Computer-Implemented Method and Non-Transitory Computer-Readable Medium

Abstract

To provide systems, methods, devices and techniques for estimating an ejection-fraction characteristic of a mammal.SOLUTION: An electrocardiogram (ECG) procedure is performed on a mammal, and a computer system obtains ECG data that describes results of the ECG over a period of time. The system provides a predictive input that is based on the ECG data to an ejection-fraction predictive model, such as a neural network or other machine-learning model. In response, the ejection-fraction predictive model processes the input to generate an estimated ejection-fraction characteristic of the mammal. The system outputs the estimated ejection-fraction characteristic of the mammal for presentation to a user.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] This specification describes computer-based techniques for analyzing physiological electrical data (eg, electrocardiogram data), such as to estimate a patient's ejection fraction characteristics.

[0002] [CROSS REFERENCE TO RELATED APPLICATIONS] This application claims priority to U.S. Patent Application No. 62 / 599,163, filed December 15, 2017, and U.S. Patent Application No. 62 / 569,268, filed October 6, 2017. The disclosures of these prior applications are considered part of the disclosure of this application and are incorporated herein in their entireties.

[0003] [background] The ejection fraction is a primary measure of cardiac health in humans and other mammals. The ejection fraction typically indicates the amount of blood ejected from the heart with each pump. With each pumping cycle (i.e., "heartbeat" or "cardiac cycle"), the heart muscle contracts and relaxes to push blood into the arterial system of the subject. As the heart relaxes, the ventricles fill with blood. Then, during the systole of the pumping cycle, a portion of the blood is sent from the ventricles through the aorta to the arterial system. However, not all of the blood that filled the ventricles during relaxation can be pumped out during contraction. The percentage of blood that is actually ejected from the ventricles during a pumping cycle is referred to as the ejection fraction. Although the ejection fraction often only pertains to measuring the portion of blood ejected from the left ventricle, for purposes of this specification, the ejection fraction can alternatively refer to the portion of blood ejected from only the left ventricle, only the right ventricle, or both ventricles, and can indicate the strength or health of the heart muscle.

[0004] Generally, an ejection fraction greater than about 50-55 percent is considered normal for humans. However, some people (often unknowingly) have a low ejection fraction, such as below the 50-55 percent range, or a very low ejection fraction, such as below 35 percent. This is called asymptomatic ventricular dysfunction (ASVD). A low or very low ejection fraction is often a marker of serious cardiac complications, such as cardiac arrest, sudden death, and various stages of heart failure. When ASVD is diagnosed, there are effective therapies to treat ASVD or prevent these complications.

[0005] Ejection fraction has traditionally been measured using echocardiography. During an echocardiogram, a sonographer uses a sound transducer to produce images of the heart using specialized techniques. Although echocardiography is non-invasive, it often requires a specialist to perform and interpret the procedure, typically requiring expensive equipment and infrastructure within the patient's healthcare provider's laboratory. Ejection fraction has also been measured using magnetic resonance imaging (MRI) techniques, computed tomography (CT) techniques, and nuclear medicine scans, all of which require specially trained personnel and expensive equipment.

[0006] [Summary] Disclosed herein are systems, methods, devices, and other techniques for estimating a mammal's cardiac ejection fraction from an electrocardiogram (ECG). An electrocardiogram is a measurement of the electrical activity of a mammal's heart. The pumping action of the heart is driven by continuous cycles of electrical polarization and depolarization of the myocardium. This electrical activity can be captured by an electrocardiogram, whereby electrodes are placed on a subject's body surface (e.g., on the subject's chest and limbs) and the electrical potential between each pair of electrodes over time is measured. The electrical signals captured by this process form an electrocardiogram. When an electrocardiogram is graphed to show the changes in electrical potential between the electrodes over time, one can see a waveform or ECG tracing that shows the polarization and depolarization of the heart during each of one or more cardiac cycles. For purposes of this specification, an electrocardiogram can include any number of leads, including the traditional 12 leads, additional leads, or as little as a single lead. Additionally, the ECG may be acquired from adhesive electrodes, conductive electrodes, capacitive electrodes, handheld electrodes, wearable / garment-based electrodes, subcutaneous electrodes, electrodes attached to an implanted device, or any combination of these.

[0007] The shape of the ECG waveform is influenced by multiple factors. The techniques disclosed herein are premised in part on the recognition that the ECG of a subject may be influenced by the cardiac ejection fraction. For example, a subject with a low or very low ejection fraction compared to another subject with a normal ejection fraction may impart minute deformations to one or more parts of the ECG waveform. Underlying diseases affecting the heart, whether from atherosclerosis, myopathy, inflammation, or valvular disorders from any source, may impair the pumping ability of the myocardium. Underlying diseases may also affect the metabolism of individual myocytes or their interconnections, leading to fibrotic deposits or infiltration of inflammatory cells, all of which result in minute electrical changes. These local cardiac electrical changes contribute to the deformations recorded on the surface ECG. Such deformations may not be visible to the naked eye, but may nevertheless be detectable using a computer-based model according to the techniques disclosed herein. Thus, the present specification describes how a subject's ECG can be used as a screening tool to predict the subject's ejection fraction characteristics. In many scenarios, ECG-based screening is preferred over measuring ejection fraction by echocardiography or other means. ECG equipment is often more widely available than echocardiography equipment and can be performed more quickly without a trained sonographer. Furthermore, consumer and home ECG equipment (e.g., single-lead ECG patch) is becoming increasingly popular. Using the techniques disclosed herein, the number of people screened for low ejection fraction and the number of screenings can be increased using ECG equipment in various settings, whether clinical or home. In some cases, if a subject's ECG screening indicates a high probability of low or very low ejection fraction, further evaluation of the condition can be performed to verify the condition, for example, by using other measurement means such as echocardiography, MRI, or CT scan.

[0008] The subject matter disclosed herein includes a computer-implemented method. The method may be implemented by a system of one or more computers in one or more locations. In some aspects, the system has one or more processors and one or more computer-readable media encoded with instructions that, when executed by the one or more processors, cause the processors to implement the method. Some aspects include only a computer-readable medium encoded with instructions that, when executed, cause the processors to implement the method.

[0009] The method can include receiving, by a system, electrocardiogram (ECG) data describing an ECG of the mammal over a period of time. The system provides predictive inputs derived from the ECG data to an ejection fraction prediction model. The ejection fraction prediction model can be used to process the predictive inputs to generate an estimated ejection fraction characteristic of the mammal. The system then provides as an output the estimated ejection fraction characteristic of the mammal.

[0010] These and other implementations can optionally further include one or more of the following features.

[0011] The ECG data can include one or more channels, each channel including a subset of the ECG data describing a respective lead of the mammal's ECG over a period of time, and the prediction input can characterize a respective lead of the ECG for each of the one or more channels of ECG data.

[0012] The ECG data can include a plurality of channels, each channel including a subset of the ECG data describing a respective lead of a plurality of leads of the ECG of the mammal during a period of time, and the prediction input can characterize the plurality of leads of the ECG for each of the plurality of channels of ECG data.

[0013] The period of time over which the ECG data describes the mammal's ECG may span multiple cardiac cycles of the mammal.

[0014] The mammal may be a human.

[0015] The ejection fraction prediction model may be a regression model, for example a logistic regression model.

[0016] The ejection fraction prediction model can be a machine learning model, such as a neural network (e.g., a feedforward neural network, a convolutional neural network, or a recurrent neural network).

[0017] The ejection fraction model can use clinical characteristics (including age, sex, weight, and / or the presence or absence of measurable medical conditions such as hypertension and diabetes) as inputs to further refine its output.

[0018] The ejection fraction model can also generate additional outputs, such as risk of mortality over a given time period (eg, risk of mortality in the next year).

[0019] The estimated ejection fraction characteristic of the mammal may be a value that represents an absolute (eg, intrinsic) estimate of the mammal's ejection fraction.

[0020] The estimated ejection fraction profile of the mammal may indicate an estimated range of the ejection fraction of the mammal.

[0021] Providing the estimated ejection fraction characteristic of the mammal as an output may include providing the estimated ejection fraction characteristic for presentation to the mammal or to a health care provider associated with the mammal.

[0022] The method can further include generating a predictive input by determining a value for one or more morphological features of the ECG of the mammal. The predictive input can be indicative of a value for the one or more morphological features of the ECG of the mammal.

[0023] The morphological features of the mammalian ECG can include at least one of T wave amplitude, P wave amplitude, P wave area, T wave area, left T wave slope, right T wave slope, left P wave slope, right P wave slope, T wave duration, P wave duration, PR interval, QRS duration, QRS amplitude, QRS area, QRS energy, QRS peak-to-peak ratio, or QT segment length.

[0024] The ejection fraction prediction model may be personalized for the mammal.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the invention pertains. Although the present invention can be practiced using methods and materials similar or equivalent to those described herein, suitable methods and materials are described below. All publications, patent applications, patents, and other references described herein are incorporated by reference in their entirety. In case of conflict, the present specification, including definitions, will control. In addition, the materials, methods, and examples are for illustrative purposes only and are not intended to be limiting.

