Optimization of The Process of Identifying Sleep-DisorderedBreathing Based on CNN and LSTM Recurrent NeuralNetworks Using PSG EEG Signals

One of the challenges of Clinical polysomnographic (PSG) Datasets is the volume of information they contain, as they can consist of hun-‎dreds to thousands of PSG records, and each PSG record contains more than a dozen clinical time series about eight hours in length. Manu-‎al analysis of such datasets is a slow and laborious process, which is highly dependent on the experience and skill of the sleep technologist ‎and consequently limits PSG-based sleep-related studies. The main objective of this project is the automatic identification of arousals due to ‎respiratory events, specifically RERA (Respiratory Effort-Related Arousal) events and events associated with apnea/hypopnea using PSG ‎electroencephalography (EEG) signals. This paper implements technique to deal with data imbalance and improve systems performance in ‎identifying Respiratory events, design and implements different classification systems for identifying arousals related to respiratory events ‎using PSG EEG signals. Also analyze the performance obtained for each of the implemented classifiers compared to the models presented ‎in the literature. For this four systems were developed based on convolutional neural networks (CNN) and Long-Short Term Memory ‎‎(LSTM) recurrent neural networks. In this context, this research aims to contribute to the optimization of the process of identifying sleep-‎disordered breathing.