is based on time characteristics, used the Catch22 model. The third method, which determines the type of respiration using the characteristics of the sinusoid, used a composite model based on two models of Hist Gradient Boosting. Results. Three methods have been developed to determine the type of human breathing. Machine learning models were trained for each of the methods to find the best accuracy result. After conducting a comparative analysis of the developed approaches, the approach with the best accuracy is determined. Conclusion. A method for determining the type of human breathing based on machine learning has been developed, the accuracy of which is 0.81.