WO2025106046A1 - Machine learning-based device for detection of heart failure cases with reduced and preserved ejection fraction - Google Patents

Abstract

The invention relates to a machine learning based device for detecting and classifying heart failure with reduced ejection fraction and heart failure with preserved ejection fraction. The invention, in particular, relates to a device that can perform a three-class classification by assessing in a single step whether an individual with symptoms of heart failure is a heart failure with reduced ejection fraction (HFrEF) or heart failure with preserved ejection fraction (HFpEF) or healthy.

Description

MACHINE LEARNING-BASED DEVICE FOR DETECTION OF HEART FAILURE CASES WITH REDUCED AND PRESERVED EJECTION FRACTION

Technical Field

The invention relates to a machine learning based device for detecting and classifying heart failure with reduced ejection fraction and heart failure with preserved ejection fraction.

In particular, the invention relates to a device capable of performing a three-class classification by assessing in a single step whether an individual with symptoms of heart failure is a case of heart failure with reduced ejection fraction (HFrEF) or heart failure with preserved ejection fraction (HFpEF) or healthy.

State of the Art

One of the priority investment areas in developed countries is medical technologies. Advances in these technologies help medical science to diagnose and treat many diseases. The leading diagnostic methods of these technological advances are medical imaging and signal processing. Signal processing is mostly based on mathematical models and has a complex structure. Heart failure (HF) can be defined as a cardiac structural or functional defect in which the heart is unable to deliver enough oxygen to meet the metabolic needs of tissues, despite normal filling pressures (or simply increased filling pressures) [1], Inadequate oxygen supply causes symptoms such as rapid fatigue, shortness of breath, pulse irregularity (arrhythmia) and negatively affects daily life. Ejection fraction (EF) is the amount of blood pumped from one ventricle to the body with each beat of the heart [2],

In a normal person, EF is between 50-70%. This means that between 50% and 70% of the blood coming into the heart is pumped to the body. Left ventricular ejection fraction (LVEF) is a measure of pumping efficiency to the systemic circulation, while right ventricular ejection fraction is a measure of pumping efficiency to the pulmonary circulation [2], LVEF is important not only as a marker of disease, but also because in most clinical trials patients are identified according to LVEF. LVEF is usually measured by echocardiography and is used as a general measure of a person's heart function. Echocardiography can be called the "gold standard" method for measuring EF. Single Photon Emission Computed Tomography (SPECT) and radionuclide ventriculography (MUGA) are also used. These devices are expensive, some methods are invasive and all require a specialist. There may also be situations where access to these devices is limited. The two most common subtypes of heart failure, defined by LVEF, are HF with reduced ejection fraction (HFrEF) and HF with preserved ejection fraction (HFpEF). The American Heart Association defines HF cases as HFrEF if LVEF < 40% and as HFpEF if LVEF > 50% [3], Main clinical trials in patients with systolic HF or HFrEF have generally included patients with EF < 40% and currently proven therapies have been shown to be effective only in this group of patients. On the other hand, studies have also been conducted in HF patients with EF >40-45% and no other cardiac disorders. In some of these patients, EF is completely normal (usually >50%) and systolic function is not significantly reduced. Therefore, the term HFpEF was developed to describe these patients [4], [5], The underlying pathophysiologic disorder in patients with HFpEF is thought to be LV diastolic dysfunction, and therefore the diagnosis of LV diastolic dysfunction is essential in the diagnosis of this type of HF. No single echocardiographic parameter is sufficiently accurate and reproducible to diagnose LV diastolic dysfunction. Therefore, a comprehensive echocardiographic examination including fully correlated two-dimensional and Doppler data is recommended [4], [6], [7], Doppler ultrasound devices are expensive, echocardiographic examination is laborious and requires a specialist physician. Because HFpEF is a complex syndrome that can be caused by structural or functional cardiac disorders rather than the presence of a single disease entity, accurate diagnosis can be difficult even for HF specialists [2], For HFrEF and HFpEF which is difficult to diagnose, require expensive devices and specialists, a method that more economical, easier to measure and advantageous for patient comfort, is needed. There is a need for a machine learning (ML) based medical decision support system that provides the physician with preliminary information within seconds whether the patient who is admitted to the hospital with symptoms and suspicion of HF has HFrEF, HFpEF or is healthy, prevents unnecessary expensive and laborious tests, and is beneficial in terms of both time and patient comfort, using only a single signal (Electrocardiography (ECG) or Photoplethysmography (PPG)).

There are several studies in the literature that use artificial intelligence (Al) and ML methods for various heart failure classifications [8]-[15], These studies have generally focused on HFrEF, with some examining only HFpEF. Classifications were made for two classes. The features used for classification were demographic information such as age, height and weight or more onerous features such as those obtained from echocardiography, magnetic resonance imaging or invasive methods. In recent years, Al algorithms have been developed to detect various heart diseases using ECG. In these studies, algorithms have been developed with Al methods using ECG for conditions such as aortic stenosis, anemia, atrial fibrillation, cardiac contractile dysfunction, congestive heart failure, arrhythmia prediction and myocardial infarction [12], [ 16]-[22] . In these studies, 12-lead ECG data were generally used. By applying Al to ECG, it has been shown that subtle changes in the QRS can be correlated with cardiac function such as myocardial fibrosis, congestive heart failure, effectiveness of diuresis therapy, etc., enabling faster and less costly assessment [23], Studies have been conducted to predict death and hospital readmission due to heart failure using Al and ML algorithms [24], In some of these studies, echocardiographic data [25], or multiple parameters such as clinical phenotyping, laboratory, ECG and echocardiography were used simultaneously [26], Congestive heart failure (CHF) encompasses HFrEF and HFpEF. A binary classification was made as presence or absence of CHF [27],

In a study by Hasumi et al. (2020), the patient class in the dataset included HFrEF, HFpEF and Coronary Artery Disease (CAD). However, the classification was based on the presence or absence of HF and the patient class was classified according to the New York Heart Association (NYHA) functional classification [28], The focus of the studies by Garcia-Escobar et al. (2022), Hasumi et al. (2020), Herman et al. (2022), Jin et al. (2022), Masetic and Subasi (2016), Plati et al. (2022) is not the diagnosis of HFrEF and HFpEF. A 10-year HF risk estimate was made for HF-REF and HF-PEF with an Al-based system developed using 12-lead ECG data [29], Here, LVEF was assumed below 50% for HFrEF and above 50% for HFpEF. That is, HF with mid-range ejection fraction (HFmrEF) (41% < LVEF < 49%) was also included in HFrEF. There is no evidence that this system can help differentiate between HFrEF and HFpEF.

PPG is a widely used clinical method to measure physiological parameters such as cardiac output and oxygen saturation. It is a good option for cardiovascular health studies due to its small size and cost effectiveness. Moreover, studies prove that PPG can be obtained independent of calibration [30], PPG has been used in cardiovascular and heart disease-related studies. Most of these have used ML algorithms. Rubins et al. evaluated groups of healthy volunteers and volunteers with cardiovascular disease (CVD) with a novel algorithm using simultaneous finger and ear PPG measurements [31], Non-invasive blood pressure estimation with PPG using ML and Al methods [32], Performance evaluation of deep learning models using photoplethysmography signals to diagnose different types of arrhythmias [33], PPG was used to screen for sleep-disordered breathing in heart failure (HF) patients [34], As an alternative to invasive methods, studies have been performed with PPG. PPG has been used to interpret the effect of L-arginine and citrulline amino acids on endothelial function in patients with right HF and stable diastolic patients [35], PPG has been found to indicate left ventricular filling pressure during the Valsalva maneuver [36], PPG used in the Valsalva maneuver has been shown to estimate pulmonary capillary wedge pressure in HF patients [37], The usefulness of PPG measured during the Valsalva maneuver has been investigated to determine whether hospitalized HF patients were at risk [38], It has been shown that mechanical change in HF patients can be accurately identified by PPG analysis [39], Since PPG is easy to measure, there are many studies on smartphones, wearable technologies and electronic devices. A novel algorithm using PPG has been developed for atrial fibrillation detection using smartphones [40], Since PPG is easily measured from smartphones and wearable devices, diabetes detection has been studied using PPG [41], A device with high specificity and good sensitivity was developed using PPG to detect left ventricular end-diastolic pressure greater than 20 mm Hg [42], The Corsano 287 wristband with PPG technology has been shown to be able to determine heart rate and RR intervals with high accuracy [43], A prototype PPG device was developed to distinguish between healthy individuals and those with CHF. Simultaneously acquired ECG and PPG signals were used to distinguish between healthy individuals and CHF patients [44], CHF includes cases of HFrEF and HFpEF. In the literature, studies on CHF are usually related to whether heart failure is present or not.

