JoWUA

Abstract

One crucial parameter for monitoring a person's health is the respiratory rate (RR). RR detection has been performed both automatically and manually in the past. We have recently focused our efforts on solving the RR detection problem with SREC machines and ML algorithms, and it appears promising that we will soon have a solution for completely automatic RR detection. Nevertheless, some deficiencies exist in the previous methods, such as the requirement for specialized sensors, low accuracy, or incompleteness of the real-time function. Motivated by these drawbacks, in the following letter, we propose a deep-learning solution to detect abnormal waveform segments in real-time from in-situ sensor data. To utilize the sensor data, we train Convolutional Neural Networks (CNNs) to extract expressive features and reliably detect RR intervals. Our approach is low-cost, simple to deploy, and able to scale without the need for many custom sensors or additional hardware. It also includes a feature selection of the input data to decrease the size of the input data while enhancing the detection accuracy. We applied our algorithm to real-time sensor data from an experiment conducted on various individuals and compared its performance with that of previous approaches. The results show that our method not only outperforms related algorithms in terms of accuracy but also in terms of real-time capability.