A reduced-power machine learning system for arrhythmia detection

Abstract

Techniques are disclosed that use characterization to reduce the impact of machine learning cardiac arrhythmia detection on the power consumption of a medical device. In one example, a medical device performs feature-based characterization of sensed electrocardiogram data from a patient to obtain cardiac features indicative of an arrhythmia episode in the patient. The medical device determines whether the cardiac features meet threshold criteria for application of a machine learning model to validate the feature-based characterization of the electrocardiogram data. In response to determining that the cardiac features meet the threshold criteria, the medical device applies the machine learning model to the sensed electrocardiogram data to verify that an arrhythmia episode has occurred or to determine a classification of the arrhythmia episode.

Description

[Technical field]

[0001] FIELD OF THE DISCLOSURE This disclosure relates generally to medical devices, and more particularly, to medical devices configured to detect arrhythmias. [Background technology]

[0002] Malignant tachyarrhythmias, such as ventricular fibrillation, are uncoordinated contractions of the myocardium of the heart's ventricles and are the arrhythmia most commonly identified in patients with cardiac arrest. If this arrhythmia continues for more than a few seconds, it can lead to cardiogenic shock and cessation of effective blood circulation. As a result, sudden cardiac death (SCD) can occur in a matter of minutes.

[0003] In patients at high risk for ventricular fibrillation, the use of implantable medical devices (IMDs), such as implantable cardioverter defibrillators (ICDs), has been found to be beneficial in preventing SCD. ICDs are battery-powered electrical shock devices that may include an electrical housing electrode (sometimes referred to as a can electrode) and are typically coupled to one or more electrical leads placed within the heart. For example, if an arrhythmia is detected, the ICD can transmit a pulse using the electrical leads to shock the heart and restore normal rhythm. Some ICDs are configured to terminate the detected tachyarrhythmia by delivery of anti-tachycardia pacing (ATP) prior to delivery of the shock. Additionally, ICDs are configured to deliver relatively high amplitude post-shock pacing after successful termination of the shockable tachyarrhythmia to support the heart as it recovers from the shock. Some ICDs also deliver bradycardia pacing, cardiac resynchronization therapy (CRT), or other forms of pacing.

[0004] Other types of medical devices may be used for diagnostic purposes. For example, an implantable or non-implantable medical device may monitor a patient's heart. A user, such as a physician, may review data generated by the medical device for occurrences of cardiac arrhythmias, such as atrial or ventricular tachyarrhythmias, or asystole. The user may diagnose a medical condition of the patient based on the occurrence of identified cardiac arrhythmias. Summary of the Invention

[0005] In accordance with the techniques of the present disclosure, a medical device system is described herein that uses characterization and machine learning to perform cardiac arrhythmia detection and classification. Specifically, the medical device system described herein can use the characterization to perform preliminary detection of cardiac arrhythmia in a patient and use only the machine learning model to validate cardiac arrhythmia episodes detected by the characterization or to classify such episodes detected by the characterization as being of a particular type of cardiac arrhythmia.

[0006] As described herein, characterization refers to the use of features obtained through signal processing for use in detecting or classifying episodic cardiac arrhythmias. Typically, characterization involves the use of designed rules to identify or extract features of electrocardiogram data, measure properties of such features, and use the measurements to detect or classify arrhythmias. For example, characterization may be used to identify features such as R waves, QRS complexes, P waves, T waves, the rate of such features, the intervals between such features, feature morphology, width or amplitude of such features, or other or other types of cardiac features or properties of features not explicitly described herein. Characterization may include feature extraction, signal filtering, peak detection, refractory analysis, or other types of signal processing, feature engineering, or detection rule development. Characterization algorithms can be optimized for real-time, embedded, and low-power applications, such as for use in implantable medical devices. However, characterization algorithms may require expert design and feature engineering to accurately detect arrhythmias in patients.

[0007] In contrast to characterization techniques for cardiac arrhythmia detection and classification, machine learning techniques may be used for cardiac arrhythmia detection and classification. As described herein, machine learning refers to the use of a machine learning model, such as a neural network or a deep learning model, which is trained on a training data set to detect cardiac arrhythmia from electrocardiogram data. Machine learning techniques contrast with characterization in that characterization relies on signal processing, and a machine learning system may "learn" the underlying features present in electrocardiogram data that indicate arrhythmia episodes without requiring knowledge or understanding of the relationship between the features and arrhythmia episodes on behalf of the system designer.

[0008] In one example, a medical device, such as an IMD, senses electrocardiogram data of a patient. The medical device performs feature-based representation of the electrocardiogram data to obtain cardiac features indicative of an arrhythmia episode in the patient. The medical device determines whether the cardiac features meet threshold criteria for application of a machine learning model to validate the feature-based representation of the electrocardiogram data. If the medical device determines that the cardiac features meet the threshold criteria for applying the machine learning model, then the medical device applies the machine learning model to the sensed electrocardiogram data to, for example, confirm that an arrhythmia episode has occurred in the patient or to detect one or more other types of arrhythmias that have occurred in the patient.

[0009] In another example, the medical device compares a first cardiac feature of the electrocardiogram data with a cardiac feature defined by an entry in the arrhythmia dictionary. In response to determining that the first cardiac feature of the electrocardiogram data is not similar to the cardiac feature defined by the entry in the arrhythmia dictionary, the medical device applies a machine learning model to determine a classification of an arrhythmia episode revealed by the first cardiac feature. The medical device may store the determined arrhythmia classification and the cardiac feature as a new entry in the arrhythmia dictionary to build the arrhythmia dictionary. Subsequently, upon detecting a second cardiac feature similar to the first cardiac feature via characterization, the medical device determines that the second cardiac feature is indicative of an arrhythmia episode of the same classification as the arrhythmia episode indicated by the first cardiac feature.

[0010] The techniques of the present disclosure may provide certain improvements in the field of cardiac arrhythmia detection and classification by medical devices. For example, the techniques of the present disclosure may use machine learning models to analyze only those cardiac features identified by the characterization as likely to present the patient with an arrhythmia episode. By using the machine learning models to validate arrhythmia detection in patients, the techniques of the present disclosure may increase the accuracy of arrhythmia detection. Furthermore, by using low power characterization to limit the use of computationally complex and power-intensive machine learning models to only the most relevant patient data, the techniques of the present disclosure may efficiently implement machine learning models to detect cardiac arrhythmia detection without adversely increasing the power usage or decreasing the battery life of such medical devices.

[0011] In one example, the disclosure provides a method for detecting cardiac electrogram data of a patient by a medical device including a processing circuit and a storage medium; performing, by the medical device, a feature-based representation of the sensed electrocardiogram data to obtain cardiac features present in the electrocardiogram data and indicative of an arrhythmia episode in the patient; determining, by the medical device and based on the feature-based representation, that the cardiac features meet threshold criteria for application of a machine learning model to verify that an arrhythmia episode has occurred in the patient; and determining, in response to determining that the cardiac features meet threshold criteria for application of the machine learning model, that the machine learning model trained using the electrocardiogram data of a plurality of patients is configured to generate a cardiac electrogram data. A method is described that includes applying a learning model to cardiac electrogram data sensed by a medical device; verifying that an arrhythmia episode has occurred in the patient based on the machine learning model; generating, in response to verifying that an arrhythmia episode has occurred in the patient by the machine learning model, a report by the medical device including an indication that an arrhythmia episode has occurred in the patient and one or more cardiac features consistent with the arrhythmia episode; and outputting, for display, the report including an indication that an arrhythmia episode has occurred in the patient by the medical device and one or more cardiac features consistent with the arrhythmia episode.

[0012] In another example, the disclosure describes a method including: sensing, by a medical device including a processing circuit and a storage medium, electrocardiogram data of a patient; performing, by the medical device, feature-based representation of the sensed electrocardiogram data to obtain cardiac features present in the electrocardiogram data; determining, by the medical device, a similarity between the obtained cardiac features and cardiac features of each entry of a plurality of entries of an arrhythmia dictionary of the medical device; and, in response to a determination that each entry of the plurality of entries of the arrhythmia dictionary includes an arrhythmia classification with a plurality of classifications of arrhythmia in the patient and cardiac features revealing an arrhythmia classification, the obtained cardiac features are not similar to the cardiac features of each entry of the plurality of entries of the arrhythmia dictionary, determining, by the medical device, that an arrhythmia episode has occurred in the patient based on a machine learning model and applying the machine learning model trained using the electrocardiogram data of the plurality of patients to the sensed electrocardiogram data; and storing, by the medical device, a first entry in the arrhythmia dictionary including a first classification of the arrhythmia episode and the obtained cardiac features.

[0013] This summary is intended to provide an overview of the subject matter described in the present disclosure. It is not intended to provide an exclusive or comprehensive description of the apparatus and methods described in detail in the accompanying drawings and description below. One or more further details are set forth in the accompanying drawings and the description below. [Brief description of the drawings]

[0014] [Figure 1] FIG. 1 is a conceptual diagram illustrating an example of a medical device system for predicting cardiac arrhythmia including a leadless implantable medical device and a patient-associated external device in accordance with the techniques of the present disclosure. [Diagram 2] FIG. 2 is a block diagram showing an example of the implantable medical device of FIG. 1. [Diagram 3] FIG. 2 is a block diagram showing a configuration example of the implantable medical device of FIG. 1. [Figure 4]4 is a flowchart illustrating an example operation according to the techniques of this disclosure. [Diagram 5] 4 is a flowchart illustrating an example operation according to the techniques of this disclosure. [Figure 6] 4 is a flowchart illustrating an example operation according to the techniques of this disclosure. [Figure 7] 4 is a flowchart illustrating an example operation according to the techniques of this disclosure. [Figure 8] 4 is a flowchart illustrating an example operation according to the techniques of this disclosure.

[0015] Like reference characters refer to like elements throughout the figures and description. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Techniques are disclosed for the efficient use of machine learning methods for cardiac arrhythmia detection in medical devices. Characterization algorithms may use sensed electrocardiogram data from a patient to perform, for example, QRS detection and / or arrhythmia detection. Such characterization algorithms may be optimized for real-time embedded, low-power applications, such as for use in implantable medical devices. However, characterization algorithms may require expert design and feature engineering to accurately detect arrhythmias in a patient.

[0017] Machine learning methods for detecting arrhythmias, such as deep learning and artificial intelligence (AI), provide a flexible platform for developing algorithms for various purposes. For example, machine learning systems can detect, for example, atrial fibrillation (AF), filter out episodes that do not indicate arrhythmias, and the like, with high accuracy without the expert design and feature engineering required for cardiac arrhythmia algorithms, such as characterization. However, machine learning systems may be computationally prohibitive for implementation in medical devices, such as IMDs or medical devices that operate on battery power. Heavy use of computationally expensive machine learning models in medical devices can impact battery life.

[0018] As described in detail herein, techniques, methods, systems, and devices are disclosed for physiological, device-based, and algorithm-based methods of conditioning the use of on-device machine learning systems to ensure efficient power usage. As described herein, a medical device system is described that enables the use of in-device machine learning arrhythmia detection, such as deep learning or AI, by medical devices that perform short-term or long-term diagnostic analysis or monitoring in a power-efficient manner to enable the use of machine learning arrhythmia detection.

[0019] FIG. 1 illustrates an example medical device system 2 environment in relation to a patient 4 and a heart 6, in accordance with certain example apparatus and methods described herein. The example technology may be used with an IMD 10, which may be leadless and in wireless communication with an external device 12, as shown in FIG. 1. In some examples, the IMD 10 may be coupled to one or more leads. In some examples, the IMD 10 may be implanted outside the thoracic cavity of the patient 4 (e.g., subcutaneously in the chest as shown in FIG. 1). The IMD 10 may be positioned near the level of the heart 6 and / or just below the sternum.