[0026] The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will be apparent from the description and drawings, and from the claims. [Brief description of the drawings]

[0027] [Figure 1] FIG. 1 is a conceptual diagram of an exemplary system for recording and processing ECG data and using the ECG data to estimate a subject's ejection fraction characteristics. [Diagram 2] 1 is a flow diagram of an exemplary process for estimating a subject's ejection fraction characteristics using ECG data. [Diagram 3] 1 is an exemplary ECG tracing for one heart beat of a patient, the ECG tracing showing the constituent segments of the heart beat and various morphological waveform features. [Figure 4]1 is a flow diagram of an exemplary process for selecting and using an appropriate ejection fraction prediction model corresponding to specified characteristics of a subject. [Diagram 5] 1 is a flow diagram of an exemplary process for using ECG-based ejection fraction estimation as a screening tool to determine whether further evaluation is warranted. [Figure 6] 1 is a flow diagram of an example process for training an ejection fraction prediction model, such as a neural network-based model. [Figure 7] 1 is a flow diagram of an exemplary process for predicting survival of a subject from a cardiac condition, such as low or very low ejection fraction, from ECG data for the subject. [Figure 8] FIG. 1 is an exemplary neural network model for estimating a subject's ejection fraction characteristics from ECG data. [Figure 9] FIG. 1 illustrates the results of a first neural network implementation of an ejection fraction prediction model that classifies patients into groups according to whether their ejection fraction is predicted to be above or below 50 percent. [Figure 10] FIG. 13 illustrates the results of a second neural network implementation of an ejection fraction prediction model that classifies patients into groups according to whether their ejection fraction is predicted to be above or below 35 percent. [Figure 11] 1 is a graph showing the correlation between model output and true ejection fraction values ​​in a study. [Figure 12] FIG. 1 is a block diagram of an example computing device that can be used to implement the systems, methods, devices, and other techniques described herein. [Figure 13] Figure 1 is a chart depicting the data mining schema for training, validation, and testing of the convolutional neural network ejection fraction prediction model. To avoid cross contamination between datasets and training the network based on a particular patient's ECG, the ECG for each patient was used only once (in only one of the groups). Patients with more than one data point were used for follow-up analysis. [Figure 14] Figure 1 shows the receiver operating characteristic curve for identifying patients with EF≦35% using a convolutional neural network. The ROC and area under the curve (AUC) were calculated using the validation and test (holdout) data sets. The identical AUC demonstrates the robustness of the algorithm to different data sets. [Figure 15] FIG. 1 is a pie chart showing the distribution of ejection fraction according to network classification. Of patients classified as having low ejection fraction, 63.5% had an ejection fraction below 50%. Of those classified as normal, 1.3% had an ejection fraction ≦35% and 90.1% had an EF ≧50%. [Figure 16] 1 is a table showing various patient characteristics and co-morbidities for patients included in an exemplary study and evaluation of ejection fraction prediction models. [Figure 17] FIG. 13 shows the long-term outcomes of patients with an EF ≥ 50% at the time of initial classification. Patients who initially had normal EF but were classified as having low EF by the network had a significantly increased risk of future LV dysfunction compared to patients who had normal EF and were classified as normal. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0028] Like references and labels in the drawings indicate like elements.

[0029] Disclosed herein are systems, methods, devices, and other techniques for estimating a cardiac ejection fraction of a mammal based on an electrocardiogram (ECG) of the mammal. In some implementations, a machine learning model, such as a neural network, is configured to process predictive inputs characterizing the ECG data and output an indication of the mammal's estimated ejection fraction. The model can be trained to account for complex combinations of features that are otherwise imperceptible to humans but that have been determined (e.g., by an iterative training process) to correlate with particular ejection fraction characteristics. Additional details regarding these and other techniques are provided below in the description of FIGS. 1-12.

[0030] 1 is a conceptual diagram of an exemplary system 100 for recording and processing ECG data and using the ECG data to estimate ejection fraction characteristics of a subject 102. For purposes of this example, the subject 102 is considered to be a human being, and more specifically, a patient of a healthcare provider. However, it should be understood that the description is not limited to this example. In other implementations, the subject 102 can be a human being not specifically associated with a healthcare provider, or can be any other mammal, and a suitable model is constructed to map the ECG data to estimate the ejection fraction for that mammal.

[0031] A set of electrodes 104 are placed on the body surface of the patient 102 to allow for recording of signals indicative of electrical activity of the patient's 102 heart during an ECG procedure. In some implementations, ten leads can be attached to the patient to perform a standard 12-lead ECG recording (e.g., some electrodes are placed on the chest near the area of ​​the heart and other electrodes are placed on the patient's extremities). A 12-lead ECG is useful for acquiring multiple channels of ECG data, with data for each channel representing a respective lead. Each lead is formed by the potential between a pair of electrodes. Due to variations in electrode positioning, different angles are formed by different pairs of electrodes for different leads, resulting in each lead providing a different view of the patient's heart's electrical activity. For example, signals from each lead can be recorded simultaneously during a period (e.g., 5, 10, or 15 seconds) to capture information regarding the timing and location of electrical activity along different radial directions.

[0032] In some implementations, ECG data can be recorded using a sensor platform or electrode configuration other than or in addition to a 12-lead ECG configuration. For example, a removable ECG patch can be attached to a patient's body surface, the patch including two or more electrodes that form one, two, or more leads from which an ECG can be recorded. In some implementations, the patient 102 can manually contact a pair of fixed external electrodes with their fingers, or the patient 102 can wear a watch, wristband, chest band, or other device that secures two or more electrodes in place on the patient 102 to sense the electrical activity of the patient's heart. The patient 102 or a healthcare provider can use a mobile computing device (e.g., smartphone 106) to configure aspects of an ECG procedure in a clinical or non-clinical setting.

[0033] The system 100 further includes a data recorder and control unit 108, an ECG processing system 110, and one or more input / output devices 112a-b. In some implementations, wires extending from each of the electrodes 104 are coupled to the data recorder and control unit 108 to enable recording of electrical signals sensed by the electrodes. The unit 108 can include, for example, an analog-to-digital (A / D) converter and other analog or digital signal conditioning circuitry, such as amplifiers, filters, etc.

[0034] The data recorder and controller unit 108 is communicatively coupled to an ECG processing system 110. The ECG processing system 110 is a system of one or more computers, which may be distributed in one or more physical locations. The system 110 may or may not be in the same physical location as the patient 102 and the data recorder and control unit 108. The system 110 may be coupled to peripheral devices such as a display screen 112a for presenting information to a user, and a mouse and keyboard 112 for receiving user input. Additionally, the system 110 may include various components 114-126 that facilitate processing of the ECG data to determine estimated ejection fraction characteristics, estimated survival rates, and present information regarding such estimates.

[0035] For example, the system 110 may include a pre-processor 114 that digitally conditions the ECG signal received from the data recorder and control unit 108. The pre-processor 114 may perform noise reduction, anti-aliasing, or other digital techniques to prepare the ECG data describing the ECG of the patient 102 for further processing.

[0036] The system 110 further includes one or more ejection fraction prediction models 118. These models 118 are generally configured to process one or more prediction inputs characterizing the patient's ECG data and generate an estimated ejection fraction characteristic for the patient 102 based on the prediction inputs. The ejection fraction characteristic may indicate, for example, an absolute estimate of the patient's ejection fraction (e.g., a particular value, such as 50 percent, 45 percent, 40 percent, 35 percent, 30 percent, or another value) or a classification of the patient's estimated ejection fraction (e.g., a normal ejection fraction greater than 50 percent, a low ejection fraction between 35-50 percent, or a very low ejection fraction below 35 percent).

[0037] As described in more detail below, the ejection fraction prediction model 118 can be a regression model, a machine learning model, or both. In some implementations, the model 118 is a feed-forward, recurrent, or convolutional neural network, or a capsule network. The neural network model can have a fully connected layer and can use an autoencoder network. In some implementations, the system 110 stores and maintains multiple ejection fraction prediction models 118. Each model 118 can correspond to a different set of patient characteristics (e.g., age, weight, sex, or other characteristics). When evaluating the ejection fraction for a new patient 102, the system 110 can select an appropriate model from the models 118 that corresponds to characteristics that match the characteristics of the patient 102. Each model 118 can be trained, for example, on data points from a patient that has the corresponding characteristics specifically for the model 118. Other types of machine learning or regression models can also be applied, such as support vector machines (SVMs), hidden Markov models (HMMs), and other linear and nonlinear systems.

