Automatic ECG analysis

Abstract

Clinical monitoring studies require feature extraction from the electrocardiogram (ECG) in order to evaluate the state of the heart of the patient. Online and offline delineation and classification of an electrocardiogram (ECG) are applied automatically. The wavelet transformation helps in the delineation of ECG. Our ECG delineation system scans and detects the waves from the ECG signal from each lead. From these waves, two lists of handcrafted features are generated, which are the amplitude of each wave, and the time duration between waves. The features extracted from the ECG are used to detect and identify heart diseases. The k-means clustering algorithm is applied to detect the strange peak that can indicate unfamiliar diseases.

Description

[0001] Automatic ECG Analysis

[0002] Field of the invention

[0003] This invention relates to electrocardiographic (ECG) systems that can sequentially identify, analyze, and classify ECG recordings, collectively known as ECG analysis. It more precisely relates to the real-time (online) delineation of such signals.

[0004] State of the art

[0005] The analysis of ECG recordings within the prior art has substantially improved throughout the years by an apparent unrelenting improvement in the sophisticated analysis system available within the industry. The noninvasive electrocardiogram (ECG) has long been used as means to diagnose diseases reflected by disturbances of the heart’s electrical activity. Beyond traditional electrocardiography, the automated processing and analysis of the ECG signal have been receiving significant attention and have witnessed substantial advances. In particular, a large body of algorithms has been proposed for the detection of the ECG characteristic waves, so-called ECG delineation, following a variety of approaches based on convolutional neural networks, artificial neural networks, or low-pass differentiation.

[0006] Traditionally, the automatic analysis of ECG signals, including filtering and delineation, was either performed online using bulky, high-performance bedside cardiac monitors, or performed offline during a postprocessing stage after ambulatory ECG recording using wearable, obtrusive ECG data loggers. Currently, however, a significant industrial and academic effort is mainly on online automatic ECG analysis, wearable, and wireless ECG monitors which are enabled in mobile cardiology systems. The resulting products are a single-lead ECG monitor, heart rate (HR), and skin temperature measurements. And the analysis of ECG performs continuous arrhythmia detection based on HR measurements only. Accordingly, state-of-the-art unobtrusive wireless mobile / ambulatory ECG monitors are single lead and limited to HR measurement and analysis to detect arrhythmia.

[0007] Summary

[0008] This summary is provided to give a simple form of concepts and techniques used in this present invention. This system provides a continuous and simultaneous recording of a wearable device and analysis of ECG signals. The present invention relates to a method for computerizing the delineation of ECG signals: applying the wavelet transform on an ECG signal to analyze and assign each wave in the ECG signal to its time point.

[0009] The significance of the wavelet transformation appears when the detection of waves starts. Wavelet transforms are mathematical tools for analyzing data where features vary over different scales. These scales are divided into two categories: approximated and detailed. The detailed scales are five levels of derivative of ECG signals. The importance of each level differs according to how each wave would be detected. For example, the third

[0010] 1

[0011] SUBSTITUTE SHEET (RULE 26) level in the wavelet is so important because it is used to detect the QRS -complex. And if this QRS-complex does not exist in the ECG signal, this indicates that the patient has asystole or ventricular fibrillation, and this patient needs to be rescued from death.

[0012] Depending on these time points, the system can calculate the amplitude of each wave and the time interval between each point and the other. These features are fed into the second model which is the diagnostic model.

[0013] The diagnostic model uses amplitudes and time intervals to detect diseases. This model is built on clinical information; such as the difference in the shape of the QRS complex, the hidden P-wave, or the T-wave that disappeared due to the ST-elevation. The doctors use these waves in the diagnosis of the patients. This model is considered a multi-label classification of an ECG signal.

[0014] After the diagnostic model accomplishes its task, the clustering algorithm starts to work. This algorithm collects similar peaks together. The goal of this algorithm is to define the aberrantly or strange peak that appears in the ECG signal of the patient. This algorithm helps the doctor in identifying the diseases that the patient has.

[0015] Detailed Description

[0016] The present invention relates to the analysis of ECG signals based on wavelet transformation.

