Detection of human respiration patterns using deep convolution neural networks

Abstract

The method for real­time recognition of respiration types (patterns) of a patient to monitor his conditions and threats to his health, which is a special case of the problem of human activities recognition (HAR), was proposed. The method is based on application of deep machine learning using the convolution neural network (CNN) to classify the chest motion speed. It was shown that the decisions, taken in this case, are coordinated with mobile medicine technology (mHealth) of the use of body sensors and smartphones for signals processing, but CNN offer important additional opportunities at improving the quality of processing the accelerometer­sensor signals in the presence of interfering signals (noise) from other sources and instrumental errors of devices. We proposed the method of transformation of one­dimensional (1D) accelerometer signals into two­dimensional (2D) graphic images, processed using CNN with multiple processing layers, due to which the accuracy of determining the respiration pattern in various situations for different physical states of patients increases compared with the case when two­dimensional accelerometer signal conversion is not used. In this case, an increase in accuracy (or quality) of determining different types of respiration occurs while maintaining a sufficient speed of performing procedures of the planned method, which allows classification of respiration types in real time. This technique was tested as a component of the Body Sensor Network (BSN) and high accuracy (88 %) of determining the patient's respiration state was established, which in combination with contextual data, obtained from other BSN nodes, makes it possible to determine the patient's state and a signal of the aggravation of their respiratory diseases.