[0038] In some implementations, the system 110 maintains one or more survival estimation models 120. The survival estimation model 120 is configured to process the ECG data, the estimated ejection fraction characteristic, or both, and generate a prognosis of the patient's estimated future survival rate (e.g., 1, 2, 5, and / or 10 year survival rate) based on the patient's ejection fraction characteristic, the prognosis being indicative of the likelihood of surviving a cardiac condition. For example, the ejection fraction characteristic model 118 may be used to generate an estimated ejection fraction characteristic for the patient. The survival estimation model 120 may then map the estimated ejection fraction characteristic to a future survival rate for the patient, as determined from empirical data. For example, the estimated 5 year survival rate for a patient with a very low ejection fraction may be lower than a patient with a normal ejection fraction. In other implementations, the survival estimation model 120 may process a prediction input characterizing the patient's ECG and derive an estimated survival rate for the patient 102 directly from the prediction input, without first starting with an estimated ejection fraction for the patient. In this manner, characteristics of the patient's ECG other than or in addition to the ejection fraction can be considered when determining the estimated survival rate for the patient 102. The survival estimation model 120 can be a regression model, a machine learning model, or both. In some implementations, the model 120 is a feed-forward, recurrent, or convolutional neural network, or a capsule network. The neural network model can have a fully connected layer and can use an autoencoder network. In some implementations, the system 110 stores and maintains multiple survival estimation models 120. Each model 120 can correspond to a different set of patient characteristics (e.g., age, weight, sex, or other characteristics). When evaluating the ejection fraction for a new patient 102, the system 110 can select an appropriate model from the models 120 that corresponds to characteristics that match the characteristics of the patient 102. Each model 120 can be trained, for example, on data points from a patient having characteristics that correspond specifically to the model 120.

[0039] A predicted input generator 116 processes the ECG data from the data recorder and control unit 108 or the ECG signal pre-processor 114 to generate a predicted input suitable for processing by the ejection fraction prediction model 118 or the survival estimation model 120. For example, the predicted input generator 116 may normalize and vectorize the ECG data from one or more channels (corresponding to one or more leads) into a format expected by the ejection fraction prediction model 118. In some implementations, the predicted input includes a time series of values ​​indicative of the amplitude of the ECG for one or more leads at each time point at a specified sampling frequency during a period spanning one or more cardiac cycles (e.g., 1, 2, 5, or 10 seconds). In some examples, the predicted input may represent an ECG signal spanning a relatively short amount of time (e.g., 1, 2, 5, or 10 seconds), while in other examples, the predicted input may represent an ECG signal for a relatively longer period of time (e.g., 30 seconds, 1 minute, 2 minutes, 5 minutes, 10 minutes, or more). In some implementations, the predicted input generator 116 averages ECGs from two or more cardiac cycles to generate an average ECG that represents an average cardiac cycle of the patient 102. In some implementations, the predicted input generator 116 selects a portion of the ECG that corresponds to a single cardiac cycle and characterizes only the selected portion of the ECG in the predicted input. The selected portion is thus representative of a cardiac cycle of the patient 102.

[0040] In some implementations, rather than processing actual time series data representing the amplitude of the ECG waveform over time, the ejection fraction prediction model 118 and the survival estimation model 120 may instead process prediction inputs indicative of the values ​​of one or more morphological features of the patient's ECG. A morphological feature is a parameter that characterizes the shape of an ECG waveform or a portion of an ECG waveform, such as a P wave, a QRS complex, or a T wave. Exemplary morphological features that may be identified in the prediction input include T wave amplitude, P wave amplitude, P wave area, T wave area, T wave left tilt, T wave right tilt, P wave left tilt, P wave right tilt, T wave duration, P wave duration, PR interval, QRS duration, QRS amplitude, QRS area, QRS energy, QRS peak-to-peak ratio, QT segment length, or a combination of two or more thereof. A number of these features are graphically illustrated in the exemplary ECG waveform 300 of FIG. 3. In some implementations, the system 110 includes a feature extractor 126 that analyzes the ECG data and determines the values ​​of any relevant morphological features to comprise prediction inputs that are processed by the ejection fraction prediction model 118.

[0041] The system 110 further includes a training subsystem 122 and a notification manager 124. The training subsystem 122 is configured to train the ejection fraction prediction model 118, the survival estimation model 120, or both. An exemplary process 600 implemented by the training subsystem 122 to train the ejection fraction prediction model 118 is described with respect to FIG. 6. The notification manager 124 is configured to provide the estimated ejection fraction characteristic, the estimated survival rate, or both as output to one or more users. In some implementations, the notification manager 124 provides the estimated ejection fraction characteristic for display on the screen 112a. In addition, one or more notification services can be registered with the computing system 110, such that the notification manager 124 puts the results of the estimated ejection fraction characteristic, the survival rate estimation, or both, into each of the registered services for presentation by one or more means (e.g., to a smartphone at a healthcare provider's clinic for visual, audible, and / or tactile presentation).

[0042] FIG. 2 is a flow diagram of an exemplary process 200 for estimating ejection fraction characteristics of a subject using ECG data. Process 200 can be implemented by a computing system, such as system 110 of FIG. 1. At step 202, the system acquires ECG data for a patient. The ECG data can include one or more channels, each channel representing an ECG signal for a respective lead of the ECG procedure, and data from each channel can be time-registered. The ECG data can span a single cardiac cycle, a portion of a cardiac cycle, or multiple cardiac cycles of the patient. At step 204, a prediction input generator of the system generates one or more prediction inputs from the ECG data and feeds them to an ejection fraction prediction model, such as ejection fraction prediction model 118. In some implementations, the prediction inputs represent a time series of values ​​for an ECG waveform. In other implementations, the prediction inputs represent values ​​of morphological features of the ECG waveform (step 206). At step 208, the system provides the prediction inputs to the ejection fraction prediction model for processing. At step 210, the ejection fraction prediction model processes the prediction input representing the patient's ECG results to generate an estimated (predicted) ejection fraction characteristic. In some implementations (212), the ejection fraction characteristic is an absolute estimate of the patient's ejection fraction that identifies a unique value for the ejection fraction. In other implementations (214), the ejection fraction characteristic is a category or range of ejection fraction values ​​(e.g., very low ejection fraction, low ejection fraction, or normal ejection fraction). For example, the ejection fraction model can be trained to classify the patient's ejection fraction into one of two, three, or more possible ejection fraction categories defined by a specified threshold ejection fraction value. A binary classification model can classify the patient's ejection fraction into two possible categories. The system then stores and / or outputs the estimated ejection fraction characteristic at step 216, e.g., for presentation to the patient or to the patient's healthcare provider.

[0043] In some implementations, the ejection fraction prediction model does not estimate a patient's current ejection fraction characteristics (as described in other implementations), but instead can be trained and configured to predict a patient's risk of developing low or very low ejection fraction at some point in the future (e.g., one month, six months, one year, two years, five years, or more into the future) by detecting subtle features or changes in the patient's ECG that indicate early onset of disease.

[0044] 3 is an illustration of an exemplary ECG tracing 300 for one heart beat of a patient. The ECG tracing shows the constituent segments of the heart beat and various morphological waveform features.

[0045] FIG. 4 is a flow diagram of an exemplary process 400 for selecting and using an appropriate ejection fraction prediction model corresponding to patient characteristics. Process 400 can be implemented by a computing system such as system 110 of FIG. 1. In some examples, a global ejection fraction prediction model can be trained and used for a wide range of patients. However, in other examples, a patient-specific personalized model can be used, or a semi-personalized model trained on ECG data and measured ejection fraction characteristics from a population of patients with the same or similar characteristics as the new patient for whom an ejection fraction estimate is desired, to determine a more accurate or reliable estimate of the patient's ejection fraction. For example, multiple ejection fraction prediction models can be generated, each corresponding to a different set of patient characteristics, such as age, height, weight, body mass index (BMI), sex, family history, any indicator of other co-morbidities such as diabetes, hypertension, hyperlipidemia, high sensitivity C-reactive protein (CRP), smoking, history of coronary artery disease, history of inflammatory disease, or combinations thereof. In step 402, the system identifies a set of characteristics for a patient for whom an estimated ejection fraction characteristic is to be determined. At step 404, the system selects one of the ejection fraction prediction models that corresponds to the identified set of characteristics for the patient. For example, for a male patient over age 50, a model trained on data from subjects with similar characteristics to the patient may be selected over other models trained on data from subjects with other characteristics. At step 406, the system generates an ejection fraction prediction using the selected ejection fraction prediction model that corresponds to the patient's characteristics. For example, the system may generate an estimated ejection fraction characteristic according to the selected model according to process 200 described with respect to FIG. 2.

