Pre-ventricular constriction system

Abstract

Systems and methods for detecting and classifying preventricular contractions (PVCs) are discussed. An exemplary system includes a sensor circuit to sense a cardiac signal of a subject, and a processor circuit to detect heartbeat from the cardiac signal and to detect PVC in a first detection mode or a different second detection mode. The first detection pattern includes a pattern in which a particular type of PVC candidate is identified using a cardiac interval or signal amplitude of the detected heartbeat, and the identified PVC candidate is classified as a PVC single shot or a plurality of consecutive PVC using a signal feature including a morphological feature of the sensed cardiac signal. The second detection mode includes detecting a particular type of PVC using the heart interval or signal amplitude of the detected heartbeat. In response to a mode switch trigger event, the first detection mode may be transitioned to a second detection mode.

Description

[0001] Related Applications

[0002] This application claims the benefit of U.S. Provisional Application No. 63 / 359,338, filed on July 8, 2022, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This document relates generally to medical devices, and more particularly to systems, devices, and methods for detecting various types of premature ventricular contractions (PVCs). Background Art

[0004] Implantable medical devices (IMDs) have been used to monitor patient health or disease states and deliver treatments. For example, an implantable cardioverter-defibrillator (ICD) can be used to monitor certain abnormal heart rhythms and deliver electrical energy to the heart to correct the abnormal rhythm. Some IMDs can be used to monitor chronic deterioration of cardiac hemodynamic performance, such as due to congestive heart failure (CHF), and provide cardiac stimulation therapy, including cardiac resynchronization therapy (CRT), to correct cardiac dyssynchrony within or between ventricles.

[0005] Some IMDs can detect arrhythmias. One common arrhythmia is premature ventricular contractions (PVCs), also called premature ventricular beats or ventricular premature contractions. A PVC is a premature heartbeat that originates in the ventricles and disrupts the normal rhythm of the heart. A typical PVC pattern includes a normal beat, an extra beat (PVC), a slight pause, and then a stronger-than-normal beat, although other PVC patterns can also be clinically identified in patients. During the pause after the PVC, the heart fills with more blood, giving the next beat extra strength. PVCs may occur randomly or regularly, such as at regular intervals. PVCs may be concentrated during periods of high density, or they may be evenly distributed throughout the day.

[0006] PVCs can occur alone or in a repetitive pattern. When a PVC occurs as a single premature beat (also called a single PVC or PVC single), the patient may describe the sensation as "palpitations" or "missed beats." The beats following the PVC may be strong enough to cause chest pain or discomfort. The most common single PVC pattern may include a long compensatory pause after the PVC. If the PVC continues to alternate with regular sinus beats, the patient has a bigeminy. Similarly, if one out of every three heartbeats is a PVC, it is called a trigeminy. Some PVCs may appear in a PVC pair pattern, characterized by two consecutive PVCs. The two PVCs in the PVC pair may come from the same ectopic lesion (single lesion PVC), or from two or more ectopic lesions (multi-lesion PVCs). A pattern of three or more (but less than a threshold number, such as thirty in one example) consecutive PVCs is called a PVC burst or non-sustained ventricular tachycardia (NSVT).

[0007] PVCs can be associated with a variety of cardiac or non-cardiac conditions. Examples include cardiomyopathy, mitral valve prolapse, and myocardial infarction. Any structural heart disease that alters the conduction pathway due to tissue changes can cause PVCs. Examples of non-cardiac conditions include hyperthyroidism, anemia, and even hypertension. Patient populations with a higher risk of cardiovascular disease and poor clinical cardiovascular markers have a higher incidence of PVCs.

[0008] Individuals with frequent PVCs or a series of PVCs may experience a fluttering sensation in the chest or neck. If PVCs are frequent enough to reduce the heart's ability to pump blood, the patient may experience weakness, dizziness, or fainting. PVC burden is a quantitative measure of a patient's PVC frequency and can be calculated as the percentage of all ventricular beats measured over a specified time period (usually 24 hours) that are PVCs of any type (single, doublet, or NSVT). While a PVC burden of less than ~5% is generally considered benign, a PVC burden of more than 10-20% may be clinically actionable. Accurate and efficient detection of various types of PVCs can help improve PVC burden estimation and appropriate management of patients with arrhythmias. Summary of the invention

[0009] Timely detection of PVCs and identification of PVC types or patterns can have important clinical significance for accurate estimation of PVC burden and clinical assessment of cardiac function in patients. Traditional PVC detection and PVC burden estimation focus on detecting single PVCs (PVC single shots) based on characteristic interval patterns, such as a shorter (shorter than the cardiac cycle during normal sinus rhythm) pre-PVC interval followed by a prolonged compensatory interval. This traditional approach has several disadvantages. First, although most single PVCs are characterized by a prolonged post-PVC compensatory interval, in some patients or under certain circumstances, single PVCs may not be accompanied by a compensatory interval, especially when the PVC is fast enough to proceed to the next sinus beat normally, or during certain atrial arrhythmias (such as atrial fibrillation). Detection that relies on prolonged compensatory intervals may miss some single PVCs and reduce the accuracy of PVC burden estimation. Second, interval-based PVC detection algorithms may produce false detections (including false positive or false negative detections), which may also reduce the accuracy of PVC burden estimation. Although more complex features such as signal morphology can be used to detect single PVCs, detection algorithms using these features generally require more computing resources. In resource- and power-conservative environments, such as real-time PVC detection via battery-powered implantable devices, morphological analysis of every heartbeat may not be feasible. Third, conventional PVC detection techniques focus on detecting single PVCs and are typically not designed or optimized for detecting patterns of other PVC types, such as PVC pairs or bursts of three or more PVCs. Although the same single PVC detection algorithm can be used to detect other PVC patterns, this approach may be less efficient and is typically computationally demanding, especially in real-time implementations. More importantly, because single PVCs and PVC pairs or PVC bursts typically have different characteristics (e.g., interval patterns such as compensatory pauses, PVC morphology, etc.), algorithms designed or optimized for detecting single PVCs may result in false-positive or false-negative detection of PVC beats in pairs or PVC bursts, thereby reducing the accuracy of PVC burden estimation.

[0010] Among other things, this document discusses systems, devices, and methods for detecting and classifying various types of PVCs. An exemplary system includes a sensor circuit for sensing a cardiac signal of a subject, and a processor circuit for detecting heartbeats from the sensed cardiac signal and detecting PVCs in a first detection mode or a different second detection mode. In the first detection mode, the processor circuit may use a cardiac interval or signal amplitude of the detected heartbeat to identify a particular type of PVC candidate. The PVC candidate may be one of a PVC single shot candidate, a PVC pair candidate, or a PVC continuous shot candidate. The processor circuit may generate a signal feature set including morphological features from the sensed cardiac signal, and use the generated signal feature set to classify the identified PVC candidate as a PVC single shot or a pattern of multiple continuous PVCs. In the second detection mode, the processor circuit may use a cardiac interval or signal amplitude of the detected heartbeat to detect a particular type of PVC. In response to a mode switching trigger event, the first detection mode may be converted to a second detection mode. The processor circuit may determine the PVC load based on the detected PVC.

[0011] Example 1 is an implantable medical device for detecting cardiac arrhythmias in a subject, comprising: a sensor circuit configured to sense a cardiac signal of the subject; and a processor circuit configured to: detect heartbeats from the sensed cardiac signals; detect premature ventricular contractions (PVCs) in a first detection mode, comprising: using the cardiac interval or signal amplitude of the detected heartbeats to identify a specific type of PVC candidate; generating a signal feature set including morphological features from the sensed cardiac signal; and using the generated signal feature set to classify the identified PVC candidates into a single PVC or a pattern of multiple consecutive PVCs; and in response to a mode switching trigger event, detecting the specific type of PVC using the cardiac interval or signal amplitude of the detected heartbeats in a second detection mode different from the first detection mode.

[0012] In Example 2, the subject matter of Example 1 may optionally include, wherein the mode switching trigger event comprises a threshold time elapsed after the PVC is detected in the first detection mode.

[0013] In Example 3, the subject matter of Example 2 optionally includes, wherein the threshold time is a number of weeks.

[0014] In Example 4, the subject matter of any one or more of Examples 1-3 may optionally include, wherein the processor circuit is configured to determine a PVC load representing a percentage of heartbeats detected as a PVC single shot or a pattern of multiple consecutive PVCs.

[0015] In Example 5, the subject matter of any one or more of Examples 1-4 optionally includes, wherein the first detection mode includes at least one trained computational model, the at least one trained computational model including corresponding weight factors for a set of signal features, the at least one trained computational model being trained to identify one or more PVC types.

[0016] In Example 6, the subject matter of Example 5 optionally includes, wherein the at least one trained computational model includes a logistic regression model, wherein, in order to classify the PVC candidates identified in the first detection mode, the processor circuit is configured to: calculate a weighted combination of the generated signal feature set, each signal feature being weighted by a corresponding weight factor; and based on a comparison of the weighted combination with a confidence threshold, classify the identified PVC candidates into a pattern of a single PVC or multiple consecutive PVCs.