[0020] In some examples, the IMD 10 may take the form of a Reveal LINQ™ Insertable Cardiac Monitor (ICM) or a Holter Cardiac Monitor, available from Medtronic plc, Dublin, Ireland. As discussed herein, the techniques of the present disclosure may be performed by an implantable device, such as the IMD 10.

[0021] In other examples, the techniques described herein may be performed by an external medical device, such as external device 12, in addition to or instead of IMD 10. Such an external medical device may be disposed external to patient 4 (e.g., disposed on the skin of patient 4) and may perform any or all of the functions described herein with respect to IMD 10. External device 12 may be a computing device configured for use in settings such as a home, clinic, or hospital, and may further be configured to communicate with IMD 10 via wireless telemetry. For example, external device 12 may couple to a remote patient monitoring system, such as Carelink® available from Medtronic plc of Dublin, Ireland. External device 12 may include, in some examples, a programmer, an external monitor, or a mobile device, such as a mobile phone, a “smart” phone, a laptop, a tablet computer, a personal digital assistant (PDA), etc. In some examples, external device 12 is a wearable electronic device, such as the SEEQ™ Mobile Cardiac Telemetry (MCT) System available from Medtronic plc, Dublin, Ireland, or another type of wearable “smart” electronic apparel, such as a “smart” watch, a “smart” patch, or “smart” glasses.

[0022] In some examples, a user, such as a physician, technician, surgeon, electrophysiologist, or other clinician, may interact with external device 12 to retrieve physiological or diagnostic information from IMD 10. In some examples, a user, such as a patient 4 or a clinician as described above, may also interact with external device 12 to program IMD 10, e.g., to select or adjust values ​​of operating parameters of IMD 10. In some examples, external device 12 serves as an access point to facilitate communication with IMD 10.

[0023] A user, such as a physician, technician, surgeon, electrophysiologist, or other clinician, may interact with the external device 12 to retrieve physiological or diagnostic information from the IMD 10. The user may also interact with the external device 12 to program the IMD 10, for example, to select values ​​for the IMD's operating parameters. The external device 12 may include a processor configured to evaluate EGMs and / or other sensed signals transmitted from the IMD 10 to the external device 12.

[0024] In any such example, the processing circuitry of the medical device system 2 may transmit patient data including electrocardiogram data of the patient 4 to a remote computer (e.g., the external device 12, or another device not shown in FIG. 1 ). In some examples, the processing circuitry of the medical device system 2 may transmit a determination that the patient 4 is experiencing an episode of cardiac arrhythmia, such as an episode of bradycardia, tachycardia, atrial fibrillation, or ventricular fibrillation.

[0025] The external device 12 may be a computing device (e.g., used in a home, outpatient, clinic, or hospital setting) for communicating with the IMD 10 via wireless telemetry. The external device 12 may include or be coupled to a remote patient monitoring system, such as Carelink® available from Medtronic plc of Dublin, Ireland. In some examples, the external device 12 may receive data, alerts, patient physiological information, or other information from the IMD 10.

[0026] The external device 12 may be used to program commands or operating parameters into the IMD 10 to control its functions (e.g., when configured as a programmer for the IMD 10). In some examples, the external device 12 may be used to interrogate the IMD 10 to retrieve data, including operational data of the device as well as physiological data stored in the IMD memory. Such interrogations may occur automatically according to a schedule and / or in response to remote or local user commands. Programmers, external monitors, and consumer devices are examples of external devices 12 that may be used to interrogate the IMD 10. Examples of communication technologies used by the IMD 10 and the external device 12 include radio frequency (RF) telemetry, which may be an RF link established via Bluetooth, WiFi, or Medical Implant Communication Service (MICS). In some examples, the external device 12 may include a user interface configured to allow the patient 4, a clinician, or another user to remotely interact with the IMD 10. In some such examples, the external device 12 and / or other devices of the medical device system 2 may be a wearable device (eg, in the form of a watch, necklace, or other wearable item).

[0027] The medical device system 2 is an example of a medical device system configured to perform cardiac arrhythmia detection, verification, and reporting. In accordance with the techniques of this disclosure, the medical device system 2 performs machine learning arrhythmia detection and characterization to detect and classify cardiac arrhythmias in the patient 4. Additional examples of the one or more other implantable or external devices may include an implantable, multi-channel cardiac pacemaker, ICD, IPG, leadless (e.g., intracardiac) pacemaker, extravascular pacemaker and / or ICD, or other IMD, or a combination of such IMDs configured to deliver CRT to the heart 6, an external therapy delivery device such as an external monitor, an external pacing or electrical stimulation device, or a drug pump. In some examples, the IMD 10 implements a machine learning system, such as a neural network, a deep learning system, or other type of predictive analytics system.

[0028] Communications circuitry in each device (e.g., IMD 10 and external device 12) of medical device system 2 may enable the devices to communicate with one another. Additionally, while one or more sensors (e.g., electrodes) are described herein as being disposed on the housing of IMD 10, in other examples, such sensors may be disposed on the housing of another device implanted inside or outside patient 4. In such examples, one or more of the other devices may include processing circuitry configured to receive signals from the electrodes or other sensors on the respective device, and / or communications circuitry configured to transmit signals from the electrodes or other sensors to another device (e.g., external device 12) or a server.

[0029] In accordance with the techniques of the present disclosure, the medical device system 2 uses characterization and machine learning to perform cardiac arrhythmia detection and classification. Specifically, a medical device, such as the IMD 10 or the external device 12, uses the characterization to perform a preliminary detection of cardiac arrhythmia in the patient 4. In some examples, the medical device applies the machine learning model to the electrocardiogram data of the patient 2 to verify that the characterization of the electrocardiogram data correctly detected an episode of cardiac arrhythmia. In some examples, the medical device applies the machine learning model to the electrocardiogram data of the patient 2 to verify that the characterization of the electrocardiogram data correctly classified the episode of cardiac arrhythmia as a particular type of arrhythmia. For ease of explanation, the following sections describe the techniques of the disclosure as performed by the IMD 10. However, the techniques of the present disclosure may be performed by other types of medical devices, such as the external device 12, or a combination of medical devices (e.g., the IMD 10 and the external device 12) operating in conjunction with each other.

[0030] In one example of the disclosed technology, the IMD 10 senses electrocardiogram data of the patient 4. The IMD 10 performs feature-based portrayal of the electrocardiogram data to obtain cardiac features indicative of an arrhythmia episode in the patient 4. The IMD 10 determines whether the cardiac features meet a threshold criterion for application of a machine learning model to validate the feature-based portrayal of the electrocardiogram data. In some examples, the IMD 10 further determines that a noise of at least one of the cardiac features is below a predetermined threshold. In some examples, the IMD 10 further determines that the patient is in a first posture state of a plurality of posture states or a first activity state of a plurality of activity states. In response to determining that the cardiac features meet the threshold criterion, the IMD 10 applies the machine learning model to the sensed electrocardiogram data to, for example, validate that an arrhythmia episode has occurred in the patient 4 or detect one or more additional types of arrhythmia that have occurred in the patient 4.

[0031] In one example of the disclosed technology, IMD 10 may classify an arrhythmia episode by comparing cardiac features consistent with the arrhythmia episode to cardiac features in an arrhythmia dictionary maintained by IMD 10. IMD 10 compares a first cardiac feature of the electrocardiogram data to cardiac features defined by an entry in the arrhythmia dictionary. For example, in response to determining that the first cardiac feature of the electrocardiogram data is similar to a cardiac feature defined by an entry in the arrhythmia dictionary, IMD 10 determines that the first cardiac feature indicates that an arrhythmia episode has occurred in patient 4, which is a classification defined by a matching entry in the arrhythmia dictionary.

[0032] As another example, in response to determining that a first cardiac feature of the electrocardiogram data is not similar to the cardiac features defined by any entry in the arrhythmia dictionary, IMD10 applies a machine learning model to determine a classification of the arrhythmia episode revealed by the first cardiac feature. IMD10 may store the determined arrhythmia classification and cardiac feature as a new entry in the arrhythmia dictionary to create an arrhythmia dictionary. Subsequently, upon detecting, via characterization, a second cardiac feature that is similar to the cardiac feature of the arrhythmia dictionary entry, IMD10 determines that the second cardiac feature is indicative of an arrhythmia episode of the same classification as an arrhythmia episode including a cardiac feature defined in the arrhythmia dictionary entry and matching the second cardiac feature.

[0033] The techniques of the present disclosure may provide certain improvements in the field of cardiac arrhythmia detection and classification by medical devices such as the IMD 10. For example, the techniques of the present disclosure may use machine learning models to analyze only electrocardiogram data identified by characterization as likely to present the patient with an arrhythmia episode. By using machine learning models to validate arrhythmia detection of patient 4 performed by characterization, the techniques of the present disclosure may leverage machine learning to increase the accuracy and flexibility of arrhythmia detection. Furthermore, by using low power characterization to limit the use of computationally complex and power-intensive machine learning models to only the most relevant patient data, the techniques of the present disclosure may efficiently implement machine learning models to detect cardiac arrhythmia detection without adversely increasing the power usage or decreasing the battery life of such medical devices.

[0034] FIG. 2 is a block diagram illustrating an example of the implantable medical device of FIG. 1. As shown in FIG. 2, the IMD 10 includes a processing circuit 50, a sensing circuit 52, a communication circuit 54, a memory 56, a sensor 58, a switching circuit 60, a characterization circuit 260, and electrodes 16A, 16B (hereinafter, "electrodes 16"), one or more of which may be disposed within a housing of the IMD 10. In some examples, the memory 56 includes computer-readable instructions that, when executed by the processing circuit 50, cause the IMD 10 and the processing circuit 50 to perform various functions attributed to the IMD 10 and the processing circuit 50 herein. The memory 56 may include any volatile, non-volatile, magnetic, optical, or electrical medium, 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 medium.

[0035] The processing circuitry 50 may include fixed function circuits and / or programmable processing circuits. The processing circuitry 50 may include any one or more of 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 analog logic circuitry. In some examples, the processing circuitry 50 may include multiple components, such as 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 circuits. The functionality attributed to the processing circuitry 50 herein may be embodied as software, firmware, hardware, or any combination thereof.

[0036] The sensing circuit 52 and the communication circuit 54 may be selectively coupled to the electrodes 16A, 16B via the switching circuit 60 as controlled by the processing circuit 50. The sensing circuit 52 may monitor signals from the electrodes 16A, 16B to monitor the electrical activity of the heart of the patient 4 of FIG. 1 and generate electrocardiogram data for the patient 4. In some examples, the processing circuit 50 performs characterization of the sensed electrocardiogram data via the characterization circuit 260 and detects an episode of cardiac arrhythmia in the patient 4. In some examples, the processing circuit 50 transmits the electrocardiogram data of the patient 4 to an external device, such as the external device 12 of FIG. 1, via the communication circuit 54. For example, the IMD 10 transmits the digitized electrocardiogram data to the external device 12 of FIG. 1 for data processing or review by a clinician. In some examples, the IMD 10 transmits one or more segments of electrocardiogram data in response to detecting an episode of arrhythmia via the characterization circuit 260. In another example, the IMD 10 transmits one or more segments of electrocardiogram data in response to an instruction from the external device 12 (e.g., when the patient 4 experiences one or more symptoms of an arrhythmia and enters a command into the external device 12 instructing the IMD 10 to upload the electrocardiogram data for analysis by a monitoring center or clinician).

[0037] In some examples, the IMD 10 includes one or more sensors 58, such as one or more accelerometers, microphones, and / or pressure sensors. The sensing circuitry 52 may monitor signals from the sensors 58 and transmit patient data obtained from the sensors 58 to an external device, such as the external device 12 of FIG. 1, for analysis. In some examples, the sensing circuitry 52 may include one or more filters and amplifiers for filtering and amplifying signals received from one or more of the electrodes 16A, 16B and / or other sensors 58. In some examples, the sensing circuitry 52 and / or the processing circuitry 50 may include a rectifier, a filter and / or an amplifier, a sense amplifier, a comparator, and / or an analog-to-digital converter.