[0046] While an ECG-based estimation of a patient's ejection fraction can be a useful screening procedure, further evaluation of the patient may be justified based on the results of the ECG-based screening procedure. FIG. 5 is a flow diagram of an exemplary process 500 for screening potentially problematic ejection fraction levels using an ECG-based approach. The process 500 may be automated, for example, performed by the computing system 110, or may be performed by a healthcare provider or other suitable individual. At step 502, an estimated ejection fraction characteristic of the patient is obtained, for example, based on the process 200 described with respect to FIG. 2. At step 504, the system determines whether the estimated ejection fraction characteristic, and optionally additional factors, meet one or more screening criteria to guide a decision as to whether further evaluation of the patient's condition is justified. For example, the estimated ejection fraction characteristic may be an absolute value indicative of the patient's predicted ejection fraction, and the screening criteria may include a threshold ejection fraction (e.g., 35 percent or 50 percent). If the patient's estimated ejection fraction is below the threshold, a follow-up procedure for further evaluation may be performed on the patient. The follow-up procedure may be, for example, an echocardiogram, an MRI, a CT scan, a nuclear medicine examination, or a combination thereof. In some implementations, if the patient's ejection fraction characteristic falls into a range of ECG values ​​that are considered risky or unsafe, further medical evaluation may be performed on the patient.

[0047] In some implementations, if a patient's ejection fraction falls below a threshold (e.g., 35 percent or 50 percent) or if the patient's ejection fraction profile falls into a range of ECG values ​​that are considered dangerous or unsafe, a therapeutic drug may be administered to the patient or a treatment may be administered to the patient. Such drugs or treatments may include beta-blockers, angiotensin receptor antagonists, statins (coronary artery disease), implantable cardioverter-defibrillators, cardiac resynchronization devices, and other well-defined therapies (see 2013 ACCF / AHA Guideline for the Management of Heart Failure. Circulation. 2013;128:e240-e327, and 2016 ACC / AHA / HFSA Focused Update on New Pharmacological Therapy for Heart Failure, Circulation. 2016;134:e282-e293). In some implementations, if the patient's ejection fraction falls below a threshold (e.g., 35 percent or 50 percent) or if the patient's ejection fraction profile falls into a range of ECG values ​​considered risky or unsafe, therapeutic intervention includes additional diagnostic testing, screening for inflammatory and other systemic conditions, and specific treatments or coronary angiograms or other imaging to detect coronary artery disease, which have clearly defined treatments.

[0048] FIG. 6 is a flow diagram of an example process 600 for training an ejection fraction prediction model, such as a neural network-based model. In some implementations, a machine learning technique, such as gradient descent (including batch gradient descent or stochastic gradient descent), can be used to train the ejection fraction model. The process 600 can be performed by a computing system, such as the training subsystem 122 described with respect to FIG. 1. At step 602, the system obtains a set of training data pairs. Each pair includes an ECG prediction input that characterizes a particular patient's ECG and a target ejection fraction characteristic for that patient. The target ejection fraction characteristic can be a "true" or measured ejection fraction characteristic determined by echocardiography or other available procedures. In some cases, the model is trained on data from many different patients. If the model is semi-personalized, the patients represented in the training set may share a common set of characteristics with each other. At step 604, the training system selects a first training data pair for processing. At step 606, the prediction input from the training data pair is provided to the ejection fraction prediction model. The ejection fraction prediction model may be, for example, a neural network machine learning model with one or more layers of perceptrons with associated weights / parameters for the activation function associated with each perceptron. These weights may be randomly initialized when training begins and may be gradually refined over time as additional training iterations are performed. At step 608, the system processes the prediction input according to the current weights / parameters of the model to generate an estimated ejection fraction characteristic. At step 610, the estimated ejection fraction characteristic is compared to the target ejection fraction characteristic to determine an output error. This output error is then back-propagated through the network using gradient descent to update the current weights / parameters of the neural network. The process then returns to step 604, where another training data pair is selected and operations 604-612 are repeated until a training termination condition occurs.

[0049] Although the ejection fraction prediction model can be trained on a large data set, the model itself can be relatively small. After training, the model can be relatively small and can predict ejection fraction characteristics with fewer computational demands than may have been required during training.

[0050] Although the process 600 generally refers to a supervised learning process, in some implementations, the ejection fraction prediction model can also be trained using unsupervised learning techniques.

[0051] FIG. 7 is a flow diagram of an example process 700 for predicting a subject's survival from a cardiac condition, such as low or very low ejection fraction, from ECG data for the subject. In some implementations, the process 700 is performed by a computing system including a survival estimation model, such as model 120 (FIG. 1) of ECG processing system 110. At step 702, the system acquires ECG data for the patient. The ECG data can include one or more channels, each channel representing an ECG signal for a respective lead of the ECG procedure, and data from each channel can be time-registered. The ECG data can span a single cardiac cycle, a portion of a cardiac cycle, or multiple cardiac cycles of the patient. At step 704, a prediction input generator of the system generates one or more prediction inputs from the ECG data and feeds them into the survival estimation model. In some implementations, the prediction inputs represent a time series of values ​​for an ECG waveform. In other implementations, the prediction inputs represent values ​​of morphological features of the ECG waveform (706). At step 708, the prediction input is provided to a survival estimation model, and at step 710, the model processes the prediction input to determine an estimated survival rate for the patient based on the patient's ECG. The estimated survival rate may be, for example, a 1 month, 1 year, 2 years, 3 years, 4 years, 5 years, or 10 years prognosis. The system may then provide an estimate of survival rate as an output (step 712) for presentation to the patient, the patient's healthcare provider, or both.

[0052] FIG. 8 is a diagram of an exemplary neural network model 800 for estimating a subject's ejection fraction characteristics from ECG data. The model 800 processes an ECG prediction input that characterizes one or more channels of a patient's ECG. As can be seen in FIG. 8, the model 800 can include an input layer of perceptrons and / or sigmoid neurons, an output layer, and one or more hidden layers between the input and output layers. For example, data from multiple channels can be processed and pooled until the data is provided to a fully connected layer of the network. The model 800 can output an estimate or classification of the ejection fraction characteristics for the patient.

[0053] In some implementations, the neural network model 800 includes multiple convolutional layers for feature extraction. Each convolutional layer can include a constant or variable length filter that focuses on the current lead of the ECG or multiple leads simultaneously (e.g., k×1 or k×m leads). Following the convolutional layers, the network can include one or more fully connected layers, each with the same or different number of neurons. Zero, one, or more recurrent neural network layers can be added for temporal feature extraction before, after, or in parallel with the convolutional layers. Skip layers and neurons may or may not be included.

[0054] FIG. 9 shows results 900 of a first neural network implementation of an ejection fraction prediction model that classifies patients into groups according to whether their ejection fraction is predicted to be above or below 50 percent.

[0055] FIG. 10 shows the results 1000 of a second neural network implementation of the ejection fraction prediction model that classifies patients into groups according to whether their ejection fraction is predicted to be above or below 35 percent.

[0056] FIG. 11 is a graph 1100 showing the correlation between model output and true ejection fraction values ​​in the study.

[0057] Although several examples of detecting a patient's impaired (low) ejection fraction using data from a standard 12-lead ECG are described herein, in other implementations, fewer leads, including single-lead ECG, may be used to effectively detect instances of low ejection fraction or identify individuals likely to be at high risk of developing a low ejection fraction within a future period (e.g., 5 years). The ejection fraction prediction model may be trained and implemented based on data from any number of ECG leads, including a single lead, any multi-lead configuration, and a standard 12-lead ECG. For example, a single-lead ECG based on lead 1 has been shown to be particularly effective in providing robust data from which the ejection fraction prediction model can generate an ejection fraction prediction. Advantageously, ejection fraction prediction based on a single-lead ECG or other non-traditional electrodes (e.g., fewer than a standard 12-lead ECG) may expand access to screening devices implementing the ejection fraction prediction model described herein. For example, a pair of electrodes can be integrated into or communicatively coupled to a smartphone or other portable electronic device, allowing an individual to conveniently capture their own ECG and quickly obtain a predicted result, for example, by touching the electrodes with their fingertips. In other implementations, the electrodes can be provided in a wearable patch for continuous monitoring over a period of time.

[0058] In certain examples, yet another advantage that can be achieved by at least some of the implementations described herein is the ability to detect low ejection fraction based on a short, single time profile of ECG data (e.g., from a single lead or multiple leads) for a patient. For example, the system only needs to process a relatively short sample of ECG data to generate a high-confidence prediction of the patient's ejection fraction status, such as whether the patient has a low EF. In some implementations, the ejection fraction prediction model is configured to predict whether the patient has a low EF (e.g., an EF less than 35, 40, or 50 percent) based on an ECG sample of 60 seconds or less. In some implementations, the ejection fraction prediction model is configured to predict whether the patient has a low EF (e.g., an EF less than 35, 40, or 50 percent) based on an ECG sample of 30 seconds or less. In some implementations, the ejection fraction prediction model is configured to predict whether the patient has a low EF (e.g., an EF less than 35, 40, or 50 percent) based on an ECG sample of 20 seconds or less. In some implementations, the ejection fraction prediction model is configured to predict whether a patient has a low EF (e.g., an EF less than 35, 40, or 50 percent) based on an ECG sample of 15 seconds or less. In some implementations, the ejection fraction prediction model is configured to predict whether a patient has a low EF (e.g., an EF less than 35, 40, or 50 percent) based on an ECG sample of 10 seconds or less. In some implementations, the ejection fraction prediction model is configured to predict whether a patient has a low EF (e.g., an EF less than 35, 40, or 50 percent) based on an ECG sample of 5 seconds or less. In some implementations, the ejection fraction prediction model is configured to predict whether a patient has a low EF (e.g., an EF less than 35, 40, or 50 percent) based on an ECG sample of 2 seconds or less. Thus, the acquisition time for the ECG can be relatively short, which is convenient for the patient.Additionally, the use of machine learning models such as convolutional neural networks trained on large data sets can enable fast processing, allowing a computing system running the model to return ejection fraction prediction results based on ECG input in a relatively short amount of time.

