A wearable defibrillator device configured to apply a machine learning algorithm

Abstract

A device is disclosed, which is configured to be worn by a patient for cardiac defibrillation, the device comprising: sensing electrodes, which are configured to sense cardiac signals of the patient; defibrillation electrodes; a therapy delivery circuit system, which is configured to deliver defibrillation therapy to the patient through the defibrillation electrodes; a communication circuit system, which is configured to receive data of at least one physiological signal of the patient from at least one sensing device separate from the device; a memory, which is configured to store the data, the cardiac signals and a machine learning algorithm; and a processing circuit system, which is configured to apply the machine learning algorithm to the data and the cardiac signals to probabilistically determine at least one state of the patient and determine whether to control the delivery of the defibrillation therapy based on the at least one probabilistically determined patient state.

Description

Technical Field

[0001] The present disclosure relates to medical devices, and more particularly, to cardiac defibrillators. Background Art

[0002] Malignant tachyarrhythmias, such as ventricular fibrillation (VF), are uncoordinated contractions of the heart muscle of the ventricles in the heart and are the most common arrhythmias in cardiac arrest patients. If such an arrhythmia persists for more than a few seconds, it can lead to cardiogenic shock and cessation of effective blood circulation. Sudden cardiac death (SCD) can occur within minutes.

[0003] A wearable automatic external defibrillator (WAED), also known as a wearable cardiac defibrillator (WCD), is an option for patients who are identified as being at risk for malignant tachyarrhythmias but for whom an implantable cardioverter-defibrillator is not indicated or desired. A WAED typically includes a strap or garment that carries its components, such as sensing and defibrillation electrodes, processing circuitry, and shock generation and sensing circuitry, which allows the patient to wear such components. A WAED typically implements a conventional AED arrhythmia detection algorithm in which the most recent seconds of electrocardiogram data are analyzed, and the algorithm does not take into account other types of sensor data that are typically not applicable to a WAED. Summary of the invention

[0004] In some aspects, the present disclosure describes examples of systems, devices, and techniques for detecting tachyarrhythmias and, in some cases, defibrillation. A patient may wear a defibrillation device, such as a WAED. The processing circuitry of the device may implement a machine learning algorithm to probabilistically determine the state of the patient, such as determining the state of the patient's tachyarrhythmia and / or co-morbidity, such as chronic obstructive pulmonary disease (COPD), as part of a tachyarrhythmia detection algorithm. The processing circuitry of the device may have a graphics processing unit (GPU) and a central processing unit (CPU) architecture, which may allow the device to implement the machine learning algorithm.

[0005] The machine learning algorithm enables the defibrillator device to implement a probabilistic determination of whether a patient's condition is normal or abnormal, such as whether the patient is experiencing or will experience a treatable tachyarrhythmia, which may also be referred to as the patient's tachyarrhythmia state, using, for example, Bayesian, random forest, and / or decision tree methods. The machine learning algorithm and GPU architecture also enable the device to make probabilistic determinations by considering a large amount of different data together and identifying patterns in the data. The data may include current and historical values ​​(or values ​​derived therefrom) of time-varying signals such as an electrocardiogram (ECG) and / or other sensed physiological or environmental signals.

[0006] In some instances, the device receives data from one or more other sensing devices that may be worn by or implanted in the patient, the data comprising a signal, value, or independent determination of the patient's state for consideration by the GPU based on a machine learning algorithm. The device may be a master device of such other sensing devices in a master / slave communication relationship. Probabilistic determinations based on various types of data and signal sources may allow for more accurate determination of the patient's state, particularly in the face of noise in one or more signals. In some instances, probabilistic determinations may allow for prediction of tachyarrhythmias before they occur, and allow for preventive therapy in response to fibrillation, such as cardiac pacing or other electrical stimulation to disrupt the heart rhythm, rather than defibrillation shocks.

[0007] The device can update the machine learning algorithm based on the collected data through the GPU. In some instances, such as during the initial training phase with patient data, the update to the machine learning algorithm can also be based on feedback from the user or other device on whether the algorithm is correct in determining the patient's state. In addition, in some instances, the remote system can implement a wider GPU architecture, such as one with a larger number of cores or otherwise having greater processing data and the ability to update the algorithm. The remote system can receive the collected data and patient status decisions from the device, update its algorithm instance based on the data and decision, and provide updates to the device. In some instances, different groups can be distinguished from each other based on different characteristics, and the remote system can use group-specific data to update the group-specific machine learning algorithm of the patient within each group. In addition, in the case where a patient wearing such a defibrillator device is later instructed to implant an ICD, the ICD's tachyarrhythmia detection algorithm can be implemented or configured based on a machine learning algorithm stored in the defibrillator device, which has learned to detect tachyarrhythmias in a specific patient over time.

[0008] In one example, a device configured to be worn by a patient for cardiac defibrillation includes: sensing electrodes configured to sense cardiac signals of the patient; defibrillation electrodes; a therapy delivery circuit system configured to deliver defibrillation therapy to the patient through the defibrillation electrodes; a communication circuit system configured to receive data of at least one physiological signal of the patient from at least one sensing device separate from the device; and a memory configured to store the data, the cardiac signal, and a machine learning algorithm. The device further includes a processing circuit system configured to apply the machine learning algorithm to the data and the cardiac signal to probabilistically determine at least one state of the patient and determine whether to control the delivery of the defibrillation therapy based on the at least one probabilistically determined state of the patient.

[0009] In another example, a method for monitoring cardiac signals and determining whether to deliver defibrillation therapy through a device configured to be worn by a patient includes: sensing the patient's cardiac signals through sensing electrodes of the device; receiving data of at least one physiological signal of the patient from at least one sensing device separate from the device through a communication circuit system of the device; storing the cardiac signals, the data, and a machine learning algorithm through a memory of the device; applying the machine learning algorithm to the data and the cardiac signals through a processing circuit system of the device to probabilistically determine at least one state of the patient; and determining through the processing circuit system whether to control the delivery of defibrillation therapy performed through the therapy delivery circuit system of the device based on the at least one probabilistically determined state of the patient.

[0010] In another example, a system for cardiac defibrillation includes a device configured to deliver defibrillation therapy, wherein the device is configured to be worn by a patient, wherein the device includes a processing circuit system, the processing circuit system including a first graphics processing unit (GPU); sensing electrodes, the sensing electrodes configured to sense a cardiac signal of the patient; defibrillation electrodes; a therapy delivery circuit system, the therapy delivery circuit system configured to deliver defibrillation therapy to the patient through the defibrillation electrodes; and a memory configured to store the cardiac signal and a machine learning algorithm. The system further includes a computing system, the computing system communicatively coupled to the device, the computing system including a second GPU, wherein the first GPU is configured to apply the machine learning algorithm to the cardiac signal, and the processing circuit system is configured to determine whether to control the delivery of the defibrillation therapy based on the result of applying the machine learning algorithm to the cardiac signal, and wherein the second GPU is configured to update the machine learning algorithm based on the cardiac signal and group data, wherein the group data includes data of cardiac signals from multiple other patients.

[0011] In another example, a system for determining a patient's tachyarrhythmia state includes: a device configured to be worn by the patient; and a sensing device separate from the device. The device includes sensing electrodes configured to sense cardiac signals of the patient; a communication circuit system configured to receive data of at least one physiological signal of the patient from the sensing device via wireless communication; a memory configured to store the data, the cardiac signal, and a machine learning algorithm; and a processing circuit system configured to apply the machine learning algorithm to the data and the cardiac signal to determine a treatable tachyarrhythmia state of the patient. The processing circuit system of the device is configured to request the data as a master device from the sensing device as a slave device according to a master / slave relationship.

[0012] The present disclosure is intended to provide an overview of the subject matter described in the present disclosure. The present disclosure is not intended to provide an exclusive or exhaustive explanation of the systems, devices, and methods described in detail in the following figures and the specification. Further details of one or more examples of the present disclosure are set forth in the following figures and the specification. Other features, objects, and advantages will become apparent from the description and drawings, as well as the claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a conceptual diagram illustrating an example system configured to apply a machine learning algorithm to a patient's data and provide cardiac defibrillation to the patient in accordance with an example of the presently disclosed technology.

[0014] Figure 2 is a conceptual diagram of another example system configured to apply a machine learning algorithm to a patient's data and provide cardiac defibrillation to the patient in accordance with an example of the presently disclosed technology.

[0015] Figure 3 is a block diagram of an example configuration of a defibrillator device that can implement a machine learning algorithm and provide cardiac defibrillation to a patient according to an example of the presently disclosed technology.

[0016] Figure 4 yes Figure 3 A block diagram of an example configuration of certain components of a defibrillator device.

[0017] Figure 5 is a block diagram of an example configuration of a sensing device according to an example of the technology of the present disclosure, which can sense a physiological phenomenon of a patient.

[0018] Figure 6is a block diagram of an example configuration of an external device configured to receive data from a defibrillation device and update a machine learning algorithm according to an example of the presently disclosed technology.

[0019] Figure 7 is a flow chart illustrating an example technique that may be implemented by a defibrillator device to apply a machine learning algorithm to data and determine whether to deliver therapy.

[0020] Figure 8 is a flow chart illustrating an example technique that may be implemented by a defibrillator device to apply a machine learning algorithm to data including data received from one or more sensing devices and determine whether to deliver therapy.

[0021] Fig. 9 is a flow chart illustrating an example technique that may be implemented by a defibrillator device to apply a machine learning algorithm to data to characterize a patient's state as normal or abnormal.

[0022] Fig.10 is a flow chart illustrating an example technique that may be implemented by an external device to determine updates to a machine learning algorithm for a different population based on population data.

[0023] Fig.11 is a flowchart showing the different learning stages of a machine learning algorithm.

[0024] Fig.12 is a flow chart illustrating an example technique for configuring a tachyarrhythmia detection algorithm of a patient's implantable cardioverter-defibrillator based on data previously evaluated by a machine learning algorithm of the patient's defibrillator device.

[0025] Fig.13 is a diagram illustrating an example technique for parsing data evaluated by a machine learning algorithm to configure a tachyarrhythmia detection algorithm for an implantable cardioverter-defibrillator.

[0026] Fig.14 is a flow chart illustrating an example technique for applying a machine learning algorithm of a defibrillator device to data to determine whether an indication of chronic obstructive pulmonary disease is provided.

[0027] Fig.15 is a graph illustrating data from a cardiac signal that a machine learning algorithm would learn to classify as normal, according to an example of the presently disclosed techniques.

[0028] Figure 16-22 is a graph illustrating data from a cardiac signal that a machine learning algorithm would learn to classify as abnormal, according to an example of the presently disclosed techniques.

[0029] Fig.23 and 24is a graph showing RR interval slope data that a machine learning algorithm according to an example of the presently disclosed technology would learn to classify as normal and abnormal, respectively.

[0030] Fig.25 is a graph illustrating data of cardiac signals that a machine learning algorithm from patients in different populations would learn to classify as abnormal, according to an example of the presently disclosed techniques.

[0031] Figure 26-28 is a graph illustrating data of cardiac signals from patients in different populations that a machine learning algorithm would learn to classify as normal, according to an example of the presently disclosed techniques.

[0032] Fig.29 is a graph illustrating data of a spirometer signal from a patient that a machine learning algorithm would learn to classify as normal according to an example of the presently disclosed techniques.

[0033] Fig.30 is a graph illustrating data from a patient that a machine learning algorithm would learn to classify as abnormal, in accordance with an example of the presently disclosed techniques.