[0038] The communications circuitry 54 may include any suitable hardware, firmware, software, or any combination thereof for communicating with another device, such as the external device 12, or another medical device or sensor, such as a pressure sensing device. Under the control of the processing circuitry 50, the communications circuitry 54 may receive downlink telemetry and transmit uplink telemetry from the external device 12 or another device with the aid of an internal or external antenna, e.g., antenna 26. In some examples, the communications circuitry 54 may communicate with the external device 12. In addition, the processing circuitry 50 may communicate with external devices (e.g., the external device 12) and networked computing devices via a computer network, such as the Medtronic CareLink® network developed by Medtronic, plc of Dublin, Ireland.

[0039] A clinician or other user may retrieve data from IMD 10 using external device 12 or by using another local or networked computing device configured to communicate with processing circuit 50 via communications circuitry 54. A clinician may also program parameters of IMD 10 using external device 12 or another local or networked computing device. In some examples, a clinician may select one or more parameters that define how IMD 10 senses electrocardiogram data of patient 4.

[0040] One or more components of IMD 10 may be coupled to a power source (not shown in FIG. 2), which may include a rechargeable or non-rechargeable battery disposed within the housing of IMD 10. The non-rechargeable battery may be selected to last for several years, while the rechargeable battery may be inductively charged from an external device, for example, on a daily or weekly basis.

[0041] According to the techniques of the present disclosure, the processing circuit 50 senses electrocardiogram data of the patient 4 via the electrodes 16 using the sensing circuit 52. In some examples, the electrocardiogram data is an ECG of the patient 4. The processing circuit 50 performs characterization of the electrocardiogram data via the characterization circuit 260 to obtain one or more cardiac features present in the electrocardiogram data. The characterization circuit 260 may further perform preliminary detection of arrhythmia episodes. In some examples, the characterization circuit 260 includes circuitry configured to perform one or more of QRS detection, refractory processing, noise processing, or characterization of the electrocardiogram data. For example, the characterization circuit 260 receives raw signals via the sensing circuit 50 and / or the sensor 58 and extracts one or more cardiac features from the raw signals. In some examples, the characterization circuit 260 identifies one or more cardiac features, such as one or more of the RR intervals present in the electrocardiogram data, an average heart rate present in the electrocardiogram data, a minimum heart rate present in the electrocardiogram data, a maximum heart rate present in the electrocardiogram data, a PR interval present in the electrocardiogram data, heart rate variability present in the electrocardiogram data, one or more amplitudes of one or more features of the ECG, or intervals between one or more features of the ECG, T-wave alternans, QRS morphology, or other types not expressly described herein.

[0042] As an example, the characterization circuit 260 identifies one or more features of a T wave of an electrocardiogram of the patient 4 to detect an episode of cardiac arrhythmia in the patient 4. In some examples, the one or more identified features are one or more amplitudes of the T wave. In some examples, the one or more identified features are a frequency of the T wave. In some examples, the one or more identified features include at least a T wave amplitude and a T wave frequency. In some examples, the characterization circuit 260 identifies one or more relative changes in the one or more identified features indicative of an episode of cardiac arrhythmia in the patient 4. In some examples, the characterization circuit 260 identifies one or more interactions between a plurality of identified features indicative of an episode of cardiac arrhythmia in the patient 4. In some examples, the characterization circuit 260 analyzes the patient data representing one or more values ​​averaged over a short period of time (e.g., about 3 minutes). For example, the electrocardiogram data may include one or more of an average frequency or an average amplitude of a T wave or a QRS wave of the electrocardiogram of the patient 4 to detect an episode of cardiac arrhythmia.

[0043] The processing circuit 50 may apply the characterization via the characterization circuit 260 to determine that one or more cardiac features are indicative of an episode of cardiac arrhythmia. In some examples, the processing circuit 50 applies the characterization via the characterization circuit 260 to classify the detected cardiac arrhythmia episode as an episode of a particular type of cardiac arrhythmia (e.g., bradycardia, tachycardia, atrial fibrillation, or ventricular fibrillation). In some examples, the processing circuit 50 performs characterization of the sensed cardiac electrogram data via the characterization circuit 260, as described in more detail below. In some examples, the characterization performed by the IMD 10 is reduced in complexity to conserve power of the IMD 10. This may enable the characterization circuit 260 to perform early or preliminary detection of cardiac arrhythmia.

[0044] Additionally, as described in more detail below, the processing circuit 50 applies the machine learning system 250 to the electrocardiogram data to verify or classify the detection of an arrhythmia episode by the characterization circuit 260. The machine learning system 250 may perform a more comprehensive and detailed analysis of the electrocardiogram data to more accurately detect cardiac arrhythmias across the characterization circuit 260, but the machine learning system 250 may require more computational resources and power for the characterization circuit 260. By using the machine learning system 250 to verify or classify the detection of an arrhythmia episode by the characterization circuit 260, the IMD 10 may take advantage of the high accuracy provided by the machine learning system 250 while minimizing the power consumption or battery life of the IMD 10. In some examples, the processing circuit 50 transmits, via the communication circuit 54, to the external device 12, one or more of the electrocardiogram data, one or more cardiac features present in the electrocardiogram data, an indication of an episode of cardiac arrhythmia verified by the machine learning system 250, or an indication of the classification of the detected episode of cardiac arrhythmia determined by the machine learning system 250.

[0045] In some examples, the machine learning model implemented by the machine learning system 250 is trained with training data including electrocardiogram data of multiple patients labeled with descriptive metadata. For example, during the training phase, the machine learning system 250 processes multiple ECG waveforms. Typically, the multiple ECG waveforms are from multiple different patients. Each ECG waveform is labeled with one or more episodes of one or more types of arrhythmia. For example, the training ECG waveform may include multiple segments, each segment labeled with a descriptor that specifies the absence of arrhythmia or the presence of a particular classification of arrhythmia (e.g., bradycardia, tachycardia, atrial fibrillation, or ventricular fibrillation). In some examples, a clinician manually labels the presence of arrhythmia in each ECG waveform. In some examples, the presence of arrhythmia in each ECG waveform is labeled according to a classification by a characterization algorithm. The machine learning system 250 can operate to convert the training data into vectors and tensors (e.g., multidimensional arrays), on which the machine learning system 250 can apply mathematical operations such as linear algebra, nonlinear, or alternative computational operations. The machine learning system 250 uses the training data 104 to teach the machine learning model to weigh different features present in the electrocardiogram data. In some examples, the machine learning system 250 uses the electrocardiogram data to teach the machine learning model to apply different coefficients that represent one or more features of the electrocardiogram as being more or less important with respect to the occurrence of a particular class of cardiac arrhythmia. By processing a large number of such ECG waveforms labeled with arrhythmia episodes, the machine learning system 250 can build and train a machine learning model to receive electrocardiogram data from patients not previously analyzed by the machine learning system 250, such as patient 4 of FIG. 1, and further process such electrocardiogram data to detect the presence or absence of various classes of arrhythmia in the patient with high accuracy. Typically, the greater the amount of electrocardiogram data on which the machine learning system 250 is trained, the greater the accuracy of the machine learning model in detecting or classifying cardiac arrhythmias in new electrocardiogram data.

[0046] After the machine learning system 250 trains the machine learning model, the machine learning system 250 may receive patient data, such as electrocardiogram data, of a particular patient, such as patient 4. The machine learning system 250 applies the trained machine learning model to the patient data to detect cardiac arrhythmia episodes of the patient 4. Further, the machine learning system 250 applies the trained machine learning model to the patient data to classify the patient's cardiac arrhythmia episodes as indicative of a particular type of arrhythmia. In some examples, the machine learning system 250 may output a preliminary determination that the cardiac arrhythmia episode is indicative of a particular type of arrhythmia, as well as an estimate of the certainty in the determination. In response to determining that the estimate of the certainty in the determination is greater than a predetermined threshold (e.g., 50%, 75%, 90%, 95%, 99%), the processing circuit 50 may classify the cardiac arrhythmia episode as a particular type of arrhythmia. As described herein, the processing circuit 50 may use the machine learning system 250 to verify that the characterization circuit 260 correctly detected the arrhythmia episode or that the characterization circuit 260 correctly classified the arrhythmia episode as a particular type.

[0047] In some examples, the machine learning system may process one or more cardiac features of the electrocardiogram data instead of the raw electrocardiogram data itself. The one or more cardiac features may be obtained via characterization performed by the IMD 10, as described above. The cardiac features may include, for example, one or more of the RR intervals present in the electrocardiogram data, the average heart rate present in the electrocardiogram data, the minimum heart rate present in the electrocardiogram data, the maximum heart rate present in the electrocardiogram data, the PR intervals present in the electrocardiogram data, the heart rate variability present in the electrocardiogram data, one or more amplitudes of one or more features of the ECG, T-wave alternans, QRS morphometry, or other types of cardiac features not explicitly described herein. In such an example implementation, the machine learning system may train the machine learning model via a plurality of training cardiac features labeled with arrhythmia episodes, instead of a plurality of ECG waveforms labeled with arrhythmia episodes as described above.

[0048] In some examples, the machine learning system 250 may apply the machine learning model to other types of data to determine that an arrhythmia episode has occurred in the patient 4. For example, the machine learning system 250 may apply the machine learning model to one or more characteristics of the electrocardiogram data that correlate to the patient's arrhythmia, the activity level of the IMD 10, the input impedance of the IMD 10, or the battery level of the IMD 10.

[0049] In a further example, the processing circuitry 50 may generate an intermediate representation of the electrocardiogram data from the electrocardiogram data. For example, the processing circuitry 50 may apply one or more signal processing, signal decomposition, wavelet decomposition, filtering, or noise reduction operations to the electrocardiogram data to generate an intermediate representation of the electrocardiogram data. In this example, the machine learning system 250 processes such intermediate representation of the electrocardiogram data to detect and classify arrhythmia episodes of the patient 4. Furthermore, the machine learning system 250 may train a machine learning model via a plurality of training intermediate representations labeled with arrhythmia episodes, instead of a plurality of raw ECG waveforms labeled with arrhythmia episodes as described above. The use of such intermediate representations of the electrocardiogram data may enable the machine learning system 250 to train and develop a lighter-weight, less computationally complex machine learning model. Furthermore, the use of such intermediate representations of the electrocardiogram data may require fewer iterations and less training data to build an accurate machine learning model, as opposed to the use of raw electrocardiogram data to train the machine learning model.

[0050] In some examples, the memory 56 includes an arrhythmia dictionary 270. In some examples, the arrhythmia dictionary 270 includes a plurality of entries. Each entry of the plurality of entries includes a classification of one or more particular types of cardiac arrhythmia (e.g., bradycardia, tachycardia, atrial fibrillation, or ventricular fibrillation). Additionally, the entry includes one or more cardiac features indicative of the classification of the cardiac arrhythmia. As described in more detail below, the processing circuit 50 uses the arrhythmia dictionary 270 to classify arrhythmia episodes detected via the characterization circuit 260 as being of a particular type of arrhythmia. Additionally, the processing circuit 50 applies the machine learning system 250 to classify detected episodes of arrhythmia for which the arrhythmia dictionary 270 does not include a corresponding entry to build robust entries in the arrhythmia dictionary 270.

[0051] Although described herein in the context of an exemplary IMD 10 sensing cardiac electrogram data of a patient 4, the techniques for cardiac arrhythmia detection disclosed herein may be used with other types of devices. For example, the techniques may be implemented in an extracardiac defibrillator coupled to electrodes outside the cardiovascular system, a transcatheter pacemaker configured to be implanted within the heart, such as, for example, the Micra™ transcatheter pacing system available from Medtronic PLC of Dublin, Ireland, an insertable cardiac monitor, such as the Reveal LINQ™ ICM, also available from Medtronic PLC of Dublin, Ireland, a neurostimulator, such as the SEEQ™ Mobile Cardiac Telemetry (MCT) system available from Medtronic plc of Dublin, Ireland, a drug delivery device, a medical device external to the patient 4, a wearable device, such as a wearable defibrillator, 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 “smart” apparel, such as “smart” glasses, a “smart” patch, or a “smart” watch.