[0059] Certain implementations have been described with respect to models that estimate or predict an individual's ejection fraction. However, it should be appreciated that these techniques can be extended more generally to models that facilitate detection of a variety of current or impending structural heart diseases, including abnormal ejection fraction, abnormal left ventricular mass (e.g., elevated left ventricular mass, low left ventricular mass), valvular heart disease, ischemic heart disease, atrial appendage abnormalities, the presence of shunts, or patent foramen ovale (PFO), enlargement of a heart chamber (e.g., left atrium, right atrium, left ventricle, right ventricle), or combinations thereof and / or other structural heart diseases.

[0060] FIG. 12 is a block diagram of computing devices 1200, 1250 that can be used as a client or one or more servers to implement the systems and methods described herein. Computing device 1200 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. Computing device 1250 is intended to represent various forms of mobile devices, such as personal digital assistants, mobile phones, smartphones, and other similar computing devices. In addition, computing device 1200 or 1250 can include a Universal Serial Bus (USB) flash drive. A USB flash drive can store an operating system and other applications. A USB flash drive can include input / output components, such as a wireless transmitter or a USB connector that can be inserted into a USB port of another computing device. The components, their connections and relationships, and functions shown are meant to be exemplary only and are not meant to limit the implementations described and / or claimed herein.

[0061] The computing device 1200 includes a processor 1202, a memory 1204, a storage device 1206, a high-speed interface 1208 that connects to the memory 1204 and a high-speed expansion port 1210, and a low-speed interface 1212 that connects to a low-speed bus 1214 and the storage device 1206. Each of the components 1202, 1204, 1206, 1208, 1210, and 1212 are interconnected using various buses and may be mounted on a common motherboard or may be mounted in other suitable manners. The processor 1202 may process instructions for execution within the computing device 1200, including instructions stored in the memory 1204 or on the storage device 1206, to display graphical information as a GUI on an external input / output device, such as a display 1216 coupled to the high-speed interface 1208. In other implementations, multiple processors and / or multiple buses may be used, as appropriate, along with multiple memories and multiple types of memories. Also, multiple computing devices 1200 may be connected, each providing some portion of the required operations (eg, as a bank of servers, a collection of blade servers, or a multi-processor system).

[0062] The memory 1204 stores information within the computing device 1200. In one implementation, the memory 1204 is one or more volatile memory units. In another implementation, the memory 1204 is one or more non-volatile memory units. The memory 1204 can also be another form of computer-readable medium, such as a magnetic or optical disk.

[0063] The storage device 1206 can provide mass storage for the computing device 1200. In one implementation, the storage device 1206 can be or include a computer-readable medium such as a floppy disk device, a hard disk device, an optical disk device or tape device, a flash memory, or other similar solid-state storage device, or an array of devices including devices in a storage area network or other configuration. The computer program product can be tangibly embodied in an information carrier. The computer program product can also include instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer or machine-readable medium, such as the memory 1204, the storage device 1206, or the memory on the processor 1202.

[0064] The high-speed controller 1208 manages bandwidth-intensive operations for the computing device 1200, and the low-speed controller 1212 manages less bandwidth-intensive operations. Such an allocation of functions is merely exemplary. In one implementation, the high-speed controller 1208 is coupled to the memory 1204, the display 1216 (e.g., via a graphics processor or accelerator), and a high-speed expansion port 1210, which can accept various expansion cards (not shown). In this implementation, the low-speed controller 1212 is coupled to the storage device 1206 and the low-speed expansion port 1214. The low-speed expansion port can include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet) and can be coupled to one or more input / output devices, such as a keyboard, pointing device, scanner, or a networking device, such as a switch or router, for example, via a network adapter.

[0065] The computing device 1200 can be implemented in a number of different forms, as shown. For example, the computing device 1200 can be implemented as a standard server 1220, or multiple times within a cluster of such servers. The computing device 1200 can also be implemented as part of a rack server system 1224. In addition, the computing device 1200 can be implemented within a personal computer, such as a laptop computer 1222. Alternatively, components from the computing device 1200 can be combined with other components in a mobile device (not shown), such as device 1250. Each such device can include one or more of the computing devices 1200, 1250, and the entire system can consist of multiple computing devices 1200, 1250 in communication with each other.

[0066] Computing device 1250 includes, among other components, a processor 1252, memory 1264, input / output devices such as a display 1254, a communication interface 1266, and a transceiver 1268. Device 1250 may also include a storage device, such as a microdrive or other device, to provide additional storage. Components 1250, 1252, 1264, 1254, 1266, and 1268 are each interconnected using various buses, and some of these components may be mounted on a common motherboard, or may be mounted in other suitable manners.

[0067] The processor 1252 can execute instructions in the computing device 1250, including instructions stored in the memory 1264. The processor can be implemented as a chipset, consisting of chips including separate analog and digital processors. Additionally, the processor can be implemented using any of a number of architectures. For example, the processor 1252 can be a CISC (Complex Instruction Set Computer) processor, a RISC (Reduced Instruction Set Computer) processor, or a MISC (Minimum Instruction Set Computer) processor. The processor can provide coordination of other components of the device 1250, such as, for example, control of a user interface, applications executed by the device 1250, and wireless communication by the device 1250.

[0068] The processor 1252 can communicate with a user via a control interface 1258 and a display interface 1256 coupled to a display 1254. The display 1254 can be, for example, a TFT (thin film transistor liquid crystal display) display or an OLED (organic light emitting diode) display, or other suitable display technology. The display interface 1256 can include suitable circuitry for driving the display 1254 to present graphical and other information to the user. The control interface 1258 can receive commands from the user and convert them for transmission to the processor 1252. Additionally, an external interface 1262 can be provided in communication with the processor 1252 to enable near area communication between the device 1250 and other devices. The external interface 1262 can provide, for example, wired communication in some implementations, or wireless communication in other implementations, and multiple interfaces can also be used.

[0069] The memory 1264 stores information within the computing device 1250. The memory 1264 can be implemented as one or more of one or more computer readable media, one or more volatile memory units, or one or more non-volatile memory units. Also, an expansion memory 1274 can be provided and connected to the device 1250 via an expansion interface 1272, which can include, for example, a SIMM (Single In-Line Memory Module) card interface. Such expansion memory 1274 can provide extra storage space for the device 1250 or can store applications or other information for the device 1250. In particular, the expansion memory 1274 can include instructions that perform or supplement the processes described above, and can also include secure information. Thus, for example, the expansion memory 1274 can be provided as a security module for the device 1250 and can be programmed with instructions that enable secure use of the device 1250. In addition, secure applications can be provided with additional information via a SIMM card, such as placing identifying information on the SIMM card so that it cannot be hacked.

[0070] The memory may include, for example, flash memory and / or NVRAM memory, as discussed below. In one implementation, a computer program product is tangibly embodied in an information carrier. The computer program product includes instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer or machine-readable medium, such as memory 1264, expansion memory 1274, or memory on processor 1252, which may be received via transceiver 1268 or external interface 1262.

[0071] The device 1250 can communicate wirelessly via a communication interface 1266, which can include digital signal processing circuitry, if necessary. The communication interface 1266 can provide communication in various modes or protocols, such as GSM voice calls, SMS, EMS, or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS, among others. Such communication can occur, for example, via a radio frequency transceiver 1268. In addition, short-range communication can occur using Bluetooth, WiFi, or other such transceivers (not shown), etc. In addition, a global positioning system (GPS) receiver module 1270 can provide additional navigation and location related wireless data to the device 1250, which can be used as appropriate by applications executing on the device 1250.

[0072] Device 1250 can also communicate audibly using voice codec 1260, which can receive spoken information from a user and convert it into usable digital information. Voice codec 1260 can also generate audible sounds for the user, such as through a speaker in a handset of device 1250. Such sounds can include sounds from a voice call, can include recorded sounds (e.g., voice messages, music files, etc.), and can also include sounds generated by applications running on device 1250.

[0073] Computing device 1250 can be implemented in a number of different forms, as shown in the figure. For example, computing device 1250 can be implemented as a mobile phone 1280. Computing device 1250 can also be implemented as part of a smartphone 1282, personal digital assistant, or other similar mobile device.

