Machine Learning Methods for Sleep Apnoea Detection Based on Pulse and Oximetry Data

Sleep apnoea is a disease that affects children and adults and can lead to cardiovascular disease, diabetes, and cognitive impairment in severe cases. The Polysomnography is recognised as the golden diagnostic method, but it is expensive and time-consuming, making it impossible to conduct widespread screening.
This thesis first introduces the background of automatic sleep apnoea detection and commonly used change detection algorithms.
Chapter 3 introduces an anomaly detection method that uses an adaptive Cumulative Sum (CUSUM) change point detection algorithm to monitor outliers in the signal. The test results of the adaptive CUSUM are compared with those of the classic CUSUM.
Chapter 4 proposes a novel framework for extracting features from sleep signals using wavelet transforms and uses the RUSBoost algorithm to address the data imbalance problem in sleep apnoea detection. This chapter evaluates classic machine learning methods, such as support vector machines (SVM), k-nearest neighbours, Dirichlet process Gaussian mixture model, and the ensemble method Random Undersampling Boosting (RUSBoost), which aim to address the data imbalance problem. In addition, this chapter utilises feature fusion techniques to evaluate the performance of single-signal detection and multi-signal detection. Chapter 5 presents and compares deep learning approaches, specifically Convolutional Neural Network (CNN), CNN with SVM and Recurrent Neural Network architectures. The signal-level fusion strategy enhances detection sensitivity significantly.

All proposed approaches are tested on public datasets in different environments, including the Apnoea-ECG database[1], the Childhood Adenotonsillectomy Trial database [2], and St Vincent hospital [3], demonstrating the effectiveness of the proposed methods in identifying apnoea events in different situations.