Detecting cardiac disease from reduced lead-set ECGs.

Abstract

The embodiments of the present disclosure utilize DNNs for ECG interpretation, where the original ECG waveforms are directly populated by the DNN with pairwise interpretation labels for training, without the need for descriptive feature extraction or rule-based criteria.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to U.S. Provisional Application No. 63 / 328,670, entitled “SINGLE ALGORITHM DETECTION OF CARDIAC CONDITIONS FROM REDUCED LEAD SET ECG,” filed April 7, 2022, the disclosure of which is incorporated herein by reference.

[0002] Aspects of the present disclosure relate to diagnosing cardiac disease based on electrocardiograms acquired with a reduced lead set, and in particular to algorithmically diagnosing cardiac disease using a reduced lead set and a single algorithm. [Background technology]

[0003] Cardiovascular disease is the leading cause of death in the world. In 2008, 30% of deaths worldwide were attributable to cardiovascular disease. It is also estimated that by 2030, more than 23 million people will die from cardiovascular disease each year. Cardiovascular disease is equally prevalent throughout first and third world populations, affecting people regardless of socioeconomic status.

[0004] Arrhythmia is a heart condition in which the electrical activity of the heart is irregular or faster than normal (tachycardia) or slower than normal (bradycardia). While many arrhythmias are not life-threatening, some can cause cardiac arrest or even sudden cardiac death. In fact, cardiac arrhythmias are one of the most common causes of death in hospital visits. Atrial fibrillation (A-fib) is the most common cardiac arrhythmia. In A-fib, electrical conduction through the ventricles is irregular and disorganized. A-fib may not cause symptoms, but is often accompanied by 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. To treat A-fib, patients may take medications to slow the heart rate or correct the heart rhythm. Patients may also take anticoagulants to prevent strokes and may undergo surgical interventions, including cardiac ablation, to treat A-fib. In another example, ECG can provide decision support for acute coronary syndromes (ACS) or other cardiac diseases by interpreting various rhythmic and morphological conditions, including myocardial infarction (MI) and ischemia.

[0005] Often, patients with A-fib (or other types of arrhythmias) are monitored over extended periods of time to manage the disease. For example, patients may be provided with a Holter monitor or other portable electrocardiogram device to continuously monitor the electrical activity of the cardiovascular system, e.g., for at least 24 hours. Such monitoring can be important in detecting conditions such as acute coronary syndromes (ACS), among others.

[0006] The American Heart Association and the European Society of Cardiology recommend that a 12-lead ECG should be obtained as soon as possible for patients with possible ACS or other cardiac disease when symptoms are present. Pre-hospital ECGs have been shown to significantly reduce time to treatment and indicate better survival. Time to first ECG is so important that it is a quality and performance indicator monitored by several regulatory agencies. According to the 2015 National Health Statistics, in the United States (US), over 7 million people visited an emergency department (ED) with a primary complaint of chest pain or symptoms related to cardiac disease. In the US, ED visits are increasing at a rate of 3.2% annually, and outside the US, ED visits are increasing at a rate of 3% to 7% annually. Summary of the Invention [Means for solving the problem]

[0007] The described embodiments and their advantages may best be understood by reference 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. [Brief description of the drawings]

[0008] [Figure 1A] FIG. 1 illustrates an exemplary system, according to some embodiments of the present disclosure. [Figure 1B] FIG. 1 is a block diagram illustrating an example system, according to some embodiments of the present disclosure. [Diagram 2] FIG. 2 illustrates various leads that may comprise an ECG, according to some embodiments of the present disclosure. [Diagram 3] FIG. 1 is a block diagram illustrating an example system, according to some embodiments of the present disclosure. [Figure 4] FIG. 1 is a block diagram illustrating the structure of a machine learning (ML) model for performing reduced lead ECG interpretation, according to some embodiments of the present disclosure. [Figure 5A] FIG. 1 is a block diagram illustrating a basic structure of a deep neural network (DNN), in accordance with some embodiments of the present disclosure. [Figure 5B] FIG. 1 is a block diagram illustrating an example convolutional neural network (CNN), in accordance with some embodiments of the present disclosure. [Figure 5C] FIG. 2 is a block diagram illustrating an example training process for a CNN, in accordance with some embodiments of the present disclosure. [Figure 6] FIG. 2 is a block diagram illustrating an example ECG rhythm classification model, according to some embodiments of the present disclosure. [Figure 7A] FIG. 2 is a block diagram illustrating an example ECG morphology classification model, in accordance with some embodiments of the present disclosure. [Figure 7B] FIG. 2 is a block diagram illustrating an ECG signal, according to some embodiments of the present disclosure. [Figure 7C] FIG. 2 is a block diagram illustrating a median beat of an ECG signal, according to some embodiments of the present disclosure. [Figure 8] FIG. 2 is a block diagram illustrating an example global ECG feature model, according to some embodiments of the present disclosure. [Figure 9] FIG. 7 illustrates example conditions for validating classification of the ECG rhythm classification model of FIG. 6, according to some embodiments of the present disclosure. [Figure 10] FIG. 1 is a flow diagram of a method for training and implementing a reduced lead ECG interpretation model according to some embodiments of the present disclosure. [Figure 11] FIG. 1 is a block diagram of an example computing device that may perform one or more of the operations described herein, in accordance with some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] In the following detailed description of the embodiments of the present disclosure, numerous specific details are set forth in order to provide a more complete understanding of the present disclosure. However, it will be apparent to those skilled in the art that the concepts within the present disclosure can be practiced without these specific details. In other instances, well-known features are not described in detail to avoid unnecessarily complicating the description.

[0010] An electrocardiogram (ECG) provides several ECG waveforms that represent the electrical activity of a person's heart. An ECG monitoring device may include a set of electrodes for recording these ECG waveforms of a patient's heart (also referred to herein as "taking an ECG"). The set of electrodes may be placed at multiple locations on the patient's skin, and the electrical signals (ECG waveforms) recorded between each electrode pair in the set of electrodes may be referred to as leads. Various numbers of leads may be used to take an ECG, and various numbers and combinations of electrodes may be used to form the various leads. Exemplary numbers of leads used to take an ECG are 1, 2, 3, 6, and 12 leads.

[0011] The ECG waveforms recorded by an ECG monitoring device (each corresponding to an ECG lead) may contain data corresponding to the electrical activity of a person's heart. A typical heartbeat may contain several variations in electrical potential, which may be classified into waves and wave complexes, including P waves, QRS complexes, T waves, and U waves, among others, as known in the art. In other words, each ECG waveform may contain P waves, QRS complexes, T waves, and U waves, among others, as known in the art. The shape and duration of these waves may be related to various characteristics of a person's heart, such as the size of the person's atrium (e.g., indicative of atrial enlargement), and may be the first source of heartbeat characteristics unique to a person. The ECG waveform may be analyzed (typically after standard filtering and "cleaning" of the signal) for various metrics that are useful for detecting cardiac events or conditions, such as detecting and characterizing cardiac arrhythmias. Such metrics may include, for example, ECG waveform amplitude and morphology (e.g., QRS complex amplitude and morphology), R-ST segment and T wave amplitude analysis, and heart rate variability (HRV).

[0012] As mentioned above, an ECG waveform is generated from measuring multiple leads (each lead 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 different morphology / amplitude). This is because the various leads may analyze the same electrical event, but each from a different angle. FIG. 1A shows a view 105 of the ECG waveform detected by each of the three leads (I, II, and III) when a three-lead ECG is taken, as well as an exploded view 110 of the ECG waveform measured by lead III showing the QRS complex. As shown, the amplitudes and morphologies of the ECG waveforms taken from leads I-III are all different, with the ECG waveform measured by lead III having the largest amplitude and the ECG waveform measured by lead I having the smallest amplitude.

[0013] There are various "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 electrodes 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), e.g., I=LA-RA. Lead II is typically the voltage between the left leg (LL) and the right arm (RA), e.g., II=LL-RA. Lead III is typically the voltage between the left leg (LL) and the left arm (LA), e.g., III=LL-LA. Augmented limb leads can also be determined from RA, RL, LL, and LA. Augmented vector right (aVR) lead is equal to RA-(LA+LL) / 2 or -(I+II) / 2.Augmented vector left (aVL) lead is equal to LA-(RA+LL) / 2 or I-II / 2.Augmented vector foot (aVF) lead is equal to LL-(RA+LA) / 2 or II-I / 2.

[0014] It should be noted that two or more lead sets may be transformed to generate a complete 12-lead ECG. Such transformation may be performed using a machine learning model (e.g., neural networks, deep learning techniques, etc.). The machine learning model may be trained using 12-lead ECG data corresponding to a population of individuals. The data may be pre-processed to filter the data in a manner appropriate for the application before being input into the machine learning model. For example, the data may be categorized according to height, sex, weight, nationality, etc. before being used to train one or more machine learning models such that the resulting one or more models are fine-tuned for a particular type of individual. In a further embodiment, the machine learning model may be further trained based on the user's own ECG data to fine-tune and personalize the model to further reduce any residual synthesis error.

