A fused-image-based approach to detect obstructive sleep apnea using a single-lead ECG and a 2D convolutional neural network

S. M. Isuru Niroshana; Xin Zhu; Keijiro Nakamura; Wenxi Chen

doi:10.1371/journal.pone.0250618

Abstract

Obstructive sleep apnea (OSA) is a common chronic sleep disorder that disrupts breathing during sleep and is associated with many other medical conditions, including hypertension, coronary heart disease, and depression. Clinically, the standard for diagnosing OSA involves nocturnal polysomnography (PSG). However, this requires expert human intervention and considerable time, which limits the availability of OSA diagnosis in public health sectors. Therefore, electrocardiogram (ECG)-based methods for OSA detection have been proposed to automate the polysomnography procedure and reduce its discomfort. So far, most of the proposed approaches rely on feature engineering, which calls for advanced expert knowledge and experience. This paper proposes a novel fused-image-based technique that detects OSA using only a single-lead ECG signal. In the proposed approach, a convolutional neural network extracts features automatically from images created with one-minute ECG segments. The proposed network comprises 37 layers, including four residual blocks, a dense layer, a dropout layer, and a soft-max layer. In this study, three time–frequency representations, namely the scalogram, the spectrogram, and the Wigner–Ville distribution, were used to investigate the effectiveness of the fused-image-based approach. We found that blending scalogram and spectrogram images further improved the system’s discriminative characteristics. Seventy ECG recordings from the PhysioNet Apnea-ECG database were used to train and evaluate the proposed model using 10-fold cross validation. The results of this study demonstrated that the proposed classifier can perform OSA detection with an average accuracy, recall, and specificity of 92.4%, 92.3%, and 92.6%, respectively, for the fused spectral images.

Citation: Niroshana SMI, Zhu X, Nakamura K, Chen W (2021) A fused-image-based approach to detect obstructive sleep apnea using a single-lead ECG and a 2D convolutional neural network. PLoS ONE 16(4): e0250618. https://doi.org/10.1371/journal.pone.0250618

Editor: Khanh N.Q. Le, Taipei Medical University, TAIWAN

Received: December 16, 2020; Accepted: April 9, 2021; Published: April 26, 2021

Copyright: © 2021 Niroshana et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: Data are available: https://www.physionet.org/.

Funding: 1. X.Z. 18K11532 JSPS Kakenhi Basic Research Fund C https://www.jsps.go.jp/english/index.html The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. 2.X.Z. 2020-P-3 Competitive Research Fund The University of Aizu https://www.u-aizu.ac.jp/ The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

Introduction

Sleep apnea (SA) is one of the most common respiratory disorder and is caused by the complete or partial discontinuation of airflow that accompanies an obstruction of the upper airway for a short time [1, 2]. A complete pause of at least 10 s in the airflow through the upper airway during sleep is usually considered as an apnea-episode.

Three types of SA can be identified, depending on the manner of breathing, namely central SA (CSA), obstructive SA (OSA), and mixed apnea [2]. CSA is caused by instability in the central nervous system and is related to blockage of the airway at the back of the throat. If the airway block is only partially disturbed, then the pathology is termed hypopnea [3]. Such a hypopnea event involves at least 10 s of shallow breathing, which lowers the air volume entering the lungs to below normal levels and causes blood-oxygen desaturation of at least 4%. An occurrence of OSA is recognized when a patient has a complete airflow pause in the upper airway for at least 10 s. Mixed apnea is identified when the apnea begins as a CSA and terminates as an OSA, showing both CSA and OSA characteristics.

Undiagnosed and untreated repetitive apneic episodes can cause a variety of health issues, including excessive daytime sleepiness, cardiovascular and neurological diseases such as memory impairment, high blood pressure, acute coronary syndrome, and congestive heart failure [4, 5]. According to previous studies [2, 6, 7], around 3–7% of adult men and 2–5% of adult women worldwide suffer severely from SA.

In clinical work, the severity of apnea and hypopnea episodes is qualitatively assessed using the apnea–hypopnea index (AHI). The AHI value is defined as the average number of apnea/hypopnea episodes occurring within one hour. In general, a subject showing an AHI value greater than five would be diagnosed with SA [3, 8]. A mild OSA is diagnosed if the AHI value lies between 5 and 15, moderate OSA patients show AHI values between 15 and 30, and severe cases have AHI values above 30 [9].

The most common diagnostic method for OSA suspects is polysomnography (PSG). In PSG, various physiological signals are acquired from sleeping patients, including airflow, respiratory effort, electroencephalogram (EEG), electrocardiogram (ECG), and oxygen saturation (SaO₂). The specific patterns in these physiological signals are then scrutinized by sleep experts to detect sleep-related disorders such as OSA. A PSG study requires dedicated nursing staff and expensive medical equipment specifically designed for PSG data acquisition. The PSG diagnosis method therefore needs dedicated supervision and is time-consuming, expensive, and uncomfortable for patients. Aiming to minimize these technical and economic complications of conventional PSG, several automatic SA detection methods have been proposed during the past two decades. These are usually based on the analysis of the cardiopulmonary (CP) bivariate signal (a combination of heart rate (HR) and respiratory rate (RR) signals), or ECG-derived respiration (EDR).

Respiratory activity is disturbed during an apnea/hypopnea episode, which causes observable variations in the RR signal. Because recording respiratory activity via sensors positioned around the nose is uncomfortable for patients, the RR signal is usually acquired indirectly via EDR signals or inductance plethysmography [10, 11]. The EDR signal is widely used in detecting many sleep-related pathologies because it accurately reflects respiratory activity. Moreover, ECG electrodes can be easily attached to the body without disturbing sleep, unlike direct respiratory sensors.

However, studies have shown that an episode of SA is more strongly associated with signal variability, including heart-rate variability (HRV), morphological variations in the ECG signal [12–14], and variations in the ECG signal’s QRS duration [15, 16]. Therefore, many studies have been carried out based on these observations, and algorithms based on morphological-variation features have tended to show improved performance [13, 17].

Our method also focuses on ECG signal variability, HRV, and morphological variations during an apneic event. It exploits a fused combination of time–frequency representations (TFRs) to intensify the HRV-based and QRS-based variations in ECG signals (scalogram and spectrogram). The presence or absence of apneic events is then detected by means of a deep convolutional neural network (CNN) using fused images created from scalogram and spectrogram representations. To the best of our knowledge, our study is one of few that detects OSA by employing fused spectral images. Compared with other existing methods, our method shows improved performance because it not only combines a variety of TFRs but also exploits recent advances in CNN-based classifiers.

The remainder of this paper is structured as follows. Section 1 summarizes previous work on the detection of OSA-related disorders. Section 2 provides a detailed description of our proposal’s data, method, data preprocessing, image creation, model training, and evaluation criteria. Our results, discussion, and conclusions are presented in Sections 3, 4, and 5, respectively.

Related work

During the previous two decades, many methods have been proposed to detect SA, using a variety of physiological signals that includes ECG, EDR, and respiratory signals [18–20]. According to Guilleminault et al. [21], the occurrence of an apneic event is related to the concomitant variation in the RR intervals in the ECG signal. To date, many studies have considered automatic OSA-detection methods using a single ECG lead. In [22], Khandoker et al. proposed an SA-detection method using features extracted from successive wavelet-coefficient of the RR intervals and the EDR signal from the R waves in the QRS complex with a support vector machine (SVM) classifier. In their study, more than 90% of test subjects were classified correctly. Song et al. [23] developed a per-segment apnea-detection method using a discriminative hidden Markov model (HMM) based on the ECG signals, where frequency-domain and time-domain features were extracted from the EDR and ECG signals. The per-segment detection accuracy of their model was 86.2% with the PhysioNet Apnea-ECG database.

Kunyang et al. [24] proposed a neural-network (NN)-based model that used an HMM for SA classification. In their work, a combination of sparse autoencoders, NNs, and HMMs was used to develop the framework. A classification accuracy of 84.7% was achieved for per-segment apnea detection. Hayano et al. [25] proposed a screening method for OSA using cyclic variations in heart rate (CVHR). The agreement between the SA and the presence or absence of CVHR in each one-minute period was found to be 83%. In the study presented in [17], Sharma et al. achieved 84.4% accuracy for one-minute ECG signals in detecting apnea using a least-squares (LS) SVM classifier with a Gaussian radial-basis-function (RBF) kernel for features derived from Hermite expansion coefficients. Later, Viswabhargav et al. in [8] proposed an apnea detection method whereby EDR and RR signals were utilized to extract sparse residual entropy (SRE) features, using an SVM classifier. In their study, an RBF-kernel-based SVM classifier achieved 85.43% sensitivity and 92.60% specificity for the SRE features.

