Ejeksiyon fraksiyonu düşük ve korunmuş kalp yetersizliği vakaları için tek sinyal kullanarak makine öğrenmesi tabanlı yeni bir tanı algoritması = A new machine learning-based diagnostic algorithm using single signal for cases of heart failure with low and preserved ejection fraction

Abstract

Kalp yetersizliği (KY) ile ilgili olarak, mortaliteyi azaltmak ve ömrü uzatmak tedavinin ana hedeflerinden biridir. Birçok klinik çalışma KY hastalarını Sol Ventriküler Ejeksiyon Fraksiyonuna (SVEF) göre tanımlar. SVEF'ye göre KY alt tiplerinden en sık görülen ikisi, düşük ejeksiyon fraksiyonlu KY (DEF-KY, SVEF ≤ %40) ve korunmuş ejeksiyon fraksiyonlu KY (KEF-KY, SVEF ≥ %50)'dir. DEF-KY tanısı daha kolaydır. Ancak KEF-KY tanısı uzmanlar için bile daha karmaşık ve zordur. SVEF normal göründüğünden (sağlıklı bireylerde de SVEF ≥ %50), bazı benzer semptomlar nedeniyle KEF-KY vakaları göğüs hastalıkları ile karıştırılabilir. SVEF'yi ölçmek için genellikle ekokardiyografi kullanılır. Bu, uzman gerektiren pahalı bir cihazdır ve cihaza ulaşmanın kısıtlı olduğu durumlar olabilir. Tedaviye ekokardiyografi yapılmadan başlanması gereken acil durumlar da olabilir. Bu tür durumların çözümü için ekonomik, pratik karar destek sistemlerine ihtiyaç duyulmaktadır. Bu tez çalışmasında, DEF-KY, sağlıklı ve KEF-KY sınıflarının olduğu üçlü sınıflandırma yapan, yalnızca 3 kablolu elektrokardiyogram (EKG) özelliklerini kullanan bir tanı algoritması ve yalnızca fotopletismografi (PPG) ve PPG'den türetilen kalp hızı değişkenliği (KHD) özelliklerini kullanan bir algoritma olmak üzere iki farklı algoritma geliştirilmiştir. Ayrıca DEF-KY ve KEF-KY ikili sınıflandırması için de sadece KHD özelliklerini kullanan başka bir algoritma da geliştirilmiştir. 25 yaş ve üzeri 61 gönüllüden 10 saniye boyunca elektrokardiyogram (EKG) ve fotopletismografi (PPG) verileri eş zamanlı olarak alınmıştır. Hem EKG hem de PPG sinyallerini temizlemek için dijital filtreler kullanılmıştır. Temizlenmiş PPG'lerden KHD'ler türetilmiştir. EKG ile yapılan üçlü sınıflandırma çalışması için dört farklı makine öğrenmesi (MÖ) algoritması kullanılmıştır. Veri setinin rasgele %80'i eğitim ve %20'si test veri seti olarak ayarlanmıştır. Bu sınıflandırmalar sonucunda elde edilen en yüksek doğruluk oranı, k-en yakın komşu (k-NN) algoritması için %100'dür. Yalnızca KHD özellikleri kullanılarak yapılan DEF-KY ve KEF-KY ikili sınıflandırması için üç farklı MÖ algoritması kullanılmıştır. Bu sınıflandırmalar sonucunda elde edilen en yüksek doğruluk oranı, Destek Vektör Makinaları (DVM) için %98.33'dir. PPG ve PPG'den türetilen KHD kullanılarak yapılan üçlü sınıflandırma için üç farklı MÖ algoritması kullanılmıştır. Yine 10 kat çapraz doğrulama ile değerlendirme yapılmıştır. Sınıflandırmalar sonucu elde edilen en yüksek doğruluk oranı, topluluk sınıflandırıcı için %87.78'dir. Bu çalışmada elde edilen sonuçlara dayanarak, sadece 3 kablolu EKG ya da sadece PPG özelliklerinin DEF-KY ve KEF-KY tanısında kullanılabileceği ve önemli sonuçlar sağlayacağı belirlenmiştir. Bu çalışma ile, tek bir sinyalle DEF-KY ve KEF-KY teşhis etme olasılığının yolu açılmıştır. Konuyla ilgili bundan sonraki araştırmalara öncülük edeceği düşünülmektedir.