[0017] In Example 7, the subject matter of any one or more of Examples 5-6 optionally includes, wherein the pattern of multiple consecutive PVCs includes a PVC pair or a PVC burst of three or more consecutive PVCs; wherein the identified PVC candidates include one or more of an identified PVC single candidate, an identified PVC pair candidate, or an identified PVC burst candidate, wherein at least one trained computational model includes one or more of a trained single detection model, a trained pair detection model, or a trained PVC burst detection model, and wherein, in order to classify the identified PVC candidates, the processor circuit is configured to apply the trained single detection model to the identified PVC single candidate, apply the trained pair detection model to the identified PVC pair candidate, and apply the trained PVC burst detection model to the identified PVC burst candidate.

[0018] In Example 8, the subject matter of Example 7 optionally includes, wherein the processor circuit is configured to: identify the PVC candidate as a PVC single-shot candidate; and apply the sensed cardiac signal to a trained single-shot detection model to generate a single-shot feature set for the identified PVC single-shot candidate, and use the single-shot feature set and the trained single-shot detection model to confirm the PVC single-shot detection.

[0019] In Example 9, the subject matter of Example 8 optionally includes, wherein the processor circuit is configured to, in response to a confirmed PVC single-shot detection, apply the sensed cardiac signal to a trained post-single-shot PVC detection model to generate a post-single-shot feature set for the heartbeat immediately following the confirmed PVC single-shot, and use the post-single-shot feature set and the trained post-single-shot PVC detection model to confirm the PVC pair detection.

[0020] In Example 10, the subject matter of Example 7 optionally includes, wherein the processor circuit is configured to: identify a PVC candidate as a PVC pair candidate based on two consecutive cardiac intervals shorter than an interval threshold; and apply the sensed cardiac signal to a trained pair detection model to generate a first feature set and a second feature set for two consecutive PVCs of the identified PVC pair candidate, respectively, and use the first feature set and the second feature set and the trained pair detection model to confirm the PVC pair detection.

[0021] In Example 11, the subject matter of Example 7 optionally includes, wherein the processor circuit is configured to: identify a PVC candidate as a PVC burst candidate based on (i) three or more consecutive cardiac intervals that are shorter than an interval threshold, or (ii) a confirmed PVC pair detection; and apply the sensed cardiac signal to a trained PVC burst detection model to generate a third feature set for a third or subsequent PVC of the PVC burst candidate, and use the third feature set and the trained PVC burst detection model to confirm the PVC burst detection.

[0022] In Example 12, the subject matter of any one or more of Examples 1-11 optionally includes, wherein the signal feature set generated in the first detection mode includes: one or more amplitude features, which include the ratio of the peak-to-peak (PP) amplitude between two consecutive heartbeats in the sensed cardiac signal; or one or more interval features, which include the difference in cardiac intervals between two consecutive heartbeats in the sensed cardiac signal.

[0023] In Example 13, the subject matter of any one or more of Examples 1-12 optionally includes, wherein the signal feature set generated in the first detection mode includes one or more morphological features, which include: a first morphological similarity measure between (i) a heartbeat in the sensed cardiac signal and (ii) a normal sinus rhythm (NSR) morphology; or a second morphological similarity measure between two consecutive heartbeats in the sensed cardiac signal.

[0024] In Example 14, the subject matter of Example 13 optionally includes, wherein one or more of the first morphological similarity measure or the second morphological similarity measure comprises a correlation measure.

[0025] In Example 15, the subject matter of any one or more of Examples 13-14 optionally includes, wherein the processor circuit is configured to: detect a change in the NSR morphology; and if the detected change in the NSR morphology is below a change threshold, classify the identified PVC candidate using the first morphology similarity metric and the second morphology similarity metric, and if the detected change in the NSR morphology exceeds the change threshold, classify the identified PVC candidate using the second morphology similarity metric instead of the first morphology similarity metric.

[0026] Example 16 is a method for detecting arrhythmia in a subject using an implantable medical device, comprising: sensing a cardiac signal of the subject using a sensor circuit; detecting a heartbeat from the cardiac signal of the subject; detecting premature ventricular contractions (PVCs) in a first detection mode, comprising: using a cardiac interval or signal amplitude of the detected heartbeat to identify a specific type of PVC candidate; generating a signal feature set including morphological features from the sensed signal; and using the generated signal feature set to classify the identified PVC candidates into a single PVC or a pattern of multiple consecutive PVCs; in response to a mode switching trigger event, using the cardiac interval or signal amplitude of the detected heartbeat to detect a specific type of PVC in a second detection mode different from the first detection mode.

[0027] In Example 17, the subject matter of Example 16 may optionally include, wherein the mode switching trigger event comprises a threshold time elapsed after the PVC is detected in the first detection mode.

[0028] In Example 18, the subject matter of any one or more of Examples 16-17 optionally includes, wherein the first detection mode includes at least one trained computational model, the at least one trained computational model including corresponding weight factors for a set of signal features, the at least one trained computational model being trained to identify one or more PVC types.

[0029] In Example 19, the subject matter of Example 18 optionally includes, wherein the pattern of multiple consecutive PVCs includes a PVC pair or a PVC burst of three or more consecutive PVCs, wherein the identified PVC candidates include one or more of an identified PVC single candidate, an identified PVC pair candidate, or an identified PVC burst candidate, wherein at least one trained computational model includes one or more of a trained single-shot detection model, a trained pair detection model, or a trained PVC burst detection model, wherein classifying the identified PVC candidates includes applying the trained single-shot detection model to the identified PVC single candidate, applying the trained pair detection model to the identified PVC pair candidate, or applying the trained PVC burst detection model to the identified PVC burst candidate.

[0030] In Example 20, the subject matter of Example 19 optionally includes: identifying a PVC candidate as a PVC single-shot candidate; applying the cardiac signal to a trained single-shot detection model to generate a single-shot feature set for the identified PVC single-shot candidate, and using the single-shot feature set and the trained single-shot detection model to confirm the PVC single-shot detection; and in response to the confirmed PVC single-shot detection, applying the cardiac signal to a trained post-single-shot PVC detection model to generate a post-single-shot feature set for the heartbeat immediately following the confirmed PVC single-shot, and using the post-single-shot feature set and the trained post-single-shot PVC detection model to confirm the PVC pair detection.

[0031] In Example 21, the subject matter of any one or more of Examples 19-20 optionally includes: identifying a PVC candidate as a PVC pair candidate based on two consecutive cardiac intervals shorter than an interval threshold; and applying the cardiac signal to a trained pair detection model to generate a first feature set and a second feature set for two consecutive PVCs of the identified PVC pair candidate, respectively, and using the first feature set and the second feature set and the trained pair detection model to confirm the PVC pair detection.

[0032] In Example 22, the subject matter of any one or more of Examples 19-21 optionally includes: identifying a PVC candidate as a PVC burst candidate based on (i) three or more consecutive cardiac intervals that are shorter than an interval threshold or (ii) a confirmed PVC pair detection; and applying the cardiac signal to a trained PVC burst detection model to generate a third feature set for a third or subsequent PVC of the PVC burst candidate, and using the third feature set and the trained PVC burst detection model to confirm the PVC burst detection.

[0033] Embodiments of the systems, devices, and methods discussed herein can improve the technology of device-based arrhythmia detection, particularly the detection of PVCs and the estimation of PVC load. According to some embodiments, multiple PVC detectors can be customized to detect or classify different PVC types (patterns), such as PVC singles, pairs, or bursts of PVCs, using respectively optimized feature sets. In some examples, the dedicated detectors can each be implemented as a corresponding computational model, which is trained separately to optimize their performance in detecting PVC patterns. In some examples, the feature set used by the detector can be dynamically determined based on, for example, the patient's physiological or physical condition (e.g., posture or physical activity state). Compared with traditional PVC detectors that focus on detecting a single PVC, the dedicated detectors described herein can help reduce false positive and false negative detections or classifications of PVC patterns (such as pairs or bursts of PVCs), thereby improving the overall accuracy of PVC load estimation. In addition, compared with applying the same single PVC detection algorithm to detect each PVC beat in a PVC pair or burst of PVCs, dedicated pairs and bursts of PVC detectors can be more accurate and efficient, especially in energy and resource sensitive systems, such as implanted medical devices.

[0034] According to some examples, layered PVC detection involves a first triggering stage that uses only cardiac interval or signal amplitude information to mark PVC candidates (single-shot candidates, linked pairs, or PVC burst candidates), and a second confirmation stage that uses more complex features (e.g., morphological features) and algorithms than the first triggering stage to confirm and classify the identified PVC candidates. Because the first triggering stage requires fewer resources, PVC candidates can be identified more efficiently; and the second confirmation stage can better utilize system resources to more accurately detect PVCs and classify them into different types, such as PVC single shots, linked pairs, or PVC bursts. Therefore, compared with traditional PVC detection techniques, the layered PVC detection discussed in this article can not only improve PVC detection and classification performance (e.g., higher sensitivity and specificity) and PVC load estimation accuracy, but also better utilize system resources and reduce power consumption to meet device life requirements in implantable medical devices, thereby improving the functionality of implantable medical devices.

[0035] Morphology-based PVC detection typically requires more system resources and consumes more power, which makes meeting the power consumption requirements of implantable medical devices more challenging. The present invention describes a unique technique for switching between two different PVC detection modes with different computational complexity and power requirements. The first detection mode involves hierarchical PVC detection using computationally and power-intensive morphology-based features. The second detection mode uses simpler features, such as cardiac intervals and / or signal amplitudes, but does not use morphological features. The transition between the first detection mode and the second detection mode can occur in response to a mode switching trigger event, such as a threshold time (e.g., one week, two weeks, or other specified duration) after PVC detection in the first detection mode. Because the second detection mode typically requires less computation and power consumption than the first detection mode, automatically switching to the second detection mode can help save battery power of the implantable medical device and extend the life of the device.

