Comparative Performance Analysis of Time-Frequency Domain Images and Raw Signal Data for Classification of ECG Signals

Turgut Özseven

Makale detayları

Comparative Performance Analysis of Time-Frequency Domain Images and Raw Signal Data for Classification of ECG Signals

Yazarlar:

Turgut Özseven

Yayıncı:

Duzce University

Ana Başlıklar:

Mühendislik

Dergi:

Düzce Üniversitesi Bilim ve Teknoloji Dergisi Vol.12 Issue 2, pp.745-761

Belge Tipi:

Makale

Yayın Tarihi:

4/2024

Belge Dili:

İngilizce

Document ID (Org No.):

2001315882

pISSN:

2148-2446

Anahtar kelimeler:

Özet:

 ECG signals are one of the most common tools used to diagnose cardiovascular diseases. ECG signals are obtained by measuring electrical changes on the skin surface. Arrhythmias occurring in the heart are diagnosed because the expert evaluates ECG signals. This diagnosis depends on the experience of the specialist and is a subjective evaluation. With the widespread use of computer-aided diagnostic systems, evaluations dependent on the experts experience are objectified, and support is provided to the physician for diagnosis. For computer-aided ECG classification, beats are detected from ECG signals, and arrhythmias are detected by analyzing the structure of these beats. In recent years, deep learning models have been successful in classifying ECG signals. The data to be used in the classification process is realized with the help of morphological features or images of the signal. The main objective of this study is to compare the classification performance of digital and visual heartbeat data for ECG signal classification. For this purpose, 1D-CNN and 2D-CNN architectures are used for the type of ECG signals. As inputs of the 1D-CNN model, numerical values of the heartbeat signal and hand-crafted features obtained from these numerical values were used. The inputs of the 2D-CNN model are the raw signal image, spectrogram, scalogram, Mel-spectrogram, GFCC, and CQT images, which are visual representations of the heartbeat signal. The results show that the numerical model of the ECG signal fails for classification, while the hand-crafted features provide 85.2 accuracy. The results obtained with the visual representation of the signal provided over 99 classification accuracy for all images. The highest success rate was 99.9 with the visualization of the raw signal. In line with these findings, the 2D-CNN architecture and the visual representation of the heartbeat signal were found to be the most suitable method for classifying ECG signals.

 ECG signals are one of the most common tools used to diagnose cardiovascular diseases. ECG signals are obtained by measuring electrical changes on the skin surface. Arrhythmias occurring in the heart are diagnosed because the expert evaluates ECG signals. This diagnosis depends on the experience of the specialist and is a subjective evaluation. With the widespread use of computer-aided diagnostic systems, evaluations dependent on the experts experience are objectified, and support is provided to the physician for diagnosis. For computer-aided ECG classification, beats are detected from ECG signals, and arrhythmias are detected by analyzing the structure of these beats. In recent years, deep learning models have been successful in classifying ECG signals. The data to be used in the classification process is realized with the help of morphological features or images of the signal. The main objective of this study is to compare the classification performance of digital and visual heartbeat data for ECG signal classification. For this purpose, 1D-CNN and 2D-CNN architectures are used for the type of ECG signals. As inputs of the 1D-CNN model, numerical values of the heartbeat signal and hand-crafted features obtained from these numerical values were used. The inputs of the 2D-CNN model are the raw signal image, spectrogram, scalogram, Mel-spectrogram, GFCC, and CQT images, which are visual representations of the heartbeat signal. The results show that the numerical model of the ECG signal fails for classification, while the hand-crafted features provide 85.2 accuracy. The results obtained with the visual representation of the signal provided over 99 classification accuracy for all images. The highest success rate was 99.9 with the visualization of the raw signal. In line with these findings, the 2D-CNN architecture and the visual representation of the heartbeat signal were found to be the most suitable method for classifying ECG signals.

 ECG signals are one of the most common tools used to diagnose cardiovascular diseases. ECG signals are obtained by measuring electrical changes on the skin surface. Arrhythmias occurring in the heart are diagnosed because the expert evaluates ECG signals. This diagnosis depends on the experience of the specialist and is a subjective evaluation. With the widespread use of computer-aided diagnostic systems, evaluations dependent on the experts experience are objectified, and support is provided to the physician for diagnosis. For computer-aided ECG classification, beats are detected from ECG signals, and arrhythmias are detected by analyzing the structure of these beats. In recent years, deep learning models have been successful in classifying ECG signals. The data to be used in the classification process is realized with the help of morphological features or images of the signal. The main objective of this study is to compare the classification performance of digital and visual heartbeat data for ECG signal classification. For this purpose, 1D-CNN and 2D-CNN architectures are used for the type of ECG signals. As inputs of the 1D-CNN model, numerical values of the heartbeat signal and hand-crafted features obtained from these numerical values were used. The inputs of the 2D-CNN model are the raw signal image, spectrogram, scalogram, Mel-spectrogram, GFCC, and CQT images, which are visual representations of the heartbeat signal. The results show that the numerical model of the ECG signal fails for classification, while the hand-crafted features provide 85.2 accuracy. The results obtained with the visual representation of the signal provided over 99 classification accuracy for all images. The highest success rate was 99.9 with the visualization of the raw signal. In line with these findings, the 2D-CNN architecture and the visual representation of the heartbeat signal were found to be the most suitable method for classifying ECG signals.