[0015] As discussed herein, for patients with possible cardiac disease, a 12-lead ECG should be obtained as soon as symptoms are present, since pre-hospital ECGs have been found to significantly reduce time to treatment and have been shown to have better survival rates. In addition, current portable ECG devices, such as Holter monitors, are typically bulky and difficult for subjects to administer without the assistance of a medical professional. For example, use of a Holter monitor requires the patient to wear a bulky device on their chest and to precisely place multiple electrode leads in precise locations on their chest. These requirements can interfere with the subject's activities, including the subject's natural movements, such as bathing and showering. Once an ECG is taken by such a device, the ECG is transmitted to the subject's physician, who then analyzes the ECG waveform and provides a diagnosis and other recommendations. Currently, this process often must be performed through hospital administrators and health care organizations, and many patients do not receive feedback in an appropriate manner.

[0016] Although 12-lead ECGs may be administered in hospitals or clinics to detect cardiac disease, it is desirable to have a simpler reduced-lead system that can be used at home, or in a doctor's office, or in an emergency department, or in pre-hospital settings to detect cardiac disease. Certain reduced-lead (e.g., 6-lead) devices exist that can mathematically construct all 12 leads (where some are "synthetic leads") based on the reduced-lead set by either linear or non-linear transformations to provide a full set of 12 leads. This composite full 12-lead data can be analyzed by a physician or a full 12-lead algorithm. Based on existing techniques to reconstruct the full 12-lead set, a natural workflow for an automated system would be as follows: 1) use a lead reconstruction algorithm to create a 12-lead set from the reduced lead set (e.g., 4-lead), and then 2) use a pathology classification algorithm that takes the reconstructed 12-leads as input and provides as output a determination (classification) of normal and various pathologies.

[0017] However, there is an inherent tradeoff with such reduced lead devices as they have less resolution / ability to detect and identify the full range of cardiac pathology using the standard 12-lead ECG criteria recommended by the American College of Cardiology (ACC / AHA). It would be desirable to have the discrimination capabilities of a 12-lead system with the simplicity of a reduced set of leads that can be used outside of the hospital (at home, in emergency response, in a doctor's office, etc.) without the tradeoff of sacrificing discrimination resolution.

[0018] 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 prediction or classification. This progression of computations through the network is called forward propagation. The input and output layers of a deep neural network are called the visible layers. The input layer is where the deep learning model takes in data to process, 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 when attempting to "subsume" multiple algorithmic steps into a single step. DNNs can essentially subsume all the algorithms used to make a diagnosis based on ECG data into a single mathematical process, using a technique called "end-to-end" learning. By utilizing this end-to-end training of what were multiple task / ML models, the resulting DNNs can provide highly accurate results, which can often exceed the performance provided by the use of multiple dedicated algorithms.

[0019] FIG. 1B illustrates a system 100 for cardiac disease management. The system 100 may be prescribed for use by a first user, e.g., the first user's physician. Alternatively, the system 100 may be used without input from a physician or other third party. The system 100 may comprise a local computing device 101 of the first user. The local computing device 101 may be loaded with a user interface, dashboard, or other subsystems of the cardiac disease management system 100. For example, the local computing device 101 may be loaded with a mobile software application ("mobile app") 101A for interfacing with the system 100. The mobile app 101A may be configured to interface with one or more biosensors (e.g., ECG monitor 103) and may comprise software and a user interface for managing biodata collected by the local computing device 101 from the one or more biosensors. The local computing device 101 may comprise any suitable computing device, such as, for example, a tablet computer, a smartphone, a server computer, a desktop computer, a laptop computer, or a body-worn computing device (e.g., a smart watch or other wearable). In some embodiments, the local computing device 101 may comprise a single computing device or may include multiple interconnected computing devices (eg, multiple servers arranged in a cluster).

[0020] The local computing device 101 may be coupled to one or more biosensors. For example, the local computing device 101 may be coupled to an ECG monitor 103 that may include a set of electrodes for recording ECG (electrocardiogram) data of the first user's heart (also referred to herein as "taking an ECG"). The ECG data may be recorded or taken using a set of electrodes placed at multiple locations on the first user's skin. The electrical signals recorded between electrode pairs may be referred to as leads, and various numbers of leads may be used to record the ECG data, and various numbers and combinations of electrodes (placed on the user's arms, legs, and chest) may be used to form the various leads. Exemplary numbers of leads used to take an ECG are 1, 2, 6, and 12 leads. 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 in the same location on the left and right arms, e.g., near the wrist. The leg electrodes 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, for example near the ankles.

[0021] FIG. 2 shows a single dipole heart model 115 with a 12-lead set with leads I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, and V6 all represented on a 6-axis system. The heart model 115 assumes a uniform heart region in all directions, changing only in magnitude and direction with cycle time. As shown in FIG. 2, there are two orthogonal planes, the frontal plane and the horizontal plane. Within each plane, there are several leads that cover the entire plane. In the frontal plane, there are two independent leads I and II, and four other derived leads III, aVR, aVL, and aVF, each 30 degrees apart. The reason the frontal plane has two independent leads is that they are far-field leads, each of which can cover a wider field of view, like a wide-angle camera lens, but provide less detail. In the horizontal plane, there are usually six independent leads, all of which are closer to the heart than the limb leads and may be called near-field leads. Following the same camera analogy, the near-field leads may act like a zoom lens, covering a smaller field of view but more accurate for localized activity such as ischemia and infarction. The two orthogonal planes are related by using a composite reference point formed by leads I and II, called the Wilson-Central-Terminal (WCT). This is defined as RA+LA+RL / 3, but both leads I and II are recorded with RA as the reference so that the voltage in RA can be considered zero, and the WCT (VW) can be calculated as lead I+lead II / 3, using RA as the reference for both leads I and II (thus assuming that RA has a potential of zero).

[0022] Returning to FIG. 1B, in some embodiments, the ECG monitor 103 may comprise a handheld ECG monitor (such as, for example, AliveCor® Inc.'s KardiaMobile® or KardiaMobile® 6L devices) with a smaller number of electrodes (e.g., 2, 3, 4, or any number less than 12 electrodes). In these embodiments, the electrodes can be used to measure a subset of the leads shown in FIG. 2, such as Lead I (e.g., voltage between the left arm and the right arm) simultaneously with Lead II (e.g., voltage between the left leg and the right arm), and Lead I simultaneously with another one of the chest leads, such as Lead V2 or V5. It should be noted that any other combination of leads is possible. If desired, additional leads can then be algorithmically derived (e.g., by the ECG monitor 103 itself or the local computing device 101) from the determined subset of leads. For example, an augmented limb lead can also be determined from values ​​measured by the LA, RA, LL, and RL electrodes. The augmented vector right (aVR) may be equal to RA-(LA+LL) / 2 or -(I+II) / 2. The augmented vector left (aVL) may be equal to LA-(RA+LL) / 2 or I-II / 2. The augmented vector foot (aVF) may 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 may utilize a machine learning (ML) model to derive a complete 12-lead set from a subset of measured leads. In some embodiments, the ECG monitor 103 may be in the form of a smartphone or a wearable device such as a smart watch. In some embodiments, the ECG monitor 103 may be a handheld sensor coupled to the local computing device 101 with an intermediate protective case / adapter.

[0023] The ECG data recorded by the ECG monitor 103 may include, for example, the electrical activity of the first user's heart. A typical heartbeat may include several variations in electrical potential, which may be categorized into waves and wave complexes, including P waves, QRS complexes, T waves, and possibly U waves, as known in the art. The shape and duration of the P waves may be related to the size of the user's atria (e.g., indicative of atrial enlargement) and may be a first source of user-specific heartbeat characteristics.

[0024] The QRS complex may correspond to the depolarization of the ventricles and can be divided into three separate 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. Also, the His / Purkinje system of the heart, which can increase the conduction velocity to coordinate the depolarization of the ventricles, may cause the QRS complex to appear "peaky" rather than rounded. The duration of the QRS complex in a healthy heart may be in the range of 60 to 100 milliseconds (ms), but may vary due to conduction abnormalities. The duration of the QRS complex may serve as another source of heartbeat characteristics unique to the user.

[0025] The duration, amplitude, and morphology of each of the Q, R, and S waves may vary from person to person, and may vary significantly, especially in users with cardiac disease or cardiac irregularities. For example, a Q wave that is greater than 1 / 3 the height of an R wave or has a duration longer than 40 ms (milliseconds) may indicate a myocardial infarction and provide a unique characteristic of the user's heart. Similarly, other healthy ratios of Q and R waves may be used to distinguish the heartbeats of different users.