[0052] FIG. 3 is a block diagram illustrating an example configuration of the implantable medical device of FIG. 1. Components in FIG. 3 are not necessarily drawn to scale, but instead may be enlarged to show detail. Specifically, FIG. 3 is a block diagram of a top view of an example configuration of the IMD 10 of FIG. 1.

[0053] FIG. 3 is a conceptual diagram illustrating an example IMD 10 that may include components substantially similar to the IMD 10 of FIG. 1. In addition to the components shown in FIGS. 1 and 2, the example IMD 10 shown in FIG. 3 may also include a wafer-scale insulating cover 74 that may help insulate electrical signals passing between the electrodes 16A, 16B on the housing 14 and the processing circuitry 50. In some examples, the insulating cover 74 may be disposed over the open housing 14 to form a housing for the components of the IMD 10B. One or more components of the IMD 10B (e.g., the antenna 26, the processing circuitry 50, the sensing circuitry 52, the communication circuitry 54, and / or the switching circuitry 60) may be formed on the underside of the insulating cover 74, such as by using flip-chip technology. The insulating cover 74 may be inverted onto the housing 14. When inverted and disposed on the housing 14, the components of the IMD 10 formed on the underside of the insulating cover 74 may be disposed within a gap 78 defined by the housing 14. The housing 14 can be formed from titanium or any other suitable material (e.g., a biocompatible material) and can have a thickness of about 200 micrometers to about 500 micrometers. These materials and dimensions are merely examples, and other materials and other thicknesses are possible for the devices of the present disclosure.

[0054] In some examples, the IMD 10 collects patient data for the patient 4, including electrocardiogram data, via the sensing circuitry 50 and / or the sensor 58. The sensor 58 may include 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 includes one or more of the patient's activity level, the patient's heart rate, the patient's posture, the patient's electrocardiogram, the patient's blood pressure, the patient's accelerometer data, or other types of patient data. In some examples, the IMD 10 uploads the patient data to the external device 12 via the communication circuitry 54, which may upload such data to a remote monitoring center or patient monitoring network. In some examples, the IMD 10 uploads the patient data daily. In some examples, the patient data includes one or more values ​​representing average measurements for the patient 4 over an extended period of time (e.g., about 24 hours to about 48 hours). In this example, the IMD 10 uploads the patient data as well as performs arrhythmia detection and classification for the patient 4 (as described below). However, in other examples, the medical device that processes the patient data to detect and / or classify arrhythmias in the patient 4 is different from the medical device that performs short-term monitoring of the patient 4. For example, the IMD 10 may perform short-term monitoring of the patient 4, and the external device 12 processes the patient data to detect and / or classify arrhythmias in the patient 4.

[0055] FIG. 4 is a flow chart illustrating an example operation in accordance with the techniques of this disclosure. For convenience, FIG. 4 is described with respect to FIG. 1. In some examples, the operation of FIG. 4 is an operation for detecting and classifying cardiac arrhythmia in patient 4. In the operation of FIG. 4, system 2 combines the accuracy of the machine learning model of machine learning system 250 in detecting cardiac arrhythmia with the low power consumption of the characterization algorithm. Although the operation of FIG. 4 is described with respect to IMD 10 of FIG. 1, in other examples, the operation of FIG. 4 may be performed by, for example, external device 12, or a combination of IMD 10 and external device 12.

[0056] As shown in Figure 4, the IMD 10 senses 402 electrocardiogram data of a patient 4. The electrocardiogram data may be, for example, a transient ECG of the patient 4 or a full disclosure ECG of the patient 4. Furthermore, the electrocardiogram data of the patient 4 may be from a single channel or a multi-channel system. For simplicity, in the example of Figure 4, the electrocardiogram data of the patient 4 is described as single channel transient ECG data.

[0057] The IMD 10 applies the characterization to the electrocardiogram data to detect one or more cardiac features (404). The IMD 10 further applies the characterization to the electrocardiogram data to detect one or more episodes of arrhythmia. For example, the IMD 10 may apply a QRS detection characterization and a noise flag to the electrocardiogram data to provide arrhythmia characteristics and / or cardiac features of the detected episodes of arrhythmia (e.g., heart rate variability during an episode of atrial fibrillation, duration of pauses). In some examples, the characterization is a one-size-fits-all type algorithm. In other examples, the characterization may be an algorithm with parameters programmed by a clinician or specific to the patient 4. In some examples, the IMD 10 applies the characterization to the electrocardiogram data to classify the detected arrhythmia episodes as a particular type of arrhythmia.

[0058] The IMD 10 further determines a trigger for a detected episode of cardiac arrhythmia. For example, an episode of cardiac arrhythmia may be detected in response to an automatic trigger, in response to a baseline trigger, or in response to input received from the patient 4. For example, a baseline trigger may not occur in response to a detected arrhythmia, but rather in response to a periodically occurring electrocardiogram segment capture operation (e.g., a segment captured upon implantation or attachment of the IMD 10, or a segment captured once every 24 hours). In some examples, the IMD 10 may apply the machine learning system 250 to validate only arrhythmia episodes detected in response to an automatic trigger, while arrhythmia episodes detected in response to a baseline trigger or patient input may be reported without validation by the machine learning system 250. For example, in response to determining that an arrhythmia episode was not detected by an automatic trigger (e.g., “No” at block 406), the IMD 10 generates (410) a report of the arrhythmia and outputs (412) the report to a clinician or to a monitoring center for review, as described below.

[0059] In some examples, the IMD 10 may use the machine learning system 250 to perform several tasks. For example, the IMD 10 may use the machine learning system 250 to verify that the detection of the cardiac arrhythmia by the characterization was appropriate. Using the machine learning system 250 in this manner may limit the computational complexity of the machine learning model and provide the clinician with information relevant to the diagnosis. For example, a machine learning model may be used with high arrhythmia detection sensitivity to verify that an automatic trigger of atrial fibrillation correlates with an episode of atrial fibrillation in the patient 4. As another example, a machine learning model may be used to ensure that only the portion of the cardiac electrogram that coincides with the duration of a verified episode of atrial fibrillation in the patient 4 is used to calculate the atrial fibrillation burden in the patient 4. As described herein, the burden of an arrhythmia episode is the ratio of the length of time of the arrhythmia episode to the length of time during which the patient was monitored by the IMD 10.

[0060] As another example, IMD 10 may use machine learning system 250 to detect the presence of other episodes of arrhythmia not detected by the characterization but consistent with arrhythmia episodes detected by the characterization. For example, IMD 10 may use machine learning system 250 to detect the presence of "high-severity" arrhythmias such as pauses, atrial fibrillation, or ventricular tachycardia. The use of machine learning models designed to detect the presence or absence of cardiac arrhythmia episodes of a range of types of arrhythmias may be of more limited computational complexity than machine learning models designed to detect instances of precise types and occurrences of cardiac arrhythmia episodes.

[0061] 4, in response to determining that a cardiac arrhythmia episode was detected in response to the automatic trigger (e.g., "Yes" at block 406), IMD10 applies the machine learning model of machine learning system 250 to the electrocardiogram data to verify the detection of the cardiac arrhythmia episode by characterization (408). In some examples, IMD10 applies the machine learning model of machine learning system 250 to the electrocardiogram data to verify the classification of the cardiac arrhythmia episode by characterization.

[0062] In some examples, arrhythmia episodes detected in response to an automatic trigger are of the same limited duration (e.g., 30 seconds before and after arrhythmia detection) and are typically shorter in duration than patient-triggered episodes. In some examples, the machine learning model receives as input at least a portion of the electrocardiogram waveform that caused the automatic trigger and the reason for the trigger (e.g., detection of atrial fibrillation, bradycardia, or pause characterization). The electrocardiogram waveform can be, for example, an ECG waveform.

[0063] In some examples, the machine learning model receives as input a preprocessed version of the electrocardiogram waveform, or a signal decomposition of the electrocardiogram waveform. For example, IMD10 may apply preprocessing to the electrocardiogram waveform by applying a downsampling or signal normalization operation to the waveform to generate an intermediate representation of the electrocardiogram. The signal decomposition is the use of wavelet decomposition bands or a frequency domain representation (e.g., a spectrogram). In some examples, IMD10 may apply preprocessing to the electrocardiogram waveform to reduce the computational complexity of the machine learning model. For example, by downsampling a record originally sampled at 200 samples / second to 100 samples / second, IMD10 may require half the computational resources to process the downsampled record as the original record. Similarly, by using the same signal decomposition used for arrhythmia detection, IMD10 may leverage precomputed data to reduce the computational complexity of the machine learning system 250, since the machine learning model does not need to learn an optimal set of signal convolutional layers.

[0064] In response to validating that the characterization detected the cardiac arrhythmia episode correctly, the IMD 10 stores and / or transmits the electrocardiogram data of the episode to a monitoring center or clinician for review. For example, the IMD 10 generates (410) a report of the arrhythmia and outputs (412) the report to the clinician or monitoring center. Typically, only cardiac arrhythmia episodes detected by the characterization and verified by the machine learning system 250 as appropriate or significant are stored and transmitted to the monitoring center and / or clinician for review. For example, if the machine learning system 250 detects an episode of bradycardia and the characterization performed on the electrocardiogram data indicates that four out of four noise-free heart beats are less than 30 beats per minute (BPM), the IMD 10 generates a report notifying a physician of the occurrence of an arrhythmia episode.

[0065] In one example, the report includes an indication that an arrhythmia episode occurred in the patient and one or more cardiac features consistent with the arrhythmia episode. In some examples, the report further includes a classification of the arrhythmia episode as a particular type of arrhythmia. In some examples, the report includes a subsection of electrocardiogram data obtained from patient 4 that is consistent with the arrhythmia episode. For example, IMD 10 may identify a subsection of electrocardiogram data for patient 4, where the subsection includes electrocardiogram data for a first period prior to the arrhythmia episode (e.g., typically less than 10 minutes prior to the onset of the arrhythmia episode), a second period during the onset of the arrhythmia episode, and a third period after the arrhythmia episode (e.g., typically less than 10 minutes after the cessation of the arrhythmia episode). In some examples, the duration of the episode may vary by device type and may further depend on the use case of the medical device, one or more settings of the medical device, or the particular type of arrhythmia sensed. For example, some types of arrhythmias self-terminate quickly (resulting in a shorter episode duration) while other types of arrhythmias are sustained, such that the recorded duration of the episode depends on the designated memory space of the medical device. Typically, the length of time of the patient's electrocardiogram data is greater than the first, second, and third time periods. Additionally, the IMD 10 identifies one or more cardiac features that coincide with the first, second, and third time periods. The IMD 10 includes in the report a subsection of the electrocardiogram data and the one or more cardiac features that coincide with the first, second, and third time periods.

[0066] In some examples, in response to the report, the IMD 10 receives one or more adjustments to one or more parameters used by the IMD 10 to sense electrocardiogram data of the patient 4. The IMD 10 performs such adjustments for subsequent sensing of electrocardiogram data of the patient 4.

[0067] FIG. 5 is a flow chart illustrating example operations in accordance with the techniques of this disclosure. For convenience, FIG. 5 is described in conjunction with FIG. 1. In some examples, the operations of FIG. 5 are operations for detecting and classifying cardiac arrhythmias in patient 4. In the operations of FIG. 5, system 2 combines the accuracy of the machine learning models of machine learning system 250 in detecting cardiac arrhythmias with the low power consumption of the characterization algorithms. Although the operations of FIG. 5 are described in conjunction with IMD 10 of FIG. 1, in other examples, the operations of FIG. 5 may be performed by, for example, external device 12, or a combination of IMD 10 and external device 12.

