Developing sensor technology and algorithm for enhancing accuracy of monitoring anomaly sleep at home

Abstract

Sleep apnoea represents one of the most prevalent respiratory disorders worldwide, affecting approximately 49.7% of men and 23.4% of women. While polysomnography (PSG) remains the gold standard for diagnosis, its complexity, high cost, and limited accessibility have motivated the development of alternative diagnostic approaches. As respiration and blood oxygen saturation (SpO2) serve as key indicators in sleep apnoea assessment, photoplethysmography (PPG) has emerged as a promising technology for non-invasive monitoring. This dissertation presents a comprehensive investigation of PPG-based sleep monitoring systems, encompassing hardware development, signal processing algorithms, and advanced machine learning techniques for abnormal sleep detection. The research establishes a systematic framework for wearable sleep monitoring through three key phases of development and validation. First, a non-invasive continuous monitoring system was developed to assess sleep apnoea via SpO2 and heart rate (HR) measurements. Various breathing experiments were performed to investigate the relationship between breathing patterns, SpO2 fluctuations, and HR variations during apnoea events. Motion artifact (MA) analysis revealed directional and frequency dependencies, various adaptive filters were implemented to compare their effectiveness in MA removal. Next, an integrated framework combining variational mode decomposition (VMD) denoising with deep learning autoencoders was developed for automated abnormal SpO2 detection. The study investigated both the relationship between architectural complexity and performance gains, as well as the trade-offs between model complexity and computational requirements. The performance of deep learning autoencoders was evaluated across datasets with different characteristics. The VMD algorithm achieved an SpO2 root-mean-square error (RMSE) of 0.862% against clinical references. Among four autoencoder architectures, the long short-term memory (LSTM) autoencoder performed best (accuracy: 0.86) by capturing temporal SpO2 dynamics during apnoea events, though with higher computational costs. Finally, the feasibility of direct raw PPG signal analysis was investigated, bypassing error-prone SpO2 calculations. A wireless multi-sensor system integrating PPG and accelerometer data with Bluetooth connectivity was developed and evaluated. Comparative analysis of nine machine learning and deep learning models revealed that neural network architectures consistently outperformed traditional methods, with the novel 1D-Branch CNN achieving optimal balance between accuracy (0.92) and computational efficiency (~25% faster than CNN-LSTM). This system demonstrates the feasibility of using raw PPG signals for reliable abnormal sleep detection in clinical applications, emerges as particularly promising, offering an effective compromise between the computational efficiency, high performance and the need for generalization in variable environments.