Convolutional neural network for automatically discriminating pause attacks detected by implantable medical devices

Abstract

Systems and methods for declaring cardiac activity pauses include a memory for storing specific executable instructions and a convolutional neural network (CNN) model including a global average pooling (GAP) layer. The one or more processors are configured to execute particular executable instructions to obtain a device-classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) for a corresponding candidate pause episode announced by the IMD. The DCA data set includes one or more pulsated cardiac activity (CA) signals sensed by the IMD. The processor (s) applies the CNN model to the DCA data set to identify a valid subset of the DCA data set that correctly characterizes the corresponding CA signal. The display is configured to present information about a valid subset of the DCA data set.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 515,359, filed on July 25, 2023, entitled “CONVOLUTIONAL NEURAL NETWORK FOR AUTOMATIC DISCRIMINATION OF PAUSE EPISODES DETECTED BY AN IMPLANTABLE MEDICALDEVICE,” the subject matter of which is incorporated herein by reference in its entirety. Technical Field

[0003] Embodiments herein generally relate to utilizing convolutional neural networks to identify and / or treat episodes of pauses in cardiac activity signals. Background Art

[0004] An implantable medical device (IMD) can be implanted in a patient to monitor heart rhythm and, in some cases, deliver therapy to treat the patient (such as with pacing pulses). An implantable cardiac monitor (ICM), a type of IMD, is a small single-lead device implanted subcutaneously in a patient to monitor heart rhythm over time. The ICM automatically records arrhythmic events (such as atrial fibrillation (AF) and pauses), as well as captures symptomatic episodes activated by the patient. The recorded data is sent to a remote monitoring platform, which notifies the clinician for further analysis and intervention.

[0005] Today, many arrhythmia detection processes are implemented within IMDs (such as, but not limited to, ICMs) that detect arrhythmias based on various criteria, such as irregularities and patterns of variation in R-wave to R-wave (RR) intervals. In some embodiments, the arrhythmia detection process is performed on a beat-by-beat basis over cardiac activity (CA) signals and analyzes characteristics of interest, such as RR intervals over a period of time. An arrhythmia episode is declared based on characteristics of interest, such as when the pattern of RR intervals of a suspected beat segment is sufficiently irregular and dissimilar to the pattern of RR intervals of a sinus beat segment. When an arrhythmia episode is detected by the ICM, the ICM stores the CA signals (e.g., electrocardiogram or EGM signals) associated with the episode as an arrhythmia episode (AE) dataset, and includes, along with the AE dataset, one or more device record (DD) tags that specify aspects of interest within the CA signals and / or the episode.

[0006] However, the arrhythmia detection process may sometimes declare a false arrhythmia episode when the patient is not experiencing an arrhythmia. When a false arrhythmia episode is declared, the ICM continues to store the CA signals associated with the episode with DD markers (albeit incorrect / false DD markers) as the AE dataset. For example, a false pause detection may occur due to various conditions such as insufficient sensing of the R wave or momentary loss of electrode tissue contact.

[0007] For certain implantable devices and conditions, large AE datasets may be stored and sent due to frequent false detections. This is particularly challenging for ICMs, where computational and power constraints limit internal algorithms, which may lead to misclassification errors. False arrhythmia detections may lead to additional review burden and resource utilization in the clinic, which often must spend considerable time reviewing the AE datasets.

[0008] Additionally, for implantable devices that provide therapy, declaring a false arrhythmia episode when the patient is not experiencing an arrhythmia may result in overtreatment, which may negatively impact the patient. In some cases, the increase in therapy may use more battery power than necessary, thereby reducing the usable life of the IMD.

[0009] There remains a need to reduce the burden on clinicians to review EGM signals and DD markers, particularly EGM signals associated with false arrhythmic episodes such as pause events, and to improve patient treatment. Summary of the invention

[0010] According to an embodiment herein, a system for announcing a pause in cardiac activity includes a memory for storing specific executable instructions and a convolutional neural network (CNN) model trained to detect pause episodes. The CNN model includes a global average pooling (GAP) layer. The system includes one or more processors configured to execute specific executable instructions to obtain a device classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) for a corresponding candidate pause episode announced by the IMD, the DCA dataset including one or more beats of cardiac activity (CA) signals sensed by the IMD, and applying the CNN model to the DCA dataset to identify a valid subset of the DCA dataset that correctly characterizes the corresponding CA signal. The system includes a display configured to present information about the valid subset of the DCA dataset.

[0011] Optionally, the one or more processors are further configured to apply the CNN model to the DCA dataset to identify an invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signal, wherein the valid subset of the DCA dataset indicates a portion of the DCA dataset indicating a true pause, and the invalid subset of the DCA dataset indicates a second portion of the DCA dataset indicating a false pause.

[0012] Optionally, the information indicates a recommendation for: i) a change in treatment of a patient associated with the IMD that would increase the DCA dataset identified as a valid subset of the DCA dataset, or ii) a change in treatment associated with the IMD that would increase the DCA dataset identified as a valid subset of the DCA dataset.

[0013] Optionally, the one or more processors are further configured to apply the CNN model to the DCA dataset to identify an invalid subset of the DCA dataset that incorrectly characterizes corresponding CA signals, wherein, in response to a change in treatment for the patient or the IMD, the one or more processors are further configured to obtain a second DCA dataset generated by the IMD, apply the CNN model to the second DCA dataset to identify a second invalid subset of the DCA dataset that incorrectly characterizes corresponding CA signals, and confirm that the second invalid subset of the DCA dataset has fewer invalid candidate pause episodes than the invalid subset of the DCA dataset.

[0014] Optionally, the CNN model includes at least two one-dimensional (1D) convolutional layers, and the GAP layer is configured to receive outputs from each of the at least two 1D convolutional layers, wherein the one or more processors are further configured to process the output of the first 1D convolutional layer using i) a rectified linear unit activation function, ii) a batch normalization function, or iii) a dropout function.

[0015] Optionally, the CNN model includes at least a first 1D convolutional layer and a second 1D convolutional layer, wherein the one or more processors are further configured to normalize the output of the first 1D convolutional layer and send the normalized output of the first 1D convolutional layer to the second 1D convolutional layer.

[0016] Optionally, the CNN model further includes a fully connected layer configured to receive input from the GAP layer.

[0017] Optionally, the CNN model outputs a confidence indicator in conjunction with each of the plurality of DCA datasets, the confidence indicator indicating a degree of confidence that the corresponding DCA dataset represents a true positive or false positive assignment of a pause.

[0018] Optionally, the one or more processors are further configured to compare the confidence indicator to a detection threshold, wherein the information further comprises information about a valid subset of the DCA dataset that exceeds the detection threshold. Optionally, in response to adjusting the detection threshold, the one or more processors are further configured to present information about a valid subset of the DCA dataset that exceeds the adjusted detection threshold.

[0019] Optionally, the CNN model includes five one-dimensional (1D) convolutional layers, and the GAP layer is configured to receive outputs from each of at least two of the 1D convolutional layers.

[0020] Optionally, the CNN model represents a model trained using an enhanced set of DCA datasets, wherein the enhanced set of DCA datasets includes a reference DCA dataset from a patient and a synthetic DCA dataset generated based on the reference DCA dataset.

[0021] Optionally, a synthetic DCA dataset is generated using: i) noise addition, ii) signal inversion, iii) amplification, iv) insertion of one or more p-waves at the previous RR interval during the pause interval, v) stretching or contracting the duration of the pause interval, or vi) two or more of noise addition, signal inversion, amplification, insertion of one or more p-waves at the previous RR interval during the pause interval, or stretching or contracting the duration of the pause interval.

[0022] Optionally, the information indicates a recommendation for adjusting a sensing parameter or a therapy parameter associated with the IMD.

[0023] Optionally, the system further comprises an IMD comprising a combination of subcutaneous electrodes configured to collect CA signals, an IMD memory configured to store program instructions, and one or more IMD processors configured to execute the program instructions to analyze the CA signals and declare a candidate pause episode based on the analysis, and generate a DCA data set comprising the corresponding CA signals. The system further comprises a transceiver configured to wirelessly transmit the DCA data set to an external device.

[0024] Optionally, the system further comprises an external device comprising a memory and one or more processors and a transceiver configured to wirelessly receive the DCA data set from the IMD.

[0025] Optionally, the system also includes a server comprising a memory and one or more processors, the memory being configured to store a set of DCA datasets, and the one or more processors being configured to apply the CNN model to the set of DCA datasets.

[0026] According to embodiments herein, a computer-implemented method for confirming a device-recorded (DD) pause episode includes obtaining, under control of one or more processors configured with specific executable instructions, a device-classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) for a corresponding candidate pause episode declared by the IMD, the DCA dataset including one or more beats of cardiac activity (CA) signals sensed by the IMD. The method applies a convolutional neural network (CNN) model trained to detect pause episodes to the DCA dataset to identify a valid subset of the DCA dataset that correctly characterizes the corresponding CA signals. The method presents information about the valid subset of the DCA dataset.

[0027] Optionally, the method further comprises applying the CNN model to the DCA dataset to identify an invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signal, wherein the valid subset of the DCA dataset indicates a portion of the DCA dataset indicating a true pause, and the invalid subset of the DCA dataset indicates a second portion of the DCA dataset indicating a false pause.

[0028] Optionally, the information indicates a recommendation for: i) a change in treatment of a patient associated with the IMD that would increase the DCA dataset identified as a valid subset of the DCA dataset, or ii) a change in treatment associated with the IMD that would increase the DCA dataset identified as a valid subset of the DCA dataset.

[0029] Optionally, the method further includes applying the CNN model to the DCA dataset to identify an invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signals. In response to a change in treatment for the patient or the IMD, the method further includes obtaining a second DCA dataset generated by the IMD; applying the CNN model to the second DCA dataset to identify a second valid subset of the DCA dataset that correctly characterizes the corresponding CA signals, and identifying a second invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signals; and confirming that the second invalid subset of the DCA dataset has fewer invalid candidate pause episodes than the invalid subset of the DCA dataset.

[0030] Optionally, the method further comprises adjusting the CNN model by inputting into a layer of the CNN model at least one of: i) age-related characteristics, ii) gender, iii) body mass index, iv) race, v) medical history, vi) medication use, vii) lifestyle factors, or viii) family history.

[0031] Optionally, wherein the CNN model comprises at least two layers, the method further comprises freezing one of the at least two layers, and updating at least one unfrozen layer of the CNN model using the additional DCA dataset.

[0032] According to an embodiment of the present invention, in Example 1, a system (200) for declaring a pause in cardiac activity includes: a memory (288) for storing specific executable instructions and a convolutional neural network (CNN) model trained to detect pause episodes, the CNN model including a global average pooling (GAP) layer; one or more processors (220) configured to execute specific executable instructions to: obtain a device classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) (212) for a corresponding candidate pause episode declared by the IMD (212), the DCA dataset including one or more beats of cardiac activity (CA) signals sensed by the IMD (212); and apply the CNN model to the DCA dataset to identify a valid subset of the DCA dataset that correctly characterizes the corresponding CA signal; and a display configured to present information about the valid subset of the DCA dataset.

[0033] Optionally, in Example 2, in the system of Example 1, the one or more processors (220) are further configured to apply the CNN model to the DCA dataset to identify an invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signal, wherein the valid subset of the DCA dataset indicates a portion of the DCA dataset indicating a true pause, and the invalid subset of the DCA dataset indicates a second portion of the DCA dataset indicating a false pause.

[0034] Optionally, in Example 3, in the system described in Example 1 or 2, the information indicates a recommendation for: i) a change in treatment of a patient (208) associated with an IMD (212) that would increase the DCA dataset identified as a valid subset of the DCA dataset, or ii) a change in treatment associated with the IMD (212) that would increase the DCA dataset identified as a valid subset of the DCA dataset.

[0035] Optionally, in Example 4, in the system of Example 3, the one or more processors (220) are further configured to apply the CNN model to the DCA dataset to identify an invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signal, wherein in response to a change in treatment for the patient (208) or the IMD (212), the one or more processors (220) are further configured to: obtain a second DCA dataset generated by the IMD (212); apply the CNN model to the second DCA dataset to identify a second invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signal; and confirm that the second invalid subset of the DCA dataset has fewer invalid candidate pause episodes than the invalid subset of the DCA dataset.

[0036] Optionally, in Example 5, in the system of any one of Examples 1 to 4, wherein the CNN model includes at least two one-dimensional (1D) convolutional layers, the GAP layer is configured to receive outputs from each of the at least two 1D convolutional layers, wherein the one or more processors (220) are further configured to process the output of the first 1D convolutional layer using i) a rectified linear unit activation function, ii) a batch normalization function, or iii) a dropout function.

[0037] Optionally, in Example 6, in the system of any one of Examples 1 to 4, wherein the CNN model includes at least a first one-dimensional (1D) and a second one-dimensional (1D) convolutional layer, wherein the one or more processors (220) are further configured to: normalize the output of the first 1D convolutional layer; and send the normalized output of the first 1D convolutional layer to the second 1D convolutional layer.

[0038] Optionally, in Example 7, in the system of any one of Examples 1 to 6, the CNN model further includes a fully connected layer configured to receive input from the GAP layer.

[0039] Optionally, in Example 8, in the system of any of Examples 1 to 7, wherein the CNN model outputs a confidence indicator in conjunction with each of the plurality of DCA datasets that indicates a degree of confidence that the corresponding DCA dataset represents a true positive or false positive designation of a pause.

[0040] Optionally, in Example 9, in the system of Example 8, wherein the one or more processors (220) are further configured to compare the confidence indicator to a detection threshold, wherein the information further includes information about a valid subset of the DCA data set that exceeds the detection threshold.

[0041] Optionally, in Example 10, in the system of Example 9, wherein, in response to adjusting the detection threshold, the one or more processors (220) are further configured to present information about the valid subset of the DCA data set that exceeds the adjusted detection threshold.

[0042] Optionally, in Example 11, in the system of any one of Examples 1 to 10, wherein the CNN model includes five one-dimensional (1D) convolutional layers, the GAP layer is configured to receive outputs from each of at least two 1D convolutional layers.

[0043] Optionally, in Example 12, in the system of any one of Examples 1 to 11, wherein the CNN model represents a model trained using an enhanced set of DCA datasets, wherein the enhanced set of DCA datasets includes a reference DCA dataset from a patient and a synthetic DCA dataset generated based on the reference DCA dataset, wherein the synthetic DCA dataset is generated using: i) noise addition, ii) signal inversion, iii) amplification, iv) inserting one or more p-waves at a previous RR interval during the pause interval, v) stretching or contracting the duration of the pause interval, or vi) two or more of noise addition, signal inversion, amplification, inserting one or more p-waves at a previous RR interval during the pause interval, or stretching or contracting the duration of the pause interval.

[0044] Optionally in Example 13, in the system of any one of Examples 1 to 12, wherein the information indicates a recommendation for adjusting a sensing parameter or a therapy parameter associated with the IMD (212).

[0045] Optionally, in Example 14, in the system described in any one of Examples 1 to 13, further comprising an IMD (212), the IMD (212) comprising: a combination of subcutaneous electrodes configured to collect CA signals (114, 126); an IMD memory (288), the IMD memory configured to store program instructions; and one or more IMD processors (220), the one or more IMD processors configured to execute program instructions to: analyze the CA signals and declare candidate pause episodes based on the analysis; and generate a DCA data set including corresponding CA signals; and a transceiver (264), the transceiver configured to wirelessly send the DCA data set to an external device (214).

[0046] According to the embodiments herein, in Example 15, a computer-implemented method for confirming a device-recorded (DD) pause episode includes: under the control of one or more processors configured with specific executable instructions, obtaining a device-classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) for a corresponding candidate pause episode declared by the IMD, the DCA dataset including one or more beats of cardiac activity (CA) signals sensed by the IMD; applying a convolutional neural network (CNN) model trained to detect pause episodes to the DCA dataset to identify a valid subset of the DCA dataset that correctly characterizes the corresponding CA signals; and presenting information about the valid subset of the DCA dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1A An implantable cardiac monitor (ICM) intended for subcutaneous implantation at a site near the heart according to embodiments herein is shown.

[0048] Figure 1B A block diagram of an ICM formed in accordance with embodiments herein is shown.

[0049] Figure 2A A simplified block diagram of a system including an IMD according to embodiments herein is shown.

[0050] Figure 2B The embodiment according to this invention is shown Figure 2A Block diagram of the system.

[0051] Figure 3 A high-level overview of a system formed in accordance with embodiments herein is shown.

[0052] Figure 4A An overview of an example convolutional neural network (CNN) model utilized in accordance with embodiments herein is shown.

[0053] Figure 4B Three different electrogram (EGM) samples from different ICM devices including pause episodes according to embodiments herein are shown.

[0054] Figure 4C Another subcutaneous EGM (SEGM) sample including a pause episode according to embodiments herein is shown.

[0055] Figure 4D An example of a fully connected layer and a global average pooling layer according to an embodiment of this article is shown.

[0056] Figure 4E A table showing exemplary hyperparameters of interest according to embodiments herein.

[0057] Figure 4F Example equations that may be used to calculate multiple metrics according to embodiments herein are shown.

[0058] Figure 5 A process for training / building a CNN model to analyze a DCA dataset with respect to arrhythmias of interest (such as pauses) according to embodiments herein is shown.

[0059] Fig. 6A A process for distinguishing between valid and invalid device classification pause episodes according to embodiments herein is shown.

[0060] Figure 6B A process for determining and / or implementing treatment for a patient using a CNN model according to an embodiment of the present invention is shown.

[0061] Figure 7 A distributed processing system according to embodiments herein is shown.

[0062] Figure 8 A system-level diagram indicating potential devices and networks utilizing the methods and systems herein is shown. DETAILED DESCRIPTION

[0063] The terms "cardiac activity signal", "cardiac activity signals", "CA signal" and "CA signals" (collectively referred to as "CA signals") are used interchangeably throughout the text and shall mean analog or digital electrical signals recorded by two or more electrodes located transvenously, subcutaneously or on the skin, wherein the electrical signals are indicative of cardiac electrical activity. Cardiac activity can be normal / healthy or abnormal / arrhythmic. Non-limiting examples of CA signals include ECG signals collected by skin electrodes, and EGM signals collected by transvenous electrodes and / or non-transvenous subcutaneous electrodes (e.g., SEGM). EGM and SEGM are used interchangeably herein.

[0064] The term "sensing vector" shall refer to a path extending between two or more physically real electrodes operating as sensing sites.

[0065] The term "subcutaneous" shall mean below the surface of the skin but not intracardiac and not intravenous.

[0066] The terms "device-classified arrhythmia dataset" and "DCA dataset" may be used interchangeably and shall mean a dataset that includes: i) a CA signal collected in response to an IMD determining that the CA signal indicates an arrhythmia of interest, and ii) one or more device-recorded flags associated with one or more features of interest in the CA signal, which are used in whole or in part by the IMD in conjunction with the determination of the arrhythmia of interest. Additionally or alternatively, a DCA dataset shall include CA signals collected in response to an IMD determining that the CA signal indicates an arrhythmia of interest, but does not include device-recorded flags. In some cases, the arrhythmia of interest is a pause episode or event.

[0067] The terms "device classified normal sinus dataset" and "DCNS dataset" may be used interchangeably and shall mean a dataset comprising: i) a CA signal collected in response to an IMD determining that the CA signal indicates normal sinus rhythm, and ii) one or more device-recorded flags associated with one or more features of interest in the CA signal that are used, in whole or in part, by the IMD in conjunction with the determination of normal sinus rhythm. Additionally or alternatively, a DCNS dataset shall include CA signals collected in response to an IMD determining that the CA signal indicates normal sinus rhythm, but excluding the device-recorded flags.

