阅读说明：本技术 用于基于人工智能的心律失常检测的数据准备 (Data preparation for artificial intelligence based arrhythmia detection ) 是由 D·R·马斯格鲁夫 N·查克拉瓦希 S·丹妮 T·D·哈达德 A·拉德克 R·卡特拉 L 于 2020-04-20 设计创作，主要内容包括：公开了用于准备数据以用于基于人工智能(AI)的心律失常检测的技术。根据本公开的技术,一种计算系统可以获得表示同一患者的心脏节律的波形的心脏电描记图(EGM)条带。另外地,所述计算系统可以对所述心脏EGM条带进行预处理。所述计算系统然后可以将深度学习模型应用于经过预处理的心脏EGM条带以生成指示所述心脏EGM条带是否表示一种或多种心律失常的一次或多次发生的心律失常数据。(Techniques for preparing data for Artificial Intelligence (AI) -based arrhythmia detection are disclosed. In accordance with the techniques of this disclosure, a computing system may obtain cardiac Electrogram (EGM) strips of waveforms representing cardiac rhythms of the same patient. Additionally, the computing system may pre-process the cardiac EGM strips. The computing system may then apply a deep learning model to the preprocessed cardiac EGM strips to generate arrhythmia data that indicates whether the cardiac EGM strips represent one or more occurrences of one or more arrhythmias.)

1. A method, comprising:

obtaining, by a computing system, one or more cardiac Electrogram (EGM) strips representing a waveform of a cardiac rhythm of a patient;

preprocessing, by the computing system, the one or more cardiac EGM strips; and

applying, by the computing system, a deep learning model to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

2. The method of claim 1, wherein preprocessing the one or more cardiac EGM strips comprises:

generating, by the computing system, one or more temporally-modified cardiac EGM strips by modifying a temporal resolution of the one or more cardiac EGM strips to match an expected temporal resolution of a deep-learning model; and

generating, by the computing system, one or more preprocessed cardiac EGM strips by subtracting an average of the samples of the one or more temporally modified cardiac EGM strips from the samples of the one or more temporally modified cardiac EGM strips.

3. The method of claim 1, wherein preprocessing the one or more cardiac EGM strips comprises:

generating, by the computing system, one or more temporally-modified cardiac EGM strips by modifying a temporal resolution of the one or more cardiac EGM strips to match an expected temporal resolution of the deep-learning model;

generating, by the computing system, one or more time-modified cardiac EGM strips by subtracting an average of the samples of the one or more time-modified cardiac EGM strips from the samples of the one or more time-modified cardiac EGM strips;

calculating, by the computing system, a moving signal envelope over a sliding window of the one or more mean-subtracted cardiac EGM strips; and

normalizing, by the computing system, the one or more mean-subtracted cardiac EGM strips based on the moving signal envelope.

4. The method of any of claims 1-3, wherein:

the arrhythmia data is first arrhythmia data;

the method further comprises:

obtaining, by the computing system, a marker channel;

pre-processing, by the computing system, the marker channel;

applying, by the computing system, a marker channel-based deep learning model to the preprocessed marker channels to generate second cardiac arrhythmia data indicative of a second set of occurrences of the one or more cardiac arrhythmias;

determining, by the computing system, whether the first cardiac arrhythmia data is consistent with the second cardiac arrhythmia data; and

at least one of:

adjusting, by the computing system, a confidence level of an occurrence of one of the arrhythmias based on whether the occurrence of the arrhythmia is in both the first set of occurrences and the second set of occurrences or only one of the first set of occurrences and the second set of occurrences; or

Extending, by the computing system, a duration of a monitoring session of a medical device that generates the one or more cardiac EGM strips based on the occurrence of the arrhythmia being in both the first set of occurrences and the second set of occurrences.

5. A method according to claim 4, wherein the marker channel is indicative of a detected QRS complex.

6. The method of any of claims 1-5, wherein:

preprocessing the one or more cardiac EGM strips comprises:

scaling, by the computing system, a waveform represented by the one or more cardiac EGM strips; and

decomposing, by the computing system, a waveform represented by the one or more cardiac EGM strips into a plurality of channels corresponding to different frequency bands, and

applying the deep learning model to the one or more preprocessed cardiac EGM bands includes applying, by the computing system, the deep learning model to the channel to generate the arrhythmia data.

7. The method of any of claims 1-6, wherein:

preprocessing the one or more cardiac EGM strips comprises:

scaling, by the computing system, a waveform represented by the one or more cardiac EGM strips; and

generating, by the computing system, a transformed waveform by transforming the scaled waveform into a two-dimensional time-frequency domain, and

applying the deep learning model to the one or more preprocessed cardiac EGM bands includes applying the deep learning model to the transformed waveforms.

8. The method of any of claims 1-7, wherein:

preprocessing the one or more cardiac EGM strips includes applying, by the computing system, a learned scaling factor to waveforms represented by the one or more cardiac EGM strips; and is

Applying the deep learning model comprises:

determining, by the computing system, QRS probability values corresponding to the plurality of time points, each of the QRS probability values indicating a respective probability that a peak of a QRS complex occurs during the time point corresponding to the QRS probability value; and

detecting, by the computing system, the set of occurrences of the one or more arrhythmias by providing the QRS probability values and the one or more preprocessed cardiac EGM strips as inputs to a neural network that generates the arrhythmia data.

9. The method of any one of claims 1-8, wherein preprocessing the one or more cardiac EGM strips comprises:

determining, by the computing system, a polarity of the one or more cardiac EGM strips; and

reversing, by the computing system, the polarity of the one or more cardiac EGM strips based on the polarity of the one or more cardiac EGM strips not being an expected polarity of the deep learning model.

10. The method of claim 9, wherein:

the deep learning model is a first deep learning model, and

determining the polarity of the one or more cardiac EGM strips includes applying, by the computing system, a second deep learning model to the one or more cardiac EGM strips to determine the polarity of the one or more cardiac EGM strips.

11. The method of any of claims 1-10, wherein the method further comprises:

obtaining, by the computing system, training input vectors, wherein each of the training input vectors comprises a segment of a training cardiac EGM band and device classification data indicative of one or more arrhythmias detected in the training cardiac EGM band;

training, by the computing system, a self-encoder based on the training input vector to reconstruct training cardiac EGM bands of the training input vector;

obtaining, by the computing system, additional device classification data;

providing, by the computing system, the one or more preprocessed cardiac EGM strips and the additional device classification data to an input layer of the self-encoder; and

determining, by the computing system, based on probability values generated by an intermediate layer of the self-encoder, whether the classification data correctly identifies an arrhythmia, if any,

wherein each of the probability values corresponds to a different arrhythmia and indicates a confidence level that the one or more preprocessed cardiac EGM strips contain the arrhythmia.

12. The method of any one of claims 1-11, wherein preprocessing the one or more cardiac EGM strips comprises one or more of:

scaling samples of a signal of the one or more cardiac EGM strips such that the samples of the signal of the one or more cardiac EGM strips are distributed within an expected sample value range of the deep learning model, or

Increasing or decreasing a sampling rate of the signals of the one or more cardiac EGM strips to match an expected sampling rate of the deep learning model.

13. A computing system, comprising:

a storage device configured to store one or more cardiac Electrogram (EGM) strips of waveforms representative of a cardiac rhythm of a patient;

one or more processing circuits configured to:

pre-processing the one or more cardiac EGM strips; and

applying a deep learning model to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

14. The computing system of claim 13, wherein the one or more processing circuits are configured to perform the method of any of claims 2-12.

15. A computer-readable storage medium having instructions stored thereon that, when executed, cause a computing system to:

obtaining one or more cardiac Electrogram (EGM) strips of a waveform representative of a cardiac rhythm of a patient;

pre-processing the one or more cardiac EGM strips; and

applying a deep learning model to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

16. The computer-readable storage medium of claim 15, wherein the instructions, when executed, cause the computing system to perform the method of any of claims 2 to 12.

Technical Field

The present disclosure relates generally to health monitoring and, more particularly, to monitoring cardiac health.

Background

Malignant tachyarrhythmias, such as ventricular fibrillation, are uncoordinated contractions of the myocardium of the ventricles in the heart and are the most common identified arrhythmias in patients with sudden cardiac arrest. If this arrhythmia lasts more than a few seconds, it can lead to cardiogenic shock and cessation of effective blood circulation. Thus, Sudden Cardiac Death (SCD) can occur within minutes.

Implantable or non-implantable medical devices may monitor cardiac arrhythmias in a patient's heart. A user, such as a physician, may view data generated by the medical device regarding the arrhythmia. The user may diagnose a medical condition of the patient based on the arrhythmia.

Disclosure of Invention

In general, this disclosure describes techniques for preparing data for Artificial Intelligence (AI) -based arrhythmia detection. As described herein, a computing system may obtain cardiac Electrogram (EGM) strips of waveforms representative of a cardiac rhythm of a patient. Additionally, the computing system may pre-process the cardiac EGM strips. The computing system may then apply a deep learning model to the preprocessed cardiac EGM strips to generate arrhythmia data that indicates whether the cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

In one aspect, the present disclosure describes a method comprising: obtaining, by a computing system, one or more cardiac Electrogram (EGM) strips representing a waveform of a cardiac rhythm of a patient; preprocessing, by the computing system, the one or more cardiac EGM strips; and applying, by the computing system, a deep learning model to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

In another aspect, the present disclosure describes a computing system comprising: a storage device configured to store one or more cardiac Electrogram (EGM) strips of waveforms representative of a cardiac rhythm of a patient; one or more processing circuits configured to: pre-processing the one or more cardiac EGM strips; and applying a deep learning model to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

In another aspect, the present disclosure describes a computer-readable storage medium having instructions stored thereon that, when executed, cause a computing system to obtain one or more cardiac Electrogram (EGM) strips representing waveforms of a cardiac rhythm of a patient; pre-processing the one or more cardiac EGM strips; and applying a deep learning model to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

This summary is intended to provide an overview of the subject matter described in this disclosure. It is not intended to provide an exclusive or exhaustive explanation of the apparatus and methods described in detail in the following figures and description. Additional details of one or more examples are set forth in the accompanying drawings and the description below.

Drawings

FIG. 1 is a block diagram illustrating a system for analyzing one or more aspects of a patient's cardiac rhythm according to one or more techniques of the present disclosure.

Fig. 2 is a conceptual diagram illustrating an Implantable Medical Device (IMD) and a lead of the system of fig. 1 in more detail.

Fig. 3 is a block diagram of an example implantable medical device according to one or more techniques of this disclosure.

Fig. 4 is a block diagram illustrating an example computing device operating in accordance with one or more techniques of this disclosure.

Fig. 5 is a flow diagram illustrating example operations of an Artificial Intelligence (AI) system in accordance with one or more techniques of this disclosure.

Fig. 6 is a flow diagram illustrating first example operations for preprocessing one or more cardiac Electrogram (EGM) strips in accordance with one or more techniques of this disclosure.

Fig. 7 is a flow diagram illustrating second example operations for preprocessing cardiac EGM strips according to one or more techniques of the present disclosure.

Fig. 8A is a conceptual diagram illustrating example waveforms and envelopes determined according to one or more techniques of the present disclosure.

Fig. 8B is a conceptual diagram illustrating an example normalized waveform generated according to one or more techniques of this disclosure.

Fig. 9 is a flow diagram illustrating example operations of an AI system according to one or more techniques of this disclosure.

Fig. 10A is a conceptual diagram illustrating a graph of an example cardiac waveform and QRS markers detected by the apparatus.

Fig. 10B is a conceptual diagram illustrating a graph of heart rate over time based on a marker channel.

Fig. 11 is a flow diagram illustrating a third example operation for preprocessing cardiac EGM strips according to one or more techniques of the present disclosure.

