Multimodal signal analysis for unobtrusive characterization of obstructive sleep apnea

Obstructive sleep apnea (OSA) is the most prevalent sleep related breathing disorder, nevertheless subjects suffering from it often remain undiagnosed due to the cumbersome diagnosis procedure. Moreover, the prevalence of OSA is increasing and a better phenotyping of patients is needed in order to prioritize treatment. The goal of this thesis was to tackle those challenges in OSA diagnosis. Additionally, two main algorithmic contributions which are generally applicable were proposed within this thesis. The binary interval coded scoring algorithm was extended to multilevel problems and novel monotonicity constraints were introduced. Moreover, improvements to the random-forest based feature selection were proposed including the use of the Cohen?s kappa value, patient independent validation, and further feature pruning steered by the correlation between features. These novel methods were applied together with classification and feature selection methods from the literature to improve the OSA diagnosis. All methods were predominantly tested on data sets collected at the sleep laboratory of the UZ Leuven. Within this PhD, I contributed to the collection and structuring of these data sets to have them ready-to-use for collaborators and future generations of students. The first application was the development of reliable, multimodal OSA screening methods based on unobtrusive measurements such as oxygen saturation (SpO2 electrocardiography (ECG), pulse photoplethysmography (PPG), and respiratory measures. An SpO2-based model was proposed which extracted features from oxygen desaturations, including novel estimates of quasi-periodicity. These were then used to identify desaturations caused by apneas using a random forest classifier. Furthermore, the use of a robust linear regression was introduced to estimate the apnea-hypopnea index (AHI) from the number of detected apneic desaturations per hour of recording. The obtained model was tested on several independent test sets and obtained OSA screening accuracies larger than 88 % over all sets. Unimodal methods based on the ECG and PPG were proposed as well. The ECG-based model included extra cardiorespiratory measures and investigated the use of the slope range as an alternative ECG-derived respiration (EDR). In addition, the method was designed such that the ECG and EDR signals could easily be replaced by unobtrusive ECG and respiratory measures. The PPG model was based on the extraction of time series known to be modulated by the respiration and/or the autonomous nervous system. Additionally, the use of instantaneous estimates of the heart rate variability using a point-process model was investigated. Different multimodal OSA detection approaches were explored. The inclusion of the respiratory effort signal in the ECG-based algorithm led to an improvement in performance. While the straightforward addition of the PPG and ECG features to the SpO2-based model was not effective. The use of a 4-state hidden Markov model was investigated as a smarter way to combine the SpO2 and PPG features. Additionally, the impact of sleep staging on the performance of these models was tested. Finally, a main contribution of this PhD was to test the developed ECG and PPG OSA detection algorithms on unobtrusive signals, including capacitively-coupled ECG and bioimpedance, and wearable PPG recordings. Both experiments showed promising results, especially using both capacitively-coupled signals, but also limitations of the current algorithms on the unobtrusive data. In the last part of this thesis, a contribution towards a better characterization of OSA patients beyond the AHI was proposed. Novel pulse oximetry markers were developed and investigated to assess the cardiovascular status of OSA subjects. It was found that patients with cardiovascular comorbidities experienced more severe oxygen desaturations and incomplete resaturations to the baseline SpO2 values. These markers were shown to be of added value to patient demographics to assess the cardiovascular status of OSA patients. The proposed PPG features, on the other hand, did not improve the patient classification. The obtained SpO2 markers are a major contribution of this PhD as they might give more physiological insights in the detrimental effects of apneas on the cardiovascular system, and they might be useful to prioritize OSA patients for treatment. These parameters were further investigated by performing statistical tests to compare the different cardiovascular conditions and the influence of drug use. Moreover, the novel multilevel interval coded scoring was used to train a model to predict the cardiovascular status of OSA patients. The final model included the age, BMI and the SpO2 parameters, and obtained good classification performances on a clinical population. In order to further validate the predictive power of these novel SpO2 markers, a clinical OSA data set including a long-term cardiovascular follow-up should be used.