[0068] The term "device-recorded signature" refers to a signature generated by an IMD to characterize one or more features of interest within a corresponding CA signal. The signature may be declared based on a number of criteria, such as signal processing, feature detection, and arrhythmia detection software and hardware running within and / or on an implantable cardiac monitor and / or implantable medical device.

[0069] The term "marker" shall mean data and / or information identified from a CA signal, which may be presented as a graphical and / or digital marker indicating one or more features within the CA signal and / or indicating one or more episodes exhibited by a cardiac event. The marker may be superimposed on the CA signal, or presented in proximity to the CA signal, and aligned in time with the CA signal. Non-limiting examples of markers may include an R wave marker, a noise marker, an activity marker, an interval marker, a refractory marker, a P wave marker, a T wave marker, a PVC marker, a sinus rhythm marker, an AF marker, a pause marker, and other arrhythmia markers. As further non-limiting examples, basic event markers may include "AF entry" to indicate the start of an AF event, "AF in" to indicate that AF is ongoing, "AF exit" to indicate that AF has terminated, "pause entry" to indicate the start of a pause event, "pause in" to indicate that a pause is ongoing, "pause exit" to indicate that a pause has terminated, "T" to indicate a tachycardia beat, "B" to indicate a bradycardia beat, "A" to indicate an asystole beat, "VS" to indicate a regular sinus beat, "tachy" to indicate a tachycardia episode, "brady" to indicate a bradycardia episode. "Asystole" indicates an asystole episode, and "patient activated" indicates a patient activated episode. Activity markers may indicate activity detected by an activity sensor during a CA signal. Noise markers may indicate entry / start, ongoing, recovery, and exit / stop of noise. Markers may be presented as symbols, dashed lines, numerical values, thickened portions of waveforms, etc. Markers may represent events, intervals, refractory periods, ICM activity, and other algorithm-related activities. For example, an interval marker, such as an RR interval and / or a pause interval (eg, a pause period), may include a numerical value indicating the duration of the interval.An AF marker indicates an atrial fibrillation rhythm.

[0070] The term "synthetic DCA dataset" shall mean a dataset comprising artificially generated or computer generated CA signals, wherein the CA signals and DD signatures are based on an actual DCA dataset collected from a patient, but wherein the CA signals were not collected from an actual patient.

[0071] The terms "beat" and "cardiac event" are used interchangeably and are intended to include both normal and abnormal events.

[0072] The terms "normal" and "sinus" are used to refer to events, features, and characteristics of or appropriate for the healthy or normal functioning of the heart.

[0073] The terms "abnormal," "arrhythmia," or "arrest" are used to refer to events, features, and characteristics of or appropriate for unhealthy or abnormal working of the heart.

[0074] The term "machine learning" shall mean an artificial intelligence algorithm that learns from various automatic or manual inputs, such as features of interest, arrhythmias previously classified by a device, observations, and / or data. Based on features of interest, arrhythmias previously classified by a device, observations, and / or data, the machine learning algorithm is adjusted in multiple iterations. For example, the machine learning algorithm is adjusted by supervised learning, unsupervised learning, and / or reinforcement learning. Non-limiting examples of machine learning algorithms are convolutional neural networks (CNNs), gradient boosted random forests, decision trees, K-means, deep learning, artificial neural networks, and the like. The machine learning model may include a CNN architecture. It should be appreciated that the network architecture may be different and / or other types of machine learning models may be utilized. As a non-limiting example, the architecture may include N network layers, each with M sublayers, followed by pooling and normalization. The architecture components may include: 1-dimensional convolutional layer ("Conv1D"), rectified linear unit ("relu") activation function, batch normalization ("BN"), and the like. The network output may be a continuous value between 0 and 1, where values ​​close to zero indicate a high confidence that the CA signal indicates a false positive of a pause, and values ​​close to 1 indicate a high confidence that the CA signal indicates a true positive of a pause.

[0075] The term "real time" refers to a time frame that occurs substantially simultaneously with the occurrence of a normal or abnormal episode or event of interest. For example, real-time processing or operation will occur during or immediately after (e.g., within minutes or seconds thereafter) an arrhythmia episode (such as a pause), as data becomes available, etc.

[0076] The term "obtain" used in conjunction with data, signals, information, etc. includes at least one of the following: i) accessing a memory of an IMD, ICM, external device, or remote server in which data, signals, information, etc. are stored, ii) receiving data, signals, information, etc. through a wireless communication link between the ICM or IMD and a local external device, iii) receiving data, signals, information, etc. at a remote server through a network connection, and / or iv) sensing or collecting signals (e.g., CA signals, impedance signals, etc.) provided on the ICM or IMD or between electrode combinations coupled to the ICM or IMD. When viewed from the perspective of the ICM or IMD, the obtaining operation may include sensing new signals in real time, and / or accessing a memory to read stored data, signals, information, etc. from a memory within the ICM or IMD. When viewed from the perspective of a local external device, the obtaining operation includes receiving data, signals, information, etc. at a transceiver of the local external device, where the data, signals, information, etc. are sent from the ICM and / or remote server. The obtaining operation may be from the perspective of a remote server, such as when receiving data, signals, information, etc. from a local external device and / or directly from the ICM at a network interface. The remote server may also obtain data, signals, information, etc. from local memory and / or from other memory (such as within a cloud storage environment) and / or from memory of a workstation or clinician's external programmer.

[0077] The terms "recommend" and "recommendation" shall mean a proposed action to be taken by a patient, healthcare provider, and / or device (e.g., IMD, local external device, remote server, programmer, smartphone) to change therapy and / or maintain a current course of action (e.g., maintenance therapy).

[0078] The term "treatment" shall mean one or more of the following: Non-limiting examples of treatment include instructing a medical professional to change a patient's medication or dosage, editing a diagnosis, clearing an alarm, changing one or more parameters within an IMD (e.g., sensing parameters, shock parameters, thresholds associated with detecting true pauses and false pauses, adjustment of a confidence indicator based on the degree of confidence that the corresponding DCA data set represents a true pause or a false pause, and / or pacing parameters). Examples of treatment may also include instructions to change, maintain, add, or stop treatment delivered by an active IMD (such as pacing therapy, ATP pacing therapy, neural stimulation therapy, mechanical circulatory support, etc.). Examples of treatment may also include dispatching an ambulance to the patient's location, instructing the patient to go to the hospital immediately, instructing the patient to schedule an appointment, instructing the patient to change a prescription, instructing the patient to undergo additional examinations (e.g., diagnostic imaging examinations, exploratory surgery, etc.), instructing the patient to change medications, diet, behavior, activities. In some cases, treatment may be implemented automatically, such as automatically adjusting sensitivity levels within a predetermined range, automatically adjusting threshold levels within a predetermined range, etc. Treatment may also determine how urgent / important the detected event is based on whether the patient has been previously diagnosed and is currently being treated, or if this is a new diagnosis and treatment initiated based on additional status input from the provider.

[0079] According to embodiments herein, methods and systems train and utilize a convolutional neural network (CNN) model with automatic classification of pause episodes to determine whether a candidate arrhythmia (e.g., pause) declared and classified by an IMD from an EGM and / or a subcutaneous EGM (SEGM) is a true positive or a false positive. The CNN model provides the highest sensitivity and specificity in order to minimize false episodes while maintaining true episodes.

[0080] The CNN model can be trained using a data set that includes determined true pause episodes and a data set that does not include pause episodes. Additional training data can be generated by enhancing the data set using one or more enhancement techniques (such as inversion, amplification, noise addition, inserting one or more p-waves during the pause interval at the previous RR interval, and / or stretching or shrinking the duration of the pause interval). More than one CNN model can be trained to detect pause events, such as by using different training data, different enhanced training data, etc.

[0081] The trained CNN model can evaluate patient data acquired by the implanted device. The trained CNN model determines the probability that an episode recorded by the implanted device is a true pause event or a false pause event. True pause events can be sent to clinicians, thereby reducing the burden of reviewing all possible pause events.

[0082] In addition, the trained CNN model can be used to adjust treatment parameters to improve the patient's treatment. Treatment parameters may include but are not limited to detection thresholds, sensitivity levels, treatment settings, etc.

[0083] The trained CNN model can be personalized for one or more patients and can be further modified through transfer learning to identify other relevant events within the patient data. Clinicians can fine-tune the CNN model based on specific patient information.

[0084] FIG. 1 shows an implantable medical device (IMD) 100 intended for subcutaneous implantation at a site near the heart. In some cases, the IMD 100 may be an ICM implanted subcutaneously at a site near the heart. It should be understood that the embodiments herein are equally applicable to other implantable medical devices, such as leadless monitoring and / or therapeutic devices, cardiac monitoring devices, pacemakers, cardioverters, cardiac rhythm management devices, defibrillators, leadless monitoring devices, leadless pacemakers, and / or leadless implantable medical devices (LIMDs), which are implanted near and / or within the heart and are configured to receive near-field and / or far-field cardiac CA signals. The IMD 100 includes a pair of spaced apart sensing electrodes 114, 126 positioned relative to the housing 102. The sensing electrodes 114, 126 provide detection of far-field electrogram signals. Many configurations of electrode arrangements are possible. For example, the electrode 114 may be located on the distal end of the IMD 100, while the electrode 126 is located on the proximal side of the IMD 100. Additionally or alternatively, the electrode 126 may be located on an opposite side, an opposite end, or elsewhere on the IMD 100. The distal electrode 114 may be formed as part of the housing 102, for example, by coating all but a portion of the housing with a non-conductive material such that the uncoated portion forms the electrode 114. In this case, the electrode 126 may be electrically isolated from the housing 102 electrode by placing it on a component separate from the housing 102, such as the head 120. Alternatively, the head 120 may be formed as an integral part of the housing 102. The head 120 includes an antenna 128 and the electrode 126. The antenna 128 is configured to wirelessly communicate with the external device 154 according to one or more predetermined wireless protocols (e.g., Bluetooth, Bluetooth low energy, Wi-Fi, etc.). The housing 102 includes various other components, such as: sensing electronics for receiving signals from the electrodes, a microprocessor for processing the signals according to an algorithm (such as the arrhythmia and pause algorithm(s) described herein), a circulating memory for temporary storage of CA data, a device memory for long-term storage of CA data upon certain triggering events (such as pause detection), sensors for detecting patient activity, and a battery for powering the components. Although FIG. 1 relates to an ICM, the embodiments disclosed herein are applicable to other implantable medical devices (IMDs), and the terms IMD and ICM are used interchangeably herein.

[0085] In at least some embodiments, IMD 100 is configured to be placed subcutaneously using a minimally invasive approach. Subcutaneous electrodes are provided on housing 102 to simplify the implantation procedure and eliminate the need for a transvenous lead system. Sensing electrodes may be located on opposite sides of the device and are designed to provide robust seizure detection through consistent contact at the sensor-tissue interface. In conjunction with recording EGM / SEGM signals, IMD 100 may be configured to be activated by the patient or automatically.

[0086] IMD 100 senses far-field subcutaneous CA signals, processes the CA signals to detect arrhythmias (such as pause events), and if a pause is detected, automatically records the CA signals in a memory for subsequent transmission to an external device. CA signal processing and pause detection are provided at least in part by algorithms embodied in or implemented by a microprocessor. IMD 100 includes one or more processors and a memory storing program instructions that direct the processor to implement pause detection and analyze cardiac activity signals collected through one or more sensing channels.

[0087] Figure 1B A block diagram of an IMD 100 formed according to embodiments herein is shown. The IMD 100 can be implemented to monitor only ventricular activity, or to monitor both ventricular activity and atrial activity through sensing circuitry. The IMD 100 has a housing 102 that houses electronic / computing components. The housing 102 (which is often referred to as a "can," "housing," "encapsulation," or "housing electrode") can be programmably selected to act as an electrode for certain sensing modes. The housing 102 also includes a connector (not shown) having at least one terminal 113 and an optional additional terminal 115. The terminals 113, 115 can be coupled to sensing electrodes disposed on or proximate to the housing 102. Optionally, more than two terminals 113, 115 can be provided to support more than two sensing electrodes, such as for a bipolar sensing scheme using the housing 102 as a reference electrode. Additionally or alternatively, the terminals 113, 115 can be connected to one or more leads having one or more electrodes disposed thereon, wherein the electrodes are located at various locations around the heart. The type and location of each electrode can vary.

[0088] IMD 100 includes a programmable microcontroller 121 that controls various operations of IMD 100, including cardiac monitoring. Microcontroller 121 includes a microprocessor (or equivalent control circuitry), RAM and / or ROM memory, logic and timing circuits, state machine circuits, and I / O circuits. Microcontroller 121 also performs the operations described herein in connection with collecting cardiac activity data and analyzing cardiac activity data.

[0089] A switch 127 is optionally provided to allow different electrode configurations to be selected under the control of the microcontroller 121. The electrode configuration switch 127 may include a plurality of switches for connecting the desired electrodes to the appropriate I / O circuitry, thereby facilitating electrode programmability. The switch 127 is controlled by a control signal from the microcontroller 121. Alternatively, the switch 127 may be omitted and the I / O circuitry connected directly to the housing electrode 114 and the second electrode 126 ( Figure 1A ).

[0090] The microcontroller 121 includes an arrhythmia detector 134 configured to analyze the cardiac activity signal to identify potential arrhythmic episodes (e.g., tachycardia, bradycardia, asystole, bradypause, pause, atrial fibrillation, etc.). As an example, the arrhythmia detector 134 can implement (one or more) arrhythmia detection algorithms, as described in U.S. Patent 8,135,456, entitled “Device and method for detecting atrial fibrillation”, issued on March 13, 2012, as described in U.S. Patent 11,559,241, entitled “Methods and systems for reducing false declarations of arrhythmias”, issued on January 24, 2023, and / or as described in U.S. Patent 11,633,141, entitled “Method and system to detect r-waves in cardiacarrhythmic patterns”, issued on April 25, 2023, the entire subject matter of which is incorporated herein by reference in its entirety. Although not shown, microcontroller 121 may also include other specialized circuits and / or firmware / software components that assist in monitoring various conditions of the patient's heart and managing pacing therapy.

[0091] The arrhythmia detector 134 of the microcontroller 121 includes an onboard RR interval irregularity (ORI) process 136 that uses RR interval irregularity to detect arrhythmia episodes (such as AF episodes). The ORI process 136 can be implemented as firmware, software, and / or circuitry. The ORI process 136 uses hidden Markov chains and Euclidean distance calculations of similarity to evaluate the transition behavior of one R wave (RR) interval to another R wave (RR) interval, and compares the patient's RR interval transitions to known RR interval transitions during atrial fibrillation (AF) and non-AF episodes obtained from the same patient and / or many patients.

[0092] The microcontroller 121 may include other arrhythmia determination circuitry / software / firmware (not shown) configured to analyze the CA signal to identify the presence of an arrhythmia, and more specifically a pause. Additional circuitry / software / firmware may also analyze the patient's motion when determining a pause.

[0093] The arrhythmia detector 134 analyzes the sensed far-field CA signals sensed along the sensing vector between the subcutaneous electrode combination for one or more beats. The arrhythmia detector 134 identifies one or more features of interest from the CA signals and determines whether the CA signals indicate normal sinus rhythm or an arrhythmia episode based on further analysis of the features of interest. When an arrhythmia episode is identified, the arrhythmia detector 134 generates one or more DD markers that are temporally aligned with the corresponding features of interest in the CA signals. The arrhythmia detector 134 forms a DCA dataset associated with the classified arrhythmia episode (e.g., a pause episode) and stores the DCA dataset in the memory 160 of the IMD 100. The arrhythmia detector 134 iteratively or periodically repeats the analysis of the incoming far-field CA signals to continuously add DCA datasets of corresponding arrhythmia episodes to form a set of DCA datasets.

[0094] IMD 100 is also equipped with a communication modem (modulator / demodulator) 140 to enable wireless communication. In one embodiment, the communication modem 140 uses high-frequency modulation, such as using RF, Bluetooth, or Bluetooth low energy telemetry protocols. The signal is sent in a high frequency range and will pass through body tissues as fluid without stimulating the heart or being felt by the patient. The communication modem 140 can be implemented as a part of the microcontroller 121 in hardware, or as software / firmware instructions programmed into the microcontroller 121 and executed by the microcontroller 121. Alternatively, the modem 140 can reside separately from the microcontroller as an independent component. The modem 140 helps to retrieve data from the remote monitoring network. The modem 140 enables timely and accurate data to be transmitted directly from the patient to an electronic device used by a physician or patient.

[0095] IMD 100 includes sensing circuit 144, which is selectively coupled to one or more electrodes performing sensing operations through switch 127 to detect cardiac activity data indicative of cardiac activity. Sensing circuit 144 may include a dedicated sense amplifier, a multiplexed amplifier, or a shared amplifier. It may also employ one or more low power precision amplifiers with programmable gain and / or automatic gain control, bandpass filtering, and threshold detection circuits to selectively sense features of interest. In one embodiment, switch 127 may be used to determine the sensing polarity of the cardiac signal by selectively closing the appropriate switch.

[0096] The output of the sensing circuit 144 is connected to the microcontroller 121, which in turn determines when to store the cardiac activity data of the CA signal (digitized by the A / D data acquisition system 150) in the memory 160. For example, when a potential arrhythmia episode is detected, the microcontroller 121 may only store the cardiac activity data (from the A / D data acquisition system 150) in the memory 160. The sensing circuit 144 receives control signals 146 from the microcontroller 121 for controlling the timing of the gain, threshold, polarization charge removal circuit (not shown), and any blocking circuits (not shown) coupled to the inputs of the sensing circuit.

[0097] Optionally, IMD 100 may include multiple sensing circuits similar to sensing circuit 144, where each sensing circuit is coupled to two or more electrodes and is controlled by microcontroller 121 to sense electrical activity detected at the corresponding two or more electrodes. Sensing circuit 144 may operate in a unipolar sensing configuration or in a bipolar sensing configuration. Alternatively, sensing circuit 144 may be removed entirely, and microcontroller 121 performs the operations described herein based on CA signals from A / D data acquisition system 150 coupled directly to the electrodes.

[0098] The IMD 100 also includes an analog-to-digital A / D data acquisition system (DAS) 150 coupled to one or more electrodes via switch 127 to sample cardiac activity signals on any desired electrode pair. The data acquisition system 150 is configured to acquire electrocardiogram (EGM) signals as CA signals, convert the raw analog data to digital data, and store the digital data as CA data for later processing and / or telemetry transmission to an external device 154 (e.g., a programmer, a local transceiver, or a diagnostic system analyzer). The data acquisition system 150 is controlled by a control signal 156 from the microcontroller 121. The EGM signals can be used as cardiac activity data for analysis for potential arrhythmic episodes (such as pauses). The ORI process 136 can be applied to signals from the sensing circuit 144 and / or the DAS 150.

[0099] As an example, the external device 154 may represent a bedside monitor that is installed in the patient's home and is used to communicate with the IMD 100 when the patient is at home, in bed, or asleep. The external device 154 may be a programmer used in the clinic to interrogate the IMD 100, retrieve data, and program detection criteria and other features. The external device 154 may be a handheld device (e.g., a smart phone, a tablet device, a laptop computer, a smart watch, etc.) that may be coupled to a remote monitoring service, a medical network, etc. via a network (e.g., the Internet). The external device 154 facilitates physician access to patient data and allows the physician to view real-time CA signals as they are collected by the IMD 100. In some cases, the external device 154 may facilitate patient access, such as viewing alarms, summaries, etc. associated with the patient data.

[0100] Microcontroller 121 is coupled to memory 160 via a suitable data / address bus 162. Programmable operating parameters used by microcontroller 121 are stored in memory 160 and are used to customize the operation of IMD 100 to suit the needs of a particular patient. Such operating parameters define, for example, detection rate thresholds, sensitivity, automatic features, pause detection criteria, AF detection criteria, activity sensing or other physiological sensors, and electrode polarity.

