| Abstract [eng] |
Sleep is a very important component of health and well-being, but its disorders, such as obstructive sleep apnea, are a serious problem that causes significant health complications. This master's thesis project conducted a comprehensive study on machine learning-based methods for monitoring and diagnosing sleep apnea using photoplethysmography (FPG) signals. The first part of the paper analyses the problem of sleep apnea, its clinical relevance and traditional diagnostic methods, such as polysomnography, and their disadvantages - high cost, complexity and patient discomfort. Non-invasive diagnostic alternatives are discussed, with particular emphasis on the potential of FPG technology due to its simplicity and convenience for use in the home environment. The second part of the paper presents detailed algorithms used in the research, their application methodology and evaluation criteria. Two main types of algorithms were analysed: one using features extracted directly from FPG signals (e.g. PPI and DAP), the other using deep learning transformer models that do not require manually extracted features. The experimental results presented in the third part show that the feature-based model, when processing 60-second intervals (0.5–7 Hz filter) and using the selected five most important features (Aoff, Adn, deltaT, Tsw10, Tdia), achieved the highest AUC – 84.3% and F1 – 77.4%. The transformer model with 10-second signals (0.5–7 Hz) recorded the best sensitivity indicator (71.2%) and AUC – 78.7%. In terms of resource consumption, the feature-based solution requires significantly less computational operations and memory, making it most suitable for embedded systems. The results obtained confirm that machine learning methods using FPG signals can be an effective alternative to traditional polysomnography and provide an opportunity for non-invasive apnea monitoring at home. |