Fig. 12 is a conceptual diagram illustrating an example original cardiac waveform and waveforms of channels corresponding to different frequency bands according to one or more techniques of the present disclosure.

Fig. 13 is a flow diagram illustrating a fourth example operation for preprocessing cardiac EGM strips according to one or more techniques of the present disclosure.

Fig. 14 is a flow diagram illustrating example operations of an AI system according to one or more techniques of this disclosure.

Fig. 15 is a flow diagram illustrating fifth example operations for preprocessing cardiac EGM strips according to one or more techniques of the present disclosure.

Fig. 16 is a conceptual diagram illustrating example operation of an AI system including an autoencoder according to one or more techniques of this disclosure.

Fig. 17 is a flow diagram illustrating example operations in which an autoencoder is used to confirm device classification in accordance with one or more techniques of the present disclosure.

Like reference numerals refer to like elements throughout the drawings and the description.

Detailed Description

FIG. 1 is a block diagram illustrating a system 10 for analyzing one or more aspects of a cardiac rhythm of a patient 14 in accordance with the techniques of the present disclosure. The system 10 includes a medical device 16. One example of such a medical device is an Implantable Medical Device (IMD), as shown in fig. 1. As illustrated by the example system 10 in fig. 1, in some examples, the medical device 16 may be, for example, an implantable cardiac monitor, an implantable cardiac pacemaker, an implantable cardioverter/defibrillator (ICD), or a pacemaker/cardioverter/defibrillator. In some examples, medical device 16 is a non-implantable medical device, such as a non-implantable cardiac monitor (e.g., a holter monitor).

In the example of fig. 1, medical device 16 is connected to leads 18, 20, and 22 and is communicatively coupled to an external device 27, which in turn is communicatively coupled to computing system 24 through a communication network 25. Medical device 16 senses electrical signals, such as cardiac Electrograms (EGMs), that accompany the depolarization and repolarization of heart 12 through electrodes on one or more leads 18, 20, and 22 or the housing of medical device 16. The medical device 16 may also deliver therapy to the heart 12 in the form of electrical signals through electrodes positioned on one or more leads 18, 20, and 22 or a housing of the medical device 16. The therapy may be pacing, cardioversion, and/or defibrillation pulses. The medical device 16 may monitor the cardiac EGM signals collected by the electrodes on the leads 18, 20, or 22 and diagnose and treat arrhythmias based on the cardiac EGM signals.

In some examples, medical device 16 includes communication circuitry 17 that includes any suitable circuitry, firmware, software, or any combination thereof, for communicating with another device, such as external device 27 of fig. 1. For example, the communication circuitry 17 may include one or more processors, memories, radios, antennas, transceivers, receivers, modulation and demodulation circuitry, filters, amplifiers, etc., for radio frequency communication with other devices, such as the computing system 24. The medical device 16 may use the communication circuitry 17 to receive downlink data to control one or more operations of the medical device 16 and/or to transmit uplink data to the external device 27.

Leads 18, 20, 22 extend into heart 12 of patient 14 to sense electrical activity of heart 12 and/or deliver electrical stimulation to heart 12. In the example shown in fig. 1, Right Ventricular (RV) lead 18 extends through one or more veins (not shown), the superior vena cava (not shown), and right atrium 26, and into right ventricle 28. Left Ventricular (LV) lead 20 extends through one or more veins, the vena cava, right atrium 26, and into coronary sinus 30 to a region adjacent to the free wall of left ventricle 32 of heart 12. Right Atrial (RA) lead 22 extends through one or more veins and the vena cava, and into right atrium 26 of heart 12.

Although the example system 10 of fig. 1 depicts a medical device 16, in other examples, the techniques of this disclosure may be applied to other types of medical devices that are not necessarily implantable. For example, a medical device in accordance with the techniques of this disclosure may include a wearable medical device or "smart" garment worn by the patient 14. For example, such a medical device may take the form of a watch worn by the patient 14 or circuitry adhesively attached to the patient 14. In another example, a medical device as described herein may include an external medical device having an implantable electrode.

In some examples, the external device 27 takes the form of an external programmer or mobile device, such as a mobile phone, a "smart" phone, a laptop computer, a tablet computer, a Personal Digital Assistant (PDA), or the like. In some examples, external device 27 is a CareLink available from Medtronic, Inc^TMA monitor. A user, such as a physician, technician, surgeon, electrophysiologist, or other clinician, may interact with the external device 27 to retrieve physiological or diagnostic information from the medical device 16. The user (e.g., the patient 14 or clinician as described above) may also interact with the external device 27 to program the medical device 16, e.g., to select or adjust values for operating parameters of the medical device 16. External device 27 may contain processing circuitry, memory, user interface, and communication circuitry capable of transmitting information to and receiving information from each of medical device 16 and computing system 24.

In some examples, computing system 24 takes the form of a handheld computing device, computer workstation, server, or other networked computing device, smart phone, tablet computer, or external programmer that includes a user interface for presenting information to a user and receiving input from a user. In some examples, computing system 24 may include one or more devices that implement a machine learning system, such as a neural network, a deep learning system, or other type of machine learning system. A user (e.g., a physician, technician, surgeon, electrophysiologist, or other clinician) may interact with the computing system 24 to retrieve physiological or diagnostic information from the medical device 16. The user may also interact with computing system 24 to program medical device 16, for example, to select values for operating parameters of the IMD. The computing system 24 may include a processor configured to evaluate the cardiac EGM (or segment thereof) and/or other sensing signals transmitted from the medical device 16 to the computing system 24.

Network 25 may include one or more computing devices (not shown), such as one or more non-edge switches, routers, hubs, gateways, security devices such as firewalls, intrusion detection and/or intrusion prevention devices, servers, computer terminals, laptops, printers, databases, wireless mobile devices such as cellular telephones or personal digital assistants, wireless access points, bridges, cable modems, application accelerators, or other network devices. Network 25 may contain one or more networks managed by a service provider and may therefore form part of a large-scale public network infrastructure, such as the internet. The network 25 may provide computing devices, such as the computing system 24 and the medical devices 16, with access to the internet and may provide a communication framework that allows the computing devices to communicate with each other. In some examples, the network 25 may be a private network that provides a communications framework that allows the computing system 24, medical devices 16, and EMR database 66 to communicate with one another, but isolates the computing system 24, medical devices 16, and EMR database 66 from external devices for security purposes. In some examples, communications between the computing system 24, the medical device 16, and the EMR database 66 are encrypted.

The external device 27 and the computing system 24 may communicate via wireless or non-wireless communication over the network 25 using any technique known in the art. In some examples, computing system 24 is a remote device that communicates with external device 27 via an intermediate device (such as a local access point, wireless router, or gateway) located in network 25. Although in the example of fig. 1, external device 27 and computing system 24 communicate over network 25, in some examples, external device 27 and computing system 24 communicate directly with each other. Examples of communication techniques may include, for example, according toOr communication by the BLE protocol. Other communication techniques are also contemplated. The computing system 24 may also communicate with one or more other external devices using a variety of known wired and wireless communication techniques.

The EMR database 66 stores EMR data for the patient 14. The EMR database 66 may include processing circuitry and one or more storage media (e.g., Random Access Memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electronically erasable programmable read-only memory (EEPROM), or flash memory.

In one example, the computing system 24 receives patient data collected by the medical device 16 of the patient 14. In some examples, the patient data includes physiological data of the patient 14, such as one or more of: activity level of the patient 14, heart rate of the patient 14, posture of the patient 14, cardiac electrogram of the patient 14, blood pressure of the patient 14, pulse transit time of the patient 14, respiration rate of the patient 14, hypopnea index or apnea of the patient 14, accelerometer data of the patient 14, characteristics of accelerometer data derived from the patient 14, such as activity count, posture, statistical control process variables, etc., raw electromyogram or cardiac EGM of the patient 14, one or more characteristics of raw electromyogram derived from the patient 14, such as heart rate variability, t-wave alternans, QRS morphology, etc., interval data and characteristics derived from interval data, heart sounds, potassium levels, blood glucose index, temperature of the patient 14, or any data derivable from the above parameterized data, or any other type of patient parameter data. In some instances, the medical device 16 or another device may automatically generate patient parameter data by processing information from one or more sensors. For example, the medical device 16 may determine, via one or more sensors, that the patient 14 has fallen, that the patient 14 is weak or sick, or that the patient 14 is suffering from sleep apnea.

In some examples, the patient data includes environmental data, such as air quality measurements, ozone levels, particle counts or contamination levels near the patient 14, ambient temperature, or time of day. In some examples, one of the medical device or the external device 27 may sense environmental data via one or more sensors. In another example, the environmental data is received by the external device 27 via an application executing on the external device 27 (e.g., a weather application) and uploaded to the computing system 24 over the network 25. In another example, the computing system 24 collects the environmental data directly from a cloud service with location-based data of the patient 14.

In some examples, the patient data includes patient symptom data uploaded by the patient 14 via an external device, such as external device 27. For example, the patient 14 may upload patient symptom data via an application executing on a smartphone. In some instances, the patient 14 may upload patient symptom data via a user interface (not depicted in fig. 1), such as through a touch screen, keyboard, graphical user interface, voice commands, and so forth.

In some examples, the patient data includes device-related data, such as one or more of impedance of one or more electrodes of the medical device, selection of electrodes, a drug delivery schedule of the medical device, history of electrical pacing therapy delivered to the patient, or diagnostic data of the medical device. In some examples, the medical device that collects patient data is an IMD. In other examples, the medical device collecting the patient data is another type of patient device, such as a wearable medical device or a mobile device (e.g., a smartphone) of the patient 14. In some examples, the computing system 24 receives patient data periodically, such as daily.

In some examples, the computing system 24 further receives EMR data for the patient 14 from the EMR database 66. EMR data may be considered another form of patient data. In some examples, the EMR data stored by the EMR database 66 may contain many different types of historical medical information about the patient 14. For example, the EMR database 66 may store a patient's medication history, a patient's surgical history, a patient's hospital history, a patient's potassium level over time, one or more laboratory test results of the patient 14, a patient's 14 cardiovascular history, or a patient's 14 comorbidities, such as atrial fibrillation, heart failure, or diabetes, as examples.

Computing system 24 may implement a cardiac EGM monitoring system that may help manage chronic cardiac disease. In accordance with the techniques of this disclosure, to implement a cardiac EGM monitoring system, computing system 24 may apply Artificial Intelligence (AI) techniques to analyze patient data, such as cardiac EGM data. Example AI techniques may include deep learning or other machine learning techniques. Neural network algorithms are an example of deep learning algorithms.

The AI system is a computing system that includes a memory and one or more processing circuits configured to perform AI techniques. In the context of fig. 1, the AI system may be the medical device 16, the computing system 24, the external device 27, or another device or device system. Thus, in this disclosure, unless otherwise indicated, the discussion of actions performed by the AI system may apply to actions performed by any of these devices.

The AI system may generate data regarding one or more aspects of the heart rhythm of the patient 14. For example, the AI system may generate arrhythmia data based at least in part on cardiac EGM strips obtained from one or more medical devices, such as medical device 16, that indicates whether the cardiac EGM strips represent one or more occurrences of one or more arrhythmias. The cardiac EGM band includes data representing the patient's cardiac rhythm over successive time periods (e.g., 30 seconds, 45 seconds, etc.). A cardiac EGM strip may include a series of samples of a waveform representing a heart rhythm. A user (e.g., a technician, physician, patient, healthcare professional, or other type of user) may view the detected occurrence of one or more arrhythmias for diagnostic purposes or as part of performing ongoing care of the patient 14. In addition to cardiac EGM strips, the AI system may use one or more other types of patient data to detect the occurrence of an arrhythmia, such as information from an electronic medical record of the patient 14.

