Detection and Classification of Obstructive Sleep Apnea Disorders: A Comparative Analysis of Various Deep Machine Learning Classifiers

Sleep Apnea (SA) disorder is mainly caused by blockage of the upper respiratory tract. It is a common respiratory disorder now-a-days. SA condition is primarily characterized by loud disruptive snoring during sleep. It also causes brief repeated cycles of cessation of breathing. It is very common among the adults but a small percentage existing children’s. Recent studies reveal that it is becoming common in children’s as well. The signs and symptoms of SA disorder includes loud excessive snoring, choking or gasping for air, restless sleep during night, excessive daytime sleepiness, early morning headaches, falling asleep during routine activities during daytime. A low sleep quality can affect several vital functions which includes temporary cessation of breathing. It may occur sometimes for more than 10 seconds or minutes. During this period, hundreds of times SA events may occur and if these events get repeated over a long time it can lead to depression, memory loss, asthma, low blood oxygen levels, daytime fatigue, a lack of energy and serious health issues like high blood sugar levels, high blood pressure levels, weakened immune system, neurological issues, liver problems and cardiovascular complications.

Human brain instructs the breathing muscles to take a breath during the sleep.

This paper is organized as follows: section 2 describes about various types of Sleep Apnea. Section 3 details the process used for diagnosis of Sleep Apnea and also about presence of OSA in children. Section 4 presents the literature survey of articles used for detection and diagnosis of Apnea related problems. Section 5 describes how ECG can be used for detection of Sleep Apnea with Deep Learning. Section 6 details about Apnea ECG database and how the features of ECG are extracted for data processing. Section 7 presents the discussion of various ML, DL classifier models used for diagnosis of Apnea with their accuracies.

The global occurrence of SA disorder has increased over the past few decades. Now it is important and required to diagnose because in short-term it can lead to poor quality of life due to lack of restful sleep. On the basis of patient’s medical history, diagnosis of SA often will be suspected, but there are several others tests to confirm the SA disorder.

3.1 Polysomnography (PSG) screening test

Brian Sun and Sally Sun [2] describes that Polysomnography test plays a significant role for the diagnosis of disorder namely Obstructive SA. In many countries Polysomnography (PSG) is widely used as a Sleep Apnea diagnosis method and tool to detect sleep dis-orders. PSG screening test also called as gold-standard procedure used to diagnose the SA is performed in well-equipped hospital laboratories by using the PSG signal. The primary objective of this test is to measure and analyze sleep signals at various stages [3]. During attended PSG screening test, a healthcare expert physically observes a person sleeping in the sleep laboratory setting. As part of PSG test, several sensors are used to measure and record the brain, cardiopulmonary, respiratory signals and muscle, blood oxygen saturation (SPO₂), breathing airflow, patient body position that includes.

An electro-encephalogram (EEG) screens, detects and measures the continuous electrical activity in diagnosing brain related disorders.

• Electro-oculogram screens eye movement – EOG
• Electrocardiogram screens and checks the heart's rhythm and electrical activity – ECG
• Measurement of airflow - via mouth and nasal
• Chest and abdominal movement's measurement
• Loudness of snoring audio recording
• Blood oxygen Saturation levels (SPO₂)
• Video monitoring of the patient
• Muscle activity monitoring using Electromyogram (EMG).

PSG screening test provides an objective measurement of sleep related ailments which includes details on flow of air, breathing effort. These ailments can be used to predict Sleep Apnea disorder. The PSG screening test records cardio-respiratory (CR) signals and also electro-physiological (EP) signals throughout the sleep night. This is one of the difficult situations for any patient to stay in Hospital’s Sleep Lab for the whole night to run the PSG screening test depicted in Figure 1. Once laboratory screening test are completed, all the recorded signals are manually examined by an expert health consultant. Finally, after cross-checking all the information with patient symptoms a diagnosis is reached but it is a time-consuming, labour intensive and error-prone. Screening is also an expensive and complex process as it needs to be performed in Sleep-lab and also requires the trained health expert’s effort. Negative side of this PSG screening test is forcing patient to stay in an laboratory environment with multiple sensors attached to the body which directly impacts the sleep quality of the patient. Therefore, under these circumstances, there is a requirement for a smart diagnostic system which uses less signals, examines maximum patients quickly.

