Detection of cardiac condition from reduced lead group ECG

Abstract

Embodiments of the present disclosure utilize a DNN for ECG interpretation, where the DNN uses pairs of interpretation markers to directly ingest an original ECG waveform for training, without the need for illustrative feature extraction or rule-based criteria.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Provisional Application No. 63 / 328,670, filed on April 7, 2022, entitled “SINGLE ALGORITHM DETECTION OF CARDIAC CONDITIONS FROM REDUCED LEAD SET ECG,” the disclosure of which is incorporated herein by reference. Technical Field

[0003] Aspects of the present disclosure relate to diagnosing a cardiac condition based on an electrocardiogram taken with a reduced lead set, and in particular, to algorithmically diagnosing a cardiac condition using a reduced lead set and a single algorithm. Background Art

[0004] Cardiovascular disease is the leading cause of death in the world. In 2008, 30% of all deaths worldwide were attributable to cardiovascular disease. Furthermore, it is estimated that by 2030, more than 23 million people will die each year from cardiovascular disease. Cardiovascular disease is equally prevalent in populations of first and third world countries and affects people regardless of socioeconomic status.

[0005] Arrhythmia is a heart condition in which the electrical activity of the heart is irregular or faster (tachycardia) or slower (bradycardia) than normal. Although many arrhythmias are not life-threatening, some arrhythmias may lead to cardiac arrest or even sudden cardiac death. In fact, cardiac arrhythmias are one of the most common causes of death when going to the hospital. Atrial fibrillation (A-fib) is the most common cardiac arrhythmia. In A-fib, electrical conduction through the heart's ventricles is irregular and chaotic. Although A-fib may not cause symptoms, it is often associated with palpitations, shortness of breath, fainting, chest pain or congestive heart failure, and also increases the risk of stroke. A-fib is usually diagnosed by taking an electrocardiogram (ECG) of the subject. In order to treat A-fib, patients can take drugs to slow down the heart rate or adjust the rhythm of the heart. Patients can also take anticoagulants to prevent stroke, or can even undergo surgical intervention including cardiac ablation to treat A-fib. In other examples, the ECG can provide decision support for acute coronary syndrome (ACS) or other cardiac conditions by interpreting various rhythm and morphological conditions including myocardial infarction (MI) and ischemia.

[0006] Typically, patients with A-fib (or other types of arrhythmias) are monitored over extended periods of time to manage the disease. For example, a patient may be provided with a Holter monitor or other dynamic electrocardiographic device to continuously monitor the electrical activity of the cardiovascular system over, for example, at least 24 hours. Such monitoring may be critical to detecting conditions such as acute coronary syndrome (ACS).

[0007] The American Heart Association and the European Society of Cardiology recommend that a 12-lead ECG be obtained as early as possible for patients with possible ACS or other cardiac conditions when symptoms are present. Prehospital ECGs have been found to significantly reduce treatment time and show better survival rates. The time to first ECG is so important that it is a quality and performance metric monitored by several regulatory agencies. According to the National Health Statistics in 2015, in the United States (US), more than 7 million people visited the Emergency Department (ED) with complaints of chest pain or symptoms related to a cardiac condition. In the United States, ED visits are growing at a rate of 3.2% per year, and outside the United States, ED visits are growing at 3% to 7% per year. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The described embodiments and their advantages may be best understood by referring to the following description taken in conjunction with the accompanying drawings which in no way limit any changes in form and detail that may be made to the described embodiments by those skilled in the art without departing from the spirit and scope of the described embodiments.

[0009] Figure 1A is a diagram illustrating an example system according to some embodiments of the present disclosure.

[0010] Figure 1B is a block diagram illustrating an example system according to some embodiments of the present disclosure.

[0011] Figure 2 is a diagram illustrating the various leads that may make up an ECG according to some embodiments of the present disclosure.

[0012] Figure 3 is a block diagram illustrating an example system according to some embodiments of the present disclosure.

[0013] Figure 4 is a block diagram illustrating the structure of a machine learning (ML) model for reduced-lead ECG interpretation according to some embodiments of the present disclosure.

[0014] Figure 5Ais a block diagram showing the basic structure of a deep neural network (DNN) according to some embodiments of the present disclosure.

[0015] Figure 5B is a block diagram illustrating an example convolutional neural network (CNN) according to some embodiments of the present disclosure.

[0016] Figure 5C is a block diagram illustrating an example training process for a CNN according to some embodiments of the present disclosure.

[0017] Figure 6 is a block diagram illustrating an example ECG rhythm classification model according to some embodiments of the present disclosure.

[0018] Fig. 7A is a block diagram illustrating an example ECG morphology classification model according to some embodiments of the present disclosure.

[0019] Figure 7B and Figure 7C are block diagrams illustrating an ECG signal and a median beat of the ECG signal, respectively, according to some embodiments of the present disclosure.

[0020] Figure 8 is a block diagram illustrating an example global ECG feature model according to some embodiments of the present disclosure.

[0021] Fig. 9 The method for verifying some embodiments of the present disclosure is shown. Figure 6 Example conditions for classification of the ECG rhythm classification model.

[0022] Fig.10 is a flow chart of a method for training and implementing a reduced-lead ECG interpretation model according to some embodiments of the present disclosure.

[0023] Fig.11 is a block diagram of an example computing device that can perform one or more of the operations described herein, according to some embodiments of the present disclosure. DETAILED DESCRIPTION

[0024] In the following detailed description of the embodiments of the present disclosure, many specific details are set forth to provide a more thorough understanding of the present disclosure. However, it will be apparent to those of ordinary skill in the art that the concepts within the present disclosure may be practiced without these specific details. In other instances, well-known features are not described in detail to avoid unnecessarily complicating the specification.

[0025] An electrocardiogram (ECG) provides multiple ECG waveforms representing the electrical activity of a person's heart. An ECG monitoring device may include an electrode set for recording these ECG waveforms (also referred to herein as "obtaining an ECG") of a patient's heart. The electrode set may be placed at multiple locations on the patient's skin, and the electrical signals (ECG waveforms) recorded between each electrode pair in the electrode set may be referred to as leads. Different numbers of leads may be used to obtain an ECG, and various leads may be formed using different numbers and combinations of electrodes. Example numbers of leads for obtaining an ECG are 1, 2, 3, 6, and 12 leads.

[0026] The ECG waveforms recorded by the ECG monitoring device (each ECG waveform corresponds to the lead of the ECG) may include data corresponding to the electrical activity of the human heart. As known in the art, a typical heartbeat may include several potential changes, which may be classified as waves and complexes, including P waves, QRS complexes, T waves, and U waves, etc. In other words, as known in the art, each ECG waveform may include P waves, QRS complexes, T waves, and U waves, etc. The shape and duration of these waves may be related to various characteristics of the human heart, such as the size of the human atrium (e.g., indicating atrial enlargement), etc., and may be the first source of human-specific heartbeat characteristics. (Usually after standard filtering and "cleaning" of the signal) the ECG waveform may be analyzed for various indicators, which are useful in detecting cardiac events or states such as cardiac arrhythmia detection and characterization. Such indicators may include, for example, ECG waveform amplitude and morphology (e.g., QRS complex amplitude and morphology), R wave ST segment and T wave amplitude analysis, and heart rate variability (HRV).

[0027] As described above, an ECG waveform is generated by measuring multiple leads (each lead is formed by a different electrode pair), and the ECG waveform obtained from each different electrode pair / lead may be different / unique (e.g., may have a different morphology / amplitude). This is because although various leads may analyze the same electrical event, each lead may do so from a different angle. Figure 1A A diagram 105 showing ECG waveforms detected by each of the three leads (I, II, and III) when a 3-lead ECG is obtained and an exploded view 110 of the ECG waveform measured by lead III showing a QRS complex. As shown in the figure, the ECG waveforms obtained from leads I to III are different in amplitude and shape, wherein the ECG waveform measured by lead III has the largest amplitude, and the ECG waveform measured by lead I has the smallest amplitude.

[0028] There are different "standard" configurations of electrode placement that can be used to place electrodes on a patient. For example, an electrode placed on the right arm may be referred to as RA. An electrode placed on the left arm may be referred to as LA. The RA and LA electrodes may be placed at the same location on the left and right arms, preferably near the wrist in some embodiments. The leg electrode may be referred to as RL for the right leg and LL for the left leg. The RL and LL electrodes may be placed at the same location on the left and right legs, preferably near the ankle in some embodiments. Lead I is typically the voltage between the left arm (LA) and the right arm (RA), for example, I=LA-RA. Lead II is typically the voltage between the left leg (LL) and the right arm (RA), for example, II=LL-RA. Lead III is typically the voltage between the left leg (LL) and the left arm (LA), for example, III=LL-LA. Enhanced limb leads may also be determined from RA, RL, LL, and LA. The enhanced vector right (aVR) lead is equal to RA-(LA+LL) / 2 or -(I+II) / 2. The augmented vector left (aVL) lead is equal to LA-(RA+LL) / 2 or I-II / 2. The augmented vector foot (aVF) lead is equal to LL-(RA+LA) / 2 or II-I / 2.

[0029] It should be noted that groups of two or more leads can be transformed to generate a complete 12-lead ECG. This transformation can be performed using a machine learning model (e.g., a neural network, deep learning techniques, etc.). The machine learning model can be trained using 12-lead ECG data corresponding to a population of individuals. The data can be preprocessed before being input into the machine learning model to filter the data in a manner suitable for the application. For example, the data can be classified according to height, gender, weight, nationality, etc. before being used to train one or more machine learning models, so that the resulting one or more models are fine-tuned for a specific type of individual. In another embodiment, the machine learning model can be further trained based on the user's own ECG data to further fine-tune and personalize the model to reduce any residual synthesis errors.