The AI system may be trained to identify one or more aspects of the patient's 14 cardiac rhythm that are of interest to a given cardiac EGM band by applying one or more deep learning models that have been trained to identify such aspects of the patient's 14 cardiac rhythm. Aspects of the cardiac rhythm of the patient 14 may include various arrhythmias, the location of such arrhythmias within one or more cardiac EGM strips (which reflect the time at which the arrhythmia occurred), morphological aspects of the arrhythmia occurrence, and the like. The cardiac rhythm classification model may be trained on cardiac EGM bands extracted from a population of subjects, individual patients, groups of patients, and in some instances other data.

By preprocessing the input data provided to the deep learning model, the performance of the deep learning model may be improved. For example, as described in this disclosure, preprocessing the data may enable a deep learning model to be used with cardiac EGM strips generated by multiple types of devices. Thus, in accordance with the techniques of this disclosure, the AI system may obtain a cardiac EGM strip of waveforms representative of the patient's cardiac rhythm. Additionally, the AI system may preprocess cardiac EGM bands. The computing system may then apply a deep learning model to the preprocessed cardiac EGM strips to generate arrhythmia data that indicates whether the cardiac EGM strips represent one or more occurrences of one or more arrhythmias. For example, the deep learning system may be trained to generate arrhythmia data that includes vectors of elements corresponding to different arrhythmias. The values of the elements in the vector indicate whether the occurrence of the corresponding arrhythmia is present in the cardiac EGM band.

Fig. 2 is a conceptual diagram illustrating medical device 16 and leads 18, 20, 22 of system 10 of fig. 1 in more detail. In the illustrated example, bipolar electrodes 40 and 42 are positioned adjacent the distal end of lead 18, and bipolar electrodes 48 and 50 are positioned adjacent the distal end of lead 22. Additionally, four electrodes 44, 45, 46, and 47 are positioned adjacent the distal end of the lead 20. Lead 20 may be referred to as a quadrupolar LV lead. In other examples, lead 20 may include more or fewer electrodes. In some examples, LV lead 20 comprises a segmented electrode, e.g., where each of a plurality of longitudinal electrode locations of the lead, such as the locations of electrodes 44, 45, 46, and 47, comprise a plurality of discrete electrodes arranged at respective circumferential locations around the circumference of the lead.

In the illustrated example, electrodes 40 and 44-48 take the form of ring electrodes, and electrodes 42 and 50 may take the form of extendable helix tip electrodes mounted telescopically within insulative electrode heads 52 and 56, respectively. Leads 18 and 22 also contain elongated electrodes 62 and 64, respectively, which may take the form of coils. In some examples, each of electrodes 40, 42, 44-48, 50, 62, and 64 is electrically coupled to a respective conductor within the lead body of its associated lead 18, 20, 22, and thereby to circuitry within medical device 16.

In some examples, the medical device 16 includes one or more housing electrodes, such as the housing electrode 4 shown in fig. 2, which may be integrally formed with an outer surface of the hermetically sealed housing 8 of the medical device 16 or otherwise coupled to the housing 8. In some examples, the housing electrode 4 is defined by an uninsulated portion of an outward facing portion of the housing 8 of the medical device 16. Other divisions between the insulated and non-insulated portions of the housing 8 may be used to define two or more housing electrodes. In some examples, the housing electrode comprises substantially all of the housing 8.

Housing 8 encloses signal generation circuitry that generates therapeutic stimulation (e.g., cardiac pacing, cardioversion, and defibrillation pulses), as well as sensing circuitry for sensing electrical signals that accompany the depolarization and repolarization of heart 12. The housing 8 may also enclose a memory for storing the sensed electrical signals. The housing 8 may also enclose communication circuitry 17 for communication between the medical device 16 and the computing system 24.

Medical device 16 senses electrical signals attendant to the depolarization and repolarization of heart 12 via electrodes 4, 40, 42, 44-48, 50, 62 and 64. The medical device 16 may sense such electrical signals via any bipolar combination of the electrodes 40, 42, 44-48, 50, 62, and 64. Further, any of the electrodes 40, 42, 44-48, 50, 62, and 64 may be combined with the housing electrode 4 for unipolar sensing.

The illustrated number and configuration of leads 18, 20, and 22 and electrodes are merely examples. Other configurations, i.e., the number and positioning of leads and electrodes, are also possible. In some examples, system 10 may include additional leads or lead segments with one or more electrodes positioned at different locations in the cardiovascular system for sensing and/or delivering therapy to patient 14. For example, instead of or in addition to the inter-cardiac leads 18, 20, and 22, the system 10 may include one or more epicardial or extravascular (e.g., subcutaneous or substernal) leads that are not positioned within the heart 12.

The medical device 16 may transmit the patient data to the computing system 24 (e.g., via the external device 27). The patient data may include data based on electrical signals detected by electrodes 4, 40, 42, 44-48, 50, 62, and/or 64. For example, the medical device 16 may gather cardiac EGM data and send it to the computing system 24. In accordance with techniques of this disclosure, an AI system, which may be implemented by computing system 24, medical device 16, or another device, may pre-process patient data and use the pre-processed patient data to determine to generate arrhythmia data or other data related to the heart rhythm of patient 14. In some examples, medical device 16 may pre-process cardiac EGM strips and computing system 24 or external device 27 may apply a deep learning model to the pre-processed cardiac EGM strips to generate arrhythmia data.

Although described herein in the context of an example medical device 16 that provides therapeutic electrical stimulation, the techniques disclosed herein may be used with other types of devices. For example, the techniques may be implemented with: an extracardiac defibrillator coupled to electrodes external to the cardiovascular system, a transcatheter pacemaker configured to be implanted within the heart, such as Micra commercially available from medton, inc^TMTranscatheter pacing systems, plug-in heart monitors, such as the Reveal LINQ also commercially available from medtronic^TMICM, neurostimulator, drug delivery device, wearable device, such as wearable cardioverter defibrillator, fitness tracker, or other wearable device, mobile device, such as mobile deviceA telephone, "smart" phone, laptop, tablet computer, Personal Digital Assistant (PDA), or "smart" apparel, such as "smart" glasses or "smart" watches.

Fig. 3 is a block diagram of an example medical device 16 in accordance with the techniques of this disclosure. In the example shown, the medical device 16 includes processing circuitry 58, memory 59, communication circuitry 17, sensing circuitry 50, therapy delivery circuitry 52, sensors 57, and a power source 54. Memory 59 contains computer readable instructions that, when executed by processing circuitry 58, cause medical device 16 and processing circuitry 58 to perform various functions attributed herein to medical device 16 and processing circuitry 58 (e.g., performing short-term prediction of cardiac arrhythmias, delivering therapy, such as anti-tachycardia pacing, bradycardia pacing, and post-shock pacing therapy, etc.). Memory 59 may include any volatile, non-volatile, magnetic, optical, or electrical media, such as Random Access Memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), Electrically Erasable Programmable ROM (EEPROM), flash memory, or any other digital or analog media.

The processing circuitry 58 may include any one or more of the following: a microprocessor, controller, Digital Signal Processor (DSP), Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA), or equivalent discrete or analog logic circuitry. In some examples, processing circuitry 58 may include multiple components (e.g., any combination of one or more microprocessors, one or more controllers, one or more DSPs, one or more ASICs, or one or more FPGAs), as well as other discrete or integrated logic circuitry. The functionality attributed herein to processing circuitry 58 may be embodied as software, firmware, hardware, or any combination thereof.

Processing circuitry 58 controls therapy delivery circuitry 52 to deliver stimulation therapy to heart 5 according to therapy parameters that may be stored in memory 59. For example, the processing circuitry 58 may control the therapy delivery circuitry 52 to deliver electrical pulses having an amplitude, pulse width, frequency, or electrode polarity specified by the therapy parameters. In this manner, therapy delivery circuitry 52 may deliver pacing pulses (e.g., ATP pulses, bradycardia pacing pulses, or post-shock pacing therapy) to heart 5 via electrodes 34 and 40. In some examples, therapy delivery circuitry 52 may deliver pacing stimuli in the form of voltage or current electrical pulses, such as ATP therapy, bradycardia therapy, or post-shock pacing therapy. In other examples, therapy delivery circuitry 52 may deliver one or more of these types of stimuli in the form of other signals (e.g., sine waves, square waves, or other substantially continuous time signals).

Therapy delivery circuitry 52 is electrically coupled to electrodes 34 and 40 carried on the housing of medical device 16. Although the medical device 16 may include only two electrodes, such as electrodes 34 and 40, in other examples, the medical device 16 may utilize three or more electrodes. The medical device 16 may use any combination of electrodes to deliver therapy and/or detect electrical signals from the patient 12. In some examples, therapy delivery circuitry 52 includes charging circuitry, one or more pulse generators, capacitors, converters, switching modules, and/or other components capable of generating and/or storing energy for delivery as pacing therapy, cardiac resynchronization therapy, other therapies, or combinations of therapies. In some examples, the therapy delivery circuitry 52 delivers the therapy as one or more electrical pulses according to one or more therapy parameter sets that define the amplitude, frequency, voltage, or current of the therapy or other parameters of the therapy.

Sensing circuitry 50 monitors signals from one or more combinations (also referred to as vectors) of two or more of electrodes 4, 40, 42, 44-48, 50, 62 (fig. 2), and 64 (fig. 2) to monitor electrical activity, impedance, or other electrical phenomena of heart 12. In some examples, the sensing circuitry 50 includes one or more analog components, digital components, or a combination thereof. In some examples, the sensing circuitry 50 includes one or more sense amplifiers, comparators, filters, rectifiers, threshold detectors, analog-to-digital converters (ADCs), and so forth. In some examples, the sensing circuitry 50 may convert the sensed signals to digital form and provide the digital signals to the processing circuitry 58 for processing or analysis. In one example, sensing circuitry 50 amplifies the signals from electrodes 4, 40, 42, 44-48, 50, 62, and 64 and converts the amplified signals to multi-bit digital signals by an ADC.

In some examples, the sensing circuitry 50 performs sensing of cardiac electrograms to determine heart rate or heart rate variability, or to detect arrhythmias (e.g., tachyarrhythmias or bradycardias) or to sense other parameters or events from cardiac electrograms. The sensing circuitry 50 may also include switching circuitry for selecting which of the available electrodes (and electrode polarities) to use for sensing cardiac activity based on which electrode combination or electrode vector is used in the current sensing configuration. Processing circuitry 58 may control switching circuitry to select the electrode that serves as the sensing electrode and its polarity. The sensing circuitry 50 may contain one or more detection channels, each of which may be coupled to a selected electrode configuration to detect cardiac signals via the electrode configuration. In some examples, the sensing circuitry 50 may compare the processed signal to a threshold to detect the presence of atrial or ventricular depolarization and indicate the presence of atrial depolarization (e.g., P-wave) or ventricular depolarization (e.g., R-wave) to the processing circuitry 58. The sensing circuitry 50 may include one or more amplifiers or other circuitry for comparing the cardiac electrogram amplitude to a threshold, which may be adjustable.

The processing circuitry 58 may include timing and control modules that may be embodied in hardware, firmware, software, or any combination thereof. The timing and control modules may comprise dedicated hardware circuitry (e.g., an ASIC) separate from other processing circuitry 58 components (e.g., a microprocessor), or software modules executed by components of the processing circuitry 58, which may be a microprocessor or an ASIC. The timing and control module may implement a programmable counter. If the medical device 16 is configured to generate and deliver bradycardia pacing pulses to the heart 12, such counters may control the basic time intervals associated with DDD, VVI, DVI, VDD, AAI, DDI, DDDR, VVIR, DVIR, VDDR, AAIR, DDIR, or other pacing modes.

