Non-invasive cardiac monitor and method for using recorded cardiac data to infer physiological characteristics of a patient - Patents.com

Abstract

To provide a non-invasive heart monitoring device which records fibrillation electrocardiographic data for inferring a physiological feature of a person such as cardiac arrhythmia.SOLUTION: A heart monitor device comprises: a water-tight housing; a surface on the housing configured to be sealably engaged to a mammal; an adhesive on the surface configured to remain continuously affixed to the mammal; at least two electrodes permanently disposed within the housing, positioned to detect continuous heartbeat signals of the mammal while the surface is sealably engaged to the mammal; a hardware processor configured to process the detected heartbeat signals with a subset of computer-executable instructions; and a transmitter. The device can process the detected heartbeat signals in a remote computing system. Some embodiments include a wearable heart monitor device for detecting the heartbeat signals for a long period of time by being attached to the mammal for a long period of time.SELECTED DRAWING: Figure 19

Description

REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Provisional Patent Application No. 62 / 975,626, filed February 12, 2020, and U.S. Provisional Patent Application No. 63 / 090,951, filed October 13, 2020, which are incorporated by reference in their entireties. There is proper legal basis, including, but not limited to, 35 U.S.C. § 119(e), for the filing date of the aforementioned application to be the effective filing date of this application. [Background technology]

[0002] For purposes of this disclosure, certain aspects, advantages, and novel attributes of various embodiments are described herein. It is to be understood that not necessarily all such advantages may be achieved in accordance with any particular embodiment. Thus, various embodiments may be implemented to achieve one or more advantages taught herein, but may not necessarily achieve other advantages that may be taught or suggested herein. Summary of the Invention

[0003] The embodiments described herein are directed to physiological monitoring devices that may be worn continuously and comfortably by a human or animal subject for at least one week or more, and more typically for two to three weeks or more. In one embodiment, the device is specifically designed to sense and record cardiac rhythm (e.g., electrocardiogram, ECG) data, although in various alternative embodiments, one or more additional physiological parameters may be sensed and recorded. Such physiological monitoring devices may include a number of features to facilitate and / or enhance the patient experience and to make the diagnosis of cardiac arrhythmias more accurate and timely.

[0004] Some embodiments include a wearable cardiac monitor device for monitoring cardiac rhythm signal data of a user, the wearable cardiac monitor device comprising: A watertight housing; a surface on the housing configured to sealably engage the mammal; and an adhesive on the surface configured to remain continuously affixed to the mammal for at least seven days without removal until monitoring is completed; at least two electrodes permanently disposed within the housing, the electrodes being arranged to detect continuous cardiac signals of the mammal while the surface is sealably engaged to the mammal; a hardware processor configured to process the detected heart rate signal according to a first subset of computer-executable instructions; A transmitter, The transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to process the output with a second subset of computer-executable instructions to infer a likelihood of occurrence of the cardiac arrhythmia.

[0005] In some embodiments, the arrhythmia comprises at least one of ventricular tachycardia, supraventricular tachycardia, ecstopy, ventricular fibrillation, or extended pauses.

[0006] In some embodiments, inferring the likelihood of occurrence of cardiac arrhythmia includes processing the output of a first subset through a second subset, the first subset processing cardiac signals continuously detected and stored for at least 24 hours.

[0007] In some embodiments, the housing includes a patient trigger configured to be depressed to initiate recording of a temporal instance of a sensed cardiac event.

[0008] In some embodiments, the electrodes are disposed entirely within the housing.

[0009] In some embodiments, the first subset of computer-executable instructions includes a first subset of layers of a neural network, and the second subset of computer-executable instructions includes a second subset of layers of the neural network.

[0010] Some embodiments include a wearable cardiac monitor device for monitoring vital sign data of a user, the wearable device comprising: an adhesive assembly including a housing and wings, the wings including electrodes configured to detect a cardiac signal from a user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of a neural network; a transmitter configured to transmit the first subset of outputs to a computing system, the computing system configured to infer a likelihood of occurrence of a cardiac arrhythmia by processing the outputs through a second subset of layers of a neural network; Includes.

[0011] In some embodiments, the computing system is further configured to determine an atrial fibrillation burden from the detected cardiac signal.

[0012] In some embodiments, atrial fibrillation burden includes the amount of time a user spends in atrial fibrillation during a period of time.

[0013] In some embodiments, atrial fibrillation burden includes the amount of time the user spends in atrial fibrillation during sleep and waking periods.

[0014] Some embodiments include a wearable device for monitoring biosignal data of a user, the wearable device comprising: an assembly including electrodes configured to detect cardiac signals from a user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of the neural network; A transmitter, The transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to process the output through a second subset of the layers of the neural network to infer a likelihood of occurrence of a cardiac arrhythmia.

[0015] In some embodiments, the hardware processor is configured to process the detected cardiac signals through a first subset of layers of the neural network continuously in substantially real-time to detect the ECG signals.

[0016] In some embodiments, the output of the first subset of layers of the neural network includes an indication of R peak.

[0017] In some embodiments, the computing system is further configured to reconstruct the cardiac signal via a second subset of layers.

[0018] In some embodiments, the outputs of the first subset of layers of the neural network have less dimensionality than the inputs of the first subset of layers of the neural network.

[0019] In some embodiments, the hardware processor is configured to select a particular neural network from the plurality of neural networks based on a characteristic.

[0020] In some embodiments, the characteristics include at least one of: a characteristic of the wearable device; a characteristic of a network between the wearable device and the computing system.

[0021] In some embodiments, the hardware processor is further configured to compress outputs of a first subset of layers of the neural network, and transmitting the outputs of the first subset to the computing system includes transmitting the compressed outputs, and the computing system is configured to decompress the transmitted data, and processing the outputs through a second subset of layers includes processing the decompressed data.

[0022] In some embodiments, the hardware processor is further configured to quantize outputs of a first subset of the layers of the neural network, and transmitting the outputs of the first subset to the computing system includes transmitting the outputs.

[0023] In some embodiments, the hardware processor is further configured to compress the digitized output, and transmitting the output of the first subset to the computing system includes transmitting the compressed output, and the computing system is configured to decompress the transmitted data, and processing the output via the second subset of layers includes processing the decompressed data.

[0024] In some embodiments, transmitting the first subset of outputs to the computing system includes transmitting digitized outputs.

[0025] Some embodiments include a monitor that infers a likelihood of a characteristic of a user; The monitor is Housing and at least two electrodes permanently disposed within the housing and positioned to detect continuous signals on the surface; a printed circuit board assembly including a hardware processor configured to process the detected signals via a first subset of computer executable instructions; A transmitter, The transmitter is configured to transmit a data output of a first subset of computer-executable instructions to a computer system, and the computer system is configured to process the data output via a second subset of computer-executable instructions to infer a likelihood of an occurrence.

[0026] In some embodiments, the dimensionality of the data output of the first subset of computer-executable instructions sent to the computing system is less than the data of the detected signals from the at least two electrodes.

[0027] In some embodiments, the data output of the first subset of computer-executable instructions is encrypted, the transmitter transmits the encrypted data output, and the computing system processes the encrypted data output via the second subset of computer-executable instructions.

[0028] In some embodiments, the transmitter transmits the data output to a computing system over a wireless communications channel.

[0029] In some embodiments, the monitor further comprises a receiver configured to receive the updated first subset of computer-executable instructions from the computing system and update the first subset of computer-executable instructions to the updated first subset of computer-executable instructions, and the hardware processor is further configured to process a signal via the updated first subset of computer-executable instructions.

[0030] In some embodiments, the monitor further includes an accelerometer configured to collect motion data, and the computing system is configured to match the motion data with the detected signal to infer a likelihood of the occurrence.

[0031] In some embodiments, the monitor is a cardiac monitor, the continuous signal is an electrocardiogram signal, and the occurrence event is an arrhythmic event.

[0032] In some embodiments, the first subset of computer-executable instructions includes a first subset of layers of a neural network, and the second subset of computer-executable instructions includes a second subset of layers of the neural network.

[0033] Some embodiments include a monitor that infers the likelihood of a user characteristic; The monitor is A watertight housing; a surface on the housing configured to sealably engage the mammal; and an adhesive on the surface configured to remain continuously affixed to the mammal without removal until monitoring is completed for at least seven days; at least two electrodes permanently disposed within the housing and positioned to detect continuous cardiac signals of the mammal while the surface is sealably engaged with the mammal; a printed circuit board assembly including a hardware processor configured to process the detected heart rate signal via a first subset of computer executable instructions; A transmitter, The transmitter is configured to transmit a data output of the first subset of computer-executable instructions to a computing system; The computing system is configured via a second subset of computer-executable instructions to process the data output to infer a likelihood of an arrhythmia occurring.

[0034] In some embodiments, the arrhythmia comprises at least one of ventricular tachycardia, supraventricular tachycardia, ectopic excitation, ventricular fibrillation, or prolonged pauses.

[0035] In some embodiments, inferring the likelihood of occurrence of cardiac arrhythmia includes processing the output of a first subset through a second subset, the first subset processing cardiac signals continuously detected and stored for at least 24 hours.

[0036] In some embodiments, the housing includes a patient trigger configured to be depressed to initiate recording of a temporal instance of a sensed cardiac event.

[0037] Some embodiments include a system for training a neural network to infer likelihoods of user features, The system includes a wearable device, Wearable devices include: an assembly including electrodes configured to detect cardiac signals from a user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of the neural network; A transmitter, The transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to process the output through a second subset of the layers of the neural network to infer a likelihood of occurrence of a cardiac arrhythmia; The system is training a first neural network to identify a first feature by processing first training data for a first time period through the first neural network; freezing weights of the first neural network; training a second neural network to identify a second characteristic by processing second training data for a second time period, the second time period being longer than the first time period, via the first neural network and the second neural network; unfreezing the weights of the first neural network; training the first and second neural networks simultaneously by processing third training data of a third time period longer than the first time period through the first and second neural networks to identify a second feature; The neural network is then trained according to the method described above.

[0038] In some embodiments, the outputs of the first subset of layers of the neural network are of a smaller dimensionality than the inputs of the first subset of layers of the neural network.

[0039] In some embodiments, the hardware processor is configured to select a particular neural network from the plurality of neural networks based on the characteristics.

[0040] In some embodiments, the characteristics include at least one of a characteristic of the wearable device, a characteristic of a network between the wearable device and the computing system.

[0041] In some embodiments, the hardware processor is further configured to compress outputs of a first subset of layers of the neural network, wherein transmitting the outputs of the first subset to the computing system includes transmitting the compressed outputs, and the computing system is configured to decompress the transmitted data, and wherein processing the outputs through the second subset of layers includes processing the decompressed data.

[0042] In some embodiments, the hardware processor is further configured to quantify outputs of a first subset of the layers of the neural network, and transmitting the outputs of the first subset to the computing system includes transmitting the outputs.

[0043] Some embodiments include a method of training a neural network to infer likelihoods of user features, The method is: training a first neural network to identify a first characteristic by processing first training data for a first time period through the first neural network; freezing weights of the first neural network; training a second neural network to identify a second characteristic by processing second training data of a second time period, the second time period being longer than the first time period, through the first and second neural networks; Unfreezing the weights of the first neural network; and simultaneously training the first and second neural networks to identify a second characteristic by processing third training data of a third time period, the third time period being longer than the first time period, via the first and second neural networks.

[0044] In some embodiments, the second training data is third training data and the second time period is a third time period.

[0045] In some embodiments, the training is based on the available processing power of the wearable device.

[0046] In some embodiments, the training is based on the available memory of the wearable device.

[0047] In some embodiments, the training is based on network availability between the wearable device and an external computing system.

[0048] In some embodiments, the training is based on a type of characteristic of the user.

[0049] In some embodiments, the type of user characteristic includes the occurrence of cardiac arrhythmia.

[0050] In certain embodiments, the computing system is further configured to provide a report, the report including the likelihood of occurrence of cardiac arrhythmia.

[0051] In certain embodiments, the report includes a graph of atrial fibrillation burden over time.

[0052] In certain embodiments, the report includes an indication of the presence of atrial fibrillation.

[0053] In certain embodiments, the report includes a monitoring period of at least 14 days.

[0054] In certain embodiments, the computing system is a server or a gateway.

[0055] In a particular embodiment, the computing system is a smartphone.

[0056] In a particular embodiment, the computing system communicates with the transmitter through the intermediary of a smartphone.

[0057] In certain embodiments, the wearable device is configured to collect secondary signals, and the hardware processor is configured to process the secondary signals through a first subset of layers of the neural network.

[0058] In certain embodiments, the wearable device further includes an accelerometer configured to measure movement of a user, and the secondary signal is accelerometer data.

[0059] In certain embodiments, the computing system is further configured to determine an atrial fibrillation burden, the atrial fibrillation burden including an amount of time spent in an atrial fibrillation state during the user's activity.

[0060] In certain embodiments, the user's movements include a first degree of movement and a second degree of movement.

[0061] In certain embodiments, the secondary signal is electrode contact integrity data.

[0062] In certain embodiments, the electrodes and hardware processor are housed within a chest strap.

[0063] In a particular embodiment, the electrodes and hardware processor are housed in a watch adapted to be worn on a human wrist.

[0064] In certain embodiments, the electrodes and hardware processor are housed within a wearable fitness band.

[0065] In certain embodiments, the arrhythmia comprises ventricular tachycardia.

[0066] In certain embodiments, the arrhythmia comprises supraventricular tachycardia.

[0067] In certain embodiments, the arrhythmia comprises ectopic excitation.

[0068] In certain embodiments, the arrhythmia comprises ventricular fibrillation.

[0069] In certain embodiments, the arrhythmia comprises a prolonged pause.

[0070] In certain embodiments, the hardware processor is configured to be detached from the wearable device and modified while detached from the wearable device.

[0071] In certain embodiments, quantifying the output includes rounding or truncating the value of the output.