[0101] In addition, memory 160 stores cardiac activity data, as well as markers and other data content associated with detecting an arrhythmia episode. The operating parameters of IMD 100 can be non-invasively programmed into memory 160 via telemetry circuit 164, which is in telemetric communication with external device 154 via communication link 166. Telemetry circuit 164 allows intracardiac electrograms and status information related to the operation of IMD 100 (such as contained in microcontroller 121 or memory 160) to be sent to external device 154 via established communication link 166. According to embodiments herein, telemetry circuit 164 transmits the DCA data set and other information related to the arrhythmia episode to external device 154.

[0102] IMD 100 may also include a magnet detection circuit (not shown) coupled to microcontroller 121 to detect when a magnet is placed on the unit. The clinician may use the magnet to perform various test functions of housing 102 and / or to signal to microcontroller 121 that external device 154 is in place to receive data or send data to microcontroller 121 via telemetry circuit 164.

[0103] IMD 100 may also include one or more physiological sensors 170. Such sensors are often referred to (in the pacemaker field) as "rate response" or "exercise" sensors. Physiological sensors 170 may also be used to detect changes in the physiological condition of the heart or diurnal changes in activity (e.g., detecting sleep and wake states). Signals generated by physiological sensors 170 are passed to microcontroller 121 for analysis and storage in memory 160, optionally in combination with cardiac activity data, markers, episode information, etc. Although shown as included within housing 102, (one or more) physiological sensors 170 may be external to housing 102 but still implanted in the patient or carried by the patient. Examples of physiological sensors may include, for example, sensors that detect motion, activity, temperature, sense respiration rate, pH of blood, ventricular gradients, activity, position / posture, minute ventilation (MV), etc.

[0104] Battery 172 provides operating power to all components in IMD 100. Battery 172 is capable of operating for long periods of time at low current consumption. Battery 172 also ideally has predictable discharge characteristics so that selective replacement time can be detected. As an example, housing 102 employs a lithium / silver vanadium oxide battery. Battery 172 can provide various life cycles (e.g., three years or more of device monitoring). In alternative embodiments, battery 172 can be rechargeable. See, e.g., U.S. Patent No. 7,294,108, Cardiac Event Micro-Recorder and Method of Implantation Thereof, the entire contents of which are incorporated herein by reference.

[0105] IMD 100 provides simple configuration data storage options to enable physicians to prioritize data based on individual patient conditions to capture important events and reduce the risk of missing unexpected events. IMD 100 can be programmable for pre-trigger event storage and post-trigger event storage. For example, IMD 100 can be automatically activated to store 10-120 seconds of CA data before an event of interest and / or store 10-120 seconds of post-CA data. Optionally, IMD 100 can provide patient-triggered activation, in which pre-event CA data and post-event CA data are stored (e.g., 1-15 minutes of pre-event storage and 1-15 minutes of post-event storage). Optionally, IMD 100 can provide manual (patient-triggered) or automatic activation for CA data. Optionally, IMD 100 can provide additional programming options (e.g., duration of asystole, bradycardia rate, tachycardia rate, tachycardia cycle count). The amount of CA data storage can vary based on the size of memory 160.

[0106] The IMD 100 can provide comprehensive and secure diagnostic data reporting, including a heart rate summary, to assist physicians in diagnosing and treating patient conditions. As an example, a report can include an attack diagnosis, attack duration, attack count, attack date / time stamp, and a heart rate histogram for automatically triggered events. The IMD 100 can be configured to be relatively small (e.g., between 2-10cc in volume), which can, among other things, reduce the risk of infection during the implantation process, allow the use of a small incision, allow the use of a smaller subcutaneous pocket, etc. The small footprint can also reduce implantation time and introduce fewer body image changes to the patient.

[0107] Figure 2A A simplified block diagram of a system 200 operating in accordance with embodiments herein is shown. System 200 includes an IMD 212 and one or more external devices or instruments. A single external device 214 is shown. IMD 212 and external device 214 are configured to wirelessly communicate with each other via a wireless communication link 216. In an embodiment, external device 214 receives communications from a data processing system remote from external device 214 and sends communications to the data processing system, the communications including physiological data acquired from patient 208, such as a DCA data set and information identifying IMD 212 and the type of data acquired by IMD 212.

[0108] The IMD 212 is implanted in the patient 208 at a location near the heart 209. The IMD 212 can be a pacemaker, an implantable cardiac monitoring device (ICM), a defibrillator, an ICM coupled to a pacemaker, etc. The IMD 212 is intended to be implanted in a subcutaneous pocket of the patient. The IMD 212 can be configured to sense cardiac signals to monitor changes in cardiac activity over time. In some embodiments, the IMD 212 can also be configured to deliver stimulation therapy to the heart 209. For example, the IMD 212 can be a dual-chamber stimulation device that can treat both fast arrhythmias and slow arrhythmias with stimulation therapy (including cardioversion, defibrillation, and pacing stimulation), as well as being able to detect heart failure, assess its severity, track its progression, and control the delivery of therapy and warnings in response thereto.

[0109] IMD 212 in the illustrated embodiment includes a body or housing 218 connected to at least one lead 219 . Figure 2A209, but in another embodiment, IMD 212 may include multiple leads. Lead 219 extends from housing 218 to heart 209 so that the distal end contacts the patient's tissue surrounding heart 209. Lead 219 includes one or more electrodes that can measure cardiac signals of heart 209 and deliver stimulation therapy to heart 209. In an embodiment, a single electrode can transmit a stimulation pulse in a stimulation mode and then can quickly switch to a monitoring mode to detect cardiac signals after the stimulation pulse.

[0110] Although the IMD 212 in the illustrated embodiment includes leads 219, one or more embodiments described herein utilize a leadless IMD that does not include any leads. For example, the IMD 212 can be a leadless pacemaker, a leadless cardiac monitoring device (ICM), etc. In general, the IMD 212 can represent a cardiac monitoring device, a pacemaker, a cardioverter, a cardiac rhythm management device, a defibrillator, a neurostimulator, a leadless monitoring device, a leadless pacemaker, etc.

[0111] The housing 218 may contain a battery, a pulse generating circuit, a communication circuit, a data storage device (e.g., a memory), and / or a control circuit including one or more processors. The control circuit system is used to receive and analyze EGM signals from the electrodes. The control circuit system may include at least one processor for processing the EGM signals according to an algorithm to make a determination about the state of the heart 209. The memory provides storage for cardiac signals and programming instructions for the control circuit for providing programming sessions. The battery powers the circuits within the housing 218. For example, the battery powers the pulse generating circuit system to generate stimulation pulses and powers the communication circuit system to communicate with the external device 214. The control circuit can generate a message to be transmitted to the external device 214 via the communication circuit. The message may include the EGM signal and / or data generated based on the EGM signal. The control circuit system is also configured to analyze a message received from the external device 214 via the communication circuit, the message including a programming package for updating the operating configuration (including settings, parameters, behaviors, etc.) of the IMD 212. The following is combined with Figure 2B An exemplary structure of IMD 212 is discussed and illustrated.

[0112] The external device 214 can represent a computing device that a patient or clinician can access or own, such as a tablet computer, a smart phone, a wearable device, a laptop computer, a desktop computer, a bedside monitor installed in a patient's home, etc. The external device 214 can alternatively be a programmer device used by a clinician in the clinic. The external device can be one of the devices listed above, which is communicatively connected to a second external device, which represents the source of a programming package intended to update the operational configuration of the IMD212. For example, the second external device can be a server or another computing device that is communicatively connected to the first external device 214 via a network connection (e.g., the Internet). Typically, the external device 214 helps doctors access patient data and allows doctors to view real-time EGM signals and update the operational configuration of the IMD212 without the clinician being close to the patient, such as helping to change the treatment completed by the IMD 212.

[0113] Figure 2B 212. A housing 218 or casing of the IMD 212 houses the electronic and / or computing components. The housing 218 also includes a connector (not shown) having at least one terminal 252 and optional additional terminals 254, 256, 258, 260. The terminals can be connected to electrodes located at various locations in and around the heart (such as on leads 219 (e.g., Figure 2A The electrodes may include various combinations of ring electrodes, tip electrodes, coil electrodes, electric shock electrodes, etc.

[0114] IMD 212 includes a programmable microcontroller 220 that controls various operations of IMD 212, including cardiac monitoring and stimulation therapy. Microcontroller 220 includes one or more processors (e.g., a microprocessor or equivalent control circuitry), RAM and / or ROM memory, logic and timing circuits, state machine circuits, and I / O circuits. Microcontroller 220 includes an arrhythmia detector 234 that is configured to detect cardiac activity data to identify potential atrial fibrillation (AF) episodes and pauses and other arrhythmias (e.g., tachycardia, bradycardia, asystole, etc.).

[0115] An electrode configuration switch 226 is optionally provided to allow different electrode configurations to be selected under the control of the microcontroller 220. The electrode configuration switch 226 may include a plurality of switches for connecting the desired electrodes to the appropriate I / O circuitry, thereby facilitating electrode programmability. The switch 226 is controlled by a control signal 228 from the microcontroller 220. Alternatively, the switch 226 may be omitted and the I / O circuitry connected directly to the housing electrode.

[0116] The IMD 212 may include a chamber pulse generator 222 that generates stimulation pulses for connecting desired electrodes to appropriate I / O circuits, thereby facilitating electrode programmability. The pulse generator 222 is controlled by the microcontroller 220 via a control signal 224. The IMD 212 includes a sensing circuit 244 selectively coupled to one or more electrodes, which performs sensing operations through a switch 226 to detect cardiac activity. The sensing circuit 244 may include a dedicated sensing amplifier, a multiplexed amplifier, or a shared amplifier. The sensing circuit 244 may operate in a unipolar sensing configuration or in a bipolar sensing configuration. The output of the sensing circuit 244 is connected to the microcontroller 220, which in turn triggers or inhibits the pulse generator 222 in response to the absence or presence of cardiac activity. The sensing circuit 244 receives a control signal 246 from the microcontroller 220 for controlling the gain, threshold, polarization, and timing of any blocking circuit (not shown) coupled to the sensing circuit.

[0117] IMD 212 also includes an analog-to-digital A / D data acquisition system (DAS) 284 coupled to one or more electrodes via switch 226 to sample cardiac signals on any desired electrode pair. A / D DAS 284 is controlled by control signal 286 from microcontroller 220.

[0118] IMD 212 is also equipped with communication circuitry or a modem (modulator / demodulator) 240 to enable wireless communication. Modem 240 enables timely and accurate data transmission directly from the patient to external device 214, and vice versa. For example, communication modem 240 is configured to establish a communication link 216 ( Figure 2A In an embodiment, communication modem 240 may receive programming packets from external device 214 conveying therapy settings for updates, changes, variations, etc. In addition to remote programming of IMD 212, wireless communication link 216 with external device 214 facilitates physician and / or patient access to patient data (such as a DCA data set generated by IMD 212).

[0119] The communication modem 240 can utilize radio frequency (RF), Bluetooth, or Bluetooth low energy telemetry protocols. The signal is sent in a high frequency range and will pass through body tissues as fluids without stimulating the heart or being felt by the patient. The communication modem 240 can be implemented in hardware as part of the microcontroller 220, or as software / firmware instructions programmed into and executed by the microcontroller 220. Alternatively, the modem 240 can reside separately from the microcontroller 220 as an independent hardware component.

[0120] The microcontroller 220 is coupled to a non-transitory data storage device, referred to herein as a memory device 288, via a suitable data / address bus 262. The memory device 288 stores programmable operating parameters used by the microcontroller 220 and / or data associated with the detection and determination of arrhythmias. The memory 288 may store a CNN model that has been trained to detect pause episodes. In an embodiment, the memory device 288 also stores the current programming settings, as well as any and all changes, modifications, updates, etc. previously made to the programming settings.

[0121] IMD 212 optionally includes one or more physiological sensors 270 that adjust pacing stimulation rate, detect changes in cardiac output, changes in cardiac physiology, and / or diurnal variations in activity (e.g., detecting sleep and wake states). Examples of physiological sensors 270 may include, for example, sensors that sense respiratory rate, pH of blood, ventricular gradients, activity, body movement, position / posture, minute ventilation (MV), etc.

[0122] Battery 272 provides operating power to all components in IMD 212. Battery 272 is capable of operating for long periods of time at low current consumption, and is capable of providing high current pulses (for capacitor charging) when a patient requires a shock pulse (e.g., in excess of 2A at voltages greater than 2V for a period of 10 seconds or longer).

[0123] IMD 212 also includes impedance measurement circuitry 274, which can be used for many things, including sensing respiratory phase. IMD 212 can also be equipped with telemetry circuitry 264, which can selectively communicate with an external device (such as device 214) when connected via a physical (e.g., wired) communication link or a wireless link. IMD 212 includes shock circuitry 280 controlled by control signals 282 generated by microcontroller 220. Shock circuitry 280 generates low energy (e.g., up to 0.5 Joules), medium energy (e.g., 0.5-10 Joules), or high energy (e.g., 11 to 40 Joules) shock pulses as controlled by microcontroller 220.

[0124] The microcontroller 220 may include other dedicated circuits and / or firmware / software components, such as a timing control (module) 232 and a morphology detector (module) 236. The timing control 232 is used to control various timing parameters, such as stimulation pulses (e.g., pacing rate, atrial-ventricular (AV) delay, interatrial conduction (AA) delay, interventricular conduction (VV) delay, etc.), and timing for tracking RR intervals, refractory periods, blanking intervals, noise detection windows, evoked response windows, alarm intervals, marker channel timing, etc. The morphology detector 236 is configured to view and analyze one or more features of the morphology of the cardiac activity signal (such as the morphology of the detected R wave) to determine whether to include or exclude one or more beats from further analysis.

[0125] Figure 3 A high-level overview of a system formed according to an embodiment of the present invention is shown. The system includes an external artificial intelligence (AI) system on a remote monitoring data server. The AI ​​system may include a CNN model as disclosed herein, but other CNN models are contemplated. Implementing the CNN model in a remote monitoring system can significantly reduce the EGM episode viewing burden and simplify clinical workflow. Another advantage is the ability to provide recommendations for adjusting one or more parameters within the IMD 100, 212 to improve the IMD's detection of true pause events.

[0126] At 300, the CA signal is analyzed by one or more arrhythmia detection algorithms in the ICM or IMD 100, 212. As discussed herein, the arrhythmia detection algorithm can detect a pause. When an arrhythmia is identified, one or more DCA data sets associated with the arrhythmia are recorded, including device recording marks that specify features of interest within the CA signal and / or identify the nature of the arrhythmia. Once the set of DCA data sets are stored in the IMD 100, 212, the set of DCA data sets are wirelessly transmitted from the IMD 100, 212 to a local external device 302 and / or a remote server 304. The data can be transmitted when a predetermined amount of DCA data sets have been collected, or on a periodic basis. The remote server 304 utilizes one or more CNN models 306 to re-analyze the uploaded set of DCA data sets. The CNN model 306 identifies valid subsets and invalid subsets (also referred to as appropriate subsets and inappropriate subsets) of the DCA data sets. The appropriate or valid subset includes DD markers that correctly characterize the corresponding CA signal, while the inappropriate or invalid subset includes DD markers that incorrectly characterize the corresponding CA signal. In other words, the appropriate or valid subset corresponds to correct / positive arrhythmias, such as true pauses, while the inappropriate or invalid subset corresponds to incorrect / negative arrhythmias, such as false pauses. Then, as discussed herein, information about the true positive or valid subset is provided to a clinician portal 308 in various forms, thereby significantly reducing the burden of reviewing all potential EGM pause episodes identified by the IMD 100. In other embodiments, information about both true positives and false positives is provided to the clinician portal 308. The clinician portal 308 can automatically suggest setting adjustments based on the results to improve the true positive detection of pauses by the IMD 100, 212.

[0127] In further embodiments, the CNN model 306 can be housed within a local external device 302 that additionally or alternatively reanalyzes the collection of uploaded DCA data sets. In this manner, information about true positives and false positives can be provided to the patient, an alert can be generated (such as to recommend that the patient contact their physician), adjustments to the settings of the IMD 100, 212 (such as within predetermined limits) can be automatically determined and sent to the IMD 100, 212 to improve true positive detection of pauses, etc.

[0128] Additionally or alternatively, the CNN model 306 can output a confidence indicator (e.g., probability, likelihood, continuous value between 0 and 1) that indicates a level or degree of confidence that the individual underlying DCA dataset indicates a true positive or false positive designation for a pause. As an example, the digital indicator can be a continuous value between 0 and 1, where a value close to zero indicates a high confidence that the subcutaneous EGM signal within the DCA dataset does not indicate a pause event, and is therefore a false positive. A value close to 1 indicates a high confidence that the subcutaneous EGM signal within the DCA dataset indicates a pause event, and is therefore a true positive. When the DCA dataset does not indicate a pause event, the DCA dataset may include a CA signal indicating normal sinus rhythm or other. For example, a CA signal within a DCA dataset may exhibit an excessively noisy signal that should not otherwise be characterized as an arrhythmia or normal sinus rhythm. A threshold value (such as a desired sensitivity level) can be used to determine excessively noisy CA signals (such as those that may result in inappropriate R-wave detection) within a particular DCA dataset.

[0129] Additionally or alternatively, the CNN model 306 may include a detection threshold that can be changed / fine-tuned by the clinician based on the clinician's needs and various factors. In the aforementioned example, where a value close to 1 indicates a high confidence of a true positive and a value close to 0 indicates a high confidence of a false positive, the detection threshold can be lowered (e.g., closer to 0) to increase the sensitivity of the CNN model 306. For example, when the detection threshold is set to 0.25, the CNN model 306 will identify more false positive DCA data sets than when the detection threshold is set to 0.75. For example, it may be desirable to apply a higher level of sensitivity in conjunction with certain types of pause events that may not occur regularly. In addition, it may be desirable to apply higher thresholds, and therefore reduce sensitivity while increasing specificity, for other types of arrhythmias that are considered less "critical" to the patient's health and may occur more frequently.

[0130] The arrhythmia detection algorithm running on the IMD 100, 212 and the CNN model(s) running on the local external device 302 and / or remote server 304 provide two discriminators that work together to form a robust arrhythmia classification system. Figure 3The system reduces false positive arrhythmias by implementing a CNN-based confirmation process to provide a second check on arrhythmias declared by the IMD 100, 212, more specifically, pause events declared by the IMD 100, 212. The processing described herein can be distributed among various devices. For example, one or more of the IMD 100, 212, the local external device 302, and / or the remote server 304 may include a memory for storing specific executable instructions and a convolutional neural network (CNN) model including a global average pooling (GAP) layer; and one or more processors configured to execute the specific executable instructions to: obtain a device classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) (e.g., IMD 100, 212) for a corresponding candidate arrhythmia episode declared by the IMD, the DCA dataset including one or more beats of cardiac activity (CA) signals sensed by the IMD; and apply the CNN model to the DCA dataset to identify a valid subset of the DCA dataset that correctly characterizes the corresponding CA signals, and to identify an invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signals. Further discussed below Figure 3 and Figure 7 The one or more devices shown in may also include a display configured to present information about at least one of the valid subset or the invalid subset of the DCA data set. Additionally or alternatively, Figure 3 One or more devices implement a CNN model that outputs a confidence indicator in conjunction with each DCA dataset, the confidence indicator indicating a confidence level of a true positive or false positive designation that the corresponding DCA dataset represents a pause.

