Analysis of Obstructive Sleep Apnoea ECG-Based Features

Obstructive sleep apnoea (OSA) is a common sleeprelated breathing disorder that usually goes undiagnosed, as it requires diagnosis by polysomnography which is a laborious and costly process. The number of people undiagnosed can be reduced if machine learning algorithms are applied to signals that are easier and more cost-effective to measure than polysomnography. This paper examines the feature space formed by features extracted from HRV time series and evaluates the TSBF algorithm, a method not yet applied within OSA research. Dimensionality reduction has been used to predict whether a classifier will be good at apnoea and non-apnoea classification. UMAP has been used for dimensionality reduction and visualisation, however, the data was complex that two distinct clusters could not be separated. There was a strong degree of overlap which meant the original 9 features from heart rate variability used as inputs for the models are not adequate for achieving good accuracy. Accuracies of 76.9 % have been reached with the ApneaECG database. This result shows that simple machine learning algorithms that are not bespoke can still be used to detect OSA. Apnoea-hypopnea index calculated for the ApneaECG records is 12.0±13.7. TSBF has also been used and showed that OSA detection is possible with no feature engineering with an accuracy of 76.4 %.