[0026] The cardiac electrical activity of a user US may also include one or more characteristic durations or intervals that can be used to distinguish different users. For example, the cardiac electrical activity may include a PR interval and an ST segment, as known in the art. 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 from 120 to 200 ms. PR intervals with different durations may indicate one or more defects in the heart, such as first degree heart block (e.g., PR interval lasting more than 200 ms), pre-excitation syndrome via an accessory pathway leading to early activation of the ventricles (e.g., PR interval lasting less than 120 ms), or another type of heart block (e.g., variable PR interval). The ST segment may be measured, for example, from the QRS complex to the T wave, starting at 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 from 80 ms to 120 ms and usually has a slight upward concavity. The combination of the ST segment length and the ST segment depression or elevation can also be used to generate characteristic information specific to each user's heart beat.

[0027] The ECG monitor 103 may be used by a user to measure their own ECG data for analysis, as discussed in further detail herein. The ECG monitor 103 may also transmit the measured ECG data and / or results of diagnostic analyses (discussed in further detail herein) to the local computing device 101 by any suitable wired or wireless connection, such as, for example, a Wi-Fi connection, a Bluetooth® connection, a near field communication (NFC) connection, an ultrasonic signal transmission connection, etc.

[0028] The ECG data may be recorded continuously by the user at regular intervals. For example, the intervals may be once a day, once a week, once a month, or some other predetermined interval. The ECG data may be recorded under similar or different circumstances, at the same or different times of the day, as described herein. The ECG data may also be recorded at the same or different times of the interval (e.g., the ECG data may be captured asynchronously). Alternatively or additionally, the ECG data may be recorded upon request by the user at various separate times, such as when the user feels chest pain or experiences other unusual or abnormal sensations, or, for example, in response to instructions from the user's physician. In another embodiment, the ECG data may be recorded continuously over a period of time.

[0029] Each ECG data record may be time-stamped and 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 app 101A) may include a user interface for data entry that allows the user to input their user characteristics. Examples of user characteristics may include age, sex, race, ethnicity, relevant medical history, location, diet (e.g., food / drink habits), medication / drug consumption, exercise patterns, sleep / rest patterns, stress, anxiety, pain, or other unusual or abnormal sensations, activities performed before, during, or after recording of the data, or any other user-specific circumstances or factors that may affect the user's ECG data. The local computing device 101 may add the user characteristics to the ECG data and store the ECG data and the user characteristics for further analysis (e.g., interpretation of the ECG). Alternatively, the local computing device 101 may transmit the ECG data and the user characteristics (collectively referred to as user data) to the cloud storage system 113. Because the user data is time-stamped or tagged, the ECG data can be matched or associated with an activity or situation of interest. As described in more detail herein, this also allows for comparison of the ECG data before, after, and during the activity or situation of interest so that effects on the ECG data can be determined and taken into account during further analysis.

[0030] When user data is transmitted by the local computing device 101 to the cloud storage system 113 for storage and analysis, such transmission may be in real time, at regular intervals such as hourly, daily, weekly, etc., and / or any interval therebetween, or may be on-demand. The local computing device 101 and the cloud storage system 113 may be coupled to each other (e.g., operatively coupled, communicatively coupled, and 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, such as a wireless carrier system that may be implemented using Wi-Fi hotspots connected to the network 140, and / or various data processing facilities, communication towers (e.g., cell towers), and the like. 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.

[0031] The accumulated user data over time for a particular user may form a time series health record for that particular user. Cloud storage system 113 may comprise any suitable type of computing device or machine having a programmable processor, including, for example, a server computer, a desktop computer, a laptop computer, a tablet computer, a smartphone, etc. In some embodiments, cloud storage system 113 may comprise a single computing device or may comprise multiple interconnected computing devices (e.g., multiple servers configured in a cloud storage cluster).

[0032] FIG. 3 illustrates a hardware block diagram of the computing device 101. 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 that may store data. The persistent storage may be a local storage unit or a remote storage unit. The persistent storage 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 may also be a monolithic / single device or a distributed set of devices. The memory 120 may be configured for long-term storage of data and may retain data between power on / off cycles of the computing device 101. The memory 120 may store user data accumulated over time, not only for the user, but also for a number of other users. The memory 120 may further include a reduced-lead ECG analysis model 120A configured to analyze the user's reduced-lead ECG data measured each time an ECG is performed to determine whether the user is experiencing a cardiac disorder, as discussed in more detail herein.

[0033] As discussed herein, mobile ECG monitors attempt to simplify the process of taking an ECG by using a reduced set of leads, e.g., any number of leads less than 12. However, such implementations must essentially sacrifice resolution or discrimination ability in detecting the full range of cardiac pathologies for the convenience of the reduced lead set. Previous work has attempted to mathematically construct all 12 leads based on the reduced lead set by either linear and non-linear transformations to provide a set of all 12 leads (wherein some are "synthetic leads"), and this composite complete 12-lead data can be analyzed by a physician or a full 12-lead algorithm. Based on this previous work to reconstruct the complete 12-lead set, a natural workflow for an ML model implemented by a mobile ECG device also involves multiple layers, including, for example, 1) using a lead reconstruction algorithm to create a 12-lead set from the reduced lead set (e.g., 4 leads), and 2) using a pathology classification algorithm that takes the reconstructed 12 leads as input and provides as output a determination (classification) of normal and various pathologies. However, the application of reduced 12-lead sets and synthetic leads may be limited because the detection and interpretation of abnormal cases are subject to ECG criteria developed for standard 12-lead ECG. There are no specific criteria developed or recommended by the ACC / AHA for reduced lead sets. This results in gaps in the accuracy of such synthetic leads, especially in localized STEMI situations.

[0034] In addition, solving the problem of providing accurate readings of reduced lead ECG requires an ML model architecture consisting of various layers (e.g., pre-processing, feature extraction, optimization, prediction, decision-making), including lead reconstruction and pathology classification, each of which processes input data in a pipelined manner. However, such multi-layer approaches may provide less than ideal results due to the complexity of the model. For example, such multi-layer architectures require each layer to be individually optimized under different criteria.

[0035] A deep neural network (DNN) is a type of model that consists of multiple layers of interconnected nodes, each layer building on the previous layer to refine and optimize predictions or classifications. This progression of computations through the network is called forward propagation. The input and output layers of a deep neural network are called the visible layers. The input layer is where the deep learning model takes in data to process, and the output layer is where the final prediction or classification is made. For example, you may have a set of photos of various pets and want to categorize them by "cat," "dog," "hamster," etc. A deep learning algorithm can determine which features (e.g., ears) are most important to distinguish each animal from another. In machine learning, this hierarchy of features is manually established by human experts. Then, through the process of gradient descent and backpropagation, the DNN adjusts and adapts its accuracy, allowing it to predict with increasing accuracy for new photos of animals. Backpropagation involves fine-tuning the weights of the neural net based on the error rate (i.e., loss) obtained in the previous epoch (i.e., iteration). Proper tuning of the weights ensures lower error rates and makes the model more reliable by increasing generalization.

[0036] DNNs can also be trained using end-to-end learning, which utilizes a DNN structure consisting of several layers to solve complex problems. For ECG processing, the input end is the ECG waveform, and the output end is the classification result. Each DNN layer (or group of layers) can be specialized to perform intermediate tasks required for such problems. For example, the convolutional layer of a DNN model (CNN) can automatically extract ECG features. However, those features are mostly distributed into vectors in an intermediate output layer called a flattening layer at the end of the CNN block. The elements in the flattening layer, also called embedding variables, can be used for either classification of rhythm and morphology interpretation, or regression to predict ECG measurements such as PR interval, QRS duration, and QT interval. End-to-end learning refers to training a possibly complex learning system represented by a single model (specifically a DNN) that represents the complete target system, bypassing the intermediate layers that are usually present in traditional pipeline designs. In other words, when two or more ML models are serving as components of a larger architecture, the DNN may be trained in an end-to-end manner that involves training all components simultaneously (i.e., training as a single network).

[0037] The embodiments of the present disclosure utilize DNNs for ECG interpretation, where the original ECG waveforms are directly populated by the DNNs using pairwise interpretation labels for training, without the need for explanatory feature extraction or rule-based criteria. As will be discussed in more detail herein, the embodiments of the present disclosure differ from traditional expert system-based criteria because the DNN model plays a key role in the entire ECG classification and ECG interpretation process.

[0038] 4A illustrates a reduced lead ECG analysis model 120A (hereafter referred to as interpretation model 120A). 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 a downsampling module 131, a lead conversion module 132, and a median beat generation module 133, as well as filtering and signal processing functions. 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 using 10 second samples of leads [I, II, v1 / v2, v4], this is for illustrative purposes and not limitation. In fact, any suitable set of input leads may be used, and the signal of each lead may have any suitable duration. Downsampling module 131 may downsample each input 4-lead ECG signal to 150 Hz, and lead transformation module 132 may determine the separate 4 limb leads [III, aVR, aVL, aVF] using a conventional lead reconstruction algorithm. Lead transformation module 132 may add the algorithmically reconstructed leads to each of the 4-lead ECG signals, resulting in 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].