[0131] Figure 1A and Figure 2A One or more of the IMDs include a combination of subcutaneous electrodes configured to collect CA signals; an IMD memory configured to store program instructions; and one or more IMD processors configured to execute the program instructions to: analyze the CA signals and declare candidate pause episodes based on the analysis; generate a DCA data set including corresponding CA signals; and a transceiver configured to wirelessly send the DCA data set to an external device. Figure 1A , Figure 2A , Figure 3 and Figure 7 One or more of the external devices in the system may include a memory and one or more processors and a transceiver, the transceiver being configured to wirelessly receive a DCA dataset from the IMD. The server may include a memory and one or more processors, the memory being configured to store a set of DCA datasets, and the one or more processors being configured to apply a CNN model to the set of DCA datasets.

[0132] Figure 4AAn overview of an example CNN model utilized according to embodiments of the present invention is shown. Although CNN model 400 includes a convolutional neural network architecture, it should be appreciated that the network architecture may be different and / or other types of machine learning models may be utilized.

[0133] A fully connected multi-layer neural network is called a multi-layer perceptron (MLP). MLP is the most commonly used supervised neural network and has been shown to be effective in learning complex systems. The MLP architecture is variable, however, it consists of several layers of neurons connected to each other in a feed-forward manner. Each neuron is a weighted sum of its inputs passed through a non-linear function. The number of layers and neurons are called hyperparameters of the neural network, and these need to be fine-tuned. Cross-validation techniques are used to find the ideal values ​​for these. The training of weight adjustments is done via back-propagation. In some cases, deeper neural networks are better at processing data. However, deeper layers can lead to the vanishing gradient problem, and special algorithms are used to address this problem.

[0134] CNN is a deep neural network (DNN) architecture that is typically trained by a gradient-based optimization algorithm. Figure 4A As shown, the architecture or structure of the CNN model 400 includes five 1-dimensional (1D) convolutional layers 402, a global average pooling (GAP) layer 404, a fully connected layer 406, and a softmax layer 408. It should be understood that other CNN models may include more or fewer convolutional layers. In some embodiments, each of the convolutional layers 402 has parameters (e.g., filter: 64, kernel: 33x1), but other sizes are contemplated.

[0135] Typically, the CNN model 400 includes a plurality of back-to-back layers connected in a feed-forward manner. The main layers include a convolutional layer 402, a normalization layer 403, a pooling layer (e.g., a GAP layer 404), and a fully connected layer 406 (e.g., a dense layer). The convolutional layer 402, the normalization layer 403, and the GAP layer 404 are responsible for extracting features, while the fully connected layer 406 is responsible for classification. The CNN model 400 excels in automatically learning and extracting relevant features from raw data, thereby eliminating the need for manual feature engineering.

[0136] The CNN model 400 is trained and validated using IMD pause EGM episodes in two datasets (e.g., a training dataset and a validation dataset). By way of example only, the training dataset and validation dataset may have a mixture of known true pause events and false pause events (e.g., the pause EGMs of the training and validation datasets have been specifically adjudicated). In some embodiments, the training dataset is ensured to have enough true and false detections to ensure that the dataset is not overwhelmed by false positives while still providing enough examples for the CNN model 400 to learn from. In some cases, a Z-score normalization function is used for each segment.

[0137] For reference, Figure 4B Three different EGM samples including pause episodes from different IMD devices are shown. The vertical axis plots the amplitude in millivolts (mV) and the horizontal axis plots the time in seconds (s). To train the CNN model 400, different approximate lengths of the SEGM before triggering are used (e.g., 4s, 6s, 8s, 10s, more than 10s, a range between 4s and 10s). As an example, Figure 4C Another SEGM sample including a pause episode according to an embodiment of the present invention is shown. Figure 4C The shaded area in the 6 s before the trigger includes the detected or suspected pause episode. The trigger may be a point in time determined by the IMD 100, 212, where the IMD 100, 212 determines a pause episode. The IMD 100, 212 may place the SEGM segment (in Figure 4C The data were saved as DCA datasets.

[0138] In some cases, a hybrid approach is used to complete pause determination, such as the approach in "Piorkowski C, Dawoud F, Qu F, Ryu K. Hybrid Human and Machine Learning Dudication Approach for EvaluatingArrhythmia Discriminator Algorithm Performance. APHRS 2019 Poster AP19-00435". In some embodiments, when a hybrid approach is used, an experienced determiner confirms the labeling. The hybrid approach utilizes machine learning techniques and human verification, is highly accurate, and significantly reduces the burden of human determination. In further embodiments, some pause events are manually determined by experienced determiners (e.g., experienced researchers).

[0139] Return to Figure 4A , the 1D input signal 420 is fed to the first convolutional layer within the five convolutional layers 402 without any feature extraction. In some embodiments, the 1D input signal 420 can be a single-lead ECG or EGM vector or a collection of 1D amplitudes across time. The 1D input signal 420 can be a SEGM of length as described above (e.g., 4s, 6s, 8s, 10s, greater than 10s, a range between 4s and 10s), and during the training phase of the CNN model 400, can be from a reference device classified arrhythmia (DCA) dataset, as discussed further below.

[0140] Convolutional layers C1 410, C2 412, C3 414, C4 416, and C5 418 each extract features at different levels from the input signal 420. Convolutional layers C2 412, C3 414, C4 416, and C5 418 each receive the output of a previous convolutional layer and / or normalization layer, and thus receive a signal based on the input signal 420. Convolutional layers C1 410, C2 412, C3 414, C4 416, and C5 418 perform a convolution operation on the output of the previous layer (or in the case of C1 410, the input signal 420) using the current convolution kernel [6, 13], as shown in the following equation, and their convolution kernel size is 33x1:

[0141] )

[0142] Among them, M k is the effective range of the convolution kernel, represents the output of the kth neuron in layer l, b k is the bias of the kth neuron in layer l, ω ik is the weight kernel between the i-th neuron in layer l-1 and the k-th neuron in layer l, and f(•) denotes the rectified linear unit (ReLU) activation function.

[0143] The output of each of the convolutional layers C1 410, C2 412, C3 414, C4 416, and C5 418 is a feature map, which can be normalized (e.g., batch normalization, batch normalization function), processed by a rectified linear unit activation function (e.g., activation: ReLU), and / or processed by a dropout function (e.g., dropout: 0.5) before the subsequent layers receive data. For example, dropout can be used to randomly remove some of the features to avoid overfitting, while activation can use a threshold to determine which features to send to the next stage.

[0144] As an example, in Figure 4AAfter applying five convolutional layers 402 in the proposed architecture shown, feature maps corresponding to each corresponding convolutional layer 410-418 are generated. To further process these feature maps, they are usually passed through additional layers such as batch normalization, activation, and dropout layers. These layers help enhance the performance and robustness of the network by normalizing the feature maps, introducing nonlinearity, and preventing overfitting. Batch normalization is a technique that normalizes the activations of the previous layer, ensuring that the inputs to the subsequent layers are standardized and less sensitive to changes in the training data. This normalization helps to speed up the training process and improve the overall stability of the network. The activation function introduces nonlinearity into the feature map, enabling the network to learn complex patterns and relationships in the data. Common activation functions include ReLU (rectified linear unit), sigmoid, and tanh. Dropout is a regularization technique used to prevent overfitting during training. It randomly sets a portion of the activations in the feature map to zero during each training iteration. This forces the network to learn more robust and generalized representations by preventing dependence on specific activations. By passing the feature maps through batch normalization, activation, and dropout layers, the network can further refine the learned representations and increase its ability to extract meaningful features from the input data.

[0145] The global average pooling (GAP) layer 404 is a pooling operation designed to replace the fully connected layer in the classic CNN. The GAP layer 404 receives the outputs from all convolutional layers 402 at the same time, such as Figure 4D As shown. The GAP layer 404 in the CNN model 400 calculates the average value of each feature map, reducing the spatial dimension of the data to a fixed-length vector. This technology provides several advantages, including dimensionality reduction, translation invariance, retention of spatial hierarchy, and increased interpretability. By reducing the dimensionality, the model becomes more efficient and less prone to overfitting. Translation invariance ensures the robustness of the model to spatial shifts. Spatial hierarchy retention captures important spatial patterns, and the resulting fixed-length vector enhances interpretability by highlighting the presence or absence of features. Overall, the GAP layer 404 improves computational efficiency, generalization, and interpretability in the CNN model 400.

[0146] In some embodiments, a feature map is generated for each corresponding category of the classification task in the last convolutional layer. Instead of adding a fully connected layer on top of the feature map, the average of each feature map is taken and the resulting vector is fed directly into the softmax layer 408.

[0147] The output of the softmax layer 408 is a continuous value between 0 and 1, where values ​​close to zero indicate high confidence that the SEGM does not correspond to an arrhythmia (e.g., pause) and is a false positive, and values ​​close to 1 indicate high confidence that it is a true positive. This enables the system to be fine-tuned based on clinical needs. For example, the detection threshold can be lowered so that the system is very sensitive to certain types of critical arrhythmias that may not occur regularly. For less critical arrhythmias that may occur more frequently, the classification threshold (e.g., detection threshold) can be increased to obtain higher specificity.

[0148] Figure 4D A comparison of a fully connected layer 430 and a GAP layer 432 according to an embodiment of the present invention is shown. One advantage of the GAP layer 432 over the fully connected layer 430 is that the GAP layer 432 is more natural to the convolutional structure by enforcing the correspondence between the feature map and the category. Therefore, the feature map can be easily interpreted as a category confidence map. Another advantage is that there are no parameters to optimize in the GAP layer 432, so overfitting is avoided in this layer. In addition, the GAP layer 432 sums spatial information, so it is more robust to spatial translation of the input. In some embodiments, the GAP layer 432 requires fewer epochs to have optimal results without overfitting.

[0149] By using a GAP layer 432, each of the activation maps in the final layer before the GAP layer 432 acts as a detector of a different pattern in the input, localized in space. To obtain a class activation map corresponding to the image, the detected patterns are transformed into detected objects. This transformation is done by noting that each node in the GAP layer 432 corresponds to a different activation map, and the weights connecting the GAP layer 432 to the final dense layer encode the contribution of each activation map to the predicted object class. To obtain a class activation map (e.g., classification), the contribution of each detected pattern in the activation map is summed, where detected patterns that are more important to the predicted object class are given more weight.

[0150] In some embodiments, with reference to the fully connected layer 430, a fully connected layer is added on top of the feature map after the last convolutional layer. Each neuron in the fully connected layer 430 is connected to each neuron in the previous layer, thereby effectively flattening the spatial structure of the feature map (e.g., a flattened vector). This allows the network to learn complex relationships between features across the entire image. The output of the fully connected layer is then fed into the soft max layer 408 for classification. With reference to the GAP layer 432, instead of using a fully connected layer, each feature map is averaged individually. This involves calculating the average value of each feature map over its spatial dimensions (e.g., width and height) to obtain a single value for each feature map. These averages are then used as a compressed representation of the feature map. The resulting vector containing the average value from each feature map is fed directly into the soft max layer 408 for classification. The use of global average pooling has several advantages. It reduces the number of parameters in the network because a fully connected layer is not required, which can help alleviate overfitting. Because the average operation is performed individually for each feature map, it also retains the spatial information present in the feature map to some extent. This may be beneficial for tasks where spatial relationships are important, such as EGM signal classification.

[0151] For the last layer, a softmax layer 408 is used as an activation function in the output layer of the CNN model 400 to predict a binomial probability distribution. The softmax formula is as follows:

[0152]

[0153] All Z i The values ​​are the elements of the input vector and can take any real value. k is the number of classes. The term at the bottom of the formula is a normalization term that ensures that all output values ​​of the function sum to 1 and thus constitute a valid probability distribution.

[0154] Deep neural networks give rise to the vanishing gradient problem. Vanishing gradient refers to the fact that in deep neural networks, the back-propagated error signal (gradient) generally decreases exponentially as a function of the distance from the last layer. In other words, the useful gradient information from the end of the network fails to reach the beginning of the network. Batch normalization is a normalization technique applied to the input (current) batch of data, which is the output of the previous convolutional layer. Omitting the strict mathematical details, batch normalization can be simply visualized as an additional layer in the network that normalizes the data (using the mean and standard deviation) before feeding it to the hidden unit activation function. When used with the ReLU activation function, it can solve the problem of vanishing gradients.

[0155] To do this, we have a pair called μ B Take the average value for each layer. This is based on the layer x iIt is calculated by dividing the sum of all values ​​of by the mean of all values ​​m.

[0156]

[0157] Then the variance It can be calculated as follows: 1. Subtract μ, which is the deviation of each value, from each value B and square it to get the squared deviation; and 2. sum the results of doing this for each value and divide by the number of values ​​m. This gives the mean or mean squared deviation:

[0158] =

[0159] The standard deviation can be calculated as the sum of the mean squared error and ε at the root. ε is a constant value as small as 0.001. This is added to avoid division by zero and also increase the variance. The variance of a population is usually larger than the variance of any sample taken from that population, especially in the case of small sample sizes where the mean is close to the peak of the population distribution, so adding a little variance to each batch helps to take that into account. Now calculate the mean and standard deviation, normalized as follows:

[0160]

[0161] The normalized values ​​are then multiplied by γ and added to β. These are learnable parameters used to further scale the normalized values. The final batch normalization values ​​are as follows:

[0162]

[0163] Batch normalization can be applied both before and after the activation function. However, in some embodiments, better results are achieved when batch normalization is applied before the activation function. During training, the mean and standard deviation are calculated using the samples in the mini-batch. However, in testing, it does not make sense to calculate new values. Therefore, batch normalization uses the running mean and running variance calculated during training. Batch normalization can lead to faster training times, and the ability to use higher learning rates without vanishing or exploding gradients is also promising.

[0164] The term "drop" refers to dropping nodes (input and hidden layers) in a NN. All forward and backward connections to the dropped nodes are temporarily removed, thereby creating a new network architecture outside the parent network. Nodes are dropped with a drop probability p. For example, if the hidden layer has 1000 neurons (nodes) and dropout is applied with a drop probability = 0.5, 500 neurons will be randomly dropped in each iteration (batch). Typically, for the input layer, the retention probability (i.e., 1-drop probability) is closer to 1. In some embodiments, the retention probability is about 0.8. For the hidden layer, the larger the drop probability, the sparser the model, where about 0.5 can be the most optimized retention probability, which indicates that 50% of the nodes are dropped. In the overfitting problem, the CNN model 400 learns statistical noise. Specifically, the main motivation for training is to reduce the loss function given all neurons. Therefore, in overfitting, units can change in a way that corrects the errors of other units. This leads to complex co-adaptation, which in turn leads to the overfitting problem because such complex co-adaptation cannot generalize on unseen data sets. If dropout is used, these units are prevented from fixing the errors of other units, thus preventing co-adaptation, since at each iteration the presence of the unit is highly unreliable. Therefore, by randomly dropping a few units (nodes), it forces the layer to take more or less responsibility for the input by taking a probabilistic approach. During inference, the dropout layer is not used. This means that all units are considered during the prediction step. However, since all units / neurons are taken from the layer, the final weights will be larger than expected, and to handle this problem, the weights are first scaled by the chosen dropout rate. This way, the network will be able to make accurate predictions. In some cases, if a unit is retained with probability p during training, the output weight of this unit is multiplied by p during the prediction phase.

[0165] Hyperparameter optimization and training

[0166] When implementing a NN, several choices (or hyperparameters) must be set before training—these range from the type of architecture to the depth and width of the layers, to the activation functions of neurons in different layers, and so on. Since arbitrary choices of hyperparameters can lead to suboptimal results, a three-way split of the data into training, validation, and test sets was created to identify reasonable architectures and parameter ranges. NN optimization can then be performed via KerasTuner and grid search, guided by these pre-established ranges. By adopting this process, a large space of possible models can be evaluated and many configurations tested.

[0167] Figure 4EA table of exemplary hyperparameters of interest and the range of available options during search according to embodiments of the present invention is shown. The hyperparameters of interest consist of activation function, dropout percentage, learning rate, learning rate decay, nodes per layer, and optimizer. Additional hyperparameters for convolutional models include the number of filters, kernel size, and regularization. It should be understood that other hyperparameters may be used.

[0168] In some embodiments, for a data set, the following process can be used. For each set of hyperparameters sampled in the search, the data can be divided into a training set (80%) and a validation set (20%). Therefore, the proposed architecture is trained on the training data set. Then, in order to evaluate the architecture, the validation accuracy score can be checked for different thresholds. The network architecture with the highest average accuracy score for all thresholds can be selected. In some cases, all or some networks can be trained for 1000 epochs using an early stopping condition. In some cases, when the accuracy of the validation set does not improve within 25 epochs, the training stops. Ten-fold cross validation can be used to train all models. This cross-validation process requires ten different trainings for a given model (e.g., reinitializing the network parameters each time), and ensures that i) different subsets of the data are used for training and testing, and ii) each data point is used as part of the training set (e.g., nine times) and in the test set (e.g., once). In some cases, when cross-validation is performed, the data partitioning used is not used during the hyperparameter search.

[0169] Data augmentation and regularization used during training

[0170] To avoid overfitting when training a neural network, the SEGM in the training set is augmented. By augmenting the data, the CNN model 400 is exposed to a wider range of examples, helping it learn more robust and invariant representations. Data augmentation is particularly useful when training deep learning models with limited labeled data because it effectively increases the effective size of the training set without incurring additional data collection or labeling costs.

[0171] In general, in machine learning, but especially for deep learning, classification accuracy tends to depend on the amount of training data; limited training data often leads to low accuracy. Data augmentation involves systematically generating new samples to enhance existing datasets by transforming existing samples in a way that increases the accuracy and stability of the classification. When trying to generalize a model to a new dataset, exposing the classifier to a changing representation of its training samples often makes the model more invariant and robust to such transformations. In some embodiments, popular techniques for data augmentation of signals fall into five categories: noise addition, generative adversarial nets (GANs), sliding windows, Fourier transforms, and other techniques such as inversion, scaling, magnification, and reduction. In some embodiments, different data augmentation methods are selected based on the dataset.

[0172] For example, data augmentation can be used to avoid overfitting and improve the performance of CNN models. To enhance a training data set using inversion, the input signal can be flipped by multiplying by -1, effectively creating a mirrored version of the original data. This enhancement technique provides an additional change in the temporal characteristics of the EGM signal (e.g., amplification factor = 2, percentage 100%). The intuition behind this data augmentation is that the IMD 100, 212 can be implanted upside down and can move in the body, resulting in a negative R peak in the data set. In some cases, inverting the original training data set effectively doubles the training data set. In some embodiments, the amplification factor used for data augmentation can be approximately 2.

[0173] As an example only, referring to noise addition, in some embodiments, a Gaussian noise layer (mean = 0, std = 0.05) can be added to alleviate overfitting, and since it is a regularization layer, it is only activated during training. In other embodiments, various other types of noise (such as Poisson noise, salt and pepper noise, etc.) can be added to the original EGM signal with different parameters (e.g., mean (μ) and standard deviation (σ)). In a further embodiment, Gaussian noise with different parameters (mean = 0, standard deviation = σ = (0.01, 0.1, 0.2, 0.5)) can be added to all channels of the original EGM signal.

[0174] In some embodiments, data enhancement may include inserting one or more p-waves at previous RR intervals during the pause interval. Figure 4C For reference, enhancement would include adding one or more p-waves of the same or different morphology to the pause duration or pause interval, starting with the last p-wave at about -3 seconds before the QRS complex, and adding one or two p-waves before the trigger using an RR interval that can be chosen (e.g., the average of the first few p-waves, between 2 and 10 cycles) to obtain an approximate p-wave morphology. This can be used to enhance a true pause episode, where no p-waves are present during the pause duration.

[0175] In other embodiments, data enhancement may include stretching or shrinking the duration of the pause interval. Figure 4C For reference, the signal can start stretching about 500ms (about -2 seconds) after the last QRS and lengthen / contract the pause duration or pause interval to change the duration of the pause interval to another physiologically valid duration, such as between 3 and 10 seconds. This can be achieved by stretching or compressing and then interpolating using splines, linear interpolation, in some cases by removing segments, or by duplicating short segments on the time axis and interleaving the duplicated segments.