[0030] As discussed herein, since pre-hospital ECG has been found to significantly reduce treatment time and show better survival rates, 12-lead ECG should be obtained as early as possible for patients who may have heart conditions when symptoms present. In addition, current dynamic ECG devices (such as Holter monitors, etc.) are usually bulky and difficult to implement for the subject without the help of medical professionals. For example, using a Holter monitor requires patients to wear a bulky device on their chests and accurately place multiple electrode leads on the precise positions of their chests. These requirements may hinder the activities of the subject, including their natural movements (such as bathing and showering, etc.). Once such a device obtains ECG, the ECG is sent to the subject's physician, who then analyzes the ECG waveform and provides diagnosis and other suggestions. At present, this process must usually be carried out by hospital managers and health management organizations, and many patients do not receive feedback in a convenient way.

[0031] A 12-lead ECG can be implemented in a hospital or medical clinic setting to detect cardiac conditions, however, it would be desirable to have a simpler reduced lead system that can be used at home, or in a doctor's office, or in an emergency center, or in the pre-hospital setting to detect cardiac conditions. There are certain reduced lead (e.g., 6-lead) devices that can mathematically construct all 12 leads based on a reduced lead set through linear or nonlinear transformations to provide a set of all 12 leads (some leads are now "synthesized leads"). This synthesized full 12-lead data can be analyzed by a physician or by a full 12-lead algorithm. Based on the existing technology of reconstructing a full 12-lead set, the natural workflow of an automated system would be as follows: 1) create a 12-lead set from a reduced lead set (say, 4 leads) using a lead reconstruction algorithm, and then 2) use a pathology classification algorithm to take the reconstructed 12 leads as input and provide a determination (classification) of normal versus various pathologies as output.

[0032] However, there are inherent tradeoffs with such reduced lead devices as they offer lower resolution / ability to detect and discriminate the full range of cardiac conditions using the standard 12-lead ECG guidelines recommended by the American College of Cardiology (ACC / AHA). It would be desirable to have the discrimination power of a 12-lead system with the simplicity of a reduced lead set that can be used outside of the hospital (at home, in an emergency, in a doctor's office, etc.) but without the tradeoff of sacrificing discrimination resolution.

[0033] A deep neural network (DNN) is a model consisting of multiple layers of interconnected nodes, each layer building on the previous layer to refine and optimize predictions or classifications. This computational progression through the network is called forward propagation. The input and output layers of a deep neural network are called visible layers. The input layer is where the deep learning model ingests data for processing, and the output layer is where the final prediction or classification is made. DNNs are known to have excellent performance in highly nonlinear and highly complex systems, especially in the case of trying to "classify" multiple algorithm steps into a single step. DNNs can use a technique called "end-to-end" learning to essentially classify all (one or more than one) algorithms used for diagnosis based on ECG data into a single mathematical process. By utilizing this end-to-end training of what is a multi-task / ML model, the resulting DNN can provide high-precision results, which can often exceed the performance provided by using multiple dedicated algorithms.

[0034] Figure 1B A system 100 for cardiac disease management is shown. The system 100 may be provided for use by a first user, such as a physician of the first user. Alternatively, the system 100 may be used without input from a physician or other third party. The system 100 may include a local computing device 101 of the first user. The local computing device 101 may be loaded with a user interface, dashboard, or other subsystem of the cardiac disease management system 100. For example, the local computing device 101 may be loaded with a mobile software application ("mobile application") 101A for interfacing with the system 100. The mobile application 101A may be configured to interface with one or more biometric sensors (e.g., an ECG monitor 103), and may include software and a user interface for managing biometric data collected by the local computing device 101 from one or more biometric sensors. For example, the local computing device 101 may include any suitable computing device such as a tablet computer, a smart phone, a server computer, a desktop computer, a laptop computer, or a body-worn computing device (e.g., a smart watch or other wearable device). In some embodiments, the local computing device 101 may include a single computing device, or may include multiple interconnected computing devices (eg, multiple servers configured in a cluster).

[0035] The local computing device 101 may be coupled to one or more biometric sensors. For example, the local computing device 101 may be coupled to an ECG monitor 103, which may include an electrode group for recording ECG (electrocardiogram) data (also referred to herein as "obtaining ECG") of the heart of the first user. ECG data may be recorded or obtained using an electrode group placed in multiple locations on the skin of the first user. The electrical signals recorded between the electrode pairs may be referred to as leads, and different numbers of leads may be used to record ECG data, and various leads may be formed using different numbers and combinations of electrodes (placed on the user's arms, legs, and chest). Example numbers of leads for obtaining ECG are 1, 2, 6, and 12 leads. The electrode placed on the right arm may be referred to as RA. The electrode placed on the left arm may be referred to as LA. The RA and LA electrodes may be placed at the same location on the left and right arms, such as near the wrist. The leg electrode may be referred to as RL for the right leg, and may be referred to as LL for the left leg. The RL and LL electrodes may be placed at the same location on the left and right legs, such as near the ankles.

[0036] Figure 2 A single dipole heart model 115 is shown with a 12-lead set including leads I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, and V6, all of which are represented on a six-axis system. The heart model 115 assumes a homogeneous cardiac field that changes only in amplitude and direction over a period of time in all directions. Figure 2As shown, there are 2 orthogonal planes, i.e., the frontal plane and the horizontal plane. In each plane, there are several leads to cover the entire plane. In the frontal plane, there are 2 independent leads I and II, and 4 other derived leads III, aVR, aVL and aVF separated by 30 degrees respectively. The reason why the frontal plane has 2 independent leads is that these 2 independent leads are far-field leads, and these far-field leads can each cover a wider viewing angle like a wide-angle camera lens but provide less details. In the horizontal plane, there are usually 6 independent leads, and these 6 independent leads are closer to the heart than the limb leads and can be called near-field leads. Following the same analogy of the camera, the near-field lead can behave like a zoom lens that covers less viewing angle but has higher accuracy for local activities (such as ischemia and infarction). The two orthogonal planes are related by using a synthetic reference point formed by leads I and II, which is called Wilson-Central-Terminal (WCT). WCT is defined as RA+LA+RL / 3, but considering that both Leads I and II are recorded with reference to RA, so that the voltage of RA can be considered to be zero, WCT(VW) can be calculated using RA as the reference for both Leads I and II (thus assuming it has zero potential) as: Lead I+Lead II / 3.

[0037] Return to reference Figure 1B In some embodiments, the ECG monitor 103 may include (e.g., such as from Company Device or 6L device, etc.) that includes a small number of electrodes (e.g., two, three, four, or any number of electrodes less than twelve). In these embodiments, the electrodes can be used to measure Figure 2A subset of leads is shown, such as measuring lead I (e.g., the voltage between the left and right arms) simultaneously with lead II (e.g., the voltage between the left leg and the right arm) and measuring lead I simultaneously with lead V2 or another lead in the chest leads (such as V5, etc.). It should be noted that any other combination of leads is possible. If necessary, additional leads can be derived algorithmically from the determined subset of leads (e.g., by the ECG monitor 103 itself or the local computing device 101). For example, enhanced limb leads can also be determined from the values ​​measured by the LA, RA, LL, and RL electrodes. The enhanced vector right (aVR) can be equal to RA-(LA+LL) / 2 or -(I+II) / 2. The enhanced vector left (aVL) can be equal to LA-(RA+LL) / 2 or I-II / 2. The enhanced vector foot (aVF) can be equal to LL-(RA+LA) / 2 or II-I / 2. In some embodiments, the ECG monitor 103 itself or the local computing device 101 can utilize a machine learning (ML) model to derive a complete 12-lead set from a measured subset of leads. In some embodiments, the ECG monitor 103 can be in the form of a smart phone or a wearable device such as a smart watch. In some embodiments, the ECG monitor 103 can be a handheld sensor coupled to the local computing device 101 through an intermediate protective case / adapter.

[0038] For example, the ECG data recorded by the ECG monitor 103 may include the electrical activity of the heart of the first user. As is known in the art, a typical heartbeat may include several electrical potential changes, which may be classified into waves and wave groups, including P waves, QRS complexes, T waves, and sometimes U waves. The shape and duration of the P wave may be related to the size of the user's atria (e.g., indicating atrial enlargement) and may be a first source of user-specific heartbeat characteristics.

[0039] The QRS complex can correspond to the depolarization of the heart's ventricles and can be separated into three different waves - the Q wave, the R wave, and the S wave. Because the ventricles contain more muscle mass than the atria, the QRS complex is larger than the P wave. In addition, the His / Purkinje heart system, which can increase the conduction velocity to coordinate the depolarization of the ventricles, can make the QRS complex appear "pointed" rather than round. The duration of the QRS complex of a healthy heart can be in the range of 60 to 100 milliseconds (ms), but can be changed due to conduction abnormalities. The duration of the QRS complex can be used as another source of user-specific heartbeat characteristics.

[0040] The duration, amplitude, and morphology of each of the Q wave, R wave, and S wave may vary among different individuals, and in particular, may differ significantly for users with heart disease or heart irregularities. For example, a Q wave greater than 1 / 3 the height of an R wave or greater than 40 ms (milliseconds) in duration may indicate a myocardial infarction and provide a unique characteristic of a user's heart. Similarly, the ratio of otherwise healthy Q waves to R waves may be used to distinguish the heartbeats of different users.

[0041] The electrical activity of the heart of the user US may also include one or more characteristic durations or intervals that can be used to distinguish different users. For example, as known in the art, the electrical activity of the heart may include a PR interval and an ST segment. The PR interval may be measured from the beginning of the P wave to the beginning of the QRS complex. The PR interval may typically last for 120 to 200 ms. PR intervals with different durations may indicate one or more defects in the heart, such as a first-degree heart block (e.g., the PR interval lasts for more than 200 ms), preexcitation syndrome via a bypass that causes early activation of the ventricle (e.g., the PR interval lasts for less than 120 ms), or other types of heart block (e.g., the PR interval is variable), etc. The ST segment may be measured from the QRS complex to the T wave, for example, starting from the junction between the QRS complex and the ST segment, and ending at the beginning of the T wave. The ST segment may typically last for 80 to 120 ms, and typically has a slight upward depression. It is also possible to generate characteristic information unique to the heartbeat of each user using a combination of the length of the ST segment and the depression or elevation of the ST segment.