 EKG sinyalleri kardiyovasküler hastalıkların klinik tanısı için kullanılan en yaygın araçlardan birisidir. Cilt yüzeyindeki elektriksel değişimlerin ölçülmesi ile EKG sinyalleri elde edilmektedir. EKG sinyallerinin uzmanın değerlendirmesi sonucu kalpte oluşan aritmiler teşhis edilmektedir. Bu teşhis uzmanın deneyimine bağlı olup subjektif bir değerlendirmedir. Bilgisayar destekli tanı sistemlerinin yaygınlaşması ile uzmanın deneyimine bağımlı değerlendirmeler objektifleşmekte ve hekime tanı için destek sağlanmaktadır. Bilgisayar destekli EKG sınıflandırma için EKG sinyallerinden atımlardan tespit edilmekte ve bu atımların yapısı incelenerek aritmiler tespit edilmektedir. Son yıllarda derin öğrenme modellerindeki yüksek başarı EKG sinyallerinin de sınıflandırılması için kullanılmaya başlanmıştır. Sınıflandırma sürecinde kullanılacak veri sinyalin morfolojik özellikleri veya görüntüsü yardımıyla gerçekleştirilmektedir. Bu çalışmanın temel amacı, EKG sinyallerinin sınıflandırılması için sayısal ve görsel kalp ritmi verilerinin sınıflandırma performanslarının karşılaştırılmasıdır. Bu amaçla, EKG sinyallerinin sınıflandırılması için 1D-CNN ve 2D-CNN mimarileri kullanılmıştır. 1D-CNN modelinin girdileri olarak kalp ritmi sinyalinin sayısal değerleri ve bu sayısal değerlerden elde edilen öznitelikler kullanılmıştır. 2D-CNN modelinin girdisi kalp ritmi sinyallinin görsel olarak temsilini içeren ham sinyal görüntüsü, spektrogram, skalogram, mel-spektrogram, GFCC ve CQT görüntüleridir. Elde edilen sonuçlar, EKG sinyallerinin sayısal temsilinin sınıflandırma için başarısız olduğunu, hand-crafted özniteliklerin 85.2 doğruluk sağladığını göstermiştir. Sinyalin görsel temsili ile elde edilen sonuçlar tüm görüntüler için 99 üzerinde sınıflandırma doğruluğu sağlamıştır. Bunlar içerisindeki en yüksek başarı ise sinyalin ham halinin görselleştirilmesi ile 99.9 olarak elde edilmiştir. Elde edilen bu bulgular doğrultusunda, EKG sinyallerinin sınıflandırılması için en uygun yöntemin 2D-CNN mimarisi ve kalp ritmi sinyalinin görsel temsili olduğunu göstermiştir.

Makale detayları

Comparative Performance Analysis of Time-Frequency Domain Images and Raw Signal Data for Classification of ECG Signals

Yazarlar:

Turgut Özseven

Yayıncı:

Duzce University

Ana Başlıklar:

Mühendislik

Dergi:

Düzce Üniversitesi Bilim ve Teknoloji Dergisi Vol.12 Issue 2, pp.745-761

Belge Tipi:

Makale

Yayın Tarihi:

4/2024

Belge Dili:

İngilizce

Document ID (Org No.):

2001315882

pISSN:

2148-2446

ECG

deep learning

heartbeat

classificationEKG

derin öğrenme

kalp ritmi

sınıflandırma

Sayı

Makale detayları

Yazarlar: Turgut Özseven

Yayıncı: Duzce University

Ana Başlıklar: Mühendislik

Dergi: Düzce Üniversitesi Bilim ve Teknoloji Dergisi Vol.12 Issue 2, pp.745-761

Belge Tipi: Makale

Yayın Tarihi: 4/2024

Belge Dili: İngilizce

Document ID (Org No.): 2001315882

pISSN: 2148-2446

Sayı

Yazarlar:

Turgut Özseven

Yayıncı:

Duzce University

Ana Başlıklar:

Mühendislik

Dergi:

Düzce Üniversitesi Bilim ve Teknoloji Dergisi Vol.12 Issue 2, pp.745-761

Belge Tipi:

Makale

Yayın Tarihi:

4/2024

Belge Dili:

İngilizce

Document ID (Org No.):

2001315882

pISSN:

2148-2446