[0072] In certain embodiments, the amount of digitization for the output is based on lossless compression performance.

[0073] In certain embodiments, the amount of quantification for the output is based on the efficiency of at least one of processing power, storage, or network usage, and the accuracy of the neural network.

[0074] In certain embodiments, the neural network is trained by applying a quantification to the outputs of a first subset of the layers of the neural network.

[0075] In a particular embodiment, the neural network is trained by applying lossless compression to the outputs of a first subset of the layers of the neural network.

[0076] In a particular embodiment, the neural network is trained by applying quantification and lossless compression to the outputs of a first subset of the layers of the neural network.

[0077] In certain embodiments, the neural network is trained without quantification and lossless compression to generate a first neural network, and the first neural network is trained applying quantification and lossless compression to outputs of a first subset of the layers of the neural network to generate a second neural network.

[0078] Some embodiments include a wearable device for monitoring the data, the wearable device comprising: a sensor configured to detect an electrocardiogram signal from a user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of the neural network; A transmitter, the transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to process the output through a second subset of the layers of the neural network to infer a likelihood of occurrence of a cardiac arrhythmia.

[0079] Some embodiments include a computing system for estimating a likelihood of occurrence of a cardiac arrhythmia, The computing system is one or more first hardware processors configured to receive data from a wearable device including a sensor configured to detect electrocardiogram signals from a user; and one or more second hardware processors configured to process the detected cardiac signals through a first subset of the layers of the neural network, process the received data through a second subset of the layers of the neural network, and receive from an output of the second subset of the layers an inference of a likelihood of an occurrence of a cardiac arrhythmia.

[0080] Some embodiments include a method of monitoring data, comprising: The method is: detecting an electrocardiogram signal from a user; processing the detected cardiac signals through a first subset of layers of a neural network; transmitting the first subset of outputs to a computing system; A computing system processes the output through a second subset of the layers of the neural network to infer the likelihood of cardiac arrhythmia occurrence.

[0081] Some embodiments include a non-transitory computer storage medium containing computer-executable instructions that, when executed by a processor, cause the processor to: detecting an electrocardiogram signal from a user; processing the detected cardiac signals through a first subset of layers of a neural network; transmitting the first subset of outputs to a computing system; The computing system processes the output through a second subset of layers of the neural network to infer the likelihood of cardiac arrhythmia occurrence.

[0082] Some embodiments include a method of training a neural network to infer likelihoods of user features, The method is: training a first neural network to identify a first characteristic by processing first training data for a first time period through the first neural network; freezing weights of the first neural network; training a second neural network to identify a second characteristic by processing second training data of a second time period, the second time period being longer than the first time period, through the first and second neural networks; unlocking the weights of the first neural network; and simultaneously training the first and second neural networks to identify a second characteristic by processing third training data of a third time period, the third time period being longer than the first time period, via the first and second neural networks.

[0083] In a particular embodiment, the second training data is third training data and the second time period is a third time period.

[0084] In certain embodiments, the training is based on the available processing power of the wearable device.

[0085] In certain embodiments, the training is based on the available memory of the wearable device.

[0086] In certain embodiments, the training is based on network availability between the wearable device and an external computing system.

[0087] In certain embodiments, the training is based on a type of characteristic of the user.

[0088] In certain embodiments, the type of user characteristic includes the occurrence of cardiac arrhythmia.

[0089] Some embodiments include a system for training a neural network to infer likelihoods of features of a user, the system including a wearable cardiac monitor device; Wearable heart monitor devices include: A watertight housing; a surface on the housing configured to sealably engage the mammal; and an adhesive on the surface configured to remain continuously affixed to the mammal without removal until monitoring is completed for at least seven days; at least two electrodes permanently disposed within the housing positioned to detect continuous cardiac signals of the mammal while the surface is sealably engaged to the mammal; a hardware processor configured to process the detected heart rate signals through a first subset of layers of a neural network; A transmitter, The transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to infer a likelihood of occurrence of a cardiac arrhythmia by processing the output through a second subset of the layers of the neural network; The system is training a first neural network to identify a first characteristic by processing first training data for a first time period through the first neural network; freezing weights of the first neural network; training a second neural network to identify a second characteristic by processing second training data for a second time period, the second time period being longer than the first time period, via the first neural network and the second neural network; Unfreezing the weights of the first neural network; concurrently training the first and second neural networks to identify a second characteristic by processing third training data for a third time period, the third time period being longer than the first time period, through the first and second neural networks; The neural network is configured to train the neural network by

[0090] Some embodiments include a wearable device for monitoring biosignal data of a user, Wearable devices include: an adhesive assembly including a housing and wings, the wings including electrodes configured to detect electrocardiac signals from a user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of the neural network; A transmitter, The transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to process the output through a second subset of the layers of the neural network to infer a QT interval.

[0091] In certain embodiments, the computing system is further configured to determine an atrial fibrillation burden from the detected cardiac signal.

[0092] In certain embodiments, atrial fibrillation burden includes the amount of time a user spends in atrial fibrillation during a period of time.

[0093] In certain embodiments, atrial fibrillation burden includes the amount of time the user spends in atrial fibrillation during sleep and waking periods.

[0094] In certain embodiments, the hardware processor is configured to process the detected electrocardiogram signal through a first subset of layers of the neural network continuously in substantially real time with the detection of the ECG signal.

[0095] In certain embodiments, the output of the first subset of layers of the neural network includes an indication of the R peak.

[0096] In certain embodiments, the computing system is further configured to reconstruct the cardiac signal via a second subset of the layers.

[0097] Some embodiments include a system for training a neural network to infer likelihoods of features of a user, the system including a wearable device; Wearable devices include: an assembly including electrodes configured to detect cardiac signals from a user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of the neural network; A transmitter, The transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to process the output through a second subset of the layers of the neural network to infer a likelihood of occurrence of a cardiac arrhythmia; The system is training a first neural network to identify a first characteristic by processing first training data for a first time period through the first neural network; freezing weights of the first neural network; training a second neural network to identify a second characteristic by processing second training data for a second time period, the second time period being longer than the first time period, via the first neural network and the second neural network; Unfreezing the weights of the first neural network; concurrently training the first and second neural networks to identify a second characteristic by processing third training data for a third time period, the third time period being longer than the first time period, through the first and second neural networks; The neural network is configured to train the neural network by

[0098] In certain embodiments, the outputs of the first subset of layers of the neural network are of a smaller dimensionality than the inputs of the first subset of layers of the neural network.

[0099] In a particular embodiment, the hardware processor is configured to select a particular neural network from the plurality of neural networks based on the characteristics.

[0100] In certain embodiments, the characteristics include at least one of a characteristic of the wearable device, a characteristic of the network between the wearable device and the computing system.

[0101] In a particular embodiment, the hardware processor is further configured to compress outputs of a first subset of layers of the neural network, and transmitting the outputs of the first subset to the computing system includes transmitting the compressed outputs, and the computing system is configured to decompress the transmitted data, and processing the outputs through a second subset of layers includes processing the decompressed data.

[0102] In certain embodiments, the hardware processor is further configured to quantify outputs of a first subset of the layers of the neural network, and transmitting the outputs of the first subset to the computing system includes transmitting the outputs.

[0103] In certain embodiments, the computing system infers the QT interval by reconstructing the ECG signal from the output of the second subset of layers of the neural network.

[0104] In certain embodiments, the computing system infers the QT interval based on the average QT interval from a window of the encoded trait that includes multiple QT intervals.

[0105] Some embodiments include a wearable device for monitoring biosignal data of a user, Wearable devices include: an adhesive assembly including a housing and wings, the wings including electrodes configured to detect electrocardiac signals from a user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of the neural network; A transmitter, The transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to generate a template beat based on a number of beats over a period of time.

[0106] Some embodiments include a wearable device for monitoring biosignal data of a user, Wearable devices include: an assembly including electrodes configured to detect cardiac signals from a user; a hardware processor configured to process the detected cardiac signal through a first subset of layers of a neural network, wherein an output of the first subset of layers of the neural network comprises an R-R peak interval, the R-R peak interval comprising an elapsed time between two successive R-peaks; A transmitter, The transmitter is configured to transmit the first subset of outputs to a computing system; The computing system is configured to infer a likelihood of occurrence of a cardiac arrhythmia by processing the RR peak intervals at the output of the first subset of the layers of the neural network through a second subset of the layers of the neural network.

[0107] In a particular embodiment, a first subset of layers of the neural network extracts RR interval subsequences using overlapping sliding windows and generates a sequence of RR peak intervals by shifting the sliding windows over the detected cardiac signal.

[0108] These and other aspects and embodiments of the invention are described in more detail below with reference to the drawings. [Brief description of the drawings]

[0109] [Figure 1A] FIG. 1A is a perspective view of a physiological monitoring device according to one embodiment. [Figure 1B]FIG. 1B is a degradation profile diagram of a physiological monitoring device according to one embodiment. [Figure 2A] FIG. 2A is a top perspective view of a printed circuit board assembly of a physiological monitoring device according to one embodiment. [Figure 2B] FIG. 2B is a bottom perspective view of a printed circuit board assembly of a physiological monitoring device according to one embodiment. [Figure 3A] FIG. 3A is a perspective view of a flexible body and gasket of a physiological monitoring device according to one embodiment. [Figure 3B] FIG. 3B is an exploded view of a flexible body and gasket of a physiological monitoring device according to one embodiment. [Figure 3C] FIG. 3C is an illustration of a flexible body and gasket of a physiological monitoring device according to one embodiment. [Figure 3D] FIG. 3D is an illustration of a flexible body and gasket of a physiological monitoring device according to one embodiment. [Figure 3E] FIG. 3E is a perspective view of a flexible body and gasket of a physiological monitoring device according to one embodiment. [Figure 4] FIG. 4 is an exploded view of a rigid housing of a physiological monitoring device according to one embodiment. [Figure 5A] FIG. 5A is a perspective view of a battery holder of a physiological monitoring device according to one embodiment. [Figure 5B] FIG. 5B is a perspective view of a battery holder of a physiological monitoring device according to one embodiment. [Figure 6A] FIG. 6A is a cross-sectional view of a physiological monitoring device according to one embodiment. [Figure 6B] FIG. 6B is a cross-sectional view of a physiological monitoring device according to one embodiment. [Figure 7] FIG. 7 is an exploded view of a physiological monitoring device that includes many options, according to one embodiment. [Figure 8A]FIG. 8A is a perspective view of a person wearing a physiological monitoring device according to one embodiment, showing how the device flexes with body movement and position. [Figure 8B] FIG. 8B is a perspective view of a person wearing a physiological monitoring device according to one embodiment, illustrating how the device flexes with body movement and position. [Figure 9A] FIG. 9A illustrates a procedure for attaching a physiological monitoring device to a patient's body according to one embodiment. [Figure 9B] FIG. 9B illustrates a procedure for attaching a physiological monitoring device to a patient's body according to an embodiment. [Figure 9C] FIG. 9C illustrates a procedure for attaching a physiological monitoring device to a patient's body according to an embodiment. [Figure 9D] FIG. 9D illustrates a procedure for attaching a physiological monitoring device to a patient's body according to one embodiment. [Figure 9E] FIG. 9E illustrates a procedure for attaching a physiological monitoring device to a patient's body according to one embodiment. [Figure 9F] FIG. 9F illustrates a procedure for attaching a physiological monitoring device to a patient's body according to one embodiment. [Figure 10] FIG. 10 is a schematic diagram of one embodiment of the heartbeat inference service. [Figure 11] FIG. 11 is a schematic diagram illustrating an embodiment of a system for extracting and transmitting data attributes from a physiological monitor. [Figure 12] FIG. 12 is a schematic diagram illustrating an embodiment of a system for extracting and transmitting data characteristics from a physiological monitor using a transmitting device. [Figure 13] FIG. 13 is a schematic diagram illustrating an embodiment of a physiological monitoring system utilizing an additional data channel. [Figure 14] FIG. 14 is a schematic diagram illustrating an embodiment of a physiological monitoring system incorporating a data filter. [Figure 15] FIG. 15 is a schematic diagram illustrating an embodiment of a wearable device system. [Figure 16] FIG. 16 is a schematic diagram illustrating an embodiment of a symptom transmission system. [Figure 17] FIG. 17 is a schematic diagram illustrating an embodiment of an asymptomatic transmission system. [Figure 18A] FIG. 18A is a graph of an embodiment illustrating ECG strip readings corresponding to NSR heart beats. [Figure 18B] FIG. 18B is a graph of an embodiment illustrating ECG strip readings corresponding to NSR heart beats. [Figure 19] Figure 19 shows the high-level architecture of the burden forecasting model. [Figure 20] FIG. 20 is a high-level architectural embodiment of FIG. [Figure 21] FIG. 21 is an embodiment of the feature extraction model of FIG. [Figure 22A] FIG. 22A is a schematic diagram of an embodiment of a system for predicting heart beat annotations using neural network encoding. [Figure 22B] FIG. 22B is a schematic diagram of an embodiment of a first and second subset of layers in a single neural network. [Figure 22C] FIG. 22C is a schematic diagram of an embodiment of a system for processing a subset of layers of a neural network on a patch. [Figure 22D] FIG. 22D is a schematic diagram of an embodiment of sending the output of a first subset of layers to a server. [Figure 22E] FIG. 22E is a schematic diagram of an embodiment for designing (and / or training) a neural network. [Figure 22F] FIG. 22F is a schematic diagram of an embodiment for designing, training, and / or selecting a neural network based on a user's current behavior. [Figure 23] FIG. 23 is a schematic diagram of an embodiment of a computer network system. [Figure 24] FIG. 24 is a schematic diagram of an embodiment of a programming and distribution module. [Diagram 25] FIG. 25 shows a lossless compression implementation with an ECG encoder. [Figure 26] FIG. 26 illustrates an embodiment of a digitizer that performs digitization on the ECG encoder output data after processing a first subset of the layers of a neural network. [Figure 27] FIG. 27 shows an embodiment with both a digitizer and lossless compression via a lossless compressor and lossless decompressor. [Figure 28A] FIG. 28A illustrates an embodiment of a process for generating a template beat. [Figure 28B] FIG. 28B illustrates an embodiment of a process for generating a template beat. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0110] The following description is directed to a number of different embodiments. However, the described embodiments can be implemented and / or varied in many different ways. For example, the described embodiments may be implemented in any suitable device, apparatus, or system for monitoring any of a number of physiological parameters. For example, the following description focuses primarily on a long-term patch-based heart rate monitoring device. In one alternative embodiment, the physiological monitoring device may be used, for example, for pulse oximetry and diagnosis of obstructive sleep apnea. The method of use of the physiological monitoring device also varies. In some cases, the device is worn for a week or less, and in other cases, the device may be worn for at least 7 days and / or for more than 7 days, for example, 14 to 21 days or more.