[0036] The improvements in PVC detection and PVC load estimation discussed herein can be achieved with little additional cost or increased system complexity. In some examples, existing system performance (e.g., high PVC detection sensitivity and specificity, etc.) can be maintained using less costly or less obtrusive systems, devices, and methods. With improved PVC detection, fewer alarms are provided, battery life can be extended, fewer unnecessary medications and procedures can be scheduled, prescribed, or provided, and overall system cost savings and power savings can be achieved compared to existing medical devices and systems.

[0037] This summary is an overview of some of the teachings of the present application and is not intended to be an exclusive or exhaustive treatment of the subject matter. More details about the subject matter can be found in the detailed description and the appended claims. Other aspects of the present disclosure will be apparent to those skilled in the art upon reading and understanding the following detailed description and viewing the accompanying drawings, each of which should not be construed as limiting. The scope of the present disclosure is defined by the appended claims and their legal equivalents. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Various embodiments are shown by way of example in the accompanying drawings. These embodiments are exemplary and are not intended to be exhaustive or exclusive embodiments of the subject matter.

[0039] Figure 1 An example of a patient management system and a portion of an environment in which the system may operate are generally shown.

[0040] Figure 2 An example of an arrhythmia detection system for detecting and classifying premature ventricular contractions (PVCs) and estimating PVC burden is generally shown.

[0041] Figures 3A-3C Example ECG tracings of various PVC types are shown.

[0042] Figure 4 is a diagram illustrating a method of identifying various types of PVC candidates from sensed heartbeats at the triggering stage of hierarchical PVC detection.

[0043] Figure 5 is a diagram illustrating a method of confirming or classifying a PVC candidate into one of the PVC types during the confirmation stage of hierarchical PVC detection.

[0044] Figure 6 is a diagram showing a method of selecting a PVC detection model based on changes in a patient's condition.

[0045] Figure 7 A block diagram of an example machine is generally shown on which any one or more of the techniques (eg, methodologies) discussed herein may be performed. DETAILED DESCRIPTION

[0046] Disclosed herein are systems, devices, and methods for detecting and classifying various types of premature ventricular contractions (PVCs). An exemplary system includes a sensor circuit for sensing a cardiac signal of a subject, and a processor circuit for detecting heartbeats from the cardiac signal and detecting PVCs in a first detection mode or a different second detection mode. The first detection mode includes using the cardiac intervals or signal amplitudes of the detected heartbeats to identify specific types of PVC candidates, and using signal features including morphological features of the sensed cardiac signal to classify the identified PVC candidates into a pattern of a single PVC or multiple consecutive PVCs. The second detection mode includes using the cardiac intervals or signal amplitudes of the detected heartbeats to detect specific types of PVCs. The transition between the first detection mode and the second detection mode can occur in response to a mode switching trigger event.

[0047] Figure 1 A portion of an environment in which an example patient management system 100 and a patient management system 200 may operate is shown. The patient management system 100 may perform a range of activities including remote patient monitoring and disease condition diagnosis. These activities may be performed near the patient 101, such as at the patient's home or office, through a centralized server, such as at a hospital, clinic, or doctor's office, or through a remote workstation, such as a secure wireless mobile computing device.

[0048] The patient management system 100 may include one or more mobile medical devices, an external system 105, and a communication link 111 that provides communication between the one or more mobile medical devices and the external system 105. The one or more mobile medical devices may include an implantable medical device (IMD) 102, a wearable medical device (WMD) 103, or one or more other implantable, leadless, subcutaneous, external, wearable, or mobile medical devices configured to monitor, sense, or detect information from the patient 101, determine physiological information about the patient 101, or provide one or more therapies to treat various conditions of the patient 101, such as one or more cardiac or non-cardiac conditions (e.g., dehydration, sleep apnea, etc.).

[0049] In one example, IMD 102 may include one or more conventional cardiac rhythm management devices implanted in the chest of a patient, having a lead system including one or more transvenous, subcutaneous, or non-invasive leads or catheters to position one or more electrodes or other sensors (e.g., heart sound sensors) in, on, or around the heart of patient 101, or in one or more other locations on the chest, abdomen, or neck. In another example, IMD 102 may include, for example, a monitor implanted subcutaneously in the chest of patient 101, IMD 102 including a housing containing a circuit system, and in some examples, including one or more sensors, such as a temperature sensor, etc.

[0050] IMD 102 may include an evaluation circuit configured to detect or determine specific physiological information of patient 101, or determine one or more conditions, or provide information or alarms to a user, such as patient 101 (e.g., patient), clinician, or one or more other caregivers or processes. IMD 102 may alternatively or additionally be configured as a therapeutic device configured to treat one or more medical conditions of patient 101. Therapy may be delivered to patient 101 via a lead system and associated electrodes or using one or more other delivery mechanisms. Therapy may include delivering one or more drugs to patient 101, such as using IMD 102 or one or more other mobile medical devices. In some examples, therapy may include cardiac resynchronization therapy for correcting dyssynchrony and improving cardiac function in patients with heart failure. In other examples, IMD 102 may include a drug delivery system, such as a drug infusion pump, to deliver drugs to patients for the management of arrhythmias or complications caused by arrhythmias, hypertension, or one or more other physiological conditions. In other examples, IMD 102 may include one or more electrodes configured to stimulate the patient's nervous system or provide stimulation to muscles of the patient's airway, etc.

[0051] WMD 103 may include one or more wearable or external medical sensors or devices (e.g., an automatic external defibrillator (AED), a Holter monitor, a patch-based device, a smart watch, a smart accessory, a wrist- or finger-worn medical device such as a finger-based photoplethysmography sensor, etc.).

[0052] The external system 105 may include a dedicated hardware / software system, such as a programmer, a remote server-based patient management system, or a system defined primarily by software running on a standard personal computer. The external system 105 may manage the patient 101 through the IMD 102 or one or more other mobile medical devices connected to the external system 105 via the communication link 111. In other examples, the IMD 102 may be connected to the WMD 103 via the communication link 111, or the WMD 103 may be connected to the external system 105. This may include, for example, programming the IMD 102 to perform one or more of the following: collecting physiological data, performing at least one self-diagnostic test (such as for device operating status), analyzing physiological data, or optionally delivering or adjusting treatment for the patient 101. In addition, the external system 105 may send information to or receive information from the IMD 102 or WMD 103 via the communication link 111. Examples of information may include real-time or stored physiological data from patient 101, diagnostic data such as detection of patient hydration status, hospitalization status, response to therapy delivered to patient 101, or device operating status of IMD 102 or WMD 103 (e.g., battery status, lead impedance, etc.). Communication link 111 may be an inductive telemetry link, a capacitive telemetry link, or a radio-frequency (RF) telemetry link, or wireless telemetry based on, for example, "strong" Bluetooth or IEEE 802.11 Wireless Fidelity "Wi-Fi" interface standards. Other configurations and combinations of patient data source interfacing are also possible.

[0053] The external system 105 may include an external device 106 near one or more mobile medical devices, and a remote device 108 at a location relatively far from one or more mobile medical devices, which communicates with the external device 106 via a communication network 107. Examples of external devices 106 may include medical device programmers. The remote device 108 may be configured to evaluate the collected patient or patient information and provide alarm notifications and other possible functions. In one example, the remote device 108 may include a centralized server that acts as a central hub for the storage and analysis of collected data. The server may be configured as a single, multiple or distributed computing and processing system. The remote device 108 may receive data from multiple patients. The data may be collected by one or more mobile medical devices and other data acquisition sensors or devices associated with the patient 101. The server may include a storage device to store the data in a patient database. The server may include an alarm analyzer circuit to evaluate the collected data to determine whether a specific alarm condition is met. The satisfaction of the alarm condition may trigger the generation of an alarm notification, such as provided by one or more human-perceivable user interfaces. In some examples, the alarm condition may be alternatively or additionally evaluated by one or more mobile medical devices (such as implantable medical devices). For example, the alarm notification may include a web page update, a telephone or pager call, an email, SMS, a text or "instant" message, and a message to the patient and a simultaneous direct notification to emergency services and clinicians. Other alarm notifications are also possible. The server may include an alarm prioritizer circuit that is configured to prioritize the alarm notifications. For example, the alarms for detected physiological events may be prioritized using a similarity metric between physiological data associated with the detected physiological event and physiological data associated with historical alarms.

[0054] The remote device 108 may also include one or more locally configured clients or remote clients securely connected to the server via the communication network 107. Examples of clients may include personal desktops, laptops, mobile devices, or other computing devices. System users, such as clinicians or other qualified medical professionals, may use the client to securely access patient data stored in a database in the server and select and prioritize patients and alerts for medical care provision. In addition to generating alert notifications, the remote device 108, including the server and interconnected clients, may also execute a follow-up regimen by sending a follow-up request to one or more mobile medical devices, or by sending a message or other communication to the patient 101 (e.g., the patient), a clinician, or an authorized third party as a compliance notification.

[0055] The communication network 107 can provide wired or wireless interconnection. In one example, the communication network 107 can be based on the Transmission Control Protocol / Internet Protocol (TCP / IP) network communication specification, but other types or combinations of network implementations are also possible. Similarly, other network topologies and arrangements are also possible.

