Systems and methods for reduced-lead electrocardiogram diagnosis using deep neural networks and rule-based systems

Abstract

A method and system for automatically diagnosing patients based on reduced-lead electrocardiograms (ECGs) using one or more deep neural networks are provided. In one embodiment, a method for automatically diagnosing patients using reduced-lead ECGs includes acquiring reduced-lead ECG data, wherein the reduced-lead ECG data includes fewer than twelve lead signals; determining the type of each of the fewer than twelve lead signals; selecting a deep neural network based on the type of each of the fewer than twelve lead signals; and mapping the fewer than twelve lead signals to a diagnosis using the deep neural network. In this way, the reduced-lead ECG data can be mapped to a diagnosis using an intelligently selected deep neural network, wherein the deep neural network is trained on reduced-lead ECG data including a set of ECG lead types identical to the acquired reduced-lead ECG data.

Description

Technical Field

[0001] The embodiments of the subject matter disclosed herein relate to the analysis of electrocardiograms (ECGs), and more specifically, to methods and systems for the automatic diagnosis of ECGs using deep neural networks and rule-based systems. Background Technology

[0002] Heart disease has become the most common disease affecting people worldwide. Millions die from heart attacks each year, and an equal number undergo coronary artery bypass grafting or balloon angioplasty for advanced heart disease. Early detection and timely treatment can reduce the probability and / or severity of such events. Early detection can improve quality of life and slow the progression of heart failure. Recording an electrocardiogram (ECG) is a non-invasive method for assessing a patient's heart condition, enabling the early detection of cardiac abnormalities. The characteristics of the ECG allow for relatively accurate and rapid diagnosis.

[0003] ECG can provide information about the normal and / or pathophysiology of the heart. Routinely, ten electrodes placed in contact with the patient's skin are used to measure twelve leads. These twelve leads provide information about the electrical potential passing through the heart along twelve different axes intersecting the heart, and the changes in the intensity of these potentials over time provide twelve different half-cycle signals. These twelve signals can then be used to assess the state of the heart and may diagnose one or more cardiac conditions, such as arrhythmias, myocardial infarction, cardiac hypertrophy, etc.

[0004] However, in some cases, it may be desirable to use fewer than twelve leads to assess a patient's cardiac physiological state. The acquisition of a twelve-lead signal relies on placing six electrodes on the patient's chest and one electrode on each of the patient's limbs. However, in some situations, placing each of the ten electrodes on the patient may be undesirable. For example, after surgery, the patient's limbs or chest may have bandages or sutures that hinder electrode placement. In another example, it may be difficult to place each of the six chest electrodes on a newborn because their chest area may be relatively small compared to an adult patient, or it may be uncomfortable / irritating for the newborn. In these cases, it may be desirable to use fewer than twelve leads to monitor the patient's cardiac state.

[0005] Conventional diagnostic methods / algorithms may rely on a 12-lead ECG and may produce inconsistent diagnoses or fail to provide a definitive diagnosis when applied to a reduced-lead ECG (as used herein, a reduced-lead ECG will be understood as including an ECG with one to eleven of the standard twelve leads / signals). One attempt to apply conventional diagnostic methods to a reduced-lead ECG involves expanding the reduced-lead ECG to a standard 12-lead signal by simulating one or more missing lead signals, thereby mimicking a standard full ECG (as used herein, a full ECG refers to a standard 12-lead ECG). Conventional diagnostic methods can then be applied to the simulated full ECG.

[0006] However, the inventors of this paper have recognized the problems with the above-mentioned methods. For example, simulated ECG leads may have varying degrees of accuracy, and therefore may produce inconsistent results when used to diagnose heart conditions using standards established based on a standard twelve-lead ECG. Furthermore, simulating ECG leads can consume significant computational resources. Therefore, there is a general need to explore techniques for increasing the diagnostic accuracy of ECG diagnostic systems using reduced-lead ECG data while reducing reliance on ECG lead simulation. Summary of the Invention

[0007] This disclosure at least partially addresses the aforementioned problems. In one embodiment, a diagnostic for a reduced-lead ECG can be generated without simulating missing leads by a method comprising the following steps: acquiring reduced-lead ECG data, wherein the reduced-lead ECG data includes fewer than twelve lead signals; determining the type of each of the fewer than twelve lead signals; selecting a deep neural network based on the type of each of the fewer than twelve lead signals; and mapping the fewer than twelve lead signals to a diagnostic using the deep neural network. In this way, a diagnostic for reduced-lead ECG data can be determined by intelligently selecting a deep neural network trained on reduced-lead ECG data that includes ECG leads of the same type as the acquired reduced-lead ECG data without simulating one or more missing lead signals.

[0008] The above-described advantages, as well as other advantages and features, of this specification will become apparent, either alone or in connection with the accompanying drawings, from the following detailed description. It should be understood that the above summary is provided to present a simplified version of the selected concepts further described in the detailed description. This is not intended to identify key or essential features of the claimed subject matter, the scope of which is uniquely defined by the claims following the detailed description. Furthermore, the claimed subject matter is not limited to embodiments that address any of the disadvantages mentioned above or in any part of this disclosure. Attached Figure Description

[0009] A better understanding of the various aspects of this disclosure can be achieved by reading the following detailed description and referring to the accompanying drawings, in which:

[0010] Figure 1 A block diagram of an exemplary implementation of an ECG system is shown;

[0011] Figure 2 This is a schematic diagram illustrating a first embodiment of a hybrid ECG diagnostic system;

[0012] Figure 3 This is a schematic diagram illustrating a second embodiment of the hybrid ECG diagnostic system;

[0013] Figure 4 This is a schematic diagram illustrating a third embodiment of the hybrid ECG diagnostic system;

[0014] Figure 5 This is a high-level flowchart illustrating a method for automatically diagnosing ECG data using a hybrid system;

[0015] Figure 6 This is a flowchart illustrating a first embodiment of a method for automatically diagnosing ECG data using a hybrid ECG diagnostic system;

[0016] Figure 7 This is a flowchart illustrating a second embodiment of a method for automatically diagnosing ECG data using a hybrid ECG diagnostic system;

[0017] Figure 8 This is a flowchart illustrating a third embodiment of a method for automatically diagnosing ECG data using a hybrid ECG diagnostic system;

[0018] Figure 9 This is a flowchart illustrating a method for training a hybrid ECG diagnostic system to automatically diagnose ECG data according to an exemplary embodiment;

[0019] Figure 10 An example ECG according to an exemplary embodiment of the present disclosure is shown, a first diagnosis of the ECG generated by a rule-based system, and a second diagnosis of the ECG generated by a hybrid ECG diagnostic system;

[0020] Figure 11 This is a schematic diagram illustrating a reduced-lead ECG diagnostic system according to an exemplary embodiment;

[0021] Figure 12 This is a flowchart illustrating a method for training a reduced-lead ECG diagnostic system to automatically diagnose reduced-lead ECG data according to an exemplary embodiment;

[0022] Figure 13 This is a flowchart illustrating a method for diagnosing reduced-lead ECG data using a reduced-lead ECG diagnostic system.

[0023] These accompanying figures illustrate specific aspects of the described system and method for automatically diagnosing reduced-lead ECG data using a reduced-lead ECG diagnostic system. Together with the following description, the figures illustrate and explain the structures, methods, and principles described herein. In these figures, the dimensions of components may be enlarged or otherwise modified for clarity. Well-known structures, materials, or operations are not shown or described in detail to avoid obscuring aspects of the described components, systems, and methods. Detailed Implementation

[0024] The following description relates to systems and methods for automatically diagnosing electrocardiograms (ECGs) using a hybrid system comprising one or more deep neural networks and rule-based systems (such as the rule-based systems described in Rawi, Atiaf Ayal, and Muzhir Shaban Al-Ani. “A rule-based expert system for automated ECG diagnosis.” International Journal of Advances in Engineering & Technology 6.4 (2013):1480, the entire contents of which are incorporated herein by reference for all purposes). Specifically, the following description relates to systems and methods for increasing the adaptability and accuracy of ECG diagnostic systems by combining one or more deep neural networks with rule-based systems to produce a hybrid system that increases diagnostic accuracy and adaptability to specific use cases. This hybrid system can also achieve accurate diagnosis of reduced-lead ECG data by learning to directly map features of reduced-lead ECG data to a diagnosis without simulating signals from missing leads.

[0025] In some implementation schemes, it can be via Figure 1 The electrocardiogram (ECG) shown is obtained from the patient by an ECG system 100. The obtained ECG data can be automatically diagnosed by an ECG processing system 102, which implements one or more hybrid ECG diagnostic systems, such as those used in… Figure 2 , Figure 3 and Figure 4 One or more of the hybrid ECG diagnostic systems 200, 300, and 400 shown are illustrated. The ECG processing system 102 may implement one or more hybrid ECG diagnostic systems to perform... Figure 5 Method 500 shown Figure 6 Method 600 shown Figure 7 The method 700 and / or shown Figure 8 One or more operations of the illustrated method 800 are used to diagnose ECG data. In a first embodiment, the ECG processing system 102 may implement a hybrid ECG diagnostic system 200 to use... Figure 6 One or more operations of the method 600 shown are used to automatically diagnose acquired ECG data. In a second embodiment, the ECG processing system 102 may implement a hybrid ECG diagnostic system 300 to use... Figure 7 One or more operations of the method 700 shown are used to automatically diagnose acquired ECG data. In a third embodiment, the ECG processing system 102 may implement a hybrid ECG diagnostic system 400 to use... Figure 8 One or more operations of the method 800 shown are used to automatically diagnose acquired ECG data. Figure 10 An embodiment is shown that can be automatically diagnosed using the above systems and methods for a 12-lead ECG, as well as a first erroneous diagnosis generated using a separate rule-based system and a second correct diagnosis generated using a hybrid ECG diagnostic system according to an embodiment of this disclosure. Figure 9 The method 900 shown trains hybrid ECG diagnostic systems 200, 300, and 400 so that the hybrid ECG diagnostic systems can be adapted to specific use cases by training on use case-specific data.

[0026] Furthermore, this disclosure provides access to deep neural network libraries (such as...) Figure 11 The deep neural network library 1104 shown selects a deep neural network (such as deep neural network 1110) to automatically diagnose reduced-lead ECGs. The selected deep neural network 1110 may include a deep neural network trained on reduced-lead ECG data and can be used for... Figure 2 The hybrid ECG diagnostic system 200 shown Figure 3 The hybrid ECG diagnostic system 300 and / or shown Figure 4 The hybrid ECG diagnostic system 400 shown is used to perform... Figure 13 The method 1300 shown illustrates one or more operations to diagnose reduced lead ECG data. The deep neural network library 1104 may include multiple different deep learning networks, depending on... Figure 12The illustrated method 1200 trains each deep learning network using different reduced lead groups. In one embodiment, a first trained deep neural network in deep neural network library 1104 may include parameters learned via training on reduced lead ECG data, wherein the reduced lead ECG data includes lead types I, II, III, aVR, aVL, and aVF (limb leads), but lacks lead types V1, V2, V3, V4, V5, and V6 (chest leads). Therefore, the first trained deep neural network may be selected from deep neural network library 1104 to map acquired reduced lead ECG data to a diagnosis, wherein the acquired reduced lead ECG data includes the same lead types (leads I, II, III, aVR, aVL, and aVF) on which the first trained deep neural network is trained. In other words, by training multiple deep neural networks using different permutations of reduced lead groups (where there are 4,094 different permutations of reduced lead groups for a standard 12-lead ECG), and storing the trained deep neural networks in a deep neural network library based on the reduced lead ECG data on which the deep neural networks were trained, and by selecting a deep neural network from the library of deep neural networks trained on reduced lead ECG data of the same type of leads as the acquired data, acquired ECG data with one or more missing leads can be effectively and accurately mapped to a diagnosis.

[0027] As used herein, lead type (also known as ECG lead type) refers to one of the twelve or fifteen conventionally measured ECG leads (I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, V6, V7, V8, and V9). In other words, a conventional twelve- or 15-lead ECG includes one signal / lead from each of the twelve or 15 different lead types. The term "lead type" is used to distinguish an ECG measurement (lead type) taken along a specific axis through the heart from a specific instance (lead) of the ECG measurement. For example, a training dataset may include thousands of different ECG leads (e.g., from thousands of different patients, or from hundreds of patients with repeated measurements), but a training dataset may include only six lead types, such as V1, V2, V3, V4, V5, and V6. In other words, each of the thousands of leads in the training dataset has a type V1, V2, V3, V4, V5, or V6, and no lead in the training dataset has a type I, II, III, aVR, aVL, or aVF.

[0028] Furthermore, as used herein, complete lead set, complete ECG lead set, complete 12-lead ECG, complete ECG, and other similar terms refer to an ECG comprising twelve different leads (each of the twelve lead types). For example, an ECG obtained from a patient comprising type I, type II, type III, type aVR, type aVL, type aVF, type V1, type V2, type V3, type V4, type V5, and type V6 leads may be referred to as a complete ECG, 12-lead ECG, complete 12-lead ECG, etc. Conversely, the terms reduced lead set, reduced lead ECG data, reduced lead data, reduced ECG lead set, and other similar terms will be understood to refer to an ECG comprising no more than eleven and no less than one lead (each of the different lead types).