[0034] Fig.31 is a graph illustrating data of a spirometer signal from an elderly patient that a machine learning algorithm would learn to classify as normal according to an example of the presently disclosed techniques.

[0035] Fig.32 is a graph showing data from a spirometer signal under varying air quality conditions. DETAILED DESCRIPTION

[0036] In some aspects, the techniques described herein include: collecting data from one or more sources; applying a machine learning algorithm to the data; determining whether to deliver therapy based on the results of applying the machine learning algorithm to the data; and if it is determined to deliver therapy, delivering the therapy. The machine learning algorithm can be implemented by a defibrillator device, such as a GPU of a WAED. The machine learning algorithm can be updated based on the patient's data and data from a larger population. The machine learning algorithm can be configured to characterize the patient data as normal or abnormal or otherwise determine one or more states of the patient.

[0037] As an example, a patient may wear a vest device (e.g., WAED) having a GPU and also having a subcutaneously implanted insertable heart monitor. Each of these devices may sense cardiac signals of the patient's heart. The vest device may control the insertable heart monitor to perform various tasks, such as transmitting the sensed signals to the vest device. In other words, the vest device may act as a master device of the insertable heart monitor in a master / slave relationship.

[0038] Based on the signal, the machine learning algorithm implemented by the GPU of the vest device can determine that the patient suffers from an arrhythmia, and the vest device can be configured to deliver therapy to the patient to treat the arrhythmia. The patient (or healthcare professional) can confirm, for example, through a user interface of the vest device or another device whether an adverse event has actually occurred. In this way, the patient and / or healthcare provider can contribute to the development of the machine learning algorithm, such as providing near real-time reinforcement for the reinforcement learning of the algorithm. In some instances, the heart monitor or another device can make an independent determination of an arrhythmia, which can provide near real-time feedback (reinforcement) for the reinforcement learning development of the machine learning algorithm. Based on previous events and signals, the GPU can update the machine learning algorithm.

[0039] Figure 1 is a conceptual diagram illustrating an example system 100 that can be used to deliver therapy to the heart of a patient 102, such as providing therapy for ventricular fibrillation. System 100 can include a device 110, one or more sensing devices 130, a network 150, and one or more external devices 170. In one instance, device 110 can be worn by patient 102. Device 110 can be a vest device that includes clothing, electronics, and electrodes, as further described herein. Device 110 can be configured to monitor the heart of patient 102 and can be configured to provide therapy. In some instances, device 110 includes one or more sensing electrodes that are configured to sense a phenomenon of patient 102 (e.g., a physiological signal such as a cardiac signal). In some instances, device 110 is a WAED.

[0040] One or more sensing devices 130 are configured to sense phenomena of patient 102 and / or the patient's environment. Device 110 may be configured to sense phenomena of the same or different patient 102 as one or more sensing devices 130. Figure 1 As shown, device 110 and one or more sensing devices 130 may communicate via one or more links 120. In some examples, link 120 may be Links, such as Low Energy (BLE) link.

[0041] In some examples, device 110 acts as a master device for one or more sensing devices 130 according to a master / slave relationship. One or more sensing devices 130 acting as slaves of device 110 can improve the overall sensitivity of a machine algorithm implemented by device 110, for example, by improving true positives while avoiding false positives due to signal interference and false signals. In addition, communication according to a master / slave relationship can limit communication-related consumption of power supplies of one or more sensing devices 130 relative to other more frequent communication schemes. In such a relationship, device 110 requests data from one or more sensing devices 130 only when needed, such as when there is a preliminary indication that the patient's status may be abnormal based on the cardiac signals (and possibly other data) sensed by device 110.

[0042] In some instances, device 110 will use a designated magnetic, radio frequency (RF), or electrical signal to wake up sensing device 130. In some instances, once a connection is established between device 110 and sensing device 130, periodic notifications may maintain the connection. In some instances, device 110 may use an organizational or globally unique identifier to distinguish between sensing devices 130. In some instances, communications between device 110 and sensing device 130 may generally be based on Or BLE protocol.

[0043] The network 150 may represent any single network or combination of networks that facilitates communication between devices. As an example, the network 150 may represent a combination of wireless and wired networks (e.g., the Internet) that facilitate communication between one or more external devices 170 and the apparatus 110 (and / or the sensing device 130). The external devices 170 and the network 150 may include a remote patient monitoring system, such as the one available from Medtronic plc of Dublin, Ireland. In some examples, the one or more external devices 170 include one or more servers and one or more personal computers with which a healthcare provider can interact through a user interface. In some examples, the system 100 includes multiple external devices 170 (e.g., a remote patient monitoring system and one or more personal computers). In some examples, the one or more external devices 170 may include a cloud-based computing system.

[0044] External device 170 may be configured to receive patient data from apparatus 110 and / or sensing device 130 and store the data in a memory. External device 170 may store data collected from a patient population. In some instances, population data includes information about patient 102, but in other instances, population data does not necessarily include information about patient 102.

[0045] As will be described in more detail below, the device 110 may include a GPU that applies a machine learning algorithm to the data of the patient 102 to, for example, determine whether fibrillation is detected and whether to deliver defibrillation therapy. The device 110 may deliver defibrillation therapy based on the determination made by the GPU. The data may include data received from one or more sensing devices 130. The GPU may update the machine learning algorithm, i.e., the algorithm may learn to better classify future patient data based on the patient data. The external device 170 may also include a GPU and determine updates to the machine learning algorithm based on the patient data and the population data, and provide the updates to the defibrillation device 110.

[0046] Figure 2 is a conceptual diagram of a system 200, the system being Figure 1 1 and 130 of the defibrillator device 110 and the sensing device 130 of the system 100. The system 200 may include the device 210, which is an example of the device 110. As shown, the system 200 includes sensing devices 230A-230D (collectively referred to as "sensing devices 230"), but in some examples, the system 200 may include fewer, more, or different sensing devices 230. In general, the patient 102 may be implanted with, wear, or interact with any one or more of the sensing devices 230. The sensing device 230 is Figure 1 An example of a sensing device 130.

[0047] System 200 can include strap 204. In some examples, strap 204 is part of device 210, and strap 204 is configured to secure a portion of device 210 to patient 102 (e.g., a battery pack connected to the electronics of device 210). Typically, device 210 is configured to be worn by patient 102.

[0048] System 200 may be configured for cardiac defibrillation. For example, system 200 includes various sensors (e.g., device 210 and one or more of sensing devices 230A-230D) configured to sense signals, such as physiological signals of patient 102 and characteristics of the patient's environment. For example, device 210 may include sensing electrodes and associated sensing circuitry configured to sense cardiac signals of patient 102. In some instances, device 210 may be a WAED.

[0049] Sensing device 230 such as sensing device 230B can include a sensor (e.g., an electrode and an associated sensing circuit system) configured to sense a cardiac signal of patient 102. Therefore, the heart of patient 102 can be the source of the signal sensed by device 210, and can also be the source of the signal sensed by sensing device 230B. However, due to factors such as the position on the body of patient 102 or the type of sensor used, the cardiac signals sensed by different devices may be slightly different (e.g., the type or degree of noise or motion artifacts). In this way, by including multiple devices, system 200 is configured to determine the state of the heart of patient 102 (e.g., normal or abnormal) based on multiple signals. In addition, using multiple devices to sense signals from the same source (e.g., from the heart of patient 102 or from other parts of patient 102) helps to verify the determination of the patient's state. In this way, system 200 can have reduced false alarms of abnormal patient states (e.g., false alarms of treatable tachyarrhythmia states) relative to systems using only one device, which can also be referred to as false positives. In some instances, false positives due to signal interference and false signals that may be specific to one device or sensing modality may be avoided. Avoiding false alarms can help avoid delivering unnecessary therapy (e.g., inappropriate shock). In these and other ways, system 200 may be more reliable than a system that relies on only one device to determine the status of patient 102.

[0050] The device 210 may include a communication circuit system configured to receive a sensed signal and / or other data derived from the sensed signal from the sensing device 230. As used herein, as an example, the term data may refer to a signal, data derived from a signal, and a determination made by any device based on a signal or other data. In some instances, the device 210 includes a memory configured to store data received from the sensing device 230, a cardiac signal sensed by the device 210, and a machine learning algorithm, among others. The device 210 includes a processing circuit system of a processing circuit system as further described herein, which may include a GPU. In some instances, the GPU is configured to apply a machine learning algorithm to data, such as one or more physiological signals received from the sensing device 230 and a cardiac signal sensed by the device 210. The processing circuit system may be configured to determine whether to control the delivery of defibrillation therapy to the patient 102 (e.g., by the therapy delivery circuit system and defibrillation electrodes of the device 210). In some instances, this determination of the processing circuit system is based on the result of applying the machine learning algorithm to the physiological signal and the cardiac signal. For example, the processing circuit system may be configured to probabilistically determine one or more states of patient 102 based on applying a machine learning algorithm to the data, and control the delivery of therapy or take one or more other actions based on the one or more determined patient states.

[0051] An example of sensing device 230 is sensing device 230A. Sensing device 230A may include a headband, a hat, etc. configured to be worn on the head of patient 102, and one or more sensors are positioned on the head. Such sensors may include electroencephalogram (EEG) sensors, such as electrodes configured to sense EEG signals and associated sensing circuit systems. The sensed EEG signal is an example of the physiological signal described above. For example, device 210 can determine the patient state based on the EEG signal to, for example, determine whether to deliver defibrillation therapy. Different waveform morphologies and timing data of EEG such as learned from patient and / or population data can be associated with supraventricular or ventricular sources of tachyarrhythmias, and therefore can be used to determine the patient's treatable tachyarrhythmia state by a machine learning algorithm, such as whether the patient state is a treatable tachyarrhythmia. The EEG signal data considered by the machine learning algorithm may include data derived from another measure of the amount and / or pattern of the variability of a Lorenz plot or an EEG signal. In some examples, the device 210 can diagnose specific cardiac conditions, such as atrial fibrillation (AF), through heart rate variability (HRV) based on ECG and EEG signal data. The device 210 can also determine one or more comorbid conditions, such as epilepsy or stroke, or other conditions that may co-occur with arrhythmias, based on the EEG signals or other data derived from the EEG signals.

[0052] Another example of sensing device 230 is sensing device 230B. Sensing device 230B can be a cardiac monitor (e.g., an implantable cardiac monitor). For example, sensing device 230B can be a Reveal®, available from Medtronic, Inc. of Dublin, Ireland. In the form of an insertable cardiac monitor (ICM). As described above, the apparatus 210 may sense a first cardiac signal from the heart of the patient 102 , while the patient device 230B senses a second cardiac signal from the heart of the patient 102 .

[0053] The sensing device 230B may also include sensors configured to sense other signals indicative of other physiological phenomena of the patient 102. For example, the sensing device 230B may be configured to sense the temperature, posture, activity, blood oxygenation, and tissue perfusion of the patient 102. Further, although primarily described herein as a cardiac monitor, in other instances, the sensing device 230B may be any implantable medical device (IMD) configured to sense one or more physiological signals of the patient and in some instances configured to deliver therapy, such as an implantable pacemaker, an ICD, a neurostimulator, an implantable pressure sensor, or a drug delivery device. The device 210 may determine the patient state based on the signals sensed by the sensing device 230B or other data derived from the sensing device, for example, to determine whether to deliver defibrillation therapy.