Tripathy et al. [26] introduced a novel method that analyzed the CP signal using fast and adaptive bivariate EMD coupled with cross time-frequency. The CP signal was formulated using both the HR and RR signals derived from the ECG signal. Their method achieved average sensitivity and specificity values of 82.27% and 78.67%, respectively, using an SVM classifier and “random forest” classifiers in a 10-fold cross-validation method.

In [27], Singh et al. proposed a method based on the heartbeat interval and EDR, where sliding-mode singular spectrum analysis was used to extract features, with sensitivity and specificity values being 82.45% and 79.72%, respectively.

In these methods, many of the features of the ECG signal used in the classification are derived manually. These include waveform parameters such as instantaneous amplitude (IA) and instantaneous frequency (IF), residual entropy features, statistical features, and other specifically derived features. Some features are derived from the QRS complex and selected manually. In some cases, much manual preprocessing is required when performing specific derivations, including EDR signal extraction and QRS approximation prior to the extraction of specific features. Moreover, most existing approaches use frequency-domain and time-domain representations and nonlinear features derived from physiological signals, where substantial knowledge and relevant experience is required.

To address this issue, Wang et al. [28] proposed a method based on a modified LeNet-5 CNN, where feature extraction is automated with an accuracy of 87.6% in the classification of OSA.

Recently, deep learning has become widely implemented in medical imaging and signal analysis because of its advances in pattern recognition and image-based studies. Researchers have also used deep-learning techniques to address ECG-related research issues such as arrhythmia detection [29–34] and other research applications [35, 36]. In these studies, deep neural networks (DNNs) were introduced successfully to extract descriptive and distinguishable features automatically from the input data, which were then used to perform the classification.

McNames et al. [37] employed spectrogram signatures calculated from ECGs via fast Fourier Transform (FT) to classify OSA. They obtained a case-based detection accuracy of 92.6%. Singh et al. [38] proposed a method based on ECG scalograms that were created via wavelet transforms to detect OSA using a DNN. Their method achieved an accuracy of 86.22% and a sensitivity of 90% in per-minute OSA classification.

In this study, we propose a novel method for OSA detection using fused images created by combining Short-Time Fourier Transform (STFT) and continuous wavelet transform (CWT) representations. A deep CNN model is employed to classify apneic and non-apneic ECG segments using the fused images as inputs. The proposed method does not use any QRS-based features or other manually derived features in performing the classification. Instead of using a limited number of features derived from the QRS complex or other EDR signals, we found that combination of two spectral images carries more discriminative information on the presence or absence of apneic events.

Materials and methods

Dataset

To evaluate the proposed method, we utilized the popular and widely used PhysioNet Apnea-ECG database provided by Dr. Thomas Penzel at Philipps University [39, 40]. The dataset comprises single-lead ECG signals from 70 subjects. The recordings are in two groups (a released set and a withheld set), each with 35 subjects. The ECG signals were recorded at a sampling rate of 100 Hz and with 16-bit resolution. Each ECG signal lasted 420–600 min with a mean of 492±32 min. Non-overlapping one-minute ECG segments were annotated as either “OSA” or “Normal,”/but no distinction was made between cases of hypopnea or apnea. The PhysioNet Apnea-ECG database includes both male and female subjects aged from 27 to 63 years with a mean of 43.8±10.8 years. The body weights of the subjects ranges from 53 to 135 kg with a mean of 86.3±22.2 kg. The sleep recordings were obtained from 25 male and 7 female volunteers, including both healthy and OSA subjects [41, 42].

Method

The TFR of a signal is often used to analyze the information embedded in a variety of signals, including physiological, speech, and geophysical signals. It can be used to identify complex and high-dimensional nonstationary properties of the signal. STFT and CWT are two of the most widely used visual representations for analyzing nonstationary signals. In particular, altered frequencies, amplitudes over time, and morphological variations in ECG signals can be better represented using STFT and CWT instead of FT (see Eqs 1 and 3). The TFR of a signal is often illustrated as a colored image (heat map) in a spectrogram or a scalogram. The spectrogram image usually comprises a visual representation of the STFT, and the scalogram image is a visual representation of the CWT [43].

STFT is used to construct the TFR of the physiological signals as a spectrogram with a constant time–frequency resolution (see Section 2, Eq 1). A constant sliding window along the time axis is employed to create a two-dimensional (2D) representation of the signal at this fixed resolution [41, 44, 45]. As a result of using a constant window, all the frequency information is analyzed at the same time–frequency resolution. In contrast to STFT, the CWT’s wavelet window is scaled and shifted during the transformation. This provides long time windows for low-frequency regions and shorter time windows for high-frequency regions. Therefore, the scalogram provides a more detailed and finer representation of the signal in both low- and high-frequency regions. The mathematical formula for calculating the wavelet coefficients is given in Eq 3. As described in this equation, a basis function, i.e., the mother wavelet ψ(t), and its scaled and dilated versions are used to decompose the time-domain signal. In both TFRs, the use of a window introduces a compromise between time localization and frequency localization.

In this study, we conducted experiments comparing four types of imaging: scalogram images, spectrogram images, images based on smoothed pseudo Wigner–Ville distribution [46, 47], and fused images (a hybrid version of CWT and STFT images, see Section 2.3), for the identification of apneic events. However, we noted that the Wigner–Ville distribution method has cross-term issues when used with nonstationary signals [48].

Our proposed apnea-detection method is based on deep learning, using a fusion of two spectral images (scalogram and spectrogram images) for one-minute ECG segments (see Fig 4). Each one-dimensional ECG segment in the time domain is transformed into more-detailed 2D forms (scalogram, spectrogram, Wigner–Ville distribution, and fused image), which are used by the CNN to perform image feature extraction and classification.

As explained in the Introduction, the motivation behind combining the two TFRs is to increase the discriminative information in newly formed images. With the Apnea-ECG database providing one-minute-based annotations, the proposed method uses one-minute ECG segments to identify apneic events. Although there are many good reasons for employing TFR or fused TFR in a variety of applications, expert knowledge is expressly needed to extract specific features from such representations. In other words, it is often not realistic to select, analyze, or identify the specific patterns or features in a TFR manually because it contains such fine and complex details. Therefore, utilizing deep learning techniques is the most promising approach to identifying TFR features intelligently and automatically.

We propose a residual learning approach for performing OSA classification, as shown schematically in Fig 1. A plain CNN is created with stacked layers of linear and nonlinear processing units. These layers enable the network to learn complex and detailed representations at different levels of abstraction [49]. A residual network differs from a plain CNN in having “skip” connections, as shown in Fig 1. Here, activations from the previous layer are reused until the next layer learns its weights. These skip connections help to mitigate gradient vanishing and degradation, which are common problems in large plain networks. Moreover, a residual network can be easily trained to learn a residual mapping with fewer stacked layers than a plain network, with substantially improved performance in image classification [50–52].

Download:

Fig 1. Proposed 2D-CNN network for OSA event detection.

“Conv(k,s,f)” denotes a convolutional layer where k, s, and f are the kernel size, stride size, and number of filters, respectively. “Max(p,s)” denotes a max-pooling layer where p and s are the pool size and stride size, respectively. The values for the filter sizes “f” in the four residual blocks are 32, 64, 96, and 128.

https://doi.org/10.1371/journal.pone.0250618.g001

Preprocessing and image creation

It is usual for a raw ECG signal recorded by a continuous cardiac monitoring process to be corrupted by various types of noise. The presence of noise-related artifacts in the TFR image may cause imprecise estimation of characteristic points and features [53]. Therefore, we performed signal denoising using three of MATLAB’s inbuilt functions: “wavedec,” “waverec,” and “cmddenoise” [46]. First, the raw one-minute segments were transformed into wavelet coefficients using “wavedec,” with the “sym8” wavelet used to perform the baseline correction. Next, the “cmddenoise” function was utilized to perform interval-dependent thresholding for the baseline-corrected signal. Fig 2 shows part of a raw ECG segment and its preprocessed waveform before being transformed into image form. The denoised dataset is transformed into TFRs (four image datasets) as illustrated schematically in Fig 3.

Download:

Fig 2. Preprocessing ECG segments.

