Obstructive sleep apnoea (OSA) is a common nocturnal breathing disorder where the upper airways collapse repetitively during sleep causing cessations in breathing. Full obstructions block all airflow and are called apnoeas while partial obstructions only partially limit the airflow and are called hypopneas. The diagnosis of OSA is based on daytime symptoms and the apnoea-hypopnea index (AHI) which is defined as the number of respiratory events per hour of sleep. The AHI is determined with overnight polysomnography (PSG), where several physiological signals such as breathing, brain electrical activity and blood oxygenation, are recorded. The respiratory events are scored manually by reviewing the recorded signals using visual scoring rules. However, the manual scoring of each event is very time consuming and therefore it is an expensive process. Currently, the AHI calculated from the scoring is used as a full-night average and thus the variation in OSA severity during the night is ignored. The event frequency can vary significantly during the night and this variation can affect the diagnosis. In addition, there is extensive night-to-night variation in AHI. Therefore, it is generally accepted that a single night of recording is not sufficient for accurate diagnosis. However, due to the high cost of manual scoring, it is not feasible to record and score multiple consecutive nights and therefore only a single monitoring night is currently used in OSA diagnostics.
In addition to the AHI, the daytime sleepiness is used in the OSA diagnosis. The gold standard test for daytime sleepiness is the multiple sleep latency test (MSLT), where the patient’s sleep latency is measured in a sleep laboratory multiple times during a single day. However, due to the cost and complexity of MSLT, it is not commonly conducted for OSA patients and instead the daytime sleepiness is determined using subjective sleep questionnaires. However, the results of these questionnaires can vary greatly between the patients due to different personal preferences and tolerances to sleepiness. In addition, the results of these subjective sleep questionnaires correlate poorly with the objective MSLT results.
Because of these limitations in OSA diagnostics, the development of more advanced and automated methods could improve the efficiency and availability of OSA diagnostics. Machine learning methods, such as artificial neural networks (ANN) can iteratively learn features from data and use these learned features to perform intelligent tasks without the need to be explicitly programmed. They can be applied to tasks where the inputs and desired outputs can be defined but the connection between them is unknown, or dependent on multiple complex factors. Therefore, machine learning solutions could also be used to solve some of the issues in sleep apnoea diagnostics.
The aim of this thesis was to enhance the diagnostics of OSA using novel computational methods. This was done by applying machine learning for automatic respiratory event scoring, estimation of OSA severity and prediction of objective daytime sleepiness. In addition, the intra-night variation in AHI and its effect on the diagnostics of OSA, was studied. An ANN trained to estimate the AHI of an OSA patient using only a peripheral blood oxygen saturation signal achieved excellent accuracy, with 91% of the patients being classified into the correct OSA class using the ANN estimated AHI. The median absolute error of the ANN estimated AHI was 0.78 events/hour. Excellent results were also achieved with another ANN trained to automatically score the individual apnoea and hypopnea events from PSG signals. The epoch-by-epoch agreement between manual and the ANN scoring was 89%. The AHI obtained from the ANN scoring was also highly correlated with the manually determined AHI with an intra-class correlation coefficient of 0.96. In addition, when an ANN was trained to automatically estimate daytime sleepiness based on PSG signals, a moderate accuracy was reached. The ANN classified the patients as sleepy or non-sleepy, with 77% accuracy. When investigating the intra-night variation in OSA severity, the frequency and duration of the obstructive events varied significantly hour-by-hour and showed overall increasing trends towards morning. Using only the AHI for the two hours when the obstructive event frequency was highest led to significantly different rearrangement of patients between the OSA severity classes. By using this two-hour-AHI for severity classification, more consistent relationship was found between the OSA severity and mortality compared to the standard full-night AHI.
In conclusion, by applying machine learning solutions in OSA diagnostics, some of the current limitations could be mitigated. The automatic analysis tools and methods presented in this thesis achieved high accuracy when compared to manual analysis. As the presented methods are simple and fast, they could enable more affordable and more widely available screening tools for OSA. In addition, the automatic analysis methods presented in this thesis could be used together with portable recording devices to efficiently monitor and analyze multiple consecutive nights and thus the errors attributable to the intra- and inter-night variation in OSA severity could be minimized. Therefore, an even better estimation of the true OSA severity could be reached.
The doctoral dissertation of MSc Sami Nikkonen, entitled Novel computational tools for enhanced diagnostics of obstructive sleep apnoea, will be examined at the Faculty of Science and Forestry. The opponent in the public examination will be Professor Esther Rodriguez-Villegas (Department of Electrical and Electronic Engineering, Imperial College London, London, UK), and the custos will be Adjunt Professor Timo Leppänen of the University of Eastern Finland.