Efficient ANN Algorithms for Sleep Apnea Detection Using Transform Methods

Sound sleep is an important parameter of health as it is directly related to the health of the heart as well of the brain and psychological health of the human. Sleep disorder or sleep apnea (SA) hampers the quality of sleep, which drastically affects the individual’s daytime work. Sleep apnea is caused by interruptions in breathing, causing the shortage of oxygen supply to the heart, in turn affecting the pumping action of the heart. Hence, the heart rate reduces, in turn causing bradycardia, i.e., slow heart condition, and this condition of reduced oxygen level in the blood is sensed by the brain that constricts the arteries to pump blood at a faster rate to supply the deficiency of oxygen level. As a result, the heart rate increases temporarily. Thus, the resulted episode of bradycardia followed by tachycardia is referred to be an apnea event. The adverse conditions of repeated apnea events cause continuous episodes of tachycardia. Heart rate is considered to be an important indicator of sleep apnea, and it can be easily captured from the ECG signal study. Hence, we propose the low-power, high-performance methods of analysis of nighttime recordings of ECG signal to note the deviations of ECG parameters in terms of its amplitudes and time segments of its wave components as compared with their ideal values. The proposed methodology comprises steps, namely, preprocessing for noise removal, QRS complex detection, extraction of features based on QRS complex reference, and classification of features using ANN for sleep apnea detection. The methods are tested using a benchmark dataset as well as own dataset of ECG signals. The experimental results are analyzed and validated by medical experts. A typical ECG signal is preprocessed for noise removal using various filter techniques. QRS detection is done by employing methods, namely, Pan–Tompkins method, Hilbert transform, and wavelet transform methods. The 30 typical ECG features are extracted keeping the QRS complex as a reference and are used for sleep apnea detection. Feature reduction is performed using principal component analysis (PCA). The features extracted from QRS complexes occurring in ECG signal recording are classified using ANN. The training algorithms used for training ANN are Levenberg–Marquardt (LM) algorithm and the scaled conjugate gradient (SCG). The performance of the proposed three algorithms is compared, wherein the wavelet transform-based feature extraction method and SCG trained ANN algorithm emerge as the best.