(a) Part of an original ECG segment. (b) The denoised and scaled version.

https://doi.org/10.1371/journal.pone.0250618.g002

Download:

Fig 3. Image dataset creation.

https://doi.org/10.1371/journal.pone.0250618.g003

First, we prepared a spectrogram dataset for evaluating the performance of the proposed model. Here, we used MATLAB’s inbuilt function “spectrogram” with a “blackman” window. While creating the spectrogram images, the window size was set to 64 (640 ms), and the overlap was set to 60 (600 ms) samples [47]. The definition of window function ω(n) is given in Eq 2. MATLAB’s “cwt” function was employed to create the scalogram images using the “Morse” analytic wavelet. The scalogram image was formed using the squared modulus of the CWT coefficients as a function of time and frequency, where the frequency is plotted on a logarithmic scale. The height and width of the created scalogram images represent the frequency and time, with the red/green/blue (RGB) colors representing the absolute values of the CWT mapped into a (three-dimensional) color map.

In this study, the “Morse” wavelet was used as the mother wavelet for the CWT of the ECG segments, given that it had already been used successfully in many research applications [36, 54, 55]. We saved both sets of images, generated with “cwt” and “spectrogram,” using the “gcf” command. Fig 4 shows both types of TFR image created for normal and apneic ECG segments.

Download:

Fig 4. One-minute ECG segments transformed into (128, 128, 3) RGB images.

(a) Scalogram image of a normal ECG segment. (b) Spectrogram image of the normal segment. (c) Scalogram image of an apnea ECG segment. (d) Spectrogram image of the apnea segment.

https://doi.org/10.1371/journal.pone.0250618.g004

For comparison purposes, we also used MATLAB to prepare a TFR involving a smoothed pseudo Wigner–Ville distribution, with time and frequency windows used for the smoothing. (1) (2) where L is the window length, x[k] is the input ECG signal, and ω is the window function. The log values of X_STFT[m, n] are used to create the RGB color image (the spectrogram image) [44]. (3) where W_x(s, τ) is the wavelet coefficient, x(t) is the ECG signal, ψ(t) is the basis function (mother wavelet) conjugate, s is the scale, and τ is the time parameter.

In creating spectral images, we pretested several window sizes, overlap lengths, windows, and other parameters, using a small amount of randomly selected data to make sure that appropriate and comprehensible images were generated, based on visual inspection and CNN performance.

After confirming the most appropriate parameters for image creation, the scalogram, spectrogram, and pseudo Wigner–Ville distribution image datasets were constructed. Finally, the fused image dataset was generated using the scalogram and spectrogram images as shown in Fig 5. Fused images were created by embedding gray-scale values of the scalogram and its matching spectrogram into three layers of an RGB image. To embed the CWT and STFT representations into one image, the gray-scale values of the scalogram image were used as the “red” component of the new image, and the “green” component was formed using the corresponding gray-scale values of the spectrogram. The “blue” layer was created by the addition of the gray-scale values of the scalogram and spectrogram. In this way, the three RGB layers of the fused image accommodated picture elements from both scalogram and spectrogram images. As shown in Fig 5, the modified image is therefore a hybrid version of the CWT and STFT images that carries more discriminative information than does the original form. In other words, each pixel or point represents the spectral presence of the ECG wave derived from both TFRs.

Download:

Fig 5. Fusing the scalogram and spectrogram for an apnea ECG segment.

(a) Gray-scaled scalogram and spectrogram images. (b) RGB components of the modified image. (c) Fused image. (d) Fused image of a normal ECG segment. (e) Fused image of an apnea ECG segment.

https://doi.org/10.1371/journal.pone.0250618.g005

Proposed model

The proposed CNN architecture is shown in Fig 1, where the CNN comprises four residual blocks sharing the same architecture but with different hyperparameters. The CNN has 37 layers, including the convolutional, max-pooling, dense, and other layers. Overall, there are 13 convolutional layers and four max-pooling layers. The model starts with 32 convolutional filters (3 × 3) followed by a rectified linear (ReLU) activation layer. The output is then passed to a series of residual blocks, as shown in Fig 1. Each residual block is formed as two consecutive convolutional layers and a skip connection through a 1 × 1 convolutional layer to restore the dimensionality. After the addition layer in each residual block, a ReLU activation layer and a 2 × 2 pooling layer with a stride size of two are utilized to summarize the feature map generated by each residual block. The number of filters (“f” in Fig 1) used in all convolutional layers in the same residual block is kept unchanged. The max-pooled output of each residual block is passed to the next residual block. The number of filters in a residual block is successively increased (32, 64, 96, and 128). The max-pooled output of the last residual block is then passed to a 0.1 dropout layer to avoid model overfitting. Adding a dropout layer is a recognized training technique whereby some nodes are dropped out randomly during training. This is a very effective regularization method, limiting overfitting and reducing the generalization error in a DNN model.

Finally, the flattened output of the dropout layer is passed to a fully connected layer with 256 units followed by “ReLU” activation. The classification layer is a soft-max layer, where the output of the network is normalized to a probability of y_k, as specified by Eq 4. The fully connected dense layer with 256 units followed by “ReLU” activation works as the classifier for the features derived from the deep stacked residual blocks. (4) where a_k is the activation (a linear weighted sum of the hidden nodes) of the k^th neuron in the soft-max layer, and y_k is the probability of the individual class.

Implementation of model training

The proposed model was implemented and trained using the MATLAB R2020a deep-learning toolbox [56]. The model was trained with graphics processing unit support (NVIDIA GEFORCE GTX 1070) using 10-fold cross validation [57]. We selected 20,000 “normal” ECG segments and 13,062 “OSA” ECG segments to train the proposed network. The image data (D) were randomly split into 10 equal subsets {f₁, f₂, f₃, …f_k, …, f₁₀}, with one subset chosen as the test dataset and the remainder used to train the model, resulting in one model for each of the 10 folds (see Fig 6).

Download:

Fig 6. Schematic diagram of the training procedure for the proposed 2D-CNN model with 10-fold cross validation.

https://doi.org/10.1371/journal.pone.0250618.g006

After determining the test and training datasets, random oversampling was performed to balance the dataset and prevent the model from being overfitted. Here, the training set comprised 18,000 normal one-minute ECG segments for each fold. To balance the dataset, we randomly copied apnea images (the minority class) so that the total number of OSA images was also 18,000. All training images, including the randomly oversampled images, were then subjected to fold-wise image augmentation using random rotation (–8 to +8 degrees), random horizontal translation (–30 to +30 pixels), vertical translation (–10 to +10 pixels), random shearing (–5 to +5 degrees), and random horizontal flipping. Small-scale augmentation was used because spectral images tend to be consistent and steady compared with normal still images taken by a camera, where high sample variation can occur, including large rotations and scaling, vivid colors, and special effects. The full training procedure is shown in Fig 6. In this procedure, the mini-batch size was 128, and the model was evaluated after every 256 iterations both to ensure that the model continued towards convergence during training and to visualize the training process (see Fig 10).

A back-propagation algorithm was used to train the whole model by optimizing the cross-entropy error E_ce between the predicted classes and the actual class, as specified in Eq 5, using the “Adam” optimizer with an initial learning rate of 0.001, as suggested in [58, 59]. Each fold was run for up to 48 epochs, until the training loss between consecutive batch updates ceased to improve. After training all models, the best model for each fold was selected for evaluation, according to its validation accuracy. (5) where y_n,k is the actual output of node k, n is the number of examples in the mini-batch, and t_n,k ϵ{0, 1} are the target outputs [60].

Performance evaluation

To evaluate the proposed model, we used overall accuracy, per-class recall (RE), per-class precision (PR), per-class specificity (SP), and per-class F₁ score (F1) as defined in Eqs (6–10), respectively [61]. As described in the Method section, the average result of the 10 folds for each performance metric was calculated to reflect the final performance of the proposed CNN. (6)

RE (also known as the probability of detection, true positive rate, or sensitivity) reflects the correctly predicted proportion of all positive samples. (7)

PR (also known as the positive predictive value) reflects the proportion of positive predictions that are actually correct. (8)

SP (also known as true negative rate) reflects the proportion of negatives that are correctly detected. (9)

The F1 score denotes the harmonic mean of PR and RE, which considers both metrics to give an optimal measure for analyzing model performance. (10) where TP, TN, FP, and FN denote true positives, true negatives, false positives, and false negatives, respectively.

Results

In this study, an image-based method for OSA detection in one-minute ECG segments is proposed. To investigate the effectiveness of the proposed method, we compared it with other existing methods. Because we used all subjects in training the network using a 10-fold cross-validation method, we compared only the per-segment OSA-detection performance. It should be noted that the test dataset was isolated from the training data before performing random oversampling and was not subjected to image augmentation.