In some examples, the processing circuitry 58 of the medical device 16 implements the AI system 300. For example, the processing circuitry 58 may apply the deep learning model to cardiac EGM strips, as described elsewhere in this disclosure. The processing circuitry 58 may implement the AI system 300 using dedicated circuitry or by executing software instructions stored on a computer-readable medium, such as the memory 59. Sensing circuitry 50 may generate cardiac EGM strips based on data received from electrodes 4, 40, 42, 44-48, 50, 62, and 64. The communication circuitry 17 may transmit the cardiac EGM strips and/or other data to the external device 27.

The memory 59 may be configured to store various operating parameters, therapy parameters, sensed and detected data, and any other information related to the therapy and treatment of the patient 12. In the example of fig. 3, memory 59 may store, for example, sensed cardiac EGMs related to detected or predicted arrhythmias, and therapy parameters defining the delivery of therapy provided by therapy delivery circuitry 52. In other instances, the memory 59 may act as a temporary buffer for storing data until the data can be uploaded to the computing system 24.

The communication circuitry 17 comprises any suitable circuitry, firmware, software, or any combination thereof, for communicating with another device (e.g., computing system 24) via the network 25 of fig. 1. For example, the communication circuitry 17 may include one or more antennas, modulation and demodulation circuitry, filters, amplifiers, etc. for radio frequency communication with other devices (e.g., computing system 24) via network 25. Under the control of processing circuitry 58, communication circuitry 17 may receive downlink telemetry from computing system 24 and transmit uplink telemetry to the computing system via an antenna, which may be internal and/or external. Processing circuitry 58 may provide data to be transmitted upstream to computing system 24 and control signals for telemetry circuitry within communication circuitry 17, for example, via an address/data bus. In some examples, communication circuitry 17 may provide the received data to processing circuitry 58 via a multiplexer.

The power source 54 may be any type of device configured to hold an electrical charge to operate the circuitry of the medical device 16. The power source 54 may be provided as a rechargeable or non-rechargeable battery. In other examples, the power source 54 may incorporate an energy scavenging system that stores electrical energy from the movement of the medical device 16 within the patient 12.

In accordance with the techniques of this disclosure, the medical device 16 collects patient data of the patient 14 via the sensing circuitry 50 and/or the sensor 57. The sensor 57 may comprise one or more sensors, such as one or more accelerometers, pressure sensors, optical sensors for O2 saturation, and the like. In some examples, the patient data comprises one or more of: activity level of the patient 14, heart rate of the patient 14, posture of the patient 14, cardiac electrogram of the patient 14 (e.g., cardiac EGM strip of the patient 14), blood pressure of the patient 14, accelerometer data of the patient 14, or other types of patient parameter data. The medical device 16 uploads patient parameter data to the computing system 24 via the communication circuitry 17 over the network 25. In some examples, the medical device 16 uploads patient parameter data to the computing system 24 on a daily basis. In some examples, the patient parameter data includes one or more values representing an average measurement of the patient 14 over an extended period of time (e.g., about 24 hours to about 48 hours). For example, one or more other devices, such as a wearable medical device or mobile device (e.g., a smartphone) of the patient 14, may collect and upload patient parameter data to the external device 27 and/or the computing system 24.

Although described herein in the context of an example medical device 16 that provides therapeutic electrical stimulation, the techniques disclosed herein for short-term prediction of cardiac arrhythmias may be used with other types of devices. For example, the techniques may be implemented with: transcatheter pacemaker configured for implantation within the heart, such as Micra, commercially available from Meindon force of Dublin, Ireland^TMTranscatheter pacing system, insertionImplantable heart monitor, such as the Reveal LINQ also commercially available from Meindon force^TMAn ICM, a neurostimulator, a drug delivery device, a wearable device, such as a wearable cardioverter defibrillator, a fitness tracker, or other wearable device, a mobile device, such as a mobile phone, a "smart" phone, a laptop, a tablet computer, a Personal Digital Assistant (PDA), or a "smart" garment, such as "smart" glasses or a "smart" watch.

Fig. 4 is a block diagram illustrating an example computing system 24 operating in accordance with one or more techniques of this disclosure. In one example, computing system 24 includes processing circuitry 402 for executing an application 424 that includes a monitoring system 450 or any other application described herein. Although shown in fig. 4 as a stand-alone computing system 24 for purposes of example, the computing system 24 may be any component or system that contains processing circuitry or other suitable computing environment for executing software instructions, and for example, need not contain one or more of the elements shown in fig. 4 (e.g., communication circuitry 406; and in some instances, components such as storage 408 may not be co-located or in the same rack as other components). In some instances, computing system 24 may be a cloud computing system distributed across multiple devices.

As shown in the example of fig. 4, computing system 24 includes processing circuitry 402, one or more input devices 404, communication circuitry 406, one or more output devices 412, one or more storage devices 408, and a User Interface (UI) device 410. In one example, computing system 24 further includes one or more application programs 424 (such as monitoring system 450) and operating system 416 that are executable by computing system 24. Each of the components 402, 404, 406, 408, 410, and 412 are coupled (physically, communicatively, and/or operatively) for inter-component communication. In some examples, the communication channel 414 may include a system bus, a network connection, an interprocess communication data structure, or any other method for communicating data. As one example, components 402, 404, 406, 408, 410, and 412 can be coupled by one or more communication channels 414.

In one example, the processing circuitry 402 is configured to implement functionality and/or process instructions for execution within the computing system 24. For example, processing circuitry 402 may be capable of processing instructions stored in storage 408. Examples of processing circuitry 402 may include any one or more of the following: a microprocessor, a controller, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or equivalent discrete or integrated logic circuitry.

One or more storage devices 408 may be configured to store information within computing system 24 during operation. In some examples, storage 408 is described as a computer-readable storage medium. In some examples, storage 408 is a temporary memory, meaning that the primary purpose of storage 408 is not long-term storage. In some examples, storage 408 is described as volatile memory, meaning that storage 408 does not maintain stored content when the computer is turned off. Examples of volatile memory include Random Access Memory (RAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), and other forms of volatile memory known in the art. In some examples, storage 408 is used to store program instructions that are executed by processing circuitry 402. In one example, storage 408 is used by software or applications 424 running on computing system 24 to temporarily store information during program execution.

In some examples, storage 408 also includes one or more computer-readable storage media. Storage 408 may be configured to store larger amounts of information than volatile memory. Storage 408 may be further configured to store information for long periods of time. In some examples, storage 408 includes non-volatile storage elements. Examples of such non-volatile storage elements include magnetic hard disks, optical disks, floppy disks, flash memory, or various forms of electrically programmable memory (EPROM) or Electrically Erasable and Programmable (EEPROM) memory. In some examples, the storage device 408 may store patient data, such as cardiac EGM strips.

In some examples, computing system 24 also includes communication circuitry 406. In one example, the computing system 24 utilizes the communication circuitry 406 to communicate with external devices, such as the medical device 16 and the EMR database 66 of fig. 1. The communication circuitry 406 may comprise a network interface card, such as an ethernet card, an optical transceiver, a radio frequency transceiver, or any other type of device that can send and receive information. Other examples of such network interfaces may include 3G, 4G, 5G, and WI-FI^TMAnd (4) radio.

In one example, the computing system 24 also includes one or more user interface devices 410. In some examples, the user interface device 410 is configured to receive input from a user through tactile, audio, or video feedback. Examples of user interface device 410 include a presence-sensitive display, a mouse, a keyboard, a voice response system, a camera, a microphone, or any other type of device for detecting commands from a user. In some examples, a presence-sensitive display includes a touch-sensitive screen.

One or more output devices 412 may also be included in the computing system 24. In some examples, the output device 412 is configured to provide output to a user using tactile, audio, or video stimuli. In one example, output device 412 includes a presence-sensitive display, a sound card, a video graphics adapter card, or any other type of device for converting signals into an appropriate form understandable to humans or machines. In some examples, output device 412 includes a display device. Additional examples of output device 412 include a speaker, a Cathode Ray Tube (CRT) monitor, a Liquid Crystal Display (LCD), or any other type of device that can generate an understandable output to a user.

The computing system 24 may contain an operating system 416. In some examples, operating system 416 controls the operation of components of computing system 24. For example, in one example, the operating system 416 facilitates communication of one or more application programs 424 and the monitoring system 450 with the processing circuitry 402, the communication circuitry 406, the storage 408, the input device 404, the user interface device 410, and the output device 412.

Application 422 may also contain program instructions and/or data that are executable by computing system 24. An example application 422 executable by the computing system 24 may include a monitoring system 450. Other additional applications not shown may alternatively or additionally be included to provide other functionality described herein and are not depicted for simplicity.

In accordance with the techniques of this disclosure, the application 424 includes a monitoring system 450. Monitoring system 450 may be configured to receive patient data, evaluate the patient data, and generate output data. For example, in one example, when monitoring system 450 determines that patient 14 (fig. 1) is likely to experience one or more arrhythmic events belonging to one or more types of arrhythmia, monitoring system 450 may generate a notification. In another example, monitoring system 450 may generate one or more maps that illustrate changes in one or more aspects of the cardiac rhythm of patient 14.

As shown in the example of fig. 4, in some examples, the monitoring system 450 may implement an AI system 451 that includes a preprocessing unit 452 and a deep learning model 454. While the remainder of the disclosure refers to the preprocessing unit 452 and the deep learning model 454, such references may apply equally to preprocessing units and deep learning models implemented in devices or systems other than the computing system 24 (e.g., in the medical device 16 or the external device 27, or a combination thereof). Further, the AI system 451, the preprocessing unit 452, and the deep learning model 454 may be implemented outside of the context of the monitoring system. The preprocessing unit 452 may preprocess input data provided as input by the AI system 451 to the deep learning model 454, including cardiac EGM strips.

In some examples, deep learning model 454 is implemented using one or more neural network systems, deep learning systems, or other types of supervised or unsupervised machine learning systems. For example, the deep learning model 454 may be implemented by a feed-forward neural network, such as a convolutional neural network, a radial basis function neural network, a recurrent neural network, a modular or associative neural network. In some examples, the AI system 451 trains the deep learning model 454 with patient data for one or more patient populations including cardiac EGM strips to generate data regarding one or more aspects of cardiac rhythms of patients in the population. In some examples, after the AI system 451 has pre-trained the deep learning model 454, the AI system 451 may further train the deep learning model 454 with patient data specific to the patient 14 or a smaller patient group.

In some examples, the AI system 451 trains the deep learning model 454 with patient data for a patient population, determines an error rate for the deep learning model 454, and then feeds the error rate back to the deep learning model 454 to allow the deep learning model 454 to update its predictions based on the error rate. In some instances, the error rate may correspond to a difference between output data determined by the deep learning model 454 based on the input data and pre-labeled output data of the same input data. In some examples, the AI system 451 may use an error function to determine the error rate. The error function may be implemented using signal processing techniques and heuristics in a manner conventionally used to detect the occurrence of arrhythmias. In some examples, monitoring system 450 may receive feedback from a user (e.g., patient 14, a clinician, or another type of person) indicating whether a detected arrhythmia occurred in patient 14 within a particular time period. In some examples, monitoring system 450 may receive a message from medical device 16 indicating that medical device 16 has detected (or has not detected) the occurrence of an arrhythmia of patient 14. In some examples, monitoring system 450 may obtain feedback in other ways, such as by periodically examining EMR data to determine whether an arrhythmia is occurring. The monitoring system 450 may update the deep learning model 454 with feedback. Thus, the training process may occur iteratively to progressively improve the data generated by the deep learning model 454 by "learning" from the correct and incorrect data generated by the deep learning model 454 in the past. Further, the training process may be used to further fine tune the deep learning model 454, which is trained using population-based data, to generate more accurate data for a particular individual. In some instances, the personnel monitoring the service may provide feedback.