[0111] Abnormal heart beats and arrhythmias are often caused by other, less serious causes, making determining whether any of these symptoms are due to arrhythmias a major challenge. Often, arrhythmias occur infrequently and / or sporadically, making a rapid and reliable diagnosis difficult. Currently, cardiac monitoring is primarily performed by short-term (one day or less) electrodes worn on the chest, such as Holter monitors. The electrodes are connected by wires to a recording device that is usually worn on a belt. The electrodes must be replaced daily, and the wires are cumbersome. Recording devices also have limited memory and recording time. Wearing a Holter monitor also often restricts the patient's mobility, preventing them from performing certain activities, such as bathing. Furthermore, Holter monitors are capital equipment and have limited availability, which often leads to supply constraints and associated delays in testing. These constraints significantly impede the diagnostic utility of the device, the patient's compliance with using the device, and the possibility of capturing all the important information. Lack of compliance and shortcomings of the device often necessitate additional equipment, follow-on monitoring, and other testing to make the correct diagnosis.

[0112] Current methods of correlating arrhythmia occurrence with symptoms using cardiac monitoring devices such as Holter monitors and cardiac event monitors are often insufficient to make an accurate diagnosis. In fact, Holter monitors have been shown to be non-diagnostic up to 90% of the time (DE Ward et al., "Evaluation of the diagnostic value of 24-hour ambulatory electrocardiogram monitoring," Biotelemetry Patient Monitoring, Vol. 7, published in 1980).

[0113] Moreover, the medical act of actually obtaining a heart rate monitoring device and initiating monitoring is usually very complicated. There are usually numerous steps involved in ordering, tracking, monitoring, retrieving, and analyzing data from such monitoring devices. Heart monitoring devices in use today are almost always ordered by a cardiologist or cardiac electrophysiologist (EP) rather than the patient's primary care physician (PCP). This is important because the PCP is often the first to see the patient and is the first physician to determine that the patient's symptoms may be due to arrhythmia. After the patient sees the PCP, the PCP schedules an appointment for the patient to see the cardiologist or EP. This appointment is usually scheduled several weeks after the initial consultation, which increases the chance of a delayed diagnosis and an undiagnosed arrhythmia episode. When the patient finally sees the cardiologist or EP, the heart rate monitoring device is usually ordered. The monitoring period can be 24-48 hours (Holter monitor) or up to one month (cardiac event monitor or mobile telemetry device). Once monitoring is finished, patients typically must return the device to the clinic, which can be an inconvenience in itself. The data is processed by the monitoring company or a technician at the hospital or office, and finally a report is sent to the cardiologist or EP for analysis. This complex process means that fewer patients are receiving heart rate monitoring than would be desirable.

[0114] To address some of these problems with cardiac monitoring, the assignee of the present application has developed various embodiments of a small, long-term wearable physiological monitoring device. One embodiment of this device is the Zio® Patch. Various embodiments are also described, for example, in U.S. Patent Nos. 8,150,502, 8,160,682, 8,244,335, 8,560,046, and 8,538,503, the entire disclosures of which are incorporated herein by reference in their entirety. In general, the physiological patch-based monitors described in the above documents are designed to fit comfortably on the patient's chest and be worn for at least one week, typically two to three weeks. The monitor continuously detects and records heart rate signal data while the device is worn, and this heart rate data is then available for processing and analysis.

[0115] Physiological Monitoring Devices 1A and 1B, perspective and exploded profile views of one embodiment of a physiological monitoring device 100 are provided. As seen in FIG. 1A, the physiological monitoring device 100 may include a flexible body 110 coupled to a watertight, rigid housing 115, and a hinged portion 132. The flexible body 110 (sometimes referred to as a "flexible substrate" or "flexible structure") typically includes two wings 130, 131 extending laterally from the rigid housing 115, and two flexible electrode traces 311, 312, each embedded in one of the wings 130, 131. Each electrode trace 311, 312 is coupled to a flexible electrode (not visible in FIG. 1A) on a bottom surface of the flexible body 110. The electrodes are configured to sense cardiac signals from a patient to which the monitoring device 100 is attached. The electrode traces 311, 312 then transmit their signals to electronics (not visible in FIG. 1A) contained in the rigid housing 115. The rigid housing 115 also typically contains a power source, such as one or more batteries.

[0116] 1B, a partial exploded view of the physiological monitoring device 100 shows in more detail the components that make up and are housed within the rigid housing 115. In this embodiment, the rigid housing 115 includes an upper housing member 140 that detachably couples with a lower housing member 145. Sandwiched between the upper housing member 140 and the lower housing member 145 are an upper gasket 370 and a lower gasket 360 (not visible in FIG. 1B, but just below the upper gasket 370). The gaskets 370, 360 help make the rigid housing member 115 watertight when assembled. A number of components of the monitoring device 100 may be housed between the upper housing member 140 and the lower housing member 145. For example, in one embodiment, the housing 115 may house a portion of the flexible body 110, a printed circuit board assembly (PCBA) 120, a battery holder 150, and two batteries 160. The printed circuit board assembly 120 is disposed within the housing 115 so as to contact the electrode traces 311, 312 and the battery 160. In various embodiments, one or more additional components may be included within or attached to the rigid housing 115. Some of these optional components are further described below with reference to additional figures.

[0117] The battery holder 150 can hold two batteries (as in the illustrated embodiment), one battery, or more than two batteries, according to various alternative embodiments. In other alternative embodiments, other power sources may be used. In the illustrated embodiment, the battery holder 150 includes a number of retention tabs 153 for retaining the battery 160 within the holder 150. Additionally, the battery holder 150 includes a number of feet 152 for establishing the correct spacing of the battery 160 from the surface of the PCBA 120 and ensuring proper contact with the spring fingers 235 and 236. The spring fingers 235 and 236 are used in this embodiment instead of soldering the battery 160 to the PCBA 120. Although soldering may be used in alternative embodiments, one advantage of the spring fingers 235 and 236 is that the battery 160 can be removed from the PCBA 120 and the holder 150 without damaging any of those components, thus allowing multiple reuse of both. Eliminating the solder connection also simplifies and speeds up the assembly and disassembly of the monitoring device 100. In some embodiments, the upper housing member 140 can function as a patient event trigger.

[0118] 2A and 2B , the printed circuit board assembly 120 (or PCBA) may include a top surface 220, a bottom surface 230, a patient trigger input 210, and spring contacts 235, 236, and 237. The patient trigger input 210 may be configured to relay a signal from a patient trigger, such as the upper housing member 140 described above, to the PCBA 120. For example, the patient trigger input 210 may be a PCB switch or button that responds to pressure from a patient trigger (e.g., a top surface of the upper housing member 140).

[0119] 3A and 3B, flexible body 110 is shown in more detail. As shown in FIG. 3A, flexible body 110 may include wings 130, 131, a thin border 133 (or "rim", or "edge") around at least a portion of each wing 130, 131, electrode traces 311, 312, and a hinge portion 132 (or "shoulder") at or near the junction of each wing 130, 131 with rigid housing 115.

[0120] 3C and 3D, when device 100 or zoom-in device portion 101 is assembled, ECG circuit interface portion 313 makes physical contact with spring fingers 237 and provides electrical communication with PCBA 120. Electrode interface portion 310 contacts hydrogel electrode 350. Thus, electrode traces 311, 312 transmit cardiac signals (and / or other physiological data in various embodiments) from electrode 350 to PCBA 120.

[0121] 3E depicts yet another embodiment in which an upper gasket 370 includes tabs 371 that protrude away from the profile of the upper housing 140 while remaining affixed to the upper substrate 300. The tabs 371 cover portions of the electrode traces 311, 312 and provide strain relief for the traces at the points of highest stress where the flex meets the rigid housing.

[0122] 4, the upper and lower housing members 140, 145 of the rigid housing 115 are shown in greater detail. The upper and lower housing members 140, 145, when coupled together with gaskets 360, 370 therebetween, may be configured to form a watertight enclosure for containing the PCBA 120, battery holder 150, battery 160, and any other components contained within the rigid housing 115. The housing members 140, 145 may be made of any suitable material for protecting the internal components, such as a water resistant plastic.

[0123] 5A, the battery holder 150 is shown in more detail. The battery holder 150 may be made of plastic or other suitable material, is configured to be attached to the PCBA 120 and then to the rigid housing 115, and is capable of holding two batteries 160 (FIG. 1B). A plurality of protrusions 152 provide a stable platform for the batteries 160 to be positioned at a fixed distance from the surface of the PCBA 120, providing proper compression of the spring contacts 235 while avoiding unwanted contact with sensitive electronic components (FIG. 5B).

[0124] 6A and 6B, a physiological monitoring device 100 is shown in a side cross-sectional view. As shown in 6A, the physiological monitoring device 100 may include a flexible body 110 coupled to a rigid housing 115. The flexible body 110 may include a top substrate layer 300, a bottom substrate layer 330, an adhesive layer 340, and an electrode 350. Electrode traces 311, 312 are also typically part of the flexible body 110 and are embedded between the top substrate layer 300 and the bottom substrate layer 330, but are not shown in FIG. 6. The flexible body 110 forms two wings 130, 131 that extend on either side of the housing 115, and a boundary 133 that surrounds at least a portion of each wing 130, 131. The rigid housing 115 may include an upper housing member 140 coupled to a lower housing member 145 to sandwich a portion of the flexible body 110 therebetween, providing a watertight sealed compartment for the PCBA 120. The upper housing member 140 may include an inner trigger member 430, and the PCBA may include a patient trigger member 210. As previously described, the lower housing member 145 may include a number of dimples 450 or divots to enhance the comfort of the monitoring device 100. In certain embodiments, an additional mechanism may be used to reduce or prevent unwanted bending of the PCBA 120. This mechanism is shown in FIG. 6B.

[0125] 7, in some embodiments, physiological monitoring device 100 may include one or more additional optional features. For example, in one embodiment, monitoring device 100 may include a removable liner 810, an upper label 820, a device identifier 830, and a lower label 840.

[0126] 8A and 8B, the physiological monitoring device 100 generally includes a hinge portion 132 at or near the junction of each wing 130, 131 and the rigid housing 115. Additionally, while each wing 130, 131 is typically adhered to the patient via an adhesive layer 340, the rigid body 115 is not adhered to the patient and is therefore free to "float" (e.g., move up and down) on the patient's skin during patient movement and position changes. In other words, as the patient's chest contracts, the rigid housing pops up or floats on the skin, thereby minimizing stress on the device 100, increasing comfort, and reducing the tendency of the wings 130, 131 to delaminate from the skin.

[0127] 9A-9F, one embodiment of a method of applying a physiological monitoring device 100 to the skin of a human subject is described. In this embodiment, prior to the first step shown in FIG. 9A, the patient's skin may be prepared, typically by shaving a small portion of the skin on the left chest where the device 100 will be placed, and scrubbing and / or cleaning the shaved area. As shown in FIG. 9A, once the patient's skin has been prepared, the first step of applying the device 100 may include removing one or both of the two adhesive covers 600 from the adhesive layer 340 on the bottom surface of the device 100, thus exposing the adhesive layer 340. As shown in FIG. 9B, the next step may be to apply the device 100 to the skin such that the adhesive layer 340 adheres to the skin at a desired location. With reference to FIG. 9C, after the device 100 has been applied to the skin, pressure may be applied to the flexible body 110 to press it against the chest to help ensure adhesion of the device 100 to the skin.

[0128] In a next step, referring to FIG. 9D, the liner 810 is removed (e.g., peeled off) from the top surface of the flexible body 110. Once the liner 810 is removed, as shown in FIG. 9E, pressure may again be applied to help ensure that the flexible body 110 adheres to the skin. Finally, as shown in FIG. 9F, the upper housing member 140 may be pressed to turn on the physiological monitoring device 140. This described method is only one embodiment. In alternative embodiments, one or more steps may be omitted and / or one or more additional steps may be added.

[0129] The physiological monitors described above and elsewhere herein may further be combined with data processing and transmission methods and systems that improve data collection from the monitor. Additionally, the methods and systems described below may improve the performance of the monitors by allowing timely transmission of clinical information while maintaining high patient compliance and ease of use of the monitors described above. For example, the data processing and transmission methods and systems described in this section or elsewhere herein may help extend the battery life of the monitors, improve the accuracy of the monitors, and / or provide other improvements and advantages described in this section or elsewhere herein.