[0029] As used herein, the term preliminary diagnosis should be understood as a diagnosis or diagnostic prediction by a deep neural network or rule-based system that is not used as the final diagnosis on its own. Instead, it is input into one or more additional deep neural networks / rule-based systems to provide additional information to the system, thereby facilitating a more accurate prediction of the final diagnosis (the terms final diagnosis and diagnosis are used interchangeably herein). In other words, the preliminary diagnosis may be generated by the hybrid system for use within the hybrid system, while the diagnosis (also known as the final diagnosis) may be output to a user, such as via a display device, or transmitted to one or more client devices.

[0030] As used herein, the terms final diagnosis, ECG diagnosis, diagnosis, and other similar terms will be understood to include an assessment of the normality and / or abnormality of one or more features of the ECG. Specifically, an ECG diagnosis may include evaluating one or more deviations of the acquired ECG from a “baseline” or “healthy” ECG and may indicate one or more types of arrhythmias, and may further indicate one or more potential expected causes of one or more types of arrhythmias. In some embodiments, the diagnosis may include a feature-by-feature evaluation of the ECG, indicating how / whether each feature deviates from the baseline ECG (e.g., normal sinus rhythm, QRS widening, abnormal repolarization, right axis deviation, left bundle branch block, etc.).

[0031] refer to Figure 1An electrocardiogram (ECG) system 100 according to an exemplary embodiment is illustrated. The ECG system 100 includes a set of electrodes 150 attached to a patient 140. The electrodes 150 are electrically coupled to a data acquisition module 118, thus enabling the data acquisition module 118 to measure ECG waveform data by determining the potential difference between two or more electrodes of the electrodes 150. The data acquisition module 118 is communicatively coupled to an ECG processing system 102 for processing, storing, and / or diagnosing the ECG data acquired by the data acquisition module 118. The ECG processing system 102 is further communicatively coupled to a display device 114 and a user input device 116, the display device being configured to display ECG data and / or diagnoses determined based on the ECG data, and the user input device enabling a user to input data into the ECG processing system 102 or interact with data in a non-transitory memory 106. Furthermore, the ECG system 100 may be communicatively coupled to one or more client devices (not shown), such as via a network connection or via the Internet.

[0032] Data acquisition module 118 is configured to acquire the time trajectory of the potential between two or more electrodes in electrode 150. In ECG system 100, electrode 150 includes ten electrodes configured to measure a twelve-lead ECG. In some embodiments, ECG system 100 may include 13 electrodes and may be configured to acquire fifteen-lead ECG data. Although those skilled in the art will well understand the meaning of a twelve-lead ECG, in short, a twelve-lead ECG includes a record of the cardiac potential as a function of time, measured along twelve different axes intersecting the heart. Thus, a twelve-lead ECG includes twelve different time-varying signals, each of which represents cardiac electrical activity measured along a different axis. Electrode 150 may include a conductive gel that contacts the patient's skin and conducts electrical signals present at the skin to the electrode. The heart of patient 140 generates an electrical signal referred to as an ECG waveform, and this electrical signal may also be referred to herein as an ECG signal, ECG lead signal, ECG, or ECG data.

[0033] Specifically, electrode 150 includes four limb electrodes: electrode 120 (and thus commonly referred to as RA) placed on the right arm of patient 140, electrode 122 (and thus commonly referred to as RL) placed on the right leg of patient 140, electrode 138 (and thus commonly referred to as LA) placed on the left arm of patient 140, and electrode 124 (and thus commonly referred to as LL) placed on the left leg of patient 140. The four limb electrodes can be used to measure cardiac potential along six different axes intersecting the heart, each potential measured along a different axis being called a lead. The four limb electrodes are configured to measure six leads referred to as I (which includes the potential difference between electrode 138 and electrode 120), II (which includes the potential difference between electrode 124 and electrode 120), III (which includes the potential difference between electrode 124 and electrode 138), aVR (which includes the potential difference between electrode 120 and the average of electrodes 124 and 138), aVL (which includes the potential difference between electrode 138 and the average of electrodes 120 and 124), and aVF (which includes the potential difference between electrode 124 and the average of electrodes 120 and 138).

[0034] In addition, electrode 150 includes six chest electrodes: electrode 126 (commonly referred to as V1), electrode 128 (commonly referred to as V2), electrode 130 (commonly referred to as V3), electrode 132 (commonly referred to as V4), electrode 134 (commonly referred to as V5), and electrode 136 (commonly referred to as V6). The six chest electrodes, in combination with the limb electrodes, are configured to measure potential via six different axes intersecting the heart in a horizontal plane (i.e., a plane perpendicular to the plane in which the six limb leads are measured). Specifically, the six chest electrodes are configured to measure six leads referred to as V1 (which includes the potential difference between electrode 126 and the average of electrodes 120, 124 and 138), V2 (which includes the potential difference between electrode 128 and the average of electrodes 120, 124 and 138), V3 (which includes the potential difference between electrode 130 and the average of electrodes 120, 124 and 138), V4 (which includes the potential difference between electrode 132 and the average of electrodes 120, 124 and 138), V5 (which includes the potential difference between electrode 134 and the average of electrodes 120, 124 and 138), and V6 (which includes the potential difference between electrode 136 and the average of electrodes 120, 124 and 138).

[0035] Although ECG system 100 includes ten electrodes and is configured to measure a complete twelve-lead ECG, it should be understood that this disclosure provides ECG systems including more or fewer than ten electrodes and / or ECGs including more or fewer than twelve leads. This disclosure also provides electrode placements different from those described for the reference electrodes 120-138 above. Specifically, portions of this disclosure relate to using a reduced ECG lead set to improve diagnostic accuracy, which can be generated by an ECG system having fewer than ten electrodes. In some embodiments, an ECG system including four limb electrodes and no chest electrodes can record six limb leads but not six chest leads. In some embodiments, an ECG system including four limb electrodes and two chest electrodes (V1 and V5) can record six limb leads and two of the six chest leads (V1 and V5), but cannot measure the other four chest leads (V2, V3, V4, and V6). In some implementations, an ECG system that includes four limb electrodes and two chest electrodes (V2 and V5) can record six limb leads and two of the six chest leads (V2 and V5), but cannot measure the other four chest leads (V1, V3, V4, and V6).

[0036] ECG data acquired by data acquisition module 118 can be transmitted to ECG processing system 102 for storage and processing (signal filtering, normalization, noise suppression, etc.). ECG data processing system 102 can be further configured to automatically diagnose the ECG data acquired by data acquisition module 118 by performing one or more operations of one or more methods disclosed herein. ECG processing system 102 includes processor 104 and nontransitory memory 106, wherein processor 104 can read instructions from nontransitory memory 106 to perform one or more operations of one or more methods stored therein. ECG processing system 102 is further communicatively coupled to display device 114 and user input device 116, which respectively enable a user to view and interact with data within ECG processing system 102.

[0037] ECG processing system 102 includes processor 104 configured to execute machine-readable instructions stored in non-transitory memory 106. Processor 104 may be single-core or multi-core, and programs executing on it may be configured for parallel or distributed processing. In some embodiments, processor 104 may optionally include individual components distributed across two or more devices that may be remotely located and / or configured for coordinated processing. In some embodiments, one or more aspects of processor 104 may be virtualized and executed by remotely accessible networked computing devices configured in a cloud computing configuration.

[0038] Non-transitory memory 106 may store deep neural network module 108, rule-based system module 110, and ECG data module 112. Deep neural network module 108 may further include deep neural network library module 180, which may store one or more trained deep neural networks, and each trained deep neural network in deep neural network library module 180 may include multiple parameters (including weights, biases, and activation functions). Deep neural network module 108 may also include instructions for implementing one or more trained deep neural networks to automatically diagnose ECG data, alone or in combination with a rule-based system stored in rule-based system module 110, by mapping ECG data to diagnostic or preliminary diagnoses using learned parameters. For example, deep neural network module 108 and rule-based system module 110 may store instructions for implementing hybrid ECG diagnostic systems (such as...). Figure 2 The hybrid ECG diagnostic system 200 shown Figure 3 The hybrid ECG diagnostic system 300 and / or shown Figure 4 The instructions of the hybrid ECG diagnostic system 400 shown are as follows. The deep neural network module 108 may include trained and / or untrained neural networks, and may also include various data or metadata related to one or more neural networks stored therein.

[0039] The deep neural network module 108 may also include instructions for training one or more deep neural networks stored therein. The deep neural network module 108 may include instructions that, when executed by the processor 104, cause the ECG processing system 102 to perform one or more operations of methods 900 and / or 1200, discussed in detail below. In some embodiments, the deep neural network module 108 includes instructions for backpropagating the error determined via a loss function to adjust the parameters of one or more deep neural networks by implementing one or more gradient descent algorithms.

[0040] In some embodiments, the deep neural network module 108 includes instructions for intelligently selecting training data pairs from the ECG data module 112. In some embodiments, the training data pairs include corresponding ECG data and a baseline truth diagnostic pair. In some embodiments, the deep neural network module 108 includes instructions for selectively removing one or more lead types from the ECG data of the training data pair to generate reduced-lead ECGs and corresponding baseline truth diagnostics, and for training the deep neural network to map features of the reduced-lead ECG data to the diagnostics without simulating one or more missing lead types. In some embodiments, the deep neural network module 108 is not located at the ECG processing system 102.

[0041] ECG processing system 102 also includes a rule-based system module 110 that stores instructions for implementing a rule-based system to diagnose ECG data based on hard-coded cardiologist-inspired / standard diagnostic data. In some embodiments, rule-based system module 110 includes instructions for extracting multiple predetermined features from ECG data and feeding the extracted features into a decision tree, wherein the decision tree includes multiple branch decision points based on the extracted features. Rule-based system module 110 can work in conjunction with deep neural network module 108 to implement a hybrid ECG diagnostic system, such as... Figure 2 , Figure 3 and Figure 4 Hybrid ECG diagnostic systems 200, 300, and 400 are shown in the diagram. In some embodiments, a preliminary diagnosis generated by the rule-based system of rule-based system module 110 can be fed into a deep neural network of deep neural network module 108, and the deep neural network can map the preliminary diagnosis to a final diagnosis. In some embodiments, the deep neural network may include a convolutional neural network and a decision network (including a multilayer perceptron), and the preliminary diagnosis generated by the rule-based system can be input into the input layer of the decision network. The decision network then maps the preliminary diagnosis to the final diagnosis by propagating the preliminary diagnosis through multiple densely connected layers until it reaches the output layer. The output layer can then output a probability score associated with each of the multiple diagnoses, and the diagnosis with the highest probability score can be displayed to the user and / or transmitted to a client device communicatively connected to ECG processing system 102. In some embodiments, the preliminary diagnosis generated by the deep neural network of deep neural network module 108 can be input into the rule-based system of rule-based system module 110 for use by the rule-based system to determine the final diagnosis.

[0042] The non-transitory storage 106 also includes an ECG data module 112, which includes ECG data acquired by one or more data acquisition modules or an ECG system. In some embodiments, the ECG data module 112 includes data acquired by the ECG system 100. In some embodiments, the ECG data module 112 may store ECG data acquired via a communicative connection to one or more data sources outside of the ECG processing system 102. The ECG data stored within the ECG data module may be organized according to one or more organizational schemes or configured as one or more data structures known in the field of data storage. In some embodiments, ECG data may be stored in the ECG data module by indexing the ECG data according to lead groups, thereby enabling rapid retrieval of ECG data by lead groups. For example, a query for ECG data including six chest leads may be performed by the ECG processing system 102 by evaluating the index of each ECG dataset included therein and returning each ECG dataset indicating that the ECG dataset includes six chest leads.

[0043] ECG data module 112 may also store multiple ECG data sets with multiple corresponding baseline truth diagnoses, wherein the baseline truth diagnoses are generated by an expert cardiologist based on the multiple ECG data sets. In some embodiments, ECG data module 112 may include multiple training data pairs, wherein each training data pair includes ECG data (including a full twelve-lead ECG or a reduced-lead ECG) and a corresponding expert-generated baseline truth diagnoses for the ECG. The training data stored within ECG data module 112 may be used as deep neural network training routines (such as the following). Figure 9 and Figure 12 A portion of the methods described in methods 900 and 1200 are queried and used by the deep neural network module 108.

[0044] Temporarily transferred to Figure 10 An exemplary implementation of ECG data that can be acquired by ECG system 100 via data acquisition module 118 and electrode 150 and stored in ECG data module 112 is shown. Figure 10The diagram shows an electrocardiogram (ECG) 1002, comprising a twelve-lead ECG with twelve different waveforms / signals representing each of the twelve lead types. The ECG waveform of a single heartbeat is often referred to as the PQRST complex because it includes the P wave, Q wave, R wave, S wave, and T wave. The ECG waveform may also include the U wave, but this may not be present in all ECGs. The P wave appears at the onset of a heartbeat, corresponding to atrial activity, while the QRST complex follows the P wave, corresponding to ventricular activity. The QRS component represents the electrical activation of the ventricles, while the T wave represents their electrical recovery. The ST segment is a relatively stationary period. In humans, the T wave has been found to be an appropriate interval for detecting alternation. That is, the level of variation in the T wave during alternating heartbeats is a good indicator of the patient's cardiac electrical stability. In some implementations, the characteristics of the ECG waveform that can be analyzed to diagnose an ECG include the amplitude of each of the P, Q, R, S, T, and U waves, and the time between one or more of the P, Q, R, S, T, and U waves, referred to as an interval or segment. For example, the time interval between the start of a P wave and the start of a QRS complex is called the PR interval and can be used to diagnose the physiological condition of the heart. In another example, the duration between the start of a Q wave and the end of a T wave (called the QT interval) can be used to diagnose the physiological condition of the heart. Other diagnostic features of ECG include heart rate, which can be determined by the intervals between heartbeats (such as the QQ interval, TT interval, etc.) and the variations in the amplitude and / or intervals of the P, Q, R, S, and T waves from heartbeat to heartbeat.