[0056] One or more of the external device 106 or remote device 108 may output the detected physiological events to a system user, such as a patient or clinician, or to a process including, for example, an instance of a computer program executable in a microprocessor. In one example, the process may include automatically generating a recommendation for antiarrhythmic therapy or a recommendation for further diagnostic testing or treatment. In one example, the external device 106 or remote device 108 may include a corresponding display unit for displaying physiological or functional signals, or an alarm, an alert, an emergency call, or other form of warning to signal the detection of an arrhythmia. In some examples, the external system 105 may include an external data processor configured to analyze physiological or functional signals received by one or more mobile medical devices and confirm or reject the detection of an arrhythmia. Computationally intensive algorithms, such as machine learning algorithms, may be implemented in the external data processor to retrospectively process data to detect arrhythmias.

[0057] Portions of one or more mobile medical devices or external systems 105 may be implemented using hardware, software, firmware, or a combination thereof. Portions of one or more mobile medical devices or external systems 105 may be implemented using dedicated circuits that may be constructed or configured to perform one or more functions, or may be implemented using general-purpose circuits that may be programmed or otherwise configured to perform one or more functions. Such general-purpose circuits may include a microprocessor or a portion thereof, a microcontroller or a portion thereof, or a programmable logic circuit, a memory circuit, a network interface, and various components for interconnecting these components. For example, among others, a "comparator" may include an electronic circuit comparator that may be constructed to perform a specific function of a comparison between two signals, or a comparator may be implemented as part of a general-purpose circuit that may be driven by a code that instructs a portion of the general-purpose circuit to perform a comparison between two signals. A "sensor" may include an electronic circuit configured to receive information and provide an electronic output representing such received information.

[0058] The therapeutic device 110 can be configured to send information to or receive information from one or more mobile medical devices or external systems 105 using a communication link 111. In one example, one or more mobile medical devices, external devices 106, or remote devices 108 can be configured to control one or more parameters of the therapeutic device 110. The external system 105 can allow programming of one or more mobile medical devices, and can receive information about one or more signals acquired by one or more mobile medical devices, such as information that can be received via the communication link 111. The external system 105 can include a local external implantable medical device programmer. The external system 105 can include a remote patient management system, which can monitor the patient's status or adjust one or more treatments, such as from a remote location.

[0059] Figure 2 An example of an arrhythmia detection system 200 is generally shown, which is configured to detect and classify premature ventricular contractions (PVCs) and estimate PVC burden. Portions of the system 200 may be included in the physiological event detector circuit 160 of the IMD 102 or WMD 103 or in the external system 105. The system 200 may include one or more of a sensor circuit 210, a processor circuit 220, and a user interface unit 230. The system 200 may also include an optional therapy circuit 240.

[0060] Sensor circuit 210 may include circuitry configured to sense physiological signals of a patient via one or more implanted, wearable, or otherwise mobile sensors or electrodes associated with the patient. The sensor may be incorporated into a mobile device such as IMD 102 or WMD 103, or otherwise associated with a mobile device. In some examples, the sensor may be incorporated into an implantable cardiac monitor (ICM) device configured for subcutaneous implantation. Examples of physiological signals may include a surface electrocardiogram (ECG) sensed from electrodes placed on the body surface, a subcutaneous ECG sensed from electrodes placed under the skin, an intracardiac electrogram (EGM) sensed from one or more electrodes on a lead system, a thoracic or cardiac impedance signal, an arterial pressure signal, a pulmonary artery pressure signal, a left atrial pressure signal, a RV pressure signal, a LV coronary pressure signal, a coronary blood temperature signal, a blood oxygen saturation signal, a heart sound signal such as sensed by a mobile accelerometer or an acoustic sensor, a physiological response to activity, an apnea-hypopnea index, one or more respiratory signals such as a respiratory rate signal or a tidal volume signal, brain natriuretic peptide (BNP), a blood routine, sodium and potassium levels, glucose levels, and other biomarkers and biochemical markers, etc. The sensor circuit 210 may include one or more subcircuits to digitize, filter, or perform other signal conditioning operations on the received physiological signals.

[0061] Processor circuit 220 coupled to sensor circuit 210 may detect and classify PVC types or patterns, such as a PVC single (i.e., a single PVC), a PVC pair including two consecutive PVCs, or a PVC burst including three or more consecutive PVCs, and determine a PVC load estimate based on the detected and classified PVCs. Figures 3A-3C Examples of various PVC types in an ECG tracing recorded from a patient are shown, such as a PVC single 310, a PVC pair 320, a PVC triple 330 (three consecutive PVCs), or a non-sustained VT 340 comprising a series of consecutive PVCs. The processor circuit 220 may be implemented as part of a microprocessor circuit, which may be a special-purpose processor such as a digital signal processor, an application specific integrated circuit (ASIC), a microprocessor, or other type of processor for processing information including body activity information. Alternatively, the microprocessor circuit may be a general-purpose processor that may receive and execute a set of instructions to perform the functions, methods, or techniques described herein.

[0062] Processor circuit 220 may include a circuit group that includes one or more other circuits or subcircuits, including a beat detector 221, a PVC candidate identifier 222, a feature extractor 224, a PVC classifier 226, a classification model selector 227, and a PVC load estimator 228. These circuits may perform the functions, methods, or techniques described herein, either alone or in combination. In one example, the hardware of the circuit group may be invariably designed to perform a specific operation (e.g., hardwired). In one example, the hardware of the circuit group may include variably connected physical components (e.g., execution units, transistors, simple circuits, etc.) that include physically modified computer-readable media (e.g., magnetic, electrical, removable placement of invariant aggregated particles, etc.) to encode instructions for a specific operation. When the physical components are connected, the basic electrical characteristics of the hardware components are changed, for example, from an insulator to a conductor, or vice versa. The instructions enable embedded hardware (e.g., an execution unit or a loading mechanism) to create a member of the circuit group in hardware via a variable connection to perform a portion of a specific operation when operating. Therefore, when the device operates, the computer-readable medium is communicatively coupled to other components of the circuit group member. In one example, any physical component may be used in more than one member of more than one circuit group. For example, in operation, an execution unit may be used in a first circuit of a first circuit group at one point in time and reused by a second circuit in the first circuit group or a third circuit in the second circuit group at a different time.

[0063] The pulsation detector 221 can detect a heartbeat using one or more physiological signals sensed by the sensor circuit 210. In one example, the sensor circuit 210 can be coupled to one or more implantable, wearable, or otherwise mobile cardiac activity sensors that are configured to sense the patient's cardiac electrical activity or mechanical activity. In another example, the physiological signals sensed by the sensor circuit 210 can be stored in a storage device such as an electronic medical record system. The pulsation detector 221 can receive physiological signals from the storage device, such as in response to a user command or a trigger event, and use the received physiological signals to detect a heartbeat.

[0064] The pulsation detector 221 can detect the heartbeat from a cardiac electrical signal, such as an ECG sensed using surface electrodes or subcutaneous electrodes, or an intracardiac EGM sensed from inside a cardiac chamber or cardiac tissue using intracardiac electrodes. The heartbeat thus detected can correspond to a QRS complex wave in an ECG or EGM, which represents ventricular depolarization. In another example, the heartbeat can be detected from a cardiac mechanical signal such as a heart sound (HS) signal, a cardiac impedance signal, or a pressure signal, the heart sound signal being sensed, for example, by sensing cardiac vibration or acoustic information using an accelerometer or a microphone, the cardiac impedance signal that varies with periodic cardiac contractions can be sensed using an impedance sensor, and the pressure signal that varies with arterial pulse can be sensed using a pressure sensor.

[0065] In some examples, the pulsation detector 221 can pre-process the received physiological signal to improve the signal quality and detect the heartbeat from the pre-processed signal. Pre-processing may include signal filtering to improve the signal-to-noise ratio. In one example, the pulsation detector 221 can determine the noise characteristics of the cardiac electrical signal sensed within the noise window, such as the noise level, and filter the sensed cardiac electrical signal using the detected noise level. In one example, the pulsation detector 221 can use an adaptive filter to iteratively eliminate or attenuate the noise. It can then be detected from the filtered signal.

[0066] The detected heartbeats may be screened by the PVC candidate identifier 222 to identify PVC candidates, a process referred to herein as the triggering phase of hierarchical PVC detection. During the triggering phase, the PVC candidate identifier 222 may use signal characteristics of the detected heartbeats, such as signal amplitude or cardiac intervals between the detected heartbeats, to identify one or more types of PVC candidates from the detected heartbeats. In one example, cardiac interval examination may include identifying a compensatory intermittent pattern characterized by a continuous "short-long" interval pattern including a shortened pre-PVC interval shorter than a baseline cycle length (e.g., during normal sinus rhythm or NSR), followed by an extended post-PVC interval longer than the baseline cycle length. In one example, the PVC candidate identifier 222 may use, among other features, the signal amplitude or cardiac intervals of the detected heartbeats to identify one or more of a PVC single candidate 222A, a PVC pair candidate 222B, or a PVC burst candidate 222C from the received heartbeats. The identification of a particular type or pattern of PVC candidates may help improve the efficiency and accuracy of subsequent PVC confirmation and classification by the PVC classifier 226. Referring to FIG. 2A and FIG. 2B , the PVC candidate identifier 222 may be used to identify a PVC single candidate 222A, a PVC pair candidate 222B, or a PVC burst candidate 222C. Figure 4 An example of discussing PVC candidate identification during the trigger phase.