[0130] Device monitoring and clinical analysis platform The systems and methods described in detail below in conjunction with the embodiments of Figures 10-17 can selectively extract, transmit, and analyze electrocardiogram signal data and other physiological data from a wearable physiological monitor such as those described above in conjunction with Figures 1-9. The systems and methods described below can improve the performance of a wearable physiological monitor that simultaneously records and transmits data via multiple means. For example, selectively transmitting extracted data can reduce power consumption since the wearable patch does not need to transmit all recorded data. By transmitting extracted data, much of the analysis can be performed remotely from the wearable device without the need for comprehensive on-board heart rate analysis, which is highly power consuming and reduces battery life. Additionally, remote analysis without the power constraints inherent in wearable devices can allow for greater sensitivity and accuracy in the analysis of data. The reduced power consumption helps improve patient compliance by extending or eliminating the need for a period between device replacement, battery replacement, or battery recharging during a monitoring cycle. The reduced battery consumption can allow for longer monitoring periods without device replacement, such as at least one week, at least two weeks, at least three weeks, or more than three weeks.

[0131] FIG. 10 illustrates a general overview of an embodiment of a system 900 for inferring heart rate information from an RR interval time series 902, such as may be generated by a continuous heart rate monitoring device 904. The RR interval time series 902 input to the system may include a series of measurements of timing intervals between successive heart beats. The RR interval time series 902 data may be extracted or received from a dedicated heart rate monitor, such as a heart rate chest strap or a heart rate watch, or a wearable health or fitness device 906, 908, incorporating heart rate sensing functionality. Alternatively, the RR interval time series 902 may be obtained from a wearable patch designed to measure an ECG signal 904 (e.g., by locating the R peaks in the ECG using a QRS detection algorithm). Additionally, the RR interval time series 902 may be inferred from an alternative physiological signal, such as obtained from photoplethysmography (PPG). In this scenario, the peak-to-peak interval time series determined from the PPG signal may be used as an accurate estimate of the RR interval time series.

[0132] In certain embodiments, the heart rate inference system may accept multiple RR interval time series measured from a given user's 918 device in addition to an individual RR interval time series 902. In certain embodiments, the heart rate inference system 910 may accept additional sources of data, commonly referred to as alternative sensor channels, in addition to the RR interval time series data to increase the accuracy and / or value of the inference results.

[0133] Current wearable sensors, such as the iRhythm ZioPatch™ 904, further described above in connection with FIGS. 1-9, are capable of recording single-lead electrocardiogram (ECG) signals for up to two weeks on a single battery charge.

[0134] Extraction, Transmission, and Processing System 11 is a schematic diagram of an embodiment of a system and method 1000 of a wearable medical sensor 1002 with transmission capabilities, similar to the system and / or method described above in connection with FIG. 10. In some embodiments, the sensor 1002, which may be any type of sensor or monitor described in this section or elsewhere herein, continuously senses an ECG or equivalent biosignal 1004 and continuously records the ECG or equivalent biosignal 1004. The collected signal 1004 may then be continuously extracted 1006 into one or more features representing exemplary features A, B, and C. The features of the ECG or equivalent biosignal are extracted to facilitate remote analysis of the signal 1004.

[0135] Once the feature extraction as described above is complete, the various features 1008 may then be transmitted 1010 to a processing device / server 1012. The features 1008 (and alternative sensor channel data and / or features as described below) are transmitted 1010 at regular intervals to a processing device 1012 that is not physically part of the sensor 1002.

[0136] In some embodiments, the transmitted features 1014 are processed by a remote processor that utilizes the data features 1014 to perform analysis via a heartbeat inference system 1016 that analyzes and identifies 1018 segments / locations that are likely to contain arrhythmias. For example, types of arrhythmias and ectopic excitation that may be identified may include pauses, second or third degree AVB, complete heart block, SVT, AF, VT, VF, bigeminy, trigeminy, and / or ectopic excitation. The reliability of the determination may be included in the identification of the heartbeat.

[0137] The identified arrhythmia location 1018 is then transmitted back to the sensor 1002 (1020). The transmission back to the sensor 1020 may be accomplished by any communication protocol / technology described herein or elsewhere herein, for example via Bluetooth. The sensor then reads the transmitted identification location 1022 and accesses (1024) an area of ​​memory corresponding to the transmitted identification location 1022 of the ECG. In some embodiments, the sensor applies additional analysis of the identified segment to further increase the confidence of the arrhythmia identification. This additional heartbeat confidence determination step 1026 allows for increased positive predictivity before the power-intensive transmission step. In an embodiment, if the confidence exceeds a defined threshold, the data segment is transmitted. If the confidence exceeds the threshold as described above, the sensor 1002 may transmit the requested ECG segment 1028 to the processing device via any transmission means described herein or elsewhere herein.

[0138] Figure 12 is a schematic diagram of one embodiment of a system and method 2000 of a wearable ECG and / or medical sensor 2002 having transmission capabilities very similar to the system and / or method 1000 described above in connection with Figure 11. The system of Figure 12 differs from the system of Figure 11 in that it includes a secondary transmitting device 2004.

[0139] Figure 13 is a schematic diagram of an embodiment of a system and method 3000 of a wearable ECG and / or medical sensor with transmission capability 3002, very similar to the systems and / or methods described above in connection with Figures 11 and 12. Figure 13 differs from Figures 11 and 12 in that Figure 13 illustrates alternative sensor channels 3004, 3006 that result in alternative output and / or extraction 3008 of characteristics 3010. Collection of data of other channels may be useful to further enhance the ECG extracted characteristics. Data from the alternative sensor channels may be transmitted in their entirety or specific characteristics 3010 of the data channels may be extracted 3008.

[0140] Figure 14 is a schematic diagram of an embodiment of a wearable ECG and / or medical sensor 4002 with transmitting capabilities system and method 4000, very similar to the systems and / or methods described above in connection with Figures 11-13. Figure 14 differs from Figures 11-13 because the embodiment of Figure 14 incorporates an additional data filter. In some embodiments, the processing unit 4004 can also filter the heartbeats 4006 identified by the heartbeat inference system 4008 by applying filter criteria, which may be derived from multiple sources.

[0141] FIG. 15 is a schematic diagram of an embodiment of a system 5000 for a consumer wearable device without full ECG detection, with some similarities to the medical sensors of FIGS. 10 to 14. The sensor 5002 does not need to be a medical grade ECG sensor, merely allowing detection of beats. In an embodiment, the sensor 5002 may continuously sense a data channel from which the heartbeat location may be derived. Possible data sources include PPG (optionally with multiple channels for increased accuracy), bioimpedance, and ECG, with no full implementation due to poor signal quality compared to the sensors of FIGS. 10 to 14. Similar to the devices of FIGS. 10 to 14, features may be extracted from this signal, such as R-peak location, R-peak overflow flag, saturation flag, respiration rate, P / T wave location, R-peak amplitude (or proxy), or ECG signal amplitude proxy.

[0142] The consumer device system 5000 without full ECG detection can advantageously enable arrhythmia analysis using consumer available heart rate sensors, thereby reducing the cost and increasing availability of the device, which in turn can enable arrhythmia screening for a larger population, including via over-the-counter screening.

[0143] FIG. 16 is a schematic diagram of one embodiment of an ECG monitoring system 6000 with symptom transmission. Such a system would include a wearable ECG sensor similar to the sensor described in connection with FIGS. 1-14. As noted above, such a sensor would continuously sense and record the ECG. Each symptom trigger by the patient can initiate the transmission of an ECG data strip.

[0144] 17 is a schematic diagram of an embodiment of an ECG monitoring system 7000 with both symptomatic and asymptomatic transmissions. In some embodiments, each non-symptomatic trigger initiates the transmission of an ECG data strip as described above.

[0145] A system for estimating burden from R-peak sequences using neural networks In some embodiments, a wearable device is disclosed. The wearable device may include a sensor for detecting a user's biosignal, such as an ECG signal. The wearable device may process the detected biosignal through a first subset of layers of a neural network. The wearable device may take the output of the first subset of layers of the neural network and transmit the output to a computing device (e.g., an external system such as a smartphone or a server) to further process the data through a second subset of layers of the same neural network to derive features of the user, such as indications of past cardiac arrhythmias, and / or predict the onset of future arrhythmias. In some embodiments, the computing device is a processor external to the wearable device or a processor within the wearable device. In certain embodiments, at least a portion of the computations, as described herein, may occur in one or more processors external to the wearable device and / or in one or more processors within the wearable device.

[0146] In some embodiments, the output of the first subset of layers is a representation of the R-peak sequence. In other embodiments, the output of the first subset of layers is a representation of the R-R interval. The wearable device can process the ECG signal through the first subset of layers of the neural network and receive the R-peak sequence from the output of the first subset of layers. The wearable device can transmit the R-peak sequence to an external computing device. The external computing device can process the R-peak sequence through a second subset of layers of the neural network and receive derived user features (e.g., indications of past cardiac arrhythmias and / or predictions of future arrhythmia occurrences, atrial fibrillation burden, etc.) as an output of the second subset of layers.

[0147] Recently, there has been a surge in research into the use of wearable sensors to detect arrhythmias such as atrial fibrillation (AFib). Wearable sensors used in such studies fall into two broad categories. The first category is electrocardiogram (ECG) monitoring devices. The second category is photoplethysmography (PPG)-based devices. When PPG-based devices are used to detect AFib, results have been mixed. For example, one study found that using heart rate and step count data from a PPG-based device, a deep neural network-based algorithm was adequately applied to a dataset of PPG data from a cohort of patients immediately before and after defibrillation for atrial fibrillation in a clinical setting, but the same approach and algorithm performed less well in a significantly broader outpatient population with persistent atrial fibrillation. There are many possible reasons for the poor results in the outpatient population, but the most influential factor appears to be the poor quality of the PPG signal for heart rate estimation in an uncontrolled environment. This has created a need for more accurate detection of AFib using heart rate data obtained from ECG signals recorded using wearable cardiac monitors, such as ECG monitoring devices.

[0148] Atrial fibrillation burden is the percentage of time a patient spends in a state of atrial fibrillation. Recently, the concept of atrial fibrillation burden has gained attention due to studies showing an association between atrial fibrillation burden and risk of ischemic stroke in adults. Therefore, cardiac monitoring solutions that can rapidly and accurately estimate a patient's atrial fibrillation burden may lead to meaningful medical interventions for patients identified as at high risk for stroke.

[0149] For clinical analysis, a patient's ECG signal recorded using a cardiac monitoring device is usually segmented into episodes of different durations and different heartbeat types. A normal, healthy patient is usually in a state of normal heartbeat called normal sinus rhythm (NSR). Various cardiac diseases manifest as irregular / abnormal patterns of heartbeats known as arrhythmias, which are translated into irregular / abnormal patterns of ECG signals. There are various types of cardiac arrhythmias, such as atrial fibrillation. A patient may alternate between a normal heartbeat and various types of cardiac arrhythmias or may be constantly in one of these abnormal heartbeats, such as atrial fibrillation. The burden of atrial fibrillation is the percentage of time a patient spends in a state of atrial fibrillation. Atrial flutter is also an arrhythmia closely related to atrial fibrillation that increases the risk of stroke. Atrial fibrillation and atrial flutter are commonly grouped together for burden calculations because they are sometimes difficult to distinguish from one another and have similar clinical significance. Thus, for purposes of this disclosure, embodiments will be described with reference to AFib, however it will be understood that some embodiments are also (or only) applicable to atrial flutter.

[0150] The QRS complex is the main hallmark of an electrocardiogram signal. The peak within the QRS complex is commonly referred to as the R-peak. In some embodiments, the R-peak signal is detected from the ECG signal. Once the R-peak is detected, the instantaneous heart rate can be estimated as the inverse of the duration between two successive R-peaks (called the RR interval). The main distinguishing signature of an atrial fibrillation episode is the large variability of the RR interval.

[0151] 18A and 18B are graphs of an embodiment illustrating readings from an ECG strip corresponding to NSR heart beats, with the dots indicating the location of the R-peak.

[0152] The electrocardiogram signal recorded from a patient can be segmented into episodes of different beat types and the burden of atrial fibrillation can be calculated by calculating the percentage of time spent in atrial fibrillation. However, if only the sequence of R-peak locations is extracted (and not the entire ECG signal) and the calculation is performed, the estimation of the AFib burden becomes a matter of a separate algorithm. Scenarios in which such an algorithm may be useful include those cases where one wants to avoid excessive calculations in the atrial fibrillation monitoring device itself (using the entire ECG signal) or those cases where one wants to use wireless communication to limit the amount of data sent from the monitoring device to a computer server to calculate the atrial fibrillation burden.

[0153] Described herein are embodiments that use heart rate data from a wearable cardiac monitor to reliably detect AFib. Some embodiments can also reliably estimate a patient's daily AFib burden.

[0154] In some embodiments, the algorithm (eg, a neural network) can estimate AFib burden over shorter analysis windows (eg, 30 minute or hourly level durations).

[0155] In some embodiments, the burden over a longer period (e.g., longer than a shorter analysis window) can be determined by accumulating the burden predictions over the analysis window. For example, the burden can be accumulated over non-overlapping analysis windows of 30 minutes duration. Given an R peak sequence over a 30 minute analysis window, an RR interval sequence within the analysis window is derived. The RR interval sequence is then converted into a sequence of RR interval subsequences by extracting RR interval subsequences from a set of possibly overlapping sliding windows of duration w (subsumed within the analysis window) and shifted by length s. The RR interval subsequences obtained from each sliding window are padded / truncated so that all subsequences have the same length M. For example, if there are Nw sliding windows within the analysis window, the dimension of the time sequence is (Nw, M). This time sequence is fed into a neural network model to perform the burden prediction. For the purposes of illustrating the concept, w was chosen as 30 seconds, s as 25 seconds, and M as 42. A portion of the neural network on the monitor (e.g., a first subset of the layers of the neural network) can process the detected cardiac signals to generate RR interval data (and / or RR interval subsequences) and transmit the RR interval data to a remote computing device. The remote computing device can process the RR interval data (e.g., by processing through a second portion of the neural network (e.g., a second subset of the layers of the neural network)) to make inferences such as the likelihood of arrhythmia. In some embodiments, the neural network can be trained to make decisions regarding use of the wearable device, such as an indication that the user should wear the device longer due to a high risk of a potential negative health event, or that the user is at low risk and can terminate the wearing period sooner.

