Classification of Episodes in the Breathing Signal Using Machine Learning

Machine learning methods, which have become increasingly popular in recent years, can be successfully used in the analysis  of biomedical signals. Open-source libraries enable the creation of artificial intelligence models using, among others, configurable neural networks. The publication presents an approach to the classification of episodes in the breathing signal using accelerometer and pulse oximeter modules. In particular, it examines the influence  of the type of network activation function and time window parameters, i.e. width and offset, on the model sensitivity. With the most optimally selected parameters, it was possible to obtain sensitivity of 85.71% for the detection of episodes in the signal and 100% sensitivity for the classification of calm breathing. The paper also discusses the possibility of creating intelligent sensors by implementing minimized machine learning models on miniature mobile devices with limited hardware resources. Moreover, it proposes a further research path, which is the development of adaptive algorithms able  to independently select optimal learning parameters.