1.png

Figure 1. Sleep test – polysomnography /image: World chronicle polysomnography market research report analysis

Brian Sun and Sally Sun [2] described that Apnea-Hypopnea Index (AHI) was computed mainly by considering the patient’s electro-encephalography or EEG test record, electro-oculography or EOG test record, electro-myography or EMG test record, and also electro-cardiography or ECG test records or activities.

• Apneic event - above 90% drop of airflow measurably for 10 seconds.
• Hypopneic-event - in between 30–90% drop of airflow which lasts minimum of 10 seconds.

3.2 Obstructive sleep apnea in children

Typically, in children's group OSA occurs up to 2 to 5% and it may occur at any age. Mostly it is observed between the ages of 2 and 6years old. The common reason of OSA in children is inflamed tonsils and adenoids that block the air flow and breathing during sleep. For the categorization of OSA in children, as tabulated in Table 1, AHI scale is adjusted [4]. Some of the literature work also uses the Respiratory Disturbance Index (RDI) to diagnose the movement of OSA.

Table 1. Apnea-Hypopnea Index (AHI), RDI severity rating chart for pediatric & adult OSA

In recent years, various research and development approaches have been suggested for detection and diagnosis of OSA disease. For PSG screening test, an exhaustive data-collection is required which is fused by multiple sources of data as discussed in the above Section-2 such as ECG, electro-encephalogram (EEG), electro-oculogram (EOG), electro-myogram (EMG), oxygen-saturation (SpO₂) and many other signals will be recorded with multiple electrodes and sensors attached to the patient. From the screening test results, it is obvious that the PSG test is typically more robust in detecting the SA [5]. But it requires significant amount of effort, time of trained professionals towards completing the screening test process. PSG screening test also has severe impact on the patient’s sleep quality as the patient may be forced to sleep for longer durations in hospital’s Sleep Labs. Due to these circumstances, complicated examination has limited the usage of PSG screening test in clinical execution and practice.

For identification of Obstructive SA, various researchers have developed new techniques over the decades. Due to the low cost, ECG signals are used as feature rich and powerful sensor signal to diagnose, detect various cardiac disorders. Moreover, it is a well-defined method which can simulate, measure the electrical activity of the heart over a period by using a set of electrodes [6]. They are used to detect the small electrical changes on the skin of human body that arise during each heartbeat.

Table 2. Main components of ECG signal

4.png

Figure 4. Classical ECG signal characteristics

Several Bio-Medical researchers have already done deep dive into the usage of ECG signals to diagnose and to detect Sleep Apnea disorder. Several methods which uses a single channel ECG signal for Sleep Apnea diagnosis have been proposed to reduce costs. This has become one of the most popular methods because of physiologically relevant signals of Sleep Apnea [7]. Its occurrence can be easily recorded. Several research papers have already discussed the necessity of examining the ECG signals for determining the Sleep Apnea disorder [8]. ECG signal can be efficiently used to examine the patient’s heart condition because it has a low amplitude typically ~0.5 mV -- 1.0mV range with a frequency range of 0.05 --150 Hz. Every ECG signal consist of 6 set of waves i.e. P, Q, R, S, T, U and also several intervals such P-R, R-R, S-T and Q-T which are used to calculate duration and amplitude levels [9]. From the ECG signal it is possible to identify different waves and intervals as described in Figure 4 and each component details are tabulated in Table 2.

During literature survey and study, it was found that most of the research articles proposed a Deep Learning/Machine Learning algorithm architecture. It is a common schematic representation of the Machine Learning system architecture to detect Obstructive Sleep Apnea event is shown in Figure 5.