[0156] Advantageously, such decoupled processing of data may have many technical advantages. For example, such selective transmission of partially processed data may reduce the power consumption of the wearable patch. The wearable patch may not need to store the entire data processing algorithm (such as the entire neural network), but may only need to obtain a subset of the neural network layers to process. Thus, data processing on the wearable device may be much faster, require less computation, and use less battery throughout the processing. Analysis on the wearable patch side may not require a full on-board heart rate analysis, which often results in very high power consumption and may reduce battery life. The reduced battery consumption may extend the use of the wearable device, such as the period between device replacement, battery recharging, etc. Furthermore, the wearable device may have a longer monitoring period to get a complete picture of the patient's heart rate.

[0157] Another advantage of remote analysis based on RR interval data is that it may allow for greater sensitivity and precision than on-patch analysis, which may be limited by memory, processing power, and battery constraints in wearable patches. Remote analysis can process the data through a wide variety of algorithms, even algorithms based on decision trees of determination and occurrence.

[0158] Furthermore, the wearable patch may transmit data with a smaller dimensionality than the entire detected electrocardiogram signal. Advantageously, less network throughput is required. The wearable patch can transmit sufficient data for remote data processing even when wireless network conditions are not optimal. Furthermore, because only a subset of the neural network needs to be uploaded, the wearable patch can be updated with updated neural network computer executable instructions or downloadable executables faster, in weaker network conditions, and more frequently.

[0159] Furthermore, due to the lower battery and power requirements, a smaller CPU or a smaller battery can be used to achieve a smaller form factor for the wearable patch. Furthermore, the reduced battery and power requirements for data processing free up the battery and power budget of the wearable device to perform other important tasks, such as running other important software processes or powering other hardware devices, such as more electrodes, other sensors such as accelerometers, or running the same hardware and software for a longer period of time.

[0160] The neural network encoding approach has the advantage that very sophisticated algorithms can be run from the data of the battery-constrained device. An example application of this is to aggregate the encoded data from the device in real time and run an algorithm that evaluates the overall risk of the subject developing a certain clinical outcome. In this way, the system can dynamically inform the user of the wearing period. For example, after the wearable device has aggregated a certain amount of data, the remote computing device can determine that the subject is at very low risk of developing a certain clinical outcome. The system can inform the user of the wearing period early. Conversely, if the computing device determines that the user is at high risk for a certain clinical outcome, the system can inform the user to wear the device longer.

[0161] Another technical advantage of the first subset of layers of the neural network outputting RR interval data transmitted for remote processing is that the algorithm for performing inference can be updated remotely without updating the wearable patch. For example, an updated algorithm or neural network can be updated on the remote computing device via an update to the second subset of layers, while the same first subset of layers can be used on the wearable device. A technical advantage is that the second subset of layers can be selected based on a decision tree, updated or replaced with other algorithms, etc., without updating the wearable patch, making the system more compatible with other algorithms.

[0162] Example of high-level architecture of the burden forecasting model The high-level architecture of the burden prediction model is shown in Figure 19. The burden prediction model takes the RR interval subsequences as input and generates an AFib burden prediction: where Nw is the number of sliding windows and M is the size of the RR interval subsequence extracted from each sliding window.

[0163] The first layer 1902 is a time-variant feature extraction layer that extracts a 32-dimensional feature vector for each RR interval subsequence obtained from the sliding window. The output of this layer is a transformed sequence of dimension (Nw, 32). This transformed sequence is then fed into two Recurrent Neural Network (RNN) layers 1904, 1906 with Long Short Term Memory (LSTM) cells. The first RNN layer 1904 returns another transformed sequence of dimension (Nw, 32), while the second RNN layer 1906 returns a single output of dimension 32. This is followed by two Dense (fully connected) layers 1908, 1910 that generate a scalar value representing the AFib burden of the analysis window.

[0164] FIG. 20 is a diagram illustrating the LSTM architecture in more detail, in accordance with a high-level architectural embodiment of FIG.

[0165] Figure 21 is an embodiment of the feature extraction model of Figure 19. As shown, the feature extraction model can be temporally distributed over a sliding window in its first layer. Also, the model for feature extraction that is temporally distributed over a sliding window can itself be another neural network with multiple layers.

[0166] At the start of Figure 19 and at step 2102 of Figure 21, the feature extraction model takes as input an RR interval subsequence of length M, which is processed by a set of 1D convolutional and Max-pooling layers at steps 1902, 1904, 1906 of Figure 19 and steps 2104, 2106, 2108, 2110, 2112, 2114 of Figure 21. This is followed by a dense (fully connected) layer that outputs a 32-dimensional feature vector at steps 1908, 1910 of Figure 19 and steps 2120, 2122 of Figure 21. The intent of this feature extraction model is to learn features indicative of the presence / absence of AFib within each sliding window of duration w.

[0167] For neural networks, the average of the absolute difference (over all samples) between the burden predictions obtained from the model and the true burden values ​​can be used as the loss function:

[0168]

number

[0169] where b is the burden prediction output by the model for the i-th window of the dataset, b is the true burden in that window, and N is the total number of samples in the dataset.

[0170] In some embodiments, the parameters of the feature extraction network (e.g., FIG. 21) may first be learned separately by adding one more dense (fully connected) layer with scalar output and sigmoid activation to the feature extraction network, and then training the resulting model to perform binary classification for detecting the presence of AFib in 30 second ECG strips.

[0171] We freeze the parameters of the resulting feature extraction network and plug the resulting feature extraction network into the burden prediction model (Figure 20). Next, we train the subsequent layers of the burden prediction model (the recurrent network layer and the fully connected layer) in the 30-minute window mentioned in Section 4.2.

[0172] The feature extractor network is trained on a larger and more diverse set of ECG strips and can directly learn features that are indicative of the presence or absence of atrial fibrillation. When training the remaining layers of the strain prediction model, the number of parameters that have to be learned is reduced, which may lead to a faster search for optimal parameters. On the other hand, when training the entire network simultaneously, the number of parameters that have to be learned simultaneously is larger and the model is exposed to a less diverse set of ECG strips. This increases the risk of overfitting and is probably the reason for the larger gap between validation loss and training loss.

[0173] System for estimating strain and / or predicting heart rate annotation using neural networks Some embodiments disclose a wearable device that can process a detected biosignal through an encoder that includes a first subset of layers of a neural network. FIG. 22A is a schematic diagram of an embodiment of a system for predicting a heartbeat annotation using neural network encoding. In some embodiments, the system can make one or more predictions, such as predicting a heartbeat annotation or estimating a burden for a heartbeat rhythm, such as atrial fibrillation and / or atrial flutter. A wearable device, such as a heart monitor patch 2202, can process an ECG input 2204 through a first subset of layers 2206 of a neural network, such as an encoder. The wearable device can receive the output of a first subset of the layers of the neural network 2206 and transmit the output to a computing device 2208 (e.g., an external system such as a smartphone or a server) for further processing the data via a decoder including a second subset of the layers of the same neural network 2210 to derive user features 2212, such as indications or predictions of past cardiac arrhythmias and / or predictions of future occurrences of arrhythmias.

[0174] In some embodiments, the first and second subsets of layers are within one neural network. The neural network can be designed such that the output of the first subset of layers can output data with a smaller dimensionality than the input to the neural network, and the output of the second subset of layers can be designed to provide an indication of user features, such as past or future predictions of AFib. FIG. 22B is a schematic diagram of an embodiment of a first and second subset of layers within a single neural network. The neural network can be trained simultaneously on both the first subset of layers 2222 and the second subset of layers 2224. For example, if the neural network has 10 hidden layers, the first four layers are processed on a wearable device such as a heart monitor patch 2226, and the output of the fourth layer, which is of a lower dimensionality than the input to the first layer (e.g., has good data compression capabilities), is sent to an external computing system such as a server 2228. The external computing system processes the output of the fourth layer through layers 5-10. The dimensionality of each layer of the neural network can be designed to output a particular data size (e.g., the output dimension of each convolutional or pooling layer). For example, the patch 2226 can include an ECG encoder 2230 that receives ECG data 2232 at 2400 bits per second (bps). The patch 2226 can process the ECG data 2232 through a first subset of the layers 2222 of the neural network (e.g., in the ECG encoder 2230) and output data 2234 of a smaller dimensionality, such as data at 128 bps. The output data 2234 can be sent to an external server 2228. The external server 2228 can include a classifier 2238 that processes the output data 2234 through a second subset of the layers 2224 of the neural network and outputs a symptom or prediction 2236 of the patient for which it is trained. The entire neural network, including the first subset 2222 and the second subset 2224 of layers, can be designed and trained as a single neural network.

[0175] In some embodiments, the output of the first subset of layers can be of a smaller dimensionality than the input to the neural network. In this way, instead of transmitting the entire ECG signal (e.g., the input to the neural network) from the wearable device to an external computing device, the wearable device can transmit a smaller amount of data to the external computing device, such as the 128 bps output data 2234 of the first subset of layers instead of the entire 2400 bps ECG signal 2232. Advantageously, less network throughput is required to derive indications of past cardiac arrhythmias and / or predict the onset of future arrhythmias.

[0176] Further, instead of processing the ECG signal on the wearable device through all layers of the neural network to derive an indication of past cardiac arrhythmias and / or predict the onset of future arrhythmias, the wearable device can process the ECG signal only a first subset of the layers of the neural network (e.g., via an encoder) via the ECG encoder 2230 and transmit the output of the first subset to an external device 2228 that processes a second subset of the layers (e.g., via a decoder or classifier 2238). FIG. 22C is a schematic diagram of an embodiment of a system for processing a subset of the layers of a neural network on a patch. The patch 2242 can process all layers of the neural network 2244 to receive an indication of a prediction of AFib. However, such processing may require high memory and power requirements (which in turn may lead to larger size requirements, larger patch sizes), and increased battery usage, leading to shorter usage times. However, in other embodiments, the patch 2246 may process a first subset 2248 of the neural network's layers, and the external server 2250 may process a second subset of the layers. Advantageously, the wearable device performs a smaller amount of calculations and requires less memory to store the data between the layers.

[0177] Additionally, the wearable device 2246 can be smaller in size because a smaller battery, processor, and other electronics are required to process only a subset of the layers of the neural network. The wearable device can have more resources to perform other tasks. For example, in one scenario, the wearable device 2242 can have enough computing power and battery to process ECG signals internally and identify AFib for 5 days. However, if the wearable device 2246 can compute a first subset of the layers of the neural network on the wearable device 2246 and the external computing device 2250 can process a second subset, the system as a whole may be able to identify both AFib and atrial flutter over a span of 20 days. Some embodiments disclosed herein can reduce network throughput, reduce computational demands, and / or reduce storage capacity of the wearable device. Advantageously, the wearable device can preserve battery life by using less of the processor.

[0178] Certain embodiments disclosed herein improve the memory requirements of the wearable device. If the wearable device processes a first subset of layers of a neural network and an external computing system processes a second subset, the wearable device needs to store less data (e.g., software executable instructions for only the first subset of layers, rather than the entire neural network). Thus, the size of the electronic file is much smaller. Advantageously, not only does the wearable device have smaller memory requirements, but the neural network can also be loaded and processed faster than a complete neural network. Furthermore, updates to the software embedded within the wearable device can be performed much faster, even under weaker network connection conditions, and can be programmed to be performed more frequently. Furthermore, more neural networks can be stored in the wearable device. For example, instead of storing a single complete neural network in memory, the wearable patch can store the first subset of layers for five neural networks in local memory.

[0179] In some embodiments, the data transmitted from the wearable device does not need to be sufficient to recover the ECG signal with full fidelity, since the purpose of the external computing device is not to reconstruct the ECG signal but to derive some user features (e.g., prediction of classified heart rate conditions such as AFib). Thus, the neural network can be trained to generate an output at one of its layers with a smaller dimensionality than the input of the ECG signal, sending much less to the external computing system to run other layers of the neural network. FIG. 22D is a schematic diagram of an embodiment for transmitting the output of a first subset of layers to a server. The patch 2262 can transmit the complete ECG signal (e.g., at 2400 bps), or other feature data, to the server 2264. However, such data can be hacked by an unexpected third party to obtain confidential data. Some embodiments disclose processing a subset of the layers of the neural network on the wearable device, providing technical advantages over processing the entire ECG signal remotely. Transmitting the entire ECG signal at 2400 bps, or a combination of the R-peak, R-interval, and other features from the wearable device to an external computing system may require more network bandwidth than if the wearable device transmitted the output of a subset of layers with less dimensionality than the ECG signal. Instead, some embodiments of the present disclosure include a patch 2266 that processes a subset of layers 2268, and outputs a first subset of layers to a server 2270 that runs the remaining layers 2272 of the neural network. Advantageously, the transmitted data may be the output of the layers of the neural network, providing enhanced data security and encryption.

[0180] Some embodiments improve the security of health data. For example, if the entire ECG signal, the R peak, the RR interval, or the final determination of the possibility of AFib were transmitted from the wearable device to an external computing system, a third party could intercept such data and identify the user's confidential health data. Instead, some embodiments disclosed herein transmit the output of an intermediate layer of a neural network (such as the output of an encoder) to be processed externally by a subsequent layer (such as a decoder). The transmitted data may be the output of a layer of a neural network that originally provides encryption of data for data transmission. Thus, even if a third party intercepts such data, they cannot decode the data or reverse engineer the original input of the neural network.

[0181] In some embodiments, a first subset of layers may include less computationally intensive layers, while a second subset of layers may include more computationally intensive layers. Advantageously, the wearable device can process the less computationally intensive layers, and an external computing system that is agnostic to the wearable device's processing (such as a server with much greater processing power and memory capacity) can process the more computationally intensive layers.