[0039] Training an ML model for classifying an ECG signal requires determining basic characteristics of an ECG signal. The model analysis stage 140 may include ML models for identifying basic characteristics of an input ECG signal, including ECG rhythm patterns and ECG waveform morphology / shape patterns. Each of the ML models in the model analysis stage 140 may be any suitable deep neural network (DNN) model, such as a convolutional neural network (CNN) or any other suitable DNN. Determining an ECG rhythm pattern requires capturing the sequential characteristics of the ECG rhythm pattern, which in turn requires effectively detecting those characteristics. Thus, the model analysis stage 140 may include an ECG rhythm classification model 141 for each of the 8-lead ECG signals. In other words, each 8-lead ECG signal may serve as an input to the corresponding ECG rhythm classification model 141. To allow the sequential characteristics of the rhythm pattern to be detected without further feature detection, an appropriate time window for the input data (in this case, a 10 second window for two 8-lead ECG signals) should be used.

[0040] Besides rhythm pattern, waveform morphology / shape pattern is another main ECG characteristic, which is usually detected using ECG amplitude and duration / interval. For waveform morphology, representative beats of various types of data series (e.g., median beat, average beat of multiple ECG cycles) can simplify and speed up the training of waveform morphology detection model, while also improving classification accuracy. Thus, the median beat generation module 133 may determine the median beat of the first and second 8-lead ECG signals, resulting in the first and second median beat 8-lead signals, respectively. The model analysis stage 140 may include an ECG morphology model 142 for each of the median beat 8-lead signals. In other words, each median beat 8-lead signal may serve as an input to the corresponding ECG morphology ML model 142. The model analysis stage 140 may further include a global ECG measurement model 143 for detecting global ECG parameters, including, for example, PR interval, QRS duration, and QT interval. Although embodiments of the present disclosure are described with respect to only a first 8-lead ECG signal and corresponding ECG rhythm classification model 141 and ECG morphology ML model 142, this is for ease of explanation, and each 8-lead ECG signal may have its own DNN model that may form a corresponding classification, as described in further detail herein.

[0041] The final decision stage 150 may include a final ECG rhythm logic 151 and a final ECG morphology logic 152, both of which may include some simple logic to validate the classification by the corresponding DNN. The final analysis combines all the DNN outputs with ECG measurements such as PR, QRSD, QT interval, ST level, etc. to refine the classification. The output of the final decision stage 150 may be a probability predicting the likelihood of various (often difficult to distinguish) cardiac pathologies (or an indication that the user's cardiac health is normal).

[0042] FIG. 5A shows the general structure 500 of each model 141, 142, and 143 used in the model analysis stage 140. Each model may include four main components. The first component may be a lead formation layer 505 that reformats the leads of the 8-lead ECG signal (or the median beat 8-lead signal) into the format required to train the model. The second component may be a set of convolutional layers 510 that may include several (e.g., 5 to 8) CNN blocks with different numbers of filters for each block. The third component may be a set of dense layers 515 that reformulate the multi-filter output of the final convolutional layer into a flattening layer and pass the result to the next dense layer with full connection (meaning that each unit of the previous layer is connected to every unit of the next layer). The fourth component may be a classification layer 520 that generates a probability for each decision.

[0043] FIG. 5B illustrates an example of a DNN 400 that can be used in combination with some embodiments of the present disclosure. The DNN 400 may comprise a trained CNN model that takes input data 402 (e.g., user data) into a convolutional layer (aka hidden layer) 403 and applies a set of trained weights or filters 404 to the input data 406 in each of the convolutional layers 403. The output of the first convolutional layer is an activation map (not shown), which is the input to a second convolutional layer, to which trained weights or filters (not shown) are applied, where the output of the subsequent convolutional layers results in activation maps that represent increasingly complex features of the input data to the first layer. After each convolutional layer, a nonlinear layer (not shown) is applied to introduce nonlinearity into the problem, which may include activation functions such as tanh, sigmoid, or ReLU. In some cases, a pooling layer (not shown) may be applied after the nonlinear layer, which is also called a downsampling layer, taking a filter of the same length and stride, applying it to the input, and outputting the maximum number in all sub-regions that the filter convolves. Other options for pooling are average pooling and L2 norm pooling. The pooling layer reduces the spatial dimension of the input volume to reduce computational cost and control overfitting. The final layer of the network is a fully connected layer, which receives the output of the last convolutional layer and outputs an n-dimensional output vector representing a predicted quantity, e.g., the probability of LVH diagnosis and the severity of LVH (if applicable). This may result in a predicted output 406 (O*), e.g., this user is likely to suffer from stage 2 LVH. As described more fully below, the trained weights 404 may be different for each of the convolutional layers 403.

[0044] FIG. 5C illustrates the process of training the DNN 400 (for simplicity, the dense and classification layers are not shown). In FIG. 5C, the convolutional layer 403 is composed of individual convolutional layers 405, 405′ to form a convolutional layer (405). n-1, and the last nth layer is a fully connected layer. It will be understood that the last layer may be more than one fully connected layer. Training examples 411 are input to the convolutional layer 403, and a nonlinear activation function (not shown) and weights 410, 410′ to 410 are input to the convolutional layer 403. n are successively applied to the training examples 411, where the output of any convolutional layer is input to the next layer, and finally the n-th fully connected layer (405). n-1 This continues in the same manner until produces an output 414. The output or prediction 414(*) is compared to a training example 414 (e.g., Wide QRS Tachycardia) to compare the output or prediction 414 with the training example 411 (I in FIG. 6B). known If the difference or loss 416 is less than some pre-set loss (e.g., if the output or prediction 414 predicts that the user is suffering from wide QRS tachycardia with a required level of accuracy), the DNN model is considered to have converged and trained.

[0045] If the DNN 400 has not converged, backpropagation techniques can be used to iterate over the weights 410 and 410 n-1 From 410 n is updated depending on how close the prediction is to the known input. Those skilled in the art will appreciate that methods other than backpropagation may be used to adjust the weights. A second training example (e.g., different user data) may be input, and the process is repeated again with the updated weights, and then the weights are updated again and so on until the nth training example (e.g., nth user data) is input. This is repeated over and over with the same n training examples until the DNN model is trained or converges to the correct output for the known input. Once the DNN model is trained, the weights 410 to 410 nare fixed and used in the trained ECG rhythm classification model 141. As described herein, there are different weights for each convolutional layer 403 and for each of the fully connected layers. The trained ECG rhythm classification model 141 is then fed user data to determine or predict what it was trained to predict / identify (e.g., rhythm classification), as described in further detail herein.

[0046] FIG. 6 shows a detailed structure of the ECG rhythm classification model 141. The convolutional layer 610 may utilize a residual block structure that utilizes a combination of convolutional and skip layers. The skip layer (also called skip connection or shortcut connection) may allow the residual block in the deep learning model to address the problem of vanishing and exploding gradients that can occur when training very deep neural networks. This is done by directly adding the input of a layer to the output of one or more subsequent layers, creating a "shortcut" for the gradient to flow through. As a result, different parts of the convolutional layer 610 are trained at different rates on different training data points based on how the error flows backward through the layers of the convolutional layer 610. FIG. 6 shows the residual block 610A, where the skip layer is used to skip the training of the convolutional layer 2 during the learning of the identity function X. ReLu may represent a nonlinear activation function.

[0047] In addition to the residual blocks, the ECG rhythm classification model 141 may include two dense layers 615 and one classification layer 620. In some embodiments, the ECG rhythm classification model 141 may further include batch normalization and dropout layers for all residual blocks and dense layers (not all batch normalization and dropout layers shown in FIG. 6) to improve its robustness or generalization once trained. The ECG rhythm classification model 141 is designed to classify rhythm decisions, an example of which is shown in Table 1 (below). The output of the classification layer 620 may have both positive and negative probabilities for each decision. In some embodiments, for each of the 8-lead ECG signals (downsampled to 150 Hz), 9.5 seconds of the 10 seconds of data may be used for input, and the remaining 0.5 seconds of data may be used for augmentation by a time shifting technique (random data shift). This random data shift is one of the data augmentation methods used to improve the generalization of a trained DNN model, making it less sensitive to the start and end times of the input ECG rhythm data.

[0048] [Table 1]

[0049] The ECG rhythm classification model 141 may be trained using any sufficiently large training data set. For example, the training data set may include 1 million labeled ECGs with another 250,000 labeled ECGs used as an independent test set. In the example of FIG. 6A, 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 one with the smallest validation error. A training weight matrix is ​​used to balance the large difference in prevalence among all decisions. For decisions with lower prevalence, higher weights are used when the training error is backpropagated.

[0050] FIG. 7A shows the detailed structure of the ECG morphology model 142. The ECG morphology model 142 may capture the morphology changes of each lead and the relative changes between leads, referred to as changes in the spatial domain. The ECG morphology model 142 may have a similar model structure to the ECG rhythm classification model 141, but has a smaller lead formation layer and less residual block layer 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 the median beats are relatively aligned around their QRS onset. Another important design consideration for the morphology model QRS region is that the variation along the time axis is relatively smaller than the rhythm signal.