[0045] In some embodiments, the nontransitory memory 106 may include components disposed on two or more devices that can be remotely located and / or configured for coordinated processing. In some embodiments, one or more aspects of the nontransitory memory 106 may include a remotely accessible networked storage device configured in a cloud computing configuration.

[0046] ECG system 100 may also include user input device 116. User input device 116 may include one or more of the following: touchscreen, keyboard, mouse, touchpad, motion-sensing camera, or other devices configured to enable a user to interact with and manipulate data within ECG processing system 102. For example, user input device 116 may enable a user to select ECG data for diagnostics.

[0047] Display device 114 may include one or more display devices utilizing virtually any type of technology. In some embodiments, display device 114 may include a computer monitor and may display unprocessed and processed ECG data. Display device 114 may be combined with processor 104, nontransitory memory 106 and / or user input device 116 in a shared housing, or it may be a peripheral display device and may include a monitor, touch screen, projector or other display devices known in the art that enable a user to view ECG data and / or interact with various data stored in nontransitory memory 106.

[0048] It should be understood that Figure 1 The ECG system 100 shown is for illustration and not limitation. Another suitable ECG processing system may include more, fewer, or different components.

[0049] Go to Figure 2 The diagram illustrates a first embodiment of a hybrid ECG diagnostic system 200. In some embodiments, the hybrid ECG diagnostic system 200 may be implemented by an ECG processing system 102, an edge device connected to the ECG processing system 102, a cloud device communicating with the ECG processing system 102, or any suitable combination thereof.

[0050] The hybrid ECG diagnostic system 200 includes a deep neural network 210 and a rule-based system 204, configured to receive ECG data 202 and generate diagnostics 216 based on the received ECG data. In some embodiments, the hybrid ECG diagnostic system may be configured to further receive auxiliary input (not shown). In some embodiments, the ECG data 202 may include a real-time stream of ECG data acquired from an ECG system (such as ECG system 100). In some embodiments, the ECG data 202 may include pre-recorded ECG data that may be transferred to the hybrid ECG diagnostic system 200 from a communicatively linked data storage device (such as ECG data module 112 in non-transitory memory 106). In some embodiments, the ECG data 202 may include a twelve-lead ECG. In some embodiments, the ECG data 202 may include fifteen-lead data. In some embodiments, the ECG data 202 may be a reduced-lead ECG including fewer than twelve leads but more than zero leads.

[0051] In the hybrid ECG diagnostic system 200, ECG data 202 is first fed into a deep neural network 210. In some embodiments, the deep neural network 210 comprises two deep neural networks connected in series: the first is a convolutional neural network 212, including multiple convolutional layers, and the second is a decision network 214, including multiple densely connected layers (therefore, the decision network 214 may also be referred to as a multilayer perceptron or MLP). The output from the output layer of the convolutional neural network 212 is received at the input layer of the decision network 214 and mapped to multiple probability scores for each of multiple predetermined diagnoses. In some embodiments, the lead sets may have been obtained from a deep neural network library (such as...) based on the lead sets used to train the deep neural network 210 to match the lead sets in the ECG data 202. Figure 11 The deep neural network 210 is intelligently selected in the deep neural network library 1104 shown. In some implementations, it may have been based on... Figure 13 One or more operations of the method 1300 shown are selected from the deep neural network 210.

[0052] A convolutional neural network 212 is configured to receive ECG data 202 at its input layer. In some embodiments, the ECG data 202 may undergo one or more preprocessing operations before being input into the convolutional neural network 212. In some embodiments, preprocessing may include one or more of ECG normalization (where the range of ECG data is mapped to a range of 0 to 1), noise reduction, binarization of intensity values, ECG waveform thinning, etc. In some embodiments, the convolutional neural network 212 is configured to receive the ECG data 202 as a 2D data object, including a first dimension of potential and a second dimension of time. In embodiments where the ECG data 202 includes a 2D data object, the input layer of the convolutional neural network 212 may be configured as a 2D grid of input nodes / neurons, wherein each input node / neuron may receive different values ​​of the 2D ECG data as input. In some embodiments, the ECG data 202 may include one or more 2D images, wherein the ECG waveform data is graphically displayed within the 2D images such that the ECG waveform data is encoded by multiple pixel intensity values ​​of the 2D images. In such implementations, the input layer of the convolutional neural network 212 may include the same number of input nodes / neurons as the pixels in one or more 2D images, thereby enabling pixel intensity values ​​to be input one-to-one to the input nodes / neurons of the input layer of the convolutional neural network 212.

[0053] After ECG data 202 is fed into convolutional neural network 212, the data propagates through one or more convolutional layers, where, in each convolutional layer, one or more filters are applied to the input ECG data 202 to produce feature maps. Each filter includes one or more weights that map a subregion of the input space to a subregion of the feature map. Typically, filters can be used to detect local bindings of features in the input data (feature maps and / or ECG data 202). Filters may include a 3×3 weight grid, where data from the 3×3 input value grid (e.g., ECG data 202 or features from previous feature maps) is mapped to a single output value by computing the dot product between the 3×3 input value grid and the 3×3 weight grid. In some implementations, each of the one or more weights of the filters may be learned during the training phase. The depth of the feature map corresponds to the number of filters applied, while the height and width of the feature map correspond to the height and width of the input ECG data 202 (or the previous feature map from which the current convolutional layer receives data). The feature map quantifies the degree of matching between one or more applied filters and each spatial region within the input data.

[0054] Data from the feature maps can be passed through nonlinear layers and / or rectified layers to produce activation maps, which can then be used as input to subsequent layers. In some implementations, combined rectified and nonlinear layers, such as rectified linear units (ReLU), can be used. The convolutional neural network 212 may also include additional convolutional layers (where feature maps or activation maps from previous convolutional layers can be treated as input and filtered / convolved again to produce feature maps / activation maps that include higher-order features / composite features), pooling layers (where each sub-region of the ECG data 202 or feature map is downsampled to produce feature maps with lower spatial resolution), dropout layers (where one or more weights or filters are discarded during training to reduce the probability of overfitting), and regularization layers, as well as fully connected / densely connected layers (where each output / activation / feature of a previous feature map or activation map is received by each node / neuron of a subsequent layer).

[0055] One or more filters in each convolutional layer of the convolutional neural network 212 can be learned during training, thus potentially eliminating the need for hard-coding by domain experts (such as cardiologists). This allows the convolutional neural network 212 to learn to recognize and extract features relevant to accurate diagnostic predictions, rather than utilizing predetermined features. The technical advantage of using a hybrid ECG diagnostic system (such as hybrid ECG diagnostic system 200) to diagnose ECG data is that the deep neural network within the hybrid ECG diagnostic system can automatically learn to recognize use-specific features relevant to diagnostic predictions by training on use-specific data, without requiring manual updates to hard-coded rules.

[0056] The convolutional neural network 212 also includes an output layer configured to output multiple features extracted / identified in the ECG data 202. In some embodiments, the output from the convolutional neural network 212 includes a feature map indicating the probability of occurrence of each of the multiple learned features at each of multiple sub-regions within the ECG data 202. It should be understood that although the terminology used to describe the feature extractor 206 and the convolutional neural network 212 is similar because both feature extractor 206 and convolutional neural network 212 determine / extract features within the ECG data, it should be understood that the convolutional neural network 212 detects learned features at each sub-region of the ECG data 202, while the feature extractor 206 extracts pre-selected / pre-defined features, as discussed in more detail herein. Furthermore, the convolutional neural network 212 and the feature extractor 206 are configured to extract / determine features in substantially different ways.

[0057] The output from convolutional neural network 212 can be fed into decision network 214. In some embodiments, the output from convolutional neural network 212 includes multiple features identified in ECG data 202. Decision network 214 includes an input layer with multiple input nodes / neurons, wherein the number of input nodes / neurons may be equal to the number of features output by convolutional neural network 212 in some embodiments. In some embodiments, the number of input nodes / neurons in the input layer of decision network 214 may be greater than the number of output features from convolutional neural network 212, thereby enabling decision network 214 to receive additional data from one or more sources.

[0058] The decision network 214 comprises an input layer, an output layer, and one or more hidden layers, each of which includes a densely / fully connected layer. The architecture of the decision network 214 can also be referred to as a multilayer perceptron (MLP). Each neuron in each hidden layer produces an output by computing the dot product between each output of each node in the previous layer using multiple weights, where each weight is uniquely associated with a neuron in the previous layer. Mathematically, this can be represented by the following equation.

[0059]

[0060] Where X i It is the i-th neuron in the previous layer, Y j W is the output / activation of the j-th neuron in the current layer. ji It represents the weight / strength of the connection between the i-th neuron in the previous layer and the j-th neuron in the current layer, and B jThis represents the bias of the j-th neuron in the current layer. In some implementations, the activation function f is, for example, the rectified linear unit (ReLU) function. In some implementations, f may include the ordinary ReLU function, the leaky ReLU function, the parametric ReLU function, etc. In some implementations, f may include the hyperbolic tangent function, the softplus function, the soft exponential function, the sigmoid function, and other activation functions known in the field of machine learning. The decision network may include virtually any number of hidden layers.

[0061] The output layer of decision network 214 includes multiple output neurons. In some embodiments, each output neuron may correspond to a different diagnosis, and the value output from the neuron in the output layer may represent / indicate a probability score of the diagnosis associated with the neuron. For example, neuron J in the output layer corresponding to diagnosis K may output a probability score between 0 and 1, where the probability score indicates the predicted probability of diagnosis K applied to the input ECG data 202. Specifically, if neuron J outputs a probability score of approximately 0, this indicates a relatively low probability of diagnosis K applied to ECG data 202, while if neuron J outputs a probability score of approximately 1, this indicates a relatively high probability of diagnosis K applied to ECG data 202. The output from the output layer of decision network 214 may include a vector of probability scores, i.e., one probability score for each of a predetermined plurality of diagnoses. The vector of probability scores generated by the output layer of decision network 214 may be referred to as a preliminary diagnosis, as the probability scores may be fed into rule-based system 204 for further analysis / diagnosis. Specifically, the preliminary diagnosis generated by deep neural network 210 is fed into feature extractor 206 and decision tree 208 of rule-based system 204.

[0062] The rule-based system 204 includes a feature extractor 206 and a decision tree 208, and can be configured to receive ECG data 202 at the feature extractor 206, and to receive preliminary diagnoses at one or both of the feature extractor 206 and the decision tree 208.

[0063] In some implementations, feature extractor 206 may perform one or more preprocessing operations on ECG data 202. In some implementations, feature extractor 206 may determine a region of interest in the received 2D ECG data and invert the color data in the region of interest to generate a black ECG signal / waveform on a white background. Feature extractor 206 may then binarize the 2D ECG data (i.e., each pixel of the 2D ECG data below an intensity threshold may be set to black, and each pixel of the 2D ECG data above an intensity threshold may be set to white, thereby producing a 2D ECG image comprising pixels with a binary intensity distribution, where each pixel of ECG data 202 is either white or black, i.e., 1 or 0). After binarizing the ECG waveform, one or more noise suppression operations may be performed to remove isolated black pixels (i.e., black pixels not connected to the ECG waveform). Feature extractor 206 may also thin the ECG waveform by setting its thickness to one pixel, thereby reducing redundant information contained in the ECG waveform. Therefore, feature extractor 206 may preprocess ECG data 202 to achieve more efficient feature extraction and diagnostics. Furthermore, after preprocessing, the information contained in ECG data 202 can be stored / processed more easily because redundant data can be reduced or eliminated.

[0064] Feature extractor 206 includes hard-coded machine-executable instructions for determining one or more predetermined features of ECG data 202. In some embodiments, the predetermined features of the ECG include the amplitude, area, and duration / interval of each wave (P, Q, R, S, T, and U), as well as the interval / duration between waves (of the same or different types). In some embodiments, feature extractor 206 is configured to determine one or more amplitudes and / or intervals of waveforms within the ECG. In some embodiments, feature extractor 206 may determine the amplitude and period of one or more of the P wave, Q wave, R wave, S wave, T wave, U wave, or a composite wave formed by a combination of two or more of the aforementioned waves. In some embodiments, feature extractor 206 may be configured to determine one or more periods / intervals within the ECG waveform, such as inter-wave intervals / durations or the duration of a single wave. The feature extractor can determine heart rate by determining the interval / duration between corresponding features of the same wave type, for example, by measuring the duration between the peaks of a first Q wave and a second Q wave. The feature extractor can average one or more intervals / durations to determine heart rate, or can extrapolate heart rate based on a single measurement interval. Furthermore, the feature extractor can determine one or more heartbeat-to-heartbeat variations of one or both of the aforementioned waveforms. Feature extractor 206 includes hard-coded predetermined ECG features, wherein the predetermined ECG features can be determined by an expert cardiologist and may include waveform features that are intuitive to humans.

