| Abstract [eng] |
Breathing is one of the important physiological processes that can occur through the nose or the mouth. Namely, frequent breathing through the mouth can become a detrimental habit that can cause various difficult-to-cure diseases. As a result, such a breathing habit should be detected as early as possible to change or cure it. A device with discrete and real-time feedback might be helpful in the formation of the right breathing habits. This Master's thesis project describes the development of a device for continuous monitoring of the breathing mode that can effectively detect and differentiate between nasal and mouth breathing, providing valuable information on breathing patterns. Identification of the mode is based on the analysis of Hjorth and sample entropy parameters derived from temperature and acceleration signals. A data registration protocol was developed and a database of 43 examples in nasal, mouth, and speech modes was recorded. The results of the pilot study demonstrate that the designed device can efficiently detect nasal and oral breathing and speech modes during continuous monitoring using NTC thermistors with a magnetic holder and sample entropy and Hjorth Activity as the extracted parameters. Two signal classification models are compared in this study: a multiclass linear discriminant analysis (LDA) model and a multiclass nonlinear error-correcting output code (ECOC) model. The agreement between the classified long-term signal and the subjectively annotated long-term signal modes during respiration is then compared. Cohen's kappa coefficient is found to exhibit the highest agreement (k = 0,4470) when using the ECOC classification model with its hyperparameters optimized. However, future development is necessary to achieve device miniaturization and algorithmic accuracy to discern speech from mouth breathing in long-term continuous monitoring. |