Some of the studies using ECG are only related to HFrEF or only to HFpEF. There is no study using ECG or PPG that evaluates these two HF subtypes simultaneously with only a single signal.

Since HF patients present with symptoms such as chest pain, shortness of breath and fatigue, especially patients with HFpEF (LVEF> 50, like healthy individuals) may be referred to the Pulmonology department. If the pulmonologist is attentive, he/she will perform B-type natriuretic peptide (BNP) and N-terminal proBNP (NT-proBNP) blood tests, which are invasive methods used for the diagnosis of HF, and these patients are diagnosed with HFpEF. However, if they are only evaluated for chest diseases and the results are normal, no action is taken and the diagnosis of HFpEF is missed. BNP and NT-proBNP blood tests are invasive and laborious methods. In addition, echocardiography is expensive, requires specialists and there may be limited access to the device. Therefore, there is a need for an alternative economical and practical diagnostic method. There are studies in the literature using heart rate variability (HRV) for HF [45]-[47] . However, in these studies, HRVs were derived from ECG.

Although HRV is mostly derived from ECG, there are also studies using PPG-derived HRVs [48]-[50], These studies are also not relevant to the diagnosis of HF. In the literature, studies using ECG or PPG signals usually used these signals as well as demographic information such as age, height, weight, echocardiographic data, medical history, and medications the patient was taking.

In addition, there is no study in the literature in which PPG was used in studies on HFrEF and HFpEF. The use of PPG signal, which can be easily obtained with a low-cost device and is advantageous for patient comfort, will also contribute to the literature.

In the current system, as mentioned above, there are devices, systems, methods and various articles related to HF, as well as patent and/or utility model applications. In the article titled "A Prototype Photoplethysmography Electronic Device that Distinguishes Congestive Heart Failure from Healthy Individuals by Applying Natural Time Analysis", a prototype electronic device is discussed. In the study, a study was conducted to distinguish between individuals with congestive heart failure (CHF) and healthy individuals. CHF is a more general definition of HF that includes cases of HFrEF and HFpEF. Unlike heart rate, which can be high or low in a healthy person and can vary on a daily basis, low LVEF is always associated with disease [51], Therefore, the diagnosis of HFrEF can be made when LVEF < 40% on echocardiography, which makes it easier to diagnose. However, the diagnosis of HFpEF is more complex and difficult even for experts. Since LVEF > 50% in HFpEF appears normal (LVEF > 50% in healthy individuals), cases of HFpEF can be confused with chest diseases due to some similar symptoms. Therefore, rather than distinguishing between CHF and healthy individuals, there is a need for a system that identifies which type of HF a person has, if they have CHF, and does so in a single step. Hence, there is a need for such a system that prevents or minimizes HF cases from being missed, thereby contributing to the reduction of deaths due to HF.

In the application numbered "TR 2022/022035", which is in the known state of the art, 1- channel electrocardiography (ECG) signals measured from three electrodes and bioimpedance (BioZ) signals measured simultaneously with the same electrodes were used together to detect HF with a deep learning algorithm. In this study, the subgroups of HF, namely HF with reduced ejection fraction (HFrEF) and HF with preserved ejection fraction (HFpEF), were not studied. Congestive HF (CHF) includes the subgroups of HFrEF and HFpEF. In HF studies in the literature, binary classification studies have generally been conducted as either having or not having CHF. The application numbered "TR 2022/022035" is also a study to determine HF, but the type of HF is not specified, so it is a study to determine whether there is CHF or not. For ECG measurement, three sensor probes are used: a positive electrode, a negative electrode and a reference electrode, located respectively at the lower left end of the heart region on the chest, the upper right end of the heart region on the chest and the lower region of the negative electrode. A deep neural network algorithm, the convolutional neural network (CNN) model, is applied in this application. Deep neural networks are popularly used methods today. They are advantageous in some areas such as image processing. However, they are known for their high hardware requirements and long algorithm runtimes.

The application numbered "CN116368578A" in the known state of the art relates to predicting the occurrence of a disease, in particular to a method and a device for predicting the probability of the future occurrence of a disease using an artificial intelligence (Al) algorithm. Its purpose is to provide a method for evaluating a factor affecting the probability of occurrence of a disease. This work relates to a method for predicting the probability of the future occurrence of a disease using an Al algorithm and a device capable of making this prediction. The main objective of the study is to present a method and system that can effectively predict the risk of developing a disease for an individual in the future, estimate the probability of disease occurrence within a certain time period on a yearly basis, and evaluate the factors affecting the probability of disease occurrence. This method also collects input data based on an individual's health checkup data, predicts the probability of disease occurrence for each year using a trained artificial intelligence model, evaluates these prediction results, and identifies the factors that most influence the probability of disease occurrence. It can also more accurately estimate disease risk by taking into account the time frame between different time periods. This method involves collecting information about the target person's health data, using a trained Al model and evaluating the results. In this way, it can more accurately predict a person's disease risk at a given point in time and assess the impact of disease-related factors. This study is not related to HF. A system model has been created that can predict risk for any disease in general. The deep neural network method of long short-term memory is used to calculate disease prediction information. Therefore, there is a need for a new technology that focuses specifically on the diagnosis of HF subtypes, uses a single signal, and supports all these processes with machine learning algorithms. The application numbered "CN114173647A" in the known state of the art relates to a method and system for assessing disease using dynamic analysis of cardiac and photoplethysmography signals. It relates to non-invasive methods for predicting and/or detecting cardiovascular diseases using cardiometry and photoplethysmography measurements. These measurements may be used alone or in combination with measurements of physiological events and various systems. They may also include information on factors such as the presence, severity, location, evolution or status of pulmonary and cardiopulmonary diseases.

The methods and systems used in this study facilitate one or more dynamic analyses that can characterize and describe the synchrony between received cardiac signals and PPG signals. These analyses can be used to predict and/or assess the presence, severity and/or location of abnormal cardiovascular conditions or diseases, including coronary artery disease, abnormal left ventricular end-diastolic pressure disease, pulmonary hypertension and its subtypes, heart failure (HF). So this study is not specifically about the diagnosis of HFrEF and HFpEF. It assesses general cardiac status. It is about a method and system for the assessment of disease using dynamic analysis of cardiac and PPG signals. That is, not a single signal, but multiple signals are used. For this reason, there is a need for a new study and technology that uses a single signal, especially for the detection of HF subtypes.

LVEF, the parameter used to diagnose HFrEF and HFpEF, is usually measured by echocardiography and is used as a general measure of cardiac function. Echocardiography can be called the "gold standard" method for LVEF measurement. It is the method commonly used in hospitals today. Single Photon Emission Computed Tomography (SPECT) and radionuclide ventriculography (MUGA) are also used. Some biomarkers such as B-type natriuretic peptide (BNP) and N-terminal proBNP (NT-proBNP) blood tests are also used to detect HF.

