06-12-2012, 04:41 PM
Apnea MedAssist
Apnea MedAssist.doc (Size: 2.43 MB / Downloads: 32)
Abstract
I would like to propose a low-cost, real-time sleep apnea monitoring system ‘‘Apnea MedAssist” for recognizing obstructive sleep apnea episodes with a high degree of accuracy for both home and clinical care applications. The fully automated system uses patient’s single channel nocturnal ECG to extract feature sets, and uses the support vector classifier (SVC) to detect apnea episodes. “Apnea MedAssist” is implemented on Android operating system (OS) based Smartphone’s, uses either the general adult subject-independent SVCmodel or subject-dependent SVCmodel, and achieves a classification F-measure of 90% and a sensitivity of 96% for the subject-independent SVC. The real-time capability comes from the use of 1-min segments of ECG epochs for feature extraction and classification. The reduced complexity of “Apnea MedAssist” comes from efficient optimization of the ECG processing, and use of techniques to reduce SVC model complexity by reducing the dimension of feature set from ECG and ECG-derived respiration signals and by reducing the number of support vectors.
Introduction
The project involves:
• Capturing data from external device
• Processing the data
• Generating reports
• Communication with smartphones
APNEA is a sleep related breathing disorder—commonly known as obstructive sleep apnea (OSA) is a common disorder that affects about 4% of the general population. People with sleep apnea literally stop breathing repeatedly during their sleep, often for 10–30 s and as many as hundreds of times during one night. Sleep apnea can be caused by complete “apnea” or partial “hypopnea” obstruction of airway , both of which can wake one up. The frequent arousals and the inability to achieve or maintain the deeper stages of sleep can lead to excessive daytime sleepiness, nonrestorative sleep, automobile accidents, personality changes, decreased memory, erectile dysfunction (impotence), and depression. OSA has also been linked to angina, nocturnal cardiac arrhythmias, myocardial infarction, and stroke.
The primary method for diagnosing OSA at present is to have the patient undergo a sleep study, known as polysomnography (PSG). A polysomnogram typically records a minimum of eleven channels of various biosignals requiring a minimum of 22 wire attachments to the patient in a specialized sleep laboratory with attended personnel. Obstructive sleep apnea is diagnosed, if the patient has an apnea index (AI) (apneic episodes per hour) greater than 5/h, or a respiratory disturbance index, the combination of apneas and hypopneas, greater than 10/h. Several treatment options exist for OSA. These include weight reduction, oral appliances, positional therapy, continuous positive airway pressure (CPAP) therapy, and surgical options. CPAP, the most common of these therapies is usually administered at bedtime through a nasal or facial mask held in place by velcro straps around the patients head.
METHODOLOGY
This paper describes the design of “Apnea MedAssist,” a reliable automated OSA-monitoring device that uses measurements from just one lead ECG sensor. Fig. 2 shows the functional flow diagram for the signal processing and episode classification implemented on an Android-based smartphone.
Subjects Database
The device and algorithms were tested using Physionet Apnea-ECG Database . The database has a total of 35 subjects’ sleep studies. The recordings were visually scored by an expert for sleep apnea/hypopnea events on the basis of respiration and oxygen on a per minute basis. The subjects’ recordings (30 men, 5 women) were arranged in three groups: Group A recordings (20 subjects) with clear occurrence of sleep apnea (100 min or more, AHI = 15), Group B (borderline) recordings (five subjects) with some degree of sleep apnea (between 5 and 99 min, 5 = AHI < 15), and Group C (control) recordings (ten subjects) of healthy subjects with no sleep apnea (fewer than 5 min, AHI < 5).
For apnea scoring, each record was divided into 1-min nonoverlapping segments . Each minute was classified as either a “nonapnea minute” or an “apnea minute.” Minutes containing either apnea or hypopnea were classified as apnea minutes. The AI is the number of apneas observed per hour, and the HI is the number of hypopneas observed per hour. The apnea–hypopnea index (AHI) is defined as the sum of AI and HI. Hours containing one to four apnea minutes (not considered to be clinically significant) are counted as hours without apnea. Table I shows the collected subjects’ data. This segment length reduces hidden apnea episodes that actually occur within the segment. Clearly, as we increase the segment length over 1 min, the actual estimate of AHI deviates considerably.
Automated ECG Processing
The ECG measurements with a sampling period of 4 m/s are segmented into 1-min epochs and then analyzed using “ECG MedAssist” signal-processing module. This is an automated wavelet-based analysis algorithm for denoising and detrending ECG signal, and detecting its characteristic points: QRS complex, P, and T waves. The wavelet transform algorithm used here is based on the undecimated lifting scheme (ULWT) . The ULWT has reduced computational cost compared to the basic finite-impulse response implementation. We use a loworder Daubechies/D4 wavelet to minimize the filter order and consequently reduce computational cost and overall filter delay. Single-decomposition phase with seven stages yields details {Ds}7s =1 and approximations C7 . Signal denoising is implemented by zeroing out D1 detail coefficients. Detrending is performed by zeroing out C7 approximation coefficients.
To separate QRS complex from P–T waves, we extract two signals by reconstructing two groups of ULWT subbands. {Ds}5s =2 are used for QRS-complex signal reconstruction (SQRS) and {Ds}7s =5 are used for P and T waves reconstruction (SPT). We also apply real-time adaptive thresholding to all details before each reconstruction stage to minimize spectral overlap between the QRS and the PT signals and also to emphasize the core details (those details defining either QRS or PT waves).
Respiration Efforts Estimate (EDR and RSA)
The morphology changes in the ECG waves allow deriving a signal proportional to the respiratory movement. In the literature, various methods were proposed to extract the surrogate EDR signal , based on R-wave amplitude, R-wave duration, QRS complex area, T-wave amplitude or T-wave area. We use the T-wave method, since it is more suited for ECG recording with low sampling rates (<250 Hz) as in our case. However, we use the R-wave amplitude method on recordings with undetected T-wave (or inverted T-wave). Fig. 3 shows a 20-min segment of one of the recordings with the computed EDR and corresponding respiratory signals obtained using other respiratory sensor measurements.
The calculated EDR sequence is analyzed to detect and remove any outliers due to false detections or missed beats using a sliding window averaging filter (window of 41 samples) with an empirically selected exclusion threshold (±70%). The extracted respiratory time series is denoted as {EDR(q) : edri, i = 10, . . . , q}.
Learning and Classification Phases
Even though SVMs can generalize well, a careful choice of kernel function is necessary to produce a classification boundary that effectively separates (+/–)OSA classes. We use the automated system, aforementioned in two phases; training and classification (see in Fig. 5). Features with most predictive value are selected during feature search. During the learning phase we find the best performing classifier kernel type and parameters, and solve theSVMQP problem to generate the set of SVs for the optimum separating hyperplane for the OSA’s two classes.
In classification phase the SVM classifies test epoch feature vector x into (+/–)OSA classes. We use here K-fold crossverification method to evaluate the performance of the classifier. The cross-validation method can prevent the overfitting problem. Here, K depends on the classifier input training dataset size for both subject-dependent (SVC-SD) and subjectindependent (SVC-SI) classifiers. For K-fold method, we divide the training set into K subsets of equal size. Sequentially one subset is tested using the SVM classifier trained on the remaining K-1 subsets. The cross-validation accuracy is the percentage of data that are correctly classified.