[0065] After preprocessing, feature extractor 206 can fit a baseline (also known as an isoline) to the ECG waveform data. In some embodiments, feature extractor 206 can determine the baseline by fitting a horizontal straight line through the maximum number of black pixels in the ECG waveform (where a horizontal line, as used herein, will be understood as a line extending substantially parallel to the X-axis of the ECG, where the X-axis is typically the time domain of the ECG) and setting that horizontal straight line as the baseline. The baseline allows feature extractor 206 to determine the amplitude and interval of each wave in the ECG. In some embodiments, the amplitude of the wave can be determined as the difference in the Y-dimensional (potential) between the wave crest and the baseline, while the interval / duration of the wave can include the difference in the X-dimensional (time) between the initial deviation of the wave from the baseline and the point on the X-axis where the wave returns to the baseline. In some embodiments, feature extractor 206 can fit each wave of the ECG data 202 with rectangles, where the base of each rectangle extends parallel to the baseline and extends at the same height (Y-dimensional) as the baseline. The height (Y-dimensional) of each rectangle corresponds to the height / amplitude (Y-dimensional) of each corresponding wave, and the width (X-dimensional) of each wave corresponds to the duration / interval of each wave. Furthermore, the feature extractor 206 can approximate the area of ​​each wave based on the area (height multiplied by width) of each corresponding rectangle. In some embodiments, the feature extractor 206 can uniquely identify each wave of the ECG waveform by first determining which wave includes the maximum amplitude and labeling that wave as an R-wave. Then, the feature extractor 206 can determine the identities of other waves (P, Q, S, T, and U) relative to the R-wave along the X-axis. For example, the feature extractor 206 can label a wave immediately preceding the R-wave as a Q-wave, and further label a wave immediately preceding the Q-wave as a P-wave. Similarly, the feature extractor 206 can label a wave immediately following the R-wave as an S-wave, and a wave immediately following the S-wave as a T-wave, and so on.

[0066] Features determined by feature extractor 206 can be fed into decision tree 208. In some embodiments, decision tree 208 includes one or more root nodes, one or more decision nodes or child nodes, and one or more leaf nodes. Features extracted by feature extractor 206 can be input into one or more root nodes, where features can be evaluated based on one or more predetermined expert criteria / heuristics. Each node, except for leaf nodes, can include one or more child nodes. The evaluation at each node determines which child node the feature is propagated to. Child nodes branching from a given node can in turn include one or more child nodes, and this pattern can continue until one or more leaf nodes are reached. Each node branching into two or more child nodes can be called a decision node because a "decision" is made there about which child node to select. At each decision node, features of ECG data 202 determined by feature extractor 206 can be evaluated based on expert criteria, and the result of the evaluation determines which child node the extracted features are propagated to. Terminal nodes (also called leaf nodes) can include one or more diagnoses. Decision tree 208 can include multiple leaf nodes, each uniquely corresponding to a different diagnose. The extracted features of ECG data 202 are propagated through the branching structure of decision tree 208 until one or more leaf nodes are reached, and a diagnosis, such as diagnosis 216, is determined based on one or more leaf nodes.

[0067] In the hybrid ECG diagnostic system 200, a preliminary diagnosis generated by a deep neural network 210 may be fed into one or both of a feature extractor 206 and a decision tree 208. The preliminary diagnosis generated by the deep neural network 210 may be used to modify one or more features determined by the feature extractor 206 and / or modify the final diagnosis 216 generated by the decision tree 208. In some embodiments, the preliminary diagnosis generated by the deep neural network 210 may be used by a rule-based system 204 to evaluate and / or modify the confidence level of the diagnosis 216. For example, if the diagnosis 216 generated by the rule-based system 204 is the same as the preliminary diagnosis generated by the deep neural network 210, the rule-based system 204 may indicate a high confidence rating for the diagnosis. Conversely, if the diagnosis 216 generated by the rule-based system 204 deviates from the preliminary diagnosis generated by the deep neural network 210, the rule-based system 204 may indicate a low confidence rating for the diagnosis. In some embodiments, the diagnosis 216 may include a confidence rating indicating the degree of consistency / inconsistency between the preliminary diagnosis generated by the deep neural network 210 and the rule-based system 204.

[0068] Diagnostic 216 can be displayed to a user via a display device such as device 114. In some implementations, diagnostic 216 can be transmitted to one or more client devices, edge computing devices, servers, cloud storage devices, etc.

[0069] Go to Figure 3 A second embodiment of a hybrid ECG diagnostic system 300 is illustrated. In some embodiments, the hybrid ECG diagnostic system 300 may be implemented by an ECG processing system 102, an edge device connected to the ECG processing system 102, a cloud device communicating with the ECG processing system 102, or any suitable combination thereof. In the hybrid ECG diagnostic system 300, a rule-based system 304 determines a preliminary diagnosis from ECG data 302 based on hard-coded expert criteria. This preliminary diagnosis, along with multiple features determined from the ECG data 302 by a convolutional neural network 312, is fed into the input layer of a decision network 314 of a deep neural network 310. The decision network 310 then maps both the multiple features generated by the convolutional neural network 312 and the preliminary diagnosis generated by the rule-based system 304 to a diagnosis 316. In this way, the deep neural network 310 is instructed to use the preliminary diagnosis generated using hard-coded expert criteria, which can be used by the decision network 314 to generate a more accurate diagnosis 316 for the ECG data 302. In other words, the decision network 314 can update / modify the preliminary diagnosis generated by the rule-based system 304 based on the learned features extracted from the ECG data 302, thereby adding additional flexibility and accuracy to the diagnosis generated by the rule-based system 304.

[0070] In some embodiments, ECG data 302 may include a real-time stream of ECG data acquired from an ECG system (such as ECG system 100). In some embodiments, ECG data 302 may include pre-recorded ECG data that may be transferred from a communicatively linked data storage device (such as ECG data module 112 in non-transitory memory 106) to a hybrid ECG diagnostic system 300. In some embodiments, ECG data 302 may be a twelve-lead ECG. In some embodiments, ECG data 302 may be a fifteen-lead ECG. In some embodiments, ECG data 302 may be a reduced-lead ECG comprising fewer than twelve leads but more than zero leads. ECG data 302 may be transmitted sequentially or in parallel to both rule-based system 304 and deep neural network 310.

[0071] The rule-based system 304 includes a feature extractor 306 and a decision tree 308, which is configured to receive ECG data 302 at the feature extractor 306 and generate / determine a preliminary diagnosis based on it. The rule-based system 304 can be configured substantially similarly to the rule-based system 204 described above; however, in some embodiments, the rule-based system 304 may not be configured to receive the preliminary diagnosis from the deep neural network 310. The feature extractor 306 can perform one or more preprocessing operations on the ECG data 302 in a manner similar to that described above with reference to feature extractor 206. After preprocessing, the feature extractor 306 can determine one or more predetermined features of the ECG data 302, as described in more detail above with reference to feature extractor 206. The features determined by the feature extractor 206 can be fed into the decision tree 308, which is configured substantially similarly to the decision tree 208 described above. In short, the extracted features of the ECG data 302 propagate through the branching structure of the decision tree 308 until one or more leaf nodes are reached, and a preliminary diagnosis is determined based on one or more leaf nodes. The preliminary diagnosis generated by the decision tree 308 can be transmitted to the input layer of the decision network 314 of the deep neural network 310.

[0072] The deep neural network 310 can be configured in a manner substantially similar to that of the deep neural network 210. The deep neural network 310 comprises two deep neural networks connected in series: the first is a convolutional neural network 312, which includes multiple convolutional layers, and the second is a decision network 314, which includes multiple densely connected layers (therefore, the decision network 314 may also be referred to as a multilayer perceptron or MLP).

[0073] The convolutional neural network 312 may include a convolutional neural network similar to the convolutional neural network 212 described above. The convolutional neural network 312 includes one or more convolutional layers (which include one or more learning filters) and is configured to receive ECG data 302 at an input layer and map the ECG data 302 to multiple features / feature maps using one or more learning filters. In some embodiments, the output generated by the convolutional neural network 312 may include feature maps, which, along with preliminary diagnoses generated by the decision tree 308, may be input to the input layer of the decision network 314.

[0074] Decision network 314 can be configured substantially similarly to decision network 214; however, decision network 314 is configured to receive the output from convolutional neural network 312 and a preliminary diagnosis generated by decision tree 308. In some embodiments, the input layer of decision network 314 may include a first plurality of neurons configured to receive a plurality of features / feature maps output by the output layer of convolutional neural network 312 and a second plurality of neurons configured to receive the preliminary diagnosis generated by decision tree 308. The technical effect of inputting the preliminary diagnosis into the input layer of decision network 314 is that deep neural network 310 can adjust / modify the preliminary diagnosis based on one or more features identified by convolutional neural network 312 in ECG data 302, wherein convolutional neural network 312 can learn to extract / identify ECG features relevant to the correct diagnosis in use case-specific data without reprogramming / recoding. The output from the output layer of convolutional neural network 312, together with the preliminary diagnosis generated by decision tree 308 of rule-based system 310, is received at the input layer of decision network 314 and mapped to diagnosis 316. In some embodiments, diagnosis 316 includes multiple probability scores for each of a plurality of predetermined diagnoses. In some embodiments, diagnosis 316 includes a vector of probability scores, wherein each row of the vector corresponds to a different diagnosis or diagnostic feature.

[0075] In some implementations, the leads can be derived from a deep neural network library (such as...) based on the leads used to train a deep neural network 310 that matches the leads in the ECG data 302. Figure 11 The deep neural network 310 is intelligently selected in the deep neural network library 1104 shown. In some implementations, it may have been based on... Figure 13 One or more operations of the method 1300 shown are selected from the deep neural network 310.

[0076] Go to Figure 4A third embodiment of a hybrid ECG diagnostic system 400 is illustrated. In some embodiments, the hybrid ECG diagnostic system 400 may be implemented by an ECG processing system 102, an edge device connected to the ECG processing system 102, a cloud device communicating with the ECG processing system 102, or any suitable combination thereof. In the hybrid ECG diagnostic system 400, ECG data 402 is fed to a rule-based system 404, which generates a first preliminary diagnosis based on the ECG data 402. The first preliminary diagnosis is then transmitted to a deep neural network 410, which maps both the first preliminary diagnosis and the ECG data 402 to a second preliminary diagnosis. The second preliminary diagnosis is then fed back to the rule-based system 404 and used to generate a diagnosis 416. In this way, the hybrid ECG diagnostic system 400 can incorporate adaptive feature mappings from deep neural networks, which have intuitive and transparent decision models that reflect expert criteria, to generate diagnoses tailored to specific use cases and have greater transparency than that achievable solely in deep neural networks.

[0077] In some embodiments, ECG data 402 may include a real-time stream of ECG data acquired from an ECG system (such as ECG system 100). In some embodiments, ECG data 402 may include pre-recorded ECG data that may be transferred from a communicatively linked data storage device (such as ECG data module 112 in non-transitory memory 106) to a hybrid ECG diagnostic system 400. In some embodiments, ECG data 402 may be a twelve-lead ECG. In some embodiments, ECG data 402 may be a fifteen-lead ECG. In some embodiments, ECG data 402 may be a reduced-lead ECG comprising fewer than twelve leads but more than zero leads. ECG data 402 may be transmitted sequentially or in parallel to both rule-based system 404 and deep neural network 410.

[0078] The rule-based system 404 includes a feature extractor 406 and a decision tree 408, which is configured to receive ECG data 402 at the feature extractor 406 and generate / determine a first preliminary diagnosis based thereon. The rule-based system 404 is further configured to incorporate a second preliminary diagnosis generated by a deep neural network 410 into one or more of the feature extractor 406 and the decision tree 408, and to determine a diagnosis 416 based on the ECG data 402 using the second preliminary diagnosis. The feature extractor 406 may perform one or more preprocessing operations on the ECG data 402 in a manner similar to that described above with reference to feature extractor 406. After preprocessing, the feature extractor 406 may determine one or more predetermined features of the ECG data 402, such as those described in more detail above with reference to feature extractor 206. In some embodiments, the second preliminary diagnosis generated by the deep neural network 410 may be used to update / modify one or more parameters of the feature extractor 406.

[0079] Features determined by feature extractor 406 can be fed into decision tree 408, which is configured in a manner substantially similar to decision tree 208 described above. In some embodiments, a second preliminary diagnosis generated by deep neural network 410 can be used to update / modify one or more parameters of decision tree 408. In short, the extracted features from ECG data 402 can propagate through the branching structure of decision tree 408 until one or more leaf nodes are reached, and a first preliminary diagnosis is determined based on one or more leaf nodes. The first preliminary diagnosis generated by decision tree 408 can be transmitted to the input layer of decision network 414 of deep neural network 410.