[0042] As discussed in further detail herein, a user may use the ECG monitor 103 to measure their ECG data for analysis. The ECG monitor 103 may also be connected to the user through, for example, a Wi-Fi connection, The measured ECG data and / or the results of the diagnostic analysis (as discussed in further detail herein) may be sent to the local computing device 101 through any suitable wired or wireless connection, such as a wireless connection, a near field communication (NFC) connection, an ultrasound signal transmission connection, or the like.

[0043] The user can continuously record ECG data at regular intervals. For example, the interval can be once a day, once a week, once a month, or some other predetermined interval. As described herein, ECG data can be recorded at the same or different times of the day, in similar or different situations. ECG data can also be recorded at the same or different times in the interval (for example, ECG data can be captured asynchronously). Alternatively or additionally, the user can record ECG data as needed at each discrete time, such as recording ECG data when the user feels chest pain or experiences other abnormal or abnormal sensations, or recording ECG data in response to instructions to do so from, for example, a user's physician, etc. In other embodiments, ECG data can be continuously recorded over a period of time.

[0044] Each ECG data record can be time-stamped and can be annotated with additional data by the user or health care provider to describe the user characteristics. For example, the local computing device 101 (e.g., its mobile application 101A) can include a user interface for data input that allows the user to enter its user characteristics. Examples of user characteristics can include age, gender, race, ethnicity, relevant medical history, location, diet (e.g., food / beverage habits), medication / drug consumption, exercise mode, sleep / rest mode, stress, anxiety, pain or other abnormal or abnormal feelings, activities performed before, during or after data recording, or any other user-specific situation or factor that may affect the user's ECG data. The local computing device 101 can supplement the user characteristics to the ECG data, and store the ECG data and user characteristics for further analysis (e.g., ECG interpretation). Alternatively, the local computing device 101 can send the ECG data (collectively referred to as user data) and user characteristics to the cloud storage system 113. Because the user data is time-stamped or tagged, the ECG data can be matched or associated with the activity or situation of interest. As described in further detail herein, this also allows comparison of ECG data before, after, and during activities or conditions of interest so that effects on the ECG data can be determined and taken into account during further analysis.

[0045] When the local computing device 101 sends user data to the cloud storage system 113 for storage and analysis, such sending may be in real time, at regular intervals such as hourly, daily, weekly, and / or any interval in between, or may be on demand. The local computing device 101 and the cloud storage system 113 may be coupled to each other (e.g., may be operatively coupled, communicatively coupled, may communicate data / messages with each other) via a network 140. The network 140 may be a public network (e.g., the Internet), a private network (e.g., a local area network (LAN) or a wide area network (WAN)), or a combination thereof. In one embodiment, the network 140 may include a wired or wireless infrastructure, which may be provided by one or more wireless communication systems (e.g., a Wi-Fi hotspot connected to the network 140 and / or a wireless carrier system that may be implemented using various data processing equipment, communication towers (e.g., cellular towers), etc.). The network 140 may carry communications (e.g., data, messages, packets, frames, etc.) between the local computing device 101 and the cloud storage system 113.

[0046] The user data accumulated over time for a particular user can form a time series health record for the particular user. The cloud storage system 113 may include any suitable type of computing device or machine with a programmable processor, including, for example, a server computer, a desktop computer, a laptop computer, a tablet computer, a smart phone, etc. In some embodiments, the cloud storage system 113 may include a single computing device, or may include multiple interconnected computing devices (e.g., multiple servers configured in the form of a cloud storage cluster).

[0047] Figure 3A hardware block diagram of a computing device 101 is shown. The computing device 101 may include hardware such as a processing device 115 (e.g., a processor, a central processing unit (CPU)), a memory 120 (e.g., a random access memory (RAM)), a storage device (e.g., a hard disk drive (HDD), a solid-state drive (SSD), etc.), and other hardware devices (e.g., a sound card, a video card, etc.). In some embodiments, the memory 120 may be a persistent storage unit capable of storing data. The persistent storage unit may be a local storage unit or a remote storage unit. The persistent storage unit may be a magnetic storage unit, an optical storage unit, a solid-state storage unit, an electronic storage unit (main memory), or a similar storage unit. The persistent storage unit may also be a monolithic / single device or a distributed collection of devices. The memory 120 may be configured for long-term storage of data and may retain data between power-on / power-off cycles of the computing device 101. The memory 120 may store user data accumulated over time for a user and a large number of other users. The memory 120 may also include a reduced-lead ECG analysis model 120A, which is configured to analyze the user's reduced-lead ECG data measured each time an ECG is taken to determine whether the user is experiencing a cardiac condition, as discussed in further detail herein.

[0048] As discussed herein, mobile ECG monitors attempt to simplify the process of obtaining an ECG by using reduced lead sets (e.g., any number of leads less than 12). However, for the convenience of reduced lead sets, such implementations must inherently sacrifice resolution or discrimination in detecting a full range of cardiac conditions. Prior work has attempted to mathematically construct all 12 leads based on the reduced lead set through linear or nonlinear transformations to provide a set of all 12 leads (some leads are now "synthesized leads"), and the synthesized full 12-lead data can be analyzed by a physician or through a full 12-lead algorithm. Based on this prior work of reconstructing the full 12-lead set, the natural workflow of the ML model implemented by the mobile ECG device will also involve multiple layers, including, for example: 1) creating a 12-lead set from a reduced lead set (say, 4 leads) using a lead reconstruction algorithm, and 2) using a condition classification algorithm to take the reconstructed 12 leads as input and provide a determination (classification) of normal versus various conditions as output. However, since ECG guidelines developed using standard 12-lead ECGs are applied for abnormality detection and interpretation, the application of reduced lead 12-lead sets and synthetic leads may be limited. The ACC / AHA has not developed or recommended specific guidelines for reduced lead sets. This leads to a gap in the accuracy of such synthetic leads, especially in localized STEMI situations.

[0049] Furthermore, solving the problem of providing accurate readings of reduced lead ECGs requires an ML model architecture consisting of different layers (e.g., preprocessing, feature extraction, optimization, prediction, decision making) including lead reconstruction and pathology classification, where each layer processes the input data in a pipelined manner. However, due to the complexity of the model, this multi-layer approach may provide less than ideal results. For example, this multi-layer architecture requires that each layer be optimized separately under different criteria.

[0050] A deep neural network (DNN) is a type of model consisting of multiple layers of interconnected nodes, each layer building on the previous layer to refine and optimize predictions or classifications. This computational progression through the network is called forward propagation. The input and output layers of a deep neural network are called visible layers. The input layer is where the deep learning model ingests data for processing, and the output layer is where the final prediction or classification is made. For example, there may be a collection of photos of different pets, and you may want to classify them by "cat", "dog", "hamster", etc. A deep learning algorithm can determine which features (e.g., ears) are most important for distinguishing each animal from other animals. In machine learning, this feature hierarchy is manually established by human experts. Then, through the process of gradient descent and back propagation, the DNN self-adjusts and adapts itself to accuracy, allowing it to predict new photos of animals with improved accuracy. Back propagation involves fine-tuning the weights of the neural network (i.e., iteration) based on the error rate (i.e., loss) obtained in the previous period (epoch). Proper adjustment of the weights ensures a lower error rate, thereby making the model reliable by improving the generalization ability of the model.

[0051] DNNs can also be trained using end-to-end learning, which uses the advantages of a DNN structure consisting of several layers to solve complex problems. For ECG processing, the input is an ECG waveform, and the output is a classification result. Each DNN layer (or layer group) can be specialized in the intermediate tasks required for such a problem. For example, the convolutional layer (CNN) of a DNN model can automatically extract ECG features. However, these features are mainly distributed in the vectors of the intermediate output layer at the end of the CNN block, which is called a flat layer. The elements in the flat layer are also called embedded variables, which can be used for classification of rhythm and morphological interpretation or for regression to predict ECG measurements such as PR intervals, QRS duration, and QT intervals. End-to-end learning refers to training a possible complex learning system represented by a single model (specifically, DNN) representing a complete target system, thereby bypassing the intermediate layers that are usually present in traditional pipeline designs. In other words, when two or more ML models are being used as components of a larger architecture, DNNs can be trained in an end-to-end manner involving training all components simultaneously (i.e., the DNN is trained as a single network).

[0052] Embodiments of the present disclosure utilize DNN for ECG interpretation, where the raw ECG waveform is directly ingested by the DNN using paired interpretation tags for training without the need for prescriptive feature extraction or rule-based criteria. As discussed in further detail herein, embodiments of the present disclosure differ from conventional expert system-based criteria because the DNN model plays an important role in the entire ECG classification and ECG interpretation process.

[0053] Figure 4 A reduced lead ECG analysis model 120A (hereinafter referred to as the interpretation model 120A) is shown. The interpretation model 120A includes a data conditioning stage 130, a model analysis stage 140, and a final determination stage 150. The data conditioning stage 130 may include filtering and signal processing functions, as well as a downsampling module 131, a lead conversion module 132, and a median beat generation module 133. For example, the interpretation model 120A may receive as input (e.g., from an ECG monitor 103) a first 4-lead ECG signal including leads [I, II, v2, v4] and a second 4-lead ECG signal including leads [I, II, v1, v4], where the signal from each lead is 10 seconds. Although shown as using a 10-second sample of leads [I, II, v1 / v2, v4], this is for illustrative purposes and not limiting. In practice, any suitable input lead group may be used, and the signal of each lead may have any suitable duration. The downsampling module 131 may downsample each input 4-lead ECG signal to 150 Hz, and the lead conversion module 132 may determine the other 4 limb leads [III, aVR, aVL, aVF] using a conventional lead reconstruction algorithm. The lead conversion module 132 may add the leads reconstructed by the algorithm to each 4-lead ECG signal, thereby obtaining two 8-lead ECG signals. The first 8-lead ECG signal may include leads [I, II, III, v2, v4, aVR, aVL, aVF], and the second 8-lead ECG signal may include leads [I, II, III, v1, v4, aVR, aVL, aVF].