Regarding heart failure (HF), reducing mortality and prolonging life is one of the main goals of treatment. HF is one of the most common causes of hospitalization and death for adults and is on the way to becoming a worldwide epidemic. There are many advances in the treatment of HF, but due to the heterogeneous pathophysiology, HF still contains many unexplained points. Many clinical studies define HF patients according to their Left Ventricular Ejection Fraction (LVEF). According to LVEF, the two subtypes of HF are HF with reduced ejection fraction (HFrEF, LVEF ≤ 40%) and HF with preserved ejection fraction (HFpEF, LVEF ≥ 50%). HF is a common and costly disease. Timely diagnosis is important for effective treatment, but diagnosis of HF at an early stage can be difficult. Early diagnosis of HF is difficult because the symptoms are not specific. The presence of atypical findings or comorbidities may complicate the diagnosis of HF. For example, in a patient diagnosed with chronic obstructive pulmonary disease (COPD), it may be unclear whether the progression of dyspnea is due to COPD or HF. Similarly, wheezing occurs when the fluid accumulated as a result of pulmonary congestion compresses the bronchioles from the outside. In the case of asthma and COPD, the symptomatology is the same, although the pressure originates inside the bronchioles. For this reason, even if the patient is treated for chest diseases, the diagnosis of HF may be overlooked. If the pulmonologist considers the possibility of HF and makes some blood tests used in the diagnosis of HF and the results are compatible with the findings of HF, the patient can be referred to a cardiologist and the case will not be missed. The rate of HFpEF is increasing among all HF patients. When HF patients in epidemiological studies are examined, approximately 50-55% of all HF cases are HFpEF. HFpEF is often seen in the elderly women patient population with a diagnosis of HF. Diagnosing HFpEF is more difficult than diagnosing HFrEF. The diagnosis of HFpEF is more complex and difficult even for specialists. The diagnosis of HFpEF is a problem that is tried to be solved in medicine. Since LVEF appears normal (LVEF ≥ 50% in healthy individuals), HFpEF cases can be confused with pulmonary diseases due to some similar symptoms. LVEF is a measure of how much of the blood coming into the left ventricle is pumped around the body. In a normal healthy individual, the maximum LVEF is 70%. In HFrEF, the left ventricle cannot contract adequately and pump enough blood to the body. Although left ventricular contraction and blood pumping appear normal in HFpEF, adequate filling cannot be achieved because left ventricular relaxation is insufficient. For this reason, the amount of blood required for the body cannot be supplied in sufficient quantity. Echocardiography is often used to measure LVEF. This is an expensive device that requires specialists and there may be situations where access to the device is limited. There may also be emergencies where treatment should be initiated without echocardiography. Economic and practical measurement and decision support systems are needed to solve such situations. Most guidelines advocate primary care physician evaluation to diagnose HF. Once a definitive diagnosis is reached, treatment can be started by a specialist or primary care physician. Primary care physicians have an important role in monitoring the general health status of patients. They are the most aware of comorbid conditions. Treatment of comorbidities may decrease HF symptoms. Patients be more easily detected by providing medical care, by doing more research based on their signs and symptoms, or by screening those at highest risk for HF. High-risk groups include those with COPD or type 2 diabetes, particularly those who are older. In these patients, symptoms may be interpreted as part of the aging process and may cause the diagnosis to be overlooked. Echocardiography is still not available to most primary care physicians. The studies carried out within the scope of this thesis are advantageous in terms of providing a portable, easily accessible and easily measurable measurement system for primary care specialists who do not have echocardiography for the screening of high-risk groups. B-type natriuretic peptide (BNP) blood tests are also performed to diagnose HF. BNP is a substance released from the ventricles into the blood in response to changes in blood pressure according to the HF level. Blood BNP levels increase as HF symptoms worsen and decrease once the condition is controlled. Even in a person with HF whose condition is under control, the BNP level is always higher than in a normal healthy person. In