Comparison of different data preprocessing methods for sleep state classification using deep learning and cardiorespiratory signals

In this thesis, an effort was made to find cardiorespiratory signal pre-processing methods that would improve the performance of existing state of the art deep models for sleep stage classification using deep cardiorespiratory signals. Four classes were defined: Awake, Light Sleep, Deep Sleep and REM. In this quest, four preprocessing methods were examined: Raw ECG method, R-to-R Interval Extraction method, Binary signal method and Low-pass filter method. For the purpose of training 400 participants were extracted from SHHS Visit 2. Participants who have overall signal quality of 7 and minimum overnight saturation higher than 85 % were taken into consideration. 400 participants were divided into training, validation and test data according to 80% : 10 % : 10 % ratios respectively. State of the art neural network from [1] paperwork had to be modified for different data pre-processing methods to cope with receptive field size oscillation. This was achieved by stride manipulation. Further on, unbalanced data set problem was occurred where Awake and Light Sleep classes are the most dominant ones. This problem was resolved by applying reciprocal class relative frequencies. In the end best Cohen Kappa score was achieved for Raw Signal Pre-processing method with value of 49.49%. In order to improve this results, bias variance trade-off methods like model capacity decrease and balancing classes by adding more examples would have to be used.