[0067] In some examples, PVC candidate identifier 222 may initially identify unified PVC pair / PVC burst candidates without distinguishing between PVC pair candidates and PVC burst candidates. When a PVC pair is confirmed by PVC classifier 226, PVC candidate identifier 222 may further identify PVC burst candidates based on the signal characteristics of the heartbeat immediately following the confirmed PVC pair ("post-pair heartbeat"). Figure 5 An example of such delayed triggering of PVC burst candidates is discussed.

[0068] The identified PVC candidates can be confirmed and further classified into one of the PVC types, such as a single PVC, a PVC pair, or a PVC burst (also referred to herein as non-sustained VT or NSVT), which is referred to herein as the confirmation phase of hierarchical PVC detection. During the confirmation phase, the feature extractor 224 can extract a feature set from the identified PVC candidates. These features can be used by the PVC classifier 226 to confirm and classify the PVC candidates. Compared with the signal characteristics (e.g., signal amplitude and intervals between heartbeats) used by the PVC candidate identifier 222 to identify the PVC candidates, the feature extractor 224 can extract more refined features, including, for example, one or more amplitude-based features, interval-based features, or morphology-based features of the heartbeat. Using such more refined features can help reduce classification errors during the confirmation phase. On the other hand, because the first triggering phase (identifying PVC candidates) requires fewer resources and consumes less power than the second confirmation phase (confirming and classifying PVC candidates), the hierarchical PVC detection discussed herein can effectively utilize system resources, reduce power consumption, and better meet the device life requirements of implantable medical devices. Examples of amplitude-based features may include the ratio of peak-to-peak (PP) amplitudes between two consecutive heartbeats (e.g., between the current beat and the immediately preceding beat, or between the current beat and the immediately following beat), or the ratio of PP amplitudes measured in two different frequency bands. Examples of interval-based features may include the difference in cardiac intervals between two consecutive heartbeats in the received cardiac signal (e.g., between the current beat and the immediately preceding beat, or between the current beat and the immediately following beat), the expected compensatory interval difference. Examples of morphology-based features may include a morphology similarity measure between two heartbeats, or a morphology similarity measure between a heartbeat and a pre-generated template beat morphology, such as one generated from a heartbeat during normal sinus rhythm (NSR), and therefore referred to as an "NSR template." In one example, correlation waveform analysis (CWA) may be used to calculate a similarity measure as a correlation between a morphology feature of a heartbeat and a morphology feature of an NSR template. In another example, dynamic correlation waveform analysis (DCWA) can be used to calculate a similarity measure as the correlation between corresponding morphological features of two consecutive heartbeats (without using an NSR template). In one example, the morphological features can include a QRS width or QRS area under the curve derived from an ECG or EGM signal portion of the heartbeat.

[0069] In some examples, feature extractor 224 may extract feature sets according to the type of identified PVC candidates. For example, feature extractor 224 may extract single-shot feature set 224A for identified PVC single-shot candidate 222A, extract pair-shot feature set 224B for identified PVC joint pair candidate 222B, or extract PVC continuous-shot feature set 224C for identified PVC continuous-shot candidate 222C. Features of one type of PVC candidate may be different from features of another type of PVC candidate. For example, two feature sets may have different numbers or types of features, such as amplitude-based features, interval-based features, or morphology-based features. In one example, joint pair features 224B may include two separate feature sets (referred to as joint pair-1 features and joint pair-2 features) generated for two consecutive PVCs in the identified PVC joint pair candidates, respectively. Joint pair-1 features and joint pair-2 features may be different from each other (for example, they may have different numbers or types of features, such as amplitude-based features, interval-based features, or morphology-based features). In one example, joint pair-1 features and joint pair-2 features may each be different from single-shot features 224A. In the example of determining PVC burst candidates 222C based on delayed triggering as described above (i.e., identifying another PVC beat immediately following a confirmed PVC pairing, the “post-pairing heartbeat”), feature extractor 224 may generate PVC burst features 224C including features extracted from the post-pairing heartbeat.

[0070] PVC classifier 226 may confirm and classify PVC candidates using the PVC candidate features extracted by feature extractor 224. In one example, PVC classifier 226 may include separate detectors for confirming different types of PVC candidates using corresponding features generated by feature extractor 224, including, for example, a single shot detector 226A for confirming single shot candidate 222A using single shot features 224A, a joint pair detector 226B for confirming joint pair candidate 222B using joint pair features 224B, or a PVC burst detector 226C for confirming PVC burst candidate 222C using PVC burst features 224C.

[0071] In one example, PVC classifier 226 (or any of detectors 226A-226C) can use at least one trained computational model to identify and classify PVC candidates. The at least one computational model can be generated using a model generator circuit, which can be included in processor circuit 220 or separate from processor circuit 220. The at least one computational model can be trained to identify and classify the PVC type (e.g., single shot, linked pair, or PVC burst) of any given PVC candidate. In some examples, the at least one computational model can be further trained to recognize corresponding feature sets 224A-224C for different types of PVC candidates 222A-222C. Feature extractor 224 can use at least one trained computational model to extract features from the signal portion corresponding to the PVC candidate.

[0072] In one example, at least one trained computational model may include a machine learning (ML) model. ML models may be trained using supervised learning or unsupervised learning. Supervised learning uses prior knowledge (e.g., examples that associate inputs with outputs or results) to learn the relationship between inputs and outputs. The goal of supervised learning is to learn a function that best approximates the relationship between training inputs and outputs given some training data sets so that the ML model can implement the same relationship to generate corresponding outputs when given inputs. Unsupervised learning is to train ML algorithms using information that is neither classified nor labeled, and allows the algorithm to act on this information without guidance. Unsupervised learning is useful in exploratory analysis because it can automatically identify structures in data. Examples of ML models may include logistic regression, naive Bayes, random forests (RF), neural networks (NN), matrix decomposition, or support vector machines (SVM), etc. In one example, the ML model has the architecture of a deep neural network. Examples of DNNs include convolutional neural networks (CNNs), recurrent neural networks (RNNs), deep belief networks (DBNs), long-term and short-term memory (LSTM) networks, transfer learning networks, or hybrid neural networks including two or more neural network models of different types or different model configurations.

[0073] By way of example and not limitation, at least one trained computational model may include at least one trained logistic regression (LG) model including respective weight factors for feature sets generated by feature extractor 224. PVC classifier 226 (or any of detectors 226A-226C) may use at least one trained LG model to compute a weighted combination of feature sets, each weighted by a respective weight factor, and classify PVC candidates based on a comparison of the weighted combination to a confidence threshold. By way of example and not limitation, single-shot detector 226A may apply a signal portion (e.g., ECG or EGM) of an identified PVC single-shot candidate 222A to a trained single-shot detection model (e.g., LG model) and confirm detection of a PVC single-shot when the weighted sum of single-shot features 224A exceeds a first confidence threshold. Similarly, the pair detector 226B may apply the signal portion of the identified PVC pair candidate 222B to a trained pair detection model (e.g., LG model) and confirm the detection of the PVC pair when the weighted sum of the pair features 224B exceeds a second confidence threshold. The PVC burst detector 226C may apply the signal portion of the identified PVC burst candidate 222C to a trained PVC burst detection model (e.g., LG model) and confirm the detection of the PVC burst when the weighted sum of the PVC burst features 224C exceeds a third confidence threshold. The first confidence threshold, the second confidence threshold, and the third confidence threshold may be different from each other and are determined during model training.

[0074] In one example, the joint pair detector 226B may confirm the joint pair candidate 222B by respectively confirming each of the two consecutive PVCs (joint pair-1 and joint pair-2) in the identified PVC joint pair candidates. The joint pair detector 226B may apply the signal portion of the joint pair-1 to a first trained computational model (e.g., LG model) to identify a feature set (joint pair-1 features as described above) and confirm the detection of the joint pair-1. Similarly, the joint pair detector 226B may apply the signal portion of the joint pair 2 to a second trained computational model (e.g., LG model) to identify a feature set (joint pair 2 features as described above) and confirm the detection of the joint pair 2.

[0075] In some examples, confirmation or classification of a PVC type (e.g., PVC pair or PVC burst) can be based on confirmation detection of another PVC type. In one example, in response to a confirmation detection of a PVC single by single detector 226A, pair detector 226B can apply the signal portion of the heartbeat immediately following the confirmed PVC single ("post-single heartbeat") to a trained computational model (e.g., LG model) to identify a feature set from the post-single heartbeat, and if the model output exceeds a confidence threshold, a post-single PVC is detected. Pair detector 226B can then confirm the detection of a PVC pair, which includes the confirmed PVC single and the detected post-single PVC. In another example, in response to a confirmed detection of a PVC pair by pair detector 226B, PVC burst detector 226C may apply a signal portion of a heartbeat immediately following the confirmed PVC pair (a "post-pair heartbeat") to a trained computational model (e.g., an LG model) to identify a feature set from the post-pair heartbeat, and if the model output exceeds a confidence threshold, a post-pair PVC is detected. PVC burst detector 226C may then confirm the detection of a PVC burst, which includes a confirmed PVC pair and a detected post-pair PVC. Figure 5 Discuss examples of confirming or classifying PVC candidates.