[0054] Training ML models to classify ECG signals requires determining the basic characteristics of ECG signals. The model analysis stage 140 may include an ML model for identifying the basic characteristics of the input ECG signal including ECG rhythm patterns and ECG waveform morphology / shape patterns. Each ML model in the model analysis stage 140 may be any appropriate deep neural network (DNN) model such as a convolutional neural network (CNN) or any other appropriate DNN. Determining the ECG rhythm pattern requires capturing the sequential characteristics of the ECG rhythm pattern, which in turn requires effective detection of these characteristics. Therefore, the model analysis stage 140 may include an ECG rhythm classification model 141 for each 8-lead ECG signal. In other words, each 8-lead ECG signal can be used as an input to the corresponding ECG rhythm classification model 141. A sufficient time window for the input data (in this case, a 10-second window for two 8-lead ECG signals) should be used to allow the sequential characteristics of the rhythm pattern to be detected without further feature detection.

[0055] In addition to rhythm patterns, waveform morphology / shape pattern is another main ECG characteristic, and is usually detected using ECG amplitude and duration / interval. For waveform morphology, different types of representative beats of data sequences (e.g., median beats, average beats of multiple ECG cycles) can simplify waveform morphology detection models and accelerate training, while also improving classification accuracy. Therefore, the median beat generation module 133 can determine the median beats of the first 8-lead ECG signal and the second 8-lead ECG signal, thereby obtaining the first median beat 8-lead signal and the second median beat 8-lead signal respectively. The model analysis stage 140 can include an ECG morphology model 142 for each median beat 8-lead signal. In other words, each median beat 8-lead signal can be used as an input to the corresponding ECG morphology ML model 142. The model analysis stage 140 can also include a global ECG measurement model 143 for detecting global ECG parameters including, for example, PR intervals, QRS durations, and QT intervals. An embodiment of the present disclosure is described only for the first 8-lead ECG signal and the corresponding ECG rhythm classification model 141 and ECG morphology ML model 142, however this is for ease of description, and each 8-lead ECG signal can have its own DNN model, which can form a corresponding classification as further described in detail herein.

[0056] The final determination stage 150 may include final ECG rhythm logic 151 and final ECG morphology logic 152, each of which may include some simple logic to verify the classification by the corresponding DNN. The final analysis combines all DNN outputs and ECG measurements such as PR interval, QRSD interval, QT interval, ST level to refine the classification. The output of the final determination stage 150 may be a probability predicting the likelihood of various (usually indistinguishable) heart disease conditions (or an indication of normal heart health of the user).

[0057] Figure 5A The general structure 500 of each model 141, 142 and 143 used in the model analysis stage 140 is shown. Each model may include 4 main components. The first component may be a lead formation layer 505, which reformats the leads of an 8-lead ECG signal (or a median beat 8-lead signal) into the format required for the training model. The second component may be a set of convolutional layers 510, which may include multiple (e.g., 5 to 8) CNN blocks, each with a different number of filters. The third component may be a set of dense layers 515, which reshape the multi-filter output of the last convolutional layer into a flat layer and pass the result to the next dense layer with full connection, where full connection means that each unit of the previous layer is connected to all units of the next layer. The fourth component may be a classification layer 520 for generating each determined probability.

[0058] Figure 5BAn example of a DNN 400 that can be used in conjunction with some embodiments of the present disclosure is shown. The DNN 400 may include a trained CNN model that takes input data 402 (e.g., user data) into a convolutional layer (also referred to as a hidden layer) 403, where a series of trained weights or filters 404 are applied to the input data 406 in each convolutional layer 403. The output of the first convolutional layer is an activation map (not shown), which is the input to the second convolutional layer to which the trained weights or filters (not shown) are applied, where the output of subsequent convolutional layers results in activation maps that represent increasingly complex features of the input data of the first layer. After each convolutional layer, a nonlinear layer (not shown) is applied to introduce nonlinearity into the problem, where the nonlinear layer may include an activation function such as tanh, sigmoid, or ReLU. In some cases, a pooling layer (not shown) (also referred to as a downsampling layer) may be applied after the nonlinear layer, which takes filters of the same length and stride and applies them to the input, and outputs the maximum number of filters in each sub-region where the convolution operation is performed. Other options for pooling are average pooling and L2 norm pooling. Pooling layers reduce the spatial dimensionality of the input volume, thereby reducing computational cost and controlling overfitting. The last layer (one or more) of the network is a fully connected layer that takes the output of the last convolutional layer and outputs an n-dimensional output vector representing the quantity to be predicted (e.g., the probability of LVH diagnosis and (if applicable) the severity of LVH, etc.). This results in a predicted output 406 (O*), for example, that the user may be suffering from stage 2 LVH. As will be described more fully below, the trained weights 404 can be different for each convolutional layer 403.

[0059] Figure 5C The process of training DNN 400 is shown (for simplicity, dense layers and classification layers are not shown). Figure 5C , the convolutional layer 403 is shown as separate convolutional layers 405, 405' to the convolutional layer (405) n-1 , and the last nth layer is a fully connected layer. It should be understood that the last layer can be more than one fully connected layer. Training examples 411 are input to convolutional layer 403, non-linear activation functions (not shown) and weights 410, 410' to 410 n Successively applied to training examples 411, where the output of any convolutional layer is fed into the next layer, and so on, until the final nth fully connected layer (405) n-1 Output 414 is generated. Output or prediction 414 (*) is compared with training example 411 (e.g., wide QRS tachycardia) to obtain output or prediction 414 and training example 411 (in Figure 5C Also shown as I known) between the output or prediction 414 and the output or prediction 414. If the difference or loss 416 is less than some preset loss (e.g., the output or prediction 414 predicts with the necessary level of accuracy that the user is suffering from a wide QRS tachycardia), then the DNN model converges and is considered trained.

[0060] If the DNN 400 has not converged, back-propagation techniques are used to update the weights 410 and 410 based on how close the predictions are to the known inputs. n-1 to 410 n Those skilled in the art will appreciate that methods other than back-propagation may be used to adjust the weights. A second training example (e.g., different user data) may be input, and the process repeated again with the updated weights, and the weights updated again, and so on, until the nth training example (e.g., nth user data) has been input. The process is repeated over and over again with the same n training examples until the DNN model is trained or converges on the correct output for the known input. Once the DNN model is trained, the weights 410 to 410 are adjusted. n is fixed and used in the trained ECG rhythm classification model 141. As discussed herein, there are different weights for each convolutional layer 403 and each fully connected layer. Then, as described in further detail herein, the trained ECG rhythm classification model 141 is fed with user data to determine or predict what it is trained to predict / recognize (e.g., rhythm classification).

[0061] Figure 6 The detailed structure of the ECG rhythm classification model 141 is shown. The convolutional layer 610 can utilize a residual block structure that utilizes a combination of convolutional layers and skip layers. Skip layers (also known as skip connections or shortcut connections) can enable residual blocks in deep learning models to solve the problems of gradient vanishing and gradient exploding that can occur when training very deep neural networks. This is done by adding the input of a layer directly to the output of one or more subsequent layers, thereby creating a "shortcut" for the gradient to flow through. As a result, different parts of the convolutional layer 610 will be trained at different rates for different training data points based on how the error flows backward through the layers of the convolutional layer 610. Figure 6 A residual block 610A is shown in which a skip layer is used to skip the training of convolutional layer 2 during the learning of the identity function X. The ReLU function may represent a non-linear activation function.

[0062] In addition to the residual block, the ECG rhythm classification model 141 may also include 2 dense layers 615 and 1 classification layer 620. In some embodiments, the ECG rhythm classification model 141 may also include a batch normalization layer and a dropout layer for each residual block and dense layer to improve the robustness or generalization ability of the ECG rhythm classification model 141 once it is trained ( Figure 6 Each batch normalization layer and dropout layer are not shown). The ECG rhythm classification model 141 is designed to classify rhythm determinations, examples of which are shown in Table 1 (below). The classification layer 620 output may have both positive and negative probabilities for each determination. In some embodiments, for each 8-lead ECG signal (downsampled to 150 Hz), 9.5 seconds of data out of 10 seconds are used for input, and the remaining 0.5 seconds of data may be used for enhancement by a time shifting technique (random data shift). This random data shift is one of the data enhancement methods used to improve the generalization ability of the trained DNN model, which will result in the DNN model being less sensitive to the start and end times of the input ECG rhythm data.

[0063] Table 1

[0064]

[0065]

[0066] Any sufficiently large training data set may be used to train the ECG rhythm classification model 141. For example, the training data set may include one million labeled ECGs and another 250,000 labeled ECGs used as a separate test set. Figure 5C In the example, the batch size is 512, the learning rate is 0.001, and the dropout rate is 0.05. A total of 20 epochs are set. The final model is selected from the one with the minimum validation error. The training weight matrix is ​​used to balance large differences in popularity among all determinations. For determinations with lower popularity, higher weights are used when backpropagating the training error.

[0067] Fig. 7AThe detailed structure of the ECG morphology model 142 is shown. The ECG morphology model 142 can capture the morphological changes of each lead and the relative changes from lead to lead (referred to as changes in the spatial domain). The ECG morphology model 142 can have a model structure similar to the ECG rhythm classification model 141, but has a smaller lead formation layer and fewer residual block layers, because the median beat 8-lead signal has a much shorter length than the 10-second duration of the 8-lead ECG signal input to the ECG rhythm classification model 141. In addition, all median beats are relatively aligned around their QRS onset. Another important design consideration for the morphology model QRS region is that the changes along the time axis are relatively smaller than the rhythm signal.