[0080] The deep neural network 410 can be configured in a manner substantially similar to that of the deep neural network 210. The deep neural network 410 comprises two deep neural networks connected in series: the first is a convolutional neural network 412, which includes multiple convolutional layers, and the second is a decision network 414, which includes multiple densely connected layers (therefore, the decision network 414 may also be referred to as a multilayer perceptron or MLP).

[0081] Convolutional Neural Network 412 may include a convolutional neural network similar to Convolutional Neural Network 212 described above. Convolutional Neural Network 412 includes one or more convolutional layers (which include one or more learning filters) and is configured to receive ECG data 402 at an input layer and map the ECG data 402 to a feature map / multiple features using the one or more learning filters. In some embodiments, the output generated by Convolutional Neural Network 412 may include a feature map, which, along with a first preliminary diagnosis generated by decision tree 408, may be input to the input layer of decision network 414.

[0082] Decision network 414 can be configured substantially similarly to decision network 214; however, decision network 414 is configured to receive the output from convolutional neural network 412 and a first preliminary diagnosis generated by decision tree 408. In some embodiments, the input layer of decision network 414 may include a first plurality of neurons configured to receive a plurality of features / feature maps output by the output layer of convolutional neural network 412 and a second plurality of neurons configured to receive the first preliminary diagnosis generated by decision tree 408. The technical effect of inputting the first preliminary diagnosis along with a plurality of features mapped from ECG data 402 by convolutional neural network 412 into the input layer of decision network 414 is that deep neural network 410 can adjust / modify the first preliminary diagnosis based on one or more features identified by convolutional neural network 412 in ECG data 402, wherein convolutional neural network 412 can learn to extract / recognize ECG features relevant to a baseline truth diagnosis in use case-specific data without reprogramming / recoding.

[0083] The output from the output layer of the convolutional neural network 412, along with a first preliminary diagnosis generated by the decision tree 408 of the rule-based system 410, is received at the input layer of the decision network 414 and mapped to a second preliminary diagnosis. In some embodiments, the second preliminary diagnosis includes multiple probability scores for each of a plurality of predetermined diagnoses. In some embodiments, the second preliminary diagnosis includes a vector of probability scores, where each row of the vector corresponds to a different diagnosis or diagnostic feature.

[0084] In some implementations, the leads can be selected from a deep neural network library (such as...) based on the leads used to train a deep neural network 410 that matches the leads in the ECG data 402. Figure 11 In the deep neural network library 1104 shown, the deep neural network 410 is intelligently selected (that is, the lead types present in the data used to train the deep neural network 410 are the same as the lead types present in the ECG data 402). In some implementations, it may have been based on... Figure 13 One or more operations of the method 1300 shown are selected from the deep neural network 410.

[0085] A second preliminary diagnosis generated by the decision network 414 of the deep neural network 410 can be fed back to the rule-based system 404, and one or more parameters of the feature extractor 406 and / or decision tree 408 can be adjusted based on it. In some embodiments, the rule-based system 404 can use the second preliminary diagnosis to determine the confidence / determinism of the final diagnosis 416. After receiving the second preliminary diagnosis generated by the deep neural network 410, the rule-based system 404 can re-evaluate the ECG data 402 to generate the diagnosis 416. The diagnosis 416 may include the final diagnosis and may be stored in memory, displayed via a display device, and / or transmitted to one or more client devices. In this way, the hybrid ECG diagnostic system 400 combines a recursive / feedback diagnostic scheme, wherein the initial diagnostic prediction of the rule-based system 404 can be used to form the diagnostic prediction of the deep neural network 410, which can then be used by the rule-based system 404 to generate the final diagnostic prediction. The technical advantage of the above-described recursive diagnostic scheme is that it can enhance diagnostic accuracy, confidence, and consistency.

[0086] Go to Figure 5 A flowchart illustrating an exemplary method 500 for diagnosing ECG data using a hybrid ECG diagnostic system is shown. In some embodiments, method 500 may be performed by ECG system 100, an edge device connected to ECG system 100, a cloud device communicating with ECG system 100, or any suitable combination thereof. Method 500 enables automated, accurate, and use-case-specific diagnosis of one or more ECGs from one or more patients.

[0087] Method 500 begins at 502, where ECG data is acquired by an ECG system. ECG data may include one or more lead signals / waveforms indicating the electrical activity of the heart over time. In some embodiments, acquiring ECG data may include measuring a twelve-lead ECG using ten electrodes in electrical contact with the patient, as referenced above. Figure 1 A more detailed description follows. In some embodiments, acquiring ECG data may include measuring a fifteen-lead ECG using thirteen electrodes. In some embodiments, acquiring ECG data may include measuring a reduced-lead ECG using fewer than ten electrodes. The acquired ECG data may be stored in memory for later processing. In some embodiments, the ECG data acquired at operation 502 includes 2D data of potential changes over time, where potential is plotted along the y-axis and time is plotted along the x-axis. ECG data acquired at operation 502 may be stored in one or more formats, including SCP-ECG, DICOM-ECG, HL7 aECG, and other storage formats known in the ECG field. Acquiring ECG data at operation 502 can occur under various conditions, such as when the patient is at rest, during exercise, or in an ambulance.

[0088] At operation 504, a diagnosis of the ECG data acquired at operation 502 is determined using one or more of hybrid ECG diagnostic systems 200, 300, and / or 400 by performing one or more of methods 600, 700, and / or 800. In a first exemplary embodiment, operation 504 includes hybrid ECG diagnostic system 200 performing one or more of method 600 to determine a diagnosis based on the acquired ECG data. In a second exemplary embodiment, operation 504 includes hybrid ECG diagnostic system 300 performing one or more of method 700 to determine a diagnosis based on the acquired ECG data. In a third embodiment, operation 504 includes hybrid ECG diagnostic system 400 performing one or more of method 800 to determine a diagnosis based on the acquired ECG data. In some embodiments, the ECG system performing method 500 may select between the first, second, and third embodiments of the hybrid ECG diagnostic system based on one or more characteristics of the acquired ECG data.

[0089] At operation 506, a diagnosis is displayed via a display device. The diagnosis may indicate whether the ECG data includes abnormalities (arrhythmias) and which features of the ECG presentation / display are abnormal. In the case of a normal / healthy ECG, the diagnosis may indicate that no arrhythmia / abnormality was detected. In some embodiments, the display device includes a graphical user interface in which the diagnosis can be displayed to a user. In some embodiments, the display device may include a device coupled to an ECG system, such as an operator-viewable display device 114 of the ECG system. The operator can view the displayed diagnosis and may take one or more actions based on it. The diagnosis may include one or more lines of text indicating one or more features of the ECG data, and a diagnostic assessment of the one or more features. (Temporarily skip to...) Figure 10 Two exemplary implementations of the diagnostics are illustrated: a first diagnostic 1004 generated using a rule-based system without a deep neural network, and a second diagnostic 1006 generated using a hybrid ECG diagnostic system discussed herein (such as hybrid ECG diagnostic systems 200, 300, and 400). Figure 10 As shown, the first diagnosis 1004 includes a text message indicating an assessment of the overall normality / abnormality of heart rate, contour (including QRS complexes, ST-T waves), and ECG 1002, as well as one or more underlying physiological mechanisms of the indicated normality / abnormality. Similarly, the second diagnosis 1006 includes a message indicating a diagnostic assessment of the overall normality / abnormality of heart rate, contour (including QRS complexes, ST-T waves), and one or more physiological characteristics of the heart, and ECG 1002, as well as one or more underlying physiological mechanisms of the indicated normality / abnormality (such as electrolyte disturbances, possible location of coronary artery obstruction, etc.).

[0090] At operation 508, the ECG system may transmit diagnostics to one or more client systems. Client systems may include one or more client devices communicatively connected to the ECG system via a wired or wireless connection or via a network (such as the Internet). In some implementations, operation 508 includes transmitting diagnostics to an email address, URL, client cloud storage account, cloud service, or other client-related account.

[0091] After operation 508, method 500 may end. In this way, method 500 enables the use of ECG data acquired through automated diagnosis by a hybrid ECG system to quickly and accurately determine a diagnosis, wherein the diagnosis can be displayed to a user (such as a doctor or healthcare professional) and / or distributed to one or more client devices.

[0092] Go to Figure 6 A flowchart of a first embodiment of a method 600 for automatically determining a diagnosis using hybrid ECG diagnostic data is shown. Method 600 may be performed as part of method 500, as indicated in operation 504. In some embodiments, method 600 may be performed by a hybrid ECG diagnostic system 200, which may be implemented by an ECG system 100, an edge device connected to the ECG system 100, a cloud device communicating with the ECG system 100, or any suitable combination thereof. Method 600 can enable automated, accurate, and use-case-specific diagnoses of one or more ECGs from one or more patients.

[0093] Method 600 begins at 602, wherein the acquired ECG data is received by a hybrid ECG diagnostic system, wherein the trained hybrid ECG diagnostic system includes rule-based systems (such as rule-based system 204 described above) and deep neural networks (such as deep neural network 210 described above). The received ECG data includes one or more ECG waveforms / signals that indicate the electrical activity of the heart over time. In some embodiments, the received ECG data may include data from a twelve-lead ECG, as referenced above. Figure 1 More detailed description. In some embodiments, the received ECG data may include data from a 15-lead ECG. In some embodiments, the received ECG data may include data from a reduced-lead ECG. In some embodiments, the received ECG data includes 2D data of potential variation over time, where potential is plotted along the y-axis and time is plotted along the x-axis. The received ECG data may be formatted as SCP-ECG, DICOM-ECG, HL7aECG, or other storage formats known in the ECG field.

[0094] At operation 604, the received ECG data is input into the input layer of a convolutional neural network (CNN) of the deep neural network and mapped to multiple features. That is, at operation 604, the CNN extracts one or more features from the ECG data using one or more convolutional layers. The input layer of the CNN may include multiple neurons / input nodes configured to receive ECG data. In some embodiments, each neuron in the input layer of the CNN may be configured to receive a single value from the ECG data, where the single value may correspond to a pixel intensity value of an image of the ECG waveform. In some embodiments, the received ECG data may include a 2D data structure, such as a matrix or array, and the number of input neurons in the input layer of the CNN may be equal to the number of entries in the matrix / array. Furthermore, the input neurons may be configured to reflect a 2D arrangement of the received 2D ECG data, thereby preserving the spatial relationships encoded within the 2D ECG data.

[0095] A convolutional neural network (CNN) comprises one or more convolutional layers, each containing one or more filters. Each filter includes multiple weights, where each weight is used to compute a corresponding value in the received ECG data or a dot product from a previous feature map. The result of the dot product between ECG data (or previous feature maps) is used to determine the current feature map. The filter weights are learned during the training phase, rather than hard-coded, enabling the CNN to learn to recognize different features relevant to diagnostic outcomes across different datasets / use cases, even if these features may not seem intuitive to human experts.

[0096] In each of one or more convolutional layers, one or more filters included in the layer pass through each sub-region of the received ECG data, wherein the size of the sub-region is predetermined, and the output of the filter applied to each sub-region is used to generate a feature map. In some embodiments, the sub-region may include a set of continuous values ​​from an image. In some embodiments, the received ECG data includes a 2D image of an ECG waveform, and the sub-region includes a predetermined number of pixel intensity values ​​from a spatially continuous region within the 2D image.

[0097] The feature map comprises multiple values ​​that indicate the degree of matching between one or more applied filters and each sub-region of the received ECG data (or a previous feature map), thereby indicating multiple features present in the received ECG data. In some implementations, higher values ​​in a sub-region of the feature map indicate a greater degree of matching between a given filter and a corresponding sub-region of the received ECG data. That is, relatively higher values ​​in the feature map indicate a relatively higher probability of the features represented by the filters present in the ECG data. The feature map may include n dimensions (where n is a positive integer greater than 1), where the (n-1)th dimensions of the feature map correspond to the dimensions of the input data, and the nth dimension of the feature map is used to store the result of each applied filter. In other words, the size of the nth dimension of the feature map corresponds to the number of filters applied to the input data.

[0098] The output from a convolutional neural network (CNN) is generated by multiple neurons in the output layer of the CNN, which receive the preceding feature map as input. In some implementations, the output from the CNN includes one or more features present in the received ECG data. In some implementations, the output from the CNN includes feature maps generated from the received ECG data using one or more convolutional layers.

[0099] At operation 606, multiple features generated by the convolutional neural network are input into the decision network of the deep neural network and mapped to a preliminary diagnosis of the received ECG data. The decision network includes an input layer comprising multiple neurons configured to receive outputs from the convolutional neural network; multiple fully / densely connected layers; and an output layer comprising multiple neurons configured to generate probability scores for each of a plurality of predetermined diagnoses. In some implementations, the number of neurons in the input layer and the number of outputs / features generated by the convolutional neural network may be the same, thereby achieving a one-to-one correspondence between the outputs from the convolutional neural network and the input neurons of the decision network. The decision network is capable of mapping multiple features generated by the convolutional neural network to a preliminary diagnosis (including one or more diagnostic assessments) using multiple learned weights and biases.