[0051] To capture the spatial domain changes, the ECG morphology model 142 may form the ECG imaging by laying out all the leads in a two-dimensional 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 may process the two-dimensional ECG signal matrix with a two-dimensional convolution filter, as in the case of image classification. This method is more effective than applying multiple one-dimensional convolution filters to each lead individually, which is used by most other ECG training models. Since the ECG is a projection onto the torso of the process of electrical depolarization and repolarization of the heart, a spatial two-dimensional signal matrix can capture this projection more effectively. Therefore, a two-dimensional convolution mode is applied to this ECG image input.

[0052] As discussed herein, the median beat generating module 133 may generate a median beat 8-lead signal based on the 8-lead ECG signal. More specifically, the median beat generating module 133 may form a QRS detection function, perform beat detection and classification, perform beat alignment, and calculate a median beat and an average beat from the dominant beat type. FIG. 7B shows an exemplary 8-lead ECG signal, and FIG. 7C shows the median beat of the exemplary 8-lead ECG signal. As can be seen, the ECG morphology features are captured at the median beat, but the data length is much shorter.

[0053] The ECG morphology model 142 may be trained using the same training dataset as the ECG rhythm classification model 141, which may include the median beat data for each training ECG. In the example of FIG. 6A, 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 one with the smallest validation error. A morphology training weight matrix is ​​also used to balance the large difference in prevalence among all the decisions. For decisions with lower prevalence, higher weights are used when the training error is backpropagated. The ECG morphology model 142 is designed to classify ECG morphology decisions, 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 decision.

[0054] [Table 2]

[0055] The global ECG measurement model 143 may measure ECG global parameters including PR interval, QRS duration, QT interval, and heart rate. These parameters are referred to as ECG global measurements because they are not lead specific and are often estimated based on a combination of multiple leads. These parameters are often detected using complex signal processing and feature detection algorithms based on domain knowledge. However, an embodiment of the present disclosure estimates these ECG parameters using a deep learning model (global ECG measurement model 143) applied to a median beat 8-lead signal generated from an 8-lead ECG signal. Since the 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 be able to also identify the absence of a valid P wave, such as in the case of atrial fibrillation.

[0056] FIG. 8 shows a detailed structure of the global ECG measurement model 143. The structure of the global ECG measurement model 143 is similar to that of the ECG morphology model 142, except that the final classification layer of the global ECG measurement model 143 is replaced by a regression layer 820. In addition, the global ECG measurement model 143 does not have the batch normalization layer of the ECG morphology model 142, since it is more useful for classification than regression. The global ECG measurement model 143 may receive as input a median beat 4 lead signal based on the lead set [I, II, v2, v4]. More specifically, the median beat generation module 133 may determine a median beat for each lead of the lead set [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 to output three intervals: PR interval, QRS duration, and QT interval, by minimizing the absolute error between the labeled (physician-estimated) intervals from the training data and the model-estimated intervals.

[0057] The global ECG measurement model 143 may also estimate global QRS onset / offset as shown in FIG. 4. The global ECG measurement model 143 may use the global QRS onset / offset and the algorithmically derived 12-lead set to determine per-lead ECG measurements. In a standard resting 12-lead analysis, in addition to the ECG global parameters described herein, per-lead ECG measurements are important. The per-lead measurements may include amplitude, duration, and polarity of all major ECG morphology components including P waves, QRS complexes, ST waves, and T waves. Since the morphology determination is primarily determined by the ECG morphology model 142, the final ECG morphology logic 152 may use per-lead ECG measurements to aid in the final classification, which to some extent mitigates the dependency on having accurate ECG measurements. In fact, the per-lead ECG measurements can help prevent some obvious classification errors when the ECG rhythm classification model 141, ECG morphology model 142, and global ECG measurement model 143 cannot capture them or generate multiple classifications from which one needs to be selected. For example, during morphology classification, if the ECG morphology model 142 classifies both LBBB and acute MI, the final ECG morphology logic 152 may make a final classification based on the QRS duration, ST segment, and possibly T waves. In another example, during ECG rhythm analysis, false positive classification of AFIB may be avoided by checking P-wave measurements including the PR interval and P-wave amplitude.

[0058] The final decision stage 150 may include logic for making correct exclusion / inclusion based on cardiology science and clinical practice. The final decision stage 150 may perform exclusion under various conditions. Both the ECG rhythm classification model 141 and the ECG morphology model 142 are trained for non-exclusive multi-class classification, so they may be able to identify / classify multiple abnormalities. In some cases, multiple abnormal conditions may be identified for the same ECG, such as ischemia and infarction, or bundle branch block and ischemia. However, in other cases, especially during ECG rhythm analysis, a particular abnormality may affect the detection of another abnormality, or they may be mutually exclusive, such as sinus rhythm and AFIB, or paced rhythm and sinus rhythm. The final decision stage 150 may operate in one of two different modes: a high specificity mode and a high sensitivity mode. The high specificity mode may include a higher threshold for the model output probability as well as more conditions based on measurements per lead. The high sensitivity mode may include a lower threshold for the model output probability (e.g., 70%). 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.

[0059] As discussed herein, the final decision 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 a dominant rhythm from the following list: paced rhythm, sinus rhythm / normal sinus rhythm, 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 to the final ECG rhythm logic 151 are the output of the ECG rhythm classification model 141, the global ECG measurements (PR interval, QRS duration, QT interval, and heart rate), and the RR interval sequence of the dominant beat. All decisions of the final ECG rhythm logic 151 are made in two modes, high specificity and high sensitivity, as discussed above.

[0060] 9 illustrates the final ECG rhythm logic 151 as various sets of conditions for finding a dominant rhythm. The final ECG rhythm logic 151 may determine a paced rhythm as the dominant rhythm if the output of the ECG rhythm classification model 141 indicates that the probability of a paced rhythm is high (e.g., higher than 0.8) and that the probability of a paced rhythm is higher than the probability of any other rhythm. In addition, the heart rate must be less than 80 bpm, the QRS duration must be less than 150 ms, and the RR variation must be small. If a paced rhythm is confirmed, the final ECG rhythm logic 151 may not perform any other checks for other rhythms.

[0061] The final ECG rhythm logic 151 may determine sinus rhythm as the dominant rhythm (i.e., confirm the sinus rhythm classification from the ECG rhythm classification model 141) if the output of the ECG rhythm classification model 141 indicates that the probability of sinus rhythm is high (e.g., higher than 0.7) or the probability of paced rhythm is lower than the probability of other rhythms. In addition, the PR interval must be greater than 80 ms or the probability of sinus rhythm must be very high (e.g., higher than 0.95). Here, a proper range of PR interval indicates that a valid P wave is present, which is an important feature of sinus rhythm. This final analysis reinforces 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, since a high specificity setting for another abnormal rhythm also means a higher sensitivity setting for sinus rhythm. To emphasize this high specificity, sinus rhythm is confirmed before other abnormal rhythms except paced rhythm. If the ECG rhythm classification model 141 outputs a very high probability of sinus / normal sinus rhythm, confirmation is provided without checking for other rhythms.

[0062] Final ECG rhythm logic 151 may determine Afib / aflutter as the dominant rhythm if the output of ECG rhythm classification model 141 indicates that the probability of Afib / aflutter is higher than other dominant rhythms (e.g., higher than 0.7) and the probability of paced rhythm is lower than other rhythms. In addition, the PR interval must be less than 30 ms and there must be no valid P waves. If the probability (Afib) is less than the probability (aflutter), final ECG rhythm logic 151 may determine that aflutter is the dominant rhythm.

[0063] If final ECG rhythm logic 151 excludes all other rhythms as the dominant rhythm, it may check whether the junctional rhythm is the dominant rhythm based on whether there is a high probability of a junctional rhythm and low RR variability from ECG rhythm classification model 141. If both of these conditions are met, final ECG rhythm logic 151 may determine that the junctional rhythm is the dominant rhythm.

[0064] If final ECG rhythm logic 151 still has not determined a dominant rhythm after checking for junctional rhythm, it may check for the presence of sinus rhythm using a more relaxed set of conditions. More specifically, final ECG rhythm logic 151 may lower the threshold for the probability of sinus rhythm to 0.6 and require that valid P waves be present and that all global measurements are within their normal ranges (e.g., PR[120, 200], QRS duration<115, QTcF<460 ms).

[0065] In some embodiments, if the final ECG rhythm logic 151 confirms sinus rhythm, it may attempt to determine that one or more associated arrhythmias are checked. More specifically, if there is a high probability of first degree atrioventricular conduction block (AVB) and a PR interval greater than 220 ms, the final ECG rhythm logic 151 may determine that a first degree AVB is present. If there is a high probability of second degree AVB, and the probability of second degree AVB is greater than the probability of third degree AVB, and the probability of second degree AVB is greater than the probability of other prominent arrhythmias, the final ECG rhythm logic 151 may determine that a second degree AVB is present. If there is a high probability of third degree AVB, and the probability of third degree AVB is greater than the probability of other arrhythmias, the final ECG rhythm logic 151 may determine that a third degree AVB is present. If there is a high probability of sinus bradycardia and the heart rate is less than 50 bpm, the final ECG rhythm logic 151 may determine that a sinus bradycardia is present. If a high probability of sinus tachycardia is present and the heart rate is greater than 100 bpm, the final ECG rhythm logic 151 may determine that sinus tachycardia is present. If a high probability of significant (severe) sinus arrhythmia is present (e.g., greater than 0.8), the final ECG rhythm logic 151 may determine that significant sinus arrhythmia is present.