The devices currently used to diagnose HF are expensive, and their maintenance are also expensive, they are not easily portable, some methods are invasive, and all require a specialist physician. Access to these devices may also be limited. Since the devices are expensive and the number of them is low in the hospitals where they are located, patients can be seen by appointment system, which can indirectly cause a delay in diagnosing the patient. In addition, even if a device is available, there may be situations where there is no specialist cardiologist. For blood tests, for example, in rural areas, there may not be a laboratory equipped to perform the relevant tests. When sending blood samples to a central laboratory, the samples may deteriorate on the way, resulting in inaccurate or no results. This can again delay the diagnosis and it may be too late for the patient.

As a result, due to the above-mentioned drawbacks and the inadequacy of existing solutions, there is a need for a decision support system that can provide the preliminary information within seconds whether a person admitted to the hospital with symptoms and suspicion of HF is HFrEF, HFpEF or healthy.

Brief Description and Objectives of the Invention

ECG and PPG data were acquired simultaneously for 10 s from volunteers aged 25 years and older. Both signals were filtered with digital filters. After the filtering step, HRVs were derived from PPGs. Then 37 features were extracted from each of ECG, PPG and HRVs. To investigate whether the features extracted from ECG alone or from PPG and PPG-derived HRVs could be used for the diagnosis algorithm without the need for demographic information or echocardiography and other detailed examinations, the relationship between features and classes was analyzed by statistical methods. The most discriminative features were identified statistically, the number of features was reduced and classification studies were performed with ML algorithms. Classifications were performed with very high accuracy.

The aim of the invention is to develop a ML-based medical decision support system using only a single signal (ECG or PPG), which is beneficial in terms of both time and patient comfort, which provides the physician or the healthcare personnel performing the test with the preliminary information within seconds whether the person admitted to the hospital with HF symptoms and suspicion is HFrEF, HFpEF or healthy, and prevents unnecessary expensive and laborious tests to the patient.

A further aim of the invention is to enable diagnosis with a single signal for cases of HFpEF, which are likely to be missed compared to the current system.

Another aim of the invention is that the device used to diagnose HF is both easy to use and easily accessible in terms of price. Because HF is a serious disease that can result in death without intervention. For example, acute heart failure (AHF) is a HF condition that requires urgent treatment with sudden onset or sudden changes in symptoms and signs. So diagnosis and treatment should be done as quickly as possible. Thanks to the invention, any health personnel in health units such as Family Health Centers where there is no specialist physician will be able to use this system, and if any HF subtype is detected as a result of the results given by the device, cases will be prevented from being missed by referring them to a specialist physician immediately. Since it is a low-cost device, it can be found in every health institution, regardless of whether it is a rural or central institution. It is a small, portable device that patients can even use in their own homes. Anyone with HF symptoms will be able to use it easily and get results in seconds, minimizing the possibility of HF cases being missed. Moreover, early diagnosis and timely treatment can improve the quality of life of HF patients and reduce HF-related death.

Conventional ML methods were used in the invention . This is because the algorithms work fast, provide results within seconds and do not require a lot of hardware. In addition, fast methods such as k-nearest neighbors (k-NN) and support vector machines are advantageous because they can be easily adapted to real-time systems.

These figures are as follows

Figure 1: View of the workflow diagram for HF detection with ECG.

Figure 2: ROC curves obtained from each algorithm for HF detection with ECG.

Figure 3: View of the workflow diagram for the HF detection with PPG and HRV.

Figure 4: View of the device.

Element Numbers in Figures

In order to better explain the machine learning based device for the detection and classification of heart failure with reduced ejection fraction and heart failure with preserved ejection fraction developed by the present invention, the parts and elements in the figures are numbered and each number is given below:

1. Device

2. Electronic card section

3. On and off switch

4. LCD screen

5. Cable connection port for PPG sensor

6. Cable connection port for ECG probes

7. Speaker 8. ECG probes and connection cable

9. PPG sensor and connection cable

Detailed Description of the Invention

The invention relates to a ML based device for detecting and classifying heart failure with reduced ejection fraction (HFrEF) and heart failure with preserved ejection fraction (HFpEF). In particular, the invention relates to a device that can determine whether a person is a case of HFrEF, a case of HFpEF, or a healthy person by evaluating them in a single step.

The system developed by the invention performs a triple classification with three classes: HFrEF, healthy and HFpEF. In other words, the system not only provides information about whether the individual with HF symptoms has CHF, but also whether the case is HFrEF, HFpEF or healthy, that is, if there is HF, which type of HF it is.

In the system developed by the invention, firstly data were collected from the patients. Biopac MP36 device was used to collect and record the data. The sampling frequency is 200 Hz. From people aged 25 years and older, ECG and PPG data were collected simultaneously in a resting state (supine). Three-electrode (Bipolar DII) ECG data (right wrist, right and left ankle) were obtained from 61 individuals for the triple classification study with ECG. PPG data were taken from the index finger of the right hand. Data from the same person at different time intervals were also included in the data set. 10 s ECG and PPG data were used. There are a total of 180 data in the mentioned dataset, 60 of which are HFrEF, 60 of which are healthy and 60 of which are HFpEF.

Digital filtering was performed with filters designed in a simulation program. At this stage, a Chebyshev Type II bandpass filter was applied to each signal in the range of 0.25 - 100 Hz. The stopband attenuation for this filter was set to 60 dB. In the second stage, a notch filter was applied to the signal between 49 and 51 Hz to remove the mains noise at 50 Hz. For this filter, the stopband attenuation was set to 60 dB. Finally, a Moving Average filter was applied. After filtering, the heart rate variability (HRV) was derived from the PPG. Features that have been previously extracted from ECG and PPG signals in the literature and used for various disease detection were extracted [48], [52], In the time domain, 21 features were extracted from the data and a simulation program was used to calculate these features. Table 1 shows these features. No Feature Equation _

1 kurtosis

2 Skewness

3 * Span Between Quarters IQR = iqr(x)

4 Geometric Mean G = (x₁ + -- + x_n)¹/ⁿ

5 Harmonic Mean H = n/(^ 1₊... _{+ n})

6 Mobility - Hjort Parameters M = Sl/S⁷-

7 Complexity - Hjort Parameters C = ((S²/S²)² - (S /S²)²)¹/²

8 Activity - Deer parameterization A = S

{ ²

9 * Maximum ^Xmax = max(%i) Xn+i ’• x tek

~ 2~

10 Median 1 / \

— I Xn + Xn I : X gift

2 \ 2 2⁺¹/

11 * Mean Absolute Deviation MAD = mad(x)

12 *Minimum T = min(%i)

13 * Central Moments CM = moment(x, 10)

19 * Singular Value Decomposition SVD = svd(x)

20 *%25 Trimmed Average TM25 = trimmean (x, 25)

21 *%50 Trimmed Average TM50 = trimmean (x, 50)

* This feature was calculated with a Matlab function. x Mean of the distribution n: e.g. number of observations

S^. Standard deviation of the first derivative of the signal

S₂: Standard deviation of the second derivative of the signal

X_rms: Root Mean Square

Table 1: Time domain features extracted from the data in the study

16 Yule-Walker and Burg method parameters were then added to these 21 features as features.

As a result, 37 features were extracted from each of the ECG, PPG and HRV data.

The following simulation program function was used for the Yule-Walker method. [a,e,rc] = aryule(x,p) This function computes the normalized autoregressive (AR) parameters (a) corresponding to a p-order model for input array x. It also returns the estimated variance (e) and the reflection coefficients (rc) of the white noise input.