[0076] In some examples, the PVC classifier 226 (or any of the detectors 226A-226C) can apply different trained computational models to detect specific types of PVCs in response to changes in the patient's condition. Different trained computational models can include different corresponding PVC features (e.g., single-shot features, linked-pair features, or PVC burst features), model architectures, or model parameters (e.g., weight factors and confidence thresholds for corresponding PVC features for the LG model). For example, features used in correlation waveform analysis (CWA), such as the correlation between morphological features of a heartbeat and morphological features of an NSR template (heartbeats during NSR), can be strong discriminants of PVCs. This is assuming that the NSR template represents an NSR beat. In cases where the NSR morphology changes dramatically (such as due to posture or heart rate-dependent factors), a reduced CWA score for an NSR beat may result in false positives for non-PVC beats, such as premature atrial contractions (PACs) or other shortened intervals. The NSR qualification check can help decide whether the NSR template and CWA-based features can be used for PVC detection without introducing excessive false positive detections. If the NSR morphology is considered to have changed or become unstable, an alternative feature set (excluding CWA-based features) may be used to confirm and classify PVC candidates. For example, if the detected change in NSR morphology is below a change threshold, both CWA-based features and DCWA-based features may be used to confirm and classify PVC candidates. If the detected change in NSR morphology exceeds a change threshold, CW-based features are excluded from the process of confirming and classifying PVC candidates, while DCWA-based features are included.

[0077] The classification model selector 227 can select a trained computational model for use by the feature extractor 224 and the PVC classifier 226. In some examples, the classification model selector 227 can detect changes in the patient's baseline condition, such as changes in NSR characteristics, and select a trained computational model for the feature extractor 224 and the PVC classifier 226 based at least on the detected changes in the patient's baseline condition. Changes in the patient's baseline condition, such as changes in NSR characteristics, can be detected periodically or in response to a triggering event, such as a change in posture or body activity detected by a posture / activity sensor (such as an accelerometer). Such periodic or event-triggered NSR eligibility checks and NSR template updates are computationally more efficient than continuous NSR eligibility checks and NSR template updates. See below for more information. Figure 6 An example of selecting a PVC detection model based on changes in NSR characteristics is discussed.

[0078] The hierarchical PVC detection as described above includes a triggering phase in which a PVC candidate is identified using the signal amplitude or cardiac interval of a detected heartbeat, and a confirmation phase in which a PVC candidate is confirmed and classified into one of the PVC types using features including morphology-based features. Since morphology-based PVC detection generally requires more system resources and consumes more power, it may be challenging to meet the power consumption requirements of implantable medical devices. In order to reduce the power consumption of implantable medical devices, in some examples, the arrhythmia detection system 200 can operate in different PVC detection modes with different computational complexities and power requirements. In one example, a first detection mode may involve computationally and power-intensive morphological features, such as hierarchical PVC detection as described above. A different second detection mode may involve simpler features, such as cardiac intervals and / or signal amplitudes, but not morphological features. The arrhythmia detection system 200 may detect PVCs in a first detection mode and switch to a second detection mode in response to a mode switching trigger event (such as a threshold time after a PVC is detected in the first detection mode). As an example and not limitation, the threshold time may be one week, two weeks, or other duration specified by the user. In some examples, the transition from the first detection mode to the second detection mode can be based on a patient condition, such as risk of a cardiac event (e.g., arrhythmia), or a PVC load estimate based on PVC detection in the first detection mode. Because the second detection mode generally requires less computation and power consumption than the first detection mode, automatically transitioning to the second detection mode can help conserve battery power of the implanted medical device and extend device life.

[0079] PVC load estimator 228 may generate PVC load estimates using the various types of confirmed and classified PVCs provided by PVC classifier 226. The PVC load represents the frequency of any type of PVC detected within a specific time period, such as a 24-hour period in one example, or a trend of 24-hour loads over a number of days. The PVC load may be calculated as the percentage of heartbeats that are classified as any type of PVC. In some examples, PVC load estimator 228 may generate PVC loads specific to a specific PVC type, such as a PVC single load, a PVC pair load, or a PVC burst load.

[0080] The user interface unit 230 may include an input device and an output device. In one example, at least a portion of the user interface unit 230 may be implemented in the external system 105. The input device may receive user programming input, such as parameters for detecting heartbeats, or user modifications of model parameters associated with PVC candidate identification, feature extraction, or PVC classification. The input device may include a keyboard, an on-screen keyboard, a mouse, a trackball, a touchpad, a touch screen, or other pointing or navigation device. The input device may enable a system user to program parameters for sensing physiological signals, detecting arrhythmias, and generating alarms, etc. The output device may include a display to display a human-perceivable presentation of detected PVCs and PVC load estimates, such as sensed physiological information and detected heartbeats, among other information. The output unit may include a printer for printing a hard copy of the detection information. The information may be presented in a table, a chart, a schematic, or any other type of text, list, or graphical presentation format. The presentation of the output information may include audio or other media formats to alert the system user of the detected arrhythmia event. In one example, the output device may generate an alarm, an alert, an emergency call, or other form of warning to signal the system user about the detected PVC.

[0081] The optional therapy circuit 240 can be configured to deliver therapy to the patient in response to the detected PVC or PVC load estimate. Examples of therapy can include electrical stimulation therapy delivered to the heart, neural tissue, other target tissue, cardioversion therapy, defibrillation therapy, or drug therapy. In some examples, the therapy circuit 240 can modify an existing therapy, such as adjusting stimulation parameters or drug dosage.

[0082] In some examples, detected PVCs or PVC load estimates can be provided to system components that perform processes such as detecting other physiological events (e.g., arrhythmias such as atrial fibrillation or ventricular tachycardia), estimating patient cardiac risk, or generating other diagnoses. For example, a high PVC load can trigger detection of atrial fibrillation or ventricular tachycardia, or adjustment of detection algorithm parameters, such as making the algorithm more sensitive or specific to certain arrhythmias. In some examples, information such as known cardiac risk or early signs of a cardiac event can trigger PVC detection and PVC load estimation. For example, a history of atrial fibrillation or a prediction of an impending arrhythmic episode can trigger PVC detection and PVC load estimation, or adjustment of a PVC detection algorithm, such as making it more sensitive to certain types of PVCs.

[0083] Figure 4is a diagram illustrating a method 400 for identifying various types of PVC candidates from sensed heartbeats during the triggering phase of a hierarchical PVC detection process. The method 400 may be implemented in and performed by the PVC candidate identifier 222. At 410, a physiological signal (e.g., an ECG or EGM) from which heartbeats are detected may be preprocessed to reject noisy intervals and heartbeats with very low amplitudes (such as below an amplitude threshold). At 420, the preprocessed signal may be checked for shortened intervals (such as shorter than a baseline cardiac interval by a specific margin during an NSR). A shortened cardiac interval is an indication of the presence of a PVC. If a shortened interval is detected at 420, then at 430, the interval pattern is checked against criteria for triggering identification of PVC single candidates, and against criteria for triggering identification of PVC pair / burst candidates. Criteria for triggering PVC single-shot candidate identification may include a compensatory interval pattern characterized by a prolonged post-ectopic compensatory interval (such that the shortened interval and the prolonged compensatory interval add up to approximately twice the pre-ectopic or baseline cardiac interval), or a pattern of longer (but not strictly compensated) post-ectopic intervals. Additionally or alternatively, criteria for triggering PVC single-shot candidate identification may include a difference or ratio between consecutive cardiac intervals exceeding a threshold indicating a pattern of a shortened pre-ectopic interval followed by a prolonged post-ectopic interval. If the criteria for triggering PVC single-shot candidate identification are met, then the PVC single-shot candidate may be identified at 442. The PVC single-shot candidate may then be confirmed using the single-shot detector 226A.

[0084] Criteria for triggering PVC pair / burst identification may include a fast beat pair of approximately twice the previous (pre-ectopic) or baseline cardiac interval, or a "slow-fast-fast" beat pattern, which indicates a pair with a longer first interval and shortened second and third intervals. If the criteria for triggering a PVC pair / burst candidate are met, a PVC pair / burst candidate may be identified at 444. The PVC pair / burst candidate may then be confirmed and classified as a PVC pair or a PVC burst using pair detector 226B or PVC burst detector 226C, respectively.

[0085] If the short interval check at 420 or the interval pattern check at 430 fails, an amplitude check can be performed at 450. This stage marks beats that have significant amplitude differences relative to the previous beat and constrains adjacent detections that may be oversensed. In one example, the amplitude check at 450 can include an interval stage that uses the average cardiac cycle calculation rather than the previous interval for interval comparison. This handles the case of a single PVC in the context of a very irregular cardiac interval, such as during an atrial tachyarrhythmia (e.g., atrial fibrillation or atrial flutter), when the previous interval may be very short. If the heartbeat passes the amplitude check, a PVC single-shot candidate can be identified at 442, which can then be confirmed using the single-shot detector 226A.

[0086] Figure 5 4 is a diagram illustrating a method 500 for confirming or classifying PVC candidates, such that a PVC single candidate or a PVC pair / burst candidate (such as those identified using method 400) is one of the PVC types in the confirmation phase of the hierarchical PVC detection process. Method 500 may include one or more processes 510 for detecting a PVC single, 520A and 520B for detecting a PVC pair, and 530 for detecting a PVC burst of three or more PVCs. Method 500 may be implemented in and performed by feature extractor 224 and PVC classifier 226.