[0066] The final ECG morphology logic 152 may include logic to augment the results of the morphology DNN model and improve the specificity of the determination of the primary morphological abnormality. The ECG morphology model 142 also operates 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 useful to focus on the primary cause of the abnormal condition rather than multiple causes. For example, in the interpretation of acute myocardial infarction (acute MI), adding ST segment changes to the probability of the DNN model is a way to improve specificity. However, if bundle branch block, especially left bundle branch block (LBBB), is confirmed, the ST segment deflection may be secondary to changes in depolarization, which may be difficult to distinguish from primary repolarization. Therefore, if LBBB or paced ECG is confirmed, the final determination of acute MI is not determined.

[0067] Final ECG morphology logic 152 may make a final morphology decision in the following order based on classifications from one or more of ECG rhythm classification model 141 and ECG morphology model 142: paced ECG, RBBB, LBBB, other intraventricular block, LVH, RVH, old MI, acute MI, other ST elevation, ischemia, QT prolongation, atrial enlargement, normal ECG. Each of the conditions discussed below is common in both high sensitivity and high specificity modes, with only the thresholds relatively different.

[0068] Final ECG morphology logic 152 may determine a paced ECG (i.e., confirm the paced ECG classification from ECG morphology model 142) if both ECG rhythm classification model 141 and ECG morphology model 142 indicate a very high probability of a paced ECG (e.g., greater than 0.95) and low RR beat-to-beat variability (RR_std / RR_mean less than 0.04). If final ECG morphology logic 152 confirms the paced ECG classification, it may skip performing any other morphology checks.

[0069] The final ECG morphology logic 152 may determine RBBB (i.e., confirm the RBBB classification from the ECG morphology model 142) 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. If the final ECG morphology logic 152 confirms the RBBB classification, it may skip performing any other morphology checks.

[0070] The final ECG morphology logic 152 may determine LBBB (i.e., confirm the LBBB classification from the ECG morphology model 142) 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. If the final ECG morphology logic 152 confirms the LBBB classification, it may skip performing any other morphology checks.

[0071] The final ECG morphology logic 152 may determine such a BBB (i.e., may confirm the classification of any other BBB from the ECG morphology model 142) if the probability of any other BBB as indicated by the classification of the ECG morphology model 142 is high (e.g., greater than 0.8) and the QRS duration is longer than 115 ms, or if the probability of such a BBB as indicated by the classification of the ECG morphology model 142 is greater than 0.90 and the QRS duration is longer than 110 ms.

[0072] Final ECG morphology logic 152 may determine LVH morphology (i.e., confirm the LVH classification from ECG morphology model 142) if the probability of LVH as indicated by the ECG morphology model 142 classification is very high (e.g., greater than 0.99), or if the probability of LVH as indicated by the ECG morphology model 142 classification is high (e.g., greater than 0.9) and any one of the V leads has a QRS deflection greater than 2500 μV or R_aVL is greater than 1000 μV (microvolts). It should be noted that the determination of LVH typically requires several ECG amplitude measurements. However, because only two precordial leads are used and the ECG is more aggressively low-pass filtered, traditional amplitude criteria for LVH are not directly applied. Instead, the output of ECG morphology model 142 is used in combination with a set of minimum amplitude and axis measurement checks. The final ECG morphology logic 152 may determine RVH morphology (i.e., confirm the RVH classification from the ECG morphology model 142) if the probability of LVH indicated by the classification of the ECG morphology model 142 is very high (e.g., greater than 0.9) and the QRS axis is greater than 90 degrees.

[0073] Final ECG morphology logic 152 may determine acute anterior MI morphology (i.e., confirm the acute anterior MI classification from ECG morphology model 142) if the probability of acute anterior MI indicated by the classification of ECG morphology model 142 is very high (e.g., greater than 0.95), at least one V lead has an ST greater than 200 μV and an ST / T ratio greater than 0.3, and at least one of the contralateral leads (limb leads) has an ST (ST depression) less than −50 μV. Alternatively, final ECG morphology logic 152 may determine acute anterior MI morphology if the probability of acute anterior MI is very high (e.g., greater than 0.99) and at least one V lead has an ST greater than 160 μV and an ST / T ratio greater than 0.3.

[0074] Final ECG morphology logic 152 may determine acute inferior MI morphology (i.e., confirm the acute inferior MI classification from ECG morphology model 142) if the probability of acute inferior MI as indicated by the classification of ECG morphology model 142 is very high (e.g., greater than 0.95) and at least one V lead has an ST greater than 100 μV and an ST / T ratio greater than 0.3. Alternatively, final ECG morphology logic 152 may determine acute inferior MI morphology if the probability of acute anterior MI is very high (e.g., greater than 0.99) and at least one inferior lead has an ST greater than 60 μV and an ST / T ratio greater than 0.3.

[0075] Final ECG morphology logic 152 may determine acute lateral MI morphology (i.e., confirm the acute lateral MI classification from ECG morphology model 142) if the probability of acute lateral MI as indicated by the classification of ECG morphology model 142 is very high (e.g., greater than 0.95) and at least one contralateral lead has an ST greater than 100 μV and an ST / T ratio greater than 0.3. Alternatively, final ECG morphology logic 152 may determine acute lateral MI morphology if the probability of acute anterior MI is very high (e.g., greater than 0.99) and at least one contralateral lead has an ST greater than 60 μV and an ST / T ratio greater than 0.3.

[0076] The final ECG morphology logic 152 may determine previous myocardial infarction (old MI) morphology (i.e., confirm the old MI classification from the ECG morphology model 142) if the probability of old MI indicated by the classification of the ECG morphology model 142 is very high (e.g., greater than 0.90) and significant Q waves are present.

[0077] The final ECG morphology logic 152 may determine ischemia morphology (i.e., confirm the ischemia classification from the ECG morphology model 142) if the probability of ischemia indicated by the classification of the ECG morphology model 142 is very high (e.g., greater than 0.90) and there is significant ST depression corresponding to the specification location calls in the lead groups, i.e., ST in anterior leads: V2, V4, inferior leads: II, III, aVF, contralateral leads: I, aVL.

[0078] Final ECG morphology logic 152 may determine QT prolongation morphology (i.e., confirm the QT prolongation classification from ECG morphology model 142) if the probability of QT prolongation indicated by the classification of ECG morphology model 142 is very high (e.g., greater than 0.90) and QTcF (Fridericia corrected QT interval) is greater than 460 ms. The only exclusion condition for QT prolongation is a paced ECG.

[0079] The final ECG morphology logic 152 may determine normal morphology (i.e., confirm the normal classification from the ECG morphology model 142) if the ECG morphology model 142 does not detect any of the abnormal morphology and rhythm conditions listed above, sinus rhythm is detected, the heart rate is greater than 50 bpm and less than 90 bpm, the QRS duration is less than 120 ms, the PR interval is less than 230 ms, the QTcF is less than 460 ms, and the P axis, QRS axis, and T axis are all greater than 0.

[0080] 10 is a flow diagram of a method 1000 for training and implementing an ML model for reduced lead set ECG interpretation according to some embodiments of the present disclosure. 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 operating / executed on a processing device), firmware (e.g., microcode), or a combination thereof. In some embodiments, method 1000 may be performed by a computing device (e.g., computing device 101 shown in FIG. 3).

[0081] In block 1005, the ECG rhythm classification model 141 may be trained using any sufficiently large training data set. For example, the training data set may include 1 million labeled ECGs with another 250,000 labeled ECGs used as an independent test set. In the example of FIG. 6A, 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 one with the smallest validation error. A training weight matrix is ​​used to balance the large difference in prevalence among all decisions. For decisions with lower prevalence, higher weights are used when the training error is backpropagated.

[0082] Referring also to FIG. 6, the convolutional layer 610 may utilize a residual block structure that utilizes a combination of convolutional and skip layers. Skip layers (also called skip connections or shortcut connections) allow the residual block in a deep learning model to address the problem 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, creating a "shortcut" for the gradient to flow through. As a result, different parts of the convolutional layer 610 are trained at different rates on different training data points based on how the error flows backward through the layers of the convolutional layer 610. FIG. 6 shows the residual block 610A, where the skip layer is used to skip the training of the convolutional layer 2 during the learning of the identity function X. ReLu may represent a nonlinear activation function.