In some examples, deep learning model 454 is implemented using a neural network. The neural network may include an input layer and an output layer, and one or more hidden layers between the input layer and the output layer. Each layer of the neural network may contain one or more artificial neurons, which will be referred to simply as neurons by the present disclosure. The input layer of the neural network includes a plurality of input neurons. The input layer may contain a separate input neuron for each sample value of a segment of the cardiac EGM strip. In some examples, the segment may be connected to a cardiac EGM strip. In other examples, the segment may be a sub-segment of a cardiac EGM band. For example, in the example where the cardiac EGM band includes samples representing a 45 second cardiac rhythm of the patient 14, the segment may include samples representing the first 10 seconds of the cardiac EGM band.

The AI system 451 may provide overlapping sections of cardiac EGM strips to the deep learning model 454. For example, the AI system 451 may provide a section including samples representing the 0 th to 10 th seconds of the cardiac EGM strip, then provide a section including samples representing the 5 th to 15 th seconds of the cardiac EGM strip, then provide a section including samples representing the 10 th to 20 th seconds of the cardiac EGM strip, and so on. In some examples, computing system 24 may provide a segment spanning two or more cardiac EGM strips. For ease of illustration, applying a deep learning model to a cardiac EGM strip may actually refer to applying a deep learning model to a segment of a cardiac EGM strip.

In some examples, the deep learning model 454 includes a Convolutional Neural Network (CNN). For example, in one example, a convolutional layer may follow an input layer of the type described above. The first convolution layer neuron can receive input from a first set of input layer neurons consisting of a given number of consecutive input layer neurons; the second convolution layer neuron can receive input from a second set of input layer neurons consisting of the same given number of consecutive input layer neurons but offset by a step length from the first input layer neuron of the first set of input layer neurons; the third convolutional layer neuron may receive input from a third set of input layer neurons consisting of the same given number of consecutive input layer neurons but offset by a stride length from the first input layer neuron of the second set of input layer neurons; and so on. The given number of consecutive input neurons and stride length are different hyper-parameters of CNN. One or more fully connected hidden layers may follow the convolutional layer.

In some examples of the disclosure, for each respective arrhythmia in a set of one or more arrhythmias, deep learning model 454 may generate data indicating whether one or more occurrences of the respective arrhythmia is represented in a segment of a cardiac EGM strip. For example, in one instance, a hidden layer of the deep learning model 454 provides input data to an output layer of the deep learning model 454. For each respective arrhythmia in a set of arrhythmias, the output layer of the deep learning model 454 includes a separate output neuron corresponding to the respective arrhythmia. The output neurons corresponding to the respective arrhythmias receive input data from individual neurons in the hidden layer of the deep learning model 454, which also correspond to the respective arrhythmia type. The data generated by the hidden layer neurons corresponding to the respective arrhythmia includes probability values indicating the probability that the occurrence of the arrhythmia has occurred in the segment of the cardiac EGM strip. The activation function of the output neuron may apply a threshold function to the probability values generated by the hidden layer neurons. For each output neuron, the threshold function may cause the output neuron to generate a first value (e.g., 1) if the probability value provided to the output neuron is greater than a threshold value, and a second value (e.g., 0) if the probability value provided to the output neuron is less than the same threshold value.

Further, in the example of the previous paragraph, the AI system 451 may use the probability values generated by the hidden layer to track the location within the cardiac EGM strip where the occurrence of the arrhythmia occurred. For example, as described above, the cardiac EGM band may be subdivided into sections and the AI system 451 provides the sections as input to the deep learning model 454. Thus, by determining which segment of the cardiac EGM strip results in the highest probability value corresponding to an arrhythmia, the AI system 451 can determine which segment is most likely to represent the occurrence of an arrhythmia.

As noted elsewhere in this disclosure, the input provided to the deep learning model 454 may contain patient data in addition to the segment of the cardiac EGM band. For example, in some instances, the patient data may additionally include data regarding a physiological state of the patient (e.g., a patient physiological state, such as activity, posture, respiration, etc.) that may also be captured by the medical device 16. Patient data corresponding to different physiological conditions (e.g., rest, night high-attitude angle rest, etc.) may be used as additional parameters for model training or input data for deep learning model 454. Use of such data may enable the AI system 450 to detect the occurrence of arrhythmias during other disease conditions (e.g., sensitive models of tachycardia during rest may be used to monitor Heart Failure (HF) patients; models of bradycardia during activity may be used to monitor patients for chronotropic dysfunction). In some examples, the monitoring system 450 receives EMR data for the patient 14 from the EMR database 66 via the communication circuitry 406. In some examples, the EMR data stored by the EMR database 66 may contain many different types of historical medical information about the patient 14. For example, the EMR database 66 may store a medication history for the patient, a surgical history for the patient, a hospital history for the patient, potassium levels for the patient over time, or one or more laboratory test results for the patient, etc. The EMR data may form part of the patient data used as input to the deep learning model 454.

In some examples, the deep learning model 454 converts the patient data into one or more vectors and tensors (e.g., a multidimensional array) representing the patient data. The deep learning model 454 may apply mathematical operations to one or more vectors and tensors to generate a mathematical representation of the patient data. The deep learning model 454 may determine different weights corresponding to the identified relationship between the patient data and the occurrence of arrhythmia. The deep learning model 454 may apply different weights to the patient data to generate probability values.

Fig. 5 is a flow chart illustrating example operations in accordance with the techniques of this disclosure. For convenience, fig. 5 is described with respect to fig. 1 and 4. The flow diagrams of the present disclosure are presented as examples. Other examples in accordance with the techniques of this disclosure may include more, fewer, or different actions, or may perform the actions in a different order or in parallel. The operations of fig. 5 may be performed by an AI system implemented on one or more of the medical device 16, the computing system 24, the external device 27, and/or other devices.

In the example of fig. 5, the AI system may obtain one or more cardiac EGM strips for the patient 14 (i.e., the current patient) (500). The AI system may obtain one or more cardiac EGM bands for the patient 14 in one or more of a variety of ways. For example, in instances in which the computing system 24 implements an AI system, the computing system 24 may obtain one or more cardiac EGM strips of the patient 14 from the medical device 16 (e.g., via the external device 27 and the network 25). In some examples in which the AI system is implemented in medical device 16, the AI system may obtain one or more cardiac EGM bands by generating one or more cardiac EGM bands based on data from the electrodes. In some examples, the AI system may obtain the one or more cardiac EGM strips for the current patient from a database (e.g., the EMR database 66) that stores the one or more cardiac EGM strips for the current patient. Other examples of obtaining cardiac EGM bands are described elsewhere in this disclosure.

Further, in the example of fig. 5, the pre-processing unit 452 may pre-process one or more cardiac EGM strips (502). The pre-processing unit 452 may pre-process one or more cardiac EGM strips in one or more of various ways. For example, fig. 6, 7, 9, 11, and 15 illustrate example ways to preprocess one or more cardiac EGM strips. In some instances, the deep learning model 454 may be trained using cardiac EGM strips with certain characteristics, such as signal polarity, signal amplitude levels and variations, hardware characteristics, and so forth. In such examples, after training, preprocessing unit 452 may preprocess the one or more cardiac EGM bands such that the one or more cardiac EGM bands have characteristics of a training cardiac EGM band. This may allow the deep learning model 454 to be used with cardiac EGM strips generated by more types of medical devices. For example, as described in detail below, fig. 7 and 15 are examples in which the preprocessing unit 452 preprocesses one or more cardiac EGM strips to conform to one or more characteristics of cardiac EGM strips on which the deep learning model 454 is trained. Further, in some instances, the deep learning model 452 may be trained using preprocessed cardiac EGM bands generated by the preprocessing unit 452. For example, in the examples of fig. 6, 7, 9, and 15, the AI system may train the deep learning model 454 using the preprocessed cardiac EGM bands.

After preprocessing unit 452 preprocesses the one or more cardiac EGM strips, the AI system may apply deep learning model 454 to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias (504). That is, the input to the deep learning model 454 may contain one or more preprocessed cardiac EGM bands or segments thereof. The deep learning model 454 may be trained in accordance with any of the examples described elsewhere in this disclosure. When the AI system applies the deep learning model 454 to one or more preprocessed cardiac EGM strips, the AI system may perform feed-forward passes of neurons through a neural network of the deep learning model 454. The arrhythmia data may be an output of a neural network.

Further, in some examples, the AI system may generate output data based on the arrhythmia data. For example, the AI system may generate a table indicating the times and types of occurrences of detected arrhythmias. The time may be determined based on a timestamp provided by the medical device 16 with the cardiac EGM strip. In another example, the AI system may generate a map that shows the rate at which patient 14 experiences the occurrence of an arrhythmia over time. The AI system may present the output data to one or more types of users. For example, the AI system may present the output data to the patient 14, a healthcare provider of the patient 14, a user of a healthcare monitoring organization, or other type of person.

Fig. 6 is a flowchart illustrating a first example operation for preprocessing one or more cardiac EGM strips in accordance with one or more techniques of the present disclosure. In the example of fig. 6, pre-processing unit 452 may generate one or more temporally modified cardiac EGM bands by modifying a temporal resolution of the one or more cardiac EGM bands to match an expected temporal resolution of the deep learning model (602). For example, the pre-processing unit 452 may modify the temporal resolution of one or more cardiac EGM strips from 128 samples/sec to 200 samples/sec. The preprocessing unit 452 may modify one or more cardiac EGM strips in one of various ways. For example, in instances in which the expected temporal resolution of the deep learning model is greater than the temporal resolution of the one or more cardiac EGM strips, the pre-processing unit 452 may interpolate samples between samples of the one or more cardiac EGM strips. In instances in which the expected temporal resolution of the deep learning model is less than the temporal resolution of the one or more cardiac EGM strips, the pre-processing unit 452 may decimate samples in the one or more cardiac EGM strips. Other techniques are possible that modify the temporal resolution of one or more cardiac EGM strips.

Additionally, the pre-processing unit 452 may generate one or more pre-processed cardiac EGM strips by subtracting an average of the samples of the one or more time-modified cardiac EGM strips from the samples of the one or more time-modified cardiac EGM strips (604). For example, the pre-processing unit 452 may determine an average of the samples of the one or more time-modified cardiac EGM strips, and then subtract the average from each of the samples of the one or more time-modified cardiac EGM strips. In another example, the pre-processing unit 452 may subtract an average of samples of the one or more original cardiac EGM strips from the one or more original cardiac EGM strips and then modify a temporal resolution of the one or more resulting cardiac EGM strips, thereby reversing the order of acts 600 and 602.

Fig. 7 is a flow diagram illustrating second example operations for preprocessing cardiac EGM strips according to one or more techniques of the present disclosure. The medical device may be used for patients with different physiological conditions (e.g., fat/muscle mass) or potential heart disease. Thus, the signal levels in the cardiac EGM bands may be different. There may also be amplitude variations in the signal level in the cardiac EGM strip of the same patient, for example, due to postural changes or Premature Ventricular Contractions (PVCs). Because there may not be enough training data to train deep learning model 454 from scratch or for transfer learning in such conditions (e.g., a deep learning model developed using an external/resting holter monitor is applied to cardiac EGM strips from an implanted monitor of an active patient), applying additional normalization to the de-averaged waveform may improve the arrhythmia and QRS detection performance of the deep learning model.

Thus, in the example of fig. 7, the pre-processing unit 452 may generate one or more temporally modified cardiac EGM bands by modifying the temporal resolution of the one or more cardiac EGM bands to match the expected temporal resolution of the deep learning model (700). Additionally, the pre-processing unit 452 may generate one or more mean-subtracted cardiac EGM strips by subtracting a mean of the samples of the one or more time-modified cardiac EGM strips from the samples of the one or more time-modified cardiac EGM strips (702). The pre-processing unit 452 may perform acts 700 and 702 in the same manner as described with respect to acts 600 and 602 of fig. 6.