[0176] In some embodiments, as the data set expands with more samples, it is beneficial to continue to adopt one or more data enhancement methods. By combining various enhancement techniques, the robustness of the trained model can be further increased, and its ability to generalize to unseen data can be improved. In addition to signal inversion, a Gaussian noise layer can be utilized during the training phase. A Gaussian noise layer refers to a layer in which random Gaussian (normal distribution) noise is added to the input data during the training process. The noise layer helps introduce random elements into the neural network, which can also be beneficial for regularization and generalization. The noise layer introduces random changes to the input data, mimicking the presence of background noise commonly encountered in EGM recordings. By including a Gaussian noise layer, (one or more) processors effectively perform a form of random data enhancement. This enhancement strategy helps train the model to make it more resilient to noise and enhances its ability to accurately classify EGM signals in real-world scenarios. In addition, by combining signal inversion and a Gaussian noise layer, the training data set is enriched with enhanced samples that capture a wider range of changes present in the EGM signal. This enhanced data set contributes to better model training, improves generalization, and enhances the model's ability to handle noisy or distorted EGM data during reasoning.

[0177] Hierarchical feature extraction

[0178] The CNN model 400 utilizes hierarchical layers of filters that automatically capture different levels of features from the input data, such as low-level features, high-level features, and feature levels on a continuum from low-level to high-level. In the case of cardiac rhythm analysis, these filters can effectively capture subtle and complex patterns associated with different types of seizures. By automatically learning and extracting relevant features, the CNN model 400 enhances the accuracy and robustness of the classification process.

[0179] Feature-level interpretations may be defined within the system and may vary based on the specific question and data set. In some embodiments, low-level features in the EGM signal may refer to basic waveform patterns (such as P waves, QRS complexes, T waves, etc.). In some cases, these features may be relatively easy to identify visually because they exhibit unique shapes and amplitudes. However, it is important to note that the visual perception of these features may vary based on an individual's expertise and familiarity with EGM interpretation, and therefore different people may report different results based on visual perception.

[0180] The order of feature extraction may depend on the problem and the network architecture used. In some CNNs, the feature extraction process is hierarchical, where lower layers learn low-level features and higher layers learn more complex and abstract features. It is common to start with low-level features and gradually move to higher-level features in a bottom-up manner. This order helps capture local patterns before capturing more global or context-dependent patterns. In this example, convolutional layer C1410 extracts lower-level features compared to convolutional layers C2-C5 412-418, and each convolutional layer C2-C5 412-418 gradually extracts higher-level features than the previous convolutional layer. In other embodiments, convolutional layer C1 410 extracts higher-level features than convolutional layer C2 412, convolutional layer C2 412 extracts higher-level features than convolutional layer C3 414, and so on. In further embodiments, each of the convolutional layers C1-C5 extracts features at different levels.

[0181] Improve data utilization

[0182] CNN model 400 can leverage large amounts of data to improve performance. With more data, model 400 can learn more diverse patterns and generalize better to new instances. This is particularly advantageous in the medical field, where access to large amounts of labeled datasets can be challenging. By leveraging more features on the CNN, such as patient-specific data, the model can capture personalized characteristics, thereby improving the accuracy of classifying heart attacks.

[0183] Performance Measurement

[0184] Common quantitative metrics can be used to evaluate the performance of classifiers. The classifications produced by the learning method can be either anomalies or normal cases, referred to as positive or negative classes, respectively. The predicted results can also be true or false, meaning a correct prediction or an incorrect prediction, respectively. For example, the classification can be summarized into four possible states: a) True Positive (TP): a correct prediction of the positive class, b) True Negative (TN): a correct prediction of the negative class, c) False Positive (FP): an incorrect prediction of the positive class, or d) False Negative (FN): an incorrect prediction of the negative class.

[0185] Based on the classification predictions (e.g., TP, TN, FP, FN), accuracy, specificity, sensitivity, precision, recall, positive predictive value (PPV), negative predictive value (NPV), and area under the curve (AUC) can be calculated. Figure 4FExample equations that can be used to calculate multiple metrics according to embodiments of the present invention are shown. Additionally or alternatively, the performance of the classification model at all classification thresholds can be evaluated by a receiver operating characteristic curve (ROC). In some embodiments, sensitivity can be used to select the best CNN model with the best hyperparameters. In other embodiments, high sensitivity, specificity, and F1 score (or other measure of model accuracy) and high AUC of the ROC curve can be used to select the best CNN model.

[0186] Personalization

[0187] CNNs have the ability to learn and adapt to individual patient data. By training the model on different datasets that include patients with different heart rhythms, the CNN model can capture patterns and nuances specific to each patient. This personalized approach enhances the accuracy of the classification algorithm by taking into account patient-specific variations and reducing false positives.

[0188] In addition, training data associated with certain conditions, unique conditions, pathologies, and / or medications can be used to train or fine-tune the CNN model 400. In some cases, certain episodes, such as (one or more) specific pause episodes, can be used to fine-tune the CNN model 400 or create a fine-tuned CNN model customized for a specific episode. In some embodiments, the CNN model 400 can be fine-tuned based on data from one or more patients. In other embodiments, an adjusted CNN model (e.g., a copy of the main class CNN model 400 that has been further adjusted to include parameters that are better customized for one or more patients) can be created and used to analyze data sent to the data server 304 by those one or more patients.

[0189] In some embodiments, one or more hyperparameters of the CNN model 400 (such as filter size, kernel size, etc.) can be adjusted to generate a fine-tuned CNN model for one or more patients.

[0190] In the personalized CNN model, the EGM data is passed through the CNN model to capture relevant features. However, in addition to the EGM data, additional information such as demographic characteristics and age can be incorporated into the personalized CNN model. These additional features can be combined with the learned EGM features at the fully connected layer 406. By incorporating this supplementary information, the personalized CNN model 400 can utilize demographic characteristics and age to enhance its performance and customize predictions for individual patients.

[0191] In some embodiments, N personalized CNN models can be created based on different patient categories or combinations of categories, as described below. A physician can evaluate and determine the most appropriate model for a particular patient based on the individual characteristics and / or combinations of characteristics of the particular patient: i) Age-related features: In addition to age itself, additional age-related features can be added, such as indicators of age groups (e.g., young, middle-aged, elderly), age ranges, or different life stages; ii) Gender: Including gender as a feature can help capture any gender-specific variations in EGM abnormalities; iii) Body mass index (BMI): BMI is a measure of body fat based on height and weight, which can provide insights into the potential impact of body composition on EGM abnormalities; iv) Race: Race can play a role in certain cardiovascular conditions, and including this feature can help capture any population-specific variation; v) Medical history: Incorporating relevant medical history features (such as hypertension, diabetes, previous heart disease, or other relevant chronic conditions) can provide valuable context for identifying EGM abnormalities; vi) Medication use: Information about medications a person is currently taking can be included, especially if certain medications are known to affect EGM readings or can indicate an underlying condition; vii) Lifestyle factors: Including features related to lifestyle choices, such as smoking status, alcohol consumption, exercise habits, or occupation, can provide additional insights into potential risk factors; and viii) Family history: Incorporating information about family history of cardiovascular disease or EGM abnormalities can help identify individuals with a higher predisposition to that condition.

[0192] In other embodiments, the physician may collaborate, such as with a data scientist, to iteratively refine the personalized CNN model (e.g., iterative refinement). The physician may provide feedback on false positives, false negatives, and areas where the CNN model may benefit from further fine-tuning to align with real-world clinical scenarios. By involving physicians throughout the programming process, the personalized CNN model may benefit from their clinical expertise and domain knowledge. This collaboration helps ensure that the personalized CNN model is aligned with medical standards, enhances accuracy, and is clinically relevant, which may improve patient care and outcomes beyond the improvements already experienced using the CNN model 400.

[0193] In further embodiments, physicians can create personalized CNN models based on more general CNN models and / or feed personalized information into personalized CNN models. For example, physicians can select options provided on, for example, but not limited to, a graphical user interface (GUI) to personalize the CNN model, select the CNN model they wish to personalize, and select one or more options (e.g., categories, features) provided on the GUI. In other cases, data can be input using voice, gestures, cameras, microphones, etc. In some cases, physicians can customize personalized CNN models as patients change medications and / or medication dosages, are diagnosed with new conditions, etc. The physician may then be able to further customize individual personalized CNN models in the personalized CNN model for one or more of the needs of a specific patient, thereby further personalizing medical care.

[0194] Transfer Learning

[0195] CNNs can benefit from transfer learning, where knowledge gained from training on one dataset can be transferred to a related task or dataset without the need to update the feature extraction of the five convolutional layers 402. In the context of IMD 100, 212, transfer learning enables the development of CNN models for automatic classification of different cardiac episodes. For example, a CNN model trained on a large dataset for pause detection can be fine-tuned and adapted to classify bradycardia episodes in IMD 100, 212. By leveraging pre-trained knowledge and common features learned from related tasks, transfer learning reduces the need for large amounts of labeled data specific to each episode type and accelerates the development of accurate algorithms.

[0196] In some embodiments, the hyperparameters of the five convolutional layers 402 of the CNN model 400 may be frozen or fixed, and the fully connected layer 406 may be fine-tuned. Transfer learning may utilize streaming data and new data from patients that have not yet been adjudicated by an expert adjudicator.

[0197] In some cases, transfer learning can be used to update the CNN model 400 periodically as additional data is received, such as weekly, monthly, etc. Updating the CNN model 400 in this manner will improve the performance of the model over time.

[0198] During transfer learning, new data travels through the parameters of the pre-trained CNN model in a process called fine-tuning. The following steps generally summarize the process of fine-tuning, but it should be understood that other steps not included may be used and some steps may be omitted:

[0199] i) Pre-trained Model: Transfer learning starts from a pre-trained CNN model (e.g., CNN model 400) that has been trained on a large dataset for a specific task (such as pause detection, image classification, natural language processing, etc.). The pre-trained CNN model contains learned parameters (e.g., weights, biases) that capture common patterns and features from the original dataset on which it was trained;

[0200] ii) Frozen layers: In transfer learning, the initial layers of the pre-trained CNN model that capture low-level and common features are usually frozen. This means that their parameters are not updated during the training process and their learned representations are retained;

[0201] iii) New data: New data, usually from a different but related task or domain, is introduced into the transfer learning pipeline. This data can be a smaller dataset specific to the new task (e.g., bradycardia episodes) or additional data from the original task (e.g., additional pause events);

[0202] iv) Forward pass: During the forward pass, new data is passed through the pre-trained CNN model starting from the input layer. As the data propagates through the network, it undergoes transformations based on the learned parameters. Frozen layers retain their original parameters and are not updated based on new data, while unfrozen layers (usually the top layers) are trainable and will be updated;

[0203] v) Backpropagation: After the forward pass, the loss or error between the predicted output and the ground truth labels is calculated. Then, backpropagation is performed to calculate the gradient of the loss with respect to the trainable parameters of the unfrozen layers;

[0204] vi) Gradient Descent: The gradients obtained from backpropagation are used to update the parameters of the unfrozen layers through an optimization algorithm (usually gradient descent). The optimizer adjusts the weights and biases of these layers to minimize the loss, making the model more suitable for the new task;

[0205] vii) Training Iterations: The process of forward pass, backpropagation, and parameter updates is repeated for multiple iterations or epochs, allowing the CNN model to gradually adapt to new data. This process refines the learned representations in the unfrozen layers to align with specific patterns and features present in the new task. By allowing new data to propagate through unfrozen layers while keeping the frozen layers intact, transfer learning leverages previously learned knowledge to accelerate training and improve performance on new tasks.

[0206] During transfer learning, the process of updating the parameters of the CNN model 400 is typically done automatically by an optimization algorithm such as gradient descent. The CNN model 400 does not manually change parameters; instead, it learns to update them based on new data and a defined optimization objective. The optimization algorithm guided by the loss function calculates the gradient of the loss with respect to the model parameters. These gradients represent the direction and magnitude of the parameter updates required to minimize the loss. The optimizer then uses these gradients to adjust the parameters. In the context of transfer learning, the parameters of the pre-trained model are typically partially frozen, and only the parameters in the unfrozen layers are updated. The frozen layers retain their learned representations and do not change during training on new tasks or datasets.

[0207] By allowing the optimization algorithm to automatically update the parameters, the model adjusts its representations and fine-tunes them to better suit new tasks or datasets. This automatic adjustment process based on new data helps the CNN model capture task-specific patterns and improve its performance. In some embodiments, the following code can be used:

[0208] #Freeze layers until a certain layer

[0209] freeze_layers_until=5

[0210] For layers in pretrained_model.layers[:freeze_layers_until]:

[0211] layer.trainable=False

[0212] Figure 5 A process for training / building a convolutional neural network (CNN) model to analyze a DCA dataset with respect to arrhythmias of interest, such as pauses, is shown in accordance with embodiments herein. Figure 5 The operations of may be implemented by hardware, firmware, circuitry, and / or one or more processors partially and / or completely housed within an IMD, a local external device, a remote server, a combination thereof, and / or more generally within a healthcare system. Optionally, Figure 5 The operations of may be implemented in part by the IMD and in part by a local external device, a remote server, or more generally within the healthcare system. For example, the IMD includes an IMD memory and one or more IMD processors, while each of the external devices / systems (EDs) (e.g., local, remote, or anywhere within the healthcare system) includes an ED memory and one or more ED processors.

[0213] At 502, one or more processors of the system obtain a set of reference DCA data sets. In some embodiments, each of the reference DCA data sets includes a far-field reference CA signal (e.g., a SEGM signal) for one or more beats sensed along a sensing vector between a combination of subcutaneous electrodes that are not positioned intravenously. The subcutaneous electrodes can be disposed on or coupled to the IMD. For example, the subcutaneous electrodes can be disposed on the housing of the IMD 100 and / or on a non-transvenous lead coupled to the subcutaneous IMD. Each of the DCA data sets also includes one or more device record (DD) markers generated by the IMD that characterize the CA signal within the corresponding DCA data set. The reference DCA data set includes CA signals of one or more reference beats that are known to include corresponding arrhythmias of interest (such as pauses). The reference DCA data set also includes CA signals of reference beats that are known to be normal and do not include the arrhythmia of interest.

[0214] Additionally or alternatively, an ICM (such as a leadless cardiac monitor in US9,949,660) senses and classifies cardiac activity. Additionally or alternatively, a leadless IMD (such as in US Pat. No. 9,216,285) senses cardiac activity. Additionally or alternatively, a transvenous IMD (such as in US Pat. No. 8,391,980) senses cardiac activity. For implantable devices that do not classify cardiac activity within the device, the cardiac activity can be sent to an external device for classification and / or further processing to include one or more DD markers, generate a reference DCA dataset including CA signals of one or more reference cardiac beats known to include pauses and / or a reference DCA dataset including CA signals of reference cardiac beats known not to include pauses.

[0215] In some embodiments, a 6-second EGM strip (other time lengths are contemplated) can be used as a reference DCA dataset, where the 6-second EGM strip is known to include pause episodes. Multiple individual 6-second EGM strips are collected at different time points for a patient, for a patient population, recorded by various device types, device placements, device orientations, etc., where each of the individual 6-second EGM strips has a CA signal known to include a corresponding arrhythmia of interest (e.g., pauses). As another example, a second set of 6-second EGM strips is obtained, where the second set includes a reference DCA dataset known to correspond to a normal rhythm. A set of reference DCA datasets can be recorded for a patient, for a patient population, and a set of reference DCA datasets can be recorded by various device types, device placements, device orientations, etc.

[0216] The reference DCA data set includes device-recorded markers (e.g., pause indicators, R wave markers, P wave markers, RR intervals, AF indicators) that identify cardiac beats sensed by the device within a series of cardiac events. The cardiac activity data may have been previously acquired and stored in a memory of an implanted or external monitoring device, an implanted or external therapy delivery device, a programmer, a workstation, a healthcare network, or other system. When a reference DCA data set has been previously acquired, the obtain operation at 502 represents accessing and reading the previously stored reference DCA data set from one or more memory locations.

[0217] The CA signal is for one or more cardiac events spanning various time periods. As an example, multiple segments or sets of cardiac activity data may be collected, each of which is for an interval of 6 seconds to 5 minutes in length. Optionally, a segment may include one or more IMD-announced pause episodes. As another example, each of the segments or sets of cardiac activity data may be collected within an interval that starts 10-60 seconds before an episode of interest (e.g., a pause episode) and ends 10-60 seconds after the episode of interest. The CA signal may include one or more arrhythmia episodes. The DCA data set obtained at 502 may include one or more detected arrhythmia episodes and / or one or more heart beats that are confirmed to be normal without arrhythmia episodes. The DCA data set obtained at 502 may correspond to a continuous series of cardiac events (e.g., 1 continuous series lasting 6 seconds to 5 minutes) and / or a separate set of cardiac events (5, 10 or more separate series, each series lasting 6 seconds to 5 minutes of cardiac events).

[0218] When the IMD detects an arrhythmia episode of interest (e.g., a pause), the IMD may automatically initiate the collection and analysis of CA signals. Additionally or alternatively, the IMD may collect and analyze CA signals in response to a user-initiated instruction or a clinician instruction. For example, a user or clinician may utilize a smart phone, programmer, or other portable device to establish a communication session with the IMD and instruct the IMD to begin collecting and analyzing cardiac signals, such as when a patient is experiencing discomfort, feels faint, has a fast heart rate, during an outpatient visit, etc.

[0219] At 504, one or more processors enhance the set of reference DCA datasets by generating synthetic DCA datasets based on the reference DCA datasets to expand the total number of DCA datasets. Each of the synthetic DCA datasets is based on a corresponding reference DCA dataset, but differs from the corresponding reference DCA dataset in some manner. Herein, the combination of the reference DCA dataset and the synthetic DCA dataset is referred to as an "enhanced set of DCA datasets" and is used when training the CNN model 400. By enhancing the set of DCA datasets, embodiments herein avoid overfitting when training the CNN model 400. The enhancement process can be performed prior to use, or in real time or substantially real time while the enhanced set of DCA datasets is provided to the CNN model 400.

[0220] The CNN model 400 can be trained based on raw CA signals from a reference DCA dataset and an enhanced set of synthetic DCA datasets. As described above, enhancements can be implemented in various ways. As non-limiting examples, enhancements such as noise addition and / or inversion can be used to generate a synthetic DCA dataset based on the reference DCA dataset. Other enhancements include amplification, inserting one or more p-waves at the previous RR interval during the pause interval, and / or stretching or shrinking the duration of the pause period.

[0221] Additionally or alternatively, other types of enhancements may be used, and the order in which the types of enhancements are applied may be changed to generate additional synthetic DCA datasets. For example, a dataset that has been noise enhanced may be further enhanced by inversion, or vice versa. In some embodiments, using more than one or more different types of enhancements may help create a greater number / quantity of synthetic or enhanced DCA datasets.

[0222] At 506, one or more processors determine whether a sufficient amount of enhancement has been applied, i.e., whether a sufficient amount of synthetic DCA datasets have been generated. If not, the process returns to 504, where additional synthetic DCA datasets are generated. Once the desired amount of synthetic DCA datasets and reference DCA datasets are available, the process moves to 508. In some embodiments, if additional reference DCA datasets are needed, the process returns to 502. The decision at 506 enables the process to build multiple types of synthetic DCA datasets based on the reference DCA datasets in order to expand the overall set of DCA datasets by a predetermined amount. For example, it may be desirable to double, triple, or otherwise amplify the original set of reference DCA datasets by a factor of X.