[0068] To capture spatial domain variations, the ECG morphology model 142 can form ECG imaging by laying out all leads in the form of a 2D matrix / image pattern with the X-axis as time and the Y-axis as the lead order of [aVL, I, aVR, II, aVF, III, v2, v4], where the limb leads are reordered in the so-called Cabrera sequence. The ECG morphology model 142 can process the 2D ECG signal matrix using 2D convolution filters as in the image classification case. This approach is more efficient than applying multiple 1D convolution filters to each lead separately, which has been used by most other ECG training models. Since the ECG is a projection of the electrical depolarization and repolarization process of the heart to the trunk, the spatial 2D signal matrix can capture the projection more effectively. Therefore, a 2D convolution pattern is applied to the ECG image input.

[0069] As discussed herein, the median beat generation module 133 can generate a median beat 8-lead signal based on an 8-lead ECG signal. More specifically, the median beat generation module 133 can form a QRS detection function, perform beat detection and classification, perform beat alignment, and calculate the median beat and average beat according to the dominant beat type. Figure 7B An example 8-lead ECG signal is shown, and Figure 7C The median beat of an example 8-lead ECG signal is shown. It can be seen that the ECG morphology features are captured in the median beat while the data length is much shorter.

[0070] The ECG morphology model 142 may be trained using the same training data set as the ECG rhythm classification model 141 (which may include median beat data for each training ECG). Figure 5CIn the example of , the batch size is 512, the learning rate is 0.001, and the drop rate is 0.05. A total of 20 epochs are set. The final model is selected from the minimum validation error. The morphology training weight matrix is ​​also used to balance large differences in popularity among all determinations. For determinations with lower popularity, higher weights are used when backpropagating the training error. The ECG morphology model 142 is designed to classify ECG morphology determinations, an example of which is shown in Table 2 (below). The classification layer of the ECG morphology model 142 has both positive and negative probabilities for each determination.

[0071] Table 2

[0072]

[0073]

[0074] The global ECG measurement model 143 can measure ECG global parameters including PR interval, QRS duration, QT interval and heart rate. Since these parameters are not lead-specific and are usually estimated based on a combination of multiple leads, these parameters are referred to as ECG global measurements. Typically, these parameters are detected using complex signal processing and feature detection algorithms based on domain knowledge. However, embodiments of the present disclosure use a deep learning model (global ECG measurement model 143) applied to the median beat 8-lead signal generated from the 8-lead ECG signal to estimate these ECG parameters. Since ECG global parameters are closely related to ECG morphology, the global ECG measurement model 143 can be trained to perform regression of the median beat 8-lead ECG signal to detect these parameters. Another purpose of training the global ECG measurement model 143 is to also be able to identify the absence of a valid P wave such as in atrial fibrillation situations.

[0075] Figure 8The detailed structure of the global ECG measurement model 143 is shown. The structure of the global ECG measurement model 143 is similar to that of the ECG morphology model 142, except that the last classification layer of the global ECG measurement model 143 is replaced by the regression layer 820. In addition, the global ECG measurement model 143 does not have the batch normalization layer of the ECG morphology model 142, because the batch normalization layer is more useful for classification than for regression. The global ECG measurement model 143 can receive a median beat 4-lead signal based on the lead group [I, II, v2, v4] as input. More specifically, the median beat generation module 133 can determine the median beat of each lead in the lead group [I, II, v2, v4] as discussed above and provide the median beat 4-lead signal to the global ECG measurement model 143. The global ECG measurement model 143 is trained by minimizing the absolute error between the labeled (physician-estimated) intervals from the training data and the model-estimated intervals to output three intervals: PR interval, QRS duration, and QT interval.

[0076] like Figure 4 As shown, global ECG measurement model 143 can also estimate global QRS start / offset.Global ECG measurement model 143 can use global QRS start / offset and 12 lead groups derived by algorithm to determine ECG measurement lead by lead.In standard resting 12 lead analysis, in addition to ECG global parameters described herein, ECG measurement lead by lead is also important.Measurement lead by lead can include amplitude, duration and polarity of all main ECG morphological components including P wave, QRS complex, ST wave and T wave.Due to mainly being determined by ECG morphological model 142, final ECG morphological logic 152 can use ECG measurement lead by lead to assist final classification, which reduces the dependence on having accurate ECG measurement to a certain extent.In fact, ECG measurement lead by lead can help prevent some obvious classification errors, in case ECG rhythm classification model 141, ECG morphological model 142 and global ECG measurement model 143 cannot capture these errors or produce multiple classifications that need to select one from them. For example, during morphology classification, if the ECG morphology model 142 classifies both LBBB and acute MI, the final ECG morphology logic 152 can make a final classification based on QRS duration, ST segment, and in some cases, T wave. In other examples, during ECG rhythm analysis, false positive classifications of AFIB can be avoided by examining P wave measurements including PR interval and P wave amplitude.

[0077] The final determination stage 150 may include logic for making the correct exclusion / inclusion based on cardiology science and clinical practice. The final determination stage 150 may exclude under various conditions. Since both the ECG rhythm classification model 141 and the ECG morphology model 142 are trained for non-exclusive multi-class classification, multiple abnormalities can be identified / classified. In some cases, multiple abnormal conditions can be identified for the same ECG, such as ischemia and infarction, or bundle branch block and ischemia, etc. However, in other cases, especially during ECG rhythm analysis, certain abnormalities may affect the detection of other abnormalities, or these abnormalities may be mutually exclusive, such as sinus rhythm and AFIB, or pacing rhythm and sinus, etc. The final determination stage 150 may operate in one of the following two different modes: high specificity mode and high sensitivity mode. The high specificity mode may include a higher threshold for the model output probability and more conditions based on lead-by-lead measurement. The high sensitivity mode may include a lower threshold (e.g., 70%) for the model output probability. In some embodiments, in high sensitivity mode, if the probability is very high (eg, greater than 95%), the model may immediately determine that the determined classification is the final classification.

[0078] As discussed herein, the final determination stage 150 may include a final ECG rhythm logic 151 and a final ECG morphology logic 152. The final ECG rhythm logic 151 may include logic to find the dominant rhythm from the following list: paced rhythm, sinus / normal sinus, atrial fibrillation, atrial flutter, wide QRS tachycardia, junctional rhythm, and undetermined rhythm. If the dominant rhythm is sinus rhythm, the final ECG rhythm logic 151 may attempt to find an associated sinus arrhythmia. The input parameters of the final ECG rhythm logic 151 are as follows: the output of the ECG rhythm classification model 141, global ECG measurements (PR interval, QRS duration, QT interval, and heart rate), and the RR interval sequence of the dominant beat. All determinations of the final ECG rhythm logic 151 are performed in 2 modes as described above: high specificity and high sensitivity.

[0079] Fig. 9 The final ECG rhythm logic 151 is shown as a set of various conditions for finding a dominant rhythm. If the output of the ECG rhythm classification model 141 indicates that the probability of a paced rhythm is high (e.g., greater than 0.8) and the probability of a paced rhythm is greater than the probability of any other rhythm, the final ECG rhythm logic 151 can determine the paced rhythm as the dominant rhythm. In addition, the heart rate must be less than 80 bpm, the QRS duration must be less than 150 ms, and the RR variability must be small. If a paced rhythm is confirmed, the final ECG rhythm logic 151 may not perform any other checks on other rhythms.

[0080] If the output of the ECG rhythm classification model 141 indicates that the probability of sinus rhythm is high (e.g., greater than 0.7) or the probability of pacing rhythm is less than the probability of other rhythms, the final ECG rhythm logic 151 can determine sinus rhythm as the dominant rhythm (i.e., confirm the sinus rhythm classification from the ECG rhythm classification model 141). In addition, the PR interval must be greater than 80ms or the probability of sinus rhythm must be very high (e.g., greater than 0.95). Here, the appropriate range of PR intervals indicates the presence of a valid P wave, which is a key feature of sinus rhythm. This final analysis is to strengthen the results of the DNN model. It should be noted that the specificity and sensitivity settings for sinus rhythm are opposite to other abnormal rhythms, because the high specificity settings for other abnormal rhythms also mean a higher sensitivity setting for sinus rhythm. In order to emphasize this high specificity, sinus rhythm is confirmed before other abnormal rhythms other than pacing rhythm. If the ECG rhythm classification model 141 outputs a very high probability of sinus / normal sinus, it is confirmed without checking other rhythms.

[0081] If the output of the ECG rhythm classification model 141 indicates that the probability of atrial fibrillation / atrial flutter is higher than other dominant rhythms (e.g., greater than 0.7) and the probability of the paced rhythm is less than other rhythms, the final ECG rhythm logic 151 can determine atrial fibrillation / atrial flutter as the dominant rhythm. In addition, the PR interval must be less than 30ms and there must be no valid P wave. If the probability of atrial fibrillation is less than the probability of atrial flutter, the final ECG rhythm logic 151 can determine that atrial flutter is the dominant rhythm.

[0082] If the final ECG rhythm logic 151 excludes all other rhythms as the dominant rhythm, it can check whether the junction rhythm is the dominant rhythm based on whether there is a high probability of the junction rhythm from the ECG rhythm classification model 141 and a low RR variation. If these two conditions are met, the final ECG rhythm logic 151 can determine that the junction rhythm is the dominant rhythm.

[0083] If the final ECG rhythm logic 151 still does not determine the dominant rhythm after checking the junction rhythm, a more relaxed set of conditions can be used to check whether sinus rhythm is present. More specifically, the final ECG rhythm logic 151 can lower the threshold of the probability of sinus rhythm to 0.6 and require that a valid P wave is present and that all global measurements are within their normal ranges (e.g., PR [120, 200], QRS duration <115, QTcF <460ms).