The following simulation program function was used for the Burg method. fa,e,rc] = arburg(x,p)

This function calculates the normalized autoregressive (AR) parameters (a) corresponding to a p-order model for the input array x. It also returns the estimated variance (e) and the reflection coefficients (rc) of the white noise input.

In the Burg method, the AR parameters corresponding to a pth-order model are calculated as in expression 1 [53],

Expression 2 is used for the reflection coefficients.

^— 2 Sn-p+i p_₁(n)ej, p_₁(n — 1)

For the error, expression 5 is used.

^p — ^f,p^^b,p (5)

For the Yule-Walker method, the parameters are calculated with the following expressions [53],

For the features used in the study subject to the invention, AR parameters corresponding to the 4th order model are used for both Yule-Walker and Burg methods.

The 16 features obtained after 21 features are shown in Table 2 with their general ranking numbers.

No Feature Icon

22 1^st AR parameter of the Yule-Walker method 4th order model a₄(l)

23 2^nd AR parameter of the Yule-Walker method 4th order model u₄(2)

24 3^rd AR parameter of the Yule-Walker method 4th order model u₄(3)

25 4^th AR parameter of the Yule-Walker method 4th order model a₄ (4)

~ , Estimated variance of the white noise input of the Yule-Walker method 4th order _c

²⁶ model ^£»

27 1^st reflection coefficient of Yule-Walker method the 4th order model rc(l)

28 2^nd reflection coefficient of Yule-Walker method the 4th order model rc(2)

29 3^rd reflection coefficient of Yule-Walker method the 4th order model rc(3)

30 1^st AR parameter of the Burg method 4th order model u_{4 1}

31 2^nd AR parameter of the Burg method 4th order model a₄₂

32 3^rd AR parameter of the Burg method 4th order model u_{4 3}

33 4^th AR parameter of the Burg method 4th order model u₄₄

34 Estimated variance of the white noise input of the Burg method 4th order model e_p

35 1^st reflection coefficient of the Burg method 4th order model rc₄

36 2^nd reflection coefficient of the Burg method 4th order model rc₂

37 3^rd reflection coefficient of the Burg method 4th order model rc₃ Table 2. Yule-Walker and Burg method parameters used as features

Statistical analysis was performed for the ECG signal to determine whether it was normally distributed. The Kolmogorov- Smirnov Normality Test is one of the most common tests used to determine whether data are normally distributed [54], Evaluation is made according to the p and h values, which are the test results. The "p" value is the statistical probability value, "h" is the value indicating the hypothesis. If h=0, the hypothesis HO is valid, and if h=l, the hypothesis Hl is valid. To assume a normal distribution, p>0.05 is required. HO indicates that there is a normal distribution and Hl indicates that there is no normal distribution. For example, the Kolmogorov-Smirnov test result for the first ECG data in the data set is p = 7.072e'²²⁸ and h=l . In other words, it is seen that the condition p<0.05 is met and hypothesis HO is rejected and hypothesis Hl is valid. Since ECG does not show a normal distribution, it can be statistically analyzed using nonparametric methods. When there are three or more classes and the data are not normally distributed, the Kruskal-Wallis test is preferred. This test is the nonparametric version of one-way ANOVA [55], This test is used to investigate whether there is a significant difference between HFrEF, healthy, and HFpEF classes in the dataset in terms of extracted features. The asymptotic significance (p-values) obtained as a result of the Kruskal-Wallis test for each feature are given in Table 3.

Feature p values Feature No p values

1 0.8337 20 0.0000

2 0.0000 21 0.0000

3 0.1095 22 0.6501

4 0.0847 23 0.7135

5 0.8753 24 0.9343

6 0.0007 25 0.9771

7 0.0000 26 0.1214

8 0.1583 27 0.0009

9 0.0014 28 0.0000

10 0.0000 29 0.1137

11 0.1784 30 0.0000

12 0.0107 31 0.0000

13 0.0464 32 0.0002

14 0.7619 33 0.1116

15 0.0261 34 0.0000

16 0.1583 35 0.0007

17 0.1583 36 0.0000

18 0.1457 37 0.0000

19 0.1574

Table 3. p values of ECG features as a result of Kruskal -Wallis test

According to the table, p < 0.05 for features 2, 6, 7, 9, 10, 12, 13, 15, 20, 21, 27, 28, 30, 31, 32, 34, 35, 36, 37. That is, these features are discriminative for at least two of the three classes compared. The Mann Whitney U test is a nonparametric test used to determine whether two sampled groups are from the same population [56], Paired Mann-Whitney-U tests were performed as post-hoc tests to determine whether the features that appeared to be distinctive as a result of the Kruskal-Wallis test were distinctive only for certain binary classes or for all three classes. Then, a different number of features from the most relevant ones were taken as input and experiments were conducted. The optimum value was determined in the relationship of fewer features and higher accuracy. As a result, high accuracy classification was achieved as a result of the algorithm using 4 features: skewness, 3rd AR parameter of the Burg method 4th order model, 2nd reflection coefficient of the Burg method 4th order model and 3rd reflection coefficient of the Burg method 4th order model. These features rank high in the Mann Whitney U test rankings.

The triple classification study with ECG was also performed using PPG and HRV features together. As explained above, PPG and HRV are not normally distributed. Therefore, nonparametric statistical methods are used. Since the data were not normally distributed and there were three classes, Kruskal-Wallis test was applied. Paired Mann-Whitney U tests were applied as post-hoc tests and it was determined which characteristics were discriminative for which binary classes. After that, experiments were conducted with different numbers and combinations of features, from the most discriminative ones. The optimum value was determined in the relationship of fewer features and higher accuracy. Significant results were obtained with a total of eight features, 5 of which were PPG features and 3 of which were HRV features. The 4th, 6th, 15th, 34th and 35th features of PPG, namely the geometric mean, Hjort mobility parameter, average curve length, estimated variance of the white noise input of the Burg method 4th order model and the 1st reflection coefficient of the Burg method 4th order model, and the 3rd, 10th and 17th features of HRV, namely the interquartile range, median and standard error, were used.

In the system developed by the invention, a ML based system was developed using features derived from a 3 -electrode ECG signal. A total of 37 features were derived from the ECG. However, since it is difficult to derive so many features in real-time systems, the number of features was reduced to improve the system. To improve the performance of the classifier, 4 features were selected with Kruskal-Wallis Test and Mann-Whitney-U Test methods and used. The study with ECG was carried out with the sequence of operations in the flow diagram in Figure-1 and the results were obtained. Statistical analysis of 37 features extracted from ECG was performed to determine whether they were discriminative for HFrEF, Healthy and HFpEF classes. 19 of the 37 features were found to be discriminative for the three classes. Then, paired Mann Whitney U tests were used to determine which features were discriminative for which pairwise groups and the features were ranked according to the Eta correlation coefficients between ECG features and classes. Mann-Whitney U test results are given in Tables 4, 5 and 6. Feature No R (ETA) R² p values

2 0.2932 0.0860 0.0004

36 0.2136 0.0456 0.0029

12 0.1960 0.0384 0.0389

7 0.1835 0.0337 0.0333

20 0.1781 0.0317 0.0035

21 0.1758 0.0309 0.0131

9 0.1734 0.0301 0.9686

6 0.1580 0.0250 0.1482

10 0.1527 0.0233 0.0269

35 0.1482 0.0220 0.1497

27 0.1468 0.0215 0.1497

30 0.1320 0.0174 0.0765

28 0.1294 0.0168 0.0967

31 0.1014 0.0103 0.2345

13 0.0918 0.0084 0.5906

15 0.0907 0.0082 0.4235

32 0.0642 0.0041 0.4174

37 0.0524 0.0027 0.7548

34 0.0518 0.0027 0.4265

Table 4. Paired Mann-Wnitney U test results for HFrEF and Healthy classes

Feature No R (ETA) R² p values

37 0.5576 0.3109 0.0000

2 0.3790 0.1437 0.0000

31 0.3381 0.1143 0.0002

30 0.3347 0.1121 0.0000

35 0.3280 0.1076 0.0002

27 0.3233 0.1045 0.0002

32 0.3144 0.0989 0.0053

6 0.2974 0.0884 0.0002

28 0.2668 0.0712 0.0056

34 0.1362 0.0186 0.0001

36 0.2350 0.0552 0.0235

7 0.1986 0.0395 0.0235

12 0.1218 0.0148 0.1976

9 0.0860 0.0074 0.0199

15 0.0839 0.0070 0.0420

10 0.0666 0.0044 0.0765

13 0.0405 0.0016 0.0837

20 0.0278 0.0008 0.0208

21 0.0236 0.0006 0.0464

Table 5. Paired Mann-Wnitney U test results for HFrEF and HFpEF classes Feature No R (ETA) R² p values