[0017] The algorithms used here are 4 algorithms; the first one is for getting the wavelet transformation to get the derivative levels for each lead, the second one is for detecting the time point for each wave, the third one is for supervised learning, and the fourth one is for unsupervised learning. Supervised learning aims to predict an output vector from an input vector, while that unsupervised learning is to get the structure or pattern between data that helps to collect the data in clusters.

[0018] The first algorithm is the wavelet transform that takes the input data X. X is an ECG matrix of size mxn, and Y is a matrix of the features extracted from each lead. For example, in figure 1, if 5 electrodes are connected to the patient (producing 7 lead connections), the ECG input data shape becomes no. samples x no. lead

[0019] So, the output data is the ECG features extracted, and its shape becomes no. leads x 5

[0020] The second algorithm is the delineation that is used for detecting the time point for each wave in the ECG. The waves of the ECG are P-waves, QRS -complexes, and T-waves. After detecting the time points by this algorithm, the algorithm will detect the amplitude of each wave and the time interval between the two waves like PR-interval, RR-interval, and so on.

[0021] 2

[0022] SUBSTITUTE SHEET (RULE 26) From figure 2, after getting the wavelet levels, they are used to detect the waves. The delineation of the ECG is applied using the sliding window method. First, the detection of the QRS complex is built on the zero crossing and the slope variation measurement from the signal. Before starting with the details of how to get the QRS complex, each peak in the signal is converted to a zero-crossing value in the derivative of this signal. To detect the QRS complex, firstly search for the zero-crossing time point, positive and negative maximum peaks in the third level of the wavelet level that represent the derivative of the signal. Then the sum of each zero crossing is calculated by summing the positive and negative maximum peak values. And based on the sum values of each zero-time point, the time points are clustered to get the time point that indicates the real QRS complex, not noise.

[0023] The previous way is applied to each lead to get the time point of the QRS complex and the number of QRS complexes. Then compare the number of QRS complexes between each lead to find the missing beat in each lead if it existed. The detection of QRS complex is done first because there are lethal conditions that do not contain QRS complexes such as asystole and ventricular fibrillation. If one of these diseases exists, the delineation algorithm will stop working.

[0024] From figure 3, The delineation of the ECG will continue to detect the remaining waves. After detecting the time point of the QRS complex in each lead, the time point of the R- wave, Q-derivative wave, and S-wave will be calculated. The time point of the QRS complex indicates the R-wave. The Q-wave will be the negative peak before the R-wave. The S-wave will be the negative peak after the R-wave. If the RSR’ (R-wave, followed by an S-wave, followed by another R-wave) state exists, the ECG delineation algorithm will detect the R’ wave as a positive peak after the R wave.

[0025] From these features, the RR-time interval will be calculated by getting the time difference between two-time points, each one indicating the position of the R-wave. And the QRS onset will be detected by getting the zero crossing before the QRS complex and in the same way, the QRS offset will be detected after the QRS complex.

[0026] Secondly, the time point of the T-wave will be calculated that exists in the half of RR- interval. The T-wave will be detected from the zero-crossing that exists in the third derivative and the second derivative of the ECG signal. But there is a special case when the ST-segment becomes high, masking the T-wave. So, the search for zero-crossing will become in the rest of the wavelet levels. After that, the T-wave onset and offset must be detected to get the remaining waves in the ECG signal. The detection of these time points is calculated by keeping track of the slope away from the peak until it becomes lower than a certain threshold set empirically.

[0027] Then the J-wave will be detected. The J point in the ECG is the point where the QRS complex joins the ST segment. It represents the approximate end of depolarization and the beginning of repolarization as determined by the surface ECG. It is represented as a zero-

[0028] 3

[0029] SUBSTITUTE SHEET (RULE 26) crossing in each of the wavelet levels that exists between the offset of the QRS complex and the onset of the T-wave.