[0068] As shown in FIG. 5, the IMD 10 senses electrocardiogram data of the patient 4 (502). The IMD 10 applies the characterization to the electrocardiogram data to detect one or more cardiac features (504). The IMD 10 further determines whether an episode of cardiac arrhythmia was detected in response to an automatic trigger (506). For example, in response to a determination that an arrhythmia episode was not detected by an automatic trigger (e.g., “NO” in block 506), such as in the case of an episode of cardiac arrhythmia detected in response to a baseline trigger, or in response to input received from the patient 4, the IMD 10 generates an arrhythmia report (510) and outputs the report to a clinician or monitoring center for review (512). The operations of steps 502, 504, 506, 510, and 512 may occur in a manner substantially similar to steps 402, 404, 406, 410, and 412, respectively, of FIG. 4.

[0069] In response to determining that an episode of cardiac arrhythmia is detected in response to the automatic trigger (e.g., “yes” at block 506), the IMD processes the electrocardiogram data to generate filtered electrocardiogram data (507). For example, the IMD 10 may adjust the electrocardiogram data, such as one or more ECG episodes, based on device and physiological parameters, such as input impedance, activity level, and posture changes, to generate the filtered electrocardiogram data. For example, during periods when the patient 4 is exhibiting high activity levels, or during periods when the IMD 10 is experiencing highly variable levels of input impedance, the characterization may erroneously detect an episode of cardiac arrhythmia due to an automatic trigger in response to noise in the electrocardiogram data. The use of filtered electrocardiogram data may enable the IMD 10 to discard these periods of noisy data that may be affected by signal artifacts, such as changes in amplitude levels. Additionally, the use of filtered electrocardiogram data may reduce the computational complexity of the machine learning model, since the machine learning model may be designed to analyze noise-free data. Additionally, the use of filtered electrocardiogram data allows the IMD10 to verify that arrhythmia episodes were correctly detected by characterization without the use of machine learning models where arrhythmia episodes may have been falsely triggered due to noise in the electrocardiogram data, further reducing power consumption by the IMD10. In some instances, automatically triggered episodes of cardiac arrhythmias associated with pause-associated syncope or bradycardia may not be suppressed by processing with the machine learning models.

[0070] The IMD 10 applies the machine learning model of the machine learning system 250 to the filtered electrocardiogram data to verify the detection of the cardiac arrhythmia episode by the characterization (508). In some examples, the IMD 10 applies the machine learning model of the machine learning system 250 to the filtered electrocardiogram data to verify the classification of the cardiac arrhythmia episode by the characterization. In response to verifying that the detection of the cardiac arrhythmia episode by the characterization is correct, the IMD 10 stores and / or transmits the electrocardiogram data of the episode to a monitoring center or clinician for review. For example, the IMD 10 generates a report of the arrhythmia (510) and outputs the report to a clinician or monitoring center (512), as described above.

[0071] FIG. 6 is a flow chart illustrating example operations according to the techniques of this disclosure. For convenience, FIG. 6 is described in conjunction with FIGS. 1 and 2. In some examples, the operations of FIG. 6 are for detecting and classifying cardiac arrhythmias in patient 4 using an arrhythmia dictionary constructed by machine learning system 250. In the operations of FIG. 6, system 2 combines the accuracy of the machine learning models of machine learning system 250 in detecting cardiac arrhythmias with the low power consumption of the characterization algorithm. Although the operations of FIG. 6 are described in conjunction with IMD 10 of FIG. 1, in other examples, the operations of FIG. 6 may be performed by, for example, external device 12, or a combination of IMD 10 and external device 12.

[0072] As shown in Figure 6, the IMD 10 senses 602 electrocardiogram data of a patient 4. The IMD 10 applies characterization to the electrocardiogram data to detect one or more cardiac features 604. The operations of steps 602 and 604 may occur in a manner substantially similar to steps 402 and 404, respectively, of Figure 4.

[0073] The IMD 10 determines whether the arrhythmia dictionary 270 includes at least one entry (606). Each entry in the arrhythmia dictionary 270 includes a classification that defines an arrhythmia episode as a particular type or set of arrhythmias, such as an episode of bradycardia, tachycardia, atrial fibrillation, ventricular fibrillation, or atrioventricular block. Each entry further includes one or more cardiac features associated with a cardiac arrhythmia episode of the corresponding classification. In response to determining that the arrhythmia dictionary 270 includes at least one entry (e.g., “yes” in block 606), the IMD 10 determines whether the cardiac features obtained via characterization are similar to the cardiac features of the entry in the arrhythmia dictionary 270 (614). For example, the IMD 10 may compare one or more parameters of the cardiac features obtained via characterization to one or more parameters of the cardiac features of the entry in the arrhythmia dictionary. In some examples, IMD10 may determine the similarity between the cardiac features obtained via characterization and the cardiac features of an entry in the arrhythmia dictionary 270 by applying computationally efficient methods such as L1 distance, percent similarity, or if-then-else rules.

[0074] For example, IMD10 determines a similarity between the L1 distance of the cardiac feature obtained via characterization and the L1 distance of the cardiac feature of each entry of the multiple entries of the arrhythmia dictionary. In another example, IMD10 determines whether a difference between at least one parameter of the cardiac feature obtained via characterization and at least one parameter of the cardiac feature of each entry of the multiple entries of the arrhythmia dictionary 270 is greater than a predetermined threshold.

[0075] In response to determining that the cardiac feature obtained via characterization is similar to the cardiac feature of the arrhythmia dictionary entry (e.g., “yes” at block 614), IMD10 determines whether the arrhythmia dictionary entry defines the cardiac feature as indicative of an arrhythmia episode of interest (616). In response to determining that the arrhythmia dictionary entry defines the cardiac feature as indicative of an arrhythmia episode of interest (e.g., “yes” at block 616), IMD10 generates a report of the arrhythmia (610) and outputs the report to a clinician or monitoring center for review (612). The operations of steps 610 and 612 may occur in a manner substantially similar to steps 410 and 412 of FIG. 4, respectively.

[0076] In response to determining that an arrhythmia dictionary entry defines the cardiac features as not indicative of an arrhythmia episode of interest (e.g., "NO" at block 616), the IMD10 may archive the sensed electrogram data for later review by a center or clinician monitoring (618). For example, if the arrhythmia episode is an episode of normal sinus rhythm (NSR), premature ventricular contraction (PVC), premature atrial contraction (PAC), or one or more signal artifacts due to noise, the IMD10 stores the features of the detected episode (e.g., mean, median RR, RR variability, mean QRS morphology, QRS morphology deviation) for later reference. In some examples, the electrogram data of the episode is not used for further processing or reporting.

[0077] In response to determining that the arrhythmia dictionary 270 does not include at least one entry (e.g., “NO” at block 606) or in response to determining that the cardiac features obtained via characterization are not similar to any entries in the arrhythmia dictionary 270 (e.g., “NO” at block 614), then IMD10 applies the machine learning model of the machine learning system 250 to the electrocardiogram data to verify the detection of the cardiac arrhythmia episode due to the characterization (608) and / or to detect any other arrhythmias of interest. The operations of step 608 may occur in a manner substantially similar to step 408 of FIG. 4. In response to verifying that the detection of the cardiac arrhythmia episode due to the characterization is correct or that other arrhythmias of interest are present, IMD10 stores the cardiac features in the arrhythmia dictionary 270 (620). For example, IMD10 defines a classification of the arrhythmia episode and creates a new entry in the arrhythmia dictionary 270 that includes one or more cardiac features obtained via characterization that indicate the classification of the arrhythmia episode. Thus, the IMD 10 may use the machine learning system 250 to build, maintain, and update the arrhythmia dictionary 270 for detecting and classifying episodes of cardiac arrhythmia. Furthermore, using the machine learning system 250 only to validate and classify the first instance of each type of arrhythmia may limit the use of the machine learning system 250, thereby conserving power of the IMD 10. In some examples, the machine learning system 250 is only used to validate and classify episodes of cardiac arrhythmia detected in response to an automatic trigger. In some examples, after storing the cardiac features as new entries in the arrhythmia dictionary 270, the IMD 10 generates (610) a report of the arrhythmia and outputs (612) the report to a clinician or monitoring center for review.

[0078] FIG. 7 is a flow chart illustrating an example operation in accordance with the techniques of the present disclosure. For convenience, FIG. 7 is described with respect to FIG. 1. In some examples, the operation of FIG. 7 is for detecting and classifying cardiac arrhythmia in patient 4. The operation of FIG. 7 may provide a tradeoff between battery life and device life for arrhythmia detection accuracy in a patient, for example, if the operation of FIG. 4 for detecting and classifying cardiac arrhythmia does not adequately trigger or detect an arrhythmic episode in patient 4. Although the operation of FIG. 7 is described with respect to IMD 10 of FIG. 1, in other examples, the operation of FIG. 7 may be performed by, for example, external device 12, or a combination of IMD 10 and external device 12.

[0079] In most patients (e.g., between about 80%-90%), the cardiac arrhythmia detection operations described, for example, by FIGS. 4 and 5 improve the accuracy of arrhythmia detection while avoiding unfavorable power consumption. However, in the remaining subpopulation, the operations described, for example, by FIGS. 4 and 5 may not provide high accuracy in arrhythmia detection due to various physiological reasons such as placement of the IMD 10 or patient characteristics (e.g., shifting or moving of the IMD 10 may change the morphology of the cardiac electrogram). For arrhythmia diagnosis, such subpopulation of patients may benefit from a more sensitive and complex arrhythmia detection scheme as opposed to removing the IMD 10 or changing the IMD 10 to another device, even though such a scheme may consume more power (and cause a commensurate decrease in the maximum battery life of the IMD 10). The operations of FIG. 7 illustrate a technique for switching the arrhythmia detection method from, for example, the operations of FIG. 4 or 5, to a more complex machine learning arrhythmia detection scheme.

[0080] As shown in FIG. 7, the IMD 10 senses cardiac electrogram data of the patient 4 (702). The operations of step 702 may occur in a manner substantially similar to step 402 of FIG. 4. The IMD 10 determines whether the IMD 10 is operating in a default cardiac arrhythmia detection mode (e.g., such as the operations of FIG. 4 or 5) (703). Typically, the IMD 10 begins operation in such a default mode. In response to determining that the IMD 10 is operating in the default mode (e.g., “Yes” in block 703), the IMD 10 applies characterization to the cardiac electrogram data to detect one or more cardiac features indicative of an arrhythmia episode (704). In some examples, the IMD 10 acquires additional device and physiological parameters (706). The IMD 10 applies a machine learning model of the machine learning system 250 to the cardiac electrogram data to verify detection of a cardiac arrhythmia episode by the characterization (708). In some examples, IMD 10 applies a machine learning model to the electrocardiogram data to detect one or more episodes of other classes of cardiac arrhythmia not detected by characterization. In some examples, IMD 10 applies a machine learning model to the electrocardiogram data and device and physiological parameters to perform validation or detection. The operations of steps 704 and 708 may occur in a manner substantially similar to steps 404 and 408 of FIG. 4.

[0081] The IMD 10 verifies (709) whether the characterization correctly detected the cardiac arrhythmia episode based on the machine learning model. In response to verifying that the characterization correctly detected the cardiac arrhythmia episode (e.g., “Yes” at block 709), the IMD 10 generates (710) a report of the arrhythmia and outputs (712) the report to a clinician or monitoring center. The operations of steps 704 and 708 may be performed in a manner substantially similar to steps 704 and 708 of FIG. 4. In some examples, the IMD 10 may continue to operate according to a default cardiac arrhythmia detection scheme.