2 0.6123 0.3750 0.0000

37 0.5713 0.3264 0.0000

30 0.5244 0.2750 0.0000

36 0.5075 0.2575 0.0000

31 0.4578 0.2096 0.0000

7 0.4171 0.1740 0.0000

28 0.3995 0.1596 0.0000

34 0.3870 0.1497 0.0000

32 0.3581 0.1283 0.0000

10 0.2605 0.0679 0.0000

21 0.1885 0.0355 0.0000

12 0.1799 0.0324 0.0055

20 0.0946 0.0089 0.0000

13 0.0917 0.0084 0.0108

9 0.0552 0.0030 0.0001

15 0.0242 0.0006 0.0131

6 0.1555 0.0242 0.0181

35 0.1718 0.0295 0.0202

27 0.1690 0.0286 0.0242

Table 6. Paired Mann-Wnitney U test results for HFpEF and Healthy classes

Since each of the three classes has its own percentage error, the value of 0.05 is divided by 3 (0.0167) (this is called Bonferroni correction) [57], When the Mann-Whitney-U test result for the relevant feature is p < 0.0167, the trait is said to be discriminative for that pair group. So, for example, in Table 6, for features 2, 7, 9, 10, 10, 12, 13, 15, 20, 21, 28, 30, 31, 32, 34, 36 and 37, p < 0.0167, so these features are discriminative for the binary classification of HFpEF and Healthy.

Then a different number of features from the most relevant ones were taken as input and experiments were performed. The optimum value was determined in the less features / high accuracy relationship. As a result, high accuracy classification was achieved as a result of algorithms using 4 features. Classification performance parameters are given in Table 7.

ML Algorithm Accuracy s_ensftj_vfty Specificity F-measure Kappa AUC k-NN 100 1 1 1 1 1

Support Vector Machines 97.22 0.9722 0.9861 0.9722 0.9583 0.9942

Decision Trees 91.67 0.9167 0.9583 0.9167 0.8750 0.9780

Ensemble Bagged Trees 97.22 0.9722 0.9861 0.9722 0.9583 0.9896

Table 7. Performance parameters of the ECG study.

In the k-NN algorithm, when 37 features were used, the AUC values for three classes were calculated as 0.938. When the number of features was reduced to 19, an increase in the AUC values of the three classes was observed. While the results were quite good for the HFpEF class, it was seen that the positive class was better distinguished for the Healthy class, and the control class was more clearly separated for the HFrEF class. When the number of features was reduced to 4, ideal results were obtained for all classes.

In the SVM algorithm, it was observed that when 37 features were used, the control class was better separated for the HFpEF and Healthy classes, while the positive class was more clearly distinguished for the HFrEF class. When the number of features was reduced to 19, the HFpEF class gave ideal results, while a slight decrease in AUC values occurred in the other classes. When the number of features was 4, ideal results were obtained for the HFrEF class, while the results for the HFpEF and Healthy classes were also very close to ideal.

In the Decision Tree algorithm, when 37 features were used, the positive class was better distinguished for the HFpEF class, while the control classes were more successfully separated for the Healthy and HFrEF classes. When the number of features was reduced to 19, a significant decrease in the AUC value was observed for the HFrEF class. However, when only 4 features were used, more successful results were obtained for all three classes.

In the classification performed using Ensemble Classifier (Ensemble Bagged Trees), when using 37 features, the HFpEF class gave ideal results, while the control class for the HFrEF class was more successfully separated. When the number of features was reduced to 19, the positive class for the HFrEF class was better separated. When 4 features were used, the HFpEF class again gave ideal results, while the control class for the HFrEF class and the positive class for the Healthy class were more clearly distinguished.

The system developed by the invention can be embedded in future or existing measurement devices. The k-NN classifier was prefered as it is widely used in industrial applications [58], Other three algorithms were also used to support the study and to see the applicability of this triple classification using 3-electrode ECG. These algorithms also have the potential to be run in embedded systems in real time [59], According to the results obtained, this classification could be performed with high accuracy and the best results were obtained with the k-NN method. Compared to other studies in the field, the classification in this study has three classes, namely HFrEF, Healthy and HFpEF. With the invention, it is possible to determine not only whether the person has HF or not, but also which type of HF the person has, if any.

In the classification study using PPG and PPG-derived HRV features, results were obtained by following the steps according to the workflow in Figure 3. After feature extraction and selection, a triad classification was performed using various classifiers with HFrEF, Healthy and HFpEF classes. The aim of this study, as with the ECG study, is to provide a medical decision support system that will accelerate and facilitate the diagnosis of the physician in cases where echocardiography is difficult to access or where treatment should be started immediately before echocardiography data, or to help existing physicians diagnose where there is no specialist physician and to refer the patient to a specialist physician in a timely manner. Results were obtained using a total of 8 features, 5 of which are PPG features and 3 of which are HRV features. PPG's geometric mean, Hjort mobility parameter, mean curve length, estimated variance of the white noise input of the Burg method 4th order model and the 1st reflection coefficient of the Burg method 4th order model, and the interquartile range, median and standard error of the HRV were used as features. The performance parameters obtained as a result of classification are given in Table 8.

ML Algorithm Accuracy s_ensftj_vfty Specificity F- measure Kappa

Support Vector Machines 82.78 0.8278 0.9139 0.8278 0.7417 k-NN 82.22 0.8222 0.9111 0.8222 0.7333

Ensemble Bagged Trees 87.78 0.8778 0.9389 0.8778 0.8167

Table 8. Classification performance parameters for the study with PPG and HRV.

Based on the results of the working method of the system developed by the invention, it has been determined that only 3 -electrode ECG or only PPG features can be used in the diagnosis of HFrEF and HFpEF and provide significant results. Easy measurement and processing of signals increases the practicality of real-time systems. Used ML methods play an important role in bridging the gaps in heart failure and have significant advantages over traditional humanbased models [60], ML provides computers with the capacity to evaluate data beyond programmed algorithms, identify patterns in data, match learned patterns with unseen data, and improve the performance of computational tasks beyond human capabilities. k-NN, one of the machine learning algorithms used in the invention, is a very simple and powerful approach to conceptually approximating a real -valued or discrete-valued objective function. Many researchers have recently demonstrated that k-NN has a high prediction accuracy for a variety of real -world systems using many different types of datasets [61], k-NN classifier was prefered because it is widely used in industrial applications [58], Likewise, Support Vector Machines (SVM) and Decision Trees methods were preferred because they do not require high hardware, are widely used, have low cost and produce results in a short time.

The device (1) developed by the invention is shown in Figure-4. This device (1) includes electronic card section (2), on and off switch (3), LCD screen (4), connection cable port for PPG sensor (5), connection cable port for ECG probes (6), speaker (7), ECG probes and connection cable (8) and PPG sensor connection cable (9). In the electronic board section (2), controls and operations are performed. The electronic board part (2) is the part where the ECG data or PPG data is recorded with the processor, digital filtering is performed, the 4 features mentioned above if the test is performed with ECG and the 8 features mentioned above if the test is performed with PPG are extracted from the filtered data, and these features are given to the trained model and the test result is obtained.

