Automatic Detection of Obstructive Sleep Apnea Using Wavelet Transform and Entropy-Based Features from Single-Lead ECG Signal
Abstract: Obstructive Sleep Apnea (OSA) is a prevalent sleep disorder, and highly affects the quality of human life. Currently, gold standard for OSA detection is Polysomnogram. Since this method is time consuming and cost inefficient, practical systems focus on the usage of electrocardiogram (ECG) signals for OSA detection. In this paper, a novel automatic OSA detection method using a singlelead ECG signal has been proposed. A non-linear feature extraction using Wavelet Transform (WT) coefficients obtained by ECG signal decomposition is employed. In addition, different classification methods are investigated. ECG signals are decomposed into 8 levels using a Symlet function as a mother Wavelet function with third-order. Then, the entropybas features including fuzzy/approximate/sample/correct conditional entropy as well as other non-linear features including interquartile range, mean absolute deviation, variance, Poincare plot and recurrence plot are extracted from WT coefficients. The best features are chosen using the automatic sequential forward feature selection algorithm. In order to assess the introduced method, 95 single-lead ECG recordings are used. SVM classifier having a RBF kernel leads to an accuracy of 94.63% (Sens: 94.43%,
Spec: 94.77%) and 95.71% (Sens: 95.83%, Spec: 95.66%) for minute-by-minute and subject-by subject classifications, respectively. The results show that applying entropy-based features for extracting hidden information of the ECG signals outperforms other available automatic OSA detection methods. The results indicate that a highly accurate OSA detection is attained by just exploiting the single-lead ECG signals. Furthermore, due to the low computational load in the proposed method, it can easily be applied to the home monitoring systems. Existing system: Also, undiagnosed and untreated OSA may lead to a high blood pressure, brain stroke, myocardial infarction, arrhythmias, and ischemia .Even thoughOSA is detectable, the most cases are still not recognized .Polysomnogram (PSG) is the gold standard for OSA detection, which is based on the comprehensive evaluation of the cardio respiratory system and sleep signals. In this method, case studies should be asleep for a couple of nights in the exclusive sleep laboratory in order to record the 16 major signals such as Electrocardiogram (ECG), Electroencephalogram (EEG), respiratory effort, airflow, and oxygen saturation (SaO2) . A PSG device needs at least 12 channels to record the data using 22 wire connectors. The large number of the necessary wire connectors in a PSG device would interrupt the sleep, which affects the OSA detection. Moreover, the PSG test is typically performed in a hospital setting and it requires the supervision of a clinical expert, factors that make PSG an uncomfortable and costly procedure. When an OSA takes place, the oxygen saturation. Proposed system: The proposed method consists of two main steps. In the first step, noisy windows are identified by the weight calculation procedure and eliminated. At this stage, after the preprocessing, the wavelet transform is applied to the single lead ECG signals. Then nonlinear features are extracted from the wavelet transform coefficients. The best set of features is selected by the sequential forward feature selection (SFFS) algorithm and fed into different classifiers in order to discriminate the apnea events from the healthy. In the second stage, AHI is calculated by dividing the total number of apnea events by the total number of minutes of actual sleep time, and then multiply by 60. Then, considering the calculated AHI, the
subjects are marked as apnea or normal. Fig.1 shows the block diagram of the proposed OSA detection method. Advantages: Evaluation of applying two new non-linear entropy-based features including fuzzy entropy and correct conditional entropy for OSA detection and their additional advantages in this field of study. ² Development of a feature selection algorithm for feature dimension reduction and lowering complexity. ² Applying different classification methods on the same collection of features in order to show the performance of each classifier and finally selecting the best one. ² Providing minute-by-minute OSA detection and AHI measurement for patients to prioritize them in the treatment stage. Disadvantages: Wavelet Transform (WT), providing a better presentation of the time-frequency domain of a signal with a different size of windows, is designed to address the nonstationary problem of the signals. In 1991, Pincus introduced the approximate entropy (ApEn) to address the problem of the short length of the data and noisy recording of physiological signals. ApEn is a statistical method used for quantization of irregularity of signals. Modules: Feature Selection: We extracted 12 features from the WT coefficients (A8, D1... D8), all mentioned above. In total, 108 features were extracted. The feature IDs and Descriptions are presented in TABLE I. To improve the performance of the used classifiers and also to reduce the required processing time, a subset of features leading to the most distinguished classes was chosen through the SFFS algorithm. To this end, SFFS algorithm with a misclassification rate criterion was used where as the first step; all the features were individually applied to the classifier to choose the best one. Then, the best feature was combined with all the features to opt for the best coupled one.
The method is continued until the most separation between the classes is attained. At the end, the P-values of the selected collection of features are calculated using the Kolmogorov-Smirnov method. Epoch Classification Performance: After selection of the appropriate collection of the features, some classifiers including SVM, LIBSVM, LS-SVM, ANN, KNN, LR, LDA, QDA, NB and Genleboost are used for the better evaluation of the proposed method. Metrics such as Fmeasure (F), accuracy, specificity, sensitivity, and the Area under Curve (AUC) are used for the performance evaluation of different classifiers. Moreover, important parameters such as Cohen’s kappa coefficient (Kappa) and Matthews correlation coefficient (MCC) are used in order to measure the agreement between raters and the quality of binary classifications respectively. In the first experiment, all of the epochs (without weighting application) are employed in epoch-based classification. In this part, both the 60s and 30s segment durations are considered and the corresponding results are presented in TABLE III and TABLE IV, respectively. As demonstrated in TABLE III, for Apnea-ECG database SVM (RBF kernel) provided the best performance with an accuracy of 92.98%, MCC of 0.85, F-measure of 0.91 and Kappa of 0.85 in the minute-by-minute classification. For UCD database, the mentioned classifier outperformed the other classifiers with an accuracy of 93.70%, MCC of 0.85, F-measure of 0.9 and Kappa of 0.85. Comparison with Other Methods: It should be mentioned that the proposed method results in an accuracy of 92.98% and 95.71% in the minute-by-minute and subject-by-subject classifications, respectively. As can be seen in TABLE VI, the proposed method outperforms other minute-by-minute based techniques which have been implemented on the Physionet Apnea-ECG database, providing an accuracy of 92.98%. As can be seen in TABLE VIII, some of the studies need a high quality dataset to show the performance of their introduced methods. In addition, some of them even remove a part of the dataset to prepare a condition for their proposed methods to achieve better results. It should be noted that the proposed method provides a high robustness since all the recordings, including recordings with low quality, are used. In addition, the most advantage of the proposed method is that it does not depend
on the extracted parameters from the ECG signals for OSA detection (QRS complex, R peak amplitude, EDR/HRV signals and etc.). The proposed method in this study can be easily used at home monitoring systems since it imposes a low computational complexity. Another advantage of the proposed method is the high achieved accuracy in the minute-by-minute classifications. The results of the subject-by-subject classification are presented in TABLE VIII.