[0074] Various implementations of the systems and techniques described herein can be realized in digital electronic circuitry, integrated circuits, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementations in one or more computer programs executable and / or interpretable on a programmable system that includes at least one programmable processor, which can be special purpose or general purpose, coupled to receive data and instructions from and transmit data and instructions to a storage system, at least one input device, and at least one output device.

[0075] These computer programs (also known as programs, software, software applications, or code) include machine instructions for a programmable processor and may be implemented in high level procedural and / or object-oriented programming languages ​​and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device used to provide machine instructions and / or data to a programmable processor (e.g., magnetic disks, optical disks, memory, programmable logic devices (PLDs)), including a machine-readable medium capable of receiving the machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0076] To provide for user interaction, the systems and techniques described herein can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user, and a keyboard and pointing device (e.g., a mouse or trackball) by which the user can provide input to the computer. Other types of devices can be used to provide for user interaction as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback), and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0077] The systems and techniques described herein can be implemented within a computing system that includes back-end components (e.g., as a data server), or includes middleware components (e.g., an application server), or includes front-end components (e.g., a client computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described herein), or includes any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communications network). Examples of communications networks include a local area network ("LAN"), a wide area network ("WAN"), a peer-to-peer network (with ad-hoc or static members), a grid computing infrastructure, and the Internet.

[0078] A computing system may include clients and servers. Clients and servers are generally remote from each other and typically interact through a communication network. The relationship of clients and servers arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0079] [Example implementation] <Summary> This example belongs to an artificial intelligence ("AI")-based study developed by the techniques described herein to detect with high fidelity patients suffering from asymptomatic low left ventricular dysfunction (ALVD) using a non-invasive 10-second digital ECG.

[0080] background ALVD occurs in 2-9% of the population, is associated with reduced quality and longevity of life, and is generally treatable when detected. The area under the curve (AUC) for the BNP screening blood test is 0.79-0.89. This study hypothesized that the use of artificial intelligence (AI) would make it possible to identify left ventricular systolic dysfunction on the ECG, a ubiquitous and inexpensive test.

[0081] method The study involved training a convolutional neural network to identify patients with ventricular dysfunction defined as an ejection fraction (EF) ≦35% using digitally stored 12-lead ECG and echocardiogram pairs from 44,959 patients in the Mayo Clinic data vault. The network was then tested on 52,870 patients reserved for external validation.

[0082] result Of 52,870 patients studied, 4,131 (7.8%) had an EF ≤ 35%. The AUC of the ROC was 0.93. Sensitivity, specificity, and accuracy were 86.3%, 85.7%, and 85.7%, respectively. Of 1335 patients with abnormal AI screens but normal EF (false positives), 147 (11%) had at least one abnormal EF in the future (5-year incidence 9.5%). This four-fold increased risk of developing a low LVEF in the future suggests that the network may be detecting early, subclinical, metabolic or structural abnormalities that are evident on the ECG.

[0083] conclusion The application of artificial intelligence to the ECG, a ubiquitous and typically low-cost test, could enable it to serve as a powerful tool to screen for left ventricular dysfunction and further identify individuals at high risk for its future development.

[0084] <Method> Data sources and study population After Institutional Review Board approval, the study acquired data from the Mayo Clinic Digital Data Vault. 163,892 adult patients (age 18 and older) were identified with at least one digital standard 10-second, 12-lead ECG acquired in the supine position between January 1994 and February 2017, and at least one transthoracic echocardiogram (TTE) acquired within 14 days of the index ECG (Figure 13). For patients with multiple ECG and TTE data sets meeting these criteria, the earliest pair was used for network creation, validation, or testing, and later TTE data was used for follow-up analysis. A preliminary proof-of-concept assessment was performed to acquire an internal study using 2200 ECG-TTE data pairs that were excluded from the main analysis, leaving a cohort of 97,829 patients, and the first ECG-TEE paired data set of these patients was used for the primary analysis.

[0085] ECGs were acquired using a General Electric-Marquette ECG machine (Marquette, WI) at a sampling rate of 500 Hz and stored for later retrieval using the MUSE data management system. Comprehensive 2- or 3-dimensional Doppler echocardiography was available to all patients. Quantitative data were recorded at the time of acquisition in a custom database (Echo Image Management System, EIMS) developed at Mayo Clinic. Left ventricular ejection fraction (EF) was routinely measured or estimated using standardized methods, and in most reports more than one value can be recorded. For the purposes of this study, the ejection fraction value used in the model was the first value available from a standard hierarchical sequence and EF was determined using three-dimensional echocardiography (Yamani H, Cai Q, Ahmad M. Three-dimensional echocardiography in evaluation of left ventricular indices. Echocardiography 2012;29:66-75), a two-plane method using the Simpson method, a two-dimensional method (Quinones MA, Waggoner AD, Reduto LA, et al. A new, simplified and accurate method for determining ejection fraction with two-dimensional echocardiography. Circulation 1981;64:744-53), M-mode, or, in the absence of any of the above, a visually estimated EF report. If the estimate was a range, the midpoint was used as a single EF value. Left ventricular EF was classified as low (≤35%), mildly reduced (35-49%), or normal (≥50%).

[0086] 1st and 2nd results The primary outcome was the ability of the AI ​​network to identify patients with an LVEF of 35% or less using only the ECG signal, a value chosen due to its clear clinical and therapeutic importance (Russo AM, Stainback RF, Bailey SR, et al. ACCF / HRS / AHA / ASE / HFSA / SCAI / SCCT / SCMR 2013 appropriate use criteria for implantable cardioverter-defibrillators and cardiac resynchronization therapy: a report of the American College of Cardiology Foundation appropriate use criteria task force, Heart Rhythm Society, American Heart Association, American Society of Echocardiography, Heart Failure Society of America, Society for Cardiovascular Angiography and Interventions, Society of Cardiovascular Computed Tomography, and Society for Cardiovascular Magnetic Resonance. J Am Coll Cardiol 2013;61:1318-68). The secondary outcome was the ability of the AI ​​network to identify individuals who had normal EF at screening but were at increased risk of later low EF during follow-up.

[0087] Overview of AI model development This work involved developing a Convolutional Neural Network (CNN) using the KERAS Framework with a TENSORFLOW® (GOOGLE, Mountain View, CA) backend and PYTHON (van Rossum G. Python tutorial, Technical Report CS-R9526. Amsterdam 1995 May). CNNs applied to images (or videos) operate in such a way that they can extract very small patterns in a data set using convolutions. Each 12-lead ECG was considered as an "image" of 12x5000 (i.e., 12 leads sampled at 500Hz with a duration of 10 seconds) (van Rossum G. Python tutorial, Technical Report CS-R9526. Amsterdam 1995 May). The network consisted of N single-lead convolutional layers, each followed by a nonlinear "Relu" activation function, a batch normalization layer (Ioffe S, Szegedy C, Batch normalization: accelerating deep network training by reducing internal covariate shift International Conference on Machine Learning 2015), and a max-pooling layer (Nagi J, Ducatelle F, Di Caro G, et al. Max-pooling convolutional neural networks for vision-based hand gesture recognition. 2011 IEEE International Conference 2011:342-7). Features extracted from each raw digital signal ECG lead were fused in another convolutional layer that had access to all leads simultaneously.Following the final convolutional layer, this data was fed into a fully connected network with two hidden layers, including a dropout layer to avoid overfitting, and an output layer that was activated using a "Softmax" function (Cristianini N, Shawe-Taylor J, New York, NY: Cambridge University Press; 1999).

[0088] AI model training Due to the large size of our dataset, we used about 50% of the dataset to train the network. This gave us a very large dataset to test the network and better evaluate its robustness (Figure 13). After the initial split into development and testing (holdout) datasets, the development dataset was further split into training data (80% of the development set) and internal validation data (20%).

[0089] For training, we fed ECGs into the network and updated the network weights using the Adam optimizer (Kingma DP, Ba J. Adam: A Method for Stochastic Optimization 2014) with binary cross entropy as the loss function. After each epoch, we tested the network using an internal validation dataset and stopped training after it was optimized. We also tuned network hyperparameters during this process and selected the network with the lowest loss value. All ECGs were low-pass filtered (100hz) to remove high frequency noise and quantization errors.

[0090] Primary outcome – AI enhanced ECG to identify low EF After selecting the best network using the validation data, a receiver operating curve (ROC) was created using the same validation set, and its area under the curve (AUC) was measured as a primary assessment of network strength. The ROC of the validation data set was used to select two thresholds for the probability of having low LVEF. The first threshold was selected by giving equal weight to sensitivity and specificity, and the second threshold was selected to give 90% sensitivity on the validation data set. The CNN model was then used on the test data to test its ability to predict low LVEF. The two thresholds were used to calculate sensitivity, specificity, and accuracy on the test data that was not used for model training or threshold selection. The impact of age and gender on the network prediction function was also evaluated by creating a separate network with those variables as inputs and training the network to determine whether age and gender can be determined from ECG alone.