[0223] As another example, different combinations may be defined to achieve different percentages of each type of enhancement within the overall data set. For example, it may be desirable that 50% of the data be a total reference DCA data set, 20%-30% be a synthetic DCA data set generated from noise addition of the original DCA data set, 10%-15% be a synthetic DCA data set generated by inversion of the original DCA data set, and the remainder be generated using both noise addition and inversion, including any and all permutations and combinations thereof. In other embodiments, a certain percentage of the synthetic DCA data set is generated by inserting one or more p-waves and / or the duration of the stretch and / or contraction pause intervals during the pause interval. It should be noted that different percentages of the reference DCA data set, synthetic DCA data sets generated in one of N enhancement modes, synthetic DCA data sets generated in two of N enhancement modes, etc. may be used. The percentages may be modified based on the type of input data, the results achieved, etc.

[0224] At 508, the reference DCA dataset and the enhanced DCA dataset are input into the convolution layer C1410 of the CNN model 400. As previously described, the reference DCA dataset and the enhanced DCA dataset may be of a predetermined length (such as 4s, between 4s and 10s, greater than 10s, etc.). The length of each of the dataset segments may vary, so it is not required that each segment has the same length.

[0225] At 510 , one or more processors train the CNN model 400 using the enhanced set of reference DCA datasets and synthetic DCA datasets to assess heart rhythm without considering device information (eg, DD markers), as discussed herein.

[0226] The trained CNN model 400 can provide various types of outputs. In one example, the CNN model 400 can be trained to simply output an indication of whether the corresponding DCA dataset is valid or invalid.

[0227] Additionally or alternatively, the CNN model 400 can be trained to output a confidence indicator (e.g., probability, likelihood, continuous value between 0 and 1) that indicates a level or degree of confidence assigned to a true positive or false positive that an individual underlying DCA dataset represents an arrhythmia (e.g., pause). As an example, the numeric indicator can be a continuous value between 0 and 1, where a value close to zero indicates a high confidence that the SEGM signal within the DCA dataset does not indicate a pause, and is therefore a false positive. A value close to 1 indicates a high confidence that the SEGM signal within the DCA dataset indicates a pause, and is therefore a true positive. When the DCA dataset does not indicate a pause, the DCA dataset may include a CA signal that indicates normal sinus rhythm or other.

[0228] Additionally or alternatively, the CNN model 400 may be trained to provide at least one of the following outputs in conjunction with each DCA dataset: i) a confidence indicator indicating a true positive or false positive designation of confidence that the corresponding DCA dataset represents a pause, ii) a confidence indicator indicating the accuracy of pause sensing implemented by the IMD, iii) a recommendation indicating a sensitivity level for the IMD to use to identify pauses in CA signals, iv) an indicator that the associated DCA dataset is a true positive, v) an indicator that the associated DCA dataset is a false positive, vi) an indicator that the associated DCA dataset is a true negative, vii) an indicator that the associated DCA dataset is a false negative, viiii) an uncertain / inconclusive DCA dataset, or ix) an invalid DCA dataset. In some embodiments, the number of results in the uncertain / inconclusive category may be minimized by adding more samples identified by the CNN as uncertain but using similarity to the training dataset, or by random sampling and increasing the training dataset size. In other embodiments, if the event is infrequent, the CNN may be unable to evaluate the event due to the infrequent event. In still further embodiments, DCA data sets with lower confidence metrics may be sent to a physician or other decision maker for review.

[0229] Additionally or alternatively, the CNN model 400 can be trained to include a detection threshold that can be changed / fine-tuned by a clinician based on clinical needs and various factors. In the aforementioned example, where a value close to 1 indicates a high confidence of a true positive and a value close to 0 indicates a high confidence of a false positive, the detection threshold can be lowered (e.g., closer to 0) to increase the sensitivity of the CNN model 400. For example, when the detection threshold is set to 0.25, the CNN model 400 will identify more false positive DCA data sets than when the detection threshold is set to 0.75. For example, it may be desirable to apply a higher level of sensitivity in conjunction with certain types of critical arrhythmias (e.g., pause events) that may not occur regularly. In addition, it may be desirable to apply a higher threshold, and thus reduce sensitivity, while increasing specificity, in conjunction with other types of pause events that are considered less "critical" to the patient's health and may occur more often.

[0230] At 512, one or more processors save the CNN model 400 for subsequent use. For example, the CNN model 400 may be saved in an IMD 100, 212 that is or will be associated with a patient, an external instrument 154, 214, and / or one or more remote systems. Figure 5The process can be implemented across a population, a medical network in another region, to build one or more CNN models, etc. Additionally or alternatively, the CNN model can be custom trained for each clinic. Training can be performed at regular intervals or as needed. Clinics can choose to train on their own data or on data from "trusted" clinics, or both. Performance across model versions can be compared, and clinics can optionally revert to an earlier model.

[0231] Additionally or alternatively, the CNN model 400 can be constructed in alternative ways and / or utilize alternative parameters / hyperparameters. For example, the CNN model can utilize (one or more) other activation functions, such as leaky ReLu, Sigmoid, tanh, etc. The network architecture can be a different number of network layers, sub-layers, hidden layers, etc., other ways of performing data augmentation. Additionally or alternatively, the CNN model 400 can be trained to perform pulsation level determination to obtain greater visibility into classification behavior. Additionally or alternatively, the CNN model 400 can be trained to alert clinicians to inappropriate sensing and / or recommend device reprogramming, changes to one or more settings, etc.

[0232] Once the master CNN model 400 is saved at 512, the CNN model 400 may be personalized at 514, such as by combining additional features with the learned EGM features at the fully connected layer 406 to generate a personalized CNN model. Any number N of personalized CNN models may be made to personalize healthcare and treatment for one or more patients.

[0233] Similarly, at 516, transfer learning may be used to train a new CNN model based on the CNN model 400 of 510 or the personalized CNN model of 514, as previously discussed.

[0234] Figure 5 The operations may be staged to be performed on the CA signals at various times, such as in real time (eg, during or shortly after a patient experiences a seizure) or at any time after the CA signals are stored. Figure 5 The operations of the IMD 100 may be performed by devices and systems at various proximity to a patient having the IMD 100. For example, the CA signal may be read from the IMD and sent to a local portable external device (e.g., a smartphone, desktop computer, laptop computer, smart watch, etc.), where the local portable external device implements the combined local operation when in close proximity to the patient. Figure 5 In other embodiments, the CA signal may be sent / transmitted to implement the combined Figure 5 Describes the operation of all or part of a remote device / system.

[0235] Fig. 6AA process for distinguishing pause episodes for valid and invalid device classifications according to embodiments herein is shown. Fig. 6A The operations of may be implemented by hardware, firmware, circuitry, and / or one or more processors partially and / or completely housed within an IMD, a local external device, a remote server, a combination thereof, and / or more generally within a healthcare system. Optionally, Fig. 6A The operations of may be implemented in part by the IMD and in part by a local external device, a remote server, or more generally within the healthcare system. For example, the IMD includes an IMD memory and one or more IMD processors, while each external device / system (ED) (e.g., local, remote, or anywhere within the healthcare system) includes an ED memory and one or more ED processors.

[0236] At 602, one or more processors of the system obtain one or more CNN models to be used in conjunction with analyzing a previously acquired DCA dataset. For example, the external device may obtain a CNN model from a collection of CNN models stored in a memory of the external device and / or at a remote server. The CNN model selected from the collection may be selected based on the type of pause in the DCA dataset, a specific CNN model fine-tuned for a specific patient or patient class, and the like. Additionally or alternatively, the CNN model may be selected based on the type or model of the IMD, the type or version of the pause detection algorithm utilized by the IMD, and the like. Additionally or alternatively, the CNN model may be selected based on the characteristics of the patient (e.g., age, sex, weight, ethnic background, medical history, type of implanted device, duration of device implantation) based on the clinician or clinic that will view the DCA dataset. In some embodiments, the same patient data may be analyzed by two or more CNN models.

[0237] At 604, one or more processors of the system obtain a set of DCA data sets generated by the IMD for a corresponding set of candidate pause episodes announced by the IMD. The DCA data sets may also be referred to as "candidate" DCA data sets because the CNN model 400 has not verified that the IMD has determined that the pause is actually occurring. Each DCA data set may correspond to a separate candidate pause episode. Additionally or alternatively, a subset of the DCA data sets may correspond to related candidate pause episodes (such as when a patient is experiencing multiple related pause episodes occurring in succession). The DCA data sets may be generated by the IMD for a period of time prior to the implementation of the operation of FIG. 6, and / or generated by the IMD in real time during the operation of FIG. 6. For example, the IMD may monitor the patient and collect DCA data sets over a period of hours, days, weeks, or longer. Periodically, previously acquired DCA data sets may be wirelessly transmitted (uploaded) from the IMD to an external device for storage and / or further transmission to a remote server. The obtaining operation at 604 may simply correspond to an external device and / or a remote server accessing a previously stored set of DCA data sets. Additionally or alternatively, the obtaining operation at 604 may include, at least in part, wirelessly receiving the DCA dataset at an external device and / or receiving the DCA dataset at a remote server 304 (eg, over a network).

[0238] Each of the DCA datasets includes far-field cardiac activity (CA) signals of one or more beats sensed by an implantable medical device (IMD) and (optionally) one or more device recording (DD) markers generated by the IMD. For example, a single DCA dataset corresponds to one pause episode and / or a series of pause episodes occurring consecutively over a short period of time. For example, one DCA dataset may extend over a series of beats (e.g., 7-10 beats). Additionally or alternatively, a DCA dataset may be defined as including CA signals within a predetermined time period (e.g., 30s) before and / or after one or more pause episodes. Additionally or alternatively, a DCA dataset may include approximately 60 seconds of EGM, where the window is selected before triggering the evaluation, which is the duration expected by the CNN based on the design of the CNN. Figure 4C Examples of triggers and windows are shown in .

[0239] In some embodiments, the DD flag is ignored by the CNN model 400. When reviewing the DCA data set in person, the clinician or other personnel may utilize the DD flag, such as to confirm the diagnosis of the CNN model 400. In other embodiments, the IMD will abandon the DD flag to generate data to be processed by the CNN model 400.

[0240] At 606, one or more processors apply the DCA dataset to the CNN model 400, and in response thereto, obtain a confidence indicator, a valid / invalid indicator, etc. from the CNN model 400, which indicates a confidence level or degree specified as a true positive or false positive that the current DCA dataset represents a pause. For example, the confidence indicator may represent a probability or likelihood that the corresponding DCA dataset corresponds to a pause or corresponds to a normal sinus rhythm. Additionally or alternatively, the confidence indicator may represent a digital indicator between 0 and 1, where a value close to zero indicates a high confidence that the subcutaneous EGM signal within the DCA dataset does not indicate a pause (e.g., normal sinus rhythm, an excessively noisy signal that should not be characterized otherwise, an abnormal rhythm indicating another condition other than a pause), and is therefore a false positive. A value close to 1 indicates a high confidence that the subcutaneous EGM signal within the DCA dataset indicates an arrhythmia, and is therefore a true positive. The CNN model 400 may output at least one of the following in conjunction with each DCA dataset: i) a confidence indicator indicating a true positive or false positive designation of confidence that the corresponding DCA dataset represents a pause; ii) a confidence indicator indicating an accuracy of pause sensing implemented by the IMD; or iii) a recommendation indicating a sensitivity level for the IMD to use in identifying R waves in CA signals. In some embodiments, the confidence indicator levels may be set for different arrhythmia types. For example, the confidence indicator level may be set in a first setting when indicating a first type of pause and in a second setting when indicating a second type of pause.

[0241] At 608, one or more processors compare the confidence indicator provided by the CNN model 400 with one or more detection thresholds. In the aforementioned example, the confidence indicator may correspond to a numerical value close to 1 indicating a high confidence of a true positive, a numerical value close to 0 indicating a high confidence of a false positive, and any value therebetween. The clinician may change / fine-tune the detection threshold based on clinical needs and various factors. The detection threshold may be lowered (e.g., closer to 0) to increase the sensitivity of the CNN model. For example, when the detection threshold is set to 0.25, the CNN model will identify a greater number of false positive DCA data sets than when the detection threshold is set to 0.75. For example, it may be desirable to apply a higher level of sensitivity in conjunction with certain types of pauses that may occur irregularly. In addition, it may be desirable to apply a higher threshold, and thus reduce sensitivity while increasing specificity, in conjunction with other types of pauses that are considered less "critical" to the patient's health and may occur more often. When the confidence indicator exceeds the threshold, the process moves to 610. When the confidence indicator does not exceed the threshold, the process moves to 612. In some embodiments, different detection thresholds may be set for different types of pauses and / or different types of arrhythmias. Different detection thresholds may also be set by practitioners, clinics, etc.

[0242] At 610, the one or more processors add the DCA dataset to a valid dataset list to begin building a valid subset of the DCA dataset. At 612, the one or more processors add the DCA dataset to an invalid dataset list to begin building an invalid subset of the DCA dataset.

[0243] At 614, the one or more processors determine whether there are additional DCA data sets to be checked from the collection. If so, the process moves to 616, the one or more processors move to the next DCA data set in the collection, and repeat the operations at 606-614. Otherwise, the process moves to 618.

[0244] The operation of FIG. 6 is iteratively repeated until all DCA datasets have been analyzed by the CNN model 400 and grouped into valid or invalid subsets.

[0245] At 618, one or more processors then present information related to at least the valid subset of the DCA dataset to the clinician. The information presented may vary. For example, the information may display and allow the clinician to analyze each of the DCA datasets in the valid subset. Additionally or alternatively, the information may include content related to the invalid subset. For example, the clinician may be informed that 80% of the DCA datasets were identified as valid by the CNN model, while 20% were identified as invalid. The information may include an option to allow the clinician to view the invalid subset of the DCA dataset.

[0246] At 620, the one or more processors provide treatment to the patient and / or provide a recommendation for treatment to the patient. In some cases, the treatment and / or the recommendation for treatment may increase the DCA data set identified as a valid subset of the DCA data set. Additionally or alternatively, steps 604-620 may be iteratively repeated in response to changes in treatment for the patient or the IMD. In some embodiments, the one or more processors may confirm that a subsequent invalid subset of the DCA data set has fewer invalid candidate pause episodes than a previous invalid subset. Further recommendations for treatment may be iteratively made based on the comparison results.

[0247] Additionally or alternatively, treatment and / or information may include observations and / or recommendations about the thresholds used in the process of FIG. 6 to allow clinicians to adjust the level of sensitivity and / or specificity. For example, the option of adjusting the detection threshold may be presented to the clinician. In response to this, the process of FIG. 6 may repeat at least the decision and subset classification at operations 608 to 614 to "reclassify" a portion of the invalid subset. For example, when a clinician is notified that 50% of the DCA data sets were initially identified as invalid, the clinician may choose to adjust the threshold (e.g., from 0.6 to 0.75). Using the adjusted threshold, the process may reclassify the invalid subset, such as notifying the clinician that only 20% of the DCA data sets are now determined to be invalid. Additionally or alternatively, a digital count of the number of DCA data sets may be provided instead of a percentage (e.g., the valid list includes 75 DCA data sets, while the invalid list includes 13).

[0248] In some embodiments, the detection threshold may be automatically adjusted (eg, within a predetermined range) to reduce the number of DCA data sets identified as invalid (eg, false positives), and the process may automatically reclassify invalid subsets.

[0249] Additionally or alternatively, the treatment and / or information may include observations or recommendations regarding potential adjustments that may be made to the parameters of the IMD to reduce the number of false positives. For example, the CNN model 400 may output an indication that the IMD may be performing inappropriate sensing (such as oversensing or undersensing P waves, R waves, T waves, etc.). As a further indication, the CNN model 400 may identify the accuracy of R-wave sensing implemented by the IMD. Additionally or alternatively, the CNN model 400 may output recommendations on how to reprogram the parameters of the IMD (such as how to reprogram the sensitivity levels / thresholds used in conjunction with identifying P waves, R waves, and T waves).

[0250] In some embodiments, sensing parameters of the IMD can be automatically adjusted (such as within a predetermined range) to modify the sensing of one or more of the P wave, R wave, and T wave to reduce the number of false positives. One or more processors can evaluate the new DCA data set to determine whether the automatically adjusted sensing parameters reduce the number of false positives.

[0251] Additionally or alternatively, one or more processors may evaluate metrics (such as true timeouts (e.g., true positives), false timeouts (e.g., false positives), true negatives, false negatives, uncertain / indeterminate, confidence scores, and / or other metrics discussed herein). The one or more processors may use these metrics to further edit the diagnosis, clear alarms (e.g., on the IMD, external instrument), and manage other presentations, such as the urgency / importance of the event based on whether the patient has been previously diagnosed and on optimal medical treatment without planned further intervention, or whether this is a diagnosis and treatment initiation based on additional status input from the provider. Initiation or adjustment of medications, treatments, patient actions, etc. may be recommended. Programmable changes to the settings of the IMD may be recommended, automatically implemented after acceptance by the practitioner, adjusted by the practitioner, etc.

[0252] Optionally, this can be repeated for different architectures of the CNN model Figure 5 and Fig. 6A For example, various aspects of the CNN model can be varied, such as the number of convolutional layers, hyperparameters, different activation functions, etc.

[0253] Figure 5 and Fig. 6A The operation can be performed at regular intervals or as needed. Clinicians can choose to train CNN models using their own DCA and DCNS datasets (which are enhanced as described herein), so that the training data can be specific to clinicians, individual clinics, medical facilities, medical networks, or others. Additionally or alternatively, clinicians can choose to train CNN models using a larger patient population of DCA and DCNS datasets, such as by utilizing datasets collected by "trusted" third parties, or a combination thereof.

[0254] Optionally, the performance between trained CNN models or CNN models based on different enhanced DCA datasets and / or different types / versions of models can be compared. For example, a clinician can apply the same enhanced DCA dataset to a new version / type of CNN model that was previously used with a previous CNN model. The clinician can then be allowed to choose whether to use the newer version / type of CNN model based on whether the performance is better than the previous model.

[0255] Although combined analysis of the DCA dataset associated with IMD-specified pause episodes describes Fig. 6A Related to the preceding examples, it should be recognized that Fig. 6A The operations may be further applied to DCNS datasets collected by the IMD and associated with heart beats specified by the IMD as exhibiting normal sinus rhythm. For example, one or more DCNS datasets may be analyzed separately or in combination with a DCA dataset.

[0256] In the foregoing examples, the processor and memory are described as being housed within an implanted device such as an ICM and / or IMD. Additionally or alternatively, the processor and memory may be housed within at least one of a local external device and a remote server.

[0257] Figure 6B A process for determining and / or implementing treatment for a patient using a CNN model according to an embodiment of the present invention is shown. Figure 6B The operations of may be implemented by hardware, firmware, circuitry, and / or one or more processors partially and / or completely housed within an IMD, a local external device, a remote server, a combination thereof, and / or more generally within a healthcare system. Optionally, Figure 6B The operations of may be implemented in part by the IMD and in part by a local external device, a remote server, or more generally within the healthcare system. For example, the IMD includes an IMD memory and one or more IMD processors, while each external device / system (ED) (e.g., local, remote, or anywhere within the healthcare system) includes an ED memory and one or more ED processors. It should be appreciated that although described in somewhat serial fashion, Figure 6B However, one or more of the operations may be performed sequentially and / or in parallel with each other. For example, various operations of IMD 100, 212 may be performed sequentially and / or in parallel with each other and / or in parallel with other functions of IMD 100, 212 and / or external devices.

[0258] Mainly described from the perspective of accommodating CNN models within IMD 100, 212 Figure 6B However, the CNN model may be housed in a local external device and / or in a device / system remote from the patient.