The device (1) can be turned on and off using the on and off button (3). The LCD screen (4) displays information on whether the individual is HFrEF, HFpEF or healthy. The data received by the PPG sensor is transferred to the device (1) via the PPG sensor's connection cable (5). From the input port (6) of the ECG probes and connection cable, the data received from the ECG probes is transferred to the device (1). In addition, the speaker (7) in the device (1) is the sound output for the device (1) to give an audible warning. It informs the user audibly at the end of the test.

With the invention, diagnosis process is accelerated. It is also very advantageous in terms of patient comfort compared to other methods. Since it is a device (1) that can be easily used by healthcare personnel who provide home healthcare services or go to cases by ambulance, it will ensure timely diagnosis of HF, timely and correct intervention and reduce deaths due to HF. With this device (1), which can be easily used by any healthcare personnel in family health centers in neighborhoods or rurals, provides timely intervention by being diagnosed the patient and immediately being referred the patient to a more equipped health institution with a specialist physician. Although LVEF appears to be at normal percentages in HFpEF, due to abnormalities in cardiac function, if HFpEF is not diagnosed in time and treatment is not started, or if it is recognized late and the diagnosis is delayed and treatment is started late, there is a possibility that it may turn into HFrEF. Since it will be a low-cost and affordable device (1), patients will be able to have it in their own homes, and continuous monitoring of risky patients will be provided in the comfort of home.

References

[1] K. Dickstein vd., “ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure 2008J”, Eur. J. Heart Fail., c. 10, sayi 10, ss. 933-989, Eki. 2008, doi: 10.1016/J.EJHEART.2008.08.005. [2] P. Ponikowski vd., “2016 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure The Task Force for the diagnosis and treatment of acute and chronic heart failure of the European Society of Cardiology (ESC)Developed with the special contribution of t”, Eur. Heart J., c. 37, sayi 27, ss. 2129-2200, Tern. 2016, doi: 10.1093/EURHEART J/EHW 128.

[3] P. A. Heidenreich vd., “2022 ACC/AHA/HFSA Guideline for the Management of Heart Failure”, J. Card. Fail., c. 28, sayi 5, ss. el-el67, May. 2022, doi: 10.1016/J.CARDFAIL.2022.02.010.

[4] W. J. Paulus vd., “How to diagnose diastolic heart failure: a consensus statement on the diagnosis of heart failure with normal left ventricular ejection fraction by the Heart Failure and Echocardiography Associations of the European Society of Cardiology”, Eur. Heart J., c. 28, sayi 20, ss. 2539- 2550, Eki. 2007, doi: 10.1093/EURHEART J/EHM037.

[5] B. A. Borlaug ve W. J. Paulus, “Heart failure with preserved ejection fraction: pathophysiology, diagnosis, and treatment”, Eur. Heart J., c. 32, sayi 6, ss. 670-679, Mar. 2011, doi : 10.1093/EURHEARTJ/EHQ426.

[6] S. F. Nagueh, K. J. Middleton, H. A. Kopelen, W. A. Zoghbi, ve M. A. Quinones, “Doppler Tissue Imaging: A Noninvasive Technique for Evaluation of Left Ventricular Relaxation and Estimation of Filling Pressures”, J. Am. Coll. Cardiol., c. 30, sayi 6, ss. 1527-1533, Kas. 1997, doi : 10.1016/S0735-1097(97)00344-6.

[7] S. F. Nagueh vd., “Recommendations for the Evaluation of Left Ventricular Diastolic Function by Echocardiography”, J. Am. Soc. Echocardiogr., c. 22, sayi 2, ss. 107-133, §ub. 2009, doi: 10.1016/J.ECH0.2008.11.023.

[8] D. J. Choi, J. J. Park, T. Ali, ve S. Lee, “Artificial intelligence for the diagnosis of heart failure”, npj Digit. Med. 2020 31, c. 3, sayi 1, ss. 1-6, Nis. 2020, doi: 10.1038/s41746-020- 0261-3.

[9] J. E. Ho vd., “Predictors of new-onset heart failure differences in preserved versus reduced ejection fraction”, Circ. Hear. Fail., c. 6, sayi 2, ss. 279-286, Mar. 2013, doi: 10.1161/CIRCHE ARTFAILURE.112.972828.

[10] B. Kolukisa vd., “Diagnosis of Coronary Heart Disease via Classification Algorithms and a New Feature Selection Methodology”, c. 1, sayi 1, 2019. [11] P. Kota vd., “Heart disease classification comparison among patients and normal subjects using machine learning and artificial neural network techniques”, Int. J. Biosens. Bioelectron., c. Volume 7, sayi Issue 3, May. 2021, doi: 10.15406/IJBSBE.2021.07.00216.

[12] J. Kwon vd., “Artificial intelligence assessment for early detection of heart failure with preserved ejection fraction based on electrocardiographic features”, Eur. Hear. J. - Digit. Heal., c. 2, sayi 1, ss. 106-116, May. 2021, doi: 10.1093/EHJDH/ZTAA015.

[13] G. B. Lim, “Al identifies distinct subgroups of HFpEF”, Nat. Rev. Cardiol. 2021 186, c. 18, sayi 6, ss. 387-387, Mar. 2021, doi: 10.1038/s41569-021- 00545-0.

[14] M. Muthunayagam ve K. Ganesan, “Identification of systolic and diastolic heart failure progression with Krawtchouk moment feature-aided Harris hawks optimized support vector machine”, Signal, Image Video Process., c. 16, sayi 1, ss. 127-135, §ub. 2022, doi: 10.1007/S 11760-021 -01978-3/FIGURES/7.

[15] S. Selvaraj vd., “Application of Diagnostic Algorithms for Heart Failure With Preserved Ejection Fraction to the Community”, JACC Hear. Fail., c. 8, sayi 8, ss. 640-653, Agu. 2020, doi: 10.1016/J.JCHF.2020.03.013.

[16] F. Y. O. Abdalla vd., “Deep convolutional neural network application to classify the ECG arrhythmia”, Signal, Image Video Process., c. 14, sayi 7, ss. 1431-1439, Eki. 2020, doi: 10.1007/S 11760-020-01688- 2/TABLES/5.

[17] Z. I. Attia vd., “Screening for cardiac contractile dysfunction using an artificial intelligence- enabled electrocardiogram”, Nat. Med. 2019 251, c. 25, sayi 1, ss. 70-74, Oca. 2019, doi: 10.1038/s41591-018-0240-2.

[18] Z. I. Attia vd., “An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction”, Lancet, c. 394, sayi 10201, ss. 861-867, Eyl. 2019, doi: 10.1016/S0140- 6736(19)31721-0.

[19] J. M. Kwon vd., “Deep learning-based algorithm for detecting aortic stenosis using electrocardiography”, J. Am. Heart Assoc., c. 9, sayi 7, s. 14717, Nis. 2020, doi: 10.1161/JAHA.119.014717.

[20] Z. Lin, Y. Gao, Y. Chen, Q. Ge, G. Mahara, ve J. Zhang, “Automated detection of myocardial infarction using robust features extracted from 12-lead ECG”, Signal, Image Video Process., c. 14, sayi 5, ss. 857- 865, Tem. 2020, doi: 10.1007/S11760-019-01617- Y/TABLES/4.

[21] A. Sbrollini vd., “Serial electrocardiography to detect newly emerging or aggravating cardiac pathology: A deep-learning approach”, Biomed. Eng. Online, c. 18, sayi 15, ss. 1-17, §ub. 2019, doi: 10.1186/S12938-019-0630-9/FIGURES/6.