[0083] In addition to the residual blocks, the ECG rhythm classification model 141 may include two dense layers 615 and one classification layer 620. In some embodiments, the ECG rhythm classification model 141 may further include batch normalization and dropout layers for all residual blocks and dense layers (not all batch normalization and dropout layers shown in FIG. 6) to improve its robustness or generalization once trained. The ECG rhythm classification model 141 is designed to classify rhythm decisions, an example of which is shown in Table 1 (below). The output of the classification layer 620 may have both positive and negative probabilities for each decision. In some embodiments, for each of the 8-lead ECG signals (downsampled to 150 Hz), 9.5 seconds of the 10 seconds of data may be used for input, and the remaining 0.5 seconds of data may be used for augmentation by a time shifting technique (random data shift). This random data shift is one of the data augmentation methods used to improve the generalization of a trained DNN model, making it less sensitive to the start and end times of the input ECG rhythm data.

[0084] In 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. In the example of FIG. 6A, 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 one with the smallest validation error. A morphology training weight matrix is ​​also used to balance the large difference in prevalence among all the decisions. For decisions with lower prevalence, higher weights are used when the training error is backpropagated. The ECG morphology model 142 is designed to classify ECG morphology decisions, 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 decision.

[0085] 7A, the ECG morphology model 142 may capture the morphology changes of each lead and the relative changes between leads, referred to as changes in the spatial domain. The ECG morphology model 142 may have a similar model structure as the ECG rhythm classification model 141, but has a smaller lead formation layer and less residual block layer 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 the median beats are relatively aligned around their QRS onset. Another important design consideration for the morphology model QRS region is that the variation along the time axis is relatively smaller than the rhythm signal.

[0086] To capture the spatial domain changes, the ECG morphology model 142 may form the ECG imaging by laying out all the leads in a two-dimensional 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 may process the two-dimensional ECG signal matrix with a two-dimensional convolution filter, as in the case of image classification. This method is more effective than applying multiple one-dimensional convolution filters to each lead individually, which is used by most other ECG training models. Since the ECG is a projection onto the torso of the process of electrical depolarization and repolarization of the heart, a spatial two-dimensional signal matrix can capture this projection more effectively. Therefore, a two-dimensional convolution mode is applied to this ECG image input.

[0087] At block 1015, a global ECG measurement model 143 may be trained. The global ECG measurement model 143 measures ECG global parameters including PR interval, QRS duration, QT interval, and heart rate. These parameters are referred to as ECG global measurements because they are not lead specific and are often estimated based on a combination of multiple leads. These parameters are often detected using complex signal processing and feature detection algorithms based on domain knowledge. However, an embodiment of the present disclosure estimates these ECG parameters using a deep learning model (global ECG measurement model 143) applied to a median beat 8-lead signal generated from an 8-lead ECG signal. Since the ECG global parameters are closely related to ECG morphology, the global ECG measurement model 143 may be trained to perform a 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 be able to also identify the absence of a valid P wave, such as in the case of atrial fibrillation.

[0088] FIG. 8 shows a detailed structure of the global ECG measurement model 143. The structure of the global ECG measurement model 143 is similar to that of the ECG morphology model 142, except that the final classification layer of the global ECG measurement model 143 is replaced by a regression layer 820. In addition, the global ECG measurement model 143 does not have the batch normalization layer of the ECG morphology model 142, since it is more useful for classification than regression. The global ECG measurement model 143 may receive as input a median beat 4 lead signal based on the lead set [I, II, v2, v4]. More specifically, the median beat generation module 133 may determine a median beat for each lead of the lead set [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 to output three intervals: PR interval, QRS duration, and QT interval, by minimizing the absolute error between the labeled (physician-estimated) intervals from the training data and the model-estimated intervals.

[0089] In block 1020, an ECG rhythm classification and an ECG morphology classification may be determined for the input ECG by ECG rhythm classification model 141 and ECG morphology model 142, respectively. In block 1025, global parameters for the input ECG may be determined by global ECG measurement model 143.

[0090] The final decision stage 150 may include a final ECG rhythm logic 151 and a final ECG morphology logic 152. In block 1030, the final ECG rhythm logic 151 may determine the dominant rhythm from the following list: paced rhythm, sinus rhythm / normal sinus rhythm, 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 to the final ECG rhythm logic 151 are the output of the ECG rhythm classification model 141, the global ECG measurements (PR interval, QRS duration, QT interval, and heart rate), and the RR interval sequence of the dominant beat. All decisions of the final ECG rhythm logic 151 are made in two modes, high specificity and high sensitivity, as discussed above.

[0091] Final ECG morphology logic 152 may make a final morphology decision in the following order based on classifications from one or more of ECG rhythm classification model 141 and ECG morphology model 142: paced ECG, RBBB, LBBB, other intraventricular block, LVH, RVH, old MI, acute MI, other ST elevation, ischemia, QT prolongation, atrial enlargement, normal ECG. Each condition discussed below is common in both high sensitivity and high specificity modes, with only the thresholds relatively different. Input parameters to final ECG morphology logic 152 are the output of ECG morphology model 142, the output of ECG rhythm classification model 141 (if applicable), and global ECG measurements (PR interval, QRS duration, QT interval, and heart rate) as follows:

[0092] FIG. 11 shows a diagrammatic representation of a machine in the exemplary form of a computer system 1100 having a set of instructions therein for causing the machine to execute any one or more of the methodologies discussed herein for performing reduced lead ECG interpretation.

[0093] In alternative embodiments, the machine may be connected (e.g., networked) to other machines in a local area network (LAN), an intranet, an extranet, or the Internet. The machine may operate as a server or client machine in a client-server network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. The machine may be a personal computer (PC), a tablet PC, a set-top box (STB), a personal digital assistant (PDA), a mobile phone, a web appliance, a server, a network router, a switch or bridge, a hub, an access point, a network access control device, or any machine capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by the machine. Furthermore, although only a single machine is shown, the term "machine" shall be construed to include any collection of machines that individually or jointly execute a set (or sets) of instructions to perform any one or more of the methodologies discussed herein. In one embodiment, the computer system 1100 may represent a server.

[0094] The exemplary computer system 1100 includes a processing device 1102, a main memory 1104 (e.g., read only memory (ROM), flash memory, dynamic random access memory (DRAM)), a static memory 1106 (e.g., flash memory, static random access memory (SRAM), etc.), and a data storage device 1118, which communicate with each other via a bus 1130. Any of the signals provided via the various buses described herein may be time multiplexed with other signals and may be provided via one or more common buses. Additionally, the interconnections between circuit components or blocks may be depicted as buses or single signal lines. Each of the buses may alternatively be a single or multiple single signal lines, and each of the single signal lines may alternatively be a bus.

[0095] The computing device 1100 may further include a network interface device 1108 that may communicate with a 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 audio 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).

[0096] The processing device 1102 represents one or more general-purpose processing devices, such as a microprocessor, a central processing unit, etc. More specifically, the processing device may 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 implementing other instruction sets, or a processor implementing a combination of instruction sets. The processing device 1102 may 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), a network processor, etc. The processing device 1102 is configured to execute instructions 1125 to perform the operations and steps discussed herein.

[0097] The data storage device 1115 may include a machine-readable storage medium 628 having stored thereon one or more sets of instructions 1125 (e.g., software) that embody any one or more of the functional methodologies described herein. The instructions 1125 may also reside, completely or at least partially, within the main memory 1104 or within the processing device 1102 during execution thereof by the computer system 1100, with the main memory 1104 and the processing device 1102 also constituting machine-readable storage media. The instructions 1125 may further be transmitted or received over a network 1120 via the network interface device 1108.

[0098] Although the machine-readable storage medium 1128 is shown to be a single medium in the exemplary embodiment, the term "machine-readable storage medium" should be interpreted to include a single medium or multiple media (e.g., centralized or distributed databases, or associated caches and servers) that store one or more sets of instructions. A machine-readable medium includes any mechanism for storing information in a form (e.g., software, processing application) readable by a machine (e.g., a computer). A machine-readable medium may include, but is not limited to, magnetic storage media (e.g., floppy diskettes), 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 another type of medium suitable for storing electronic instructions.

[0099] The foregoing description sets forth numerous specific details, such as examples of specific systems, components, methods, and the like, to provide a good understanding of some 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 instances, well-known components or methods have not been described in detail or have been presented in simple block diagrams to avoid unnecessarily obscuring the present disclosure. Thus, the specific details described are merely exemplary. Particular embodiments may differ from these exemplary details and still be considered to be within the scope of the present disclosure.

[0100] Additionally, some embodiments may be practiced in a distributed computing environment where the machine-readable medium is stored on or executed by two or more computer systems. In addition, information transferred between computer systems may be pulled or pushed over a communications medium connecting the computer systems.

[0101] Embodiments of the claimed subject matter include, but are not limited to, various operations described herein. These operations may be performed by hardware components, software, firmware, or combinations thereof.

[0102] Although the operations of the methods herein are shown and described in a particular order, the order of operations of each method may be changed such that certain operations may be performed in the reverse order, or such that certain operations may be performed at least in part concurrently with other operations. In alternative embodiments, instructions of separate operations or sub-operations may be intermittent or interleaved.