[0087] Method 500 begins at 501 for receiving a PVC single shot candidate, such as generated at step 442 of method 400, or begins at 502 for receiving a PVC pair / burst candidate, such as generated at step 444 of method 400. For the PVC single shot candidate received at 501, the PVC single shot detection process 510 includes extracting a single shot feature set using, for example, feature extractor 224 at 512. Examples of single shot features may include amplitude-based features, such as the ratio of peak-to-peak (PP) amplitudes measured in two different frequency bands, and morphology-based features, such as CWA-based features (e.g., correlation between morphological features of a heartbeat and morphological features of an NSR template) or DCWA-based features (e.g., correlation between corresponding morphological features of two consecutive heartbeats). At 514, the single shot detector 226A may be used to detect the PVC single shot. In one example, a signal portion (e.g., ECG or EGM) of a received PVC single-shot candidate may be applied to a trained single-shot detection model (e.g., LG model) to extract single-shot features at 512 , and the PVC single-shot is confirmed at 514 when a weighted sum of the single-shot features exceeds a first confidence threshold.

[0088] The PVC shot detected at 514 may then be used in a PVC pair detection process 516 for detecting PVC pairs, such as using the pair detector 226B. At 516, the signal portion of the heartbeat immediately following the detected PVC shot (the "post-shot heartbeat") may be applied to a trained computational model (e.g., an LG model) to identify a set of features from the post-shot heartbeat. The extracted features may include, for example, DCWA-based morphological features and CWA-based morphological features, among other features. If the model output exceeds a confidence threshold, the post-shot PVC is detected. At 518, a PVC pair is detected that includes the PVC shot detected at 514 and the subsequent post-shot PVC detected at 516.

[0089] The PVC pair detected at 518 can also be used in a PVC burst detection process 530, which is used to detect a PVC burst of three or more PVCs, such as using the PVC burst detector 226C. At 532, a PVC burst candidate can be identified based on an analysis of the signal characteristics of the heartbeat immediately following the PVC pair detected at 518 (the "post-pair heartbeat"). The analysis of the signal characteristics can include the consistency of the RR intervals of the cardiac intervals. If the post-pair heartbeat meets the interval consistency condition, then the PVC burst candidate is considered to be detected. At 534, the signal portion of the "post-pair heartbeat" can be applied to a trained computational model (e.g., an LG model) to identify a feature set from the post-pair heartbeat. Among other features, the extracted PVC burst features can include, for example, CWA-based morphological features, the ratio of PP amplitudes measured in two different frequency bands, QRS width, or interval-based features, such as the difference between two consecutive cardiac intervals. If the model output exceeds a confidence threshold, the post-pair PVC is detected. At 536, a PVC burst is detected that includes the PVC pair detected at 518 and a subsequent PVC following the pair detected.

[0090] For a PVC pair / burst candidate including two consecutive PVC candidates (pair-1 and pair-2) received at 502, the PVC pair detection process 520B may include extracting features from pair-1 of the PVC pair candidate at 521, applying a signal portion of pair-1 to a first trained computational model (e.g., LG model) at 522 to confirm the detection of pair-1, extracting features from pair-2 of the PVC pair candidate at 523, and applying a signal portion of pair-2 to a second trained computational model (e.g., LG model) at 524 to confirm the detection of pair-2. The pair-1 features and pair-2 features may be different from each other (e.g., they may have different numbers or types of features, such as amplitude-based features, interval-based features, or morphology-based features). In the event that both pair-1 and pair-2 are confirmed as PVCs, at 525, the PVC pair is considered detected.

[0091] The PVC pair detected at 525, although detected using a different process 520B than the PVC pair obtained by process 520A at 518, can be similarly used in a PVC burst detection process 530 for detecting a PVC burst of three or more PVCs, as described above. The resulting PVC burst detected at 536 includes the PVC pair detected at 525 and the subsequent detected post-pair PVC.

[0092] The PVC singles detected at 514, the PVC pairs detected at 518 and 525, and the PVC bursts detected at 536 may then be used to determine a PVC load estimate at 540, such as using the PVC load estimator 228, as described above with reference to FIG. Figure 2 To avoid double counting of PVC events, if the PVC single shot detected at 514 is identified as the leading PVC in the PVC pair detected at 518, and the PVC pair at 518 is not identified as part of the PVC burst detected at 536, then the PVC pair at 518, rather than the PVC single shot at 514, is counted in the PVC load estimate at 540. Similarly, if the PVC pair detected at 518 or 525 is further identified as part of the PVC burst detected at 536, then only the PVC burst at 536, rather than the PVC pair at 518 or 525, is counted in the PVC load estimate at 540. The PVC load estimate may be provided to the user, such as displayed on a display of the user interface unit 230. In one example, based on the determined PVC load, an alert or recommendation for further evaluation or treatment may be presented to the user. In some examples, treatment may be initiated or adjusted in response to the detected PVC or the PVC load estimate.

[0093] Figure 6 is a diagram illustrating a method 600 for selecting a PVC detection model based on changes in patient conditions, such as NSR morphology. The method 600 may be implemented in and performed by the classification model selector 227 .

[0094] At 610, a triggering event such as a change in posture or body activity can be detected. At 620, in response to the detected triggering event, stability of the NSR morphology can be detected, such as by using a correlation between a previously generated NSR template and a plurality of NSR beats preceding the current PVC candidate. In one example, an average of the correlations can be used to determine the stability of the NSR morphology.

[0095] At 630, the NSR stability may be compared to a threshold to determine whether the NSR is stable (i.e., there is no significant change in the NSR morphology). If the average correlation exceeds the threshold, a stable NSR morphology is indicated. At 642, a computational model utilizing the NSR morphology, such as CWA-based features (e.g., correlation between morphological features of a heartbeat and morphological features of an NSR template), may be used by the PVC classifier 226 (or any one of the detectors 226A-226C) to confirm and classify PVC candidates. In some examples, in addition to the CWA-based features, at 642, DCWA-based features (e.g., correlation between corresponding morphological features of two consecutive heartbeats) may also be included in the computational model. However, if the average correlation is below the threshold, the NSR morphology is considered to change significantly over time and is therefore unstable. At 644, the NSR morphology and CWA-based features may be excluded from the computational model used by the PVC classifier 226 (or any one of the detectors 226A-226C) to confirm and classify PVC candidates. In some examples, instead of CWA-based features, at 644, DCWA-based features (eg, correlation between corresponding morphological features of two consecutive heartbeats) may be included in the computational model.

[0096] Figure 7 A block diagram of an example machine 700 is generally shown on which any one or more of the techniques (eg, methods) discussed herein may be performed. Portions of this description may be applied to the computing framework of various portions of the arrhythmia detection system 200 .

[0097] In alternative embodiments, the machine 700 can be run as a standalone device, or can be connected (e.g., networked) to other machines. In a networked deployment, the machine 700 can be run as a server machine, a client machine, or both in a server-client network environment. In one example, the machine 700 can act as a peer machine in a peer-to-peer (P2P) (or other distributed) network environment. The machine 700 can be a personal computer (PC), a tablet computer, a set-top box (STB), a personal digital assistant (PDA), a mobile phone, a web home appliance, a network router, a switch or a bridge, or any machine capable of executing instructions (sequential or other) specifying the actions to be taken by the machine. In addition, although only a single machine is shown, the term "machine" should also be deemed to include any machine collection that executes a set (or multiple sets) of instructions individually or jointly to perform any one or more of the methods discussed herein, such as cloud computing, software as a service (SaaS), and other computer cluster configurations.

[0098] As described herein, examples may include logic or multiple components or mechanisms, or may be operated by them. A circuit group is a collection of circuits implemented in a tangible entity including hardware (e.g., simple circuits, gates, logic, etc.). Circuit group membership can change flexibly over time and with changes in the underlying hardware. A circuit group includes members that can perform specified operations individually or in combination when operating. In one example, the hardware of the circuit group can be designed to perform specific operations (e.g., hardwired) invariably. In one example, the hardware of the circuit group may include variably connected physical components (e.g., execution units, transistors, simple circuits, etc.), including physically modified computer-readable media (e.g., magnetic, electrical, movable placement of invariant aggregated particles, etc.) to encode instructions for specific operations. When connecting physical components, the basic electrical properties of the hardware components are changed, for example, from insulators to conductors, and vice versa. Instructions enable embedded hardware (e.g., execution units or loading mechanisms) to create members of the circuit group in hardware via variable connections to perform a portion of a specific operation when operating. Therefore, when the device is operating, the computer-readable medium is communicatively coupled to other components of the circuit group members. In one example, any physical component may be used in more than one member of more than one circuit group. For example, in operation, an execution unit may be used in a first circuit in a first circuit group at one point in time and reused by a second circuit in the first circuit group or a third circuit in the second circuit group at a different time.

[0099] The machine (e.g., computer system) 700 may include a hardware processor 702 (e.g., a central processing unit (CPU), a graphics processing unit (GPU), a hardware processor core, or any combination thereof), a main memory 704, and a static memory 706, some or all of which may communicate with each other via an interconnect (e.g., a bus) 708. The machine 700 may also include a display unit 710 (e.g., a raster display, a vector display, a holographic display, etc.), an alphanumeric input device 712 (e.g., a keyboard), and a user interface (UI) navigation device 714 (e.g., a mouse). In one example, the display unit 710, the input device 712, and the UI navigation device 714 may be a touch screen display. The machine 700 may also include a storage device (e.g., a drive unit) 716, a signal generating device 718 (e.g., a speaker), a network interface device 720, and one or more sensors 721, such as a global positioning system (GPS) sensor, a compass, an accelerometer, or other sensors. The machine 700 may include an output controller 728, such as a serial (e.g., universal serial bus (USB), parallel or other wired or wireless (e.g., infrared (IR), near field communication (NFC)), etc.) connection to communicate or control one or more peripheral devices (e.g., printer, card reader, etc.).

