A RR interval based automated apnea detection approach using residual network

Highlights

• Method achieved a good result of 94.4% accuracy for apnea detection in segment.

• Firstly employ deep residual network for apnea detectionbased on low-sample-rate data.

• Weights of network can be fine-tuned on clinical.

• Strong adaptivity. Model also works well using ECG-derived respiration (EDR) signal.

Abstract

Background and Objective

Apnea is one of the most common conditions that causes sleep-disorder breathing. With growing number of patients worldwide, more and more patients suffer from complications of apnea. But most of them stay untreated due to the complex and time-consuming polysomnography (PSG) diagnosis method. Effective and precise diagnosis support system using electrocardiograph (ECG) is required. In this paper, we propose an approach using residual network to detect apnea based on RR intervals (intervals between R-peaks of ECG signal).

Methods

In our model, we apply residual network to represent information carried by RR intervals. Moreover, we proposed a novel perspective, called dynamic autoregressive representation, to provide interpretation of representing the RR intervals by convolutional layers.

Results

This approach is tested for per-segment apnea detection using publicly available dataset on Physionet. 30 overnight recordings are used for training and 5 for testing. We achieve a good result of 94.4% accuracy, 93.0% sensitivity and 94.9% specificity. This result outperform other prevalent methods based on RR intervals. This model also shows its good adaptivity while using ECG-derived respiration signal (EDR) in experiments. Its extensiveness is evaluated and compared in experiments. The proposed model is also compared with deep neural networks using original ECG signals for apnea detection, and it achieves better result using fewer input samples.

Conclusions

We develop a deep residual network to detect apnea on low-sample-rate RR intervals. The result suggests a possibility of representing RR intervals by neural network. The model showed strong adaptivity when using EDR input.

Introduction

Apnea is a common condition that causes sleep-disorder breathing. According to the survey, this condition affects people around the world over 3 decades. The apnea patients took 2% of women and 4% of men in middle-aged work force in 1993 [1]. The prevalence of apnea came to be 4% of US population in 2003 [2], and finally reached 6% of adults worldwide in 2008, while it is predicted to keep rising [3]. The growing amount of apnea patients without diagnosis and interventions are exposed to the risk of cardiovascular diseases like hypertension and stroke [4], [5] as well as mentally suffering from clinic depression which may harm their social life [6]. Many experiments were launched to study the detriment of apnea, and the result suggests relevance between apnea and several native physiological phenomena and diseases. For example, apnea plays an important role in obesity [7]; the insomnia and narcolepsy brought by apnea may lead vehicle accidents [8]; people with sleep apnea was demonstrated to be in danger during operation under anesthesia [9], and significant high mortality risk was found on untreated apnea patients, despite age, sex and BMI [3].

Apnea is a kind of sleep-disordered breathing. It happens in rapid eye movement sleep while airway is completely blocked by the musculus genioglossus, musculus geniohyoideus and fatty tissue for at least 10 s. Apnea can be classified by the variety of respiratory effort. The central sleep apnea (CSA) is characterized by repeated breathing cessation for lack of respiratory effort [10]. The obstructive sleep apnea (OSA) happens when the patients stop breathing unrelated with the respiratory effort. OSA can happen hundreds of times per night in the sleep of those who meet the diagnostic criteria for sleep apnea syndrome, and finally resulted in the mentioned diseases [11].

Sleep apnea is long-term, detrimental but treatable under diagnosis. Therapies like positive airway pressure (PAP) treatment and palatopharyngoplasty (PPP) treatment are effective under early diagnosis [12], [13], [14], [15]. Hence, prompt diagnosis is essential in the treating process of apnea. Major indexes to diagnose the severity of apnea syndrome including apnea index (AI)(number of apnea per hour sleep), hypopnea index (HI)(number of hypopnea per hour sleep), apnea/ hypopnea index (AHI)(number of apnea and hypopnea per hour sleep) and apnea lasting time. These indexs are traditionally calculated by polysomnography (PSG) signals [16], [17], [18] which includes electrocardiogram (ECG), electroencephalogram (EEG), electrocardiogram (EOG), electromyography (EMG), respiratory effort, airflow and oxygen saturation (SaO₂) [13], [14], [19]. During the process of collecting PSG signal and detecting apnea, the patient should sleep with intrusive armamentarium overnight. Moreover, medical specialists and clinic environment are also required. The traditional apnea detection process is too complex, therefore new algorithms which can perform rapidly and accurately on non-intrusive device is required. One effective solution to this problem is using single lead ECG signals to detect apnea.

