Wireless Channel Modelling for Identifying Six Types of Respiratory Patterns With SDR Sensing and Deep Multilayer Perceptron

The novel variants of COVID-19 are still causing concern across the world. The disease primarily affects the respiratory system, resulting in abnormal breathing patterns. A healthy person usually has a breathing rate between 10 to 20 per minute. A respiration rate outside of this range is considered irregular.

The medical and scientific communities have been trying to detect and identify irregular respiration patterns for early diagnosis of diseases. With radio frequency technologies like RADAR and Wi-Fi sensors, human activities and vital signs can be monitored and detected non-invasively. The latest Software-defined radios (SDRs) offer a flexible, upgradeable, and robust RF alternative for monitoring human activities.

The proposed respiratory pattern monitoring system integrates an SDR platform with deep machine learning algorithms. The proposed prototype could detect and monitor six distinct respiratory patterns.

The researchers employed Universal Software Radio Peripheral (USRP) model “2922” for the SDR system’s RF transceiver. The omnidirectional antenna of USRP broadcasts signals continuously, which travels via many multi-paths to reach the receiver in a closed environment. The signal gets modified by the reflection and diffraction from any person present. The modified signal contains information about the body posture, movements, and respiratory patterns called channel state information (CSI). This CSI signal received by the USRP is processed to select the subcarrier frequencies from the channel most sensitive to the breathing patterns. A wavelet filter removes the random data points with extremely high or low values. Further, a moving average filter of window size 8 removes the high-frequency noise.

The prototype system used a Deep Multilayer Perceptron (DMLP) machine learning algorithm to recognize breathing patterns. The DMLP, a supervised deep learning technique, generates multiple outputs from a set of inputs using a feedforward artificial neural network. A multilayer perceptron is a neural network that links several layers of a directed graph, with the signal flowing in one direction only. A backpropagation technique trains the network by calculating the error function gradient via the chain rule in the multi-layered network.

For simulation and training of the system, the researchers collected datasets for six breathing patterns from 5 participants. Under a supervised environment, the participants performed different types of breathing for 30 seconds each.

These breathing patterns include Eupnea (Normal breathing rate), Biot (deep breathing followed by no respiration), Bradypnea (slower than regular), Sighing (multiple intense breathing), Tachypnea (rapid and shallow respiration), and Kussmaul (fast and deep breathing).

Each dataset contains 3650×3500 data points. The machine learning classifier trains using only 50% of each dataset. The system uses the rest of the data to test and validate the classification. The proposed SDR- Deep Multilayer Perceptron algorithm achieved over 95% precision for all breathing patterns. Similarly, the system performed more than 95% in the F1-score in each case.

The proposed system is reliable enough to detect and monitor abnormal breathing patterns in a controlled environment. More real-time data acquisition from a diverse group of patients is needed to make the model more versatile and reliable.