[0259] At 652, one or more processors obtain one or more DCA data sets generated by the IMD that have been declared by the IMD as including candidate pause episodes. The DCA data sets may also be referred to as "candidate" DCA data sets because the CNN model 400 has not yet verified that the IMD has determined that a pause is actually occurring. Each DCA data set may correspond to a separate candidate pause episode. Additionally or alternatively, a subset of the DCA data sets may correspond to related candidate pause episodes (such as when a patient is experiencing multiple related pause episodes occurring in succession). The DCA data sets may be generated by the IMD for a period of time before being applied to the CNN model, or may be applied to the CNN model once announced by the IMD (e.g., in real time or near real time). As described herein, each of the DCA data sets corresponds to one or more declared pause episodes, may be of varying lengths of time and include a series of cardiac beats, include or exclude DD markers, etc.

[0260] Similar to Fig. 6A 606, at 654, one or more processors apply (one or more) DCA datasets to the CNN model 400, and in response thereto, obtain a confidence indicator, a valid / invalid indicator, etc. from the CNN model 400, which indicates the confidence level or degree of true positive or false positive designation that the current DCA dataset represents a pause. In addition, for example, the CNN model 400 can output at least one of the following in conjunction with each DCA dataset: i) a confidence indicator indicating the confidence level of the true positive or false positive designation that the corresponding DCA dataset represents a pause; ii) a confidence indicator indicating the accuracy of pause sensing implemented by the IMD; or iii) a recommendation indicating the sensitivity level of the IMD for identifying R waves in CA signals. In some embodiments, confidence indicator levels can be set for different arrhythmia types. For example, the confidence indicator level can be set in a first setting when indicating a first type of pause, and set in a second setting when indicating a second type of pause.

[0261] Similar to Fig. 6A 608, at 656, one or more processors compare the confidence indicator(s) provided by the CNN model 400 to one or more detection thresholds. In the aforementioned example, the confidence indicator may correspond to a numerical value close to 1 indicating a high confidence for a true positive, a numerical value close to 0 indicating a high confidence for a false positive, and any value therebetween. In some embodiments, when the confidence indicator exceeds the threshold, the process moves to 658, and the one or more processors add the DCA dataset to a valid dataset list to begin building a valid subset of the DCA dataset. When the confidence indicator does not exceed the threshold, the process moves to 660, where the one or more processors add the DCA dataset to an invalid dataset list to begin building an invalid subset of the DCA dataset. In some cases, valid subsets and invalid subsets may not be collected.

[0262] The method may return to 652 to continue evaluating the acquired / available one or more DCA data sets.

[0263] At 662, one or more processors provide treatment to the patient. In some embodiments, one or more processors may adjust or fine-tune the detection threshold within predetermined limits based on the output of the CNN model to increase or decrease sensitivity as discussed herein. For example, the detection threshold may be adjusted to keep the percentage of DCA data sets indicating a valid within predetermined limits. This provides the advantage of preventing overtreatment or undertreatment of the patient. In other embodiments, one or more processors may adjust (such as within predetermined limits) the sensitivity level used when sensing P waves, R waves, and / or T waves based on the output of the CNN model. This provides the advantage of preventing oversensing or undersensing of P waves, R waves, and / or T waves and reducing the number of false positives. In further embodiments, one or more processors may adjust the treatment settings within predetermined limits based on the output of the CNN model. For example, the sensing parameters of the IMD may be adjusted, pacing may be adjusted or temporarily suspended, and the like.

[0264] 662 may be repeated as the CNN model outputs data over time. A record of any adjustments, and optionally the DCA data set, valid subset, and / or invalid subset may be retained in memory by the IMD and / or external device until they are sent to a healthcare provider for review. Treatment pauses provide the advantage of quickly addressing a patient's medical issues and / or changes. By setting predetermined limits, there is no need to wait for a clinician to review and reprogram the patient's IMD. In some embodiments, the IMD and / or external device may send a message, such as to a remote system, to indicate that a setting has been adjusted. In other embodiments, a message may be sent if an adjustment beyond a predetermined limit appears warranted.

[0265] Figure 7 A distributed processing system 700 is shown according to embodiments herein. The distributed processing system 700 includes a server 702 connected to a database 704, a programmer 706, a local monitoring device 708 (e.g., a cell phone, tablet, computer, other medical monitoring device), and a user workstation 710 electrically connected to a network 712. Figure 7 Any processor-based component in (eg, IMD 703, workstation 710, cellular phone 714, local monitoring device 716, server 702, programmer 706) may perform the processes discussed herein.

[0266] The network 712 can provide cloud-based services, voice transmission over IP (VoIP) gateways, local plain old telephone service (POTS), public switched telephone network (PSTN), cellular phone-based networks, etc. through the Internet. Alternatively, the communication system can be a local area network (LAN), a medical campus area network (CAN), a metropolitan area network (MAN), or a wide area network (WAM). The communication system is used to provide a network that facilitates the transmission / reception of data and other information between local and remote devices (relative to patients). The server 702 is a computer system that provides services to other computing devices on the network 712. The server 702 can control the communication of information, (such as CNN models, DCA data sets, synthetic DCA data sets, CA signals, motion data, pause episode information, bradycardia episode information, arrhythmia episode information, markers, heart rate, and device settings). The server 702 is connected to the network 712 to transmit information between the programmer 706, local monitoring devices 708, 716, user workstations 710, cellular phones 714, and databases 704. For example, the server 702 can receive DCA datasets from various clinics, medical networks, individual patients, etc., and use the DCA datasets and synthetic DCA datasets to train new CNN models, update existing versions of CNN models, personalize CNN models, apply transfer learning to CNN modules, and add further outputs to existing and new CNN models. The server 702 can also push new CNN models, updated versions of CNN models, and / or personalized CNN models to various other devices, such as Figure 7 The database 704 stores information such as DCA data sets, synthesized DCA data sets, pause episode information, arrhythmia episode information, pause statistics, arrhythmia statistics, diagnosis, DD markers, CA signals, heart rate, device settings, etc. for patient populations and individually for individual patients, individual physicians, individual clinics, individual medical networks, etc. The server 702 can be implemented in conjunction with Figure 5 The CNN training operations described and / or the use of CNN models to analyze DCA datasets, such as in combination with Fig. 6A and Figure 6BAs described herein. Database 704 also maintains trained, updated and personalized CNN models as described herein. CNN models and other information are downloaded to database 704 via server 702, or alternatively, information is uploaded to server 702 from database 704. Programmer 706 can reside in the patient's home, hospital, or doctor's office. Programmer 706 can communicate wirelessly with ICM or IMD 703, and utilize protocols such as Bluetooth, GSM, infrared wireless LAN, HIPERLAN, 3G, satellite, and circuit and packet data protocols, etc. Alternatively, a telemetry "stick" connection can be used to connect programmer 706 to IMD 703. Programmer 706 can acquire cardiac signals (e.g., ECG) from surface ECG electrodes 722 on a person, electrogram (e.g., EGM) signals from IMD 703, and / or CA data, pause episode information, arrhythmia episode information, pause statistics, arrhythmia statistics, diagnosis, markers, CA signal waveforms, device settings from IMD 703. Programmer 706 connects to network 712 , such as via the Internet, to upload information acquired from surface ECG unit 720 or IMD 703 to server 702 .

[0267] The local monitoring device 708 interfaces with the communication system to upload one or more of DCA datasets, synthetic DCA datasets, CA signals, motion data, pause episode information, arrhythmia episode information, pause statistics, arrhythmia statistics, diagnostics, markers, heart rate, sensitivity profile parameter settings, and detection thresholds to the server 702. In one embodiment, the surface ECG unit 720 and the IMD 703 have a bidirectional connection 724 with the local RF monitoring device 708 via a wireless connection. The local monitoring device 708 is capable of acquiring surface ECG signals from the ECG leads 722, as well as DCA / CA datasets and other information from the IMD 703. On the other hand, the local monitoring device 708 can download data and information discussed herein (such as CNN models) from the database 704 to the IMD 703.

[0268] A physician or medical staff member may utilize a user workstation 710, a cell phone 714, and / or a programmer 706 to connect to the network 712 to download CNN models, DCA datasets, synthesized DCA datasets, CA signals, motion data, and other information discussed herein from a database 704, from a local monitoring device 708, 716, from an IMD 703, or otherwise. Once downloaded, the user workstation 710 may process the DCA datasets, CA signals, and motion data in accordance with one or more of the operations described above. The user workstation 710, the cell phone 714, and / or the programmer 706 may be used to present information about at least one of a valid subset or an invalid subset of the DCA dataset to the physician via display, printing, etc., present information about suggested adjustments to device settings, a history of automatically implemented adjustments to device settings, etc. Additionally or alternatively, the user workstation 710, cellular telephone 714, and / or programmer 706 may be used to display, in conjunction with each DCA data set, at least one of: i) a confidence indicator indicating a true positive or false positive designation of confidence that the corresponding DCA data set represents a pause; ii) a confidence indicator indicating the accuracy of pause sensing implemented by the IMD; or iii) a recommendation indicating a sensitivity level for the IMD to use in identifying R waves in CA signals.

[0269] The user workstation 710 , cell phone 714 , and / or programmer 706 may upload / push settings (eg, sensitivity profile parameter settings), IMD instructions, other information, and notifications to the cell phone 714 , local monitoring devices 708 , 716 , programmer 706 , server 702 , and / or IMD 703 .

[0270] The process described herein in conjunction with reducing false positive pause events and / or identifying and implementing treatment can be accomplished by Figure 7 The processes described herein may be distributed over one or more of the devices shown, including but not limited to IMD 703, programmer 706, local monitoring devices 708, 716, user workstation 710, cell phone 714, and server 702. Figure 7 For example, Figure 7 One or more of the devices shown in the may include a memory for storing specific executable instructions and one or more CNN models including a GAP layer trained to detect pause episodes; and one or more processors configured to execute specific executable instructions to: obtain a device classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) for corresponding candidate pause episodes declared by the IMD, the DCA dataset including one or more beats of cardiac activity (CA) signals sensed by the IMD; and apply the CNN model to the DCA dataset to identify a valid subset of the DCA dataset that correctly characterizes the corresponding CA signals. Figure 7The illustrated one or more devices may also include a display configured to present information regarding the valid subset of the DCA data set.

[0271] Additionally or alternatively, wherein the one or more processors are further configured to identify an invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signal, wherein the valid subset of the DCA dataset indicates a portion of the DCA dataset indicating a true pause, and the invalid subset of the DCA dataset indicates a second portion of the DCA dataset indicating a false pause. Additionally or alternatively, the information indicates a recommendation for: i) a change in treatment of the patient associated with the IMD that would increase the DCA dataset identified as the valid subset of the DCA dataset, or ii) a change in treatment associated with the IMD that would increase the DCA dataset identified as the valid subset of the DCA dataset. Additionally or alternatively, wherein the one or more processors are further configured to apply the CNN model to the DCA dataset to identify an invalid subset of the DCA dataset that incorrectly characterizes corresponding CA signals, wherein, in response to a change in treatment for the patient or the IMD, the one or more processors are further configured to obtain a second DCA dataset generated by the IMD, apply the CNN model to the second DCA dataset to identify a second invalid subset of the DCA dataset that incorrectly characterizes corresponding CA signals, and confirm that the second invalid subset of the DCA dataset has fewer invalid candidate pause episodes than the invalid subset of the DCA dataset.

[0272] Additionally or alternatively, the CNN model includes at least two one-dimensional (1D) convolutional layers, and the GAP layer is configured to receive outputs from each of the at least two 1D convolutional layers. Additionally or alternatively, the CNN model includes five 1D convolutional layers. Additionally or alternatively, the at least two 1D convolutional layers automatically capture different levels of features associated with the pause from input data associated with the DCA dataset.

[0273] Additionally or alternatively, one or more processors may also be configured to process the output of the first 1D convolutional layer using i) a rectified linear unit activation function, ii) a batch normalization function, or iii) a discard function. Additionally or alternatively, the CNN model includes at least a first 1D convolutional layer and a second 1D convolutional layer, and the one or more processors may also be configured to: normalize the output of the first 1D convolutional layer and send the normalized output of the first 1D convolutional layer to the second 1D convolutional layer. Additionally or alternatively, the CNN model also includes a fully connected layer configured to receive input from a GAP layer. Additionally or alternatively, the CNN model outputs a confidence indicator in conjunction with each of a plurality of DCA datasets, the confidence indicator indicating the degree of confidence specified by the corresponding DCA dataset indicating a paused true positive or false positive. Additionally or alternatively, the one or more processors are also configured to compare the confidence indicator with a detection threshold, wherein the information also includes information about a valid subset of the DCA dataset that exceeds the detection threshold. Additionally or alternatively, in response to the adjustment of the detection threshold, the one or more processors are further configured to present information about a valid subset of the DCA data set that exceeds the adjusted detection threshold. Additionally or alternatively, the CNN model also includes a softmax layer configured to output a confidence indicator. Additionally or alternatively, the CNN model represents a model trained using an enhanced set of DCA data sets, wherein the enhanced set of DCA data sets includes a reference DCA data set from a patient and a synthetic DCA data set generated based on the reference DCA data set. Additionally or alternatively, wherein the synthetic DCA data set is generated using: i) noise addition, ii) signal inversion, iii) amplification, iv) inserting one or more p-waves at a previous RR interval during a pause interval, v) stretching or contracting the duration of the pause interval, or vi) two or more of noise addition, signal inversion, amplification, inserting one or more p-waves at a previous RR interval during a pause interval, or stretching or contracting the duration of the pause interval. Additionally or alternatively, wherein the information indicates a recommendation for adjusting a sensing parameter or a treatment parameter associated with an IMD. Additionally or alternatively, the system further comprises an IMD comprising a combination of subcutaneous electrodes configured to collect CA signals, an IMD memory configured to store program instructions, and one or more IMD processors configured to execute the program instructions to analyze the CA signals and declare a candidate pause episode based on the analysis and generate a DCA data set including the corresponding CA signals; and the system further comprises a transceiver configured to wirelessly transmit the DCA data set to an external device. Additionally or alternatively, Figure 7 The system also includes an external device including a memory and one or more processors and a transceiver configured to wirelessly receive the DCA data set from the IMD. Additionally or alternatively, Figure 7The system also includes a server comprising a memory and one or more processors, the memory being configured to store a set of DCA datasets, the one or more processors being configured to apply the CNN model to the set of DCA datasets.

[0274] Figure 7 The system also includes: under the control of one or more processors configured with specific executable instructions, a method for building a convolutional neural network (CNN) model to confirm a device-recorded (DD) pause episode, comprising: obtaining a set of reference device-classified arrhythmia (DCA) data sets associated with the device-announced pause episode, the reference DCA data set comprising one or more beats of cardiac activity (CA) signals sensed by electrodes of an implantable medical device (IMD); generating a synthetic DCA data set based on the reference DCA data set to form an enhanced set of DCA data sets, wherein the generating includes processing the reference DCA data set using i) noise addition or ii) signal inversion; and applying the DCA data set and the enhanced set of DCA data sets to the CNN model to train the CNN model. In other embodiments, the generating may include amplification, inserting one or more p-waves at a previous RR interval during the pause interval, and / or stretching or contracting the duration of the pause interval, and / or other enhancement processes discussed herein and / or known in the art.

[0275] Additionally or alternatively, the one or more processors of the system are further configured to apply a first enhanced set of DCA datasets representing valid DCA datasets that correctly characterize the corresponding CA signals, and apply a second enhanced set of DCA datasets representing invalid DCA datasets that incorrectly characterize the corresponding CA signals. Additionally or alternatively, the one or more processors of the system are further configured to automatically capture different levels of features associated with the pause from input data associated with the DCA dataset and the enhanced set of DCA datasets using at least two convolutional layers. Additionally or alternatively, the one or more processors of the system are further configured to receive outputs from the at least two 1D convolutional layers using a global average pooling (GAP) layer. Additionally or alternatively, the one or more processors of the system are further configured to fine-tune the CNN model by inputting at least one of the following into a layer of the CNN model: i) age-related features, ii) gender, iii) body mass index, iv) race, v) medical history, vi) medication use, vii) lifestyle factors, or viii) family history. Additionally or alternatively, one or more processors of the system are further configured to freeze at least one layer of the CNN model and update at least one unfrozen layer of the CNN model using the additional DCA dataset.

[0276] According to the new and unique aspects of this document, IMD 100, 212, 703, cellular phone 714, local monitoring device 708, programmer 706, server 702 and / or (one or more) other processors and / or processing systems deliver specific treatment for a paused medical condition (e.g., pause episode) based on the output of the CNN model. The pause event detected by the IMD is further evaluated by the CNN model, and the true pause event is separated from the false pause event. The treatment can be based on an indicator threshold, based on how many percentages of potential pause events are determined to be true pauses, for recommending a setting change of the IMD, for recommending the action and / or behavior of the patient and / or clinician, etc. For example, the clinician uses true pauses and / or false pauses and any other output of the CNN model 400 to prescribe / change the patient's treatment (e.g., prescribe a new drug, change a drug, change a diet, modify a behavior, recommend implanting another and / or different IMD, change the programming parameters of an already implanted IMD). In another example, the IMD can automatically implement changes to the programming parameters of the IMD based on the output of the CNN model. Thus, the methods and systems herein provide practical applications for treating medical conditions in which patients pause.

[0277] Figure 8 A system-level diagram indicating potential devices and networks utilizing the methods and systems herein is shown. For example, IMD 802 can be used to collect DCA data sets. IMD 802 can provide DCA data sets (CA signals, DD markers) as well as sensitivity levels and optional motion data to various local external devices (such as tablet devices 804, smart phones 806, bedside monitoring devices 808, smart watches, etc.). Devices 804-808 include displays to present various types of CA signals, DD markers, statistics (e.g., valid percentages or numbers, invalid percentages or numbers), diagnoses, recommendations for adjustments to IMD sensing / treatment parameters, and other information described herein, and / or summaries or lists of automatic adjustments to IMD sensing / treatment parameters. IMD 802 can transmit DCA data sets to devices 804, 806, and 808 via various types of wireless communication links. IMD 802 can utilize various communication protocols and be activated in various ways, such as via Bluetooth, Bluetooth low energy, Wi-Fi, or other wireless protocols. Additionally or alternatively, when magnetic device 810 is held next to a patient, a magnetic field from device 810 may activate IMD 802 to transmit the DCA data set and arrhythmia data to one or more of devices 804 - 808 .

[0278] According to new and unique aspects of the present invention, methods and systems are described for training and utilizing a machine learning model (such as a CNN model) to determine whether a candidate arrhythmia (e.g., pause) declared and classified by an IMD from a subcutaneous EGM (SEGM) is a true positive or a false positive. Initial performance evaluations indicate that the CNN model can reduce the false positive burden associated with pauses declared by an IMD by improving sensitivity and specificity and improving positive predictive value metrics.

[0279] Embodiments may be implemented using all or part of the methods and systems described in U.S. patent application 2022 / 0117538, entitled “Methods and systems to confirm device classified arrhythmias utilizing machine learning models,” filed on June 8, 2021, and U.S. patent application 2022 / 0304612, entitled “Methods and systems for predicting arrhythmia risk utilizing machine learning models,” filed on December 17, 2023, the entire contents of which are incorporated herein by reference.