For classification of the Obstructive Sleep Apnea the Deep Learning architecture is segmented it to 6 stages that includes:

1. Single channel ECG database collection
2. Input signal pre-processing
3. Applying Deep Leaning techniques for feature extraction and selection
4. Classifier stage
5. Post-processing and OSA severity.

5.png

Figure 5. Generic block-diagram: ECG signals based OSA disorder classification using different AI/ML classifiers

5.1 Description of ECG database

For analyzing the ECG signals, Apnea-ECG Database [10] is one of the most commonly used data bases by several researchers. Philipp’s University, Marburg, Germany provided 70 ECG recordings with one-minute annotations which are freely available. These ECG signals were sampled at 100 Hz sampling rate and the length of the recordings varies between 401minutes and 587 minutes. In this paper, research articles that used Physionet’s CinC challenge-2000 database which is a another widely used public ECG database for Sleep Apnea research [10] were also analyzed. This CinC challenge-2000 database contains total of 70 records but are separated as training database and test vectors of size with 35 records. The main objective of CinC challenge-2000 data base is to check the regular ECG events with 1 minute duration along with the apneic events, if any, that are labeled as normal and OSA affected.

5.2 Preprocessing

After copying the raw ECG signals from well-known “The Apnea-ECG database” improve the quality of ECG signal. Therefore, a filtering process should be applied as these signals are noisy. So before building the ML training database it is necessary to remove noise. Typically, these noises are 60 Hz power line interference. The type of filtering operations can be done using Digital Signal Processing (DSP) filtering applications. Most of the researchers have applied an IIR filtering process to remove the noise.

5.3 Feature extraction

As described in the block-diagram of Figure 5, after completion of filtering on ECG signals, the next general methodology used in most of the research is extracting features from filtered ECG signal [11]. These set features will be used to determine the presence of OSA from the test signal.

Therefore, feature extraction techniques have been mostly applied in AI/Machine Learning applications related to Sleep Apnea detection.

Totally 11 features are used/extracted from ECG data signal. They are viz. Average heart rate, mean R-R interval distance, Root means square distance of successive R-R interval, Number of R peaks in ECG that differ more than 50 millisecond, percentage NN50, Standard deviation of R-R series, Standard deviation of heart rate, Sample entropy, Power spectral entropy, Average heart rate variability, Heart rate variability.

5.4 Neural networks - deep learning classifiers

Machine learning is one of the best alternate options because from the provided raw input data it can automatically learn and extract features. As described in section 2, it was found that most of the Sleep Apnea detection research area is happening around various deep machine learning techniques. The major difference between the classical neural networks and the deep learning algorithm techniques are the latest computing techniques. It enables deep machine learning to build networks with a large number of layers so that deep learning algorithms can explore more complex data and increased volumes of data. During the comparative analysis of various ML, Deep learning classifiers course, we have come across several Machine-Learning classifiers like SVM, DNN, RNN, CNN, 1D-CNN, CNN-LSTM, multi scale dilation attention 1-D convolution neural network (MSDA-1DCNN), RNN-LSTM, LS-SVM, Bi-LTSM, Bi-GRU and ANN-TW MLP. Especially convolution neural networks (CNNs), Long Short-Term Memory (LTSMs) classifiers are the well designed, developed classified and also most commonly used classifiers to achieve excellent performance. Further, these two ML and DL networks (CNNs, LSTMs) have gained lot of success due to its complex data processing and excellent outputs in various domains.

5.5 Post-processing

CNNs, LSTMs Deep Learning algorithm was used to determine Sleep Apnea events from classifier’s output. After getting the classification as output, the very next step is performing post-processing. In this step Apnea-Hypopnea Index (AHI) is calculated and if the AHI index values vary between 5 to 15 shows a mild OSA symptom. Moderate OSA can be diagnosed with 15 to 30 AHI index and above 30 or more AHI events per hour can be treated as severe OSA as shown in Table 1.