AI-driven lightweight real-time SDR sensing system for anomalous respiration identification using ensemble learning

In less than three years, more than six million fatalities have been reported worldwide due to the coronavirus pandemic. COVID-19 has been contained within a broad range due to restrictions and effective vaccinations. However, there is a greater risk of pandemics in the future, which can cause similar circumstances as the coronavirus. One of the most serious symptoms of coronavirus is rapid respiration decline that can lead to mortality in a short period. This situation, along with other respiratory conditions such as asthma and pneumonia, can be fatal. Such a condition requires a reliable, intelligent, and secure system that is not only contactless but also lightweight to be executed in real-time. Wireless sensing technology is the ultimate solution for modern healthcare systems as it eliminates close interactions with infected individuals. In this paper, a lightweight real-time solution for anomalous respiration identification is provided using the radio-frequency sensing device USRP and the ensemble learning approach extra-trees. A wireless software-defined radio platform is used to acquire human respiration data based on the change in the channel state information. To improve the performance of the trained models, the respiration data is utilised to produce large simulated data sets using the curve fitting technique. The final data set consists of eight distinct types of respiration: eupnea, bradypnea, tachypnea, sighing, biot, Cheyne-stokes, Kussmaul, and central sleep apnea. The ensemble learning approach: extra-trees are trained, validated, and tested. The results showed that the proposed platform is lightweight and highly accurate in identifying several respirations in a static setting.