[0100] At operation 608, the rule-based system receives a preliminary diagnosis generated by a deep neural network and ECG data received at operation 602, and determines a diagnosis based on it. In some embodiments, the preliminary diagnosis generated by the deep neural network can be used to adjust one or more criteria of the rule-based system. In some embodiments, the preliminary diagnosis can be used to select one or more types of predetermined features for extraction by the feature extractor of the rule-based system. In some embodiments, the preliminary diagnosis, together with the features extracted by the feature extractor, can be fed into the decision tree of the rule-based system and used to determine the final diagnosis of the received ECG data. In some embodiments, the preliminary diagnosis generated by the deep neural network can be used to generate confidence scores for certain diagnoses or multiple confidence scores for multiple diagnoses (such as both ventricular bundle branch block and ischemia).

[0101] After operation 608, method 600 can terminate. In this way, method 600 can use a hybrid system including a deep neural network and a rule-based system to determine the diagnosis of the received ECG data, wherein the ECG data is mapped to a preliminary diagnosis by the deep neural network, and the preliminary diagnosis, along with the ECG data, is used by the rule-based system to determine the final diagnosis. Therefore, method 600 enables the use of a preliminary diagnosis determined by an adaptive deep neural network to fine-tune the diagnostic predictions of the rule-based system, wherein the deep neural network can be trained on use case-specific data.

[0102] Go to Figure 7 A flowchart of a second embodiment of a method 700 for automatically determining a diagnosis using hybrid ECG diagnostic data is shown. Method 700 may be performed as part of method 500, as indicated at operation 504. In some embodiments, method 700 may be performed by a hybrid ECG diagnostic system 300, which may be implemented by ECG system 100, an edge device connected to ECG system 100, a cloud device communicating with ECG system 100, or any suitable combination thereof. Method 700 can enable automated, accurate, and use-case-specific diagnoses of one or more ECGs from one or more patients.

[0103] Method 700 begins at 702, wherein the acquired ECG data is received by a hybrid ECG diagnostic system, wherein the hybrid ECG diagnostic system includes a rule-based system (such as the rule-based system 304 described above) and a deep neural network (such as the deep neural network 310 described above). The received ECG data includes one or more ECG waveforms / signals that indicate the electrical activity of the heart over time. In some embodiments, the received ECG data may include data from a twelve-lead ECG, as referenced above. Figure 1More detailed description. In some embodiments, the received ECG data may include data from a 15-lead ECG. In some embodiments, the received ECG data may include data from a reduced-lead ECG. In some embodiments, the received ECG data includes 2D data of potential variation over time, with potential plotted along the y-axis and time plotted along the x-axis. The received ECG data may be formatted as SCP-ECG, DICOM-ECG, HL7 aECG, or other storage formats known in the ECG field.

[0104] At operation 704, the received ECG data is input into a rule-based system and used by the rule-based system to determine a preliminary diagnosis. In some embodiments, inputting the received ECG data into the rule-based system includes inputting the received ECG data into a feature extractor of the rule-based system. In some embodiments, the feature extractor may perform one or more preprocessing operations on the input ECG data before extracting features. The feature extractor may be configured as described above with reference to feature extractor 306 to determine one or more predetermined features present in the received ECG data. In some embodiments, the predetermined features may include one or more amplitudes, intervals / durations, and / or inter-amplitudes / durations.

[0105] The extracted features can be fed into a decision tree (such as decision tree 308 described above). The decision tree can determine the path from the root node to the leaf node based on one or more decisions made at one or more decision nodes, where the decisions are based on hard-coded expert criteria used to evaluate one or more extracted features determined by the feature extractor. Leaf nodes can include one or more diagnostic assessments of the received ECG data and can be used to determine preliminary diagnoses for input into the decision network of the deep neural network.

[0106] At operation 706, the received ECG data is input into the input layer of a convolutional neural network (such as convolutional neural network 312) of a deep neural network and mapped to multiple features. That is, at operation 706, the convolutional neural network extracts one or more features from the ECG data using one or more convolutional layers. The input layer of the convolutional neural network may include multiple neurons configured to receive ECG data. The convolutional neural network also includes one or more convolutional layers, each including one or more filters. The output from the convolutional neural network is generated by multiple neurons in the output layer of the convolutional neural network. In some embodiments, the output from the convolutional neural network includes one or more learned features present in the received ECG data. In some embodiments, the learned features may include one or more wave features or a combination of wave features. In some embodiments, the output from the convolutional neural network includes feature maps generated from the received ECG data using one or more convolutional layers.

[0107] At operation 708, multiple features output by the convolutional neural network, along with a preliminary diagnosis determined by a rule-based system, are input into a decision network of a deep neural network (such as decision network 314) and mapped to a final diagnosis of the received ECG data. The decision network includes an input layer comprising multiple neurons configured to receive the output from the convolutional neural network and the preliminary diagnosis determined by the rule-based system. In some embodiments, the input layer of the decision network includes a first plurality of neurons configured to receive multiple features output from the convolutional neural network and a second plurality of neurons configured to receive the preliminary diagnosis determined by the rule-based system. The decision network also includes one or more fully / densely connected layers and an output layer comprising multiple neurons configured to generate a probability score for each of a plurality of predetermined diagnoses. A final diagnosis can be determined based on the output of the decision network. In some embodiments, the decision network can output a probability score for each of the plurality of predetermined diagnoses, wherein the probability score indicates the likelihood of a relevant diagnosis, and a diagnosis (or multiple diagnoses) can be selected as the final diagnosis based on the relevant probability scores.

[0108] After operation 708, method 700 can be completed. According to... Figure 7 The technical advantage of the method described in determining the diagnosis of ECG data lies in the fact that the decision network can be informed of an initial / preliminary diagnosis based on expert criteria. This offers several advantages, including reducing the decision network's sensitivity to variability in the received ECG data. Specifically, the decision network can be made less sensitive to minor changes in ECG data by incorporating a second diagnostic prediction using different diagnostic methods.

[0109] Go to Figure 8A flowchart of a third embodiment of a method 800 for automatically determining a diagnosis using hybrid ECG diagnostic data using a hybrid ECG diagnostic system is shown. Method 800 may be performed as part of method 500, as indicated in operation 504. In some embodiments, method 800 may be performed by a hybrid ECG diagnostic system 400, which may be implemented by ECG system 100, an edge device connected to ECG system 100, a cloud device communicating with ECG system 100, or any suitable combination thereof. Method 800 can enable automated, accurate, and use-case-specific diagnoses of one or more ECGs from one or more patients.

[0110] Method 800 begins at 802, wherein the acquired ECG data is received by a hybrid ECG diagnostic system, wherein the hybrid ECG diagnostic system includes a rule-based system (such as the rule-based system 404 described above) and a deep neural network (such as the deep neural network 410 described above). The received ECG data includes one or more ECG waveforms / signals that indicate the electrical activity of the heart over time. In some embodiments, the received ECG data may include data from a twelve-lead ECG, as referenced above. Figure 1 More detailed description. In some embodiments, the received ECG data may include data from a 15-lead ECG. In some embodiments, the received ECG data may include data from a reduced-lead ECG. In some embodiments, the received ECG data includes 2D data of potential variation over time, with potential plotted along the y-axis and time plotted along the x-axis. The received ECG data may be formatted as SCP-ECG, DICOM-ECG, HL7 aECG, or other storage formats known in the ECG field.

[0111] At operation 804, the received ECG data is input into a rule-based system and used by the rule-based system to determine a first preliminary diagnosis. In some embodiments, inputting the received ECG data into the rule-based system includes inputting the received ECG data into a feature extractor of the rule-based system. In some embodiments, the feature extractor may perform one or more preprocessing operations on the input ECG data before extracting features. The feature extractor may be configured as described above with reference to feature extractor 406 to determine one or more predetermined features present in the received ECG data. In some embodiments, the predetermined features may include one or more amplitudes, intervals / durations, and / or inter-amplitudes / durations.

[0112] The extracted features can be fed into a decision tree (such as decision tree 408 described above). The decision tree can determine the path from the root node to the leaf node based on one or more decisions made at one or more decision nodes, where the decisions are based on hard-coded expert criteria used to evaluate one or more extracted features determined by the feature extractor. Leaf nodes can include one or more diagnostic assessments of the received ECG data and can be used to determine preliminary diagnoses for input into the decision network of the deep neural network.

[0113] At operation 806, the received ECG data is input into the input layer of a convolutional neural network (such as convolutional neural network 412) of a deep neural network and mapped to multiple features. That is, at operation 806, the convolutional neural network extracts one or more features from the ECG data using one or more convolutional layers. The input layer of the convolutional neural network may include multiple neurons configured to receive ECG data. The convolutional neural network also includes one or more convolutional layers, each including one or more filters. The output from the convolutional neural network is generated by multiple neurons in the output layer of the convolutional neural network. In some embodiments, the output from the convolutional neural network includes one or more learned features present in the received ECG data. In some embodiments, the learned features may include one or more wave features or a combination of wave features. In some embodiments, the output from the convolutional neural network includes feature maps generated from the received ECG data using one or more convolutional layers.

[0114] At operation 808, multiple features output by the convolutional neural network, along with a first preliminary diagnosis determined by a rule-based system, are input into a decision network of a deep neural network (such as decision network 414) and mapped to a second preliminary diagnosis of the received ECG data. The decision network includes an input layer comprising multiple neurons configured to receive the output from the convolutional neural network and the first preliminary diagnosis determined by the rule-based system. In some embodiments, the input layer of the decision network includes a first plurality of neurons configured to receive multiple features output from the convolutional neural network and a second plurality of neurons configured to receive the first preliminary diagnosis determined by the rule-based system. The decision network also includes one or more fully / densely connected layers and an output layer comprising multiple neurons configured to generate a probability score for each of a plurality of predetermined diagnoses. The second preliminary diagnosis can be determined based on the output of the decision network. In some embodiments, the decision network can output a probability score for each of the plurality of predetermined diagnoses, wherein the probability score indicates the likelihood of a relevant diagnosis, and a diagnosis (or diagnoses) selected as the second preliminary diagnosis can be chosen based on the relevant probability scores.

[0115] At operation 810, the rule-based system receives a second preliminary diagnosis generated by a deep neural network and ECG data received at operation 802, and determines a final diagnosis based on these diagnoses. In some embodiments, the second preliminary diagnosis generated by the deep neural network can be used to adjust one or more parameters of the rule-based system. In some embodiments, the preliminary diagnosis can be used to select one or more types of predetermined features for extraction by the feature extractor of the rule-based system. In some embodiments, the second preliminary diagnosis, along with the features extracted by the feature extractor, can be fed into the decision tree of the rule-based system and used to determine the final diagnosis of the received ECG data.

[0116] After operation 810, method 800 can terminate. In this way, method 800 employs a recursive diagnostic scheme, in which a deep neural network can be used to map a first preliminary diagnosis determined by a rule-based system to a second preliminary diagnosis. The second preliminary diagnosis can then be used by the rule-based system to generate a final diagnosis. The technical advantage of the above recursive diagnostic scheme is that it can enhance diagnostic accuracy, confidence, and consistency.

[0117] refer to Figure 9 This demonstrates the methods used to train deep neural networks (such as...). Figure 2 , Figure 3 and Figure 4 The flowcharts for method 900 (using deep neural networks 210, 310, and 410, respectively) are shown in the figures. Method 900 can be derived from... Figure 1 The ECG system 100 shown is used for implementation. In some embodiments, method 900 may be implemented by a deep neural network module 108 stored in the non-transitory memory 106 of the ECG processing system 102. Although this document trains both the convolutional neural network and the decision network simultaneously and uses the same training dataset, it should be understood that this disclosure provides for training the convolutional neural network and the decision network separately using different training datasets.

[0118] Method 900 begins at operation 902, wherein training data pairs from multiple training data pairs may be fed into the input layer of a deep neural network, wherein the training data pairs include ECG data and corresponding benchmark truth diagnostics, and wherein the input layer of the deep neural network includes the input layer of a convolutional neural network cascaded to a decision network. In some embodiments, the training data pairs and multiple training data pairs may be stored in an ECG processing system, such as ECG data module 112 in ECG processing system 102. In other embodiments, the training data pairs may be obtained via a communication link between the ECG processing system and an external storage device (such as via an Internet connection to a remote server). In some embodiments, benchmark truth diagnostics are generated via expert analysis and / or via empirical analysis.

[0119] In some implementations, training data pairs are intelligently selected from stored ECG data based on the intended use case of the deep neural network. In some implementations, the ECG training data pairs fed to the deep neural network may include ECG data and a baseline truth diagnosis determined for the ECG data, as well as one or more metadata entries related to the ECG training data pair. The metadata may include details about the ECG training data pair, such as details of one or more patients (e.g., demographic details, health details, etc.), details of one or more ECG systems, and one or more ECG conditions (e.g., resting ECG, exercise ECG, post-infarction ECG, etc.). The metadata can enable automatic and intelligent selection of training data pairs based on one or more of the included details, thereby enabling the training of use case-specific deep neural networks. In some implementations, training data pairs intelligently selected based on included metadata can be used to train a deep neural network model for analyzing ECG data from elderly patients, where the included metadata indicates that the patient's age is greater than a threshold age.

[0120] At operation 904, a preliminary diagnosis determined by a rule-based system (such as rule-based systems 204, 304, and 404 described above) based on ECG data can be fed into a deep neural network. In some embodiments, the preliminary diagnosis can be fed into the input layer of a decision network connected in series with a convolutional neural network, wherein the input layer of the decision network is further configured to receive multiple features determined by the convolutional neural network based on ECG training data.