[0082] In response to verifying that the characterization incorrectly detected the cardiac arrhythmia episode (e.g., "NO" at block 709), IMD 10 increments (716) a counter of improper instances of detected cardiac arrhythmia. In some examples, IMD 10 increments the counter if the characterization incorrectly determined that a cardiac arrhythmia episode occurred in patient 4. In some examples, IMD 10 increments the counter if the characterization incorrectly classified a cardiac arrhythmia episode in patient 4 as an episode of a particular type of arrhythmia.

[0083] If the characterization persistently detects episodes of cardiac arrhythmia or episodes that do not include arrhythmia according to the machine learning model, the IMD 10 switches from the default cardiac arrhythmia detection mode to a second mode that uses the machine learning model for arrhythmia detection. For example, the IMD determines whether an inappropriate instance of a detected cardiac arrhythmia counter is greater than a predetermined threshold (718). If the counter is not greater than the predetermined threshold (e.g., “no” at block 718), the IMD 10 continues to operate according to the default cardiac arrhythmia detection scheme. If the counter is greater than the predetermined threshold (e.g., “yes” at block 718), the IMD 10 switches to an alternative arrhythmia detection mode (720). In some examples, the IMD 10 does not switch detection modes until physiological and device characteristics are met. For example, the IMD 10 may postpone switching detection modes until the device is adequately compliant, the patient 4 is at a low or medium activity level, and a patient trigger for arrhythmia is active. In some examples, the IMD 10 turns off machine learning arrhythmia detection during a patient-initiated trigger that captures electrocardiogram data.

[0084] During a subsequent procedure, the IMD 10 senses electrocardiogram data of the patient 4 (702). In response to determining that the IMD 10 is not operating in a default mode (e.g., “No” in block 703), the IMD processes the electrocardiogram data to generate filtered electrocardiogram data (707). The operations of step 707 may occur in a manner substantially similar to step 507 of FIG. 5. The IMD 10 applies a second machine learning model of the machine learning system 250 to the filtered electrocardiogram data to detect episodes of cardiac arrhythmia present in the filtered electrocardiogram data (724). In some examples, the second machine learning model receives as input at least a portion of an electrocardiogram waveform, one or more device parameters, or one or more physiological parameters. The electrocardiogram waveform may be, for example, an ECG waveform.

[0085] Because the second machine learning model is more complex than the cardiac arrhythmia characterization detection, the second machine learning model may adversely affect the power consumption and battery life of the IMD 10. Thus, the IMD 10 updates its battery life estimate (726). In some examples, the IMD 10 may adjust the frequency or periodicity of transmissions to the external device 12. For example, to conserve battery life, the IMD 10 may switch from uploading patient data, such as sensed electrocardiogram data, daily to uploading such data weekly, extending the life of the device. The IMD 10 generates (710) a report of the arrhythmia and outputs (712) the report to a clinician or monitoring center.

[0086] FIG. 8 is a flow chart illustrating example operations in accordance with the techniques of this disclosure. For convenience, FIG. 8 is described in conjunction with FIGS. 1 and 2. In some examples, the operations of FIG. 8 are for detecting and classifying cardiac arrhythmias in patient 4 using an arrhythmia dictionary constructed by machine learning system 250. In the operations of FIG. 8, system 2 combines the accuracy of the machine learning models of machine learning system 250 in detecting cardiac arrhythmias with the low power consumption of the characterization algorithm. Although the operations of FIG. 8 are described in conjunction with IMD 10 of FIG. 1, in other examples, the operations of FIG. 8 may be performed by, for example, external device 12, or a combination of IMD 10 and external device 12.

[0087] In some applications, such as syncope and atrial fibrillation monitoring, the arrhythmia of interest (e.g., pauses, sinus fibrillation, atrial fibrillation) and the corresponding cardiac features of such arrhythmias are known to the clinician. Thus, the techniques of the present disclosure can implement appropriate algorithms on the IMD 10 that use automatic triggering for arrhythmia detection. However, for other types of applications, such as post-myocardial infarction monitoring (e.g., where characteristics such as premature ventricular contraction (PVC) burden, ST segment, etc. may be of interest), or post-transcatheter aortic valve replacement (TAVAR) monitoring, the monitoring may not know the exact arrhythmia or condition of interest. Thus, the operations of FIG. 8 implement a general machine learning model that can be used across multiple applications without developing a specific arrhythmia detection algorithm.

[0088] The operations of Figure 8 may be substantially similar to the operations of Figure 6. However, as discussed in more detail below, the operations of Figure 8 apply a machine learning system to determine whether cardiac features obtained via characterization are similar to cardiac features of an entry in arrhythmia dictionary 270, and to determine whether the cardiac features obtained via characterization are indicative of an arrhythmia episode of the same classification as the classification of the entry in arrhythmia dictionary 270.

[0089] As shown in FIG. 8, the IMD 10 senses electrocardiogram data of the patient 4 (802). The IMD 10 applies characterization to the electrocardiogram data to detect one or more cardiac features (804). The operations of steps 802 and 84 may occur in a manner substantially similar to steps 402 and 404 of FIG. 4, respectively. In some examples, the characterization performed by the IMD 10 is tuned for high sensitivity. For example, the IMD 10 may implement a sensitive auto-trigger algorithm to identify episodes of interest. In some examples, the characterization is configured to classify segments of the electrocardiogram of the patient 4 as arrhythmia episodes of interest where the segments exhibit RR rates of 90 beats per minute or more or less than 60 beats per minute, RR fluctuations of more than 50 milliseconds, or morphological fluctuations in the absence of noise. In some examples, the characterization is configured to classify segments of the electrocardiogram of the patient 4 as arrhythmia episodes of interest where the segments are non-noise, non-NSR duration.

[0090] The IMD 10 determines (806) whether the arrhythmia dictionary 270 includes at least one entry. Each entry in the arrhythmia dictionary 270 includes a classification that defines an arrhythmia episode as a particular type of arrhythmia, such as an episode of bradycardia, tachycardia, atrial fibrillation, ventricular fibrillation, or atrioventricular fibrillation block. Each entry further includes one or more cardiac features associated with a cardiac arrhythmia episode of the corresponding classification.

[0091] In response to determining that arrhythmia dictionary 270 includes at least one entry (e.g., “yes” in block 806), IMD10 applies machine learning system 250 to determine (816) whether the cardiac features obtained via characterization are similar to the cardiac features of an entry in arrhythmia dictionary 270. In some examples, machine learning system 250 determines whether the cardiac features obtained via characterization are similar to the cardiac features of an entry in arrhythmia dictionary 270. In some examples, machine learning system 250 may not determine a particular type of arrhythmia, but detects noise or changes in the cardiac electrogram other than the NSR or an entry in arrhythmia dictionary 270.

[0092] In response to determining that the cardiac feature obtained via characterization is similar to the cardiac feature of the entry in the arrhythmia dictionary 270 (e.g., “Yes” at block 814), the IMD 10 determines that the cardiac feature obtained via characterization is an arrhythmia episode of the same classification as the classification of the entry in the arrhythmia dictionary 270. In some examples, the IMD 10 updates (818) an arrhythmia episode counter of the classification of the entry in the arrhythmia dictionary 270 to track the total number of arrhythmia episodes of that classification that have occurred in the patient 4. The IMD 10 stores (820) the value of the counter with the entry in the arrhythmia dictionary 270. After storing the value of the counter in the arrhythmia dictionary 270, the IMD 10 generates (810) a report of the arrhythmia and outputs (812) the report to a clinician or monitoring center for review. The operations of steps 810 and 812 may be performed in a manner substantially similar to steps 610 and 612 of FIG. 6, respectively.

[0093] In response to determining that the arrhythmia dictionary 270 does not include at least one entry (e.g., “no” at block 806) or in response to determining that the cardiac features obtained via characterization are not similar to any entries in the arrhythmia dictionary 270 (e.g., “no” at block 814), IMD10 then applies a machine learning model of the machine learning system 250 to the electrocardiogram data to characterize the cardiac arrhythmia episode (808). For example, IMD10 applies the machine learning model to verify that a cardiac arrhythmia episode has occurred and / or to determine a classification of the cardiac arrhythmia episode. In some examples, the machine learning model is a deep learning model that characterizes the episode with “deep features” (e.g., the state of the final deep learning arrhythmia detection model before softmax classification).

[0094] The IMD 10 stores (820) the cardiac features in the arrhythmia dictionary 270 as a new entry in the arrhythmia dictionary 270 along with the value of a counter of the number of occurrences of the cardiac arrhythmia type (e.g., "1" for the corresponding first occurrence to the new entry). In some examples, the IMD 10 stores deep features of the newly detected arrhythmia episode in the entry. After creating the new entry and storing the value of the counter in the arrhythmia dictionary 270, the IMD 10 generates (810) a report of the arrhythmia and outputs (812) the report to a clinician or monitoring center for review. For example, the IMD 10 transmits the cardiac electrogram data corresponding to the detected episode for expert review and validation by a clinician. The review is performed by an expert who can look beyond typical arrhythmias for changes in the signal related to other patient features such as changes in medications.

[0095] The following examples may illustrate one or more aspects of the disclosure.

[0096] Example 1. 11. A method comprising: sensing, by a medical device including a processing circuit and a storage medium, electrocardiogram data of a patient; performing, by the medical device, feature-based portrayal of the sensed electrocardiogram data to obtain cardiac features present in the electrocardiogram data and indicative of an arrhythmia episode of the patient; determining, by the medical device and based on the feature-based portrayal, that the cardiac features satisfy threshold criteria for application of a machine learning model to verify that an arrhythmia episode has occurred in the patient; applying, by the medical device, a machine learning model trained using electrocardiogram data of a plurality of patients to the sensed electrocardiogram data to verify, based on the machine learning model, that an arrhythmia episode has occurred in the patient; and generating, by the medical device, a report including an indication that an arrhythmia episode has occurred in the patient and one or more cardiac features consistent with the arrhythmia episode in response to verifying by the machine learning model that an arrhythmia episode has occurred in the patient; and outputting, by the medical device, the report for display including an indication that an arrhythmia episode has occurred in the patient and one or more cardiac features consistent with the arrhythmia episode.

[0097] Example 2. The method of example 1, wherein the arrhythmia episode in the patient is a first classification arrhythmia episode in the patient, and applying a machine learning model to the sensed cardiac electrogram data to verify that an arrhythmia episode occurred in the patient, the method further comprising, in response to determining that a first classification arrhythmia episode occurred in the patient based on the feature-based representation, applying a machine learning model to the sensed cardiac electrogram data to determine that a second classification arrhythmia episode occurred in the patient, to generate a report including an indication that an arrhythmia episode occurred in the patient and one or more of the cardiac features consistent with an arrhythmia episode, the report including an indication that a first classification arrhythmia episode occurred in the patient, an indication that a second classification arrhythmia episode occurred in the patient, and one or more of the cardiac features consistent with an arrhythmia episode of the first classification.

[0098] Example 3. The method of any of Examples 1 or 2, wherein the electrocardiogram data comprises an electrocardiogram (ECG) of the patient.

[0099] Example 4. The method of any of Examples 1-3, wherein performing feature-based representation of the electrocardiogram data to obtain cardiac features present in the electrocardiogram data includes performing at least one of QRS detection, refractory processing, noise processing, or representation of the electrocardiogram data to obtain cardiac features present in the electrocardiogram data.

[0100] Example 5. The method of any of Examples 1-4, comprising applying a machine learning model to verify that an episode of arrhythmia has occurred in the patient, and applying a machine learning model to verify that an episode of at least one of bradycardia, tachycardia, atrial fibrillation, ventricular fibrillation, or atrioventricular block has occurred in the patient.

[0101] Example 6. The method of any of Examples 1-5, wherein the cardiac features present in the electrocardiogram data are one or more of an RR interval present in the electrocardiogram data, an average heart rate present in the electrocardiogram data, a minimum heart rate present in the electrocardiogram data, a maximum heart rate present in the electrocardiogram data, a PR interval present in the electrocardiogram data, a heart rate variability present in the electrocardiogram data, one or more amplitudes of one or more features of an electrocardiogram (ECG), or an interval between one or more features of an ECG.