[0054] Another example of sensing device 230 is sensing device 230C. Sensing device 230C can be worn on patient 102, such as on a limb (e.g., an arm or wrist). Sensing device 230C is a wearable device, such as a watch or activity monitor that includes one or more sensors. Sensing device 230C can be configured to sense the heart rate, activity, and concentrations of various substances in the fluid of patient 102, such as salt in blood or sweat. Device 210 can determine the patient state based on the signal sensed by sensing device 230C or other data derived from the sensing device, for example, to determine whether to deliver defibrillation therapy.

[0055] Another example of sensing device 230 is sensing device 230D. Sensing device 230D can be a mobile computing device (e.g., a mobile phone). In this way, sensing device 230D can include a microphone, a camera, a processing circuit system, and other components that can be used to determine information about patient 102 or related to the patient. For example, patient 102 can take photos of food or beverages consumed by patient 102. As another example, sensing device 230D can use GPS or other positioning technology to track the location of patient 102. System 200 can automatically track these data and determine the patient state from the data. For example, system 200 can consider information about the food consumption of patient 102, such as when the patient consumes a specific type of food (e.g., electrolytes) that may affect the health of patient 102 above the normal amount (e.g., above the previously determined baseline).

[0056] The location data collected by sensing device 230D (or another sensing device 230 or device 210) can be used by the systems described herein in various ways. For example, the location data can indicate the activities performed by the patient and the environment or environmental conditions to which the patient is exposed. Device 210 can use this data when determining the status of the patient, such as with respect to treatable tachyarrhythmias, respiratory disorders (such as COPD), or other comorbidities that may be particularly affected by the environment to which the patient is exposed. Further, data collected for a given patient can be tagged with location data. In some examples, external device 170 ( Figure 1 ) can use location data to segment different patient groups for population-based learning via machine learning algorithms.

[0057] Another example of sensing device 230 is a cardiovascular pressure monitoring device. The cardiovascular pressure monitoring device can be an implantable pressure monitoring device that can be implanted in a chamber of the heart, a pulmonary artery, or another cardiovascular location, such as the pressure monitoring device described in the commonly-assigned U.S. Patent No. 8,864,676 of Beasley et al. and U.S. Patent Publication No. 2016 / 0220825 of Greenhut et al. In other examples, the cardiovascular pressure monitoring device can be an external pressure monitoring device, such as a blood pressure measurement system based on a cuff. As an example, the cardiovascular pressure signal generated by such a sensor can include values ​​of systolic pressure, diastolic pressure, or mean pressure, including pulmonary artery pressure or peripheral vascular pressure. The pressure signal (e.g., the morphology or trend of the signal) can indicate a current or predicted treatable tachyarrhythmia state or comorbidity, such as heart failure and COPD.

[0058] In some examples, one or more sensing devices 230 can provide signals or values ​​for pulse rate, oxygen saturation, and respiratory rate. In some examples, the sensing devices include an integrated pulse oximetry system that provides values ​​or signals for these parameters. Trends or patterns in these parameters can indicate current or predicted treatable tachyarrhythmic states or comorbidities such as heart failure and COPD.

[0059] Although primarily described herein as implanted or wearable, the sensing device may be any device configured to collect data about a patient or the patient's environment. Another example of the sensing device 230 may be a glucose monitor configured to sense a physiological signal (e.g., glucose concentration) of the patient 102. Other examples of the sensing device 230 include a spirometer, a medical imaging device, a food scale, or an air quality sensing device.

[0060] In general, as further described herein, a machine learning algorithm can use signals sensed by one or more sensing devices or other data derived from such signals to determine a state of patient 102. For example, device 210 can determine a tachyarrhythmia state of the patient based on the determined state, such as whether the patient is experiencing or will likely experience a treatable tachyarrhythmia, and whether to deliver defibrillation or another therapy configured to terminate or prevent tachyarrhythmia. Device 210 can be configured to deliver therapy based on the determination. In some instances, device 210 can be configured to additionally determine whether a comorbidity, such as COPD, has manifested in patient 102 based on applying a machine learning algorithm to such data.

[0061] Figure 3 is a block diagram of an example configuration of a defibrillator device 310, which may be Figure 1 Device 110 or Figure 2 310. The device 310 may include a communication circuit system 312, a memory 314, a user interface 318, a processing circuit system 322, a power supply 328, a sensing circuit system 330, a sensing electrode 332, a therapy delivery circuit system 334, and a defibrillation electrode 336. The sensing circuit system 330 is electrically coupled to the sensing electrode 332, and the therapy delivery circuit system 334 is electrically coupled to the defibrillation electrode 336. The processing circuit system 322 may include a GPU 324 and a CPU 326. The memory 313 may be configured to store one or more algorithms 316.

[0062] Communications circuitry 312 includes circuitry for communicating with another device (e.g., any device described herein, such as Figure 1 The sensing device 130 or Figure 1The communication circuit system 312 may be configured to transmit or receive radio frequency ("RF") signals via an antenna (not shown), or to transmit or receive other signals via a wired connection to another device or network. The communication circuit system 312 may be configured to perform such communications with other devices such as the sensing device 130 or 230 or the external device 170. As an example, the communication circuit system may include resistors, inductors, capacitors, amplifiers, and / or transistors configured to generate, modulate, filter, and / or demodulate signals according to any of a variety of communication protocols. In some instances, the communication circuit system 312 may be coupled to one or more electrodes and configured with similar circuit systems to transmit and receive signals through the electrodes for tissue conduction communication (TCC).

[0063] In general, any device described herein may be configured to communicate with any other device (e.g., any device described herein may include communication circuitry similar to the communication circuitry of communication circuitry 312). In some examples, communication circuitry 312 may be configured to communicate with any other device via To communicate (e.g., transmit or receive RF signal). For example, communication circuit system 312 can be configured to transmit or receive BLE radio signals. In some cases, communication circuit system 312 includes a BLE module. In some examples, multiple components of system 100 are connected through Connect to each other.

[0064] In some examples, communication circuitry 312 includes input circuitry configured to receive a signal from another device. In some examples, communication circuitry 312 includes output circuitry configured to transmit information from apparatus 310 to another device. For example, communication circuitry 312 may include circuitry configured to transmit information about a patient's status to external device 170.

[0065] The communication circuitry 312 may communicate with another device or network (e.g., Figure 1 The communication circuit system 312 may also send a command to another device to send information to the device 310. For example, the communication circuit system 312 may be configured to send a command to the sensing device 130 or 230 to instruct the sensing device to transmit a physiological signal or other signal or data derived therefrom to the device 310.

[0066] Memory 314 may store instructions that cause processing circuit system 322 to provide the functionality described herein, as well as information used by processing circuit system 322 to provide the functionality attributed to device 310 as described herein. Memory 314 may include any volatile, non-volatile, magnetic, optical, or electrical media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), electrically erasable programmable ROM (EEPROM), flash memory, or any other digital or analog media. Memory 314 may also include removable memory portions that may be used to provide memory updates or increases in memory capacity. Removable memory may also allow patient data to be easily transferred to another computing device or removed before device 310 is used for another purpose (e.g., upgrading, cleaning, or adjusting the size of a garment of device 310).

[0067] The memory 314 may be configured to store one or more algorithms 316. For example, the algorithm 316 may include a machine learning algorithm. The machine learning algorithm may be configured to determine one or more states of the patient 102. In one example, the machine learning algorithm is applied to data about the patient 102, such as a physiological signal or data derived from the signal, by a processing circuit system 322, such as a GPU 324. The machine learning algorithm may be automatically updated by the processing circuit system 322 based on new data stored in the memory 314, such as to continue learning. In the case of a signal, the data may include amplitude and time information (e.g., an electrocardiogram (ECG) with measured voltages over time). Typically, the data may include values ​​that change over time, and the machine learning algorithm may take into account features of such signals or changes in such features over time.

[0068] The memory 314 may be configured to store data about the patient 102 or the patient's environment. In some instances, such data is sensed by the sensing device 130 or 230. For example, the memory 314 may be configured to store data about the patient's environment at a particular time, such as indicated by location data, fluid status information (e.g., hydration level or edema), and patient health record information. The memory 314 may be configured to store data about the patient 102 in a table, list, or graphical format.

[0069] The user interface 318 may be configured or otherwise operable to receive input from a user such as a patient 102 or a healthcare provider. The user interface 318 may be configured to display information to the user. For example, the user interface 318 may include one or more lights, a display, a motor configured to provide a vibration alarm (e.g., similar to the "vibration" mode on a smartphone), a speaker configured to indicate an alarm to the user. The patient 102 may interact with the user interface 318, which may include a display configured to present a graphical user interface to the patient, and a keypad or another mechanism for receiving input from the patient.

[0070] In some instances, the patient 102 or a caregiver interacts with the user interface 318 to provide information to the system 100. For example, the patient 102 or the caregiver can facilitate the development of the machine learning algorithm by providing feedback about events. For example, the user interface 318 can indicate to the patient 102 that a therapy is indicated based on the state of the patient 102, but the patient 102 can override the delivery of the therapy (e.g., if the patient 102 knows that the therapy should not be indicated or is detected as a "false positive"). For example, the user interface 318 includes one or more buttons (e.g., digital or physical buttons). Although in Figure 3 310, but the user interface through which a patient or another user can provide feedback (also referred to as reinforcement) for reinforcement learning by the machine learning algorithm can be a user interface provided by any one or more computing devices. For example, the user interface can be a user interface of a smart phone, a wearable device (e.g., a smart watch), or a dedicated device for communicating with the device 310.

[0071] The patient or caregiver provides information about false alarms through a user interface, and the machine learning algorithm can be adjusted based on the new information about the false alarms. Reinforcement from the user can be in near real time, such as once the patient or caregiver reacts to a determination made by the machine learning algorithm. As an example, adjustments to the machine learning algorithm can include adjusting weights and / or connections between nodes.

[0072] Processing circuitry 322 may be configured to perform the techniques described herein. Processing circuitry 322 may include fixed-function processing circuitry or programmable processing circuitry and may include, for example, one or more of a microprocessor, a controller, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), discrete logic circuitry, or any other processing circuitry configured to provide the functionality attributed to processing circuitry 322 herein that may be embodied as firmware, hardware, software, or any combination thereof. Processing circuitry 322 may control Figure 1In some examples, processing circuit system 322 can control other devices (eg, patient device 130 or 230).

[0073] In some examples, processing circuit system 322 includes GPU 324. In general, GPU 324 can be configured to apply machine learning algorithms to data collected by device 310, such as by sensing circuit system 330 and / or from sensing device 130 or 230, which data can include patient signals (e.g., one or more physiological signals from the sensing device and a cardiac signal sensed by sensing circuit system 330). GPU 324 can be configured to automatically update the machine learning algorithm, such as based on the collected information, updates determined by external device 170 based on population data, determinations made by other devices (such as sensing device 130 or 230), or patient or other user input. GPU 324 can include multiple parallel cores, which enable data sets containing a variety of different data from a variety of sources to be applied to the machine learning algorithm in parallel.

[0074] refer to Figure 1 , external device 170 may also include a GPU. In one example, GPU 324 of device 310 may include a first GPU. The GPU of the external device may include a second GPU. In some examples, any processing circuitry described herein including first and second GPUs may be configured to perform techniques attributed to processing circuitry 322 or processing circuitry of an external device.