[0280] Embodiments may be implemented in conjunction with one or more implantable medical devices (IMDs). Non-limiting examples of IMDs include one or more of an implantable leadless monitoring and / or therapeutic device and / or an alternative implantable medical device. For example, an IMD may represent a cardiac monitoring device, a pacemaker, a cardioverter, a cardiac rhythm management device, a defibrillator, a leadless monitoring device, a leadless pacemaker, and the like. Additionally or alternatively, an IMD may be a leadless implantable medical device (LIMD) that includes one or more structural and / or functional aspects of the devices described in U.S. Patent 9,216,285 “LEADLESS IMPLANTABLE MEDICAL DEVICE HAVING REMOVABLE AND FIXED COMPONENTS” and U.S. Patent 8,831,747 “LEADLESS NEUROSTIMULATION DEVICE AND METHOD INCLUDING THE SAME”, the entire contents of which are incorporated herein by reference. Additionally or alternatively, the IMD may include one or more structural and / or functional aspects of the device(s) described in U.S. Patent 8,391,980, “METHOD AND SYSTEM FOR IDENTIFYING A POTENTIAL LEAD FAILURE IN AN IMPLANTABLE MEDICAL DEVICE” and U.S. Patent 9,232,485, “SYSTEM AND METHOD FOR SELECTIVELY COMMUNICATING WITH ANIMPLANTABLE MEDICAL DEVICE,” the entire contents of which are incorporated herein by reference.Additionally or alternatively, the IMD may be a subcutaneous IMD that includes one or more structural and / or functional aspects of the device(s) described in the following patents: U.S. Patent No. 10,765,860, filed on May 7, 2018, entitled “SUBCUTANEOUS IMPLANTATION MEDICAL DEVICE WITH MULTIPLE PARASTERAL-ANTORITY ELECTRODES”; U.S. Patent No. 10,722,704, filed on May 7, 2018, entitled “IMPLANTABLE MEDICAL SYSTEMS AND METHODS INCLUDING PULSE GENERATORS AND LEADS”; U.S. Patent No. 11,045,643, filed on May 7, 2018, entitled “SINGLE SITE IMPLANTATION METHODS FOR MEDICAL DEVICES HAVING MULTIPLE LEADS”, the entire contents of which are incorporated herein by reference.Additionally or alternatively, the IMD may be a leadless cardiac monitor (ICM) that includes one or more structural and / or functional aspects of the devices described in the following patents: U.S. Patent 9,949,660, entitled “METHOD AND SYSTEM TO DISCRIMINATE RHYTHM PATTERNS IN CARDIAC ACTIVITY,” issued April 24, 2018; U.S. Patent 10,729,346, entitled “METHOD AND SYSTEM FOR SECOND PASSCONFIRMATION OF DETECTED CARDIAC ARRHYTHMIC PATTERNS”; U.S. Patent 11,020,036, entitled “METHOD AND SYSTEM TO DETECT R-WAVES IN CARDIAC ARRHYTHMIC PATTERNS”; U.S. Patent 11,020,036, entitled “METHOD AND SYSTEM TO DETECT PREMATURE VENTRICULAR CONTRACTIONS IN CARDIAC ACTIVITY SIGNALS"; U.S. Patent 10,874,322 entitled "ADJUSTABLE ANTENNA SYSTEM TO COMMUNICATE WITH AN IMPLANTABLE MEDICAL DEVICE AND METHODFOR USING SAME"; and U.S. Published Application 2021 / 0020294, filed on July 16, 2020, entitled "METHODS, DEVICES AND SYSTEMS FOR HOLISTIC INTEGRATED HEALTHCARE PATIENT MANAGEMENT," the entire contents of which are incorporated herein by reference. In addition, according to the embodiments of this document, one or more combinations of IMDs from the above-incorporated patents and applications can be utilized.

[0281] All references, including publications, patent applications, and patents, cited herein are hereby incorporated by reference to the same extent as if each reference were individually and specifically indicated to be incorporated by reference and were set forth in its entirety herein.

[0282] Conclusion

[0283] The various methods shown in the figures and described herein represent exemplary embodiments of the methods. These methods can be implemented in software, hardware, or a combination thereof. In the various methods, the order of the steps can be changed, and various elements can be added, reordered, combined, omitted, modified, etc. The various steps can be performed automatically (e.g., without direct prompting by user input) and / or programmatically (e.g., according to program instructions).

[0284] It will be apparent to those skilled in the art having the benefit of this disclosure that various modifications and variations may be made. It is intended to cover all such modifications and variations, and therefore, the above description should be regarded as illustrative rather than restrictive.

[0285] Various embodiments of the present disclosure utilize at least one network familiar to those skilled in the art to support communications using any of a variety of commercially available protocols, such as Transmission Control Protocol / Internet Protocol ("TCP / IP"), User Datagram Protocol ("UDP"), protocols operating in various layers of the Open Systems Interconnection ("OSI") model, File Transfer Protocol ("FTP"), Universal Plug and Play ("UPnP"), Network File System ("NFS"), Common Internet File System ("CIFS"), and AppleTalk. The network can be, for example, a local area network, a wide area network, a virtual private network, the Internet, an intranet, an extranet, a public switched telephone network, an infrared network, a wireless network, a satellite network, and any combination thereof.

[0286] In an embodiment utilizing a web server, the web server may run any of a variety of server or middle-tier applications, including a hypertext transfer protocol ("HTTP") server, an FTP server, a common gateway interface ("CGI") server, a data server, a Java server, an Apache server, and a business application server. The server(s) may also be capable of executing a program or script in response to a request from a user device, such as by executing one or more network applications, which may be implemented as one or more scripts or programs written in any programming language (e.g., C, C#, or C++) or any scripting language (e.g., Ruby, PHP, Perl, Python, or TCL) and combinations thereof. The server(s) may also include a database server, including but not limited to those commercially available from Oracle®, Microsoft®, Sybase®, SAS®, and IBM®, as well as open source servers such as MySQL, Postgres, SQLite, MongoDB, and any other server capable of storing, retrieving, and accessing structured or unstructured data. The database server may include a table-based server, a document-based server, an unstructured server, a relational server, a non-relational server, or a combination of these and / or other database servers.

[0287] The environment may include various data repositories and other memories and storage media, as described above. These may reside in various locations, such as on any or all of the storage media in one or more local (and / or resident in one or more of the computers) or remote computers on the network. In a particular set of embodiments, the information may reside in a storage area network ("SAN") familiar to those skilled in the art. Similarly, any necessary files for performing functions attributed to a computer, server, or other network device may be appropriately stored locally and / or remotely. In the case where the system includes computerized devices, each such device may include hardware elements that may be electrically coupled via a bus, including, for example, at least one central processing unit ("CPU" or "processor"), at least one input device (e.g., a mouse, keyboard, controller, touch screen, or keypad), and at least one output device (e.g., a display device, printer, or speaker). Such a system may also include one or more storage devices, such as disk drives, optical storage devices, and solid-state storage devices, such as random access memory ("RAM") or read-only memory ("ROM"), as well as removable media devices, memory cards, flash memory cards, and the like.

[0288] Such a device may also include a computer-readable storage medium reader, a communication device (e.g., a modem, a network card (wireless or wired), an infrared communication device, etc.), and a working memory as described above. The computer-readable storage medium reader may be connected to or configured to receive a computer-readable storage medium, which represents a remote, local, fixed and / or removable storage device and a storage medium for temporarily and / or more permanently containing, storing, sending and retrieving computer-readable information. The system and various devices will also typically include a plurality of software applications, modules, services or other elements located within at least one working memory device, including an operating system and an application program, such as a client application or a web browser. It should be understood that alternative embodiments may have many variations different from the above-described embodiments. For example, custom hardware may also be used and / or specific elements may be implemented in hardware, software (including portable software, such as applets), or both. In addition, connections to other computing devices such as network input / output devices may be employed.

[0289] Various embodiments may also include receiving, sending or storing instructions and / or data implemented according to the foregoing description on a computer-readable medium. Storage media and computer-readable media for containing code or portions of code may include any suitable media known or used in the art, including storage media and communication media, such as but not limited to volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing and / or sending information (such as computer-readable instructions, data structures, program modules or other data), including RAM, ROM, electrically erasable programmable read-only memory ("EEPROM"), flash memory or other memory technology, compact disk read-only memory ("CD-ROM"), digital versatile disk (DVD) or other optical storage device, magnetic cassette, magnetic tape, magnetic disk storage device or other magnetic storage device or any other medium that can be used to store the desired information and can be accessed by system devices. Based on the disclosure and teachings provided herein, a person of ordinary skill in the art will understand other ways and / or methods of implementing various embodiments.

[0290] The units / modules / applications herein may include any processor-based or microprocessor-based systems, including systems using microcontrollers, reduced instruction set computers (RISC), application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), logic circuits, and any other circuits or processors capable of performing the functions described herein. Additionally or alternatively, the modules / controllers herein may represent circuit modules that may be implemented as hardware with relevant instructions for performing the operations described herein (e.g., software stored on tangible and non-transitory computer-readable storage media (such as computer hard drives, ROMs, RAMs, etc.)). The above examples are exemplary only, and are therefore not intended to limit the definition and / or meaning of the term "controller" in any way. The units / modules / applications herein may execute a set of instructions stored in one or more storage elements in order to process data. The storage elements may also store data or other information as desired or required. The storage element may be in the form of an information source or physical memory element within the module / controller herein. The set of instructions may include various commands that instruct the modules / applications herein to perform specific operations (such as the methods and processes of the various embodiments of the subject matter described herein). The set of instructions may be in the form of a software program. The software may be in various forms, such as system software or application software. Furthermore, the software may be in the form of a collection of separate programs or modules, a program module within a larger program, or a portion of a program module. The software may also include modular programming in the form of object-oriented programming. The processing of input data by a processing machine may be in response to user commands, or in response to the results of previous processing, or in response to a request made by another processing machine.

[0291] Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense. It will, however, be evident that various modifications and changes may be made thereto without departing from the broader spirit and scope of the invention as set forth in the claims.

[0292] Other variations are within the scope of the present disclosure. Thus, while the disclosed technology is susceptible to various modifications and alternative constructions, certain illustrated embodiments thereof are shown in the drawings and have been described in detail above. However, it should be understood that it is not intended to limit the invention to the specific form or forms disclosed, but on the contrary, it is intended to cover all modifications, alternative constructions and equivalents that fall within the spirit and scope of the invention as defined in the appended claims.

[0293] Unless otherwise indicated herein or clearly contradicted by context, the use of the terms "a" and "an" and "the" and similar referents in the context of describing the disclosed embodiments (particularly in the context of the appended claims) should be interpreted as covering both the singular and the plural. Unless otherwise indicated, the terms "comprising", "having", "including" and "containing" should be interpreted as open-ended terms (i.e., meaning "including but not limited to") unless otherwise indicated. The term "connected" when unmodified and referring to a physical connection should be interpreted as partially or completely contained therein, attached to or joined together, even if there is something in between. Unless otherwise indicated herein, the recitation of numerical ranges herein is intended to serve only as a shorthand method of referring individually to each individual value falling within the range, and each individual value is incorporated into the specification as if it were individually recited herein. Unless otherwise indicated or contradicted by context, the use of the term "set" (e.g., "set of items") or "subset" should be interpreted as a non-empty set that includes one or more members. Furthermore, unless otherwise stated or contradicted by context, the term "subset" of a corresponding set does not necessarily mean a proper subset of the corresponding set, but rather a subset and a corresponding set may be equal.

[0294] Unless otherwise specified herein or clearly contradicted by context, the operations of the processes described herein may be performed in any suitable order. The processes described herein (or variations and / or combinations thereof) may be performed under the control of one or more computer systems configured with executable instructions, and may be implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) that is jointly executed by hardware or a combination thereof on one or more processors. The code may be stored on a computer-readable storage medium, for example, in the form of a computer program that includes multiple instructions that may be executed by one or more processors. The computer-readable storage medium may be non-transitory.

[0295] All references, including publications, patent applications, and patents, cited herein are hereby incorporated by reference to the same extent as if each reference were individually and specifically indicated to be incorporated by reference and were set forth in its entirety herein.

[0296] It should be understood that the subject matter described herein is not limited in its application to the details of the construction and arrangement of the components set forth in the description herein or shown in the accompanying drawings. The subject matter described herein can have other embodiments and can be practiced or executed in various ways. In addition, it should be understood that the wording and terminology used herein are for descriptive purposes and should not be considered as limiting. The use of "including", "comprising" or "having" and variations thereof herein is intended to cover the items listed thereafter and their equivalents and additional items.

[0297] It should be understood that the above description is intended to be illustrative and not limiting. For example, the above embodiments (and / or aspects thereof) may be used in combination with each other. In addition, many modifications may be made to adapt specific situations or materials to the teachings of the present invention without departing from the scope of the present invention. Although the dimensions, material types, and physical properties described herein are intended to define the parameters of the present invention, they are by no means limiting, but exemplary embodiments. Many other embodiments will be apparent to those skilled in the art after reading the above description. Therefore, the scope of the present invention should be determined with reference to the appended claims and the full range of equivalents to which these claims are entitled. In the appended claims, the terms "including" and "in which" are used as the plain English equivalents of the corresponding terms "comprising" and "wherein". In addition, in the appended claims, the terms "first", "second", and "third", etc., are used only as labels and are not intended to impose numerical requirements on their objects. Furthermore, the appended claim limitations are not written in means-plus-function format and are not intended to be interpreted under 35 USC 112(f) unless and until such claim limitations expressly use the phrase "means for..." followed by a recitation of functionality without further structure.

Claims

1. A system for announcing cardiac arrest, comprising: A memory for storing specific executable instructions and a convolutional neural network (CNN) model trained to detect pause episodes, the CNN model including a global average pooling (GAP) layer; one or more processors configured to execute the specific executable instructions to: obtaining a device classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) for corresponding candidate pause episodes declared by the IMD, the DCA dataset comprising one or more beat-by-beat cardiac activity (CA) signals sensed by the IMD; as well as applying the CNN model to the DCA dataset to identify a valid subset of the DCA dataset that correctly characterizes corresponding CA signals; as well as A display is configured to present information about the valid subset of the DCA data set.

2. The system according to claim 1, wherein: The one or more processors are further configured to apply the CNN model to the DCA dataset to identify an invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signal, wherein the valid subset of the DCA dataset indicates a portion of the DCA dataset indicating a true pause and the invalid subset of the DCA dataset indicates a second portion of the DCA dataset indicating a false pause.

3. The system according to claim 1, wherein: The information indicates a recommendation for: i) a change in treatment of a patient associated with the IMD that would increase the DCA dataset identified as the valid subset of the DCA dataset, or ii) a change in treatment associated with the IMD that would increase the DCA dataset identified as the valid subset of the DCA dataset.

4. The system according to claim 3, wherein: The one or more processors are further configured to apply the CNN model to the DCA dataset to identify invalid subsets of the DCA dataset that incorrectly characterize corresponding CA signals, wherein, in response to a change in therapy for the patient or the IMD, the one or more processors are further configured to: obtaining a second DCA data set generated by the IMD; applying the CNN model to the second DCA dataset to identify a second invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signals; and It is determined that the second invalid subset of the DCA dataset has fewer invalid candidate pause episodes than the invalid subset of the DCA dataset.

5. The system according to claim 1, wherein: The CNN model comprises at least two one-dimensional (1D) convolutional layers, the GAP layer being configured to receive outputs from each of the at least two 1D convolutional layers, wherein the one or more processors are further configured to process the output of the first 1D convolutional layer using i) a rectified linear unit activation function, ii) a batch normalization function, or iii) a dropout function.

6. The system according to claim 1, wherein: The CNN model comprises at least a first 1D convolutional layer and a second 1D convolutional layer, wherein the one or more processors are further configured to: Normalizing the output of the first 1D convolutional layer; and The normalized output of the first 1D convolutional layer is sent to the second 1D convolutional layer.

7. The system according to claim 1, wherein: The CNN model also includes a fully connected layer configured to receive input from the GAP layer.

8. The system according to claim 1, wherein: The CNN model outputs a confidence indicator in conjunction with each of the plurality of DCA datasets, the confidence indicator indicating a degree of confidence that the corresponding DCA dataset represents a true positive or false positive designation of a pause.

9. The system according to claim 8, wherein: The one or more processors are further configured to compare the confidence indicator to a detection threshold, wherein the information further includes information about the valid subset of the DCA data set that exceeds the detection threshold.

10. The system of claim 9, wherein in response to adjusting the detection threshold, the one or more processors are further configured to present information about the valid subset of the DCA data set that exceeds the adjusted detection threshold.

11. The system according to claim 1, wherein: The CNN model includes five one-dimensional (1D) convolutional layers, and the GAP layer is configured to receive outputs from each of at least two 1D convolutional layers.

12. The system of claim 1, wherein: The CNN model represents a model trained using an enhanced set of DCA datasets, wherein the enhanced set of DCA datasets includes a reference DCA dataset from a patient and a synthetic DCA dataset generated based on the reference DCA dataset.

13. The system according to claim 12, wherein: The synthetic DCA dataset is generated using: i) noise addition, ii) signal inversion, iii) amplification, iv) insertion of one or more p-waves at the previous RR interval during the pause interval, v) stretching or contracting the duration of the pause interval, or vi) two or more of noise addition, signal inversion, amplification, insertion of one or more p-waves at the previous RR interval during the pause interval, or stretching or contracting the duration of the pause interval.

14. The system of claim 1, wherein: The information indicates a recommendation for adjusting a sensing parameter or a therapy parameter associated with the IMD.

15. The system of claim 1, further comprising the IMD, the IMD comprising: a combination of subcutaneous electrodes, wherein the combination of subcutaneous electrodes is configured to collect the CA signal; an IMD memory configured to store program instructions; as well as one or more IMD processors configured to execute the program instructions to: analyzing the CA signal and declaring a candidate pause episode based on the analysis; as well as generating the DCA dataset comprising the corresponding CA signals; as well as A transceiver is configured to wirelessly transmit the DCA data set to an external device.

16. The system of claim 1, further comprising an external device comprising the memory and the one or more processors and a transceiver configured to wirelessly receive the DCA data set from the IMD.

17. The system of claim 1, further comprising a server comprising the memory and the one or more processors, the memory being configured to store the set of DCA datasets, the one or more processors being configured to apply the CNN model to the set of DCA datasets.

18. A computer-implemented method for confirming a device-recorded (DD) pause episode, comprising: Under the control of one or more processors configured with specific executable instructions, obtaining a device classified arrhythmia (DCA) dataset generated by an implantable medical device (IMD) for corresponding candidate pause episodes declared by the IMD, the DCA dataset comprising one or more beat-by-beat cardiac activity (CA) signals sensed by the IMD; applying a convolutional neural network (CNN) model trained to detect pause bouts to the DCA dataset to identify a valid subset of the DCA dataset that correctly characterizes corresponding CA signals; as well as Information about the valid subset of the DCA data set is presented.

19. The method of claim 18, further comprising applying the CNN model to the DCA dataset to identify invalid subsets of the DCA dataset that incorrectly characterize the corresponding CA signals, wherein The valid subset of the DCA data set indicates a portion of the DCA data set indicating a true timeout, and the invalid subset of the DCA data set indicates a second portion of the DCA data set indicating a false timeout.

20. The method according to claim 18, wherein: The information indicates a recommendation for: i) a change in treatment of a patient associated with the IMD that would increase the DCA dataset identified as the valid subset of the DCA dataset, or ii) a change in treatment associated with the IMD that would increase the DCA dataset identified as the valid subset of the DCA dataset.

21. The method according to claim 20, further comprising: applying the CNN model to the DCA dataset to identify invalid subsets of the DCA dataset that incorrectly characterize the corresponding CA signals; as well as In response to a change in therapy for the patient or the IMD, the method further comprises: obtaining a second DCA data set generated by the IMD; applying the CNN model to the second DCA dataset to identify a second valid subset of the DCA dataset that correctly characterizes the corresponding CA signals, and to identify a second invalid subset of the DCA dataset that incorrectly characterizes the corresponding CA signals; and It is determined that the second invalid subset of the DCA dataset has fewer invalid candidate pause episodes than the invalid subset of the DCA dataset.

22. The method of claim 18, further comprising adjusting the CNN model by inputting into a layer of the CNN model at least one of: i) age-related characteristics, ii) gender, iii) body mass index, iv) race, v) medical history, vi) medication use, vii) lifestyle factors, or viii) family history.

23. The method according to claim 18, wherein: The CNN model includes at least two layers, and the method further includes: freezing one of the at least two layers; and At least one unfrozen layer of the CNN model is updated using the additional DCA dataset.

Images