Sleep apnea events recognition based on polysomnographic signals recordings: a machine learning approach

Sleep diseases are one of the major causes of physical and psychological problems for workers, resulting in high financial losses in terms of direct, indirect, related and intangible costs. This thesis will focus on a particular type of sleep-related disorder, the Sleep Apnea-Hypopnea Syndrome (SAHS) which causes numerous involuntary respiratory pauses during the night (“apneic events”) leading to a drop of blood oxygen saturation with consequent subject awakening and reduction of sleep quality. There are mainly three forms of sleep apnea: 1)The Obstructive Sleep Apnea (OSA), which is characterized by an upper airway airflow reduction caused by the collapse of the soft tissues in the back of the throat and the tongue; 2) The Central Sleep Apnea (CSA), which is characterized by the absence of respiratory effort and, thus, the absence of airflow; 3) The Mixed Sleep Apnea (MSA), which is a combination of the previous two. Respectively, they represent the 84%, the 0.4% and the 15% of the total cases in U.S. and Europe. The gold standard diagnosis system for sleep apnea is Polysomnography (PSG) that is conducted in special sleep units during an entire night. Numerous physiological signals are recorded during the PSG, and then analyzed by experts that divide the entire signal into epochs and assign to each epoch a sleep stage and an apneic score to specify the subject’s condition during sleep. Manual revision of the PSG recordings requires a considerable amount of time and effort, it is not error free, and together with the sleep units saturation elevates the economic burden of SAHS. In this perspective, a system based on artificial intelligence could offer support to the experts. The main purpose of the present work is to develop an automatic sleep apnea detection algorithm based on only a subset of the signals recorded during a PSG. In particular, it includes only signals from non-invasive sensors that are: EEGs (C3-M2 and O1-M2), EOGs, ECG, SpO2, thoracic volume, audio recording and body position. The dataset used for the present work is the Wisconsin Sleep Cohort DB. It includes more than 2’500 subjects’ PSG (with an average length of 6.13h ± 0.97h) and 3 textual files with useful information. The dataset was cleaned removing potential outliers and a set of 130 features was computed for each subject using the Matlab® working environment. Then, a MANOVA was performed in order to understand how many clusters could be distinguished among the 4 considered (normal, CSA, OSA, MSA), based on the data variability. Only 3 classes were distinguishable, since the MSA is very less represented and with features similar to those of OSA and CSA, therefore it was excluded from the dataset. Finally, different machine learning models were trained, such as Decision Trees, Naïve Bayes, Discriminant Analysis, SVM and K-NN, and underwent hyperparameters optimization according to built-in MatLab routines. In conclusion, among the models trained, the most efficient classifier manages to distinguish between normal sleep epochs and apnea epochs. Moreover, a cascading classifier was tested to distinguish between different apnea forms after separating normal epochs from apnea epochs.