[0075] Based on learning from, for example, a data set about patient 102, a machine learning algorithm can discover patterns and relationships among several independent and interdependent variables that can be derived from the data. Continuous consideration of these variables or other learned variables from continuous updates of the algorithm can allow the machine learning algorithm to probabilistically determine, for example, classify, diagnose and / or predict the state of the patient. As an example, the machine learning algorithm can be configured to employ any one or more of Bayesian, random forest, decision tree, linear regression, deep learning, neural network and / or dimensionality reduction techniques. In some instances, the result of applying the machine learning algorithm to the data, such as one or more physiological signals from sensing device 130 or 230 and / or a cardiac signal from sensing circuit system 330, includes a classification of the data, such as whether the cardiac signal and the physiological signal (alone or together) are normal or abnormal, or indicate one or more other states of the patient, such as whether a treatable tachyarrhythmia is indicated or predicted or whether one or more comorbidities are indicated or predicted.

[0076] In some instances, the processing circuit system includes a CPU 326. In general, the CPU 326 can be configured to control the activities of components of the device 310, such as the communication circuit system 312, the user interface 318, the sensing circuit system 330, and the therapy delivery circuit system 324. The CPU 326 can also be configured to obtain data for consideration by the GPU 324 and control the functions of the GPU 324. In some instances, the CPU 326 is configured to execute an arrhythmia detection algorithm (e.g., a treatable tachyarrhythmia detection algorithm). The arrhythmia detection algorithm can be stored in the memory 314. In some instances, the CPU 326 is configured to update the tachyarrhythmia detection algorithm based on the results of applying the machine learning algorithm to the data (e.g., the physiological signals sensed by the patient device and the cardiac signals sensed by the device). The CPU 326 is configured, for example, to control the delivery of defibrillation therapy (e.g., whether to deliver defibrillation) based on the results of applying the machine learning algorithm to the physiological signal cardiac signal. In some instances, CPU 326 may be configured to control the delivery of other therapies by therapy delivery circuitry 334, such as therapy configured to prevent a predicted arrhythmia, based on the results of applying the machine learning algorithm to the physiological and cardiac signals.

[0077] The power supply 328 delivers operating power to the various components of the device 310. The power supply 328 may include a small rechargeable or non-rechargeable battery and power generation circuitry for generating operating power. Recharging may be accomplished through a wired connection to a voltage source or through proximal inductive interaction between an external charger and an inductive charging coil within the device 310. In some examples, the power supply 328 may include a replaceable battery, and the device 310 may be configured so that the patient 102 can access a cavity in which the power supply 328 is stored to replace the battery.

[0078] In some examples, the sensing circuit system 330 can be configured to generate a signal. In some examples, the sensing circuit system 330 includes amplifiers, filters, analog-to-digital converters, and other circuit systems configured to generate and condition signals for receipt by the processing circuit system 322. The sensing circuit system 330 can be coupled to the sensing electrodes 332 and can receive cardiac electrical signals, for example, from various combinations of two or more sensing electrodes 332. The sensing circuit system 330 can be configured to sense cardiac events associated with depolarization and repolarization of cardiac tissue.

[0079] The sensing circuit system 330 may include one or more sensing channels, each of which may be selectively coupled to a corresponding combination of sensing electrodes 332 to detect electrical activity of the heart. Different sensing electrodes 332 may be positioned within the device 310 at locations that effectively measure cardiac signals of the patient 102. Each sensing channel may be configured to amplify, filter, and rectify cardiac electrical signals received from selected electrodes coupled to the corresponding sensing channel to detect cardiac events (such as R-waves), and may also include an analog-to-digital converter to provide a digitized representation of the time-varying signal to the processing circuit system 322. The processing circuit system 322, such as the GPU 326, may determine from the sensed cardiac signal whether the cardiac signal indicates a normal state or an abnormal state, and if abnormal, determine whether the cardiac signal indicates one of a plurality of abnormal sub-classifications, such as bradycardia, treatable tachyarrhythmia, syncope, from signal noise (e.g., 60 Hz noise), motion artifacts, or signal loss. In some instances, in addition to sensing cardiac signals, the device 310 may include sensors for sensing any of the various physiological or other signals described herein.

[0080] The data set applied to the machine learning algorithm can include various data from various sources, and the data includes values ​​that vary over time. In the case of a signal, the data can include amplitude and time information, such as an electrocardiogram (ECG) with a measured voltage over time. Typically, the machine learning algorithm can consider the characteristics of the signal formed by the value that varies over time or the change of such features over time. For example, in the case of a cardiac signal, the variable may be related to the characteristics of the cardiac signal, such as P wave, R wave, QRS complex, ST segment, QT interval and T wave and heart rate and heart rate variability. The variable can be related to the morphology of the signal, such as the maximum or minimum amplitude or width of the slope, the area under the curve or the morphological characteristics of the signal. Other variables related to the morphology of the signal can include values ​​or features identified by Fourier or wavelet transforms, turning point algorithms or other signal transformation or decomposition techniques. The variable can include differences over time and / or relative to the baseline of such features, the rate of change of such features or the differences of such features between different sensing vectors. Signals having characteristics that can be evaluated by a machine learning algorithm in this way are not limited to physiological signals, but may instead be the value of any measurable or derivable parameter over time or other parameters derivable from the ECG, such as RR interval length or variability, which values ​​may form a trend-based curve.

[0081] The therapy delivery circuit system 334 can be coupled to the defibrillation electrode 336. In some instances, the therapy delivery circuit system 334 is configured to deliver a therapy shock to the heart of the patient 102, such as based on a determination that the patient 102 has an abnormal heart state (such as an occurring or predicted treatable tachyarrhythmia). For example, if the patient 102 suffers from fibrillation or other treatable tachyarrhythmia, the device 310 can deliver therapy to the patient 102 through the defibrillation electrode 336. In some instances, the therapy delivery circuit system 334 can be configured to deliver various types of therapy (e.g., pacing and / or defibrillation, or in some instances, drug therapy). In some instances, the therapy is preventive, such as pacing or drug therapy configured for a treatable tachyarrhythmia predicted by a machine learning algorithm. In instances where the therapy includes electrical pulses, such as pacing or defibrillation, the therapy delivery circuit system 334 can include one or more capacitors, charge pumps, current sources, or other signal generation circuit systems, and switch circuit systems for coupling the signal to the electrode 336. In instances where the therapy includes drug therapy, therapy delivery circuitry 334 may include circuitry configured to generate a signal to drive a pump.

[0082] Figure 4 is a block diagram illustrating an example configuration of certain components of device 310. Figure 4 As shown, device 310 may include various elements connected to each other by flexible material 438. Specifically, GPU 324 is connected to circuit system 400 and circuit system 400 is connected to power supply 328 by flexible material 438. Circuit system 400 may include communication circuit system 312 and CPU 326, and in some examples, sensing circuit system 330 and therapy delivery circuit system 334 ( Figure 4 ). Generally, the separation of the relatively large GPU 326 and power supply 328 from other components of the device 310 and the flexible material connections can provide the device 310 with a particular form factor, such as a device that can be more easily worn by the patient 102.

[0083] Flexible material 438 can provide electrical connections between components of the device, for example, can include conductors extending through, on, or between layers of flexible material. In some examples, flexible material 438 includes a rigid flexible electronic substrate construction. Such a construction can enable various form factors for device 410, such as a body band (e.g., a wearable band with embedded electronics that can include shoulder straps) or a headband with sensors integrated therein.

[0084] Figure 5 is a block diagram of an example configuration of a sensing device 530, which may be Figure 1 The sensing device 130 and / or Figure 2 In the example shown and similar to the example with respect to Figure 2 The sensing device 530 may be a Reveal® sensing device 230B as described above, available from Medtronic plc in Dublin, Ireland. In other examples, the patient device 530 may be any of the sensing devices described herein (e.g., Figure 2 The sensing device described herein) or other types of sensing devices.

[0085] Figure 6 is a block diagram of an external device 670, which may be Figure 1 An example of an external device 170. Figure 5 and 6 Both, the following elements can have Figure 3 The similar elements described herein may have at least the same configuration and functionality. For example, communication circuit system 512 and communication circuit system 612 may have the same or similar functionality as communication circuit system 312. For example, memory 515 and memory 614 may have the same or similar functionality as memory 314. Although one or more algorithms 616 are shown as being stored in memory 614, memory 614 may be configured to store other information, such as any information described herein. Similarly, memory 514 may be configured to store any information described herein.

[0086] Processing circuitry 522 and processing circuitry 622 may have the same or similar functionality as processing circuitry 322. Processing circuitry 622 is illustrated as including GPU 624, which may have the same or similar functionality as GPU 324 (in some instances, except as noted herein), and CPU 626, which may have the same or similar functionality as CPU 326. Power supplies 528 and power supplies 628 may have the same or similar functionality as power supply 328. User interface 618 may have the same or similar functionality as user interface 318 or any other interface described herein.

[0087] The sensing circuit system 530 and the sensing circuit system 630 may have the same or similar functions as the sensing circuit system 330. The one or more sensors 540 and the one or more sensors 640 may be used, for example, with respect to Figure 2The patient device 230 of the example is in the form of a sensor described. In some instances, such sensors may include accelerometers, optical sensors, pressure sensors, air or blood flow sensors, temperature sensors, or air quality sensors. One or more electrodes 532 may have the same or similar functions as the sensing electrodes 332, for example, the sensing circuit system 530 may be configured to sense physiological signals of the patient 102, such as cardiac signals, through the electrodes. In some instances, the cardiac signals include information such as about the RR interval, the amplitude of the R-wave, the QRS width, or the variability of the RR interval. In one instance, the input to the machine learning algorithm includes these data from the cardiac signals.

[0088] In some examples, external device 670 and / or sensing device 530 may be configured to detect information about the environment of patient 102. For example, external device 670 and sensing device 530 may be configured to determine the temperature and / or quality of the air in the environment in which patient 102 is located.

[0089] In some examples, the memory 514 of the sensing device is configured to store a tachyarrhythmia detection algorithm, and the processing circuit system 522 executes the tachyarrhythmia detection algorithm. The processing circuit system 522 can store in the memory 514 an indication of whether a tachyarrhythmia is detected at various times. The processing circuit system 522 of the sensing device 530 can provide such indications to the device 310 via the communication circuit system 512. As described herein, the GPU 324 of the device 310 can use these indications of the determination of tachyarrhythmia by the sensing device 530 as input for a machine learning algorithm to assist in classifying the current patient state, and / or as feedback for adaptation of the machine learning algorithm.

[0090] Rather than continuous or periodic communication, the device 310 and the sensing device 530 can communicate in a master / slave relationship as a master and a slave, respectively. Communication according to a master / slave relationship can limit communication-related consumption of the power supply 528 of the sensing device 530 relative to other more frequent communication schemes. In such a relationship, the device 310 only requests data from the sensing device 530 for use by the machine learning algorithm when needed, such as when there is a preliminary indication that the patient's state may be abnormal based on the cardiac signal sensed by the device 310 (and possibly other data).

[0091] The algorithm 616 (eg, Figure 6624 can include one or more machine learning algorithms, and the GPU 624 can be configured to implement the machine learning algorithms. In some instances, the GPU 624 is more computationally capable than the GPU 324 of the device 310. The greater computational power may be due to various factors, such as the number of cores or clock speed. In the case of the external device 670, then particularly in the case where the device 310 is wearable by the patient 102, the size of the processing circuitry, control power consumption, and heat may be less of a concern than with the device 310.

[0092] GPU 624 may be configured to implement a relatively more computationally intensive version of a machine learning algorithm than GPU 324, which may implement a “lite” version of a machine learning algorithm. In some instances, GPU 624 updates the machine learning algorithm differently than GPU 324. For example, GPU 624 may update the machine learning algorithm based on population data, and in doing so, update the machine learning algorithm based on a significantly larger dataset than GPU 324. CPU 626 may control communication circuitry 612 to send the determined update to device 310, and GPU 326 may update its instance of the machine learning algorithm as indicated by external device 670.