More and more apnea detection methods using single lead signal were proposed. There are studies using SaO₂ signal [20] and airflow signal [21], while more studies were carried out using single lead ECG signal in order to meet the non-intrusive demand and hardware conditions of wearable mobile equipment. The possibility of diagnosing apnea by single lead ECG was demonstrated in 1984 [15], then Physionet hold the competition called CinC Challenge 2000 [22], [23], which provided ECG data with minute-by-minute label of apnea [24], [25]. The dataset is publicly available in Physionet after the challenge. But only the training set’s label is presented (for 35 recordings). Recent years, more and more methods using single lead ECG to detect apnea occurring in minute were proposed and tested on Physionet dataset.

Some researchers attached their attention to design various features. For example, in [26], 7 features were extracted from ECG-derived respiration signal (EDR). Features extracted by an autoregressive model were used as the measurement in k nearest neighbors (KNN) classifier to accomplish more than 85% accuracy in [27]. Surrel et al. invented an apnea detection system based on wearable sensors using SVM classifier and 88 features extracted from several ECG-derived time-series. In [28], Bsoul et al. implement a real-time sleep apnea monitoring system with 111 features derived from time domain and spectral domain of R-peak interval (RR). Novel sparse residual entropy features were proposed by Tripathy et al. for sleep apnea detection reporting good performance in subject-specific validation[29].

Meanwhile, more and more classifiers was employed or invented to detect apnea. For example, the extreme learning machine were conducted by Tripathy using features extracted from the intrinsic band functions of both EDR and heart rate variability (HRV) signals for apnea detection [30]. Hidden Markov model for ECG was designed and utilized with 11 features in [31]. The adaptive boosting (AdaBoost) was adopted with normal inverse Gaussian parameters in [32]. Janbakhshi et al. integrate 5 classifiers for apnea detection [33]. Zhao et al. built a frequency network and measured the distance of HRV with dynamic time warping (DTW). In their work, only 1 feature of HRV was used to achieve accuracy of 90.1%, sensitivity of 88.29% and specificity of 90.5% [34].

Diagnosis support systems are very helpful for practitioners in ECG inspection tasks. But it is difficult to describe the morphological changes by artificial features [35], [36], since the transient and long-term abnormalities appear alternately with no regularity [37]. To seize the complex cardiac abnormalities, some deep neural networks are built and trained using ECG signals [38], [39], [40], [41], [42], [43]. Their researches provided precedents of employing convolutional neural networks (CNN) to detect disease using ECG signals. In apnea detection tasks, directly feeding original ECG signals to deep neural networks is adopted by some researchers [44], [45], [46], but the ECG signals sampled at 100Hz or so limited the depth of neural network. How to build deeper networks for apnea detection became a problem. Mendez et al. had proposed autoregressive model for apnea detection [27]. In their model, the HRV series were represented by the polynomial with decaying factor, whereby the convolution computation in CNN can provide adaptive factor in polynomial. This motivated us to employ CNN to RR intervals. The sample rate of RR intervals is around 1Hz, and there are several physiological research reveal the abnormalities in RR intervals when sleep apnea occurs [47], [48], [49], [50]. The low-sample-rate RR intervals can not only provide information of sleep apnea but also allow the neural network to go deeper. This motivated us to conduct a deep neural network using the low-sample-rate RR interval series, (Fig. 1).

In this paper, we propose an approach to detect apnea with residual network using RR intervals. In addition, we provide a novel perspective of representing RR intervals via convolutional layers, and it is called dynamic autoregressive representation. In traditional algorithms, the work of designing numerous artificial feature is time-consume. Besides, the complex calculation of features is not friendly for real-time system and wearable equipments. In our model, both artificial features and classifiers are supplanted by the neural network. All the computations of CNN are the multiplications of vectors. The computation speed of pre-trained CNN is still acceptable in theory. After training on graphic processing unit (GPU), the model of neural networks can be saved. Moreover, the network compression technologies make it possible to apply the deep algorithm on wearable equipment or smart phone [51], [52]. An advantage of represent ECG by network is flexibility. All the weights in networks can be fine tuned on clinical, but the artificial designed features can rarely be changed after employment. In addition, the neural networks get all the knowledge from the data. The information reduction through feature extraction is avoided.

This paper is organized as follows. In Section 2, we describe the dataset used in research, the pre-process step and model of CNN, along with the perspective called dynamic autoregressive representation. We evaluate the robustness and extensiveness by experiments in Section 3. In Section 4, we give a discussion of result and compare the proposed model with other approaches using RR intervals and ECG. Section 5 is the conclusion.