Further, in the example of fig. 7, to normalize varying amplitude levels, the pre-processing unit 452 may calculate a moving signal envelope over a sliding window of one or more cardiac EGM strips minus an average (704). The sliding window may be 0.5 seconds in duration, 1 second in duration, 1.5 seconds in duration, or have another duration. In some examples, the envelope may be the standard deviation of the waveform samples in a moving window or the 99 th and 1 st percentiles of the waveform samples in the moving window, or the maximum and minimum values of the waveform samples in the moving window.

Fig. 8A is a conceptual diagram illustrating example waveforms and envelopes determined according to one or more techniques of the present disclosure. Specifically, fig. 8A shows an example of a waveform (solid line) and envelope computed over a 1-second moving window with 99 th (upper dashed line) and 1 st (lower dashed line) percentile samples in the moving window.

Further, in the example of fig. 7, the pre-processing unit 452 may normalize 706 the cardiac EGM strip minus the average value based on the moving signal envelope. For example, in one example, let d indicate the waveform, dU indicates the upper envelope, and dL indicates the lower envelope (e.g., as shown in fig. 8A). In this example, the normalized waveform may be sampled as d/(dU-dL). Fig. 8B is a conceptual diagram illustrating an example normalized waveform generated according to one or more techniques of this disclosure. Note that in fig. 8B, the waveform QRS amplitude variation is minimized. In the example of fig. 7, the AI system may provide the normalized subtracted-average cardiac EGM strip as an input to the deep learning model 454.

In other examples, the pre-processing unit 452 may perform additional pre-processing actions. For example, the pre-processing unit 452 may normalize the entire waveform by its standard deviation, normalize the waveform by a constant gain factor, and/or perform other actions.

Fig. 9 is a flow diagram illustrating example operations of an AI system according to one or more techniques of this disclosure. The above example relates to a deep learning model applied only on cardiac EGM waveforms. Because some devices (e.g., medical device 16) have limited memory and battery power, the devices store and transmit limited waveform segments. However, such devices may also generate and store additional marker channel information. The marker channel information may correspond to a time period during and prior to a time period corresponding to the recorded cardiac EGM waveform. Marker channel information may provide additional arrhythmia diagnostics. For example, in one example, a marker channel may indicate a detected QRS complex. Thus, in this example, the marker channel may comprise a series of samples that each indicate whether a QRS complex is detected during a time period corresponding to the samples. From this data, the R-R rate of the heart rhythm of the patient 14 may be determined. The example of fig. 9 relates to arrhythmia detection using marker channels to validate and extend the deep learning model 454.

As shown in the example of fig. 9, the AI system uses marker channels to confirm and extend arrhythmia detection for the deep learning model 454. The waveform channel (e.g., cardiac EGM strip) and marker channel can be processed by two separate models: a.) waveform-based deep learning model 454 and B.) marker channel-based model. If the arrhythmias detected by the two models coincide during the duration of the waveform, the AI system increases the confidence of the arrhythmia detection and extends the time course of the arrhythmia detection just beyond the recorded waveform duration. In other words, the AI system may cause the device that generated the waveform channel and the marker channel to continue generating the waveform channel. However, if the two models do not agree, the AI system may adjust the confidence of arrhythmia detection (e.g., if the marker-based models have high reliability), but not extend the time course of arrhythmia detection. In other words, the AI system may cause the device that generates the waveform channel and the marker channel to stop generating the waveform channel. Not generating a waveform channel may reduce power consumption of the device.

In the example of fig. 9, the AI system may obtain one or more cardiac EGM bands and marker channels (900). In some examples, the marker channel indicates a QRS complex detected in the heart rhythm of patient 14. The pre-processing unit 452 may resample and normalize one or more cardiac EGM strips, e.g., as described with respect to fig. 7 (902). Additionally, the AI system may apply a deep learning model 454 (i.e., a waveform-based deep learning model) to the one or more preprocessed cardiac EGM strips to generate a first set of arrhythmia data (904). The AI system may apply the deep learning model 454 in the same manner as described with respect to fig. 5.

Further, in the example of fig. 9, the pre-processing unit 452 may pre-process the marker channel (906). For example, the pre-processing unit 452 may determine R-R intervals, average heart rate, one or more moving average heart rate trends, outlier removal, marker-based lorentz plots, heart rate variability, or other types of data based on the marker channels. Additionally, the preprocessing unit 452 may apply a marker channel-based deep learning model to the preprocessed marker channels to generate second arrhythmia data indicative of a second set of occurrences of the one or more arrhythmias (908). In this example, the marker channel-based deep learning model may be implemented as a neural network trained to identify the occurrence of arrhythmias in the preprocessed marker channels.

In the example of fig. 9, the AI system may determine whether the first set of arrhythmia data and the second set of arrhythmia data are consistent (910). For example, the AI system may check whether there are any detected arrhythmia events that are not in the combination of the first and second sets of arrhythmia data. Further, the AI system may adjust a confidence level of an occurrence of one of the arrhythmias based on whether the occurrence of the arrhythmia is in both the first set of occurrences and the second set of occurrences or only one of the first set of occurrences and the second set of occurrences. For example, if the first set of arrhythmia data and the second set of arrhythmia data are consistent for a detected arrhythmia occurrence ("yes" branch of 910), the AI system may increase a confidence level of the detected arrhythmia occurrence and extend arrhythmia detection (912). For example, the AI system may cause the medical device 16 to continue generating cardiac EGM strips. On the other hand, if the first set of arrhythmia data and the second set of arrhythmia data are inconsistent with respect to the detected arrhythmia occurrence (the "NO" branch of 910), the AI system may adjust a confidence level of the detected arrhythmia occurrence and not extend arrhythmia detection (914). For example, the AI system may cause the medical device 16 to stop generating cardiac EGM strips. Thus, the AI system may perform one or more of the following: adjusting a confidence level of an occurrence of one of the arrhythmias based on whether the occurrence of the arrhythmia is in both the first set of occurrences and the second set of occurrences or only one of the first set of occurrences and the second set of occurrences, the duration of a monitoring session of the medical device generating one or more cardiac EGM strips based on whether the occurrence of the arrhythmia is in the first set of occurrences and the second set of occurrences.

The confidence level may be used in various ways. For example, the monitoring system 450 may output a confidence level for display. In this example, displaying the confidence level may help a monitoring professional or healthcare professional determine how to take action on information about the arrhythmia. In some examples, if the confidence level is below a predetermined threshold, monitoring system 450 does not present information about the arrhythmia to the user.

Further, in some examples, the medical device 16 generates a cardiac EGM strip of limited duration (e.g., 45 seconds) upon detecting the occurrence of an arrhythmia or patient trigger. The diagnostic device has a patient trigger (e.g., a button on the device) that the patient can depress when the symptoms are felt. These are considered symptomatic ECG episodes. Around the time of patient activation/triggering, the device will capture the expanded waveform signal that is transmitted for viewing. Additionally, in such examples, the medical device 16 generates the marker channel for a more extended duration of time (e.g., 5 minutes before and 5 minutes after the occurrence of the arrhythmia or patient trigger is detected). Consistent with the example of fig. 9, the AI system may pre-process cardiac EGM strips (902) and apply a deep learning model 454 to the pre-processed cardiac EGM strips to generate arrhythmia data (904). Further, in such instances, the AI system may apply a marker channel-based model to a portion of the marker channel markers corresponding to times before and during the time corresponding to the cardiac EGM band (908). In such instances, the AI system may determine whether the outputs of the waveform-based deep learning model 454 and the marker channel-based deep learning model are consistent. For example, if the deep learning model 454 determines that the waveform of the cardiac EGM strip contains the occurrence of ventricular tachycardia and the marker-based deep learning model detects a high rate of episodes with low RR variability, the AI system may determine that the output of the waveform-based deep learning model 454 is consistent with the output of the marker channel-based deep learning model.

Based on whether the outputs of the two models are consistent, the AI system can adjust the likelihood of global arrhythmia occurrence in the waveform. For example, if there is correspondence between the marker channel-based model and the waveform-based deep learning model 454 (e.g., the marker channel detects greater than or equal to 120 times per minute (BPM) of tachyarrhythmia and the waveform-based deep learning model 45 detects Ventricular Tachycardia (VT) for the duration of the recorded waveform), the monitoring system 450 can use the trends of the marker channel to present information about what occurred before the onset of the arrhythmia detected in the cardiac EGM band. For example, monitoring system 450 may output a graph showing the performance of the patient's heart prior to the onset of an arrhythmia, such as the graph shown in fig. 10B. This may help the physician diagnose the heart condition of the patient 14.

In one example, the medical device 16 may generate one or more cardiac EGM strips representing a waveform of finite duration in response to an on-board detection algorithm that determines that an arrhythmia is occurring. The medical device 16 itself may implement an on-board detection algorithm. In one example, an on-board detection algorithm may determine that a 45 second tachycardia episode occurred from 10:30:00 am to 10:30:45 am. In this example, the medical device 16 may provide one or more cardiac EGM strips representing a 45 second waveform to the AI system to report an arrhythmia. Further, in this example, the AI system may apply the waveform-based deep learning model 454 to determine whether one or more cardiac EGM strips contain any arrhythmias during a 45 second time period (i.e., from 10:30:00 am to 10:30:45 am). Monitoring system 450 may report any arrhythmia detected by deep learning model 454 based on one or more cardiac EGM strips. Additionally, in some instances, the medical device 16 may generate marker channels covering a longer duration (e.g., from 10:20:00 a.m. to 10:35:00 a.m.). The medical device 16 may also provide this marker channel to the AI system for arrhythmia reporting. The marker channel-based model of the AI system may use the marker channels to detect arrhythmias during and adjacent to a time period corresponding to one or more cardiac EGM strips (e.g., 10:29:00 am to 10:31:00 am). If the arrhythmia detected by the marker channel-based model is consistent with the arrhythmia detected by the waveform-based deep learning model 454 during and in a time period adjacent to the time period corresponding to the one or more cardiac EGM strips, the monitoring system 450 may output an indication of the arrhythmia detected by the waveform-based deep learning model 454 based on the one or more cardiac EGM strips over a time period of 10:30:00 am to 10:30:45 am. Additionally, the monitoring system 450 may output additional information from the marker channel (from 10:20:00 am to 10:35:00 am), such as average HR and HR variability trends before, during, and after the recorded waveform duration. In this manner, monitoring system 450 may output information that provides additional arrhythmia-related information to the physician. For example, a physician may be able to see if there is a sudden onset of tachycardia.

In some examples, if the medical device 16 implements both the waveform-based deep learning model 454 and the marker channel-based model, the medical device 16 may determine whether there is a correspondence between arrhythmias identified by the waveform-based deep learning model 454 and the marker channel-based model over a period of time, and the medical device 16 may keep recording waveforms for a longer duration. If the medical device 16 determines that there is no correspondence, the medical device 16 may stop recording/deleting this waveform segment.

Fig. 10A is a conceptual diagram illustrating a graph of an example cardiac waveform and QRS markers detected by the apparatus. Fig. 10B is a conceptual diagram illustrating a graph of heart rate over time based on a marker channel. Fig. 10A and 10B are for the same data and are time aligned on the x-axis. In the example of fig. 10A, the marker channel information is shown as a series of dots arranged on a horizontal line. The dark jagged lines starting around 95 th second are the waveforms represented by one or more cardiac EGM bands. In fig. 10A, the markers are stored on the medical device 16 from a time period prior to the limited waveform segment. Generating and storing marker data may require less computational resources than cardiac EGM strips. As shown in the example of fig. 13B, the waveform-based and marker channel-based deep learning models agree that detection is of Ventricular Tachycardia (VT) occurring at about 130 second marker, and the time course of arrhythmia detection extends from a 120-130 second-only segment (within the waveform duration) to about 2 minutes prior to the onset of arrhythmia; this indicates a rapid increase in HR and VT events.