[0093] The machine learning algorithm implemented by the GPU 324 of the device 310 can be updated in various ways. For example, the GPU 324 can automatically update the machine learning algorithm based on the data it receives and considers. In some cases, learning can be reinforced by independent determinations such as feedback from a user or from another device (e.g., the sensing device 530). In addition, the GPU 324 can update the machine learning algorithm as indicated by the update determined by the GPU 624 when considering the group data. In this way, the machine learning algorithm implemented by the GPU 324 is updated based on the group data. In some instances, the population-based updates to the machine learning algorithm can include new weights for the variables of the normal variance factors applied to different groups. Typically, updates can also include updates to the graph structure of the algorithm, or updates to the parameters / variables considered by the algorithm, which may be more intensive and determined by the GPU 624.

[0094] The CPU 626 may receive data sets from a plurality of different devices 310 for a plurality of different patients 102 via the communication circuitry 612. The CPU 626 may also receive or otherwise determine information indicating into which patient population or populations the received data should be classified. The GPU 624 may implement a plurality of population-specific machine learning algorithms and apply the data for a given population to the algorithm for that population in order to determine updates to the machine learning algorithms for the devices 310 for the patients in those populations.

[0095] Different groups may have different characteristics. For example, each group may be defined by a unique combination of values ​​for multiple characteristics that distinguish the groups. Example characteristics include age, sex, location, body mass index (BMI), weight, blood pressure (ranges such as low, ideal, pre-high or critically high, and high), respiration, glucose levels, history of adverse medical events, and the presence of other comorbid medical conditions.

[0096] The normal and / or abnormal ranges of various parameters or variables that are input to the machine learning algorithm may vary between different patient populations. For example, resting heart rate for bradycardia is expected to increase with age, and resting normal heart rate is expected to increase with BMI and blood pressure. Adapting a machine learning algorithm based on a data set of patients with similar characteristics to each other can increase the algorithm's ability to correctly characterize the state of a given patient within a population.

[0097] Figure 7 is a flow chart illustrating an example technique that may be implemented by defibrillator device 310 to apply a machine learning algorithm to data and determine whether to deliver therapy. For example, device 310 receives data (700). The data includes a cardiac signal sensed by device 310. The data may also include data from a sensing device, such as another cardiac signal or other physiological signal or data derived therefrom.

[0098] The device 310 applies a machine learning algorithm to the data to characterize the data and therefore one or more states of the patient as normal or abnormal (702). In some instances, the machine learning algorithm may indicate whether a treatable tachyarrhythmia state is abnormal, such as indicating the presence of fibrillation or another shockable tachyarrhythmia. In some instances, the machine learning algorithm may additionally or alternatively indicate whether the state of one or more other conditions (e.g., comorbidities) of the patient is normal or abnormal, such as whether or not the onset or exacerbation of COPD or another comorbidity is present or predicted. The device 310 decides whether to provide therapy based on the characterization, for example, in response to certain abnormal characterizations, and in some cases, decides which of a plurality of therapies to apply (704). The therapy may be a defibrillation shock. If the characterization is a predicted treatable tachyarrhythmia, the therapy may be cardiac pacing or another therapy configured to prevent the predicted tachyarrhythmia.

[0099] Device 310 also updates the machine learning algorithm based on the considered data (706). For example, GPU 324 can update the machine learning algorithm autonomously and / or based on feedback from a user or another device (e.g., sensing device 230). As described herein, GPU 324 can also update the machine learning algorithm based on population data, such as based on population-based updates received from external device 670. Population-based updates can occur less frequently than autonomous and reinforced updates.

[0100] Figure 8 is a flow chart illustrating an example technique that may be implemented by a defibrillator device 310 (referred to in the figures as a WAED) to apply a machine learning algorithm to data including data received from one or more sensing devices 530 and determine whether to deliver therapy. Figure 8 In the example of FIG. 3 , GPU 324 applies a machine learning algorithm to a cardiac signal sensed by sensing circuit system 330 via electrode 332 (800). Based on applying the machine learning algorithm to the cardiac signal, GPU 324 determines whether the preliminary characterization of the patient's state is normal (802). If the preliminary characterization is normal (“yes” of 802), GPU 324 proceeds to apply the algorithm to a new data set derived from the cardiac signal sensed by device 310 (800).

[0101] If the preliminary characterization is not normal ("No" of 802), CPU 326 receives data from one or more sensing devices 530 via communication circuit system 312 (804). For example, device 310 acting as a master in a master / slave relationship can initiate a communication session with one or more sensing devices 530 to command the sensing devices to provide data. The data may include cardiac signals or other physiological or environmental signals sensed by the sensing devices, data derived therefrom, and / or determinations made by the sensing devices based on the signals / data.

[0102] The GPU 324 applies a machine learning algorithm to the cardiac signal and data from one or more sensing devices 530 (806). Based on this application of the algorithm, the GPU 324 characterizes a state, such as a treatable tachyarrhythmia state, as normal or abnormal (808). The CPU 326 determines whether to provide therapy to the patient 102 based on the characterization (810). If the CPU 326 determines that therapy should be delivered ("yes" of 810), the CPU 326 can control the therapy delivery circuit system 334 to deliver therapy, such as pacing or defibrillation, through the electrodes 336 (812). In some instances, the CPU 326 can take other actions based on the characterization, such as changing the sensing vector or the settings of an adjustable filter or amplifier to improve noise sources based on the signal being characterized as noisy. Although in Figure 8 Not shown, but GPU 324 can update the machine learning algorithm, for example, based on the characterization of the data and, in some cases, feedback received from a user or other device regarding the characterization of the data (such as whether a patient canceled delivery of therapy).

[0103] although Figure 7 and 8Examples include delivery of therapy in response to an abnormal manifestation, such as a defibrillation shock, but therapy need not be delivered in all examples of the technology according to the present disclosure. In some examples, the machine learning algorithm determines the patient state without the device responsively delivering therapy. Such determinations can be recorded in the memory of the device, transmitted to an external device over a network, and / or presented to a user through a user interface. For example, during the reinforcement learning phase of the machine learning algorithm, the determination of the patient state without therapy delivery may occur.

[0104] Fig. 9 is a flow chart illustrating an example technique that may be implemented by the defibrillator device 310 to apply a machine learning algorithm to data to characterize a patient state as normal or abnormal. Fig. 9 In some examples, GPU 324 applies a machine learning algorithm to the data (900) and determines whether the data indicates a normal state (902). As examples, the machine learning algorithm can employ Bayesian, random forest, decision tree, linear regression, deep learning, neural network, and / or dimensionality reduction techniques. In some examples, the machine learning algorithm employs a Bayesian network with different weights for different factors (e.g., variables) that may be derived from the ECG and other data. In some examples, the spectrum of values ​​represented by the heat map, such as about Figure 15-32 Those spectra described can be used to associate weights with some outputs. If the data indicates a normal state ("yes" of 902), the GPU 324 returns the normal characterization to, for example, the CPU 326 (904). In some instances, the machine learning algorithm is probabilistic and is configured to indicate an abnormal state if the probability of a normal state is lower than a certain percentage, such as 95%, 99%, 99.9%, or any range between any of these values. These values ​​are instances, and other (e.g., lower) percentages can be used in some instances. The sensitivity and specificity of the machine learning algorithm may increase over time because the algorithm learns based on how many learning cycles and the population size from which the data for learning is derived. In some instances, the probability may initially have a first value, such as 95%, but increase to another value, such as 99.9% or greater, over time.

[0105] Example variables that can be derived from ECG data and considered by the machine learning algorithm include R wave amplitude, RR interval length, RR interval variability, QRS width or slope, area under the curve, or other morphological features of the signal formed by the values ​​of R wave amplitude, RR interval length, RR interval variability, or QRS width over time. In some instances, a rolling window of values ​​for each variable is considered (e.g., three minutes or some other length). In some instances, a window of normal training set data for the variable will have values ​​that meet a boundary condition that is at least a threshold amount of time (e.g., a percentage) during the window, such as 99%. The following is about Figure 15-32 Examples of normal and abnormal data sets and associated boundary conditions for different variables are discussed.

[0106] In some examples, if the signature is abnormal ("No" of 902), the algorithm considers the data (and in some cases additional data) to determine an abnormal sub-signature (906). The algorithm structure for determining an abnormal sub-signature can be a fault decision tree. Possible abnormal sub-signatures include bradycardia, a treatable tachyarrhythmia, an untreatable tachyarrhythmia (e.g., atrial fibrillation or another supraventricular tachyarrhythmia), syncope, 60 Hz noise, motion artifact, signal loss, lead peeling, or high RR interval variability, such as due to atrial fibrillation.

[0107] Fig.10 is a flow chart illustrating an example technique that may be implemented by an external device 670 to determine updates to a machine learning algorithm for a different population based on population data. Fig.10In an example of the present invention, an external device 670 receives data from multiple different devices of multiple different patients (1000). The external device 670, such as a CPU 626, classifies the data into multiple different patient groups (1002). As described herein, different patient groups can be distinguished from each other based on age, gender, location, obesity classification, weight or body mass index (BMI), blood pressure, respiration, glucose level, medical history, the presence of other comorbid medical conditions (such as COPD or diabetes), the type and / or model of defibrillator equipment, and one or more types and / or one or more models of available sensing devices. Based on the classified data, the GPU 624 determines updates to different machine learning algorithms for different patient groups (1004). As described herein, as an example, updates to the machine learning algorithm can include new weights for variables, updates to the graph structure of the algorithm, or updates to parameters / variables considered by the algorithm. The external device 670, such as a CPU 626 and the communication circuit system 612 transmit updates for each group to each device in the group (1006). Updates may be provided to the defibrillator device through any wired or wireless connection (eg, through the Internet) and in some cases through one of the sensing (slave) devices acting as an intermediary.

[0108] Fig.11 6 is a flow chart illustrating different learning stages of a machine learning algorithm. Initially, a machine learning algorithm may be designed and trained, for example, using one or more external devices 670. Initial training may include applying a number of validated data sets to the algorithm, such as data sets validated by domain experts as having a particular classification (1100). In some instances, the validated data sets are data sets from a population with particular characteristics, and the machine learning algorithm is intended to be used in a device for patients within the population.

[0109] When the device 310 is used by the patient, the machine learning algorithm can undergo a period of reinforcement learning (1102). During reinforcement learning, the device 310 can receive feedback from the user through the user interface 318 and / or from another device (e.g., the sensing device 530) regarding whether the machine learning algorithm's determination of the state is correct. In some instances, new data (including new patterns in variables and relationships between variables) available from the patient (or a population of similar patients) can be reviewed and classified by one or more doctors for additional reinforcement learning by the machine learning algorithm. After a period of time, such as after the reinforcement learning has reached the desired sensitivity and specificity levels in the algorithm, human intervention may no longer be required. The machine learning algorithm can continue to learn autonomously (1104).

[0110] Fig.12is a flow chart illustrating an example technique for configuring a tachyarrhythmia detection algorithm of a patient's implantable cardioverter-defibrillator (ICD) based on data previously evaluated by a machine learning algorithm of the patient's defibrillation device 310 . Fig.12 The example techniques of may be performed by external device 670 and / or one or more other computing devices.