[0182] In some embodiments, the wearable patch can perform pre-processing on the detected cardiac signal before processing the data through the first subset of layers of the neural network. For example, the wearable patch can perform downsampling, normalization, digital filtering, and / or such pre-processing. Additionally, the wearable patch can perform more complex pre-processing, such as discrete wavelet transform, continuous wavelet transform, discrete Fourier transform, and discrete cosine transform. These transforms can be beneficial in lossy compression schemes, as they transform the data in such a way that many values ​​in the data are very close to zero. They can then create an encoding that simply preserves the values ​​with the highest amplitude, reducing errors in reconstructing the signal.

[0183] The wearable patch can simply apply an algorithm that applies a discrete wavelet transform or a discrete cosine transform and makes inferences on the detected cardiac data. However, such an approach may not be able to capture finer features important for end applications such as ECG heart rate classification, since higher amplitude features are favored. In some embodiments, the system can train a neural network encoder on this transformed data set (e.g., pre-processed cardiac data) to more intelligently encode the data required for the end application. Advantageously, this may be beneficial for some applications, which may lead to more powerful or simpler neural network encoder designs with fewer parameters compared to those that operate on raw signals, since the transformation may already have performed certain operations on the signals required for the end application.

[0184] In some embodiments, the neural network may be trained and / or designed to transmit an optimal data dimensionality from the wearable device to an external computing system. FIG. 22E is a schematic diagram of an embodiment of designing (and / or training) a neural network. The patch 2281 may process a first subset of layers 2284A, 2286A, 2288A of the first neural network 2284, the second neural network 2286, and / or the third neural network 2288. The patch 2281 may transmit the output of the first subset of layers to the server 2282. The server may process a second subset of layers 2284B, 2286B, 2288B of the first neural network 2284, the second neural network 2286, and / or the third neural network 2288.

[0185] The neural network can be trained and / or designed based on one or more factors, such as the available processing power and / or memory of the wearable device (for purposes of this disclosure, the neural network is described as being trained, but the attributes can be applied to design or process data through the neural network, and vice versa). For example, if the wearable device has low processing power, the neural network can be trained to perform less computations and / or processing through fewer layers of the neural network on the wearable device and more processing on the external computing system. Patch 2281 can apply a third neural network 2288 that includes fewer computations and / or layers of the neural network.

[0186] In some embodiments, the neural network may be trained depending on the availability of the network between the wearable device and the computing system. For example, the neural network may be trained to process more layers and / or transfer a smaller data dimensionality from the wearable device to the external computing system when the network throughput is low. The patch 2281 may apply a second neural network 2286 including a smaller data dimensionality to be transferred to the server 2282 for when the network connection is weak (or below a threshold).

[0187] In some embodiments, the neural network may be trained according to the available battery life of the wearable device. For example, the neural network may be trained to process fewer layers on the wearable device if the battery life is lower. The patch 2281 may apply a third neural network 2288 that includes fewer calculations and / or layers of the neural network.

[0188] In some embodiments, the neural network can be trained according to a desired outcome. For example, the patch 2281 can apply a first neural network 2284 that includes more outputs of the neural network, such as more features of the user. For example, if the patient is critically ill or has symptoms of AFib, the patch 2281 can apply the first neural network 2284 and process more layers and nodes through a first subset of layers 2284A so that the server 2282 can output more information through an output of a second subset of layers 2284B.

[0189] In some embodiments, the software of the patch 2281 may include instructions for a particular neural network designed (and / or trained) for a particular application. For example, one neural network may be selected for the patch 2281 with certain technical constraints (e.g., memory or processor) and serve a particular purpose (e.g., predicting AFib), and another neural network may be selected for a different patch for different features and purposes. For example, accurate detection of AFib may require more data from the patch than another desired purpose. Thus, a patch for detecting AFib may have software that computes more layers than another patch.

[0190] In some embodiments, the software of patch 2281 may be designed (and / or trained) based on a performance metric. For example, the software may select a neural network designed to produce a certain accuracy in detection and / or a certain false alarm rate.

[0191] In some embodiments, the software of patch 2281 can include multiple neural networks. Patch 2281 can select a particular neural network from among the multiple networks. For example, the selection can be based on a neural network designed for cases of low battery power or network connectivity (e.g., switching to a neural network that does less processing on the patch or transmits less data over the network).

[0192] In some embodiments, the neural network may be trained to recreate an original input signal to the neural network, such as an ECG signal, R-R interval information, R-peak information, metrics derived from the ECG signal, accelerometer data, impedance, temperature, sound (e.g., snoring), ambient light, and / or the like. Patch 2281 may transmit data of a smaller data dimensionality to server 2282, which recreates the original input signal, or a derivative thereof, such as the original input signal sampled at a lower frequency, a lower quality signal, a signal with a certain threshold for reconstruction error, and / or the like.

[0193] In some embodiments, the neural network may be trained according to the user's current activity. FIG. 22F is a schematic diagram of an embodiment for designing, training, and / or selecting a neural network based on the user's current activity. For example, the user may be sleeping 2296 or running 2292, or undergoing a certain medical procedure that may require different types of peak sensitivity. There are some specific activities that may require higher accuracy or precision so that irregularities in the heartbeat can be effectively detected. If the activity requires higher R-peak accuracy, the neural network may be trained so that the wearable device may transmit higher resolution data (e.g., for a sleeping user 2296, the signal level 2298 is less affected by motion artifacts than for a running user 2292, and therefore the ECG signal can be relied on more confidently. Thus, a high R-peak accuracy may be preferred for a running user 2292 for more accurate predictions, and a more robust encoding process may be used for the running user 2292 to ensure data accuracy). In contrast, for a running user 2292, the R signal level 2294 may be clear (e.g., a clear R peak is exhibited), and therefore a lower R peak sensitivity may be sufficient. In this scenario, the patch may select a neural network designed to not require higher R peak sensitivity, which may lead to advantages as described herein, such as reduced processing power or network throughput.

[0194] Training of models, such as artificial intelligence models, is necessarily rooted in computer technology, where such models are trained with training data and then refined by applying the model to new user data sets. Such training involves complex processing requiring a lot of processor computing and long periods of time on large training data sets, typically performed by huge server systems. Training of a model may require logistic regression or forward / backward propagation of training data, which may include input data used to adjust the parameters of the model and expected output values. Such training is a framework of machine learning algorithms, where the model is applied to new data (such as new biometric data) and allows the model to make the predictions it was trained on based on the weights or scores adjusted during training. Such training can reduce false positives and improve AFib detection performance.

[0195] In some embodiments, the wearable device may store multiple neural networks and apply one neural network based on one or more characteristics of the wearable device, the network, the user's current activity, etc. For example, the wearable device may process a first neural network when battery life is low and a second neural network when network throughput is high.

[0196] In some embodiments, the wearable device can continuously process the ECG signal through a first subset of layers of the neural network as the wearable device detects the ECG signal in real time. The wearable device can continuously transmit the output of the first subset of layers to an external computing device, which can process the received output through a second subset of layers of the neural network to derive a feature (e.g., cardiac arrhythmia) of the user in substantially real time. Advantageously, the external computing device can identify cardiac arrhythmias occurring in the user in substantially real time while reducing network throughput, memory requirements, and processing power requirements. Additionally, more powerful algorithms and neural networks are now possible that would not have been possible to run entirely on the wearable device itself due to device constraints, and on the external computing system due to input requirements.

[0197] In some embodiments, the neural network may include a first neural network and a second neural network. The first neural network (e.g., a time-distributed feature extraction model) may receive as input one or more types of data of a user's ECG signal, R-peak data, RR interval subsequences (e.g., RR intervals in a 30-second sliding window), encoded features, and / or the like, as further described herein. The first neural network may process the input RR interval subsequences and generate an atrial fibrillation (AF) feature for each sliding window. The output of the first neural network may be fed to a second neural network including a recurrent layer having long-short-term memory (LSTM) cells. The LSTM cells may process the AF features of the current sliding window and past sliding windows to process the data using a transfer learning approach, and over the course of multiple sliding windows, the LSTM may advantageously predict the onset of cardiac arrhythmia considering data longer than a 30-second window.

[0198] In some embodiments, the first neural network and the second neural network are trained in stages. The first neural network can be trained first. Training data for 30 seconds can be fed to the first neural network, and the weights can be adjusted based on the expected output of the first neural network. After the first neural network is trained, the weights of the first neural network can be frozen and the second neural network can be trained. The second neural network can be trained based on a longer interval of data than the data fed to the first neural network in a 30 second window, and the output is fed to the second neural network with an LSTM cell. The second neural network can make predictions of AFib over a longer data set, and the predicted output can be used to adjust the weights of the second neural network. Once the training of the second neural network is complete, the weights of the first neural network can be unfrozen and both the first and second neural networks can be trained simultaneously with additional training data.

[0199] As wearable biosignal processing applications proliferate, there is an increasing need to improve the sophistication and practicality of these applications. A typical scenario is a small, battery-powered wearable device equipped with sensors that measure biosignals, process them, and optionally transmit them in real-time or near real-time. Since the size of the device affects comfort, which in turn affects adaptability and therefore usefulness, it is advantageous to make the device as small as possible. This imposes several constraints on the system (mainly battery life and processing power) that may result in a negative impact on the user's experience, either by requiring the user to charge the battery or by reducing the quality of the analysis and adaptability. These constraints limit the sophistication of the processing that can take place on the device, so an option that is utilized in many applications is to offload the processing to less constrained systems, such as smartphones, cloud servers, or base stations.

[0200] Within this paradigm, there is a trade-off between the amount and type of data that is offloaded for additional processing. Ideally, full fidelity raw biosignals would be offloaded in real-time, maximizing the sophistication of processing in the least constrained part of the system. Unfortunately, with current technology limitations, this is usually not feasible due to battery costs and technical constraints to transmit full fidelity signals in real-time or in a timely manner. A way to overcome this limitation is to transmit a smaller representation of the signal, minimizing power requirements and maximizing the time between battery recharges or replacements, or the potential monitoring period of the device when powered by a single battery charge.

[0201] There are many strategies for creating this alternative representation. Depending on the application, custom designed algorithms can be used to refine the attributes uploaded for further processing. One example of this strategy is a mobile cardiac telemetry (MCT) device or a wearable activity tracker that can detect and upload the location of the heartbeat obtained by algorithms running on the device, which can then be further analyzed to detect cardiac arrhythmias by more advanced algorithms on a cloud server. Another example is a wearable device that uses an accelerometer to analyze gait to assess activity levels and uploads the results of a step counting algorithm on the device to assess risk of falling. The limitations of this set of approaches are that the features selected may not be optimal for the end use, and the device-side algorithms may be more computationally intensive than necessary.

[0202] Another common approach would be to compress the raw biosignal data in a way that allows for signal reconstruction. Commonly used techniques include lossless and lossy compression methods. While lossless compression allows for optimal processing on cloud-based servers, it generally does not provide a high enough compression ratio to reduce the burden of data transmission to a sufficient extent. Lossy compression techniques, such as wavelet-based compression, can achieve relatively high levels of compression, even when considering the amount of error introduced during reconstruction. While this strategy may be effective for certain applications, there is no guarantee that the reconstructed signal will retain the information required for the end use, as the objective is only to minimize the signal reconstruction error. This type of objective is primarily driven by high amplitude signal features, and may ignore or distort low amplitude features. In some applications, relatively low amplitude signal features may be important to the end purpose, such as the subtle P waves in an ECG associated with atrial fibrillation, which are important for detecting heartbeats such as atrial fibrillation or complete heart block.

[0203] The systems and methods disclosed herein achieve an optimal balance of all elements in the system: device processing burden, system transmission burden, and end-use practicality. This is achieved by utilizing neural network encoding, which can be used to generate schemes that compress raw biosignal data in a computationally easy manner while retaining the information required for a particular end application. A neural network can encode a signal into a compressed form and reconstruct the signal. More generally, a neural network architecture can be built into an encoder portion and a decoder portion.

[0204] In this application, the encoder section is responsible for compressing the raw biosignals, and should run on the target device, resulting in an output with smaller dimensionality than the input.

[0205] The decoder section is responsible for generating predictions for the target application starting from the encoded layers of the neural network and may be deployed elsewhere than the target device, e.g., in smartphones, cloud servers, communication hubs / gateways, etc. The encoder and decoder sections are trained as one end-to-end neural network, which effectively creates an optimal encoder and decoder simultaneously, since they are jointly optimized for the same purpose. This approach maximizes the flexibility of the system design, leading to optimal implementations. For example, the encoder section can be designed to target a specific computational complexity or compression ratio, and the output of the neural network can be selected exactly as required for the end application. This technique is ideal for any real-time and near real-time signal processing application that requires high computational complexity and has a mechanism to offload the processing.

[0206] Devices that utilize electrocardiograms (ECG), photoplethysmography (PPG), and audio signals to detect the electrical and structural status of the heart in real-time or near real-time face these challenges. Certain cardiac arrhythmias manifest themselves in subtle ways, for example, low amplitude P-wave patterns in ECG recordings, often requiring sophisticated and sensitive detection algorithms. To maximize clinical utility, heartbeats may be classified into up to 15 classes including noise, which not only depends on beat-to-beat patterns but also on morphological differences that are often subtle, especially in single-lead ECG applications. Existing devices may simply offload the entire raw signal, requiring frequent charging or battery changes, use compression techniques to offload processing, or detect commonly understood features of the ECG signal in the device and upload them to another system for further processing. Features may include the location of ECG morphological features such as the QRS complex, P waves, etc. The limited computational power and battery life of these devices means that algorithms have limited precision in their ability to detect these morphological features. Neural network encoding can be employed to transmit a compressed representation of the ECG signal. This representation can be optimized for a specific end use, such as identifying specific cardiac arrhythmias, identifying ECG morphological features such as QRS complexes, classifying beat types, and determining heart rate.