[0121] At operation 906, the deep neural network maps the input ECG data to a diagnosis by passing the ECG data through one or more convolutional layers of a convolutional neural network to extract multiple features from the ECG data, and by using a decision network to map the multiple features (and, in some embodiments, a preliminary diagnosis determined by a rule-based system) to a diagnosis. In some embodiments, the output of the deep neural network includes a vector of probability scores, where each row of the vector corresponds to a predetermined diagnosis, and where the probability score in the row corresponding to the predetermined diagnosis indicates the likelihood that the predetermined diagnosis is applicable to the ECG data.

[0122] At operation 908, the ECG processing system calculates the difference between the diagnosis generated by the deep neural network and the baseline truth diagnosis included in the ECG training data pair. In some embodiments, the baseline truth probability score for each diagnosis included in the baseline truth diagnosis may be 1, and the baseline truth probability score for each diagnosis not included in the baseline truth diagnosis may be 0, and calculating the difference between the diagnosis generated by the deep neural network and the baseline truth diagnosis may include calculating the difference between the predicted probability score and the baseline truth probability score for each diagnosis. In some embodiments, calculating the difference between the diagnosis and the baseline truth diagnosis may include inputting the calculated difference into a loss function, wherein the loss function may include one or more mathematical operations performed on the calculated difference.

[0123] At operation 910, the weights and biases of the deep neural network are adjusted based on the differences calculated at operation 908. The differences are backpropagated through layers of the decision network and convolutional network, starting from the output layer of the decision network and proceeding layer by layer to the input layer of the convolutional neural network. The weights and biases of each layer of the deep neural network can be adjusted based on the gradient (a first-order partial derivative or an approximation of the first-order partial derivative of the error function, relative to the weights or biases being evaluated). Each weight (and bias) of the deep neural network is then updated by adding the negative product of the gradients determined (or approximated) for the weights (or biases) to a predetermined step size. Method 900 can then terminate. It should be noted that method 900 can be repeated until the weights and biases of the deep neural network converge, the rate of change of the weights and / or biases of the deep neural network in each iteration of method 900 is below a threshold, or a validation error below a predetermined validation error threshold is obtained (determined using a validation dataset different from the training dataset). Once the deep neural network is complete, it can be stored in non-transitory memory. In some implementations, the trained deep neural network can be stored in a deep neural network library (such as...). Figure 1 The deep neural network library module 180 shown is included.

[0124] In this way, method 900 enables the hybrid ECG diagnostic system to learn the mapping from ECG data to diagnostics using intelligently selected ECG training data pairs. Method 900 allows the hybrid ECG diagnostic system to be more easily adapted to specific use cases by training on use-specific data.

[0125] Go to Figure 10An exemplary embodiment of ECG data 1002 (also referred to herein as electrocardiogram 1002) is shown, wherein a first diagnosis 1004 of the ECG data 1002 is generated independently by a rule-based system, and a second diagnosis 1006 of the ECG data 1002 is generated by a hybrid ECG diagnostic system according to an exemplary embodiment of the present disclosure. ECG data 1002 can be acquired by an ECG system (such as ECG system 100) using ten electrodes positioned to make electrical contact with the patient's skin. ECG data 1002 can be stored in a non-transitory memory, such as in ECG data module 112. ECG data 1002 includes a twelve-lead ECG, comprising twelve different waveforms / signals measured simultaneously for the same patient. Specifically, ECG data 1002 includes lead types I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, and V6, which can be referenced above. Figure 1 The measurement is described in more detail using ten electrodes.

[0126] Two exemplary implementations of the diagnostics are in Figure 10 As shown, the first diagnosis 1004 is generated using a rule-based system without a deep neural network component, while the second diagnosis 1006 is generated using a hybrid ECG diagnostic system (such as hybrid ECG diagnostic systems 200, 300, and 400) discussed herein. The first diagnosis 1004 includes a text message indicating an assessment of the overall normal / abnormality of sinus rhythm, QRS complex, and ECG data 1002, as well as one or more underlying physiological mechanisms of the indicated normal / abnormality. Similarly, the second diagnosis 1006 includes a message indicating a diagnostic assessment of the overall normal / abnormality of sinus rhythm, one or more physiological characteristics of the heart, and ECG 1002, as well as one or more underlying physiological mechanisms of the indicated normal / abnormality. The second diagnosis 1006 correctly identifies ECG data 1002 indicating right axis deviation and left bundle branch block, a determination not made by the first diagnosis 1004. Therefore, the diagnostic accuracy of the diagnosis determined via the hybrid ECG diagnostic system is demonstrated to be improved.

[0127] Go to Figure 11An exemplary embodiment of a deep neural network library 1104 is illustrated. The deep neural network library 1104 may be stored in the non-transitory memory of one or more computing systems. In one embodiment, the deep neural network library 1104 may be stored in a deep neural network library module 180 within the non-transitory memory 106 of an ECG processing system 102. The deep neural network library may be used in conjunction with one or more systems or methods described herein. In some embodiments, deep neural networks 210, 310, and 410, respectively shown in hybrid ECG diagnostic systems 200, 300, and 400, may be intelligently selected from deep neural network libraries such as deep neural network library 1104 based on the lead types present in the ECG data to be diagnosed. In some embodiments, in response to receiving reduced-lead ECG data to be diagnosed, the hybrid ECG diagnostic system may automatically select a trained deep neural network from a deep neural network library based on one or more metadata of the trained deep neural network. In one embodiment, in response to receiving twelve-lead ECG data with missing leads V1-V6, a deep neural network trained using the ECG data with missing leads V1-V6 may be selected from the deep neural network library. In this way, hybrid ECG diagnostic systems can determine diagnoses based on reduced lead ECG data without simulating one or more missing leads.

[0128] Deep neural network library 1104 includes multiple trained deep neural networks, including deep neural network 1110 and deep neural network 1130. In some embodiments, the deep neural networks included in deep neural network library 1104 include parameters learned via training on reduced-lead ECG data. Deep neural network library 1104 includes multiple trained deep neural networks, as indicated by the ellipsis between deep neural network 1110 and deep neural network 1130. In some embodiments, deep neural network library 1104 may include one or more trained deep neural networks for each reduced-lead group arrangement of a 12-lead ECG, wherein there are 4,094 reduced-lead group arrangements for a 12-lead ECG; therefore, in some embodiments, 4,094 different trained deep neural networks may be stored in deep neural network library 1104. In some embodiments, deep neural network library 1104 may include one or more trained deep neural networks for each reduced-lead group arrangement of a 15-lead ECG, wherein there are 32,766 reduced-lead group arrangements for a 15-lead ECG. As used in this article, reduced lead group permutations include different lead groups generated for a set of n lead types, where each permutation may include 1 to n-1 lead types (excluding reduced lead groups that contain all missing leads or none of the missing leads, as these are unimportant permutations).

[0129] The deep neural network 1110, comprising a cascaded convolutional neural network 1112 and a decision network 1114, may include multiple parameters (weights and biases) learned during training using a first reduced lead group (such as method 1200 described below). Similarly, the deep neural network 1130, comprising a convolutional neural network 1132 and a decision network 1134, may include multiple parameters learned during training using a second reduced lead group, wherein the first and second reduced lead groups are different (i.e., there is no one-to-one correspondence between the lead types of the first and second reduced lead groups). The deep neural networks 1110 and 1130 may be indexed based on one or more metadata entries, enabling rapid querying and selection of trained deep neural networks based on received ECG data. In some embodiments, deep neural networks stored in a deep neural network library 1104 may be indexed based on reduced lead group data on which they are trained, enabling automatic selection of deep neural networks based on received reduced lead ECG data.

[0130] In other words, by training multiple deep neural networks using different arrangements of reduced lead groups, and storing the trained deep neural networks in a deep neural network library based on the reduced lead ECG data on which the deep neural networks were trained, the acquired ECG data with one or more missing leads can be effectively and accurately mapped to a diagnosis by selecting a deep neural network from the deep neural network library trained on reduced lead ECG data of the same type of missing and acquired data.

[0131] refer to Figure 12 This demonstrates the methods used to train deep neural networks (such as...). Figure 2 , Figure 3 and Figure 4 The flowcharts for deep neural networks 210, 310, and 410 (shown respectively) are used to map reduced-lead ECG data to a diagnostic method 1200. Method 1200 can be derived from... Figure 1 The ECG system 100 shown is used for implementation. In some embodiments, method 1200 may be implemented by a deep neural network module 108 stored in the non-transitory memory 106 of the ECG processing system 102. Although this document trains both the convolutional neural network and the decision network simultaneously and uses the same training dataset, it should be understood that this disclosure provides for training the convolutional neural network and the decision network separately using different training datasets.

[0132] Method 1200 begins with operation 1202, in which ECG training data pairs comprising a twelve-lead group and corresponding baseline truth diagnostics are selected. In some embodiments, the ECG training data pairs may be selected from multiple training data pairs stored in non-transitory memory (such as in ECG data module 112 in ECG processing system 102). In other embodiments, the training data pairs may be obtained via a communication link between the ECG system and an external storage device (such as via an Internet connection to a remote server). In some embodiments, the baseline truth diagnostics are generated via expert analysis and / or via empirical analysis.

[0133] In some implementations, ECG training data pairs are intelligently selected from stored ECG data based on the intended use case of the deep neural network. In some implementations, the ECG training data may include 12-lead ECG data and a baseline truth diagnosis determined for the 12-lead ECG data, as well as one or more metadata entries related to the ECG training data pair. The metadata may include details about the ECG training data pair, such as details of one or more patients (e.g., demographic details, health details, etc.), details / settings of one or more ECG systems, and one or more ECG conditions (e.g., resting ECG, exercise ECG, post-infarction ECG, etc.). The metadata can enable automatic and intelligent selection of ECG training data pairs based on one or more of the included details, thereby enabling the training of a use case-specific deep neural network.

[0134] At operation 1204, one or more leads are selectively removed from the twelve-lead group to produce a reduced lead group. In some implementations, the reduced lead group may include one to eleven different lead types; that is, the twelve-lead group includes fewer than twelve leads and more than zero leads, each with a different lead type. In some implementations, each ECG lead included in the ECG training data pair includes an indication of its lead type. In some implementations, the number and type of leads to be removed may be selected by the user or determined automatically according to the training routine.

[0135] At operation 1206, the reduced lead group is fed into the deep neural network. In some implementations, at operation 1206, a reduced lead group comprising fewer than twelve leads is input into the input layer of a convolutional neural network cascaded to the decision network.

[0136] At operation 1208, a preliminary diagnosis determined by a rule-based system (such as rule-based systems 204, 304, and 404 described above) based on reduced-lead ECG data can be fed into a deep neural network. In some embodiments, the preliminary diagnosis can be fed into the input layer of a decision network connected in series with a convolutional neural network, wherein the input layer of the decision network is further configured to receive multiple features determined by the convolutional neural network based on ECG training data.

[0137] At operation 1208, the deep neural network maps the reduced lead group to a diagnosis by passing ECG data through one or more convolutional layers of a convolutional neural network to extract multiple features from the reduced lead group, and by using a decision network to map the multiple extracted features (and, in some embodiments, a preliminary diagnosis determined by a rule-based system) to the diagnosis. In some embodiments, the output of the deep neural network includes a vector of probability scores, where each row of the vector corresponds to a predetermined diagnosis, and where the probability score in the row corresponding to the predetermined diagnosis indicates the likelihood that the predetermined diagnosis is applicable to the reduced lead group.

[0138] At operation 1212, the difference between the diagnosis generated by the deep neural network and the baseline truth diagnosis included in the ECG training data pair is calculated. In some embodiments, the baseline truth probability score for each diagnosis included in the baseline truth diagnosis may be 1, and the baseline truth probability score for each diagnosis not included in the baseline truth diagnosis may be 0, and calculating the difference between the diagnosis generated by the deep neural network and the baseline truth diagnosis may include calculating the difference between the predicted probability score and the baseline truth probability score for each diagnosis. In some embodiments, calculating the difference between the diagnosis and the baseline truth diagnosis may include using a loss function, wherein the loss function may include one or more mathematical operations performed on the diagnosis and the baseline truth diagnosis.

[0139] At operation 1214, the weights and biases of the deep neural network are adjusted based on the differences calculated at operation 1212. These differences propagate back through the layers of the decision network and the convolutional neural network, starting from the output layer of the decision network and proceeding layer by layer to the input layer of the convolutional neural network. The weights and biases of each layer of the deep neural network can be adjusted based on the gradient (calculated as the first-order partial derivative of the error, or an approximation of the first-order partial derivative, relative to the weights or biases being evaluated). Each weight (and bias) of the deep neural network can then be updated by subtracting the product of the gradient (determined or approximated for the weights or biases being adjusted) multiplied by a predetermined step size.