this thesis, two different algorithms were developed as a diagnostic algorithm using only 3-lead electrocardiogram (ECG) features and a diagnostic algorithm using only photoplethysmography (PPG) and PPG-derived heart rate variability (HRV) features, for triple classification of classes of HFrEF, healthy and HFpEF. In addition, an algorithm using only HRV features has been developed for the binary classification of HFrEF and HFpEF. This thesis study aimed to determine with a single signal whether the people who admitted to the hospital with HF symptoms are a HFrEF case, healthy or a HFpEF case. ECG and PPG data were taken simultaneously for 10 seconds from 61 volunteers aged 25 and over. To clean both ECG and PPG signals, Chebyshev Type II band-pass filter in the range of 0.25-100 Hz, notch filter in the range of 49-51 Hz for 50 Hz mains noise and moving average filter were used as digital filters. HRVs were derived from cleared PPGs. 21 features have been extracted from both ECG, PPG, and HRVs at time domain. To these features, 16 more features have been added, consisting of Burg method and Yule-Walker method output parameters. Thus, 37 features have been obtained from each signal. Afterwards, the features have been analyzed statistically. The distributions of the ECG, PPG and HRV are not normal. For this reason, statistical analyzes are made with non-parametric tests. The Mann Whitney U test is a nonparametric test that performed to establish whether the two groups belong to the same population. Kruskal Wallis test, also, is a non-parametric test method used in cases where the data are not normally distributed and there are three or more classes. Therefore, these two tests were preferred. The number of features decreased after the Kruskal Wallis test and further decreased with the experiments performed after the Mann Whitney U test. For the triple classification study using ECG, after the Kruskal Wallis test, the number of features decreased from 37 to 19. The features are ranked from most relevant to least relevant according to their Eta correlation coefficients. With the experiments performed after the Mann Whitney U test, the final result was reached in the relationship of higher accuracy / fewer feature. The highest accuracy classification has been made with 4 ECG features which were at the top of rankings. Four different machine learning (ML) algorithms, namely k-NN, Support Vector Machines (SVM), Decision Trees and Bag Trees Ensemble Classifier, were used. 80% of the dataset is randomly set as training and 20% as test dataset. The accuracy rates obtained as a result of these classifications are: 100% for k-NN, 97.22% for SVM, 91.66% for Decision Trees and 97.22% for ensemble classifier. For HFrEF and HFpEF binary classification using only HRV features, after the Mann Whitney U test, the number of features decreased from 37 to 21. The features are ranked from most relevant to least relevant according to their Eta correlation coefficients. A different number of highest ranking features were taken as input and classification studies were carried out. As a result, the best results were obtained from the classification using the features in the top three of the rankings. Three different ML algorithms, namely SVM, k-NN and Decision Trees, were used. Evaluation was made with 10-fold cross validation. The accuracy rates obtained as a result of these classifications are: 98.33% for SVM, 96.67% for k-NN and 95.83% for Decision Trees. For triple classification using PPG and PPG-derived HRV, The dataset included a total of 74 features, 37 of which were PPG features and 37 were HRV features. After the Kruskal Wallis test, the number of features decreased from 74 to 58. With the experiments performed after the Mann Whitney U test, the final result was reached in the relationship of higher accuracy / fewer feature. The classification with highest accuracy has been made using 8 features, five of which were PPG features and three of which were HRV features. Three different ML algorithms, namely k-NN, SVM and ensemble classifier, were used. Again, evaluation was made with 10-fold cross-validation. The accuracy rates obtained as a result of the classifications are: 82.22% for k-NN, 82.78% for SVM and 87.78% for ensemble classifier. Based on the results obtained in this study, it was determined that only 3-lead ECG or only PPG features can be used in the diagnosis of HFrEF and HFpEF and will provide important results. This study paved the way for the possibility of diagnosing HFrEF and HFpEF with a single signal. This is the point where the thesis study contributes to the literature. It is thought that it will lead the future researches on the subject.