[0207] In an example embodiment of this approach for detecting arrhythmias in real-time or near real-time from ECG recordings, a neural network can first be trained using a set of 30-second ECG strips sampled at, for example, 200 Hz and with 14 different heartbeat labels for each sample. As preprocessing, the 200 Hz signal is downsampled by a factor of 3 with a moving average filter and conditioned with an IIR high-pass filter with a 2 Hz cutoff. This input is also scaled to have unit variance. This leaves information in the range of approximately 2-33 Hz, which contains the majority of the ECG morphological information. This results in an input dimensionality of the signal to the neural network of 2000x1. As mentioned before, the purpose of the first section of the neural network is to compress the signal, so the dimensionality of the input signal needs to be reduced.

[0208] In one embodiment, the encoding portion of the neural network has a sequence of convolutional and pooling layers, with each pooling layer reducing the time dimensionality by a factor of 2. Repeating the sequence of convolutional and pooling layers three times reduces the time dimensionality by a factor of 8. Also in this embodiment, the final convolutional layer of the encoder has 2 filters, which reduces the output dimensionality to 250x2, and is referred to as the encoded output.

[0209] The next layer of the network employs a strategy to restore the size of the time dimension to the original size of the network input. This can be accomplished using several transposed convolution layers (sometimes referred to as deconvolution layers). The layers of the network up to this point may be pre-trained to predict the original input signal, in one embodiment. In one embodiment, the remaining layers include residual convolution filter blocks. In this example, 12 sequential blocks are used, each block including two convolution layers, with a residual connection from input to output. These blocks also halve the time dimension every other block. After these last blocks, a temporally distributed dense layer is evaluated at each time point to predict a unique heartbeat class. The above example describes one embodiment of a framework for training an algorithm that encodes ECG information in a way that preserves diagnostic information in a compressed format.

[0210] In some embodiments, the encoding process can be implemented to persistently generate ECG features. Using the neural network of the above example, a 6000 sample long ECG strip creates 250 examples of feature vectors with two dimensions. This can be thought of as resulting in two encoded features for every 24 individual ECG samples. Because the nature of this process is causal, these features can be persistently created without having a concept of 30 second ECG segments. These features can be scaled so that they can be stored in a compact integer representation, and lossless compression can be applied to further reduce the memory footprint of the features. Using a continuous stream of encoded features, a server-side algorithm can employ many different strategies to detect events of interest using the encoded features.

[0211] In one embodiment for detecting segments containing arrhythmias (CHB, VT, AFIB / AFL, SVT, AVB, etc.), the original multi-class algorithm used to train the ECG encoding can be evaluated with a 30 second context of encoded features generated on the patch. The algorithm can be evaluated on a rolling window of 30 second context with or without overlap. Segments with a high probability of containing arrhythmia can be used to determine whether an event of interest is present.

[0212] In one embodiment, the encoding portion of the neural network may use a combination of convolutional, pooling, and recurrent layers, such as LSTM, to reduce the dimensionality of the signal.

[0213] In one embodiment for detecting VT (ventricular tachycardia) events, a neural network can be trained using the encoded features to predict VT sequences in ECG segments of approximately 1 second duration. The advantage of this is that a distinction can be made between more severe and less severe segments to determine which events are more interesting. For example, longer segments may be prioritized over shorter segments. This algorithm can be used to determine the burden of VT episodes that meet more specific length criteria. Conventional VT detection methods (especially in battery-constrained implantable devices) rely on beat patterns that suddenly increase the heart rate.

[0214] For example, reliance on beat detection such as the widely used Pan-Thompkins algorithm often performs poorly on wide QRS beats, necessarily limiting the ability to reliably identify arrhythmia beat patterns such as VT. Even when a complete ECG is present and morphology information can be examined by signal processing techniques, the computational burden is high for battery-operated devices, requiring devices to be recharged one or more times daily and / or to have batteries replaced in both the sensor and the gateway or mobile phone unit. For example, many current mobile cardiac telemetry (MCT) modalities struggle with this requirement, which is passed on to the patient, resulting in poor compliance, reduced time available from the diagnostic device, and reduced clinical value.

[0215] In one embodiment for assessing AF burden, the original neural network trained to create the ECG encoding can be evaluated over a 30 second context to determine if the segment contains atrial fibrillation (AFIB) or atrial flutter (AFL), and the sum of these times can be used to calculate the burden. Furthermore, for atrial flutter, existing modalities often struggle to distinguish it from other regular heartbeats, such as normal sinus, because there is no irregular beat pattern characteristic of atrial fibrillation. This requires detailed examination of the beat-to-beat signal, which the encoded characteristics allow. Of course, this is possible with a full ECG, but it requires more computational power to process the information, which often takes the form of frequency domain analysis, which in turn results in a device that is difficult for patients to use, with charging or battery replacement, as described in the MCT embodiment above.

[0216] In another embodiment for assessing the burden of AF, the encoded traits can be fed into a recurrent neural network over a larger context, e.g., 30 minutes, to predict the burden of AF in a similar manner as described above.

[0217] In another embodiment for characterizing VT, the encoded features can be fed into the network in a larger context, e.g., 30 minutes, to predict the total duration of VT and / or the number of distinct VT episodes present.

[0218] In one embodiment for determining heart rate, a neural network can be trained to use the encoded features to predict the heart rate within a time segment, such as 3 seconds. This neural network can be used to predict the number of beats within a time segment over a sliding window, which can then be used to estimate the heart rate at various times. This heart rate information can be used to further refine or prioritize events of interest detected by other algorithms. A similar approach can be used to estimate heart rate variability instead of or in addition to heart rate.

[0219] In one embodiment for determining the number of ectopic beats, a neural network can be trained to use the encoded features to predict the number of ectopic beats contained in a time segment, such as 3 seconds. This neural network can be used to estimate the total number of ectopic beats over a window of, for example, a 1-hour, 12-hour, or 24-hour period. Alternatively, in combination with the cardiac beat embodiment described above, this embodiment can be used to estimate the ratio of ectopic beats to normal beats, for example, expressed as a percentage in a 1-hour, 12-hour, or 24-hour period. This embodiment, broadly described above, can be applied to either ventricular or supraventricular ectopic beats, and / or respective couplets and triplets.

[0220] In one embodiment, R-peak locations can be approximated from the time series reconstructed from the encoded features. These R-peak locations can be used by themselves to calculate heart rate, to enhance the accuracy of heart rate estimates from alternative methods, or to detect cardiac pause events, defined as a clinically significant gap in time between two venticular contractions, for example, greater than 3 seconds. As with heart rate, this information can be used to further filter or prioritize events of interest for which ECGs should be transmitted and reported. In another embodiment, the reconstructed signal itself can be evaluated along with portions of the complete ECG to provide more context, for example, whether a low, high, or irregular heart rate continues from the end of an ECG segment. This context information can aid the process of manually requesting a fuller resolution ECG from the patch based on visual inspection, leading to an overall higher diagnostic acuity backed by human quality assurance.

[0221] In an alternative embodiment, heart rates based on these R-peak locations can be plotted along with the selected ECG strip in the clinical report in a manner that provides greater temporal context for the events shown in the report (e.g., 8-second or 16-second strips). Heart rates can be plotted either as instantaneous rates (e.g., based on the interval between two R-peaks) or as temporal heart rate trends with each data point derived from multiple R-peaks. In another embodiment, average, maximum, and minimum heart rate trends can be plotted simultaneously to provide even greater clinical context.

[0222] In some embodiments, a wearable device utilizing PPG and accelerometers may be used to detect atrial fibrillation (AFIB) episodes and estimate AFIB burden levels. Sensors in this configuration very commonly suffer from poor signal quality due to environmental factors such as improper wearing, device motion, or poor physiological conditions. For this reason, it is advantageous to utilize full fidelity raw sensor measurements for AFIB detection and AFIB burden estimation that can understand noise and prevent algorithms from predicting false positives.

[0223] Existing strategies can detect beat-to-beat intervals on the device and use this information to infer AFIB episodes. Signal quality metrics can be calculated separately and uploaded. Problems arise with this strategy when reading erroneous beat-to-beat intervals when the signal quality is poor. Using the above strategies, a neural network can be developed to predict AFIB episodes or AFIB burden from raw PPG / accelerometer signals, which also encodes the signals in a compressed representation. The network can learn to encode the information in a way that is least likely to result in erroneous predictions due to poor signal quality. The encoding can be performed on the device, and the decoding section can be evaluated on a more powerful server by uploading the encoded data. In addition to identifying AFIB episodes and AFIB burden, encoded sensor data from wearable PPG devices could be used to characterize pulse rate or detect other cardiac arrhythmias such as ventricular tachycardia, pauses, and atrioventicular block. It can also be used to detect health conditions such as diabetes, hypertension, depression, congestive heart failure, and sleep apnea.

[0224] In another embodiment, the encoded features can be optimized to describe the waveform components of an ECG beat, such as P waves, T waves, and / or segments between points in the waveform, in addition to the R peak. A neural network can be trained with the encoded features to estimate the PR interval, the pattern of which can identify subtle classifications of AV block (1st degree, 2nd degree Mobitz II, Wenckebach, 2nd degree Mobitz II) or characterize Wolff-Parkinson-White syndrome with inference of QRS width. It may be desirable to transmit these less life-threatening silent arrhythmias as well, especially in event recorders and devices that do not continuously record and store all signals like ILRs, and in some MCTs where all signals are not fully analyzed.

[0225] Similarly, neural networks can be trained with the encoded features to estimate QT intervals and reveal clinically significant findings such as Long QT Syndrome, which can be monitored over time. For example, this QT measurement method can be used for drug titration, where patients adjust drug dosages to determine the optimal combination of drug efficacy and safety for an individual. Neural networks can also be trained to detect changes in the ST segment to monitor for ST elevation and ST depression, which can be warning signs of myocardial infarction.

[0226] In another embodiment, the encoding scheme can be dynamically adjusted and adapted as the system receives the signal and learns more about its characteristics. It can also be swapped out for another scheme if the target heart rate changes for a particular patient. For example, instead of a general arrhythmia detection method, the focus may be on detecting atrial fibrillation in neuropsychiatric applications such as post-stroke monitoring, and regardless of heart rate or episode duration. In this case, a pair of encoding and decoding networks can be used that are pre-trained and optimized for sensitive detection of atrial fibrillation. These different encoders can be learned offline and updated over the air in firmware when changes are needed, or can be optionally pre-programmed into embedded devices to activate when required. In an alternative embodiment, algorithms optimized for either sensitivity or specificity performance can be implemented in a similar manner.

[0227] In certain embodiments, the input signals may include channels from accelerometer and / or gyroscope axes, which often occupy a large amount of space due to the number of axes multiplied by the required sampling rate (e.g., 20 Hz to allow for inference of steps). 3-axis accelerometer and / or 3-axis gyroscope signal traits may be reduced to a smaller number of dimensions and contain valuable information that is not limited to predefined physical indicators such as amplitude. Indeed, encoding features that can only be obtained by higher frequency sampling may allow finer discrimination between activities. The encoded accelerometer and / or gyroscope traits may be used by themselves to determine sleep stages and activity levels, steps, direction, activity type, etc. The encoded accelerometer and / or gyroscope traits, or traits of these signals encoded with ECG, may be used to further distinguish between signals generated by physiological events such as arrhythmias and signals of motion artifacts. The same data may also be useful as one input (activity level trends) to assess the health status of heart failure patients, for example, in the form of ambulatory monitoring of decompensation events after discharge from the hospital.

[0228] In other embodiments, impedance information or other measurements indicative of skin contact, such as galvanic skin response, can be encoded along with or independently of the ECG. Neural networks can be trained to detect lead off regions and, in combination with ECG characteristics, can be used to filter out segments of unreliable signals. Because the same impedance value can have different signal quality across patients and conditions, simple state-machine type algorithms using only impedance values ​​can result in lead on conditions when in fact it is a non-physiological signal, or conversely, lead off when in fact it is an ECG signal and can be analyzed. The former is particularly problematic in ambulatory devices that are sometimes prone to noise that mimics arrhythmias, resulting in false positive signals, although some of the problem can be mitigated by hardware design and sizing of flexible materials such as patches. Decoder algorithms that use the encoding features of the ECG and raw or encoded features of secondary signals such as impedance and accelerometers can further eliminate these false positives and further ensure that impedance failures do not cause the actual ECG signal to be lost.

[0229] In certain embodiments, if there is a suspected arrhythmia detected using the encoded features, it may be beneficial to further analyze using the full resolution ECG on the original device. The power savings enabled by the encoder implementation should allow the decoder-based algorithm to request either transmission or on-board analysis of the portion of the full resolution ECG stored on the device. It should be computationally feasible to run another algorithm derived from a convolutional neural network (CNN) trained on the labeled full resolution ECG on the restricted region of the original ECG.

[0230] In certain embodiments, the encoded traits can be used to learn and predict arrhythmias, such as atrial fibrillation, and other health conditions that may occur in the near future, allowing for early intervention.

[0231] In certain embodiments, the processing, i.e., either the encoding or the subsequent layers using the encoded attributes (decoding), may be performed on an intermediate device, such as a smartphone or gateway device, that is transmitted before a long-range connection such as a cell or satellite. In certain use cases, for example, when there is no need to write the raw signal to the device memory, which is often a heavy part of the battery consumption in these limited power modalities, both the encoding and decoding may occur on the device itself, such as a patch, smart watch, or other wearable that can continuously monitor ECG or other signals, and only the encoded signal may be written. This approach may be useful in scenarios where storage capacity is limited but algorithms need to run on data stored in memory (e.g., data from ECGs that have occurred in the past).