Figs 7 and 8 show the validation accuracies for each fold during the 10-fold cross validation and the confusion matrices, respectively. Fig 9 shows interquartile range (IQR) plots for performance metrics calculated across all folds.

Download:

Fig 7. Distributions of validation accuracy for TFR images and fused images over 10 folds.

https://doi.org/10.1371/journal.pone.0250618.g007

Download:

Fig 8. Confusion matrices for per-segment apnea detection, with classwise PR and RE shown in the bottom and right-hand boxes, respectively: (a) Wigner–Ville distribution images, (b) scalogram images, (c) spectrogram images, and (d) fused images.

https://doi.org/10.1371/journal.pone.0250618.g008

Download:

Fig 9. IQR plots of PR, RE, and F1 for apnea detection obtained across all folds.

The center line indicates the median, the box limits indicate the upper and lower quartiles, the whiskers indicate 1.5 × IQR, and × indicates the mean. The images are Wigner–Ville distribution images (wg), scalogram images (sc), spectrogram images (sp), or fused images (fu).

https://doi.org/10.1371/journal.pone.0250618.g009

Table 1 shows the overall macro average of performance metrics for the proposed model with 10-fold cross validation. The table shows that the performance measures are very similar for all image types other than the Wigner–Ville distribution images.

Download:

Table 1. Overall performance in per-segment apnea detection TFR images and fused images.

https://doi.org/10.1371/journal.pone.0250618.t001

When considering the validation accuracy for all folds, as shown in Fig 7, there is no great variation between folds, which indicates that the model can be generalizable to other datasets.

As shown in Fig 9, the other performance metrics, including the F₁ score, also show very small variation across the folds. Although the means of the performance metrics are nearly identical for scalogram and spectrogram images, the performance metrics show slightly higher variability across the folds in the scalogram case. The overall accuracy and F1 scores for per-segment OSA detection with scalogram images are 90.5% and 88.3%, respectively. The same measures for the spectrogram dataset are 90.5% and 88.4%. However, the proposed CNN model achieves the highest performance for the fused images, achieving 92.4% overall accuracy and a 90.6% F1 score. The variability of all measures for the fused images is slightly higher than for the corresponding scalograms and spectrograms. The lowest performance occurs in all cases for the Wigner–Ville distribution images, with the greatest variation in performance metrics across the folds.

Fig 10 shows an accuracy loss plot for the weakest classifier of fused images. The learning curves confirm that the parameters selected for image creation and the CNN are appropriate for discriminating between OSA and normal ECG segments.

Download:

Fig 10. Accuracy-loss graph of the proposed CNN (for the lowest-performing model).

https://doi.org/10.1371/journal.pone.0250618.g010

Robustness evaluation

The PhysioNet Apnea-ECG database is a relatively small dataset, with the withheld dataset and the training dataset containing 35 recordings each. Therefore, using a single withheld dataset for validation might be unfair, given that we are training a deep-learning model, which requires more data than the other machine-learning methods. To address this, we used 10-fold cross validation to test the robustness of the proposed CNN with the entire dataset (70 recordings), which was randomly divided into 10 subsets, as shown in Fig 6. Fig 11 shows the average accuracy and F1 score for per-segment OSA detection, with their 95% confidence intervals (CIs) calculated for 10 cross-validation steps. According to Fig 11, the proposed CNN model demonstrates consistent performance for all image types (excluding the Wigner–Ville distribution) in terms of validation accuracy and F1 score with a small 95% CI. The model obtained accuracies of 90.5±0.3%, 90.5±0.3%, and 92.4±0.5% for per-segment OSA detection using scalogram, spectrogram, and fused images, respectively. Similarly, the F1 scores were 88.3±0.3%, 88.4±0.3%, and 90.6±0.6%, for the scalogram, spectrogram, and fused images, respectively. For the pseudo Wigner–Ville images, significantly weaker performance (accuracy = 82.99±3.49% and F1 score = 79.19±7.31%) is observed because the Wigner–Ville distribution method has cross-term issues.

Download:

Fig 11. Overall 10-fold cross-validation results for per-segment apnea detection with Wigner–Ville distribution, scalogram, spectrogram, and fused images.

Black lines indicate the corresponding 95% confidence interval.

https://doi.org/10.1371/journal.pone.0250618.g011

Comparison with existing methods

Because the PhysioNet Apnea-ECG database has been available for some time, several automatic OSA-detection approaches exist in the literature. Here, we compare our method with those that also used the PhysioNet Apnea-ECG database. Note that we do not consider per-recording detection performance because we trained our network via 10-fold cross validation after aggregating the entire data set.

Table 2 summarizes the performance of the proposed CNN method relative to other existing methods, with respect to per-segment OSA detection. As shown in the table, our method achieved the best performance in terms of overall accuracy, sensitivity, and specificity. In particular, we can compare our method with that of Tao Wang et al. [62]. Their method obtained an overall accuracy, sensitivity, and specificity of 87.3%, 85.1%, and 88.7%, respectively for the withheld dataset, whereas our method achieved an average accuracy, sensitivity, and specificity of 92.4%, 92.3%, and 92.6%, respectively, representing a significant performance improvement with the same dataset. Furthermore considering the robustness evaluation, we improved the accuracy by ≈6%, with a smaller CI (≈±0.45%) (for 10-fold cross validation) than their ±1.5% evaluated using 7-fold cross validation for the entire dataset.

Download:

Table 2. Performance comparison of proposed and previous methods for per-segment apnea detection.

https://doi.org/10.1371/journal.pone.0250618.t002

In addition, when compared with most previous methods [5, 17, 23, 24] using the same dataset, our method has the best apnea-detection confidence, as shown in Table 2 for all performance metrics. For example, Singh et al. [38] proposed an image-based OSA-detection method that used a CNN model based on AlexNet. This method obtained a validating accuracy of 86.2%, a sensitivity of 90%, and a specificity of 83.8% using scalogram images (227, 227, 3). Although their model showed comparatively good sensitivity, our model achieved much better performance in terms of validation accuracy and specificity in addition to sensitivity. Their prediction model is a plain DNN model with five convolutional layers compared with our better-performing model that contains four residual blocks.

Although other recent approaches [8, 17, 26, 27] have performed well for this dataset, our approach has demonstrated better performance metrics because the proposed method adopts recent advances in deep learning that enable the most appropriate features to be extracted automatically. In these other studies, specific features are derived via the ECG or EDR signal, forcing the classifier to depend on manually derived features. However, it should be noted that deep learning models require balanced datasets for optimal performance, unlike other models. Another possible drawback of the proposed method is that it requires conversion of the time-domain signal into two separate TFRs before making its predictions.

Discussion

The proposed CNN performs well because it is trained using the entire dataset with 10-fold cross validation, which avoids overfitting and provides greater sample variation in the training. In addition, we randomly oversample the apnea images and perform image augmentation, which balances the dataset and improves the number of examples by creating modified versions of the images, including the fused images. The augmented training dataset helps to create skillful models and improves the ability to generalize the model for unseen data. Most importantly, the fused images we used in our study provide a satisfactory blend of discriminative features, where both CWT and STFT features are hybridized.

In contrast to other methods, note that our model detects OSA segments without segmenting the QRS complexes in ECG signals, which provides a robust approach to OSA detection. Moreover, the model can be adapted to predict apneic events using arbitrarily long ECG segments (e.g., 10 s or 20 s) because we use RGB images that can be generated for any ECG-segment length and resized to (128, 128, 3) for input to the model.

Conclusions

The purpose of this study is to describe the implementation of a robust automatic OSA-detection method based on fused TFR images. Our CNN model for per-segment OSA detection detects OSA events using images corresponding to one-minute ECG segments in any of the Wigner–Ville distribution, scalogram, spectrogram, or fused-image formats. The results for accuracy and other performance metrics demonstrate that our model not only labels apneic events automatically but also outperforms existing methods for automatic OSA detection. Our model achieved an overall accuracy of 92.4% for fused images created from scalogram and spectrogram images. Another important aspect of this work is that the model can be used with arbitrary ECG-segment lengths because the ECG segments are converted to RGB images before input into the prediction model. Moreover, no manual feature extraction is required, which depends on the experience and specific domain knowledge of the researchers. Because our model is based on a single-lead ECG channel, it could be used in wearable electronics or a smart home-monitoring systems. This would be cheaper and more convenient than having to use a conventional sleep-study environment. However, our method has some limitations. Because the PhysioNet Apnea-ECG database provides only two types of annotation (apnea and normal), the proposed model cannot classify apnea subtypes (e.g., hypopnea). In future work, we aim to extend our model to detect these different types of apnea. In addition, we will consider modifying the proposed approach using multiple apnea datasets and will investigate different fusing techniques to improve the performance.

