Detecting Types of Sleep Apneas through Nonlinear Features of Electrocardiogram and Electroencephalogram Signals

Sleep apnea is a common disease among women, and mainly men. The most dangerous complication of this disorder is heart stroke. Other complications include insufficient sleep and resulting daytime tiredness and illness that affect the individual's activities during the day, disrupt their life. Therefore, identifying this disease is important.

We used Electroencephalogram (EEG) and Electrocardiogram (ECG) channels from the data of 25 patients with sleep apnea, for each type of sleep apnea, 8 nonlinear-like features, including fractal dimension, correlation dimension, certainty, recurrence rate, mean diagonal lines, the entropy of recursive quantification analysis, sample Entropy, and Shannon entropy were extracted. Then, feature matrices were sorted using principal component analysis in the order of linear combination of features, and the 20 selected features were chosen, normalized using common methods, and fed to different classifiers. Two 5-class and 2-class classification methods were assessed. In the 5-classification, three classifiers were used; the support vector machine, k-nearest neighbor, and multilayer perceptron.

The results showed that the highest mean validity, accuracy, sensitivity, and specificity for the SVM classifier was 88.45%, 88.35%, 88.33%, and 88.32%, respectively. In the 2-class approach, in addition to the mentioned classifiers, linear discriminant analysis, Bayes, and majority voting were used, and each class was considered against all classes. The highest average validity, average accuracy, average sensitivity, average specificity using the majority rule voting was 94.35%, 94.30%, 94.32%, and 94.15% respectively.

When the results of classifiers are combined with the majority voting method, the validity of identifying the classes increases. The average validity for this method was obtained at 94.42%, which was higher than several other studies. It is recommended that databases with a larger sample size be used. This would lead to increased reliability of the proposed analysis method. Moreover, using novel deep-learning-based methods could help obtain better results.