| dc.description.abstract | Respiratory rate (RR) is an important vital sign for diagnosing and treating a number of medical conditions. Current respiration monitoring systems require that a special device is continuously attached to the human body. However, contactless respiration monitoring systems have recently been developed to overcome this inconvenience. Research has shown that channel state information (CSI) measured by WiFi devices can be used for estimating RR. Although pattern-based respiration detection has been used to extract RR from periodic changes in CSI, systems based on this method do not perform well when channel conditions are not favorable. This thesis highlights newly introduced learning-based approaches used for RR estimation. Off-the-shelf WiFi devices were used to collect fine-grained wireless CSI data, which was then used to train and evaluate machine learning models.
Results show that classification algorithms, including KNN, SVM, Random Forest, Logistic Regression and MLP, achieve over 96% accuracy when predicting RR. Regression models were compared to an existing pattern-based system, demonstrating that the majority of regression models have better performance when estimating RR. For instance, Logistic Regression’s Root Mean Square Error (RMSE) is 0.35, while pattern-based system’s RMSE is 2.7. It is important to note that classification and regression models cannot be generalized, nor can they accurately predict respiratory rate using the data collected from a new and previously unseen subject. To improve and make the models more generalizable, data used to train the models must be collected from a larger number of subjects. | en_US |