Supporting information

S1 File.

https://doi.org/10.1371/journal.pone.0250618.s001

(PDF)

Acknowledgments

The authors would like to thank Dr. Thomas Penzel of Phillips University, Marburg, Germany, for contributing the Apnea-ECG Database.

References

1. Quan S, Gillin JC, Littner M, Shepard J. Sleep-related breathing disorders in adults: Recommendations for syndrome definition and measurement techniques in clinical research. editorials. Sleep (New York, NY). 1999;22(5):662–689.
- View Article
- Google Scholar
2. Punjabi NM. The epidemiology of adult obstructive sleep apnea. Proceedings of the American Thoracic Society. 2008;5(2):136–143. pmid:18250205
- View Article
- PubMed/NCBI
- Google Scholar
3. Gubbi J, Khandoker A, Palaniswami M. Classification of sleep apnea types using wavelet packet analysis of short-term ECG signals. Journal of clinical monitoring and computing. 2012;26(1):1–11. pmid:22190269
- View Article
- PubMed/NCBI
- Google Scholar
4. Khandoker AH, Palaniswami M, Karmakar CK. Support vector machines for automated recognition of obstructive sleep apnea syndrome from ECG recordings. IEEE transactions on information technology in biomedicine. 2008;13(1):37–48.
- View Article
- Google Scholar
5. Varon C, Caicedo A, Testelmans D, Buyse B, Van Huffel S. A novel algorithm for the automatic detection of sleep apnea from single-lead ECG. IEEE Transactions on Biomedical Engineering. 2015;62(9):2269–2278. pmid:25879836
- View Article
- PubMed/NCBI
- Google Scholar
6. Coleman J. Complications of snoring, upper airway resistance syndrome, and obstructive sleep apnea syndrome in adults. Otolaryngologic Clinics of North America. 1999;32(2):223–234. pmid:10385533
- View Article
- PubMed/NCBI
- Google Scholar
7. Nieto FJ, Young TB, Lind BK, Shahar E, Samet JM, Redline S, et al. Association of sleep-disordered breathing, sleep apnea, and hypertension in a large community-based study. Jama. 2000;283(14):1829–1836. pmid:10770144
- View Article
- PubMed/NCBI
- Google Scholar
8. Viswabhargav CS, Tripathy R, Acharya UR. Automated detection of sleep apnea using sparse residual entropy features with various dictionaries extracted from heart rate and EDR signals. Computers in biology and medicine. 2019;108:20–30. pmid:31003176
- View Article
- PubMed/NCBI
- Google Scholar
9. Ruehland WR, Rochford PD, O’Donoghue FJ, Pierce RJ, Singh P, Thornton AT. The new AASM criteria for scoring hypopneas: impact on the apnea hypopnea index. sleep. 2009;32(2):150–157. pmid:19238801
- View Article
- PubMed/NCBI
- Google Scholar
10. Sadr N, de Chazal P. A comparison of three ECG-derived respiration methods for sleep apnoea detection. Biomedical Physics & Engineering Express. 2019;5(2):025027.
- View Article
- Google Scholar
11. Roebuck A, Monasterio V, Gederi E, Osipov M, Behar J, Malhotra A, et al. A review of signals used in sleep analysis. Physiological measurement. 2013;35(1):R1. pmid:24346125
- View Article
- PubMed/NCBI
- Google Scholar
12. Boudaoud S, Rix H, Meste O, Heneghan C, O’Brien C. Corrected integral shape averaging applied to obstructive sleep apnea detection from the electrocardiogram. EURASIP Journal on Advances in Signal Processing. 2007;2007:1–12.
- View Article
- Google Scholar
13. Penzel T, McNames J, De Chazal P, Raymond B, Murray A, Moody G. Systematic comparison of different algorithms for apnoea detection based on electrocardiogram recordings. Medical and Biological Engineering and Computing. 2002;40(4):402–407. pmid:12227626
- View Article
- PubMed/NCBI
- Google Scholar
14. Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on Image Processing. 2017;26(7):3142–3155. pmid:28166495
- View Article
- PubMed/NCBI
- Google Scholar
15. Bacharova L, Triantafyllou E, Vazaios C, Tomeckova I, Paranicova I, Tkacova R. The effect of obstructive sleep apnea on QRS complex morphology. Journal of electrocardiology. 2015;48(2):164–170. pmid:25541278
- View Article
- PubMed/NCBI
- Google Scholar
16. Gupta S, Cepeda-Valery B, Romero-Corral A, Shamsuzzaman A, Somers VK, Pressman GS. Association between QRS duration and obstructive sleep apnea. Journal of Clinical Sleep Medicine. 2012;8(6):649–654. pmid:23243398
- View Article
- PubMed/NCBI
- Google Scholar
17. Sharma H, Sharma K. An algorithm for sleep apnea detection from single-lead ECG using Hermite basis functions. Computers in biology and medicine. 2016;77:116–124. pmid:27543782
- View Article
- PubMed/NCBI
- Google Scholar
18. Almazaydeh L, Elleithy K, Faezipour M, Abushakra A. Apnea detection based on respiratory signal classification. Procedia Computer Science. 2013;21:310–316.
- View Article
- Google Scholar
19. Smruthy A, Suchetha M. Real-Time Classification of Healthy and Apnea Subjects Using ECG Signals With Variational Mode Decomposition. IEEE Sensors Journal. 2017;17(10):3092–3099.
- View Article
- Google Scholar
20. Sezgin N, Tagluk ME. Energy based feature extraction for classification of sleep apnea syndrome. Computers in biology and medicine. 2009;39(11):1043–1050. pmid:19762012
- View Article
- PubMed/NCBI
- Google Scholar
21. Guilleminault C, Winkle R, Connolly S, Melvin K, Tilkian A. Cyclical variation of the heart rate in sleep apnoea syndrome: mechanisms, and usefulness of 24 h electrocardiography as a screening technique. The Lancet. 1984;323(8369):126–131. pmid:6140442
- View Article
- PubMed/NCBI
- Google Scholar
22. Khandoker AH, Palaniswami M, Karmakar CK. Support vector machines for automated recognition of obstructive sleep apnea syndrome from ECG recordings. IEEE transactions on information technology in biomedicine. 2008;13(1):37–48.
- View Article
- Google Scholar
23. Song C, Liu K, Zhang X, Chen L, Xian X. An Obstructive Sleep Apnea Detection Approach Using a Discriminative Hidden Markov Model From ECG Signals. IEEE Transactions on Biomedical Engineering. 2016;63(7):1532–1542. pmid:26560867
- View Article
- PubMed/NCBI
- Google Scholar
24. Li K, Pan W, Li Y, Jiang Q, Liu G. A method to detect sleep apnea based on deep neural network and hidden markov model using single-lead ECG signal. Neurocomputing. 2018;294:94–101.
- View Article
- Google Scholar
25. Hayano J, Watanabe E, Saito Y, Sasaki F, Fujimoto K, Nomiyama T, et al. Screening for obstructive sleep apnea by cyclic variation of heart rate. Circulation: Arrhythmia and Electrophysiology. 2011;4(1):64–72. pmid:21075771
- View Article
- PubMed/NCBI
- Google Scholar
26. Tripathy R, Gajbhiye P, Acharya UR. Automated sleep apnea detection from cardio-pulmonary signal using bivariate fast and adaptive EMD coupled with cross time–frequency analysis. Computers in Biology and Medicine. 2020;120:103769. pmid:32421659
- View Article
- PubMed/NCBI
- Google Scholar
27. Singh H, Tripathy RK, Pachori RB. Detection of sleep apnea from heart beat interval and ECG derived respiration signals using sliding mode singular spectrum analysis. Digital Signal Processing. 2020;104:102796.
- View Article
- Google Scholar
28. Wang T, Lu C, Shen G, Hong F. Sleep apnea detection from a single-lead ECG signal with automatic feature-extraction through a modified LeNet-5 convolutional neural network. PeerJ. 2019;7:e7731. pmid:31579607
- View Article
- PubMed/NCBI
- Google Scholar
29. Chandra BS, Sastry CS, Jana S. Robust heartbeat detection from multimodal data via CNN-based generalizable information fusion. IEEE Transactions on Biomedical Engineering. 2018;66(3):710–717. pmid:30004868
- View Article
- PubMed/NCBI
- Google Scholar
30. Isin A, Ozdalili S. Cardiac arrhythmia detection using deep learning. Procedia computer science. 2017;120:268–275.
- View Article
- Google Scholar
31. Hammad M, Iliyasu AM, Subasi A, Ho ES, Abd El-Latif AA. A Multitier Deep Learning Model for Arrhythmia Detection. IEEE Transactions on Instrumentation and Measurement. 2020;70:1–9.
- View Article
- Google Scholar
32. Sedik A, Iliyasu AM, El-Rahiem A, Abdel Samea ME, Abdel-Raheem A, Hammad M, et al. Deploying machine and deep learning models for efficient data-augmented detection of COVID-19 infections. Viruses. 2020;12(7):769. pmid:32708803
- View Article
- PubMed/NCBI
- Google Scholar
33. Hammad M, Alkinani MH, Gupta B, Abd El-Latif AA. Myocardial infarction detection based on deep neural network on imbalanced data. Multimedia Systems. 2021; p. 1–13.
- View Article
- Google Scholar
34. Alghamdi A, Hammad M, Ugail H, Abdel-Raheem A, Muhammad K, Khalifa HS, et al. Detection of myocardial infarction based on novel deep transfer learning methods for urban healthcare in smart cities. Multimedia tools and applications. 2020; p. 1–22.
- View Article
- Google Scholar
35. Türk Ö, Özerdem MS. Epilepsy detection by using scalogram based convolutional neural network from EEG signals. Brain sciences. 2019;9(5):115. pmid:31109020
- View Article
- PubMed/NCBI
- Google Scholar
36. Byeon YH, Pan SB, Kwak KC. Intelligent deep models based on scalograms of electrocardiogram signals for biometrics. Sensors. 2019;19(4):935. pmid:30813332
- View Article
- PubMed/NCBI
- Google Scholar
37. McNames J, Fraser A. Obstructive sleep apnea classification based on spectrogram patterns in the electrocardiogram. In: Computers in Cardiology 2000. Vol. 27 (Cat. 00CH37163). IEEE; 2000. p. 749–752.
38. Singh SA, Majumder S. A novel approach osa detection using single-lead ECG scalogram based on deep neural network. Journal of Mechanics in Medicine and Biology. 2019;19(04):1950026.
- View Article
- Google Scholar
39. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. circulation. 2000;101(23):e215–e220. pmid:10851218
- View Article
- PubMed/NCBI
- Google Scholar
40. Penzel T, Moody GB, Mark RG, Goldberger AL, Peter JH. The apnea-ECG database. In: Computers in Cardiology 2000. Vol. 27 (Cat. 00CH37163). IEEE; 2000. p. 255–258.
41. Almazaydeh L, Elleithy K, Faezipour M. Detection of obstructive sleep apnea through ECG signal features. In: 2012 IEEE International Conference on Electro/Information Technology; 2012. p. 1–6.
42. de Chazal P, Penzel T, Heneghan C. Automated detection of obstructive sleep apnoea at different time scales using the electrocardiogram. Physiological measurement. 2004;25(4):967. pmid:15382835
- View Article
- PubMed/NCBI
- Google Scholar
43. Sejdic E, Djurovic I, Stankovic L. Quantitative Performance Analysis of Scalogram as Instantaneous Frequency Estimator. IEEE Transactions on Signal Processing. 2008;56(8):3837–3845.
- View Article
- Google Scholar
44. Ullah A, Anwar SM, Bilal M, Mehmood RM. Classification of Arrhythmia by Using Deep Learning with 2-D ECG Spectral Image Representation. Remote Sensing. 2020;12(10):1685.
- View Article
- Google Scholar
45. Huang J, Chen B, Yao B, He W. ECG arrhythmia classification using STFT-based spectrogram and convolutional neural network. IEEE Access. 2019;7:92871–92880.
- View Article
- Google Scholar
46. The MathWorks Inc. Wavelet Toolbox; 2020. Available from: https://www.mathworks.com/help/wavelet/.
47. The MathWorks Inc. Signal Processing Toolbox; 2020. Available from: https://www.mathworks.com/help/signal/.
48. Boashash B. Time-frequency signal analysis and processing: a comprehensive reference. Academic Press; 2015.
49. Khan A, Sohail A, Zahoora U, Qureshi AS. A survey of the recent architectures of deep convolutional neural networks. Artificial Intelligence Review. 2020;53(8):5455–5516.
- View Article
- Google Scholar
50. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2016. p. 770–778.
51. Khan A, Sohail A, Zahoora U, Qureshi AS. A survey of the recent architectures of deep convolutional neural networks. Artificial Intelligence Review. 2020;53(8):5455–5516.
- View Article
- Google Scholar
52. Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on Image Processing. 2017;26(7):3142–3155. pmid:28166495
- View Article
- PubMed/NCBI
- Google Scholar
53. Kumar P, Sharma VK. Detection and classification of ECG noises using decomposition on mixed codebook for quality analysis. Healthcare Technology Letters. 2020;7(1):18–24. pmid:32190336
- View Article
- PubMed/NCBI
- Google Scholar
54. Ren Z, Pandit V, Qian K, Yang Z, Zhang Z, Schuller B. Deep sequential image features on acoustic scene classification. In: Proc. DCASE Workshop, Munich, Germany; 2017. p. 113–117.
55. Jayalakshmy S, Sudha GF. Scalogram based prediction model for respiratory disorders using optimized convolutional neural networks. Artificial Intelligence in Medicine. 2020;103:101809. pmid:32143805
- View Article
- PubMed/NCBI
- Google Scholar
56. The MathWorks Inc. Deep Learning Toolbox; 2020. Available from: https://www.mathworks.com/help/deeplearning/.
57. Duda RO, Hart PE, et al. Pattern classification and scene analysis. vol. 3. Wiley New York; 1973.
58. Choi D, Shallue CJ, Nado Z, Lee J, Maddison CJ, Dahl GE. On Empirical Comparisons of Optimizers for Deep Learning. arXiv preprint arXiv:191005446. 2019;.
59. Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980. 2014;.
60. Xu SS, Mak MW, Cheung CC. Towards End-to-End ECG Classification With Raw Signal Extraction and Deep Neural Networks. IEEE Journal of Biomedical and Health Informatics. 2019;23(4):1574–1584. pmid:30235153
- View Article
- PubMed/NCBI
- Google Scholar
61. Powers D. Evaluation: From Precision, Recall and F-Factor to ROC, Informedness, Markedness & Correlation. Mach Learn Technol. 2008;2.
- View Article
- Google Scholar
62. Wang T, Lu C, Shen G. Detection of Sleep Apnea from Single-Lead ECG Signal Using a Time Window Artificial Neural Network. BioMed research international. 2019;2019. pmid:31950061
- View Article
- PubMed/NCBI
- Google Scholar