[0111] according to Fig.12 , indicating that patient 102 has an ICD implanted, at which point patient 102 will likely stop using device 310 (1200). External device 670 retrieves a data set that is characterized by a machine learning algorithm of device 310 of patient 102 as abnormal and, in some cases, as a tachyarrhythmia (1202). External device 670 processes the data set to identify features in the data that are detectable by a tachyarrhythmia detection algorithm of the ICD (1204). The tachyarrhythmia algorithm of the ICD may not include a GPU or a machine learning algorithm and may not be able to detect tachyarrhythmias in the same manner as device 310. As an example of identifying features of a data set detectable by an ICD, external device 670 may divide an ECG waveform of a particular length from a data set classified by device 310 (or a sequence or signal of variable values ​​derived from the ECG waveform) into multiple shorter sequential waveforms or smaller sets of sequential values ​​of a variable. External device 670 may identify features in sequential waveforms or sets of values ​​that the ICD may detect in sequence (rather than in parallel as may have been done by GPU 324) to detect tachyarrhythmias. External device 670 configures the ICD tachyarrhythmia detection algorithm based on the identified features (1206).

[0112] Fig.13 is a diagram illustrating an example technique for parsing data evaluated by a machine learning algorithm to configure a tachyarrhythmia detection algorithm for an implantable cardioverter-defibrillator. Specifically, Fig.13 A heat map of two minutes of consecutive R-wave amplitude values ​​is shown. In the heat map, the R-wave amplitude values ​​are binned within the 0.1 minute intervals in which the R-wave occurs and within one of ten 0.1 mV amplitude ranges between 0 and 1 mV. In this example heat map, each 0.1 minute interval contains six R-waves and, accordingly, six R-wave amplitude values.

[0113] Fig.13The R-wave amplitude values ​​in the heat map of can represent a set of consecutive R-wave amplitude values ​​that would be characterized by the machine learning algorithm as abnormal—particularly a sequence of R-wave amplitude values ​​in the more variable amplitude region 1210 of the heat map. Although the data considered by the machine learning algorithm is presented herein in the form of a heat map, the consideration of the data by the machine learning algorithm according to the techniques described herein does not necessarily involve converting the data into a heat map format. Rather, the data is presented in a heat map format to facilitate illustrating certain features of the data.

[0114] like Fig.13 As shown, the processing circuit system (e.g., the processing circuit system 622 of the external device 670) can parse the continuous amplitude values ​​represented by the area 1210 into shorter segments represented by boxes 1212-1218. The processing circuit system can further determine the characteristics of the data that the ICD can identify based on the parsed data segments. For example, based on the data represented by boxes 1212-1218, the processing circuit system can determine the threshold level of R wave amplitude variability, indicating or predicting an abnormal state above the threshold level, such as a treatable tachyarrhythmia state. As an example, the threshold level of variability determined from the segments represented by boxes 1212-1218 can be a threshold amount of variability within a specific time period, a threshold amount of variability in each of N continuous time periods, and / or a threshold amount of variability in M ​​continuous time periods in N continuous time periods. The threshold level of variability can be programmed into the ICD of a specific patient and used by the ICD, for example, together with other tachyarrhythmia detection or discrimination techniques, to identify treatable tachyarrhythmias.

[0115] Processing circuitry (e.g., processing circuitry 322 of device 310) may use Fig.13 The inverse of the parsing techniques presented is used to merge shorter data segments, for example from sensing device 230, together into a larger data set, such as for a trend-based curve applied to a machine learning algorithm. In some instances, the length of a continuous data value segment stored by sensing device 230 and transmitted to device 310 may be limited, for example due to the memory, communication, or other capabilities of sensing device 230. When device 310 receives data from sensing device 310, for example as requested according to a master-slave relationship, the shorter segments may be time-stamped, and the processing circuit system may combine the shorter segments to form a larger data set applied to the machine learning algorithm. In some instances, any known data padding technique may be applied to the data received from sensing device 230 to produce a data set with sufficient value for consideration by the machine learning algorithm.

[0116] Fig.14is a flow chart illustrating an example technique for applying a machine learning algorithm of a defibrillator device 310 to data to determine whether an indication of chronic obstructive pulmonary disease is provided. Although the device 310 may be primarily used to detect and treat tachyarrhythmias, the GPU 324, the ability to collect a variety of data from the sensing device 530, and the ability to adapt the machine learning algorithm based on population data collected by the external device 670 may enable the device 310 to become a platform to additionally or alternatively determine whether the patient 120 is experiencing various comorbid conditions or disorders. One example comorbidity that may be assessed by the device 310 is COPD.

[0117] according to Fig.14 In the example of FIG. 1 , device 310 receives spirometer data and environmental data from one or more sensing devices 530 (e.g., a spirometer and an environmental sensor) (1300). GPU 324 applies a machine learning algorithm (or a different portion thereof for assessing COPD) to the data (1302). Figure 28-30 Example data that can assess COPD is described. The machine learning algorithm characterizes the data (1304). In some examples, the GPU 324 compares each of the spirometer and environmental data to one or more thresholds and determines the COPD state of the patient using a decision tree based on whether the thresholds or a particular combination of thresholds for different data are met. The CPU 326 determines whether to provide an indication of COPD for the patient 102 to the external device 670, for example, via the user interface 318 or via a network, based on the characterization (1306). The GPU 324 updates the machine learning algorithm based on the data and the characterization (1308).

[0118] Fig.15 is a graph showing data from a cardiac signal that a machine learning algorithm will learn to classify as normal (e.g., a treatable tachyarrhythmia state is normal or a non-treatable tachyarrhythmia) according to an example of the presently disclosed techniques. More specifically, Fig.15 It is a heat map drawing of the combination of RR interval and R wave amplitude that appears in the normal data set, wherein each beat is binned in one of the 0.1mV amplitude bins between 0 and 1mV and one of the 0.1 second RR interval length bins between 0 and 2 seconds. Region 1400 shows the distribution of such values ​​in the normal data set. Specifically, region 1400 contains consistent R wave amplitudes within the 0.1mV range (including 0.5mV), and beats distributed in the 0.1 second RR interval length bins between 0.5 and 1 second as shown, wherein the maximum number of beats has an RR interval length between about 0.7 and 0.9 seconds. The distribution or variability of R wave amplitude, RR interval, R wave amplitude as a function of RR interval, and R wave amplitude as a function of RR interval are examples of variables that machine learning algorithms can consider to characterize patient data.

[0119] Figure 16-22 is a graph showing data from a cardiac signal that a machine learning algorithm would learn to classify as abnormal according to an example of the presently disclosed techniques. For example, Fig.16 1402 is a heat map plot of the occurrence of a combination of RR interval and R wave amplitude in an abnormal data set, and more specifically a data set that would be classified as bradycardia, wherein each beat is binned within one of the 0.1 mV amplitude bins between 0 and 1 mV and within one of the 0.1 second RR interval length bins between 0 and 3 seconds. Region 1402 illustrates the distribution of such values ​​in the bradycardia set. Specifically, region 1402 contains some variation in R wave amplitude, and as shown, beats distributed within 0.1 second RR interval length bins between 1.5 and 2.4 seconds, with the largest number of beats having RR interval lengths between approximately 0.7 and 0.9 seconds. The pattern of the combination of RR interval and R wave amplitude values ​​associated with bradycardia may be scattered or checkered as illustrated by region 1402 and relative to Fig.15 The normal data set presented is towards larger RR interval values ​​(e.g. Fig.16 to the right in the image.

[0120] As another example, Fig.17 1404 is a heat map plot of the occurrence of a combination of RR interval and R wave amplitude in an abnormal data set, and more specifically a data set that would be classified as a tachyarrhythmia (e.g., a treatable tachyarrhythmia), where each beat is binned within one of the 0.1 mV amplitude bins between 0 and 1 mV and within one of the 0.1 second RR interval length bins between 0 and 2 seconds. Region 1404 illustrates the distribution of such values ​​in the tachyarrhythmia data set. Specifically, region 1404 contains significant variation in R wave amplitude between 0.3 and 0.6 mV, particularly at shorter RR interval lengths and a large number of beats with RR interval lengths between 0.3 and 0.5 seconds, which is consistent with the majority of beats having RR interval lengths greater than 0.6 seconds. Fig.15 In contrast to the normal data set. Fig.17 The pattern exhibited can be described as containing a significant grouping of beats with shorter than normal RR interval lengths and higher than normal R-wave amplitude variability, e.g. Fig.17 The left side of the heat map.

[0121] although Fig.17The dataset is generally presented as one that would be classified as a tachyarrhythmia, but the dataset may have distinguishing characteristics (e.g., apparent in a heat map) that would allow classification of specific types of treatable or untreatable tachyarrhythmias, such as ventricular fibrillation, ventricular tachycardia, or supraventricular-tachycardia. Such different patterns of beat groupings of different tachyarrhythmias may be apparent, for example, in a heat map that plots discrete variables derived from the ECG (e.g., RR interval length and R wave amplitude) relative to waveform morphology characteristics (e.g., slope, area under the curve, or values ​​from a transform or turning point algorithm).

[0122] As another example, Fig.18 is a heat map plot of the combination of RR interval and R wave amplitude that occurs in an abnormal dataset and more specifically a dataset that is classified as syncope, characterized by a sudden rate and amplitude change in the heart signal followed by a flat line. Fig.18 In the example of , each beat is binned into one of the 0.1 mV amplitude bins between 0 and 1 mV and one of the 0.1 second RR interval length bins between 0 and 3 seconds. Region 1406 illustrates the distribution of such values ​​in the syncope dataset. Specifically, region 1406 contains significant variation in R wave amplitude, and two distinct clusters of beats with RR interval lengths between 0.7 and 1.2 seconds and greater than 2.9 seconds, respectively. Syncope beats with significantly longer RR interval lengths also have lower R wave amplitudes.

[0123] As another example, Fig.19 is a heat map plot of the combination of RR interval and R wave amplitude that occurs in an abnormal dataset and more specifically a dataset classified as 60 Hz noise, characterized by slight sinusoidal-like variations in R wave amplitude. Fig.19 In the example of , each beat is binned into one of the 0.1 mV amplitude bins between 0 and 1 mV and one of the 0.1 second RR interval length bins between 0 and 3 seconds. Region 1408 shows the distribution of such values ​​in the 60 Hz noisy data set, characterized by Fig.15 The normal data set showed a pattern of greater R-wave amplitude variability, but the RR interval length was consistent with that of the normal data set.

[0124] As another example, Fig. 20 is a heat map plot of the combination of RR interval and R wave amplitude that occurs in abnormal datasets and more specifically datasets that are classified as having signal loss (e.g., due to one or more electrodes failing, being removed, or otherwise becoming inoperable), characterized by noise associated with loss and reconnection of signal sources. Fig. 20In the example of , each beat is binned into one of the 0.1 mV amplitude bins between 0 and 1 mV and one of the 0.1 second RR interval length bins between 0 and 3 seconds. Region 1410 illustrates the distribution of such values ​​in the signal loss data set, containing distinct clusters of different R wave amplitudes associated with the loss and normal data, but RR interval lengths consistent with those of the normal data set.

[0125] As another example, Fig.21 It is a heat map plot of the combination of RR interval and R wave amplitude that occurs in abnormal datasets and more specifically in datasets classified as electrode peeling, characterized by a gradual decrease in amplitude until signal loss. Fig.21 , each beat is binned into one of the 0.1 mV amplitude bins between 0 and 1 mV and one of the 0.1 second RR interval length bins between 0 and 3 seconds. Region 1412 illustrates the distribution of such values ​​in the signal loss data set, including a large number of beats with lower R wave amplitudes and longer RR interval lengths associated with signal loss (e.g., due to lost beats).