[0084] In some embodiments, if the final ECG rhythm logic 151 confirms sinus rhythm, an attempt may be made to determine that one or more associated arrhythmias have been examined. More specifically, if there is a high probability of first-degree atrioventricular block (AVB) and a PR interval greater than 220ms, the final ECG rhythm logic 151 may determine that there is a first-degree AVB. If there is a high probability of a second-degree AVB, the probability of the second-degree AVB is greater than the probability of the third-degree AVB, and the probability of the second-degree AVB is greater than the probability of other significant arrhythmias, the final ECG rhythm logic 151 may determine that there is a second-degree AVB. If there is a high probability of a third-degree AVB and the probability of the third-degree AVB is greater than the probability of other arrhythmias, the final ECG rhythm logic 151 may determine that there is a third-degree AVB. If there is a high probability of sinus bradycardia and the heart rate is less than 50bpm, the final ECG rhythm logic 151 may determine that there is sinus bradycardia. If there is a high probability of sinus tachycardia and the heart rate is greater than 100 bpm, the final ECG rhythm logic 151 can determine that sinus tachycardia is present. If there is a high probability (e.g., greater than 0.8) of significant (severe) sinus arrhythmia, the final ECG rhythm logic 151 can determine that significant sinus arrhythmia is present.

[0085] The final ECG morphology logic 152 may include logic to strengthen the results of the morphology DNN model and improve the specificity of the determination of the main morphology abnormality. The ECG morphology model 142 also works in a non-exclusive mode, which means that multiple abnormal conditions can be classified for the same ECG. Although multiple conditions may occur physiologically, it may be helpful to focus on the main source of the abnormal condition rather than multiple sources. For example, in the interpretation of acute myocardial infarction (acute MI), adding ST segment changes on top of the DNN model probability is a way to improve specificity. However, if bundle branch block is confirmed, especially left bundle branch block (LBBB), the ST segment deflection may be secondary to depolarization changes, which may be difficult to distinguish from primary repolarization. Therefore, if LBBB or pacing ECG is confirmed, the final determination of acute MI is uncertain.

[0086] The final ECG morphology logic 152 may make a final morphology determination based on one or more of the classifications from the ECG rhythm classification model 141 and the ECG morphology model 142 in the following order: paced ECG, RBBB, LBBB, other intraventricular block, LVH, RVH, old MI, acute MI, other ST elevation, ischemia, long QT, atrial enlargement, normal ECG. The conditions discussed below are common to both high sensitivity and high specificity modes, with only the thresholds being relatively different.

[0087] If both the ECG rhythm classification model 141 and the ECG morphology model 142 indicate that the probability of a paced ECG is very high (e.g., greater than 0.95) and the RR beat variation is small (RR_std / RR_mean is less than 0.04), then the final ECG morphology logic 152 may determine a paced ECG (i.e., confirm the paced ECG classification from the ECG morphology model 142). If the final ECG morphology logic 152 confirms the paced ECG classification, then checking for any other morphologies may be skipped.

[0088] If the probability of RBBB indicated by the ECG morphology model 142 classification is high (e.g., greater than 0.8) and the QRS duration is greater than 120 ms, or if the probability of RBBB indicated by the ECG morphology model 142 classification is greater than 0.95 and the QRS duration is greater than 110 ms, the final ECG morphology logic 152 may determine RBBB (i.e., confirm the RBBB classification from the ECG morphology model 142). If the final ECG morphology logic 152 confirms the RBBB classification, then checking for any other morphologies may be skipped.

[0089] If the probability of LBBB indicated by the ECG morphology model 142 classification is high (e.g., greater than 0.8) and the QRS duration is greater than 125 ms, or if the probability of LBBB indicated by the ECG morphology model 142 classification is greater than 0.95 and the QRS duration is greater than 115 ms, the final ECG morphology logic 152 may determine LBBB (i.e., confirm the LBBB classification from the ECG morphology model 142). If the final ECG morphology logic 152 confirms the LBBB classification, then checking for any other morphologies may be skipped.

[0090] If the probability of any other BBB indicated by the ECG morphology model 142 classification is high (e.g., greater than 0.8) and the QRS duration is greater than 115 ms, or if the probability of any other BBB indicated by the ECG morphology model 142 classification is greater than 0.90 and the QRS duration is greater than 110 ms, the final ECG morphology logic 152 can determine such a BBB (i.e., confirm the classification of such a BBB from the ECG morphology model 142).

[0091] If the probability of LVH indicated by the ECG morphology model 142 classification is very high (e.g., greater than 0.99), or if the probability of LVH indicated by the ECG morphology model 142 classification is high (e.g., greater than 0.9) and the QRS deflection of one of the V leads is greater than 2500uV (microvolts) or R_aVL is greater than 1000uV, then the final ECG morphology logic 152 can determine the LVH morphology (i.e., confirm the LVH classification from the ECG morphology model 142). It should be noted that LVH determination usually requires some ECG amplitude measurements. However, since only 2 precordial leads are used and the ECG is more aggressively low-pass filtered, conventional amplitude criteria for LVH are not directly applied. Instead, the ECG morphology model 142 output is used in conjunction with a set of minimum amplitude and axis measurement checks. If the probability of LVH indicated by the ECG morphology model 142 classification is very high (eg, greater than 0.9) and the QRS axis is greater than 90 degrees, the final ECG morphology logic 152 may determine RVH morphology (ie, confirm the RVH classification from the ECG morphology model 142).

[0092] If the probability of an acute anterior MI indicated by the ECG morphology model 142 classification is very high (e.g., greater than 0.95), the ST of at least one V lead is greater than 200uV and the ST / T ratio is greater than 0.3, and the ST of at least one complementary lead (limb lead) is less than -50uV (ST depression), then the final ECG morphology logic 152 may determine an acute anterior MI morphology (i.e., confirm the acute anterior MI classification from the ECG morphology model 142). Alternatively, if the probability of an acute anterior MI is very high (e.g., greater than 0.99) and the ST of at least one V lead is greater than 160uV and the ST / T ratio is greater than 0.3, then the final ECG morphology logic 152 may determine an acute anterior MI morphology.

[0093] If the probability of an acute inferior MI as indicated by the ECG morphology model 142 classification is very high (e.g., greater than 0.95), and the ST of at least one V lead is greater than 100 uV and the ST / T ratio is greater than 0.3, then the final ECG morphology logic 152 may determine an acute inferior MI morphology (i.e., confirm the acute inferior MI classification from the ECG morphology model 142). Alternatively, if the probability of an acute anterior MI is very high (e.g., greater than 0.99), and the ST of at least one inferior lead is greater than 60 uV and the ST / T ratio is greater than 0.3, then the final ECG morphology logic 152 may determine an acute inferior MI morphology.

[0094] If the probability of an acute lateral MI indicated by the ECG morphology model 142 classification is very high (e.g., greater than 0.95), and the ST of at least one lateral lead is greater than 100 uV and the ST / T ratio is greater than 0.3, then the final ECG morphology logic 152 may determine an acute lateral MI morphology (i.e., confirm the acute lateral MI classification from the ECG morphology model 142). Alternatively, if the probability of an acute anterior MI is very high (e.g., greater than 0.99), and the ST of at least one lateral lead is greater than 60 uV and the ST / T ratio is greater than 0.3, then the final ECG morphology logic 152 may determine an acute lateral MI morphology.

[0095] If the probability of old MI indicated by the ECG morphology model 142 classification is very high (e.g., greater than 0.90) and there are significant Q waves, the final ECG morphology logic 152 can determine a previous myocardial infarction (old MI) morphology (i.e., confirm the old MI classification from the ECG morphology model 142).

[0096] If the probability of ischemia indicated by the ECG morphology model 142 classification is very high (e.g., greater than 0.90), and there is significant ST depression corresponding to the specified location calls in the following lead groups: anterior wall leads: ST in V2, V4; inferior wall leads: II, III, aVF; lateral wall leads: I, aVL, then the final ECG morphology logic 152 can determine the ischemic morphology (i.e., confirm the ischemic classification from the ECG morphology model 142).

[0097] If the probability of QT prolongation indicated by the ECG morphology model 142 classification is very high (e.g., greater than 0.90), and the QTcF (QT interval corrected by the Fridericia formula) is greater than 460 ms, then the final ECG morphology logic 152 may determine a QT prolongation morphology (i.e., confirm the QT prolongation classification from the ECG morphology model 142). The only exclusionary condition for QT prolongation is a paced ECG.

[0098] If the ECG morphology model 142 does not detect any of the abnormal morphology conditions and rhythm conditions listed above, sinus rhythm is detected, the heart rate is above 50bpm and below 90bpm, the QRS duration is not longer than 120ms, the PR interval is less than 230ms, the QTcF is less than 460ms, and the P, QRS and T axes are all greater than 0, then the final ECG morphology logic 152 can determine a normal morphology (i.e., confirm the normal classification from the ECG morphology model 142).

[0099] Fig.101 is a flow chart of a method 1000 for training and implementing an ML model for reducing lead group ECG interpretation according to some embodiments of the present disclosure. The method 1000 may be performed by processing logic, which may include hardware (e.g., circuitry, dedicated logic, programmable logic, processor, processing device, central processing unit (CPU), system on chip (SoC), etc.), software (e.g., instructions running / executing on a processing device), firmware (e.g., microcode), or a combination thereof. In some embodiments, the method 1000 may be performed by a computing device (e.g., Figure 3 The method 1000 is performed using the computing device 101 shown.

[0100] At block 1005, any sufficiently large training data set may be used to train the ECG rhythm classification model 141. For example, the training data set may include one million labeled ECGs and an additional 250,000 labeled ECGs used as an independent test set. Figure 5C In the example, the batch size is 512, the learning rate is 0.001, and the dropout rate is 0.05. A total of 20 epochs are set. The final model is selected based on the minimum validation error. The training weight matrix is ​​used to balance large differences in popularity among all determinations. For determinations with lower popularity, higher weights are used when backpropagating the training error.