[ref1] 1. Quan S, Gillin JC, Littner M, Shepard J. Sleep-related breathing disorders in adults: Recommendations for syndrome definition and measurement techniques in clinical research. editorials. Sleep (New York, NY). 1999;22(5):662–689.
View Article
Google Scholar

[2] View Article

[3] Google Scholar

[ref2] 2. Punjabi NM. The epidemiology of adult obstructive sleep apnea. Proceedings of the American Thoracic Society. 2008;5(2):136–143. pmid:18250205
View Article
PubMed/NCBI
Google Scholar

[5] View Article

[6] PubMed/NCBI

[7] Google Scholar

[ref3] 3. Gubbi J, Khandoker A, Palaniswami M. Classification of sleep apnea types using wavelet packet analysis of short-term ECG signals. Journal of clinical monitoring and computing. 2012;26(1):1–11. pmid:22190269
View Article
PubMed/NCBI
Google Scholar

[9] View Article

[10] PubMed/NCBI

[11] Google Scholar

[ref4] 4. Khandoker AH, Palaniswami M, Karmakar CK. Support vector machines for automated recognition of obstructive sleep apnea syndrome from ECG recordings. IEEE transactions on information technology in biomedicine. 2008;13(1):37–48.
View Article
Google Scholar

[13] View Article

[14] Google Scholar

[ref5] 5. Varon C, Caicedo A, Testelmans D, Buyse B, Van Huffel S. A novel algorithm for the automatic detection of sleep apnea from single-lead ECG. IEEE Transactions on Biomedical Engineering. 2015;62(9):2269–2278. pmid:25879836
View Article
PubMed/NCBI
Google Scholar