[22] Y. Zhang ve M. Xia, “Application of Deep Neural Network for Congestive Heart Failure Detection Using ECG Signals”, J. Phys. Conf. Ser., c. 1642, sayi 1, s. 012021, Eyl. 2020, doi: 10.1088/1742-6596/1642/1/012021.

[23] A. Garcia-Escobar, S. Vera-Vera, A. Jurado-Roman, | Santiago Jimenez-Valero, G. Galeote, ve R. Moreno, “Subtle QRS changes are associated with reduced ejection fraction, diastolic dysfunction, and heart failure development and therapy responsiveness: Applications for artificial intelligence to ECG”, Ann. Noninvasive Electrocar diol., c. 00, s. el2998, Tem. 2022, doi: 10.1111/ANEC.12998.

[24] D. K. Plati vd., “Machine Learning Techniques for Predicting and Managing Heart Failure”, Predict. Hear. Fail., ss. 189-226, Nis. 2022, doi: 10.1002/9781119813040.CH9.

[25] X. Jin vd., “Left atrial structure and function in heart failure with reduced (HFrEF) versus preserved ejection fraction (HFpEF): systematic review and meta-analysis”, Heart Fail. Rev., c. 1, ss. 1-23, Oca. 2022, doi: 10.1007/S 10741-021-10204- 8/FIGURES/2.

[26] R. Herman vd., “Utilizing longitudinal data in assessing all-cause mortality in patients hospitalized with heart failure”, ESC Hear. Fail., Haz. 2022, doi: 10.1002/EHF2.14011.

[27] Z. Masetic ve A. Subasi, “Congestive heart failure detection using random forest classifier”, Comput. Methods Programs Biomed., c. 130, ss. 54-64, Tem. 2016, doi: 10.1016/J.CMPB.2016.03.020.

[28] E. Hasumi vd., “Heart failure grading using single-lead electrocardiography”, medRxiv, s. 2020.10.08.20209700, Eki. 2020, doi: 10.1101/2020.10.08.20209700.

[29] I. Karabayir vd., “PO-703-04 ECG-AI CAN PREDICT RISK FOR HEART FAILURE WITH BOTH PRESERVED AND REDUCED EJECTION FRACTION”, Hear. Rhythm, c. 19, sayi 5, s. S447, May. 2022, doi: 10.1016/j.hrthm.2022.03.1058.

[30] H. Harini, L. S. Krithika, M. Shalini, S. Swaminathan, ve N. Mohan Madhu, “Design and Implementation of a Calibration - Free Pulse Oximeter”, IFMBE Proc., c. 43, ss. 100-103, 2014, doi: 10.1007/978-3-319-02913-9_26. [31] U. Rubins, A. Grabovskis, J. Grube, ve I. Kukulis, “Photoplethysmography Analysis of Artery Properties in Patients with Cardiovascular Diseases”, IFMBE Proc., c. 20 IFMBE, ss. 319-322, 2008, doi: 10.1007/978-3-540-69367-3_85.

[32] C. El-Hajj ve P. A. Kyriacou, “A review of machine learning techniques in photoplethysmography for the non-invasive cuff-less measurement of blood pressure”, Biomed. Signal Process. Control, c. 58, s. 101870, Nis. 2020, doi: 10.1016/J.BSPC.2020.101870.

[33] T. R. Kulkarni ve N. D. Dushyanth, “Performance evaluation of deep learning models in detection of different types of arrhythmia using photo plethysmography signals”, Int. J. Inf. Technol., c. 13, sayi 6, ss. 2209-2214, Ara. 2021, doi: 10.1007/S41870-021- 00795- 8/FIGURES/3.

[34] S. Sharma vd., “Photopl ethysmographic Signal to Screen Sleep-Disordered Breathing in Hospitalized Heart Failure Patients: Feasibility of a Prospective Clinical Pathway”, JACC Hear. Fail., c. 3, sayi 9, ss. 725-731, Eyl. 2015, doi: 10.1016/J.JCHF.2015.04.015.

[35] A. Orea-Tejeda vd., “The effect of L-arginine and citrulline on endothelial function in patients in heart failure with preserved ejection fraction”, Cardiol. J., c. 17, sayi 5, ss. 464-470,

2010, Eri§im: Mar. 18, 2022. [Qevrimigi]. Available at: http s : // pu bmed . n cbi . n! . ni h . sov

[36] H. A. Silber vd., “Finger photoplethysmography during the Valsalva maneuver reflects left ventricular filling pressure”, Am. J. Physiol. Heart Circ. Physiol., c. 302, sayi 10, May. 2012, doi: 10.1152/ AJPHEART.00609.2011.

[37] N. A. Gilotra vd. /‘Usefulness of Pulse Amplitude Changes During the Valsalva Maneuver Measured Using Finger Photoplethysmography to Identify Elevated Pulmonary Capillary Wedge Pressure in Patients With Heart Failure”, Am. J. Cardiol., c. 120, sayi 6, ss. 966-972, Eyl. 2017, doi: 10.1016/J.AMJCARD.2017.06.029.

[38] N. A. Gilotra vd., “Usefulness of Noninvasively Measured Pulse Amplitude Changes During the Valsalva Maneuver to Identify Hospitalized Heart Failure Patients at Risk of 30- Day Heart Failure Events (from the PRESSURE-HF Study)”, Am. J. Cardiol., c. 125, sayi 6, ss. 916-923, Mar. 2020, doi: 10.1016/J.AMJC ARD.2019.12.027.

[39] T. Besleaga vd., “Non-Invasive Detection of Mechanical Altemans Utilizing Photoplethysmography”, IEEE J. Biomed. Heal. Informatics, c. 23, sayi 6, ss. 2409-2416, Kas. 2019, doi: 10.1109/JBHI.2018.2882550. [40] T. Schack, Y. Safi Harb, M. Muma, ve A. M. Zoubir, “Computationally efficient algorithm for photoplethysmography-based atrial fibrillation detection using smartphones”, Proc. Annu. Int. Conf IEEE Eng. Med. Biol. Soc. EMBS, ss. 104-108, Eyl. 2017, doi: 10.1109/EMBC.2017.8036773.

[41] R. Avram vd., “PREDICTING DIABETES FROM PHOTOPLETHYSMOGRAPHY USING DEEP LEARNING”, J.Am. Coll. Cardiol., c. 73, sayi 9, s. 16, Mar. 2019, doi: 10.1016/S0735- 1097( 19)33778-7.

[42] P. Galiatsatos vd., “Usefulness of a Noninvasive Device to Identify Elevated Left Ventricular Filling Pressure Using Finger Photoplethysmography During a Valsalva Maneuver”, Am. J. Cardiol., c. 119, sayi 7, ss. 1053-1060, Nis. 2017, doi: 10.1016/J.AMJCARD.2016.11.063.

[43] S. Blok, M. A. Piek, 1.1. Tulevski, G. A. Somsen, ve M. M. Winter, “The accuracy of heartbeat detection using photoplethysmography technology in cardiac patients”, J. Electrocardiol., c. 67, ss. 148-157, Tern. 2021, doi: 10.1016/J.JELECTROCARD.2021.06.009.

[44] G. Baldoumas vd., “Remote sensing natural time analysis of heartbeat data by means of a portable photoplethysmography device”, https:/ doi. org/10.1080/2150704X.2020.1847351, c. 42, sayi 6, ss. 2292-2302, 2020, doi: 10.1080/2150704X.2020.1847351.

[45] A. Musialik-Lydka, B. Sredniawa, ve S. Pasyk, “Heart rate variability in heart failure”, Kardiol. Pol. (Polish Hear. Journal), c. 58, sayi 1, ss. 14-16, 2003, Eri§im: Haz. 12, 2022. [Qevrimigi]. Available at: https:/ journals.viamedica.pl/kardiologia_polska/article/vie w/82267.