[0103] The above description of illustrated implementations of the invention, including those described in the Abstract, are not intended to be exhaustive or to limit the invention to the precise forms disclosed. Specific implementations of the invention and examples thereof have been described herein for illustrative purposes, but as one skilled in the art will recognize, various equivalent modifications are possible within the scope of the invention. The word "example" or "exemplary" is used herein to mean serving as an example, illustration, or illustration. Any aspect or design described herein as "example" or "exemplary" should not necessarily be construed as preferred or advantageous over other aspects or designs. Rather, use of the word "example" or "exemplary" is intended to present a concept in a concrete manner. 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 of the natural inclusive permutations. That is, "X includes A or B" is satisfied under any of the above examples if X includes A, if X includes B, or if X includes both A and B. In addition, the articles "a" and "an" used in this application and the appended claims should generally be construed to mean "one or more" unless otherwise specified or unless the context clearly directs to the singular form. Furthermore, the use of the terms "embodiment" or "one embodiment" or "implementation" or "one implementation" throughout is not intended to refer to the same embodiment or implementation unless so described. Furthermore, the terms "first", "second", "third", "fourth", etc. used herein are intended as labels to distinguish different elements and do not necessarily have a sequential meaning according to their numerical designations.

[0104] It will be appreciated that variations of the above-disclosed and other features and functions, or alternatives thereof, may be combined into many other different systems or applications. Various presently unforeseen or unanticipated alternatives, modifications, variations, or improvements may subsequently be made by those skilled in the art, which are also intended to be encompassed by the following claims. The claims may encompass embodiments in hardware, software, or combinations thereof. [Explanation of symbols]

[0105] 100 Systems 101 Local computing device, computing device 101A Mobile Software Applications, Mobile Apps 103 ECG Monitor 105 Views 110 Exploded View 113 Cloud Storage System 115 Single dipole heart model, heart model, processing device 120 Memory 120A Simplified Lead ECG Analysis Model, Interpretation Model 130 Data Adjustment Stage 131 Downsampling Module 132 Inductive Conversion Module 133 Median Beat Generation Module 140 Network, model analysis stage 141 ECG rhythm classification model, model 142 ECG morphology model, ECG morphology ML model, model 143 Global ECG Measurement Model, Model 150 Final decision stage 151 Final ECG Tuning Logic 152 Final ECG Morphology Logic 400 DNN 402 Input Data 403 Convolutional Layer 404 Trained Weights or Filters, Trained Weights 405 Convolutional Layer 405' Convolutional layer 405 (n-1) Convolutional Layer 406 Input Data 410 Weight 410' Weight 410 n Weight 411 Training Example 414 Output, Output or Prediction 414(*) Output or Prediction 416 Difference, Gain or Loss 500 Structure 505 Inductive cambium 510 Convolutional Layer 515 Dense layer 520 classification layer 610 Convolutional Layer 610A Residual Block 615 Dense layer 620 classification layer 628 Machine-readable storage medium 820 Recurrent Layer 1000 ways 1100 Computer systems, computing devices 1102 Processing device 1104 Main memory 1106 Static Memory 1108 Network Interface Device 1110 Video Display Unit 1112 Alphanumeric Input Device 1114 Cursor Control Device 1115 Data storage device 1116 Acoustic signal generating device 1118 Data storage device 1120 Network 1125 Command 1128 Machine-readable storage medium 1130 Bus

Claims

1. 1. A method comprising: training a first deep neural network (DNN) to classify electrocardiogram (ECG) rhythm features, where a training weight matrix is ​​used to balance prevalence differences between the classifications during training; training a second DNN to classify ECG morphological features; training a third DNN to identify global ECG parameters, where both the first DNN and the second DNN comprise a classification layer and the third DNN comprises a regression layer; Implementing the first DNN, the second DNN, and the third DNN as an ensemble DNN for ECG interpretation; A method comprising:

2. In response to receiving the ECG, classifying rhythm characteristics of the ECG using the first DNN; classifying morphological characteristics of the ECG using the second DNN; Identifying global parameters of the ECG using the third DNN Steps and applying a set of decision logic to the ECG rhythm characteristics and the ECG morphology characteristics to diagnose one or more disorders indicated by the ECG based at least in part on the ECG global parameters; The method of claim 1, further comprising:

3. the classification of the rhythm characteristics of the ECG includes a probability for each of a set of rhythms; the classification of the morphological characteristics of the ECG includes a probability for each of a set of morphologies. The method of claim 1.

4. the set of decision logic comprising: a first set of criteria for validating the classification of the rhythm characteristic of the ECG based at least in part on the global parameters of the ECG; a second set of criteria for validating the classification of the morphology characteristic of the ECG based at least in part on the global parameters and the validated classification of the rhythm characteristic of the ECG.

3. The method of claim 2, comprising:

5. 4. The method of claim 3, wherein one or more of the first DNN, the second DNN, and the third DNN comprise a convolutional neural network (CNN).

6. 2. The method of claim 1, wherein each of the first DNN, the second DNN, and the third DNN comprises a convolutional block having a residual block structure that combines a set of convolutional layers and a set of skip layers.

7. Using a random data shifting technique to further generalize each of the first DNN, the second DNN, and the third DNN.

6. The method of claim 5, further comprising:

8. 1. A system comprising: an electrocardiogram (ECG) monitor configured to perform an ECG on the person and transmit said ECG; 1. A computing device comprising: in response to receiving the ECG from the ECG monitor, classifying rhythm characteristics of the ECG using a first deep neural network (DNN); classifying morphological features of the ECG using a second DNN; identifying global parameters of the ECG using a third DNN; applying a set of decision logic to the ECG rhythm characteristics and the ECG morphology characteristics to diagnose one or more disorders indicated by the ECG based at least in part on the ECG global parameters; A computing device configured to A system comprising:

9. the processing device comprising: training the first deep neural network (DNN) to classify electrocardiogram (ECG) rhythm characteristics, wherein a training weight matrix is ​​used to balance prevalence differences between classifications during training; training a second DNN to classify ECG morphological features; training the third DNN to identify global ECG parameters, wherein both the first DNN and the second DNN comprise a classification layer and the third DNN comprises a regression layer; The system of claim 8 , further comprising:

10. the classification of the rhythm characteristics of the ECG includes a probability for each of a set of rhythms; the classification of the morphological characteristics of the ECG includes a probability for each of a set of morphologies. The system of claim 8.

11. the set of decision logic comprising: a first set of criteria for validating the classification of the rhythm characteristic of the ECG based at least in part on the global parameters of the ECG; a second set of criteria for validating the classification of the morphology characteristic of the ECG based at least in part on the global parameters and the validated classification of the rhythm characteristic of the ECG. The system of claim 8, comprising:

12. 11. The system of claim 10, wherein one or more of the first DNN, the second DNN, and the third DNN comprise a convolutional neural network (CNN).

13. 10. The system of claim 8, wherein the first DNN, the second DNN, and the third DNN each comprise a convolutional block having a residual block structure that combines a set of convolutional layers and a set of skip layers.

14. the processing device comprising:

13. The system of claim 12, further comprising: a random data shifting technique for further generalizing each of the first DNN, the second DNN, and the third DNN.

15. A non-transitory computer-readable medium having instructions stored thereon, the instructions, when executed by a processing device, causing the processing device to: training a first deep neural network (DNN) to classify electrocardiogram (ECG) rhythm characteristics; training a second DNN to classify ECG morphological features; training a third DNN to identify global ECG parameters; In response to receiving the ECG, classifying rhythm characteristics of the ECG using the first DNN; classifying morphological characteristics of the ECG using the second DNN; Identifying global parameters of the ECG using the third DNN And, applying a set of decision logic to the ECG rhythm characteristics and the ECG morphology characteristics to diagnose one or more disorders indicated by the ECG based at least in part on the ECG global parameters; A non-transitory computer-readable medium for causing

16. the classification of the rhythm characteristics of the ECG includes a probability for each of a set of rhythms; the classification of the morphological characteristics of the ECG includes a probability for each of a set of morphologies.

16. The non-transitory computer-readable medium of claim 15.

17. the set of decision logic comprising: a first set of criteria for validating the classification of the rhythm characteristic of the ECG based at least in part on the global parameters of the ECG; a second set of criteria for validating the classification of the morphology characteristic of the ECG based at least in part on the global parameters and the validated classification of the rhythm characteristic of the ECG.

16. The non-transitory computer-readable medium of claim 15, comprising:

18. 20. The non-transitory computer-readable medium of claim 17, wherein one or more of the first DNN, the second DNN, and the third DNN comprise a convolutional neural network (CNN).

19. 16. The non-transitory computer-readable medium of claim 15, wherein each of the first DNN, the second DNN, and the third DNN comprises a convolutional block having a residual block structure that combines a set of convolutional layers and a set of skip layers.

20. the processing device comprising:

20. The non-transitory computer-readable medium of claim 17, further comprising: using a random data shifting technique to further generalize each of the first DNN, the second DNN, and the third DNN.

Images