[0140] At operation 1216, the ECG processing system determines whether the validation error of the deep neural network is below a threshold. The validation error can be determined using a validation dataset comprising multiple reduced lead groups and corresponding baseline truth diagnoses, including data in the validation dataset that differs from data included in the training dataset. The validation error can be determined by calculating the differences between multiple diagnoses predicted by the deep neural network based on the validation data and the corresponding multiple baseline truth diagnoses in essentially the same manner as described above with respect to operation 1212. If at operation 1216, it is determined that the validation error is not less than the threshold, then method 1200 proceeds to operation 1202, where a new ECG training data pair is selected, and operations 1204-1216 are performed on the new ECG training data pair.

[0141] However, if at 1216 the ECG processing system determines that the validation error of the deep neural network is below a threshold, then method 1200 proceeds to operation 1218, which includes storing the trained deep neural network in a deep neural network library (such as...). Figure 11 The deep neural network library 1104 shown is used in some embodiments. Storing the trained deep neural network in the deep neural network library includes indexing the trained deep neural network using one or more metadata entries, where the metadata may include the types of ECG leads present in the reduced lead groups used to train the deep neural network. The advantage of indexing the stored trained deep neural network based on the types of leads present in the reduced lead groups used during the training phase of the deep neural network is that the ECG processing system can quickly query and select the trained deep neural network for use in a hybrid ECG diagnostic system, which will be used to diagnose reduced lead ECG data containing the same reduced lead groups as the selected trained deep neural network. After operation 1218, method 1200 may end.

[0142] Go to Figure 13 A flowchart illustrating an exemplary method 1300 for diagnosing reduced-lead ECG data using a hybrid ECG diagnostic system is shown. In some embodiments, method 1300 may be performed by ECG system 100, an edge device connected to ECG system 100, a cloud device communicating with ECG system 100, or any suitable combination thereof. Method 1300 enables automated, accurate, and use-case-specific diagnosis of one or more reduced-lead ECGs for one or more patients.

[0143] Method 1300 begins with operation 1302, wherein reduced-lead ECG data is acquired by an ECG system. Reduced-lead ECG data may include one or more lead signals / waveforms indicating the electrical activity of the heart over time. Reduced-lead ECG data may lack one or more lead signals. In some embodiments, reduced-lead ECG data may include fewer than twelve leads and more than zero leads. In some embodiments, acquiring reduced-lead ECG data may include measuring fewer than twelve leads using fewer than ten electrodes in electrical contact with the patient, as referenced above. Figure 1 A more detailed description follows. The acquired ECG data may be stored in memory for later processing. In some embodiments, the reduced-lead ECG data acquired at operation 1302 includes 2D data of potential changes over time, with potential plotted along the y-axis and time plotted along the x-axis. The reduced-lead ECG data acquired at operation 1302 may be stored in one or more formats, including SCP-ECG, DICOM-ECG, HL7aECG, and other storage formats known in the ECG field. Acquiring reduced-lead ECG data at operation 1302 can occur under various conditions, such as when the patient is at rest, during exercise, or in an ambulance.

[0144] At operation 1304, the missing lead types from the acquired reduced-lead ECG data are determined. In some embodiments, each ECG lead included in the reduced-lead ECG data may include a lead type identifier, wherein the lead type identifier may include a piece of metadata associated with the reduced-lead ECG data. Determining the missing lead types may include comparing the lead type identifiers included in the reduced-lead ECG data with a standard lead type group. In some embodiments, determining which lead types are missing from the acquired reduced-lead ECG data may include comparing a set of lead type identifiers from the reduced-lead ECG data with a standard 12-lead group. In some embodiments, determining which lead types are missing from the acquired reduced-lead ECG data may include comparing a set of lead type identifiers from the reduced-lead ECG data with a standard 15-lead group.

[0145] At operation 1306, a deep neural network for diagnosing reduced-lead ECG data is selected from a deep neural network library based on the type of missing leads determined at operation 1304. The selected deep neural network can be used in hybrid ECG diagnostic systems (such as those referenced above). Figure 2 , Figure 3 , Figure 4Among those described, diagnostics are automatically determined from the acquired reduced-lead ECG data. In some embodiments, selecting a deep neural network to train from a deep neural network library includes querying the deep neural network library using one or more missing lead types determined for the reduced-lead ECG data, and selecting a deep neural network that matches the missing lead types from the ECG training data used to train the deep neural network based on the missing ECG lead types from the reduced-lead ECG data. In some embodiments, selecting a deep neural network to train from a deep neural network library includes querying the deep neural network library using one or more leads included in the reduced-lead ECG data, and selecting a deep neural network that matches the lead types present in the acquired reduced-lead ECG data based on the lead types present in the ECG training data used to train the deep neural network. Deep neural networks within a deep neural network library may be indexed based on one or more attributes of the ECG training data used during the training phase. In one embodiment, deep neural networks within a deep neural network library may be indexed based on lead types present (and / or missing) in the ECG training data used during the training phase. By selecting a deep neural network trained using the same reduced lead group as the acquired reduced lead ECG data, more accurate and consistent diagnoses can be made without simulating one or more missing leads.

[0146] At operation 1308, one or more of the hybrid ECG diagnostic systems 200, 300, and / or 400 are used, employing the deep neural network selected at operation 1306, to determine the diagnosis of the reduced-lead ECG data acquired at operation 1302. The hybrid ECG diagnostic system may use one or more of the steps of methods 600, 700, and 800, described in more detail above, to determine the diagnosis of the reduced-lead ECG data. In a first exemplary embodiment, operation 1308 includes the hybrid ECG diagnostic system 200 performing one or more operations of method 600 to determine a diagnosis based on the acquired ECG data. In a second exemplary embodiment, operation 1308 includes the hybrid ECG diagnostic system 300 performing one or more operations of method 700 to determine a diagnosis based on the acquired ECG data. In a third embodiment, operation 1308 includes the hybrid ECG diagnostic system 400 performing one or more operations of method 800 to determine a diagnosis based on the acquired ECG data. In some implementations, the ECG system execution method 1300 may select between a first implementation, a second implementation, and a third implementation of the hybrid ECG diagnostic system based on one or more characteristics of the acquired ECG data.

[0147] In this way, method 1300 intelligently selects a deep neural network from a deep neural network library based on one or more lead types missing from the acquired reduced-lead ECG data and incorporates the selected deep neural network into a hybrid ECG diagnostic system, enabling the acquisition of reduced-lead ECGs to be quickly, accurately, and computationally efficiently mapped to a diagnosis using the hybrid diagnostic system.

[0148] When describing elements of various embodiments of this disclosure, the terms “an,” “a,” and “the” are intended to refer to one or more of these elements. The terms “first,” “second,” etc., do not indicate any order, quantity, or importance, but are used to distinguish one element from another. The terms “comprising,” “including,” and “having” are intended to be inclusive and mean that additional elements may exist in addition to the listed elements. As used herein, the terms “connected to,” “linked to,” etc., indicate that an object (e.g., a material, element, structure, component, etc.) may be connected to or linked to another object, regardless of whether the object is directly connected to or linked to the other object, or whether one or more intervening objects exist between the object and the other. Furthermore, it should be understood that references to “an embodiment” or “an embodiment” of this disclosure are not intended to be construed as excluding the existence of additional embodiments that also include the referenced features.

[0149] In addition to any modifications previously indicated, those skilled in the art can devise many other variations and alternative arrangements without departing from the spirit and scope of this description, and the appended claims are intended to cover such modifications and arrangements. Therefore, although the information has been described above in a specific and detailed manner in conjunction with what is currently considered to be the most practical and preferred aspects, it will be apparent to those skilled in the art that many modifications can be made without departing from the principles and concepts set forth herein, including but not limited to changes in form, function, mode of operation, and use. Likewise, as used herein, in all respects, examples and embodiments are intended to be illustrative only and should not be construed as restrictive in any way.

Claims

1. A method, the method comprising: Acquire reduced-lead ECG data (reduced-lead ECG data), wherein the reduced-lead ECG data includes signals from fewer than twelve leads; Determine the type of each of the fewer than twelve lead signals; Determine the type of lead missing from the reduced lead ECG data; A deep neural network is selected based on the type of each of the fewer than twelve lead signals and the type of the missing lead; and The deep neural network is used to map the fewer than twelve leads to a diagnosis.

2. The method of claim 1, wherein the type of each of the fewer than twelve leads is one of I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, and V6, and wherein the type of each of the fewer than twelve leads is different from each of the other types of the fewer than twelve leads.

3. The method according to claim 1, wherein the deep neural network includes a convolutional neural network and a decision network.

4. The method of claim 3, wherein mapping the fewer than twelve lead signals to the diagnosis using the deep neural network comprises: The convolutional neural network is used to map the fewer than twelve lead signals to multiple features; as well as The decision network is used to map the multiple features from the convolutional neural network to the diagnosis.

5. The method according to claim 4, further comprising: Using a rule-based system to determine a preliminary diagnosis based on fewer than twelve leads, and wherein using the decision network to map the plurality of features from the convolutional neural network to the diagnosis further includes: The preliminary diagnosis is input into the input layer of the decision network; and The decision network is used to map both the preliminary diagnosis and the multiple features to the diagnosis.

6. The method of claim 1, wherein the deep neural network is stored in a deep neural network library and indexed based on a training dataset used to train the deep neural network.

7. The method of claim 6, wherein selecting the deep neural network based on the type of each of the fewer than twelve leads comprises matching the type of each of the fewer than twelve leads with an index of the deep neural network, wherein the index indicates the type of lead used to train the deep neural network.

8. The method of claim 1, wherein acquiring reduced-lead ECG data comprises measuring the fewer than twelve lead signals using fewer than 10 electrodes.

9. A method for training a deep neural network to automatically diagnose reduced-lead electrocardiograms (ECGs), the method comprising: Select an ECG training data pair that includes lead signal groups and a baseline truth diagnosis corresponding to the lead signal groups; Selectively remove one or more lead signals from the lead signal group to produce a reduced lead signal group; Determine the type of each lead signal in the reduced lead signal group; Determine the type of lead missing from the reduced lead signal group; The reduced lead signal group is fed into the deep neural network based on the type of each lead signal in the reduced lead signal group and the type of missing lead; Map the reduced lead signal group to the diagnostics; Calculate the difference between the diagnosis and the baseline truth diagnosis; and Adjust one or more parameters of the deep neural network based on the calculated differences.

10. The method according to claim 9, further comprising: Determine the verification error of the deep neural network; as well as The following process is used to respond when the verification error is below a predetermined threshold: Store the deep neural network in a deep neural network library; and The deep neural network is indexed in the deep neural network library based on one or more removed lead signals.

11. The method of claim 9, wherein the lead signal group is a 12-lead electrocardiogram of the patient, and wherein the baseline truth diagnosis corresponding to the lead signal group includes an expert-generated diagnosis determined for the patient.

12. The method of claim 9, wherein adjusting one or more parameters of the deep neural network based on the calculated differences comprises performing backpropagation using the calculated differences to adjust a plurality of weights and / or a plurality of biases of the deep neural network.

13. The method of claim 9, wherein the deep neural network comprises a convolutional neural network cascaded to a decision network, and wherein mapping the reduced lead signal group to the diagnosis comprises: The convolutional neural network is used to map the reduced lead signal group to multiple features; as well as The decision network is used to map the multiple features from the convolutional neural network to the diagnosis.

14. The method according to claim 13, further comprising: Using a rule-based system to determine a preliminary diagnosis based on the reduced lead signal group, wherein using the decision network to map the plurality of features from the convolutional neural network to the diagnosis further includes: The preliminary diagnosis is input into the input layer of the decision network; and The decision network is used to map the preliminary diagnosis and the multiple features to the diagnosis.

15. An electrocardiogram (ECG) processing system, the ECG processing system comprising: Display devices; The memory stores deep neural network libraries and instructions; and A processor, communicatively connected to the memory and the display device, and configured, when executing the instructions, to: Acquire reduced-lead ECG data, wherein the reduced-lead ECG data includes signals from fewer than twelve leads; Determine the type of each of the fewer than twelve lead signals; Determine the type of lead missing from the reduced lead ECG data; A deep neural network is selected from the deep neural network library based on the type of each of the fewer than twelve lead signals and the type of the missing lead; The deep neural network is used to map the fewer than twelve lead signals to a diagnosis; and The diagnosis is displayed via the display device.

16. The system of claim 15, further comprising fewer than 10 electrodes, wherein the system is configured to acquire reduced-lead ECG data using the fewer than 10 electrodes.

17. The system of claim 15, wherein the deep neural network comprises a convolutional neural network and a decision network.

18. The system of claim 17, wherein the processor is configured to map the fewer than twelve leads to the diagnosis using the deep neural network through the following process: The convolutional neural network is used to map the fewer than twelve lead signals to multiple features; and The decision network is used to map the multiple features from the convolutional neural network to the diagnosis.

19. The system of claim 17, wherein the memory further comprises a rule-based ECG diagnostic system, and wherein the processor is further configured to map the fewer than twelve lead signals to the diagnostic using the deep neural network through the following process: The convolutional neural network is used to map the fewer than twelve lead signals to multiple features; The rule-based ECG diagnostic system is used to determine a preliminary diagnosis based on fewer than twelve lead signals; and The decision network is used to map both the preliminary diagnosis and the multiple features to the diagnosis.

20. The system of claim 15, wherein the deep neural network library comprises 1 to 4,094 trained deep neural networks, each of the trained deep neural networks comprising parameters learned by training on training data comprising different reduced ECG lead groups.

Images