Detection of sleep apnea from analysis of single channel ECG recordings
Journal Name:
- Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi
Keywords (Original Language):
| Author Name | University of Author |
|---|---|
Abstract (2. Language):
Obstructive sleep apnea (OSA) is a common sleeprelated
breathing disorder. The prevalence of OSA
syndrome is known to be 4% in adult men and 2% in
adult women. Polysomnography is the gold standard
method used to diagnose OSA syndrome. This
method is performed by assessment of large number
of physiologic parameters recorded concurrently
throughout the night in hospitals' sleep laboratories.
However, this diagnostic method is expensive and
laborious. This indicates that cheaper methods,
which may be an alternative to conventional
diagnostic methods, are needed. In this study, an
alternative pattern recognition system that can
detect sleep apnea (OSA) from analysis of single
channel ECG recordings is proposed. The proposed
system consists of three basic stages: preprocessing,
feature extraction and classification. 60
ECG recordings collected from OSA syndrome and
normal subjects, each of approximately 8 hours in
duration, were used in the study. Of the records, 40
were obtained from patients with OSA and 20 from
normal subjects.
In the pre-processing stage of the proposed system,
the heart rate variability (HRV) of the ECG records
was determined by using an algorithm based on
discrete wavelet transform. Here, the wavelet
transform was used to decompose the ECG signal
into sub-bands. The D2 detail sub-band obtained by
this decomposition result was scanned with a
maximum finding algorithm to determine the R
peaks, and then the HRV pattern of the ECG was
obtained by calculating the distances between the R
peaks that follow each other.
In the feature extraction stage, the feature vectors
were determined by applying discrete wavelet
transform and entropy calculation methods to the
HRV pattern of ECG recording. In this stage, with
using Daubechies 4 (db4) wavelet, HRV pattern of
ECG recording was decomposed into the details
D1-D12 sub-bands and one final approximation,
A12 sub band. Then, entropy value of each details
sub-bands was calculated by using norm entropy.
Thus, each of the ECG recordings is represented
with a feature vector consisting of 12 samples.
In the classification stage, two different classifiers
were used to distinguish ECG recordings with OSA
from normal ECG recordings through the specified
feature vectors. One of these is a multilayer
perceptron (MLP) classifier with an artificial neural
network. This classifier, also known as back
propagation algorithm, consists of input and output
and intermediate layers. Except for the input layer,
each layer consists of artificial neurons with nonlinear
activation function. The MLP utilizes a
supervised learning strategy called back
propagation for training the network. Another
classifier used in the study is the support vector
machine (SVM). SVM is a relatively new technique
used for classification and regression tasks.
Standard SVM moves the input data set, which
cannot be linearly classified, into a high
dimensional feature space by means of kernel
functions. Then, in this space, the optimal separating
plane is determined to separate the input data from
each other. This optimal separation plane can be
determined by using the solution of a quadratic
programming problem.
The performance of the classifiers was evaluated
according to the 5-fold cross-validation test.
According to this evaluation, the accuracy
performance of the SVM classifier was 98.3% and
the accuracy performance of the MLP classifier was
96.7%. Based on the obtained results, it is
considered that the proposed system has potential
for recognition of patients with suspected OSA by
using ECG recordings.
With the proposed system, it can be determined only
whether the person is an OSA patient. However, the
severity of the OSA syndrome cannot be determined
with this system. This problem is the drawback of the
system. In the future work, we will focus on solving
this problem.
Bookmark/Search this post with
Abstract (Original Language):
Bu çalışmada, tek kanallı EKG kayıtlarının analizinden obstrüktif uyku apne (OUA) tespitini yapabilen bir
otomatik örüntü tanıma sistemi önerilmiştir. Çalışmada her biri ortalama 8 saat süreden oluşan 60 EKG
kaydı kullanılmıştır. Kayıtların 40’ı OUA hastası deneklerden 20’si ise normal deneklerden alınmıştır.
Önerilen sistem ön işlem, özellik çıkarımı ve sınıflandırma olmak üzere üç temel aşamadan oluşur. Ön işlem
aşamasında dalgacık dönüşümü tabanlı bir algoritma kullanılarak EKG kayıtlarının kalp hızı değişkenliği
(KHD) olarak da adlandırılan R-R aralıkları değişkenliği belirlenmiştir. Özellik çıkarım aşamasında ise
dalgacık dönüşümü ve entropi hesaplama yöntemleri belirlenmiş olan KHD örüntülerine uygulanarak EKG
kayıtlarını temsil eden özellik vektörleri çıkarılmıştır. Sınıflandırma aşamasında ise Destek Vektör Makinesi
(DVM) ve Yapay Sinir Ağı (YSA) sınıflandırıcıları kullanılarak OUA hastası ve normal EKG kayıtları
belirlenen özellik vektörleri üzerinden birbirinden ayırt edilmiştir. Sınıflandırıcıların başarımı 5 katlı çapraz
doğrulama testine göre değerlendirilmiştir. Bu değerlendirmeye göre DVM sınıflandırıcısının doğruluk
başarımı % 98.3 ve YSA sınıflandırıcısının doğruluk başarımı ise %96.7 olarak gerçeklenmiştir. Elde edilen
sonuçlar göz önüne alındığında önerilen sistemin OUA değerlendirilmesinde uzman hekime ön tanı imkânı
sağlayabileceği düşünülmektedir.
- 1