[0030] Then the P-wave will be detected. The P-wave is the first positive deflection on the ECG and represents atrial depolarization. In the normal ECG signal, the P-wave time point is detected from the offset of the T-wave to the onset of the QRS complex. This time point is the zero-crossing point that exists in the fifth level of the wavelet transformation. The P- wave has different morphologies such as P-mitrale, P pulmonale, inverted P-wave, far P- wave, and normal. The P-mitrale is notched and its duration is longer than 120 ms. The detection of this type in P-wave is used to identify the left atrial enlargement. In the P pulmonale state, the P wave is tall and peaked P-wave in inferior leads, VI, and V2. The existence of this type indicates right atrial enlargement. The inverted P-wave is usually manifested in the inferior leads, accompanied by a PR interval of less than 120 ms. This type of P-wave is helpful in identifying atrial rhythms. The detection of an inverted P-wave is done by checking the slopes around the P-wave. In some diseases like heart block 2 and heart block 3, there is a P-wave without the QRS complex. So, we always check the presence of P-wave and change in the RR interval in the window to check whether heart block type 2 or 3 exists. Then we differentiate between heart block 2 and heart block 3 by calculating the variance in the RR interval and the variance of PP interval in the time window of 5 seconds and comparing the two variances. If the two variances are sufficiently different, the Automatic Delineation method will identify it as heart block3. If one of the RR intervals is greater than double of each RR interval, and if a far P-wave and increasing PR interval exist, then the automatic delineation method will classify it as heart block 2.

[0031] As shown in figure 4, from these time points of each wave, the calculation of the amplitude and duration of each wave is done. After the delineation algorithm is done, the output matrix will contain all the handcraft features from each lead that will be used as input in classifying diseases. These features included ECG features in time domain; amplitudes and durations of P-wave, Q-wave, R-wave, S-wave, and T-wave, time intervals like RR- interval, PR-interval, QT-interval, and wavelet features like zero-crossing peaks, positive peaks, and negative peaks. The classification algorithm is supervised learning which is the machine learning task of learning a function that maps an input to an output based on example input-output pairs. It infers a function from labeled training data consisting of a set of training examples. In supervised learning, each example is a pair consisting of an input object (typically a vector) and the desired output value (also called the supervisory signal). A supervised learning algorithm analyzes the training data and produces an inferred function, which can be used for mapping new examples. Before talking about the algorithm, let’s represent the list of diseases that existed in the system. These diseases are divided according to the severity of diseases into 3 categories: low, medium, and high.

[0032] These diseases are premature ventricular contraction, atrial flutter, right bundle branch block, left bundle branch block, atrial fibrillation, premature atrial contraction, heart block 1, heart block 2, and heart block 3. After that, the detection of the ST segment and calculating the level of segment and type of the STEMI (ST-elevation myocardial

[0033] 4

[0034] SUBSTITUTE SHEET (RULE 26) infarction) pattern from ST-segment elevation. Then the ST level algorithm’s validation starts to check the ST level’s accuracy.

[0035] The fourth algorithm is the clustering one, based on unsupervised learning. The clustering algorithm is an unsupervised method, where the input is not a labeled one, and problemsolving is based on the experience that the algorithm gains out of solving similar problems as a training schedule. This algorithm is used to collect the peaks that are similar to each other. The similarity between peaks depends on the handcrafted features that are extracted from the delineation of the ECG. These features are the amplitude and duration of the time of each wave and the time interval between each QRS complex.

[0036] This algorithm aims to explore the strange or different shapes of the QRS complex because this algorithm helps doctors to define the existence of diseases that are not familiar. This algorithm works online during ECG monitoring and offline after the patient session ends. This algorithm works on web and mobile applications to help doctors in diagnosing strange diseases at any time. For example, this algorithm helps to differentiate between premature ventricular fibrillation and aberrantly premature atrial contraction. A premature ventricular complex (PVC) is a premature beat arising from an ectopic focus within the ventricles. An aberrantly premature atrial complex (PAC) may be conducted aberrantly in a right or left bundle pattern, which can be a challenge to differentiate them from PVCs. To solve this conflict, we make the algorithm check for the existence of some characteristics like the existence of P-wave and the compensatory pause after the QRS complex. If the P-wave exists with compensatory pause and the remaining characteristics of PVC, it will identify the QRS complex as aberrantly PAC. If the compensatory pause and the remaining characteristics exist but the P-wave does not exist, it will classify the QRS complex as PVC.