[0091] Secondary outcome – AI enhanced ECG to predict future low EF We hypothesized that early in the course of any disease that affects EF, the ECG signal would show subtle abnormal patterns due to metabolic and structural derangements that did not affect a sufficient amount of myocardium to cause a decline in EF. We further hypothesized that the CNN would classify as abnormal some of these cases that initially appear as false-positive tests but may become true-positive tests over time. To test this hypothesis, we designed a substudy to identify patients for whom: 1) the network predicted that the patient had low EF; 2) the patient had an echocardiogram performed within 14 days that demonstrated a normal EF (≥50%) indicating a false-positive finding by the algorithm; and 3) at least one additional echocardiogram (not used for training or testing) was available at a later date. True-negative cases (both the algorithm and clinical judgement were consistent with not having low EF) were used to create a control group. For the control group, subjects were selected to have an LVEF ≥50%. Kaplan-Meier analysis was used to show the incidence of low EF in terms of true-negatives versus false-positives over time. Cox proportional hazards regression was then used to estimate the hazard for low EF after adjustment for age and sex. In addition, a sensitivity analysis was performed by categorizing the predicted probability for low EF to determine whether there was a monotonic pattern in the predicted probability for future development of low EF.

[0092] statistical considerations For measures of diagnostic performance (AUC ROC, sensitivity), the sample size was sufficiently large that normal confidence intervals (CIs) are expected to have widths less than 0.5%. These CIs are therefore not reported with the estimates due to their precision. Continuous data are presented as mean + / - SD. Two-sided p-values ​​are presented for Cox models. Survival analyses were performed using SAS version 9.4. CNNs were trained using KERAS (version 2.0.3) and TENSORFLOW (version 1.0.1).

[0093] Research support The study was conceived, funded, and performed entirely by Mayo Clinic. It received no industrial support of any kind.

[0094] <Result> Study population A total of 625,326 patients who underwent ECG-TTE pairs were screened to identify the study cohort selected for analysis (Figure 13). The first ECG-TTE data pairs from patients with ECG and echocardiography performed within 2 weeks of each other constituted the analysis dataset of 97,829 patients, with 35,970 in the training set, 8,989 in the validation set, and 52,870 in the holdout test set. No patient fell into more than one group (Figure 13). The mean age of the overall patient population was 61.8±16.5 years, and 7.8% of the population had an EF≦35%. Table 1 shows patient characteristics for the training, validation, and test sets. In the study dataset, 4,131 patients (7.8%) had an EF ≤ 35%, 6,740 patients (12.7%) had an LVEF > 35% and < 50%, and 41,999 patients (79.5%) had an LVEF ≥ 50%. More than 89% of TTEs were performed within 1 day of the index ECG.

[0095] Primary result In Figure 14, the AUC for the holdout test data set was 0.93, identical to the AUC for the internal validation data set (0.93). When choosing a threshold without preference for sensitivity, the overall accuracy was 85.7%, the specificity was 85.7%, the sensitivity was 86.3%, and the negative predictive value was 98.7%. Using a threshold to give a sensitivity of 90% on the validation set and applying the algorithm to the test data set, the sensitivity was 89.1%, the specificity was 83%, the overall accuracy was 83.5%, and the negative predictive value was 98.9%. When patients without known co-morbidities (see Figure 16) were analyzed separately by the network, the AUC increased to 0.98, with a sensitivity of 95.6%, a specificity of 92.4%, a negative predictive value of 99.8%, and an accuracy of 92.5%.

[0096] Network performance was not improved by the additional input of age and gender. To understand if the network was "judging" age and gender based on ECG alone, the network was retrained to predict age and gender from ECG. Age was predicted to be 8.7±6 years, and gender was correctly judged from ECG 87% percent with an AUC of 0.94. This shows that the network was able to reliably judge age and gender from ECG alone, explaining why its performance was not affected by age or gender, since the network should have been able to account for these variables.

[0097] Distribution of EF values ​​according to AI algorithm classification When choosing a threshold without preference for sensitivity (i.e., a threshold that gives equal sensitivity and specificity using the validation data), 10,544 (19.9%) of the patients in the test set were identified by the network as having a low EF. Of these 10,544 patients, 33.8% had an EF below 35%, 29.5% had an EF between 36 and 50%, and 36.6% had a normal EF. Within the group identified by the network as normal, 1.3% had an EF below 35%, 8.6% had an LVEF between 36 and 50%, and the remainder (90.1%) had a normal LVEF (Figure 17).

[0098] Secondary outcome – AI enhanced ECG to predict future low EF Among patients who were identified by the network as having normal EF and who also had a confirmatory concurrent normal EF by echocardiography ("true negatives"), 11,515 underwent follow-up echocardiography. Of these true negative patients, 302 developed low EF at a median (IQR) follow-up of 3.8 (1.4–7.5) years (Figure 4: 10-year incidence of 4.4%). In contrast, 1,335 patients were indicated by the network as having low EF, but concurrent echocardiography demonstrated normal EF ("false positives"). Of these 1,335 patients who had an initial "false positive" result, 147 (Figure 4: 10-year incidence of 20.8%) developed left ventricular dysfunction during a median follow-up of 3.4 (IQR 1.2–6.8) years. This represented a four-fold increased risk of future low EF when the AI ​​algorithm defined the ECG as abnormal (age- and sex-adjusted, HR=4.1 [3.3-5.0], p<0.001), suggesting that the network identified ECG abnormalities before overt ventricular dysfunction emerged.

[0099] <Discussion> Left ventricular systolic dysfunction is associated with decreased quality of life, increased morbidity, and increased mortality (McDonagh TA, McDonald K, Maisel AS. Screening for asymptomatic left ventricular dysfunction using B-type natriuretic peptide. Congest Heart Fail 2008;14:5-8). Major cardiovascular societies endorse evidence-based treatments that improve symptoms and survival once detected (Al-Khatib SM, Stevenson WG, Ackerman MJ, et al. 2017 AHA / ACC / HRS Guideline for Management of Patients With Ventricular Arrhythmias and the Prevention of Sudden Cardiac Death: Executive Summary: A Report of the American College of Cardiology / American Heart Association Task Force on Clinical Practice Guidelines and the Heart Rhythm Society. Circulation 2017; Yancy CW, Jessup M, Bozkurt B, et al. 2013 ACCF / AHA guideline for the management of heart failure: a report of the American College of Cardiology Foundation / American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol 2013;62:e147-239).However, validated screens for ventricular dysfunction are lacking. (McDonagh TA, McDonald K, Maisel AS. Screening for asymptomatic left ventricular dysfunction using B-type natriuretic peptide. Congest Heart Fail 2008;14:5-8; Redfield MM, Rodeheffer RJ, Jacobsen SJ, Mahoney DW, Bailey KR, Burnett JC, Jr. Plasma brain natriuretic peptide to detect preclinical ventricular systolic or diastolic dysfunction: a community-based study. Circulation 2004;109:3176-81). The study found that applying artificial intelligence using convolutional neural networks to standard 12-lead ECG, a common clinical test that is cheap and widely available, allowed the detection of left ventricular dysfunction with an AUC of 0.93. The performance of this test compares favorably with other common screening tests such as prostate-specific antigen for prostate cancer (0.92), mammography for breast cancer (0.85), and cervical cytology for cervical cancer (0.70). In contrast to BNP, the accuracy was not affected by age or sex. Importantly, in addition to effectively identifying individuals suffering from ventricular systolic dysfunction, the network also predicted those who initially had normal LV function but would later develop low EF. Patients with normal EF but abnormal network screening ("false positives") had a four-fold increased risk of developing ventricular dysfunction in the next five years (10% risk at five years). This suggests that the network detected early, subclinical, metabolic, or structural abnormalities that would manifest themselves on the ECG.

