Almost two decades ago, sleep apnea was considered to be a medically inquisitive subject of little significance and snoring was merely the subject of humor rather than one of serious investigation. Recently, countless studies have been published, concerning all possible aspects of sleep apnea syndrome. Epidemiological and clinical research revealed that the obstructive sleep apnea syndrome may be considered as a major public health problem and that the sequel of the syndrome may have devastating consequences for the lives of those affected.
The primary objective of this research is to extract useful features from the respiratory signal to classify sleep disorders that relies on respiratory effort signal and recording the activity autonomously with minimal user intervention. The signal processing is performed with the aid of second order AR model and the features are extracted using a adaptive improved Burg's algorithm. The extracted features are classified by means of artificial neural networks and these are readily made implementable into digital signal processor.
The objective of this work is to investigate the feasibility of performing reliable respiratory signal classification for detecting sleep disorder breathing using Artificial intelligence techniques. The primary signal of interest is respiratory signal. A respiratory signal provides valuable information to study the lung function and sleep disorders. There are three classifications of respiratory signals: normal, sleep disorder breathing or sleep apnea, and motion artifact. Of the three, sleep apnea is the most prevalent and it will be the focus of this thesis. Sleep apnea is characterized by instances of periodic loss of breath or choking sensations throughout the night causing frequent interruptions to our restorative sleep. When sleep apnea is kept
vi
untreated, it increases the risk of High blood pressure, Heart attack, Obesity, Cancer and Diabetes, and thus increases the risk for worsening Heart failure.
Respiratory signals are chosen for investigation because they can be obtained non-invasively and they are user friendly. The proposed work entirely depends on the direct method which deals only with a respiratory signal as this approach increases the reliability and classification accuracy to the optimum level. This research reveals the most valuable information that respiratory signals alone are sufficient and perform even better than the combined respiratory and ECG signals.
The core part of the research comprises three subdivisions namely, noise removal, feature extraction and classification. While recording the respiratory signal, in addition to the actual signal, unwanted signals may also get recorded. These unwanted signals direct the classifier to a false diagnosis of apnea. Hence in order to obtain a noiseless respiratory signal an adaptive filter using ANFIS is developed to remove the artifacts present in the respiratory signal data.
Secondly, the features of respiratory signal are extracted using different feature extraction methods like Spectral analysis using Autoregressive (AR) Model and adaptive improved Burg's algorithm. The key features which convey fine, diminutive and accurate details like Energy Index (EI), Respiration Frequency (FZX), Dominant Frequency (FAR) and Strength of Dominant Frequency (STR) are extracted from the signal for classification.
The third, based on the extracted features, the respiratory signal is classified using different neural networks such as a Multilayer Perceptron Network (MLP), Radial Basis Function (RBF) network and Nonlinear Autoregressive exogenous model (NARX) network. The classification
performance related to accuracy, low mean square error and low computational complexity of neural network architecture such as RBF, NARX and MLP are compared, analyzed and summarized.