[46] Y. I§ler ve M. Kuntalp, “Combining classical HRV indices with wavelet entropy measures improves to performance in diagnosing congestive heart failure”, Comput. Biol. Med., c. 37, sayi 10, ss. 1502-1510, Eki. 2007, doi: 10.1016/J.COMPBIOMED.2007.01.012.

[47] S. Guzzetti, R. Magatelli, E. Borroni, ve S. Mezzetti, “Heart rate variability in chronic heart failure”, Auton. Neurosci., c. 90, sayi 1-2, ss. 102-105, Tern. 2001, doi: 10.1016/S1566- 0702(01)00274-0.

[48] M. K. Ugar, M. R. Bozkurt, C. Bilgin, ve K. Polat, “Automatic sleep staging in obstructive sleep apnea patients using photoplethysmography, heart rate variability signal and machine learning techniques”, Neural Comput. Appl., c. 29, sayi 8, ss. 1-16, Nis. 2018, doi: 10.1007/S00521-016-2365- X/FIGURES/10. [49] M. R. Bozkurt, M. K. Ugar, F. Bozkurt, ve C. Bilgin, “In obstructive sleep apnea patients, automatic determination of respiratory arrests by photoplethysmography signal and heart rate variability”, Australas. Phys. Eng. Sci. Med., c. 42, sayi 4, ss. 959-979, Ara. 2019, doi: 10.1007/S 13246-019-00796-9/FIGURES/l 1.

[50] M. Recep Bozkurt, M. Kiir§ad Ugar, F. Bozkurt, ve C. Bilgin, “Development of hybrid artificial intelligence based automatic sleep/awake detection; Development of hybrid artificial intelligence based automatic sleep/awake detection”, IET Sci., 2020, doi: 10.1049/iet- smt.2019.0034.

[51] A. M. Maceira, S. K. Prasad, M. Khan, ve D. J. Pennell, “Normalized left ventricular systolic and diastolic function by steady state free precession cardiovascular magnetic resonance”, J. Cardiovasc. Magn. Reson., c. 8, sayi 3, ss. 417-426, 2006, doi: 10.1080/10976640600572889.

[52] M. Recep Bozkurt, • Muhammed, K. Ugar, • Ferda Bozkurt, ve • Cahit Bilgin, “In obstructive sleep apnea patients, automatic determination of respiratory arrests by photoplethysmography signal and heart rate variability”, c. 42, ss. 959-979, 2019, doi: 10.1007/S13246-019-00796-9.

[53] M. R. Bozkurt, “EMG Sinyallerinin Modern Ybntemlerle Oni§lenmesi ve Simflandinlmasi”, Sakarya University, 2007.

[54] K. M. Ramachandran ve C. P. Tsokos, “Hypothesis Testing”, Math. Stat, with Appl. R, ss. 311-369, Oca. 2015, doi: 10.1016/B978-0-12-417113-8.00006-0

[55] P. E. McKight ve J. Najab, “Kruskal-Wallis Test”, Corsini Encycl. Psychol., ss. 1-1, Oca. 2010, doi: 10.1002/9780470479216.CORPSY0491.

[56] P. E. McKnight ve J. Najab, “Mann-Whitney U Test”, Corsini Encycl. Psychol., ss. 1-1, Oca. 2010, doi: 10.1002/9780470479216.CORPSY0524.

[57] W. Haynes, “Bonferroni Correction”, Encycl. Syst. Biol., ss. 154-154, 2013, doi: 10.1007/978-l-4419-9863-7_1213

[58] C. Kier ve T. Aach, “Predicting the benefit of sample size extension in multiclass k-NN classification”, Proc. - Int. Conf. Pattern Recognit., c. 3, ss. 332-335, 2006, doi: 10.1109/ICPR.2006.942. [59] K. Z. Haigh, A. M. Mackay, M. R. Cook, ve L. G. Lin, “Machine Learning for Embedded Systems : A Case Study”, BBN Technol., 2015, Eri§im: Kas. 02, 2022. [Qevrimigi], Available at: http://www.cs.cmu.edu/~khaigh/papers/2015- HaighTechReport-Embedded.pdf.

[60] T. Averbuch vd., “Applications of artificial intelligence and machine learning in heart failure”, Eur. Hear. J. - Digit. Heal., Tern. 2022, doi: 10.1093/ehjdh/ztac025.

[61] S. S. Aung, I. Nagayama, ve S. Tamaki, “Regional distance-based k-NN classification”, ICIIBMS 2017 - 2nd Int. Conf. Intell. Informatics Biomed. Sci., c. 2018-January, ss. 56-62, §ub. 2018, doi: 10.1109/ICIIBMS.2017.8279719.

T1

Claims

1. A device (1) that detects and classifies heart failure with reduced ejection fraction (HFrEF) and heart failure with preserved ejection fraction (HFpEF) characterized in that it comprises

• An input port (6), where ECG signals of individuals are acquired and transmitted to the device (1) via ECG probes and connecting cable (8), and input port (5) of the connecting cable of PPG sensor, where PPG data of individuals are acquired and transmitted to the device (1) via PPG sensor and connecting cable (9),

• An electronic card section (2) where ECG or PPG data is recorded, digital filtering is performed, from the filtered data, 4 features are extracted if ECG testing is performed, 8 features if PPG testing is performed, these features are given to the trained model and the test result is obtained.

2. A device according to claim 1, characterized in that it comprises a speaker (7) that informs the user in cases where the device (1) will give a warning and indicates audibly that the test is over as a result of the test.

3. A device according to claim 1, characterized in that it comprises an LCD display (4) for the user to view information indicating whether the individual has heart failure with reduced ejection fraction, or heart failure with preserved ejection fraction, or is healthy.

4. An apparatus according to claim 1 , characterized in that it comprises an on and off switch (3) for switching on and off of the apparatus (1) by the user.

5. A device according to claim 1, characterized in that 5 of the 8 features extracted when testing with PPG are geometric mean, Hjort mobility parameter, mean curve length, Burg method 4th order model's estimated variance of white noise input and Burg method 4th order model's 1st reflection coefficient, 3 of them are PPG derived HRV's interquartile range, median and standard error features.

6. A device according to claim 1, characterized in that the 4 features extracted when testing with ECG are skewness, Burg method 4th order model's 3rd AR parameter, Burg method 4th order model's 2nd reflection coefficient and Burg method 4th order model's 3rd reflection coefficient features.

7. A method of operation of a device (1) that detects and classifies heart failure with reduced ejection fraction and heart failure with preserved ejection fraction characterized in that it comprises the process steps • For the detection and classification of HF, if ECG testing is to be performed, ECG data are obtained from the right wrist, right and left ankle of the subject via ECG probes and connection cable (8) or for the detection and classification of HF, if PPG testing is to be performed, PPG data are obtained from the index finger of the subject's right hand via PPG sensor connection cable (9),

• Recording of ECG data or PPG data,

• Digital filtering,

• By a processor, if testing with PPG, HRV is derived from filtered PPG and extracting a total of eight features, 5 from PPG and 3 from HRV, or if testing with ECG, extracting 4 features from filtered data,

• Processing the extracted features by the system and obtaining the test result.

8. The method of operation of the device (1) for detecting and classifying heart failure with reduced and preserved ejection fraction according to claim 7, characterized in that the process step of performing digital filtering comprises the process steps of

• Applying a Chebyshev Type II bandpass filter, in the range of 0.25 - 100 Hz, to the signal,

• Applying a notch filter in the range of 49 - 51 Hz to the signal to remove mains noise at 50 Hz,

• Applying a moving average filter.