[16] View Article

[17] PubMed/NCBI

[18] Google Scholar

[ref6] 6. Coleman J. Complications of snoring, upper airway resistance syndrome, and obstructive sleep apnea syndrome in adults. Otolaryngologic Clinics of North America. 1999;32(2):223–234. pmid:10385533
View Article
PubMed/NCBI
Google Scholar

[20] View Article

[21] PubMed/NCBI

[22] Google Scholar

[ref7] 7. Nieto FJ, Young TB, Lind BK, Shahar E, Samet JM, Redline S, et al. Association of sleep-disordered breathing, sleep apnea, and hypertension in a large community-based study. Jama. 2000;283(14):1829–1836. pmid:10770144
View Article
PubMed/NCBI
Google Scholar

[24] View Article

[25] PubMed/NCBI

[26] Google Scholar

[ref8] 8. Viswabhargav CS, Tripathy R, Acharya UR. Automated detection of sleep apnea using sparse residual entropy features with various dictionaries extracted from heart rate and EDR signals. Computers in biology and medicine. 2019;108:20–30. pmid:31003176
View Article
PubMed/NCBI
Google Scholar

[28] View Article

[29] PubMed/NCBI

[30] Google Scholar

[ref9] 9. Ruehland WR, Rochford PD, O’Donoghue FJ, Pierce RJ, Singh P, Thornton AT. The new AASM criteria for scoring hypopneas: impact on the apnea hypopnea index. sleep. 2009;32(2):150–157. pmid:19238801
View Article
PubMed/NCBI
Google Scholar

[32] View Article

[33] PubMed/NCBI

[34] Google Scholar

[ref10] 10. Sadr N, de Chazal P. A comparison of three ECG-derived respiration methods for sleep apnoea detection. Biomedical Physics & Engineering Express. 2019;5(2):025027.
View Article
Google Scholar

[36] View Article

[37] Google Scholar

[ref11] 11. Roebuck A, Monasterio V, Gederi E, Osipov M, Behar J, Malhotra A, et al. A review of signals used in sleep analysis. Physiological measurement. 2013;35(1):R1. pmid:24346125
View Article
PubMed/NCBI
Google Scholar

[39] View Article

[40] PubMed/NCBI

[41] Google Scholar

[ref12] 12. Boudaoud S, Rix H, Meste O, Heneghan C, O’Brien C. Corrected integral shape averaging applied to obstructive sleep apnea detection from the electrocardiogram. EURASIP Journal on Advances in Signal Processing. 2007;2007:1–12.
View Article
Google Scholar

[43] View Article

[44] Google Scholar

[ref13] 13. Penzel T, McNames J, De Chazal P, Raymond B, Murray A, Moody G. Systematic comparison of different algorithms for apnoea detection based on electrocardiogram recordings. Medical and Biological Engineering and Computing. 2002;40(4):402–407. pmid:12227626
View Article
PubMed/NCBI
Google Scholar

[46] View Article

[47] PubMed/NCBI

[48] Google Scholar

[ref14] 14. Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on Image Processing. 2017;26(7):3142–3155. pmid:28166495
View Article
PubMed/NCBI
Google Scholar

[50] View Article

[51] PubMed/NCBI

[52] Google Scholar

[ref15] 15. Bacharova L, Triantafyllou E, Vazaios C, Tomeckova I, Paranicova I, Tkacova R. The effect of obstructive sleep apnea on QRS complex morphology. Journal of electrocardiology. 2015;48(2):164–170. pmid:25541278
View Article
PubMed/NCBI
Google Scholar

[54] View Article

[55] PubMed/NCBI

[56] Google Scholar

[ref16] 16. Gupta S, Cepeda-Valery B, Romero-Corral A, Shamsuzzaman A, Somers VK, Pressman GS. Association between QRS duration and obstructive sleep apnea. Journal of Clinical Sleep Medicine. 2012;8(6):649–654. pmid:23243398
View Article
PubMed/NCBI
Google Scholar

[58] View Article

[59] PubMed/NCBI

[60] Google Scholar

[ref17] 17. Sharma H, Sharma K. An algorithm for sleep apnea detection from single-lead ECG using Hermite basis functions. Computers in biology and medicine. 2016;77:116–124. pmid:27543782
View Article
PubMed/NCBI
Google Scholar

[62] View Article

[63] PubMed/NCBI

[64] Google Scholar

[ref18] 18. Almazaydeh L, Elleithy K, Faezipour M, Abushakra A. Apnea detection based on respiratory signal classification. Procedia Computer Science. 2013;21:310–316.
View Article
Google Scholar

[66] View Article

[67] Google Scholar

[ref19] 19. Smruthy A, Suchetha M. Real-Time Classification of Healthy and Apnea Subjects Using ECG Signals With Variational Mode Decomposition. IEEE Sensors Journal. 2017;17(10):3092–3099.
View Article
Google Scholar

[69] View Article

[70] Google Scholar

[ref20] 20. Sezgin N, Tagluk ME. Energy based feature extraction for classification of sleep apnea syndrome. Computers in biology and medicine. 2009;39(11):1043–1050. pmid:19762012
View Article
PubMed/NCBI
Google Scholar

[72] View Article

[73] PubMed/NCBI

[74] Google Scholar

[ref21] 21. Guilleminault C, Winkle R, Connolly S, Melvin K, Tilkian A. Cyclical variation of the heart rate in sleep apnoea syndrome: mechanisms, and usefulness of 24 h electrocardiography as a screening technique. The Lancet. 1984;323(8369):126–131. pmid:6140442
View Article
PubMed/NCBI
Google Scholar

[76] View Article

[77] PubMed/NCBI

[78] Google Scholar

[ref22] 22. Khandoker AH, Palaniswami M, Karmakar CK. Support vector machines for automated recognition of obstructive sleep apnea syndrome from ECG recordings. IEEE transactions on information technology in biomedicine. 2008;13(1):37–48.
View Article
Google Scholar

[80] View Article

[81] Google Scholar

[ref23] 23. Song C, Liu K, Zhang X, Chen L, Xian X. An Obstructive Sleep Apnea Detection Approach Using a Discriminative Hidden Markov Model From ECG Signals. IEEE Transactions on Biomedical Engineering. 2016;63(7):1532–1542. pmid:26560867
View Article
PubMed/NCBI
Google Scholar

[83] View Article

[84] PubMed/NCBI

[85] Google Scholar

[ref24] 24. Li K, Pan W, Li Y, Jiang Q, Liu G. A method to detect sleep apnea based on deep neural network and hidden markov model using single-lead ECG signal. Neurocomputing. 2018;294:94–101.
View Article
Google Scholar

[87] View Article

[88] Google Scholar

[ref25] 25. Hayano J, Watanabe E, Saito Y, Sasaki F, Fujimoto K, Nomiyama T, et al. Screening for obstructive sleep apnea by cyclic variation of heart rate. Circulation: Arrhythmia and Electrophysiology. 2011;4(1):64–72. pmid:21075771
View Article
PubMed/NCBI
Google Scholar

[90] View Article

[91] PubMed/NCBI

[92] Google Scholar

[ref26] 26. Tripathy R, Gajbhiye P, Acharya UR. Automated sleep apnea detection from cardio-pulmonary signal using bivariate fast and adaptive EMD coupled with cross time–frequency analysis. Computers in Biology and Medicine. 2020;120:103769. pmid:32421659
View Article
PubMed/NCBI
Google Scholar

[94] View Article

[95] PubMed/NCBI

[96] Google Scholar

[ref27] 27. Singh H, Tripathy RK, Pachori RB. Detection of sleep apnea from heart beat interval and ECG derived respiration signals using sliding mode singular spectrum analysis. Digital Signal Processing. 2020;104:102796.
View Article
Google Scholar

[98] View Article

[99] Google Scholar

[ref28] 28. Wang T, Lu C, Shen G, Hong F. Sleep apnea detection from a single-lead ECG signal with automatic feature-extraction through a modified LeNet-5 convolutional neural network. PeerJ. 2019;7:e7731. pmid:31579607
View Article
PubMed/NCBI
Google Scholar

[101] View Article

[102] PubMed/NCBI

[103] Google Scholar

[ref29] 29. Chandra BS, Sastry CS, Jana S. Robust heartbeat detection from multimodal data via CNN-based generalizable information fusion. IEEE Transactions on Biomedical Engineering. 2018;66(3):710–717. pmid:30004868
View Article
PubMed/NCBI
Google Scholar

[105] View Article

[106] PubMed/NCBI

[107] Google Scholar

[ref30] 30. Isin A, Ozdalili S. Cardiac arrhythmia detection using deep learning. Procedia computer science. 2017;120:268–275.
View Article
Google Scholar