[0100] The storage device 716 may include a machine-readable medium 722 on which is stored one or more sets of data structures or instructions 724 (e.g., software) that embody or are utilized by any one or more of the techniques or functionality described herein. The instructions 724 may also reside, completely or at least partially, within the main memory 704, within the static memory 706, or within the hardware processor 702 during execution of the instructions 724 by the machine 700. In one example, one or any combination of the hardware processor 702, the main memory 704, the static memory 706, or the storage device 716 may constitute a machine-readable medium.

[0101] Although the machine-readable medium 722 is illustrated as a single medium, the term “machine-readable medium” may include a single medium or multiple media (eg, a centralized or distributed database and / or associated caches and servers) configured to store one or more instructions 724 .

[0102] The term "machine-readable medium" may include any medium capable of storing, encoding or carrying instructions executed by the machine 700 and causing the machine 700 to perform any one or more of the techniques disclosed herein, or any medium capable of storing, encoding or carrying data structures used by or associated with these instructions. Non-limiting machine-readable medium examples may include solid-state memory and optical and magnetic media. In one example, a large-scale machine-readable medium includes a machine-readable medium having a plurality of particles having a constant (e.g., stationary) mass. Therefore, a large-scale machine-readable medium is not a transient propagation signal. Specific examples of large-scale machine-readable media may include: non-volatile memory, such as semiconductor memory devices (e.g., Electrically Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM)) and flash memory devices; disks, such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM and DVD-ROM optical disks.

[0103] The instructions 724 may also be transmitted or received over the communication network 726 using a transmission medium via the network interface device 720 using any of a variety of transmission protocols (e.g., frame relay, internet protocol (IP), transmission control protocol (TCP), user datagram protocol (UDP), hypertext transfer protocol (HTTP), etc.). Example communication networks may include a local area network (LAN), a wide area network (WAN), a packet data network (e.g., the Internet), a mobile phone network (e.g., a cellular network), a plain old telephone (POTS) network, and a wireless data network (e.g., a wireless network known as a wireless network). The Institute of Electrical and Electronics Engineers (IEEE) 802.11 family of standards, known as The IEEE 802.16 family of standards), the IEEE 802.15.4 family of standards, peer-to-peer (P2P) networks, etc. In one example, the network interface device 720 may include one or more physical jacks (e.g., Ethernet, coaxial cable, or telephone jacks) or one or more antennas to connect to the communication network 726. In one example, the network interface device 720 may include multiple antennas to wirelessly communicate using at least one of single-input multiple-output (SIMO), multiple-input multiple-output (MIMO), or multiple-input single-output (MISO) technology. The term "transmission medium" should be deemed to include any intangible medium that is capable of storing, encoding, or carrying instructions executed by the machine 700, and includes digital or analog communication signals or other intangible media to facilitate the communication of such software.

[0104] Various embodiments are shown in the above-mentioned figures. One or more features from one or more of these embodiments may be combined to form other embodiments.

[0105] The method examples described herein may be implemented at least in part by a machine or computer. Some examples may include a computer-readable medium or machine-readable medium encoded with instructions that are operable to configure an electronic device or system to perform the methods described in the above examples. Implementations of such methods may include code, such as microcode, assembly language code, high-level language code, etc. Such code may include computer-readable instructions for performing various methods. The code may form part of a computer program product. In addition, during execution or at other times, the code may be tangibly stored on one or more volatile or non-volatile computer-readable media.

[0106] The foregoing detailed description is intended to be illustrative rather than limiting.The scope of the disclosure should, therefore, be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled.

Claims

1. An implantable medical device for detecting cardiac arrhythmia in a subject, comprising: a sensor circuit configured to sense a cardiac signal of the subject; as well as A processor circuit configured to: detecting a heartbeat from the sensed cardiac signal; Detection of premature ventricular contractions (PVCs) in the first detection mode includes: Using the cardiac intervals or signal amplitudes of the detected heartbeats to identify specific types of PVC candidates; generating a signal feature set comprising morphological features from the sensed cardiac signal; and and Using the generated signal feature set, classifying the identified PVC candidates into a pattern of a single PVC or multiple consecutive PVCs; and In response to a mode switching trigger event, the specific type of PVC is detected using the cardiac interval or the signal amplitude of the detected heartbeat in a second detection mode different from the first detection mode.

2. The implantable medical device according to claim 1, wherein: The mode switching trigger event includes a threshold time elapsed after the PVC is detected in the first detection mode.

3. The implantable medical device according to claim 2, wherein: The threshold time is a number of weeks.

4. The implantable medical device according to any one of claims 1 to 3, wherein: The processor circuit is configured to determine a PVC load representing a percentage of heartbeats detected as a single shot of the PVC or a pattern of the plurality of consecutive PVCs.

5. The implantable medical device according to any one of claims 1 to 4, wherein: The first detection mode includes at least one trained computational model including respective weighting factors for the set of signal features, the at least one trained computational model being trained to identify one or more PVC types.

6. The implantable medical device according to claim 5, wherein: The at least one trained computational model comprises a logistic regression model, In order to classify the PVC candidates identified in the first detection mode, the processor circuit is configured to: calculating a weighted combination of the generated signal feature sets, each signal feature being weighted by the corresponding weight factor; and Based on the comparison between the weighted combination and the confidence threshold, the identified PVC candidate is classified as the PVC single shot or the pattern of the multiple consecutive PVCs.

7. An implantable medical device according to any one of claims 5 to 6, in, The pattern of the plurality of consecutive PVCs includes a PVC pair or a PVC burst of three or more consecutive PVCs; The identified PVC candidates include one or more of the identified PVC single-transmission candidates, the identified PVC pair candidates, or the identified PVC continuous-transmission candidates. The at least one trained computing model comprises one or more of a trained single shot detection model, a trained joint pair detection model, or a trained PVC burst detection model, In order to classify the identified PVC candidates, the processor circuit is configured to apply the trained single-shot detection model to the identified PVC single-shot candidates, apply the trained pair detection model to the identified PVC pair candidates, or apply the trained PVC continuous-shot detection model to the identified PVC continuous-shot candidates.

8. The implantable medical device according to claim 7, wherein: The processor circuit is configured to: Identify the PVC candidate as the PVC single-issue candidate; as well as The sensed cardiac signal is applied to the trained shot detection model to generate a shot feature set for the identified PVC shot candidate, and the PVC shot detection is confirmed using the shot feature set and the trained shot detection model.

9. The implantable medical device according to claim 8, wherein: The processor circuit is configured to, in response to a confirmed PVC single-shot detection, apply the sensed cardiac signal to a trained post-single-shot PVC detection model to generate a post-single-shot feature set for the heartbeat immediately following the confirmed PVC single-shot, and use the post-single-shot feature set and the trained post-single-shot PVC detection model to confirm the PVC pair detection.

10. The implantable medical device according to claim 7, wherein: The processor circuit is configured to: Based on two consecutive cardiac intervals shorter than an interval threshold, identifying the PVC candidate as the PVC pair candidate; as well as The sensed cardiac signal is applied to the trained pair detection model to generate a first feature set and a second feature set for two consecutive PVCs in the identified PVC pair candidates, respectively, and the PVC pair detection is confirmed using the first feature set and the second feature set and the trained pair detection model.

11. The implantable medical device according to any one of claims 7, wherein: The processor circuit is configured to: identifying the PVC candidate as the PVC burst candidate based on (i) three or more consecutive cardiac intervals that are shorter than an interval threshold, or (ii) a confirmed PVC pair detection; as well as The sensed cardiac signal is applied to the trained PVC burst detection model to generate a third feature set for a third or subsequent PVC of the PVC burst candidate, and the PVC burst detection is confirmed using the third feature set and the trained PVC burst detection model.

12. An implantable medical device according to any one of claims 1 to 11, wherein: The signal feature set generated in the first detection mode includes: one or more amplitude characteristics including a ratio of peak-to-peak (PP) amplitudes between two consecutive heart beats in the sensed cardiac signal; or One or more interval characteristics including a difference in cardiac interval between two consecutive heart beats in the sensed cardiac signal.

13. The implantable medical device according to claim 1, wherein: The signal feature set generated in the first detection mode includes one or more morphological features, and the one or more morphological features include: a first morphology similarity measure between (i) a heartbeat in a sensed cardiac signal and (ii) a normal sinus rhythm (NSR) morphology; or A second morphology similarity measure between two consecutive heart beats in the sensed cardiac signal.

14. The implantable medical device according to claim 13, wherein: One or more of the first morphological similarity measure or the second morphological similarity measure comprises a correlation measure.

15. An implantable medical device according to any one of claims 13-14, wherein: The processor circuit is configured to: Detecting changes in NSR morphology; and If the detected change in the NSR morphology is below a change threshold, the identified PVC candidate is classified using the first morphology similarity metric and the second morphology similarity metric, and if the detected change in the NSR morphology exceeds the change threshold, the identified PVC candidate is classified using the second morphology similarity metric instead of the first morphology similarity metric.

Images