[0102] Example 7. The method of any of Examples 1-6, wherein the machine learning model trained using electrocardiogram data of a plurality of patients comprises a machine learning model trained using a plurality of electrocardiogram (ECG) waveforms, each ECG waveform labeled with an episode of one or more classifications in a patient of the plurality of patients.

[0103] Example 8. The method of any of Examples 1-7, comprising determining that at least one of a physiological parameter of the patient or a parameter of the medical device meets a threshold criterion to determine that the cardiac feature meets the threshold criterion.

[0104] Example 9. The method of any of Examples 1-8, wherein applying the machine learning model to the sensed electrocardiogram data further includes applying the machine learning model to at least one of one or more features of the sensed electrocardiogram data that correlate to an arrhythmia in the patient, an activity level of the medical device, an input impedance of the medical equipment, or a battery level of the medical device.

[0105] Example 10. The method of any of Examples 1-9, wherein applying the machine learning model to the sensed electrocardiogram data includes applying the machine learning model to the sensed electrocardiogram data in response to determining that the cardiac features meet a threshold criterion and determining that a noise in at least one of the cardiac features is below a predetermined threshold.

[0106] Example 11. The method of any of Examples 1-10, wherein applying the machine learning model to the sensed electrocardiogram data includes applying the machine learning model to the sensed electrocardiogram data in response to determining that the cardiac feature satisfies a threshold criterion and determining that the patient is in a first postural state of the plurality of postural states.

[0107] Example 12. The method of any of Examples 1-11, wherein applying the machine learning model to the sensed electrocardiogram data includes applying the machine learning model to the sensed electrocardiogram data in response to determining that the cardiac feature meets a threshold criterion and determining that the patient is in a first activity-state of the plurality of activity-states.

[0108] Example 13. The method of any of Examples 1-12, the method including receiving, from the medical device and a user, adjustments to the feature-based representation of the electrocardiogram data in response to outputting a report comprising an indication that an arrhythmia episode has occurred in the patient and one or more cardiac features consistent with the arrhythmia episode, and performing, in response to the adjustments, the feature-based representation of the electrocardiogram data to obtain a second cardiac feature present in the electrocardiogram data.

[0109] Example 14. The method of any of Examples 1-13, wherein the patient's electrocardiogram data includes an electrocardiogram (ECG) of the patient, and the method includes identifying a subsection of the patient's ECG, the subsection including ECG data for a first time period before the arrhythmia episode, a second time period during the arrhythmia episode, and a third time period after the arrhythmia episode, the subsection including a length of time of the patient's ECG being greater than the first, second, and third time periods, identifying the one or more of the cardiac features consistent with the first, second, and third time periods, and including the subsection of the ECG and the one or more of the cardiac features consistent with the first, second, and third time periods in the report to generate a report including an indication that an arrhythmia episode has occurred in the patient and one or more of the cardiac features consistent with the arrhythmia episode.

[0110] Example 15. The method of any of Examples 1-14, wherein determining based on the feature-based representation that the cardiac feature meets a threshold criterion for application of a machine learning model to verify that an arrhythmia episode has occurred in the patient includes determining based on the feature-based representation that the cardiac feature is indicative of a first classification of arrhythmia episode has occurred in the patient, and applying the machine learning model to the sensed cardiac electrogram data to verify that an arrhythmia episode has occurred in the patient includes applying the machine learning model to the sensed cardiac electrogram data to verify a determination based on the feature-based representation that the cardiac feature is indicative of a first classification of arrhythmia episode.

[0111] Example 16. The method of any of Examples 1-15, further comprising determining a burden of arrhythmic episodes in the patient in response to verifying that an arrhythmic episode has occurred in the patient, wherein the burden of arrhythmic episodes in the patient comprises a ratio of a length of time of the arrhythmic episode to a length of time during which the patient is being monitored by the medical device.

[0112] Example 17. The method of any of Examples 1-16, wherein the method further includes processing, by the medical device, the sensed electrocardiogram data to generate filtered electrocardiogram data, and wherein applying the machine learning model to the sensed electrocardiogram data to verify that an arrhythmia episode has occurred in the patient includes applying the machine learning model to the filtered electrocardiogram data to verify that an arrhythmia episode has occurred in the patient.

[0113] Example 18. The method of example 17, wherein processing the sensed electrocardiogram data to generate the filtered electrocardiogram data includes determining at least one of a period of high level of patient activity or a period of highly varying input impedance of the medical device, and in response to determining at least one of the period of high level of patient activity or the period of highly varying input impedance of the medical device, discarding at least a portion of the sensed electrocardiogram data that coincides with at least one of the period of high level of patient activity or the period of highly varying input impedance of the medical device to generate the filtered electrocardiogram data.

[0114] Example 19. The method of example 18, wherein discarding at least a portion of the sensed electrocardiogram data that coincides with at least one of a period of high level patient activity or a period of highly variable input impedance includes discarding one or more of a sensed input impedance of the medical device, a sensed activity level of the patient, or a sensed postural change of the patient that coincides with at least one of a period of highly variable input impedance of the medical device.

[0115] Example 20. The method according to any one of Examples 1 to 19, further comprising: applying a machine learning model to the sensed electrocardiogram data to verify that an arrhythmia episode has occurred in the patient, based on the machine learning model, a first determination that an arrhythmia episode has not occurred in the patient, or a second determination that a different type of arrhythmia episode has occurred in the patient, based on the machine learning model, and in response to at least one of the first determination and the second determination, updating, by the medical device, a counter of false detection of an arrhythmia episode in the patient, and in response to determining that the value of the counter is greater than a predetermined threshold, switching from performing feature-based representation of the sensed electrocardiogram data to acquiring cardiac features present in the electrocardiogram data and indicative of the arrhythmia episode in the patient, and applying a second machine learning model trained using electrocardiogram data of a plurality of patients to the sensed electrocardiogram data to acquire cardiac features present in the electrocardiogram data and indicative of the arrhythmia episode in the patient, based on the machine learning model.

[0116] Example 21. The method of example 20, wherein sensing the patient's electrocardiogram data includes sensing first electrocardiogram data of the patient, the method including: sensing, by the medical device, second electrocardiogram data of the patient; applying a second machine learning model to the sensed second electrocardiogram data to obtain, based on the machine learning model, second cardiac features present in the second electrocardiogram data and indicative of a second arrhythmia episode in the patient; in response to obtaining, by the second machine learning model, the second cardiac features indicative of the second arrhythmia episode in the patient, generating, by the medical device, a second report including an indication that the second arrhythmia episode has occurred in the patient and one or more of the second cardiac features consistent with the second arrhythmia episode; outputting, by the medical device, for display, the second report including an indication that the second arrhythmia episode has occurred in the patient and one or more of the second cardiac features consistent with the second arrhythmia episode.

[0117] Example 22. The method of any of Examples 21 or 22, further comprising updating, by the medical device, an estimate of power consumption by the medical device in response to determining that the value of the counter is greater than a predetermined threshold, and outputting, by the medical device and for display, the estimate of power consumption by the medical device.

[0118] Example 23. The method of any of Examples 1-22, further comprising processing, by the medical device, the sensed electrocardiogram data to generate an intermediate representation of the sensed electrocardiogram data, and applying the machine learning model trained using the electrocardiogram data of the plurality of patients to the sensed electrocardiogram data to verify that an arrhythmia episode has occurred in the patient comprises applying the machine learning model trained using the intermediate representations of the electrocardiogram data of the plurality of patients to the intermediate representations of the sensed electrocardiogram data and cardiac features present in the electrocardiogram data to verify that an arrhythmia episode has occurred in the patient.

[0119] Example 24. The method of example 23, wherein processing the sensed electrocardiogram data to generate the intermediate representation of the sensed electrocardiogram data includes at least one of applying a filter to the sensed electrocardiogram data and performing signal decomposition on the sensed electrocardiogram data.

[0120] Example 25. The method of example 24, wherein performing signal decomposition on the sensed electrocardiogram data includes performing wavelet decomposition on the sensed electrocardiogram data.

[0121] Example 26. A method, comprising: sensing electrocardiogram data of a patient by a medical device including a processing circuit and a storage medium; performing, by the medical device, feature-based representation of the sensed electrocardiogram data to obtain cardiac features present in the electrocardiogram data; determining, by the medical device, a similarity between the obtained cardiac features and each of a plurality of entries of an arrhythmia dictionary of the medical device; each of the plurality of entries of the arrhythmia dictionary includes a cardiac feature that reveals a classification of an arrhythmia with a plurality of classifications of arrhythmia in the patient, and in response to determining that the obtained cardiac feature is not similar to each of the plurality of entries of the arrhythmia dictionary, determining, by the medical device, that an arrhythmia episode has occurred in the patient based on a machine learning model, applying the machine learning model trained using the electrocardiogram data of the plurality of patients to the sensed electrocardiogram data; and storing, by the medical device, a first entry in the arrhythmia dictionary including a first classification of the arrhythmia episode and the obtained cardiac feature.

[0122] Example 27. The method of example 26, further comprising, after storing an entry including the first classification of the arrhythmia episode and the acquired cardiac features, generating, by the medical device, a report including an indication that the patient has experienced an arrhythmia episode of the first classification and the one or more acquired cardiac features consistent with the arrhythmia episode of the first classification, and outputting, by the medical device and for display, the report including an indication that the patient has experienced an arrhythmia episode of the first classification and the one or more acquired cardiac features consistent with the arrhythmia episode of the first classification.

[0123] Example 28. The method of any of Examples 26 or 27, comprising: sensing, by a medical device including a processing circuit and a storage medium, second electrocardiogram data of the patient; performing, by the medical device, feature-based representation of the sensed second electrocardiogram data to obtain second cardiac features present in the second electrocardiogram data; determining, by the medical device, that the second cardiac feature is similar to a cardiac feature of a first entry of a plurality of entries of an arrhythmia dictionary of the medical device; in response to determining that the second cardiac feature is similar to the cardiac feature of the first entry of the plurality of entries of the arrhythmia dictionary, generating, by the medical device, a report including an indication of an arrhythmia episode of the first classification and one or more of the second cardiac features consistent with the arrhythmia episode of the first classification; and outputting, by the medical device and for display, the report including an indication that an arrhythmia episode of the first classification has occurred in the patient and one or more of the second cardiac features consistent with the arrhythmia episode of the first classification.

[0124] Example 29. The method of any of Examples 26-28, wherein determining a similarity between the acquired cardiac feature and the cardiac feature of each entry of the plurality of entries of the arrhythmia dictionary of the medical device includes determining that an L1 distance of the acquired cardiac feature is not similar to an L1 distance of the cardiac feature of each entry of the plurality of entries of the arrhythmia dictionary of the medical device.

[0125] Example 30. The method of any of Examples 26-28, wherein determining the similarity between the acquired cardiac features and the cardiac features of each entry of the plurality of entries of the arrhythmia dictionary of the medical device includes determining that a difference between at least one parameter of the acquired cardiac features and at least one parameter of the cardiac features of each entry of the plurality of entries of the arrhythmia dictionary of the medical device is greater than a predetermined threshold.

[0126] Example 31. The method of any of Examples 26-28, wherein determining the similarity of the acquired cardiac features to the cardiac features of each entry of the plurality of entries of the arrhythmia dictionary of the medical device includes applying a second machine learning model, trained using electrocardiogram data for the plurality of patients' electrocardiogram data, to the acquired cardiac features and the cardiac features of each entry of the plurality of entries of the arrhythmia dictionary of the medical device to determine that the acquired cardiac features are not similar to the cardiac features of each entry of the plurality of entries of the arrhythmia dictionary of the medical device.

