DERİN ÖĞRENME TABANLI HİBRİT MODEL TASARIMI İLE EKG SİNYAL SINIFLANDIRMASINDA ARİTMİ TESPİTİ
Yükleniyor...
Dosyalar
Tarih
2022-12
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Elektrokardiyografi (EKG), cilt üzerine yerleştirilen elektrotlar sayesinde kalbin elektriksel aktivitesini algılayarak elde edilen kalp grafiğinin okunup tanı konulması işlemidir. Elde edilen verilerle birlikte, hastanın fizik muayenesi, hastanın öyküsü ve diğer etmenler de göz önüne alınarak kalp ve damar hastalıkları teşhisi konulabilmektedir. Kalp hastalıkları hayati risk taşıyan ve insanların hayat kalitesini düşüren hastalıklardandır. Bu sebeplerle erken tanı önemli bir rol oynamaktadır. Makine öğrenmesi ve derin öğrenme yöntemleri sayesinde medikal alanda önemli gelişmeler elde edilmiştir. Günümüzde derin öğrenme yöntemi ile EKG cihazından elde edilen verilerdeki anormallikleri yorumlayıp sınıflandırarak ani ölümlerin önüne geçmek hedeflenmektedir. Bu çalışmada amaç, derin öğrenme mimarilerinden olan Konvolüsyonel Sinir Ağları (Convolutional Neural Network, CNN) ve Uzun Kısa Vadeli Bellek (Long Short-Term Memory, LSTM) modelleri kullanılarak hibrit bir sistem oluşturulup EKG sinyalleri üzerinde aritmileri tahmin etmek için bir çözüm üretmektir. Çalışmada Massachusetts Institute of Technology – Boston’s Beth Israel Hospital (MIT-BIH) aritmi veri setinde paylaşılmış olan 109.446 adet kalp sinyalleri 187 sütunda sayıllaştırılmış olup N: Non-ecotic beats (normal beat), S: Supraventricular ectopic beats, V: Ventricular ectopic beats, F: Fusion beats, Q: Unknown beats olmak üzere beş sınıfa ayrılmıştır. Veri setinde bulunan beş sınıftan veri sayısı en az olan F ve S sınıfı tek bir sınıf olarak birleştirilip Others isminde yeni sınıf olarak değerlendirilmiştir. Tüm sınıflar için CNN modelinde bulunan accuracy, specitivity, sensitivity, F1-score, Matthews correlation coefficient (MCC), Eğri Altındaki Alan (Area Under Curve, AUC) score ortalama değerleri sırasıyla %98.15, %98.58, %98.20, %98.09, 0.96, 0.997’dir. Tüm sınıflar için LSTM modelinde bulunan accuracy, specitivity, sensitivity, F1-score, MCC, AUC score ortalama değerleri sırasıyla %98.42, %98.67, %98.46, %98.34, 0.97, 0.997’dir. Tüm sınıflar için geliştirilen hibrit model ile bulunan accuracy, specitivity, sensitivity, F1-score, MCC, AUC score ortalama değerleri sırasıyla %98.48, %98.78, %98.58, %98.54, 0.97, 0.997’dir.
Electrocardiography (ECG) is the process of reading and diagnosing the heart graph obtained by sensing the electrical activity of the heart by means of electrodes placed on the skin. Together with the data obtained, the diagnosis of cardiovascular diseases can be made by considering the patient's physical examination, the patient's history and other factors. Heart diseases are life-threatening diseases that reduce the quality of people's lives. For these reasons, early diagnosis plays an important role. Thanks to machine learning and deep learning methods, important developments have been achieved in the medical field. Recently, it is aimed to prevent sudden deaths by interpreting and classifying the abnormalities in the data obtained from the ECG device with the deep learning method. The aim of this study is to create a hybrid system using Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) models, which are deep learning architectures, and to produce a solution to predict arrhythmias on ECG signals. In this study, 109,446 heart signals shared in the MIT-BIH arrhythmia dataset was quantified in 187 columns and N: Non-ecotic beats (normal beat), S: Supraventricular ectopic beats, V: Ventricular ectopic beats, F: Fusion beats, Q: Unknown beats. divided into five classes. F and S classes, which have the least amount of data from the five classes in the data set, were combined as a single class and evaluated as a new class named Others. The average values of accuracy, specitivity, sensitivity, F1-score, MCC, AUC score found in the CNN model for all classes are 98.15%, 98.58%, 98.20%, 98.09%, 0.96, and 0.997, respectively. The average values of accuracy, specitivity, sensitivity, F1-score, MCC, AUC score found in the LSTM model for all classes are 98.42%, 98.67%, 98.46%, 98.34%, 0.97, and 0.997, respectively. The average values of accuracy, specitivity, sensitivity, F1-score, MCC and AUC score found by the hybrid model developed for all classes are 98.48%, 98.78%, 98.58%, 98.54%, 0.97, 0.997, respectively."
Electrocardiography (ECG) is the process of reading and diagnosing the heart graph obtained by sensing the electrical activity of the heart by means of electrodes placed on the skin. Together with the data obtained, the diagnosis of cardiovascular diseases can be made by considering the patient's physical examination, the patient's history and other factors. Heart diseases are life-threatening diseases that reduce the quality of people's lives. For these reasons, early diagnosis plays an important role. Thanks to machine learning and deep learning methods, important developments have been achieved in the medical field. Recently, it is aimed to prevent sudden deaths by interpreting and classifying the abnormalities in the data obtained from the ECG device with the deep learning method. The aim of this study is to create a hybrid system using Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) models, which are deep learning architectures, and to produce a solution to predict arrhythmias on ECG signals. In this study, 109,446 heart signals shared in the MIT-BIH arrhythmia dataset was quantified in 187 columns and N: Non-ecotic beats (normal beat), S: Supraventricular ectopic beats, V: Ventricular ectopic beats, F: Fusion beats, Q: Unknown beats. divided into five classes. F and S classes, which have the least amount of data from the five classes in the data set, were combined as a single class and evaluated as a new class named Others. The average values of accuracy, specitivity, sensitivity, F1-score, MCC, AUC score found in the CNN model for all classes are 98.15%, 98.58%, 98.20%, 98.09%, 0.96, and 0.997, respectively. The average values of accuracy, specitivity, sensitivity, F1-score, MCC, AUC score found in the LSTM model for all classes are 98.42%, 98.67%, 98.46%, 98.34%, 0.97, and 0.997, respectively. The average values of accuracy, specitivity, sensitivity, F1-score, MCC and AUC score found by the hybrid model developed for all classes are 98.48%, 98.78%, 98.58%, 98.54%, 0.97, 0.997, respectively."
Açıklama
Anahtar Kelimeler
Aritmi, EKG sinyali, CNN, LSTM, derin öğrenme, Arrhythmia, ECG signal, CNN, LSTM, deep learning