[0101] Also refer to Figure 6 , the convolutional layer 610 can utilize a residual block structure that utilizes a combination of convolutional layers and skip layers. Skip layers (also known as skip connections or shortcut connections) can enable residual blocks in deep learning models to address the problems of vanishing and exploding gradients that can occur when training very deep neural networks. This is done by adding the input of a layer directly to the output of one or more subsequent layers, thereby creating a "shortcut" for the gradient to flow through. As a result, different parts of the convolutional layer 610 will be trained at different rates for different training data points based on how the error flows backward through the layers of the convolutional layer 610. Figure 6 A residual block 610A is shown in which a skip layer is used to skip the training of convolutional layer 2 during the learning of the identity function X. The ReLU function may represent a non-linear activation function.

[0102] In addition to the residual block, the ECG rhythm classification model 141 may also include 2 dense layers 615 and 1 classification layer 620. In some embodiments, the ECG rhythm classification model 141 may also include a batch normalization layer and a dropout layer for each residual block and dense layer to improve the robustness or generalization ability of the ECG rhythm classification model 141 once it is trained ( Figure 6Each batch normalization layer and dropout layer are not shown). The ECG rhythm classification model 141 is designed to classify rhythm determinations, examples of which are shown in Table 1 (below). The classification layer 620 output may have both positive and negative probabilities for each determination. In some embodiments, for each 8-lead ECG signal (downsampled to 150 Hz), 9.5 seconds of data out of 10 seconds are used for input, and the remaining 0.5 seconds of data may be used for enhancement by a time shifting technique (random data shift). This random data shift is one of the data enhancement methods used to improve the generalization ability of the trained DNN model, which will result in the DNN model being less sensitive to the start and end times of the input ECG rhythm data.

[0103] At block 1010, the ECG morphology model 142 may be trained using the same training data set as the ECG rhythm classification model 141 (which may include median beat data for each training ECG). Figure 5C In the example of , the batch size is 512, the learning rate is 0.001, and the drop rate is 0.05. A total of 20 epochs are set. The final model is selected based on the minimum validation error. The morphology training weight matrix is ​​also used to balance large differences in popularity among all determinations. For determinations with lower popularity, higher weights are used when backpropagating the training error. The ECG morphology model 142 is designed to classify ECG morphology determinations, an example of which is shown in Table 2 (below). The classification layer of the ECG morphology model 142 has both positive and negative probabilities for each determination.

[0104] Also refer to Fig. 7A , the ECG morphology model 142 can capture the morphological changes of each lead and the relative changes from lead to lead (referred to as changes in the spatial domain). The ECG morphology model 142 can have a model structure similar to the ECG rhythm classification model 141, but with a smaller lead formation layer and fewer residual block layers, because the median beat 8-lead signal has a much shorter length than the 10-second duration of the 8-lead ECG signal input to the ECG rhythm classification model 141. In addition, all median beats are relatively aligned around their QRS starts. Another important design consideration for the morphology model QRS region is that the changes along the time axis are relatively smaller than the rhythm signal.

[0105] To capture spatial domain variations, the ECG morphology model 142 can form ECG imaging by laying out all leads in the form of a 2D matrix / image pattern with the X-axis as time and the Y-axis as the lead order of [aVL, I, aVR, II, aVF, III, v2, v4], where the limb leads are reordered in the so-called Cabrera sequence. The ECG morphology model 142 can process the 2D ECG signal matrix using 2D convolution filters as in the image classification case. This approach is more efficient than applying multiple 1D convolution filters to each lead separately, which has been used by most other ECG training models. Since the ECG is a projection of the electrical depolarization and repolarization process of the heart to the trunk, the spatial 2D signal matrix can capture the projection more effectively. Therefore, a 2D convolution pattern is applied to the ECG image input.

[0106] At box 1015, a global ECG measurement model 143 can be trained. The global ECG measurement model 143 measures ECG global parameters including PR interval, QRS duration, QT interval and heart rate. Since these parameters are not lead-specific and are usually estimated based on a combination of multiple leads, these parameters are referred to as ECG global measurements. Typically, these parameters are detected using complex signal processing and feature detection algorithms based on domain knowledge. However, embodiments of the present disclosure use a deep learning model (global ECG measurement model 143) applied to a median beat 8-lead signal generated from an 8-lead ECG signal to estimate these ECG parameters. Since ECG global parameters are closely related to ECG morphology, a global ECG measurement model 143 can be trained to perform regression of the median beat 8-lead ECG signal to detect these parameters. Another purpose of training the global ECG measurement model 143 is to also be able to identify the absence of a valid P wave such as in atrial fibrillation situations.

[0107] Figure 8The detailed structure of the global ECG measurement model 143 is shown. The structure of the global ECG measurement model 143 is similar to that of the ECG morphology model 142, except that the last classification layer of the global ECG measurement model 143 is replaced by the regression layer 820. In addition, the global ECG measurement model 143 does not have the batch normalization layer of the ECG morphology model 142, because the batch normalization layer is more useful for classification than for regression. The global ECG measurement model 143 can receive a median beat 4-lead signal based on the lead group [I, II, v2, v4] as input. More specifically, the median beat generation module 133 can determine the median beat of each lead in the lead group [I, II, v2, v4] as discussed above and provide the median beat 4-lead signal to the global ECG measurement model 143. The global ECG measurement model 143 is trained by minimizing the absolute error between the labeled (physician-estimated) intervals from the training data and the model-estimated intervals to output three intervals: PR interval, QRS duration, and QT interval.

[0108] At block 1020 , ECG rhythm and ECG morphology classification may be determined for the input ECG by the ECG rhythm classification model 141 and the ECG morphology model 142 , respectively. At block 1025 , global parameters may be determined for the input ECG by the global ECG measurement model 143 .

[0109] The final determination stage 150 may include a final ECG rhythm logic 151 and a final ECG morphology logic 152. At box 1030, the final ECG rhythm logic 151 may determine the dominant rhythm from the following list: paced rhythm, sinus / normal sinus, atrial fibrillation, atrial flutter, wide QRS tachycardia, junctional rhythm, and undetermined rhythm. If the dominant rhythm is sinus rhythm, the final ECG rhythm logic 151 may attempt to find an associated sinus arrhythmia. The input parameters of the final ECG rhythm logic 151 are as follows: the output of the ECG rhythm classification model 141, global ECG measurements (PR interval, QRS duration, QT interval, and heart rate), and the RR interval sequence of the dominant beat. All determinations of the final ECG rhythm logic 151 are performed in 2 modes as described above: high specificity and high sensitivity.

[0110] The final ECG morphology logic 152 can make a final morphology determination based on one or more classifications from the ECG rhythm classification model 141 and the ECG morphology model 142 in the following order: Pace ECG, RBBB, LBBB, Other Intraventricular Block, LVH, RVH, Old MI, Acute MI, Other ST Elevation, Ischemia, Long QT, Atrial Enlargement, Normal ECG. The conditions discussed below are common to both the high sensitivity and high specificity modes, with only the thresholds being relatively different. The input parameters to the final ECG morphology logic 152 are as follows: the output of the ECG morphology model 142, the output of the ECG rhythm classification model 141 (in some cases), and global ECG measurements (PR interval, QRS duration, QT interval, and heart rate).

[0111] Fig.11 A diagrammatic representation of a machine in the example form of a computer system 1100 is shown within which a set of instructions for causing the machine to perform any one or more of the methods discussed herein for performing reduced-lead ECG interpretation are shown.

[0112] In alternative embodiments, the machine can be connected (e.g., networked) to other machines in a local area network (LAN), an intranet, an extranet, or the Internet. The machine can operate with the ability of a server or client machine in a client-server network environment, or operate as a peer machine in a peer-to-peer (or distributed) network environment. The machine can be a personal computer (PC), a tablet PC, a set-top box (STB), a personal digital assistant (PDA), a cellular phone, a web device, a server, a network router, a switch or a bridge, a hub, an access point, a network access control device, or any machine that can perform (sequentially or otherwise) an instruction set for specifying the action to be taken by the machine. In addition, although only a single machine is shown, the term "machine" should also be considered to include any set of machines that perform an instruction set (or multiple instruction sets) individually or jointly to perform any one or more than one method discussed herein. In one embodiment, computer system 1100 can represent a server.

[0113] The exemplary computer system 1100 includes a processing device 1102, a main memory 1104 (e.g., a read-only memory (ROM), a flash memory, a dynamic random access memory (DRAM)), a static memory 1106 (e.g., a flash memory, a static random access memory (SRAM), etc.), and a data storage device 1118 that communicate with each other via a bus 1130. Any signals provided on the various buses described herein may be time multiplexed with other signals and provided on one or more common buses. Additionally, the interconnections between circuit components or blocks may be shown as buses or single signal lines. Alternatively, each of the buses may be one or more single signal lines, and each of the single signal lines may alternatively be a bus.

[0114] The computing device 1100 may also include a network interface device 1108 that can communicate with the network 1120. The computing device 1100 may also include a video display unit 1110 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)), an alphanumeric input device 1112 (e.g., a keyboard), a cursor control device 1114 (e.g., a mouse), and an acoustic signal generating device 1116 (e.g., a speaker). In one embodiment, the video display unit 1110, the alphanumeric input device 1112, and the cursor control device 1114 may be combined into a single component or device (e.g., an LCD touch screen).

[0115] Processing device 1102 represents one or more general-purpose processing devices, such as a microprocessor or a central processing unit. More specifically, the processing device can be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computer (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, or a processor that implements other instruction sets, or a processor that implements a combination of instruction sets. Processing device 1102 can also be one or more special-purpose processing devices, such as an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), or a network processor. Processing device 1102 is configured to execute instructions 1125 for performing the operations and steps discussed herein.