Although other examples described in this disclosure use deep learning model 452 directly on the preprocessed cardiac EGM strip, in other examples, the AI system may perform deep learning QRS and arrhythmia detection on the transformed signal (e.g., in the frequency domain). For example, in one example, a neural network may have several initial convolutional layers to extract low-level features of an input signal. In some instances, the AI system may learn the convolution kernel/filter parameters directly from the data. In some instances, to reduce computational complexity while maintaining performance, the deeply learned data may be pre-processed using existing expert knowledge. For example, as part of preprocessing one or more cardiac EGM strips, the AI system may decompose a waveform identified by the one or more cardiac EGM strips into decomposed waveforms corresponding to different frequency bands. Different frequency bands may correspond to different aspects of the ECG signal, such as a high frequency QRS signal, a low frequency P-wave signal, etc. In this example, the AI system may generate arrhythmia data by applying the deep learning model 452 to the decomposed waveform.

Fig. 11 is a flow diagram illustrating a third example operation for preprocessing cardiac EGM strips according to one or more techniques of the present disclosure. In the example of fig. 11, as part of preprocessing the one or more cardiac EGM strips, the preprocessing unit 452 may scale the waveforms represented by the one or more cardiac EGM strips (1100). Scaling in this example, as well as other portions of this disclosure, may be used to limit the range of EGM amplitude scaling over which the model needs to be trained. Additionally, in the example of fig. 11, the pre-processing unit 452 may decompose the waveform represented by one or more cardiac EGM bands into multiple channels corresponding to different frequency bands (1002). Subsequently, as part of applying the deep learning model 454 to the segment of the preprocessed cardiac EGM strip, the AI system may apply the deep learning model to the channel to generate arrhythmia data.

Thus, in the example of fig. 11, after waveform scaling, the signal is decomposed into multiple channels, where each channel corresponds to a different frequency band. Fig. 12 is a conceptual diagram illustrating an example original cardiac waveform and waveforms of channels corresponding to different frequency bands according to one or more techniques of the present disclosure. Fig. 12 shows an example in which instead of the original waveform 1200 for deep learning, 3 channels (1202, 1204, and 1206) using (i) stationary wavelet decomposition and (ii) band specific time delays for QRS alignment are used that originate from the original waveform 1200. It should be noted that channel 1202 is primarily composed of high frequency features (generally corresponding to QRS segments) and frequency band 1206 is composed of low frequency features (e.g., p-waves). Thus, instead of having the 1x N vector as input to the deep learning model 454, a 3x N matrix (with a stationary wavelet decomposition and alignment signals) is used as input. Having preprocessed signals may help speed up learning and/or reduce the complexity of the deep learning model 454.

Fig. 13 is a flow diagram illustrating a fourth example operation for preprocessing cardiac EGM strips according to one or more techniques of the present disclosure. In the example of fig. 13, as part of preprocessing the one or more cardiac EGM strips, the preprocessing unit 452 may scale the waveforms represented by the one or more cardiac EGM strips (1300). The pre-processing unit 452 may scale the waveform according to any of the examples provided above with respect to fig. 11.

Further, in the example of fig. 13, the pre-processing unit 452 may generate one or more transformed cardiac EGM strips by transforming the scaled waveforms into a two-dimensional time-frequency domain (1302). For example, the pre-processing unit 452 may transform the scaled waveform into a spectrogram. Subsequently, the AI system may apply the deep learning model 454 to the one or more preprocessed cardiac EGM strips by applying the deep learning model to the one or more transformed cardiac EGM strips. The monitoring system 450 may output an image of the spectrogram for display. Further, in some instances, the deep learning model 454 may be implemented in a manner similar to the image recognition deep learning model. For example, the deep learning model 454 may include convolutional layers that apply individual filters to the spectrogram. Further, in some instances, by presenting the entire spectrogram to the deep learning model 454, the AI system may enable the deep learning model 454 to determine an optimal set of frequencies and features for this set of signals.

Fig. 14 is a conceptual diagram illustrating example operation of an AI system according to one or more techniques of this disclosure. In the example of fig. 14, the pre-processing unit 452 may pre-process cardiac EGM strips (1400). For example, in the example of fig. 14, the pre-processing unit 452 may apply the learned scaling factor to the waveform represented by the cardiac EGM strip. In some examples, the AI system may learn the scaling factors by repeatedly testing different scaling factors with training data and determining which scaling factor results have the best performance in correctly detecting arrhythmias.

The AI system may provide the preprocessed cardiac EGM strips as input to the deep neural network 1402. The deep neural network 1402 may be part of a deep learning model 454.

In the example of fig. 14, as part of applying the deep learning model 454, the AI system may perform QRS detection (1402). As part of performing QRS detection, deep neural network 1402 may determine QRS probability values corresponding to multiple points in time. Each of the QRS probability values indicates a respective probability that a peak of the QRS complex occurs during a point in time corresponding to the QRS probability value. For example, a set of initial layers of the deep neural network may determine QRS probability values based on the preprocessed EGM bands.

Additionally, the deep neural network 1402 may perform arrhythmia classification based on the preprocessed cardiac EGM bands and QRS probability values (1406). That is, the deep neural network 1402 may detect a set of occurrences of one or more arrhythmias by providing QRS probability values and preprocessed cardiac EGM strips as inputs to a neural network that generates arrhythmia data. Thus, in this step, QRS probability values are combined with the deeper layers of the deep learning model 454 to detect arrhythmic events.

As shown in the example of fig. 14, the AI system may output QRS probability values and occurrence times, respectively, and arrhythmia data indicating whether and where cardiac EGM strips represent one or more occurrences of one or more arrhythmias. This may provide the advantage of providing two separate paths for QRS, and arrhythmia detection is interpretable, as QRS is typically used by ECG readers/technicians trained in the art as the first step in arrhythmia determination. Furthermore, 2-way interaction between the QRS and arrhythmia detection module may help provide arrhythmia-specific QRS detection and enhanced QRS detection based on the detected occurrence of arrhythmia (e.g., if the QRS model flags PVC/bigeminy but not which, and if the arrhythmia model claims its PVC, then the rhythm is bigeminy, otherwise it is T-wave oversensing).

Fig. 15 is a flow diagram illustrating fifth example operations for preprocessing cardiac EGM strips according to one or more techniques of the present disclosure. In the example of fig. 15, the operations include signal polarity and signal characteristic pre-processing in accordance with one or more techniques of this disclosure. In some examples, deep learning model 454 is trained with data from some type of hardware and cardiac EGM signal characteristics. For example, a user may want to train the deep learning model 454 using a large data set. The data set may contain a waveform shape that is "upright," but the deep learning model 454 may also need to be used with devices where the waveform shape is not always upright. For example, some devices generate a cardiac EGM signal in which the R-wave is initially deflected in a negative direction, and other devices may generate a cardiac EGM signal in which the R-wave is initially deflected in a positive direction. To address such situations, the deep learning model 454 may be trained on a set of original waveforms and polarity-reversed versions thereof, both of which have the same arrhythmia content. In cases where the AI system needs to use a pre-existing deep learning model (i.e., where retraining is not possible), the waveform signal can be transformed to meet the input characteristics of deep learning.

Thus, in the example of fig. 15, the AI system may determine the polarity of the cardiac EGM band (1500). The AI system may then determine whether the polarity of the cardiac EGM band is the same as the expected polarity of the deep learning model 454 (1502). In response to determining that the polarity of the cardiac EGM band is not the expected polarity of the deep learning model 454 ("no" of 1502), the AI system can reverse the polarity of the cardiac EGM band (1504). Otherwise, the AI system does not invert the polarity of the cardiac EGM band (1506).

The AI system may determine the polarity of the cardiac EGM band in one of various ways. For example, in one example, when the medical device 16 is implanted in the patient 14, the implanting physician may program the settings to indicate whether the polarity is reversed. In this example, the medical device 16 may include data in the cardiac EGM strip generated by the medical device 16 that is indicative of the polarity of the cardiac EGM strip generated by the medical device 16.

In another example, when viewing and analyzing data from the medical device 16, such as cardiac EGM strips, at the monitoring center, the monitoring center technician may flag the data if the waveform morphology of the patient 14 is inverted. For short-term monitoring, such labeling may only need to be done at the beginning of the monitoring. For long-term applications, such morphological marking may be done periodically (e.g., every month) to account for any device drift.

In another example of determining the polarity of the cardiac EGM band, the AI system may use the P-wave and T-wave morphology to estimate whether the morphology is inverted. That is, the P-wave and T-wave start and end cardiac cycles always deflect in the same direction. Thus, based on the initial detection of the P-wave and T-wave, the AI system may determine the polarity of the cardiac EGM band.

In another example of determining the polarity of a cardiac EGM strip, the AI system may use a deep learning model to detect waveform polarity as a precursor for arrhythmia detection by applying a deep learning model (e.g., deep learning model 454). Thus, in this example, the deep learning model 454 may be considered a first deep learning model and the AI system may apply additional deep learning models to the cardiac EGM strip to determine the polarity of the cardiac EGM strip. Additional deep learning models may include artificial neural networks trained to classify the polarity of cardiac EGM strips.

In a related example, the AI system may use a deep-learning similarity model to check whether the input signal morphology is similar to the arrhythmia detection model and the signal morphology needed to perform the appropriate signal transformation. The deep learning similarity model may take a plurality of waveforms as inputs and may generate output data indicating whether the waveforms are similar. The waveform in this example may be the waveform represented by a cardiac EGM strip. In this example, "similarity" herein may refer to the same morphology/polarity. In other words, the deep-learning similarity model may compare waveforms from cardiac EGM strips of multiple devices to determine whether the waveforms have similar morphology. The deep-learning similarity model may be implemented as a neural network.

In addition to or instead of the polarity of the cardiac EGM band, the AI system may modify the properties of the cardiac EGM band. For example, the AI system may generate device classification data indicative of the category of the device that generated the cardiac EGM strip. Different classes of devices may have different hardware characteristics (e.g., bandwidth of the input signal). Accordingly, it may be advantageous to re-filter and/or transform the signals of cardiac EGM bands to match the input characteristics of the deep learning model 454 before applying the deep learning model to the cardiac EGM bands. For example, the AI system may filter the signal of the cardiac EGM strip to change the bandwidth of the signal. For example, in this example, the AI system may scale the signal samples of the cardiac EGM strip such that the samples are distributed within the expected sample value range of the deep learning model 454. In some examples, as part of filtering the signals of the cardiac EGM band to change the bandwidth of the signals, the AI system may increase or decrease the sampling rate of the signals of the cardiac EGM band to match the expected sampling rate of the deep learning model 454. The AI system may use interpolation to adjust the sampling rate. Due to the different characteristics of different classes of devices, the deep learning model 454 may not be able to correctly identify the occurrence of arrhythmias in cardiac EGM strips generated by multiple types of devices.

Fig. 16 is a conceptual diagram illustrating example operation of an AI system including an autoencoder according to one or more techniques of this disclosure. In the example of fig. 16, the AI system uses automatic encoding to determine device classification. In other words, the AI system can use automatic encoding to check whether the classification data assigned to the cardiac EGM strip is correct. The classification data assigned to the cardiac EGM strip may include data generated by another device (e.g., medical device 16) indicative of arrhythmias detected in the cardiac EGM strip.