[0037] Brief Description of the Drawings

[0038] Figure 1. illustrates the ECG signal from each lead when only 5 electrodes are used.

[0039] Figure 2. illustrates a P-wave, a T-wave, and a QRS-complex in the ECG pattern.

[0040] Figure 3. illustrates a PR interval, an ST interval, an RR interval, and a QT interval in an ECG pattern.

[0041] Figure 4. illustrates a block diagram that describes the steps of delineation of the ECG signals.

[0042] 5

[0043] SUBSTITUTE SHEET (RULE 26)

Claims

Claims A method of automatic ECG analysis using ECG sensors or electrodes, which comprises: a) getting the wavelet transformation of the ECG signal for each lead, b) detecting the time points of the P-wave, QRS complex, and T-wave of ECG signal for each lead, c) analyzing the time points to calculate the amplitude and duration of each wave, d) classifying the QRS peaks and detecting the heart diseases, e) calculating the ST segment level and validating its value, and f) clustering the QRS peaks to detect the strange or uncommon peaks that can indicate some diseases online and offline. The automatic ECG analysis method of claim 1, where the ECG delineation method gets the wavelet transformation of the ECG signal is used i. to de-noise the ECG signals, ii. to decompose the signal into detailed components and approximate components of different scales, and iii. to detect time points based on the wavelet transformation levels. The automatic ECG analysis method of claims 1 through 2, where the ECG delineation method is applied using the window method to detect the time points of P-wave, QRS complex, and T-wave of ECG signal for each lead continuously in real-time. The automatic ECG analysis method of claims 1 through 3, where the ECG delineation method detects the time point P-wave, QRS complex, and T-wave of the ECG signal for each lead, comprises: i. dividing the ECG signal into 3 regions; P-wave indicating atrial depolarization, the QRS complex corresponding to the depolarization of the right and left ventricles of the heart contraction, and the T-wave as the positive deflection after each QRS complex, representing ventricular repolarization, ii. each wave on the ECG has a zero-crossing value that can be detected from each wavelet transformation level, especially from the detailed components, iii. each wave in the QRS complex has two peaks that are positive and negative in the derivative signals. These peaks are used to detect the state of the wave whether it is normal or inverted. The automatic ECG analysis method of claims 1 through 4, where the ECG delineation method analyzes the time points to calculate the amplitude and duration of each wave in each lead, comprises: i. calculating the amplitude of P-wave, Q-wave, R-wave, S-wave, and T-wave with respect to the baseline, and ii. calculating the duration of each wave as well as intervals: QRS duration, PR interval, RR interval, QT interval, P-wave duration, Q-wave duration, R-wave duration, and S-wave duration. The automatic ECG analysis method of claims 1 through 5, where the ECG classification method classifies the heart disease that appeared in the QRS complex based on ECG features, uses:6SUBSTITUTE SHEET (RULE 26)i. ECG features from time domain extracted from automatic ECG delineation; amplitude and duration of each wave in the QRS complex, P wave, Q wave, R wave, S wave, T wave, ii. ECG features from time domain extracted from the ECG delineation; the time interval between every two waves; RR Interval, PR Interval, QT Interval, and iii. ECG features from wavelet domain extracted from wavelet transformation; zero crossing values, positive peaks, negative peaks. . The automatic ECG analysis method of claims 1 through 6, where the ECG classification method classifies the QRS peaks and detects the heart diseases, comprises: i. using the handcrafted features of claim 6 that are extracted from the delineation of the ECG as input for this unique algorithm, and ii. each change in the morphology of the QRS complex due to disease changes in the ECG features in the time domain and wavelet domain. . The automatic ECG analysis method of claims 1 through 7, where the ECG delineation method calculates the ST segment level and validates its value, comprises: i. detecting the J wave point in each QRS complex, and ii. calculating and validating the ST value in each QRS complex. . The automatic ECG analysis method of claims 1 to 8, where the ECG clustering method clusters the QRS peaks to detect the strange or uncommon peaks that can indicate some chronic diseases online during ECG monitoring and offline after the patient session ends and from the web and mobile application, comprises: i. using ECG features to calculate the similarity between ECG peaks, ii. collecting the peaks that are similar to each other, and iii. making another cluster to collect strange ECG peaks.