[0116] The data storage device 1118 may include a machine-readable storage medium 1128 having stored thereon one or more sets of instructions 1125 (e.g., software) embodying any one or more methods of functionality described herein. The instructions 1125 may also reside, in whole or in part, in the main memory 1104 or in the processing device 1102 during execution of the instructions 1125 by the computer system 1100; the main memory 1104 and the processing device 1102 also constitute machine-readable storage media. The instructions 1125 may also be sent or received over the network 1120 via the network interface device 1108.

[0117] Although the machine-readable storage medium 1128 is shown as a single medium in the exemplary embodiment, the term "machine-readable storage medium" should be taken to include a single medium or multiple media (e.g., a centralized or distributed database, or associated caches and servers) for storing one or more sets of instructions. Machine-readable media include any mechanism for storing information in a form (e.g., software, processing application) readable by a machine (e.g., a computer). Machine-readable media may include, but are not limited to, magnetic storage media (e.g., floppy disks); optical storage media (e.g., CD-ROMs); magneto-optical storage media; read-only memory (ROM); random access memory (RAM); erasable programmable memory (e.g., EPROM and EEPROM); flash memory; or other types of media suitable for storing electronic instructions.

[0118] The previous descriptions set forth many specific details, such as examples of specific systems, components, methods, etc., to provide a good understanding of several embodiments of the present disclosure. However, it will be apparent to those skilled in the art that at least some embodiments of the present disclosure may be practiced without these specific details. In other examples, well-known components or methods are not described in detail or presented in a simple block diagram format to avoid unnecessarily obscuring the present disclosure. Therefore, the specific details set forth are merely exemplary. Specific embodiments may be different from these exemplary details and are still considered to be within the scope of the present disclosure.

[0119] Additionally, some embodiments may be practiced in a distributed computing environment where the machine-readable medium is stored on more than one computer system and / or executed by more than one computer system. In addition, information transmitted between computer systems may be pulled or pushed across the communication medium connecting the computer systems.

[0120] Embodiments of the claimed subject matter include, but are not limited to, the various operations described herein. These operations may be performed by hardware components, software, firmware, or a combination thereof.

[0121] Although the operation of the method herein is shown and described in a particular order, the order of the operation of each method can be changed so that some operations can be performed in reverse order, or some operations can be performed at least partially in parallel with other operations. In another embodiment, the instructions or sub-operations of different operations can be intermittent or alternating.

[0122] The above description of the exemplary implementation of the present invention (including the content described in the abstract) is not intended to be exhaustive or to limit the present invention to the precise form disclosed. Although specific implementations and examples of the present invention are described here for the purpose of illustration, various equivalent modifications within the scope of the present invention are possible as will be recognized by those skilled in the relevant art. The word "example" or "exemplary" is used here to mean to be used as an example, instance or illustration. Any aspect or design described here as "example" or "exemplary" is not necessarily to be interpreted as being preferred or advantageous over other aspects or designs. On the contrary, the use of the word "example" or "exemplary" is intended to present concepts in a concrete way. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or". That is, unless otherwise specified or clear from the context, "X includes A or B" is intended to mean any natural inclusive arrangement. That is, if X includes A; X includes B; or X includes both A and B, then "X includes A or B" is satisfied in any of the foregoing cases. In addition, the articles "a" and "an" used in the present application and the appended claims should generally be interpreted as meaning "one or more than one", unless otherwise specified or clearly pointed to the singular form from the context. In addition, the use of the term "embodiment" or "one embodiment" or "implementation" or "an implementation" throughout the text is not intended to mean the same embodiment or implementation, unless described as such. In addition, the terms "first", "second", "third", "fourth", etc. as used herein are meant to distinguish between different elements, and may not necessarily have the ordinal meaning according to their numerical designations.

[0123] It should be understood that variations or alternatives to the above disclosed and other features and functions may be combined into many other different systems or applications. Various alternatives, modifications, variations or improvements not currently foreseen or expected may be subsequently made by those skilled in the art, which are also intended to be covered by the appended claims. The claims may cover embodiments in hardware, software or a combination thereof.

Claims

1. A method comprising: Training a first deep neural network (i.e., a first DNN) to classify rhythm characteristics of an electrocardiogram (i.e., ECG), wherein a training weight matrix is ​​used to balance differences in popularity between the classifications during training; training a second DNN to classify morphological characteristics of the ECG; training a third DNN to identify global parameters of the ECG, wherein both the first DNN and the second DNN include classification layers, and the third DNN includes a regression layer; and The first DNN, the second DNN, and the third DNN are implemented as an integrated DNN for ECG interpretation.

2. The method according to claim 1, further comprising: In response to receiving an ECG: classifying the rhythm characteristics of the ECG using the first DNN; classifying morphological characteristics of the ECG using the second DNN; and identifying global parameters of the ECG using the third DNN; as well as A set of decision logic is applied to the rhythm characteristics of the ECG and the morphology characteristics of the ECG to diagnose one or more conditions indicated by the ECG based at least in part on global parameters of the ECG.

3. The method according to claim 1, wherein: The classification of the rhythm characteristics of the ECG includes the probability of each rhythm in the rhythm set; and The classification of the morphological characteristics of the ECG includes the probability of each morphology in the morphology set.

4. The method according to claim 2, wherein: The decision logic set includes: a first set of criteria for validating classification of a rhythm characteristic of the ECG based at least in part on a global parameter of the ECG; and A second set of criteria for validating the classification of the morphological characteristic of the ECG based at least in part on the global parameter of the ECG and the validated classification of the rhythm characteristic.

5. The method according to claim 3, wherein: One or more of the first DNN, the second DNN and the third DNN includes a convolutional neural network, i.e., CNN.

6. The method according to claim 1, wherein: Each of the first DNN, the second DNN, and the third DNN includes a convolution block having a residual block structure, and the residual block structure combines a set of convolutional layers and a set of skip layers.

7. The method according to claim 5, further comprising: A random data shift technique is used to further generalize each of the first DNN, the second DNN, and the third DNN.

8. A system comprising: An electrocardiogram monitor is an ECG monitor configured to perform an ECG of a person and transmit said ECG; as well as A computing device configured to: In response to receiving the ECG from the ECG monitor: classifying the rhythm characteristics of the ECG using a first deep neural network, i.e., a first DNN; classifying the morphological characteristics of the ECG using a second DNN; as well as identifying global parameters of the ECG using a third DNN; as well as A set of decision logic is applied to the rhythm characteristics of the ECG and the morphology characteristics of the ECG to diagnose one or more conditions indicated by the ECG based at least in part on global parameters of the ECG.

9. The system according to claim 8, wherein: The processing device is also used for: Training the first deep neural network (i.e., the first DNN) to classify rhythm characteristics of an electrocardiogram (i.e., ECG), wherein a training weight matrix is ​​used to balance differences in popularity between the classifications during training; training the second DNN to classify morphological characteristics of the ECG; The third DNN is trained to identify global parameters of the ECG, wherein both the first DNN and the second DNN include classification layers and the third DNN includes a regression layer.

10. The system according to claim 8, wherein: The classification of the rhythm characteristics of the ECG includes the probability of each rhythm in the rhythm set; and The classification of the morphological characteristics of the ECG includes the probability of each morphology in the morphology set.

11. The system according to claim 8, wherein: The decision logic set includes: a first set of criteria for validating classification of a rhythm characteristic of the ECG based at least in part on a global parameter of the ECG; and A second set of criteria for validating the classification of the morphological characteristic of the ECG based at least in part on the global parameter of the ECG and the validated classification of the rhythm characteristic.

12. The system according to claim 10, wherein: One or more of the first DNN, the second DNN and the third DNN includes a convolutional neural network, i.e., CNN.

13. The system according to claim 8, wherein: Each of the first DNN, the second DNN, and the third DNN includes a convolution block having a residual block structure, and the residual block structure combines a set of convolutional layers and a set of skip layers.

14. The system according to claim 12, wherein: The processing device is also used for: A random data shift technique is used to further generalize each of the first DNN, the second DNN, and the third DNN.

15. A non-transitory computer readable medium having stored thereon instructions which, when executed by a processing device, cause the processing device to: Training a first deep neural network (i.e., a first DNN) to classify rhythm characteristics of an electrocardiogram (i.e., an ECG); training a second DNN to classify morphological characteristics of the ECG; Training a third DNN to identify global ECG parameters; In response to receiving an ECG: classifying the rhythm characteristics of the ECG using the first DNN; classifying morphological characteristics of the ECG using the second DNN; and identifying global parameters of the ECG using the third DNN; as well as A set of decision logic is applied to the rhythm characteristics of the ECG and the morphology characteristics of the ECG to diagnose one or more conditions indicated by the ECG based at least in part on global parameters of the ECG.

16. The non-transitory computer readable medium of claim 15, wherein: The classification of the rhythm characteristics of the ECG includes the probability of each rhythm in the rhythm set; and The classification of the morphological characteristics of the ECG includes the probability of each morphology in the morphology set.

17. The non-transitory computer readable medium of claim 15, wherein: The decision logic set includes: a first set of criteria for validating classification of a rhythm characteristic of the ECG based at least in part on a global parameter of the ECG; and A second set of criteria for validating the classification of the morphological characteristic of the ECG based at least in part on the global parameter of the ECG and the validated classification of the rhythm characteristic.

18. The non-transitory computer readable medium of claim 17, wherein: One or more of the first DNN, the second DNN and the third DNN includes a convolutional neural network, i.e., CNN.

19. The non-transitory computer readable medium of claim 15, wherein: Each of the first DNN, the second DNN, and the third DNN includes a convolution block having a residual block structure, and the residual block structure combines a set of convolutional layers and a set of skip layers.

20. The non-transitory computer readable medium of claim 17, wherein: The processing device is also used for: A random data shift technique is used to further generalize each of the first DNN, the second DNN, and the third DNN.

Images