Obstructive sleep apnea detection: exploring neural activity variations through statistical analysis and machine learning

Sleep disorders are a widespread health condition. Around 40% of adults in the United States experience sleep-related breathing issues, including obstructive sleep apnea (OSA). OSA consists in a reduction or interruption of airflow, due to airway obstruction. It is associated with various clinical conditions, health issues and alterations in the sleep electroencephalogram (EEG). Polysomnography (PSG), a medical overnight test that monitors various physiological activities during sleep, is the basis for OSA diagnosis. EEG is an essential component of PSG, used for classification of sleep stages and neural event detection, including arousals. The primary hypothesis of this thesis aimed to investigate differences in neural activity between healthy and OSA subjects. This research involved exploring variations in neural function by analysing the relative power bands of Power Spectral Density (PSD) extracted from EEG signal. The scope of previous researches was extended by using Sleep Heart Health Study database, exploiting Dirichlet regression to quantify the effects of demographic and clinical predictors, such as age, gender and apnea-hypopnoea index (AHI), on PSD and their differences between healthy and OSA subjects. Moreover, using age, gender, and overall-night PSD as predictors, several machine learning methods were trained to classify subjects' health status as healthy or OSA, aiming to explore differences between health states. Extensive data pre-processing and utilization of the multi-taper method enabled PSD features extraction. However, the exploratory analysis and data visualization revealed substantial variability in PSD estimates among subject categories, marked by notable standard deviations that could obscure distinctions between healthy and OSA subjects. Distinct models were generated using Dirichlet regression for individual sleep stages and overall night data, reflecting varying levels of complexity of input data characteristics. Predictions evaluations were conducted both on individual sleep stages and on overall night data. Predictions for relative power bands varied based on AHI values and age. Particularly, holding age constant, predictions for δ1 band increased with higher AHI values, while predictions for δ2 band slightly decreased with higher AHI values. In addition, when holding AHI constant, predictions for δ1 band decreased with increasing age, influencing other relative power bands. Predictions to distinguish healthy from OSA subjects exhibited excessive variability, hindering the drawing of clear conclusions from the ratio between the two classes coefficients. This may be related to model underfitting, suggesting the need for a more comprehensive model. Machine learning techniques employed overall night data only. Classifiers' performance varied between experiments. Moderate performance was achieved when considering the entire dataset, with classifier performance rarely exceeding 60% in balanced accuracy. Significant enhancement was evident when focusing solely on healthy and severe OSA subjects, showing balanced accuracy at 75% and AUROC and AUPRC values over 0.8. Various classifiers, including Lasso Logistic Regression, Random Forest and Extreme Gradient Boosting, displayed notable performance, contributing to successful subject classification. K-fold cross-validation confirmed stable and consistent performance of the classifiers in both experiments. Further research is required to overcome current limitations and examine the results of this study.