[0100] Congestive heart failure afflicts over 5 million people in the United States alone and consumes over $30 billion in medical costs annually (Heidenreich PA, Albert NM, Allen LA, et al., Forecasting the impact of heart failure in the United States: a policy statement from the American Heart Association. Circ Heart Fail 2013;6:606-19; Mozaffarian D, Benjamin EJ, Go AS, et al., Heart disease and stroke statistics--2015 update: a report from the American Heart Association. Circulation 2015;131:e29-322). Early detection and prevention are paramount in health care. Asymptomatic left ventricular dysfunction affects over 7 million Americans and many more worldwide. We found that 6% of patients in our population had low EF, which is consistent with previous studies (McDonagh TA, McDonald K, Maisel AS. Screening for asymptomatic left ventricular dysfunction using B-type natriuretic peptide. Congest Heart Fail 2008;14:5-8). BNP and NT-BNP have been proposed for the detection of left ventricular systolic dysfunction.Bhalla and coworkers evaluated BNP to screen for systolic and diastolic dysfunction and found that the AUC was 0.60 for BnP and 0.70 for NTproBNP, with the results improving to 0.70 and 0.73 with the addition of impedance cardiography (Bhalla V, Isakson S, Bhalla MA, et al. Diagnostic ability of B-type natriuretic peptide and impedance cardiography: testing to identify left ventricular dysfunction in hypertensive patients. Am J Hypertens 2005;18:73S-81S). A study from the Mayo Clinic from Olmsted County evaluating individuals aged 45 years and older found that the AUC was higher in individuals with more severe systolic dysfunction (0.82-0.92) than in those with other systolic dysfunctions (0.51-0.74). Furthermore, the optimal discrimination level for BNP varied by age and gender. In contrast, this study found good AI network performance that was invariant to age and gender. This was verified by creating a network that was able to determine the age and gender of patients within the range of 8±6.3 years based on ECG alone with 87% accuracy.Previous literature has documented ECG changes with age and gender (Daly C, Clemens F, Lopez Sendon JL, et al. Gender differences in the management and clinical outcome of stable angina. Circulation 2006;113:490-8; Khane RS, Surdi AD, Bhatkar RS. Changes in ECG pattern with advancing age. J Basic Clin Physiol Pharmacol 2011;22:97-101; Kuo TB, Lin T, Yang CC, Li CL, Chen CF, Chou P. Effect of aging on gender differences in neural control of heart rate. Am J Physiol 1999;277:H2233-9; Stramba-Badiale M, Locati EH, Martinelli A, Courville J, Schwartz PJ. Gender and the relationship between ventricular repolarization and cardiac cycle length during 24-h Holter recordings. Eur Heart J 1997;18:1000-6). During the training process, the network integrated and accounted for age and gender characteristics in a way that did not affect its discriminatory ability to identify ventricular dysfunction. This ability appears to be inherent to the neural network selection.

[0101] The specific ECG features used by the unsupervised convolutional neural network to classify individuals as having or not having low EF are not known due to the nature of the neural network, but by training the network on a large cohort of approximately 45,000 ECG and EF data pairs, the network was exposed to a sufficient number of ECG variants to reliably classify individuals with low ejection fraction, as demonstrated by an AUC of 0.93 when tested on a population of 52,870 individuals. In contrast to traditional applications of neural networks in medicine to mimic human skills, such as identifying mammographic lesions (Salazar-Licea LA, Pedraza-Ortega JC, Pastrana-Palma A, Aceves-Fernandez MA. Location of mammograms ROI's and reduction of false-positive. Comput Methods Programs Biomed 2017;143:97-111), this work expands the use of AI and extends capabilities beyond human skills.

[0102] A key property of our network is that it uses 12-lead ECG as its input, an inexpensive, standard, and ubiquitous test. Many rural areas of the United States (Gruca TS, Pyo TH, Nelson GC, Providing Cardiology Care in Rural Areas Through Visiting Consultant Clinics. J Am Heart Assoc 2016;5) and developing countries have limited access to cardiac care and imaging.The availability of portable, inexpensive tests for ventricular systolic dysfunction allows optimal use of limited imaging resources and allows individuals to benefit from proven therapies such as beta-adrenergic blockers, angiotensin receptor antagonists, and, when available, implantable devices (defibrillators and cardiac resynchronization systems) (Yancy CW, Jessup M, Bozkurt B, et al., 2013 ACCF / AHA guideline for the management of heart failure: a report of the American College of Cardiology Foundation / American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol 2013;62:e147-239; Yancy CW, Jessup M, Bozkurt B, et al., 2017 ACC / AHA / HFSA Focused Update of the 2013 ACCF / AHA Guideline for the Management of Heart Failure: A Report of the American College of Cardiology Foundation / American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol 2013;62:e147-239; Cardiology / American Heart Association Task Force on Clinical Practice Guidelines and the Heart Failure Society of America. J Am Coll Cardiol 2017;70:776-803).With the advent of smartphone-compatible electrodes (Yasin OZ, Attia Z, Dillon JJ, et al. Noninvasive blood potassium measurement using signal-processed, single-lead ecg acquired from a handheld smartphone. J Electrocardiol 2017;50:620-5), mobile applications could enable use in resource-limited areas. Additionally, the software-based nature of the test "samples" for the network implemented in this study allows for continuous feedback and refinement, allowing for rapid distribution of system improvements.

[0103] In this study, we chose an EF of 35% or less as the detection threshold due to the well-established results and therapeutic impact of this value (Yancy CW, Jessup M, Bozkurt B, et al. 2017 ACC / AHA / HFSA Focused Update of the 2013 ACCF / AHA Guideline for the Management of Heart Failure: A Report of the American College of Cardiology / American Heart Association Task Force on Clinical Practice Guidelines and the Heart Failure Society of America. J Am Coll Cardiol 2017;70:776-803). However, discrimination of an EF below 50% is also clinically significant. Thus, 45% of the “false positive” values ​​are indeed clinically meaningful (Figure 15). In addition, the network can be easily configured to use a different detection threshold instead, for example discrimination of an EF<40%. ECG-TTE pairs were not acquired simultaneously. However, the study included a dataset of over 100,000 ECG-TTE pairs, with over 89% of the TTEs performed within 24 hours of the ECG, so the chance of error related to time delays is small. The convolutional neural network in this study was specifically trained to detect low EF, i.e., diastolic dysfunction, and not heart failure. However, evidence-based therapies that reduce morbidity and mortality have been established for the treatment of low EF, which may be asymptomatic and therefore whose detection is a high priority.

[0104] Although several implementations have been described in detail above, other modifications are possible. Additionally, the logic flow depicted in the figures does not require the particular order or sequence shown to achieve the desired results. Other steps can be provided or steps can be omitted from the described flow, and other components can be added to or removed from the described system. Thus, other implementations are within the scope of the following claims.

Claims

1. receiving, by a system comprising one or more computers, electrocardiogram (ECG) data describing a subject's ECG over a period of time; processing the ECG data to generate predicted inputs; providing the prediction input to a prediction model generated by processing a plurality of training data to predict a prognosis of structural heart disease for the subject; outputting a prediction of structural heart disease for the subject; and Including, the prediction input includes a time series of values ​​indicative of the amplitude of the electrocardiogram for one or more leads at each time point; the predictive model includes a convolutional neural network; the convolutional neural network includes a plurality of single-lead convolutional layers and a convolutional layer having access to all leads of the electrocardiogram; Computer-implemented method.

2. Processing a plurality of training data to predict structural heart disease for the subject includes: obtaining one or more ECG training data pairs, each ECG training data pair in the one or more ECG training data pairs including a training prediction input; selecting an ECG training data pair; providing the training prediction inputs to the predictive model from selected ECG training data pairs; processing the training prediction input with the predictive model to generate a prediction of structural heart disease; comparing said prediction of structural heart disease to a target structural heart disease value to determine an output error; updating at least one parameter of the predictive model using the output error; The computer-implemented method of claim 1 , comprising:

3. Processing the ECG data to generate a predicted input includes: normalizing and vectorizing the ECG data from one or more channels; generating the predicted input based on the normalized and vectorized ECG data; The computer-implemented method of claim 1 , comprising:

4. The computer-implemented method of claim 1 , wherein the prediction input comprises a series of values ​​indicative of an amplitude of the ECG data for at least one lead over one or more cardiac cycles.

5. The computer-implemented method of claim 1 , wherein the prediction input comprises one or more values ​​of one or more morphological features of the ECG data.

6. The computer-implemented method of claim 1 , wherein the ECG data is from one or more channels, each of the one or more channels associated with one or more ECG leads.

7. The computer-implemented method of claim 1 , further comprising estimating an ejection fraction value from the predictive input.

8. The computer-implemented method of claim 7 , wherein the predictive model processes the estimated ejection fraction values ​​to generate a prediction of structural heart disease.

9. The computer-implemented method of claim 1 , wherein the prediction of structural heart disease comprises a determination of current or impending structural heart disease.

10. The computer-implemented method of claim 1 , wherein the predictive model is a regression model.

11. The computer-implemented method of claim 1 , wherein the predictive model is a machine learning model.

12. 10. The computer-implemented method of claim 1, further comprising a notification service, the notification service putting the prediction of structural heart disease into a display for presentation to the subject.

13. One or more non-transitory computer-readable media encoded with instructions that, when executed by one or more processors of a system, cause the one or more processors to perform operations, the operations including: receiving, by a system comprising one or more computers, electrocardiogram (ECG) data describing a subject's ECG over a period of time; processing the ECG data to generate predicted inputs; providing the prediction input to a prediction model generated by processing a plurality of training data to predict a prognosis of structural heart disease for the subject; outputting a prediction of structural heart disease for the subject; and Including, the prediction input includes a time series of values ​​indicative of the amplitude of the electrocardiogram for one or more leads at each time point; the predictive model includes a convolutional neural network; the convolutional neural network includes a plurality of single-lead convolutional layers and a convolutional layer having access to all leads of the electrocardiogram; One or more non-transitory computer-readable media.

Images