[0126] As another example, Fig. 22 is a heat map plot of the combination of RR interval and R wave amplitude that occurs in abnormal datasets and more specifically in datasets that are classified as having RR interval variability (e.g., due to atrial fibrillation). Fig. 22 In the example of , each beat is binned into one of the 0.1 mV amplitude bins between 0 and 1 mV and one of the 0.1 second RR interval length bins between 0 and 3 seconds. Region 1414 illustrates the distribution of such values ​​in the RR interval variability data set, including rapid and relatively large changes in the time interval between the peak of the R wave and its amplitude, such as due to sporadic conduction of rapid atrial depolarizations of atrial fibrillation. The pattern illustrated by region 1414 includes Fig.15 The normal data set shows a larger number of beats within a wider RR interval range.

[0127] Fig.23 and 24 is a graph showing RR interval slope data that a machine learning algorithm according to an example of the technology disclosed herein will classify as normal and abnormal, respectively. Specifically, Fig.23 and 24 The RR interval slope values ​​measured for consecutive beats during a two minute period are shown, where each beat is binned into the 0.1 minute bin in which it occurs and the length of the mean or median RR interval is X1. Fig.23 As shown, the normal data set has consistent RR interval slope values ​​within the range of 0.1*X1 including X1. Fig.24As shown, the abnormal data set contains significantly greater changes in the RR interval slope over time, with RR interval values ​​between 0.8*X1 and 1.4*X1.

[0128] Fig.25 is a graph showing data from patients in different groups that a machine learning algorithm would learn to classify as abnormal cardiac signals according to an example of the presently disclosed techniques. Fig.25 In the example of , each beat is binned within one of the 0.1 mV amplitude bins between 0 and 1 mV and one of the 0.1 second RR interval length bins between 0 and 3 seconds. More specifically, region 1416 shows the distribution of bradycardia data sets in relatively young patients, while region 1418 shows the distribution of bradycardia data sets in relatively older (e.g., elderly) patients. With increasing age, a slightly elevated resting heart rate is expected, and bradycardia data for older patients similarly contains shorter RR intervals. Consistent with this expectation, the beat pattern shown in region 1418 is shifted to shorter RR intervals (e.g., Fig.25 To account for these differences between patient populations, a machine learning algorithm can be adapted using a dataset from patients in the population with characteristics that match those of the patient.

[0129] Figure 26-28 is a graph illustrating data of cardiac signals from patients in different populations that a machine learning algorithm would learn to classify as normal, according to an example of the presently disclosed techniques. Figure 26-28 is a heat map plot of the combinations of RR intervals and R wave amplitudes that appear in the data. Figure 26-28 In the example of , each beat is binned into one of 0.1 mV amplitude bins between 0 and 1 mV and between 0 and 3 seconds (although Fig.28 Only the portion of the heat map between 0 and 1.4 s is shown within one of the 0.1 s RR interval length bins.

[0130] For example, Fig.26 Region 1420 shows the data distribution for patients with average or normal BMI, while region 1422 shows the data distribution for patients with increased BMI. It is expected that resting heart rate increases with increasing BMI, which is reflected in the relative shift of, for example, about 0.1 seconds in the location of the maximum population value combination between regions 1420 and 1422. Specifically, region 1420 contains 100 beats with RR intervals in the range of about 0.8 seconds to about 0.9 seconds, while region 1422 contains 90 beats with RR intervals in the range of about 0.7 seconds to about 0.8 seconds. Similarly, Fig. 27Regions 1424 and 1426 in FIG. 1 show the data distribution for males and females, respectively, where females are expected to have a resting rate that is approximately 5-10 bpm higher than males. Fig.28 Data in region 1428 that would be expected from a patient with higher blood pressure is shown. Fig.28 Also shown is a region 1430 where data for patients with lower blood pressure is expected to appear in the heat map. Fig.28 It shows how increased blood pressure is associated with increased heart rate and how they are co-factors of each other. Again, to account for these differences between patient groups, a data set from patients in a group with characteristics that match the patient can be used to adapt the machine learning algorithm. Characteristics that can distinguish patient groups include gender, age, BMI, weight, and blood pressure. Further, the machine learning algorithm can use the expected differences in the data shown in areas 1428 and 1428 to determine the status of comorbidities related to the patient's blood pressure.

[0131] Fig.29 is a graph showing data from a spirometer signal that a machine learning algorithm from a patient would learn to classify as normal according to an example of the presently disclosed technology. Specifically, Fig.29 A heat map showing volume values ​​changing over time during exhalation into a spirometer is shown. Region 1432 shows the distribution of such values ​​that would be classified as normal.

[0132] The values ​​that can be determined from the spirometer signal include the forced expiratory volume in one second (FEV1) and the forced vital capacity (FVC), which is the amount of air that can be forcefully exhaled from the lungs after the deepest breath possible. FVC is essentially equivalent lung capacity and can be calculated as the integral or area under the curve of the signal from the spirometer during this exhalation. FEV1 can be calculated similarly, but only based on the first second of the signal. The ratio of FEV1 to FVC can also be determined. Normal values ​​for these parameters are FEV1 and FVC greater than 80%, and FEV1 / FVC greater than 70%.

[0133] Fig.30 is a graph showing data from a patient's spirometer signal that a machine learning algorithm would learn to classify as abnormal according to an example of the presently disclosed technology. Specifically, Fig.30 A heat map showing the volume values ​​changing over time during exhalation into a spirometer. Region 1434 shows the distribution of such values ​​that would be classified as abnormal. The distribution of region 1434 contains Fig.29Area 1432 of lower volume values, and associated with about 35% FVC and about 20% FEV1, with a ratio of about 50%. Typically, COPD will demonstrate a decrease in FEV1 and FVC below 80%, and a decrease in FEV1 / FVC below 70%.

[0134] Fig.31 is a graph showing data from a spirometer signal from an elderly patient that a machine learning algorithm would learn to classify as normal according to an example of the presently disclosed technology. Specifically, Fig.31 A heat map of volume values ​​over time during exhalation into a spirometer is shown. Region 1436 shows the distribution of such values ​​that would be classified as normal in elderly patients. The distribution in region 1436 is associated with an FEV1 / FVC ratio of approximately 65%. To account for such differences in spirometry differences between groups with different characteristics (such as age), a machine learning algorithm for classifying COPD can be adapted based on a population-specific dataset, as described herein.

[0135] Fig.32 is a graph showing data from a spirometer signal under varying air quality conditions (e.g., with varying air particle counts). Certain ranges of particle counts are unhealthy for sensitive patients (such as those with COPD), while particle counts above this range need to be considered unhealthy for otherwise healthy patients. However, in general, increased particle counts may have an effect on normal human lung volume, as well as being associated with COPD. Region 1438 shows the distribution of spirometer data with fluctuating particle counts, characterized by greater variance and rapid changes. Fluctuations in particle counts may be associated with fluctuations in FEV1, FVC, and FEV1 / FVC. A machine learning algorithm for classifying COPD may consider particle counts along with other data (such as spirometer data) to explain such fluctuations.

[0136] In some instances, machine learning as described herein is an aspect of artificial intelligence. The systems and techniques described herein may include artificial intelligence, including reasoning, natural language processing, machine learning, and planning. Typically, machine learning may include iterative learning cycles, such as supervised learning, unsupervised learning, reinforcement learning, and deep learning networks.

[0137] The techniques described in the present disclosure may be implemented, at least in part, in hardware, software, firmware, or any combination thereof. For example, various aspects of the described techniques may be implemented within one or more processors, such as fixed-function processing circuitry and / or programmable processing circuitry, including one or more microprocessors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or any other equivalent integrated or discrete logic circuitry, and any combination of such components. The term "processor" or "processing circuitry" may generally refer to any of the aforementioned logic circuitry, alone or in combination with other logic circuitry, or any other equivalent circuitry. A control unit including hardware may also perform one or more of the techniques of the present disclosure.

[0138] Such hardware, software, and firmware may be implemented in the same device or in separate devices to support the various operations and functions described in this disclosure. In addition, any of the described units, modules, or components may be implemented together or individually as discrete but interoperable logic devices. Depicting different features as modules or units is intended to highlight different functional aspects, and does not necessarily imply that such modules or units must be implemented by separate hardware or software components. On the contrary, the functions associated with one or more modules or units may be performed by separate hardware or software components, or integrated in common or separate hardware or software components.

[0139] The techniques described in the present disclosure may also be embodied or encoded in a computer-readable medium (such as a computer-readable storage medium) containing instructions. The instructions embedded or encoded in the computer-readable storage medium may cause a programmable processor or other processor to perform the method, for example, when executing the instructions. The computer-readable storage medium may include a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, a hard disk, a CD-ROM, a floppy disk, a cassette, a magnetic medium, an optical medium, or other non-transitory computer-readable medium.

[0140] Various aspects of the disclosure have been described. These and other aspects are within the scope of the following claims and examples.

[0141] Example 19. A method for monitoring cardiac signals and determining whether to deliver defibrillation therapy through a device configured to be worn by a patient, the method comprising sensing the patient's cardiac signals through sensing electrodes of the device; receiving data of at least one physiological signal of the patient from at least one sensing device separate from the device through a communication circuit system of the device; storing the cardiac signals, the data, and a machine learning algorithm through a memory of the device; applying the machine learning algorithm to the data and the cardiac signals through a processing circuit system of the device to probabilistically determine at least one state of the patient; and determining through the processing circuit system whether to control the delivery of defibrillation therapy performed through the therapy delivery circuit system of the device based on the at least one probabilistically determined state of the patient.

[0142] Example 20. A method according to Example 19, wherein the at least one sensing device includes a subcutaneous implantable cardiac monitor, the subcutaneous implantable cardiac monitor includes a plurality of sensing electrodes for sensing the physiological signal, wherein the cardiac signal includes a first cardiac signal and the physiological signal includes a second cardiac signal.

[0143] Example 21. A method according to any of Examples 19 or 20, wherein the machine learning algorithm is configured to automatically update itself based on at least one of the data or the cardiac signal.

[0144] Example 22. A method according to any one of Examples 19 to 21, wherein the machine learning algorithm is configured to automatically update itself based on input from the patient and / or healthcare provider.

[0145] Example 23. A method according to any one of Examples 19 to 22, wherein the data of the physiological signal includes a determination by the sensing device as to whether the physiological signal indicates a tachyarrhythmia, wherein the machine learning algorithm is configured to automatically update itself based on the determination made by the sensing device.

[0146] Example 24. A method according to any one of Examples 19 to 23, wherein the at least one sensing device comprises at least one of: a glucose monitor, wherein the physiological signal comprises the glucose concentration in the patient; an electroencephalogram (EEG) sensor, wherein the physiological signal comprises brain activity; a pressure monitoring device, wherein the physiological signal comprises a cardiovascular pressure signal; a pulse oximetry device, wherein the physiological signal comprises at least one of pulse rate, oxygen saturation, or respiratory rate; a patient activity tracker, wherein the physiological signal comprises physical activity of the patient; or a position tracker configured to determine the position of the patient.

[0147] Example 25. A method according to any one of Examples 19 to 24, wherein the device comprises a wearable automatic external defibrillator (WAED), the WAED comprising a garment configured to be worn by the patient, wherein the sensing electrodes and the defibrillation electrodes are coupled to the garment.