0. An automatic ECG analysis system using ECG sensors or electrodes, which comprises: a) getting the wavelet transformation of the ECG signal for each lead, b) detecting the time points of the P-wave, QRS complex, and T-wave of ECG signal for each lead, c) analyzing the time points to calculate the amplitude and duration of each wave, d) classifying the QRS peaks and detecting the heart diseases, e) calculating the ST segment level and validating its value, and f) clustering the QRS peaks to detect strange or uncommon peaks that can indicate some diseases online and offline.

1. The automatic ECG analysis system of claim 10, where the ECG delineation method gets the wavelet transformation of the ECG signal is used: i. to de-noise the ECG signals, ii. to decompose the signal into detailed components and approximate components of different scales, and iii. to detect time points based on the wavelet transformation levels.

2. The automatic ECG analysis system of claims 10 through 11, where the ECG delineation method is applied using the window method to detect the time points of P-7SUBSTITUTE SHEET (RULE 26)wave, QRS complex, and T-wave of ECG signal for each lead continuously in realtime. . The automatic ECG analysis system of claims 10 through 12, where the ECG delineation method detects the time point P-wave, QRS complex, and T-wave of ECG signal for each lead, comprises: i. dividing the ECG signal into 3 regions; P-wave indicating atrial depolarization, the QRS complex corresponding to the depolarization of the right and left ventricles of the heart contraction, and the T-wave as the positive deflection after each QRS complex, representing ventricular repolarization, ii. each wave on the ECG has a zero-crossing value that can be detected from each wavelet transformation level, especially from the detailed components, iii.each wave in the QRS complex has two peaks that are positive and negative in the derivative signals. These peaks are used to detect the state of the wave whether it is normal or inverted. . The automatic ECG analysis system of claims 10 through 13, where the ECG delineation method analyzes the time points to calculate the amplitude and duration of each wave in each lead, comprises: i. calculating the amplitude of P-wave, Q-wave, R-wave, S-wave, and T-wave with respect to the baseline, and ii. calculating the duration of each wave as well as intervals: QRS duration, PR interval, RR interval, QT interval, P-wave duration, Q-wave duration, R-wave duration, and S-wave duration. . The automatic ECG analysis System of claims 10 through 14, where the ECG classification method classifies the disease that appeared in the QRS complex based on ECG features, uses: i. ECG features from time domain extracted from the ECG delineation; amplitude and duration of each wave in the QRS complex, P wave, Q wave, R wave, S wave, T wave, ii. ECG features from time domain extracted from automatic ECG delineation; the time interval between every two waves; RR interval, PR interval, QT interval, and iii. ECG features from wavelet domain extracted from wavelet transformation; zero crossing values, positive peaks, and negative peaks. . The automatic ECG analysis system of claims 10 through 15, where the ECG classification method classifies the QRS peaks and detects the heart diseases, comprises: i. using the handcrafted features of claim 15 that are extracted from the delineation of ECG as input for this unique algorithm, and ii. each change in the morphology of the QRS complex due to disease changes in the ECG features in the time domain and wavelet domain. . The automatic ECG analysis system of claims 10 through 16, where the ECG delineation method calculates the ST segment level and validates its value, comprises: i. detecting the J-wave point in each QRS complex, and8SUBSTITUTE SHEET (RULE 26)ii. calculating and validating the ST value in each QRS complex. The automatic ECG analysis system of claims 10 through 17, where the ECG clustering method clusters the QRS peaks to detect the strange or uncommon peaks that can indicate some chronic diseases online during ECG monitoring and offline after the patient session ends and from the web and mobile application, comprises: i. using ECG features to calculate the similarity between ECG peaks, ii. collecting the peaks that are similar to each other, and iii. making another cluster to collect strange ECG peaks.9SUBSTITUTE SHEET (RULE 26)

Images