[109] View Article

[110] Google Scholar

[ref31] 31. Hammad M, Iliyasu AM, Subasi A, Ho ES, Abd El-Latif AA. A Multitier Deep Learning Model for Arrhythmia Detection. IEEE Transactions on Instrumentation and Measurement. 2020;70:1–9.
View Article
Google Scholar

[112] View Article

[113] Google Scholar

[ref32] 32. Sedik A, Iliyasu AM, El-Rahiem A, Abdel Samea ME, Abdel-Raheem A, Hammad M, et al. Deploying machine and deep learning models for efficient data-augmented detection of COVID-19 infections. Viruses. 2020;12(7):769. pmid:32708803
View Article
PubMed/NCBI
Google Scholar

[115] View Article

[116] PubMed/NCBI

[117] Google Scholar

[ref33] 33. Hammad M, Alkinani MH, Gupta B, Abd El-Latif AA. Myocardial infarction detection based on deep neural network on imbalanced data. Multimedia Systems. 2021; p. 1–13.
View Article
Google Scholar

[119] View Article

[120] Google Scholar

[ref34] 34. Alghamdi A, Hammad M, Ugail H, Abdel-Raheem A, Muhammad K, Khalifa HS, et al. Detection of myocardial infarction based on novel deep transfer learning methods for urban healthcare in smart cities. Multimedia tools and applications. 2020; p. 1–22.
View Article
Google Scholar

[122] View Article

[123] Google Scholar

[ref35] 35. Türk Ö, Özerdem MS. Epilepsy detection by using scalogram based convolutional neural network from EEG signals. Brain sciences. 2019;9(5):115. pmid:31109020
View Article
PubMed/NCBI
Google Scholar

[125] View Article

[126] PubMed/NCBI

[127] Google Scholar

[ref36] 36. Byeon YH, Pan SB, Kwak KC. Intelligent deep models based on scalograms of electrocardiogram signals for biometrics. Sensors. 2019;19(4):935. pmid:30813332
View Article
PubMed/NCBI
Google Scholar

[129] View Article

[130] PubMed/NCBI

[131] Google Scholar

[ref37] 37. McNames J, Fraser A. Obstructive sleep apnea classification based on spectrogram patterns in the electrocardiogram. In: Computers in Cardiology 2000. Vol. 27 (Cat. 00CH37163). IEEE; 2000. p. 749–752.

[ref38] 38. Singh SA, Majumder S. A novel approach osa detection using single-lead ECG scalogram based on deep neural network. Journal of Mechanics in Medicine and Biology. 2019;19(04):1950026.
View Article
Google Scholar

[134] View Article

[135] Google Scholar

[ref39] 39. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. circulation. 2000;101(23):e215–e220. pmid:10851218
View Article
PubMed/NCBI
Google Scholar

[137] View Article

[138] PubMed/NCBI

[139] Google Scholar

[ref40] 40. Penzel T, Moody GB, Mark RG, Goldberger AL, Peter JH. The apnea-ECG database. In: Computers in Cardiology 2000. Vol. 27 (Cat. 00CH37163). IEEE; 2000. p. 255–258.

[ref41] 41. Almazaydeh L, Elleithy K, Faezipour M. Detection of obstructive sleep apnea through ECG signal features. In: 2012 IEEE International Conference on Electro/Information Technology; 2012. p. 1–6.

[ref42] 42. de Chazal P, Penzel T, Heneghan C. Automated detection of obstructive sleep apnoea at different time scales using the electrocardiogram. Physiological measurement. 2004;25(4):967. pmid:15382835
View Article
PubMed/NCBI
Google Scholar

[143] View Article

[144] PubMed/NCBI

[145] Google Scholar

[ref43] 43. Sejdic E, Djurovic I, Stankovic L. Quantitative Performance Analysis of Scalogram as Instantaneous Frequency Estimator. IEEE Transactions on Signal Processing. 2008;56(8):3837–3845.
View Article
Google Scholar

[147] View Article

[148] Google Scholar

[ref44] 44. Ullah A, Anwar SM, Bilal M, Mehmood RM. Classification of Arrhythmia by Using Deep Learning with 2-D ECG Spectral Image Representation. Remote Sensing. 2020;12(10):1685.
View Article
Google Scholar

[150] View Article

[151] Google Scholar

[ref45] 45. Huang J, Chen B, Yao B, He W. ECG arrhythmia classification using STFT-based spectrogram and convolutional neural network. IEEE Access. 2019;7:92871–92880.
View Article
Google Scholar

[153] View Article

[154] Google Scholar

[ref46] 46. The MathWorks Inc. Wavelet Toolbox; 2020. Available from: https://www.mathworks.com/help/wavelet/.

[ref47] 47. The MathWorks Inc. Signal Processing Toolbox; 2020. Available from: https://www.mathworks.com/help/signal/.

[ref48] 48. Boashash B. Time-frequency signal analysis and processing: a comprehensive reference. Academic Press; 2015.

[ref49] 49. Khan A, Sohail A, Zahoora U, Qureshi AS. A survey of the recent architectures of deep convolutional neural networks. Artificial Intelligence Review. 2020;53(8):5455–5516.
View Article
Google Scholar

[159] View Article

[160] Google Scholar

[ref50] 50. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2016. p. 770–778.

[ref51] 51. Khan A, Sohail A, Zahoora U, Qureshi AS. A survey of the recent architectures of deep convolutional neural networks. Artificial Intelligence Review. 2020;53(8):5455–5516.
View Article
Google Scholar

[163] View Article

[164] Google Scholar

[ref52] 52. Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on Image Processing. 2017;26(7):3142–3155. pmid:28166495
View Article
PubMed/NCBI
Google Scholar

[166] View Article

[167] PubMed/NCBI

[168] Google Scholar

[ref53] 53. Kumar P, Sharma VK. Detection and classification of ECG noises using decomposition on mixed codebook for quality analysis. Healthcare Technology Letters. 2020;7(1):18–24. pmid:32190336
View Article
PubMed/NCBI
Google Scholar

[170] View Article

[171] PubMed/NCBI

[172] Google Scholar

[ref54] 54. Ren Z, Pandit V, Qian K, Yang Z, Zhang Z, Schuller B. Deep sequential image features on acoustic scene classification. In: Proc. DCASE Workshop, Munich, Germany; 2017. p. 113–117.

[ref55] 55. Jayalakshmy S, Sudha GF. Scalogram based prediction model for respiratory disorders using optimized convolutional neural networks. Artificial Intelligence in Medicine. 2020;103:101809. pmid:32143805
View Article
PubMed/NCBI
Google Scholar

[175] View Article

[176] PubMed/NCBI

[177] Google Scholar

[ref56] 56. The MathWorks Inc. Deep Learning Toolbox; 2020. Available from: https://www.mathworks.com/help/deeplearning/.

[ref57] 57. Duda RO, Hart PE, et al. Pattern classification and scene analysis. vol. 3. Wiley New York; 1973.

[ref58] 58. Choi D, Shallue CJ, Nado Z, Lee J, Maddison CJ, Dahl GE. On Empirical Comparisons of Optimizers for Deep Learning. arXiv preprint arXiv:191005446. 2019;.

[ref59] 59. Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980. 2014;.

[ref60] 60. Xu SS, Mak MW, Cheung CC. Towards End-to-End ECG Classification With Raw Signal Extraction and Deep Neural Networks. IEEE Journal of Biomedical and Health Informatics. 2019;23(4):1574–1584. pmid:30235153
View Article
PubMed/NCBI
Google Scholar

[183] View Article

[184] PubMed/NCBI

[185] Google Scholar

[ref61] 61. Powers D. Evaluation: From Precision, Recall and F-Factor to ROC, Informedness, Markedness & Correlation. Mach Learn Technol. 2008;2.
View Article
Google Scholar

[187] View Article

[188] Google Scholar

[ref62] 62. Wang T, Lu C, Shen G. Detection of Sleep Apnea from Single-Lead ECG Signal Using a Time Window Artificial Neural Network. BioMed research international. 2019;2019. pmid:31950061
View Article
PubMed/NCBI
Google Scholar

[190] View Article

[191] PubMed/NCBI

[192] Google Scholar

Figures

Abstract

Introduction

Related work

Materials and methods

Dataset

Method

Preprocessing and image creation

Proposed model

Implementation of model training

Performance evaluation

Results

Robustness evaluation

Comparison with existing methods

Discussion

Conclusions

Supporting information

S1 File.

Acknowledgments

References