Improving Machine Learning Classification Accuracy for Breathing Abnormalities by Enhancing Dataset

Abstract

The recent severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), also known as coronavirus disease (COVID)-19, has appeared as a global pandemic with a high mortality rate. The main complication of COVID-19 is rapid respirational deterioration, which may cause life-threatening pneumonia conditions. Global healthcare systems are currently facing a scarcity of resources to assist critical patients simultaneously. Indeed, non-critical patients are mostly advised to self-isolate or quarantine themselves at home. However, there are limited healthcare services available during self-isolation at home. According to research, nearly 20–30% of COVID patients require hospitalization, while almost 5–12% of patients may require intensive care due to severe health conditions. This pandemic requires global healthcare systems that are intelligent, secure, and reliable. Tremendous efforts have been made already to develop non-contact sensing technologies for the diagnosis of COVID-19. The most significant early indication of COVID-19 is rapid and abnormal breathing. In this research work, RF-based technology is used to collect real-time breathing abnormalities data. Subsequently, based on this data, a large dataset of simulated breathing abnormalities is generated using the curve fitting technique for developing a machine learning (ML) classification model. The advantages of generating simulated breathing abnormalities data are two-fold; it will help counter the daunting and time-consuming task of real-time data collection and improve the ML model accuracy. Several ML algorithms are exploited to classify eight breathing abnormalities: eupnea, bradypnea, tachypnea, Biot, sighing, Kussmaul, Cheyne–Stokes, and central sleep apnea (CSA). The performance of ML algorithms is evaluated based on accuracy, prediction speed, and training time for real-time breathing data and simulated breathing data. The results show that the proposed platform for real-time data classifies breathing patterns with a maximum accuracy of 97.5%, whereas by introducing simulated breathing data, the accuracy increases up to 99.3%. This work has a notable medical impact, as the introduced method mitigates the challenge of data collection to build a realistic model of a large dataset during the pandemic.

Radio frequency sensing and its innovative applications in diverse sectors: A comprehensive study

“…The data were trained, tested, and validated using time-domain CSI data statistical features and 89.73% accuracy was found from the K-nearest machine learning algorithm. Rehman et al (2021) used COVID and non-COVID scenarios, in which real-time data were collected of abnormal breathing patterns during the pandemic using a non-contact SDR platform and applied an ML model for classifying the collected data with more than 90% accuracy. Taylor et al (2020) proposed human motion detection.…”

Non-Contact Supervision of COVID-19 Breathing Behaviour With FMCW Radar and Stacked Ensemble Learning Model in Real-Time

“…Hold breathing signals have a tiny amplitude that is almost imperceptible. Respiration of bradypnea patients has the same characteristics as hold breathing signals [45].…”

Development of an Intelligent Real-Time Multiperson Respiratory Illnesses Sensing System Using SDR Technology

“…In addition, a scenario involving monitoring through walls is considered. [44] uses SDR to implement a multi-frequency, continuous-wave radar system for monitoring breathing patterns at predetermined distances. The channel frequency response (CFR) is used in [45] to identify minute variations in OFDM subcarriers caused by human motion over wireless channels.…”