[0127] Example 32. The method of any of Examples 26-31, further comprising: in response to determining that the acquired cardiac features are similar to the cardiac features of a first entry of the plurality of entries in the arrhythmia dictionary of the medical device, updating, by the medical device, a counter of arrhythmia episodes of a first classification defined by the first entry of the plurality of entries in the arrhythmia dictionary of the medical device; and storing, by the medical device, the counter in the first entry of the plurality of entries in the arrhythmia dictionary of the medical device.

[0128] Example 33. The method according to any of Examples 26-32, further comprising: determining, by the medical device, that the acquired cardiac features are not similar to the patient's normal sinus rhythm (NSR) or noise; and in response to determining that the acquired cardiac features are not similar to the cardiac features of each entry of the multiple entries of the arrhythmia dictionary, applying a machine learning model to the sensed electrocardiogram data to determine that an episode of a first class arrhythmia has occurred in the patient, comprising: in response to determining that the acquired cardiac features are not similar to the cardiac features of each entry of the multiple entries of the arrhythmia dictionary and in response to determining that the acquired cardiac features are not similar to the patient's NSR, applying a machine learning model to the sensed electrocardiogram data to determine that an episode of a first class arrhythmia has occurred in the patient.

[0129] In some examples, the disclosed techniques include a system including means for performing any of the methods described herein. In some examples, the disclosed techniques include a computer-readable medium including instructions that cause a processing circuit to perform any of the methods described herein.

[0130] It should be understood that the various aspects disclosed herein can be combined in different combinations than those specifically presented in the description and accompanying drawings. It should also be understood that certain acts or events of any of the processes or methods described herein may be performed in different orders, or may be added, combined, or omitted entirely, depending on the embodiment (e.g., not all described acts or events may be necessary to perform the technology). Furthermore, although certain aspects of the disclosure are described for clarity as being performed by a single module, unit, or circuit, it should be understood that the technology of the disclosure may be performed by a combination of units, modules, or circuits associated with, for example, a medical device device.

[0131] In one or more examples, the described techniques can be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions can be stored as one or more instructions or codes on a computer-readable medium and executed by a hardware-based processing unit. The computer-readable medium can include non-transitory computer-readable media corresponding to tangible media, such as data storage media (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).

[0132] 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 terms "processor" or "processing circuitry" as used herein may refer to any of the foregoing structures, or any other physical structure suitable for implementing the techniques described. Also, the techniques may be implemented entirely in one or more circuits or logic elements.

Claims

1. 1. A medical device comprising: A storage medium; a processing circuit operably coupled to the storage medium; and Sensing electrocardiogram data of a patient; performing feature-based representation of the sensed electrocardiogram data to obtain cardiac features present in the electrocardiogram data and indicative of an arrhythmia episode in the patient; determining, based on the feature-based description, that the cardiac features meet a threshold criterion for application of a machine learning model to verify that the arrhythmia episode has occurred in the patient; responsive to determining that the cardiac features satisfy the threshold criteria for application of the machine learning model, applying the machine learning model, trained using electrocardiogram data of a plurality of patients, to the sensed electrocardiogram data to verify, based on the machine learning model, that the arrhythmia episode occurred in the patient; generating a report in response to verifying by the machine learning model that the arrhythmia episode occurred in the patient, the report including an indication that the arrhythmia episode occurred in the patient and one or more of the cardiac features consistent with the arrhythmia episode; generating a report regarding the arrhythmia episode in the patient without using the machine learning model in response to determining that the cardiac features do not meet the threshold criteria for application of the machine learning model; outputting said report for display; and a processing circuit configured to: applying the machine learning model to the sensed electrocardiogram data to verify that the arrhythmia episode has occurred in the patient includes at least one of: a first determination based on the machine learning model that the arrhythmia episode has not occurred in the patient; or a second determination based on the machine learning model that a different type of arrhythmia episode has occurred in the patient; updating, by the medical device, a counter of false detections of arrhythmia episodes in the patient in response to the at least one of the first determination and the second determination; responsive to determining that the value of the counter is greater than a predetermined threshold, switch from performing the feature-based representation of the sensed electrocardiogram data to obtaining cardiac features present in the electrocardiogram data and indicative of an arrhythmia episode in the patient, and apply a second machine learning model, trained using electrocardiogram data of a plurality of patients, to the sensed electrocardiogram data to obtain cardiac features present in the electrocardiogram data and indicative of an arrhythmia episode in the patient based on the machine learning model.

1. A medical device comprising:

2. 1. A medical device comprising: A storage medium; a processing circuit operably coupled to the storage medium; and Sensing electrocardiogram data of a patient; performing feature-based representation of the sensed electrocardiogram data to obtain cardiac features present in the electrocardiogram data and indicative of an arrhythmia episode in the patient; determining, based on the feature-based description, that the cardiac features meet a threshold criterion for application of a machine learning model to verify that the arrhythmia episode has occurred in the patient; responsive to determining that the cardiac features satisfy the threshold criteria for application of the machine learning model, applying the machine learning model, trained using electrocardiogram data of a plurality of patients, to the sensed electrocardiogram data to verify, based on the machine learning model, that the arrhythmia episode occurred in the patient; In response to verifying by the machine learning model that the arrhythmia episode has occurred in the patient, generating a report including an indication that the arrhythmia episode has occurred in the patient and one or more of the cardiac features consistent with the arrhythmia episode; and outputting, for display, the report including an indication that the arrhythmia episode has occurred in the patient and one or more of the cardiac features consistent with the arrhythmia episode; the arrhythmic episode in the patient is a first classification arrhythmic episode in the patient; to apply the machine learning model to the sensed electrocardiogram data to verify that the arrhythmia episode occurred in the patient, the processing circuitry is configured to apply the machine learning model to the sensed electrocardiogram data to verify that the arrhythmia episode of the first classification occurred in the patient; the processing circuitry is further configured, in response to determining based on the feature-based description that an arrhythmia episode of the first classification has occurred in the patient, to apply the machine learning model to the sensed electrocardiogram data to determine that an arrhythmia episode of a second classification has occurred in the patient; and one or more of the cardiac features consistent with the arrhythmia episode, wherein the processing circuit is configured to generate a report including an indication that the arrhythmia episode of the first classification has occurred in the patient, an indication that the arrhythmia episode of the second classification has occurred in the patient, and the one or more of the cardiac features consistent with the arrhythmia episode of the first classification.

3. 3. The medical device of claim 1 or 2, wherein the processing circuitry is configured to perform at least one of QRS detection, refractory processing, noise processing, or rendering of the electrocardiogram data to obtain cardiac features present in the electrocardiogram data, in order to perform feature-based rendering of the electrocardiogram data to obtain cardiac features present in the electrocardiogram data.

4. The machine learning model is applied to determine whether the arrhythmia episode occurred in the patient.

4. The medical device of claim 1, wherein the processing circuitry is configured to apply the machine learning model to verify that an episode of at least one of bradycardia, tachycardia, atrial fibrillation, ventricular fibrillation, or atrioventricular block has occurred in the patient ....

5. 5. The medical device of claim 1, wherein the machine learning model trained using electrocardiogram data of the plurality of patients comprises a machine learning model trained using a plurality of electrocardiogram (ECG) waveforms, each ECG waveform labeled with one or more arrhythmia episodes of one or more classes in a patient of the plurality of patients.

6. The medical device of any of claims 1 to 5, wherein to determine that the cardiac feature satisfies the threshold criterion, the processing circuit is configured to determine that at least one of a physiological parameter of the patient or a parameter of the medical device satisfies the threshold criterion.

7. The medical device of any one of claims 1 to 6, wherein the processing circuitry is configured to apply the machine learning model to the sensed electrocardiogram data in response to determining that the cardiac features satisfy the threshold criterion and determining that noise in at least one of the cardiac features is below a predetermined threshold.

8. 8. The medical device of claim 1, wherein the processing circuitry is configured to apply the machine learning model to the sensed electrocardiogram data in response to determining that the cardiac features meet the threshold criteria and determining that the patient is in a first posture state of a plurality of posture states, said processing circuitry being configured to apply the machine learning model to the sensed electrocardiogram data in response to determining that the cardiac features meet the threshold criteria and determining that the patient is in a first posture state of a plurality of posture states.

9. the electrocardiogram data of the patient includes an electrocardiogram (ECG) of the patient; to generate a report including the indication that the arrhythmia episode has occurred in the patient and the one or more of the cardiac features consistent with the arrhythmia episode, identifying a subsection of the ECG of the patient, the subsection including ECG data for a first time period prior to the arrhythmic episode, a second time period during the arrhythmic episode, and a third time period after the arrhythmic episode, the length of time of the ECG of the patient being greater than the first, second, and third time periods; identifying one or more of the cardiac features consistent with the first, second, and third time periods; and including in the report the subsection of the ECG and the one or more of the cardiac features that coincide with the first, second, and third time periods.

10. the processing circuitry is configured to determine, based on the feature-based representation, that the cardiac features are indicative that a first classification of arrhythmia episode has occurred in the patient, to determine, based on the feature-based representation, that the cardiac features satisfy a threshold criterion for application of the machine learning model to verify that the arrhythmia episode has occurred in the patient; 10. The medical device of claim 1, wherein the processing circuitry is configured to apply the machine learning model to the sensed electrocardiogram data to verify that the arrhythmia episode has occurred in the patient, the processing circuitry being configured to apply the machine learning model to the sensed electrocardiogram data to verify the determination that the cardiac features are indicative of the arrhythmia episode of the first classification based on the feature-based representation.

11. the processing circuitry is further configured to process the sensed electrocardiogram data to generate filtered electrocardiogram data; 11. The medical device of claim 1, wherein the processing circuitry is configured to apply the machine learning model to the filtered electrocardiogram data to verify that the arrhythmia episode has occurred in the patient, and to apply the machine learning model to the sensed electrocardiogram data to verify that the arrhythmia episode has occurred in the patient.

12. 1. A computing device comprising: means for sensing electrocardiogram data of a patient; means for performing feature based representation of the sensed electrocardiogram data to obtain cardiac features present in the electrocardiogram data and indicative of an arrhythmia episode in the patient; means for determining, based on the feature-based description, that the cardiac features meet a threshold criterion for application of a machine learning model to verify that the arrhythmia episode has occurred in the patient; means for applying the machine learning model, trained using electrocardiogram data of a plurality of patients, to the sensed electrocardiogram data in response to determining that the cardiac feature satisfies the threshold criteria for application of the machine learning model to verify, based on the machine learning model, that the arrhythmia episode occurred in the patient; a means for generating a report in response to verifying by the machine learning model that the arrhythmia episode occurred in the patient, the report including an indication that the arrhythmia episode occurred in the patient and one or more of the cardiac features consistent with the arrhythmia episode; responsive to determining that the cardiac features do not meet the threshold criteria for application of the machine learning model, generating a report regarding the arrhythmia in the patient without using the machine learning model; means for outputting the report for display; applying the machine learning model to the sensed electrocardiogram data to verify that the arrhythmia episode has occurred in the patient includes at least one of: a first determination based on the machine learning model that the arrhythmia episode has not occurred in the patient; or a second determination based on the machine learning model that a different type of arrhythmia episode has occurred in the patient; and the computing device further comprises: means for updating, by the computing device, a counter of false detections of arrhythmia episodes in the patient in response to the at least one of the first determination and the second determination; responsive to determining that a value of the counter is greater than a predetermined threshold, means for switching from performing the feature based representation of the sensed electrocardiogram data to acquiring cardiac features present in the electrocardiogram data that are indicative of an arrhythmia episode in the patient, means configured to apply a second machine learning model, trained using electrocardiogram data of a plurality of patients, to the sensed electrocardiogram data to obtain, based on the machine learning model, cardiac features present in the electrocardiogram data and indicative of an arrhythmia episode in the patient; A computing device may include:

13. A system comprising a medical device according to any one of claims 1 to 11 or a computing device according to claim 12.

Images