Sleep apnea events detection from polysomnogram studies using deep learning techniques

The significant raising of deep learning techniques has made us discover its capability in the detection of Sleep Apnea (SA). Sleep Apnea is a respiratory issue described by irregular breathing samples during sleep, including loud snoring and successive waking. Uncountable sleep apnea cases are under-evaluated because of the remoteness, burden, or cost of Polysomnogram in sleep laboratories. Thus, automated detection has been applied using different machine learning methods. In this thesis, we proved the clinical presumption by targeting specific physiological signals, which are Chest, Abdominal, Oxygen Saturation (SpO2). We were able to detect sleep apnea events from those signals and normal events as well. Two databases from Physionet are used in this thesis. The first database was a competition in 2018, called You Snooze You Win (YSYW), and the other database was also a competition in 2000, called the Apnea-ECG. We used the first one for training and testing, and for generalization, we used the second database for testing also. Epochs of 60 seconds in size have been handled as input datasets. Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) were included to be exploited in this task. Our results showed that the hybrid model consisting of Convolution Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) outperformed the other deep learning techniques. The top overall accuracy reached on the first database is 97%, and the second database is 92%. Finally, we discussed the obtained results with other related researches.