[0148] Example 26. A method according to any one of Examples 19 to 25, wherein the at least one probabilistically determined patient state includes at least one of whether the patient state is normal or whether the patient state is a treatable tachyarrhythmia.

[0149] Example 27. A method according to any one of Examples 19 to 26, wherein the probabilistically determined abnormal patient state further includes a sub-classification of one of the following: bradycardia, treatable tachyarrhythmia, syncope, 60 Hz noise, motion artifact, or signal loss.

[0150] Example 28. A method according to any one of Examples 19 to 27, wherein at least one of the probabilistically determined patient states includes a predicted tachyarrhythmia patient state, the therapy delivery circuit system is configured to deliver therapy configured to prevent tachyarrhythmias, and the processing circuit system is further configured to determine whether to deliver the therapy configured to prevent tachyarrhythmias based on the predicted tachyarrhythmia state.

[0151] Example 29. A method according to any one of Examples 19 to 28, wherein the at least one probabilistically determined patient state includes at least one comorbid state of the patient.

[0152] Example 30. A method according to any one of Examples 19 to 29, wherein the at least one comorbid condition of the patient includes the patient's COPD status, and the at least one sensing device includes a spirometer configured to generate a spirometer signal; and an air quality sensor configured to generate an air quality signal, wherein the processing circuit system is further configured to determine the COPD status of the patient based on applying the machine learning algorithm to the spirometer signal and the air quality signal.

[0153] Example 31. A method according to any one of Examples 19 to 30, wherein based on the at least one comorbid condition of the patient, the processing circuit system is configured to provide instructions to the patient.

[0154] Example 32. A method according to any one of Examples 19 to 31, wherein the input from the cardiac signal to the machine learning algorithm includes one or more of the RR interval, the amplitude of the R wave, the QRS width, or the RR interval variability.

[0155] Example 33. A method according to any one of Examples 19 to 32, wherein the input to the machine learning algorithm from the data of the cardiac signal and at least one physiological signal includes one or more of: the magnitude of a feature of the signal, the frequency of the feature, the morphology of the feature, or changes in the magnitude, frequency, or morphology over time.

[0156] Example 34. A method according to any one of Examples 19 to 33, wherein the therapy delivery circuit system is configured to deliver therapy configured to prevent tachyarrhythmias, and the processing circuit system is further configured to determine whether to deliver the therapy configured to prevent tachyarrhythmias based on applying the machine learning algorithm to the cardiac signal and the physiological signal.

[0157] Example 35. A method according to any one of Examples 19 to 34, further comprising controlling, by the processing circuit system, the delivery of the defibrillation therapy based on the results of applying the machine learning algorithm to the data and the cardiac signal.

[0158] Example 36. A method according to any one of Examples 19 to 35, wherein the processing circuit system of the device includes a graphics processing unit (GPU), and the GPU is configured to apply the machine learning algorithm to the data and the cardiac signal to probabilistically determine at least one of the conditions of the patient.

[0159] Example 37. A system for cardiac defibrillation, the system comprising a device configured to deliver defibrillation therapy, wherein the device is configured to be worn by a patient, wherein the device comprises: a processing circuit system comprising a first graphics processing unit (GPU); sensing electrodes configured to sense cardiac signals of the patient; defibrillation electrodes; a therapy delivery circuit system configured to deliver defibrillation therapy to the patient via the defibrillation electrodes; and a memory configured to store the cardiac signals and a machine learning algorithm; and a computing system communicatively coupled to the device, the computing system comprising a second GPU,

[0160] wherein the first GPU is configured to apply the machine learning algorithm to the cardiac signal, and the processing circuit system is configured to determine whether to control the delivery of the defibrillation therapy based on a result of applying the machine learning algorithm to the cardiac signal, and wherein the second GPU is configured to update the machine learning algorithm based on the cardiac signal and population data, wherein the population data includes data of cardiac signals from multiple other patients.

[0161] Example 38. The system of example claim 37, wherein the plurality of other patients comprises a subset of the other patients having one or more characteristics that match corresponding characteristics of the patient.

[0162] Example 39. A system according to any of Examples 37 to 38, further comprising an implantable cardioverter-defibrillator (ICD), the ICD configured to store and execute a tachyarrhythmia detection algorithm, wherein the computing system is configured to configure the tachyarrhythmia detection algorithm for the patient based on cardiac signals classified by the machine learning algorithm.

[0163] Example 40. The system of any of Examples 37 to 39, wherein the second GPU comprises at least one of a greater number of parallel cores or a greater clock speed than the first GPU.

[0164] Example 41. A system according to any of Examples 37 to 40, wherein the second GPU executes a more computationally complex version of the machine learning algorithm than the first GPU.

[0165] Example 42. A system according to any of Examples 37 to 41, wherein the processing circuit system of the device is configured to control the therapy delivery circuit system to deliver the defibrillation therapy based on the results of applying the machine learning algorithm to the data and the cardiac signal.

[0166] Example 43. A system for determining a patient's treatable tachyarrhythmia state, the system comprising a device configured to be worn by the patient and a sensing device, the sensing device being separate from the device, the device comprising: sensing electrodes configured to sense cardiac signals of the patient; a communication circuit system configured to receive data of at least one physiological signal of the patient from the sensing device via wireless communication; a memory configured to store the data, the cardiac signals, and a machine learning algorithm; and a processing circuit system configured to apply the machine learning algorithm to the data and the cardiac signals to determine the patient's treatable tachyarrhythmia state, wherein the processing circuit system of the device is configured to request the data as a master device from the sensing device as a slave device according to a master / slave relationship.

[0167] Example 44. A system according to Example 43, wherein the processing circuit system is configured to preliminarily determine the treatable tachyarrhythmia state based on the cardiac signal and request the data from the sensing device based on the preliminarily determination.

[0168] Example 45. A system according to any one of Examples 43 to 44, wherein the sensing device includes a subcutaneous implantable cardiac monitor, the subcutaneous implantable cardiac monitor including a plurality of sensing electrodes for sensing the physiological signal, wherein the cardiac signal includes a first cardiac signal and the physiological signal includes a second cardiac signal.

[0169] Example 46. A system according to any of Examples 43 to 45, wherein the device includes defibrillation electrodes configured to deliver defibrillation therapy to the patient, and wherein the processing circuit system is configured to control whether the defibrillation therapy is delivered to the patient based on the treatable tachyarrhythmia state of the patient.

[0170] Example 47. A system according to any one of Examples 43 to 46, wherein the processing circuit system of the device includes a graphics processing unit (GPU) configured to apply the machine learning algorithm to the data and the cardiac signal to determine the treatable tachyarrhythmia state of the patient.

Claims

1. A device configured to be worn by a patient for cardiac defibrillation, the device comprising: a sensing electrode configured to sense a first cardiac signal of the patient; Defibrillation electrodes; a therapy delivery circuitry configured to deliver defibrillation therapy to the patient via the defibrillation electrodes; a communication circuit system configured to receive data of at least one physiological signal of the patient other than the first cardiac signal from at least one sensing device separate from the apparatus; a memory configured to store the data received by the communication circuitry, the first cardiac signal, and a machine learning algorithm; as well as processing circuitry configured to: applying the machine learning algorithm to the first cardiac signal to probabilistically determine a preliminary characterization of at least one condition of the patient; determining, based on a machine learning algorithm applied to the first cardiac signal, that the preliminary characterization of at least one condition of the patient is abnormal; Based on the preliminary characterization being abnormal, applying the machine learning algorithm to both the data received by the communication circuitry and the first cardiac signal to probabilistically determine the at least one condition of the patient; and A determination is made whether to control delivery of the defibrillation therapy based on the at least one probabilistically determined state of the patient.

2. The device of claim 1 , wherein the machine learning algorithm is configured to automatically update itself based on at least one of the data received by the communication circuit system, the first cardiac signal, and input from the patient or a healthcare provider.

3. An apparatus according to claim 1 or 2, wherein the data of the physiological signal includes a determination by the sensing device as to whether the physiological signal indicates a rapid arrhythmia, wherein the machine learning algorithm is configured to automatically update itself based on the determination made by the sensing device.

4. The apparatus according to claim 1 or 2, wherein the at least one sensing device comprises at least one of the following: a subcutaneously implantable cardiac monitor comprising a plurality of sensing electrodes for sensing the physiological signal; a glucose monitor, wherein the physiological signal comprises a glucose concentration in the patient; an electroencephalogram (EEG) sensor, wherein the physiological signal comprises brain activity; A pressure monitoring device, wherein the physiological signal comprises a cardiovascular pressure signal; a pulse oximetry device, wherein the physiological signal comprises at least one of pulse rate, oxygen saturation, or respiratory rate; a patient activity tracker, wherein the physiological signals include physical activity of the patient; or A position tracker is configured to determine a position of the patient.

5. The apparatus of claim 4, wherein the patient activity tracker comprises a wearable sensing device configured to be worn on the patient.

6. The device of claim 1 or 2, wherein the device comprises a wearable automatic external defibrillator comprising a garment configured to be worn by the patient, wherein the sensing electrodes and the defibrillation electrodes are coupled to the garment.

7. The apparatus of claim 1 or 2, wherein the at least one probabilistically determined patient state comprises at least one of: whether the patient's condition is normal; or Whether the patient state is abnormal, wherein the probabilistically determined abnormal patient state includes one or more of the subcategories of one of the following: bradycardia, treatable tachyarrhythmia, syncope, 60 Hz noise, motion artifact, signal loss, or at least one comorbid condition of the patient.

8. An apparatus according to claim 1 or 2, wherein the at least one probabilistically determined patient state includes a predicted treatable tachyarrhythmia patient state, the therapy delivery circuit system is configured to deliver therapy configured to prevent tachyarrhythmia, and the processing circuit system is further configured to determine whether to deliver the therapy configured to prevent tachyarrhythmia based on the predicted treatable tachyarrhythmia state.

9. The apparatus according to claim 1 or 2, wherein the at least one comorbid condition of the patient comprises the COPD status of the patient, and the at least one sensing device comprises: a spirometer configured to generate a spirometer signal; as well as an air quality sensor configured to generate an air quality signal, Wherein the processing circuitry is further configured to determine the COPD status of the patient based on applying the machine learning algorithm to the spirometer signal and the air quality signal.

10. The apparatus of claim 7, wherein based on the at least one comorbid condition of the patient, the processing circuitry is configured to provide instructions to the patient.

11. The apparatus of claim 1 or 2, wherein the input from the cardiac signal to the machine learning algorithm comprises one or more of the RR interval, the amplitude of the R wave, the QRS width, or the RR interval variability.

12. A device according to claim 1 or 2, wherein the input to the machine learning algorithm from the first cardiac signal and the data of at least one physiological signal received by the communication circuit system includes one or more of the following: the magnitude of a feature of the signal, the frequency of the feature, the morphology of the feature, or changes in the magnitude, frequency, or morphology over time.

13. The apparatus of claim 1 or 2, wherein the processing circuitry is configured to control the therapy delivery circuitry to deliver the defibrillation therapy based on a result of applying the machine learning algorithm to the data and the first cardiac signal.

14. The apparatus according to claim 1 or 2, wherein the processing circuit system comprises a graphics processing unit (GPU), and the GPU is configured to apply the machine learning algorithm to the data and the first cardiac signal to probabilistically determine the at least one state of the patient.

15. The apparatus of claim 1, wherein the data received by the communication circuit system comprises a second cardiac signal sensed by a sensing device (230B) separate from the apparatus.

Images