As shown in the example of fig. 16, the input to the AI system may include one or more preprocessed cardiac EGM strips 1600 and classification data 1602. The preprocessed cardiac EGM strip 1600 represents the waveform of the cardiac rhythm of the patient 14. The pretreated cardiac EGM strip 1600 may be pretreated in one or more of a variety of ways, including any of the examples provided elsewhere in accordance with this disclosure. Classification data 1602 may include data indicative of arrhythmia classes detected in cardiac EGM strips, based on which preprocessed cardiac EGM strip 1600 is based. For example, classification data 1602 may indicate an occurrence of cardiac EGM strips determined by medical device 16 to contain atrial fibrillation.

The self-encoder 1604 classifies the probability that the classified data 1602 is correct. The autoencoder 1604 may be implemented as a deep neural network. The deep neural network of the self-encoder 1604 includes an input layer 1604, a set of hidden layers, and an output layer. The self-encoder 1604 is trained such that when the preprocessed cardiac EGM strips and the set of classification data are provided as inputs to the self-encoder 1604, the output layer outputs a reconstructed version of the preprocessed cardiac EGM strips. Self-encoder 1604 may be trained according to one of various techniques known in the art for training self-encoders. For example, the AI system may compare the reconstructed version of the preprocessed cardiac EGM strip with the original preprocessed cardiac EGM strip 1600 to determine error values that may be used in a back propagation algorithm to update parameters of the deep neural network of the self-encoder 1604. As shown in the example of fig. 16, the AI system may provide one or more of the preprocessed cardiac EGM strip 1600, or segments thereof, and the device classification data 1602 to the input layer 1606 of the deep neural network from the encoder 1604.

The output from one of the hidden layers of the deep neural network of the encoder 1604 may generate a probability value 1606. Each of the probability values corresponds to a different category of arrhythmia and indicates a level of confidence that the classification data 1602 correctly identifies the arrhythmia category in the cardiac EGM strip. In some examples, to speed up training of the deep neural network from the encoder 1604, layers up to intermediate layers are pre-trained separately from subsequent layers of the deep neural network from the encoder 1604.

Further, in the example of fig. 16, the AI system may use probability values 1606 to determine whether classification data 1602 correctly identifies an arrhythmia in the cardiac EGM strip. For example, the AI system may identify the highest of the probability values 1606 and compare the arrhythmia corresponding to the highest of the probability values 1606 with the arrhythmia indicated by the classification data 1602 to determine whether the classification data 1602 indicates the arrhythmia corresponding to the highest of the probability values 1606.

Fig. 17 is a flow diagram illustrating example operations in which an autoencoder is used to confirm device classification in accordance with one or more techniques of the present disclosure. As shown in the example of fig. 17 and consistent with the example of fig. 16, the AI system may obtain a training input vector (1700). Each of the training input vectors includes a segment of a training cardiac EGM band and device classification data indicative of one or more arrhythmias detected in the training cardiac EGM band. The AI system may train the self-encoder 1604 based on the training input vector to reconstruct a training cardiac EGM band of the training input vector (1702).

Further, in the example of fig. 17, the AI system may obtain additional classification data (1704). For example, in one example, the AI system may receive additional classification data from a medical device, such as medical device 16.

The AI system may provide the segment and classification data of the preprocessed cardiac EGM strip to the input layer 1606 of the self-encoder 1604 (1706). Further, the AI system can determine whether the classification data correctly identifies an arrhythmia (if any) in the preprocessed cardiac EGM strip based on the probability values generated by the middle layer from the encoder 1604 (1708). Each of the probability values corresponds to a different arrhythmia and indicates a confidence level that the preprocessed cardiac EGM strips contain an arrhythmia.

The following is a set of non-limiting examples of one or more techniques in accordance with the present disclosure.

Example 1. a method, comprising: obtaining, by a computing system, one or more cardiac Electrogram (EGM) strips representing a waveform of a cardiac rhythm of a patient; preprocessing, by the computing system, the one or more cardiac EGM strips; and applying, by the computing system, a deep learning model to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

Example 2. the method of example 1, wherein preprocessing the one or more cardiac EGM strips comprises: generating, by the computing system, one or more temporally-modified cardiac EGM strips by modifying a temporal resolution of the one or more cardiac EGM strips to match an expected temporal resolution of a deep-learning model; and generating, by the computing system, one or more preprocessed cardiac EGM bands by subtracting an average of the samples of the one or more temporal modified cardiac EGM bands from the samples of the one or more temporal modified cardiac EGM bands.

Example 3. the method of example 1, wherein preprocessing the one or more cardiac EGM strips comprises:

generating, by the computing system, one or more temporally-modified cardiac EGM strips by modifying a temporal resolution of the one or more cardiac EGM strips to match an expected temporal resolution of the deep-learning model; generating, by the computing system, one or more time-modified cardiac EGM strips by subtracting an average of the samples of the one or more time-modified cardiac EGM strips from the samples of the one or more time-modified cardiac EGM strips; calculating, by the computing system, a moving signal envelope over a sliding window of the one or more mean-subtracted cardiac EGM strips; and normalizing, by the computing system, the one or more mean-subtracted cardiac EGM strips based on the moving signal envelope.

Example 4. the method of any one of examples 1-3, wherein: the arrhythmia data is first arrhythmia data; the method further comprises: obtaining, by the computing system, a marker channel; pre-processing, by the computing system, the marker channel; applying, by the computing system, a marker channel-based deep learning model to the preprocessed marker channels to generate second cardiac arrhythmia data indicative of a second set of occurrences of the one or more cardiac arrhythmias; determining, by the computing system, whether the first cardiac arrhythmia data is consistent with the second cardiac arrhythmia data; and at least one of: adjusting, by the computing system, a confidence level of an occurrence of one of the arrhythmias based on whether the occurrence of the arrhythmia is in both the first set of occurrences and the second set of occurrences or only one of the first set of occurrences and the second set of occurrences; or extending, by the computing system, a duration of a monitoring session of a medical device generating the one or more cardiac EGM strips based on the occurrence of the arrhythmia being in both the first set of occurrences and the second set of occurrences.

Example 5. the method of example 4, wherein the marker channel indicates a detected QRS complex.

Example 6. the method of any one of examples 1-5, wherein: preprocessing the one or more cardiac EGM strips comprises: scaling, by the computing system, a waveform represented by the one or more cardiac EGM strips; and decomposing, by the computing system, a waveform represented by the one or more cardiac EGM strips into a plurality of channels corresponding to different frequency bands, and applying the deep learning model to the one or more preprocessed cardiac EGM strips comprises applying, by the computing system, the deep learning model to the channels to generate the arrhythmia data.

Example 7. the method of any one of examples 1-6, wherein: preprocessing the one or more cardiac EGM strips comprises: scaling, by the computing system, a waveform represented by the one or more cardiac EGM strips; generating, by the computing system, a transformed waveform by transforming the scaled waveform into a two-dimensional time-frequency domain, and applying the deep learning model to the one or more preprocessed cardiac EGM bands includes applying the deep learning model to the transformed waveform.

Example 8. the method of any one of examples 1-7, wherein: preprocessing the one or more cardiac EGM strips includes applying, by the computing system, learned scaling factors to waveforms represented by the one or more cardiac EGM strips, and applying the deep learning model includes: determining, by the computing system, QRS probability values corresponding to the plurality of time points, each of the QRS probability values indicating a respective probability that a peak of a QRS complex occurs during the time point corresponding to the QRS probability value; and detecting the set of occurrences of the one or more arrhythmias detected by the computing system by providing the QRS probability value and the one or more preprocessed cardiac EGM strips as inputs to a neural network that generates the arrhythmia data.

Example 9. the method of any of examples 1-8, wherein preprocessing the one or more cardiac EGM strips comprises: determining, by the computing system, a polarity of the one or more cardiac EGM strips; and reversing, by the computing system, the polarity of the one or more cardiac EGM strips based on the polarity of the one or more cardiac EGM strips not being an expected polarity of the deep learning model.

Example 10. the method of example 9, wherein: the deep learning model is a first deep learning model, and determining the polarity of the one or more cardiac EGM strips comprises applying, by the computing system, a second deep learning model to the one or more cardiac EGM strips to determine the polarity of the one or more cardiac EGM strips.

Example 11 the method of any one of examples 1-10, wherein the method further comprises: obtaining, by the computing system, training input vectors, wherein each of the training input vectors comprises a segment of a training cardiac EGM band and device classification data indicative of one or more arrhythmias detected in the training cardiac EGM band; training, by the computing system, a self-encoder based on the training input vector to reconstruct training cardiac EGM bands of the training input vector; obtaining, by the computing system, additional device classification data; providing, by the computing system, the one or more preprocessed cardiac EGM strips and the additional device classification data to an input layer of the self-encoder; and determining, by the computing system, whether the classification data correctly identifies an arrhythmia, if any, in the preprocessed cardiac EGM strips based on probability values generated by a middle layer of the self-encoder, wherein each of the probability values corresponds to a different arrhythmia and indicates a confidence level that the one or more preprocessed cardiac EGM strips contain the arrhythmia.

Example 12. the method of any of examples 1-11, wherein preprocessing the one or more cardiac EGM strips comprises one or more of: scaling samples of a signal of the one or more cardiac EGM strips such that the samples of the signal of the one or more cardiac EGM strips are distributed within an expected range of sample values of the deep learning model, or increasing or decreasing a sampling rate of the signal of the one or more cardiac EGM strips to match an expected sampling rate of the deep learning model.

An example 13, a computing system, comprising: a storage device configured to store one or more cardiac Electrogram (EGM) strips of waveforms representative of a cardiac rhythm of a patient; one or more processing circuits configured to: pre-processing the one or more cardiac EGM strips; and applying a deep learning model to the one or more preprocessed cardiac EGM strips to generate arrhythmia data indicating whether the one or more cardiac EGM strips represent one or more occurrences of one or more arrhythmias.

Example 14. the computing system of example 13, further configured to perform the method of any of examples 2-12.

Example 15 a computer-readable storage medium having instructions stored thereon that, when executed, cause a computing system to perform the method of any of examples 1-12.

Example 16. a method as described in the specification.

In some examples, the techniques of this disclosure include a system comprising means for performing any of the methods described herein. In some examples, the techniques of this disclosure include a computer-readable medium comprising instructions that cause processing circuitry to perform any of the methods described herein.

It should be understood that the various aspects and examples disclosed herein may be combined in different combinations than those specifically set forth in the description and drawings. It will also be understood that the acts or events of any process or method described herein can be performed in a different order, may be added, merged, or eliminated entirely, according to examples (e.g., all described acts and events may not be necessary for the performance of the techniques). Additionally, although certain aspects of the disclosure are described as being performed by a single module, unit or circuit for clarity, it should be understood that the techniques of this disclosure may be performed by a combination of units, modules or circuits related to, for example, a medical device.

In one or more examples, the described techniques may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored on a computer-readable medium in the form of one or more instructions or code and may be executed by a hardware-based processing unit. The computer-readable medium may include a non-transitory computer-readable medium corresponding to a tangible medium such as a data storage medium (e.g., RAM, ROM, EEPROM, flash memory, or any other medium that can be used to store desired program code in the form of instructions or data structures and that can be accessed by a computer).

The instructions may be executed by one or more processors, such as one or more Digital Signal Processors (DSPs), general purpose microprocessors, Application Specific Integrated Circuits (ASICs), field programmable logic arrays (FPGAs), or other equivalent integrated or discrete logic circuitry. Thus, the term "processor" or "processing circuitry" as used herein may refer to any of the foregoing structures or any other physical structure suitable for implementing the described techniques. Furthermore, the techniques may be fully implemented in one or more circuits or logic elements.

Various examples have been described. These and other examples are within the scope of the following claims.