[0232] In certain embodiments, the technology may be applied to other surface or transcutaneous biosignals other than ECG, such as PPG from the wrist, earlobe, or chest. Wearable devices (often consumer grade) feature the capture of PPG signals to trend heart rate and potentially screen for atrial fibrillation. These health insight features are often limited to very limited time detection, such as 30 seconds, and require relatively high power optical detection, for example with green light, to reliably capture the waveform through different skin types and motion. The entirety of the PPG cannot be stored or transmitted for contextual analysis of details beyond simple beat detection, or for aggregation into more long-term, clinically useful metrics, such as atrial fibrillation burden, when the PPG waveform contains identifying features specific to arrhythmic blood flow.

[0233] In some embodiments, PPG and / or ECG waveforms are encoded to infer blood pressure and other cardiac metrics such as blood perfusion, blood volume, cardiac output, ejection fraction, valvular function and / or health, where traditionally calculated beat-to-beat intervals, or even fiducial point intervals, may not be the most useful attributes. It will be appreciated that although certain embodiments are described with PPG, ECG may also be used where applicable, and vice versa.

[0234] In certain embodiments, this technology can be applied to even more data-heavy signals such as EEG. Wearable applications employing EEG are often limited to laboratory use due to the number of channels and the size of the device. EEG-trained encoding capabilities could allow computationally intensive decoder algorithms to be offloaded to a server, potentially enabling portability of devices that convert brain activity into communication, for example.

[0235] In some embodiments, the neural network encoding may be implemented in dedicated hardware, such as an FPGA IC, to further reduce battery consumption of the target device. The neural network encoder may be further optimized for embedded applications by utilizing integer or binary arithmetic.

[0236] In some embodiments, image or video streams resulting in marker or object detection or physiological measurements can use neural network encoding to offload processing to more powerful systems. Example applications are: Object detection in vehicles and security monitoring systems Emotional responses from facial images Physiological measurements such as heart rate and respiration based on video streams of the face and chest Seizure detection Fall Detection Augmented Reality Monitoring the infant for notable events (crying, breathing problems).

[0237] In some embodiments, the audio data may be encoded for purposes of speech event identification or real-time audio translation, respiratory monitoring, diagnosing respiratory disorders, and detecting snoring or apnea.

[0238] In some embodiments, ultrasound data may be encoded for purposes such as identifying contractions during pregnancy, detecting and characterizing fetal heartbeats, and identifying abnormal physiology such as tumors, strokes, blocked arteries, infections, etc.

[0239] Computing system and method In some embodiments, the above-described systems, tools, and methods of using same enable interactivity and data collection performed by a computing system 13000. FIG. 23 is a block diagram illustrating an embodiment in which the computing system 13000 is in communication with a network 13002 and also with various external computing systems 13004, such as a wearable system 13005, a gateway device 13006, which are themselves in communication with the network 13002. The computing system 13000 may be used to implement the systems and methods described herein. Although the external systems 13004 are shown as grouped, it will be appreciated that each system may be external and / or remotely located from one another.

[0240] In some embodiments, the computing system 13000 includes one or more computing devices, such as, for example, a server, a laptop computer, a mobile device (e.g., a smartphone, a smartwatch, a tablet, a personal digital assistant), a kiosk, an automobile console, or a media player. In one embodiment, the computing device 13000 includes one or more central processing units (CPUs) 13105, each of which may include a conventional or proprietary microprocessor. The computing device 13000 further includes one or more memories 13130, such as a random access memory (RAM) for temporary storage of information, one or more read-only memories (ROM) for permanent storage of information, and one or more mass storage devices 13120, such as a hard drive, a diskette, a solid-state drive, or an optical media storage device. In certain embodiments, a processing device, a cloud server, a server, or a gateway device may be implemented as the computing system 1300. In one embodiment, the modules of the computing system 13000 are connected to the computer using a standard-based bus system. In alternative embodiments, the standards-based bus system may be implemented with, for example, Peripheral Component Interconnect (PCI), MicroChannel, Small Computer computing system Interface (SCSI), Industrial Standard Architecture (ISA), and Extended ISA (EISA) architectures. Additionally, the functionality provided in the components and modules of computing device 13000 may be combined into fewer components and modules or may be separated into additional components and modules.

[0241] Computing device 13000 may be controlled and coordinated by operating system software, such as, for example, iOS, Windows XP, Windows Vista, Windows 7, Windows 8, Windows 10, Windows Server, Embedded Windows, Unix, Linux, Ubuntu Linux, SunOS, Solaris, Blackberry OS, Android, or other operating systems. In a Macintosh system, the operating system may be any available operating system, such as MAC OS X. In other embodiments, computing device 13000 may be controlled by its own operating system. A conventional operating system controls and schedules computer processes for execution, performs memory management, provides file systems, networking, I / O services, provides a user interface, such as a graphical user interface (GUI), among other functions.

[0242] An exemplary computing device 13000 may include one or more I / O interfaces and devices 13110, such as, for example, a touchpad or touch screen, but may also include a keyboard, a mouse, and a printer. In one embodiment, the I / O interfaces and devices 13110 include one or more display devices (such as a touch screen or monitor) that enable visual presentation of data to a user. More specifically, the display devices may provide for display of, for example, GUIs, application software data, and multimedia presentations. The computing system 13000 may also include one or more multimedia devices 13140, such as, for example, a camera, a speaker, a video card, a graphics accelerator, and a microphone.

[0243] The I / O interfaces and devices 13110 in one embodiment of the computing system and application tools may provide a communication interface to various external devices. In one embodiment, the computing device 13000 is electronically coupled to a network 13002, including one or more of a local area network, a wide area network, and / or the Internet, via a wired, wireless, or combination of wired and wireless communication link 13115. The network 13002 may communicate with various sensors, computing devices, and / or other electronic devices via wired or wireless communication links.

[0244] In some embodiments, the filter criteria, signals, and data may be processed by the heartbeat inference module application tool and provided to the computing system 13000 from one or more data sources 13010 over the network 13002 in accordance with the methods and systems described herein. The data sources may include one or more internal and / or external databases, data sources, and physical data stores. The data sources 13010, the external computing system 13004, and the heartbeat interface module 13190 may include databases for storing data (e.g., characteristic data, raw signal data, patient data) in accordance with the above-described systems and methods, databases for storing data processed in accordance with the above-described systems and methods (e.g., data to send to the sensor, data to send to the clinician), and the like. In one embodiment of FIG. 24, the sensor data 14050 may, in some embodiments, store data received from the sensor, data received from the clinician, and the like. The rules database 14060 may, in some embodiments, store data (e.g., instructions, preferences, profiles) that establish threshold parameters for analyzing the characteristic data. In some embodiments, one or more of the databases or data sources may be implemented using a relational database, such as Sybase, Oracle, CodeBase, MySQL, SQLite, and Microsoft® SQL Server, as well as other types of databases, such as, for example, flat file databases, entity-relationship databases, object-oriented databases, NoSQL databases, and / or record-based databases.

[0245] The computing system in one embodiment includes a heart rate interface module 13190 that may be stored in the mass storage device 13120 as executable software code executed by the CPU 13105. The heart rate interface module 13190 may include a feature module 14010, an alternative data module 14020, an inference module 14030, a feedback module 14040, a sensor data database 14050, and a rules database 14060. These modules may include, by way of example only, software components, object-oriented software components, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables, and other components. These modules are also configured to perform the processes disclosed herein, including, in some embodiments, the processes described with respect to FIGS. 10-17.

[0246] Lossless compression and / or digitization In some embodiments of the present disclosure, the system may perform lossless compression on the output of the first set of layers of the neural network before transmitting the data to a remote server to process the data through the second set of layers of the neural network. A lossless compression algorithm may allow the original data to be completely reconstructed from the compressed data. In some embodiments, the system may perform compression that allows the remote server to reconstruct the compressed data with a certain threshold accuracy, thus reducing loss.

[0247] FIG. 25 illustrates an example of lossless compression by an ECG encoder. A patch 2226 can receive a user's ECG data 2232. The monitor (e.g., patch 2226) can encode the data (e.g., via an ECG encoder 2230) through a first set of neural network layers 2222 and send the encoded output to a lossless compressor 2502. The lossless compressor 2502 can compress the data to a smaller dimensionality and send it to a remote server 2228. The remote server 2228 can decompress the data via a lossless decompressor 2504. The remote server 2228 can then process the decompressed data through a second subset of neural network layers 2224 via a classifier 2238.

[0248] In one example, a lossless compression scheme can be used to map between the output encoded traits and shorter length representations. For an 8-bit integer, a value of 0 as a binary representation "00000000" can be the output encoded trait of a first subset of layers of a neural network from the ECG encoder 2230, but a lossless compression scheme can change the 8-bit number representation of "00000000" to something else. A lossless compression scheme can have a mapping between values ​​output from the ECG encoder 2230 and compressed values. For example, a "00000000" 8-bit number representation from the ECG encoder 2230 can be represented by "10", which requires only 2 bits instead of 8 bits. A lossless compression scheme can use a short bit length code for the more common output encoded traits, while a long bit length code is used for the more unusual encoded traits. Advantageously, less data is sent to the remote server. For example, if there are many more 0 values ​​output from the ECG encoder 2230, the lossless compression scheme may map the 0 values ​​to "10." However, if there are very few 433 values ​​output from the ECG encoder 2230, the lossless compression scheme may map the 433 value, which has a binary representation of "110110001," to "5523."

[0249] In some embodiments, the lossless decompressor 2504 may obtain predetermined mapped values. In some embodiments, the lossless decompressor 2504 may obtain mapped values ​​that are static and do not change. The lossless decompressor 2504 may receive the mapped values ​​from the lossless compressor 2502. The remote server 2228 may map the compressed data (e.g., the mapped values) to actual values ​​corresponding to the ECG encoder 2230. For example, if the lossless decompressor 2504 receives the values ​​"10" and "5523", the lossless decompressor 2504 may map these values ​​to "0" and "443" and output these values ​​accordingly. These values ​​may then be processed through a second subset of layers of the neural network via the classifier 2238.

[0250] Advantageously, the dimensionality of data transmitted from the patch to the server in a system that (1) encodes data via a neural network and (2) compresses the encoded data can be smaller than a system (1) with only a neural network encoder (no lossless compression). Furthermore, with lossless compression, more data can be transmitted to the remote server 2228 while maintaining the integrity of the data itself, since the data can be reconstructed via a lossless decompressor. In one example, if there are plots of zero values ​​that are very close to each other, the system can transmit a representative representation of the plot of zero values, rather than having to transmit all the zero values ​​separately. Advantageously, the data transmitted from the patch to the server can use less network throughput and can operate under tighter network constraints without sacrificing the accuracy and performance of the neural network. Furthermore, such compression can be a form of encryption that can further enhance the privacy and security of the data as it travels over various networks from the patch 2226 to the remote server 2228.

[0251] As described herein, some embodiments of the present disclosure train a neural network to compress the data, such as through an ECG encoder 2230, through a first set of layers of the neural network to reduce the dimensionality of the output. The first set of layers of the neural network is trained to output a certain dimensionality. The compressed data is transmitted to a remote device that applies the output of the first set of layers to a second set of layers of the neural network, which may include steps described herein, such as decoding the data or determining physiological characteristics of the data.

[0252] In conventional neural networks, the system uses large numerical representations such as floating-point numbers (e.g., 32-bit data). Floating-point numbers can represent very large and very small values ​​with high precision, up to many decimal places. These conventional systems use such large numerical representations to train the neural network and process data with great accuracy. However, using such large numerical representations makes data storage, processing, and transmission inefficient.

[0253] Some embodiments of the present disclosure mitigate such shortcomings by digitizing the data. FIG. 26 illustrates a digitizer 2602 for performing digitization on the output data of the ECG encoder 2230 after processing a first subset of layers 2222 in the neural network. The patch 2226 can receive the user's ECG data 2232. The monitor (e.g., patch 2226) can encode the data through a first set of layers 2222 of the neural network (e.g., through the ECG encoder 2230) and send the encoded output to the digitizer 2602. The digitizer 2602 can digitize the data (e.g., truncate or round the data) and send it to a remote server 2228. The remote server 2228 can then process the digitized data through a second subset of layers 2224 of the neural network.

[0254] Such quantification can be optimized to have better lossless compression performance while balancing the overall performance of the neural network. The system can quantify the output of the first set of layers of the neural network before sending the data to an external server. In some embodiments, quantifying the data can include rounding (e.g., rounding to the nearest integer), truncation, or reducing the number of bits for data representation.

[0255] In some embodiments, the amount of quantification can be determined to optimize the balance between efficiency and accuracy. More aggressive quantification (e.g., truncating more bits from the data) can reduce the size of the data representation to fewer bits, but can adversely affect the accuracy of the neural network predictions. However, less aggressive quantification (e.g., truncating fewer bits from the data) can increase accuracy, but at the expense of increased storage and transmission requirements of the monitoring device on the person, increased storage requirements due to having to store a larger data representation at both the patch 2226 and the remote server 2228, and increased network transmission of data between the patch 2226 and the remote server 2228.

[0256] More aggressive quantification may introduce errors into the quantified data, which may lead to degradation of the performance of the algorithms that use quantification, as noted herein. Although the neural network will function to some extent despite the errors introduced by quantification, there will be a limit where the errors significantly degrade the performance of the algorithm. Thus, the amount of quantification can be optimized based on the degradation of hardware efficiency and the accuracy of the neural network in classification via the classifier 2238.

[0257] In some embodiments, the neural network is trained to allow for a more aggressive quantification scheme. For example, the neural network can be trained to maximize lossless compression performance. The neural network can be trained by introducing quantification during training to reduce the degradation associated with quantification. In some embodiments, the neural network is first trained without quantification. The trained neural network is then trained again with quantification. Advantageously, the training of the neural network takes quantification into account. This is particularly relevant here when the neural network is split into a first set of layers executed on the patch 2226 and a second set of layers executed on the remote server 2228. Advantageously, the result of such training can include a neural network and / or quantification that can better represent low amplitude signals from quantification. Rather, the neural network can adjust the value of its output (e.g., increase the amplitude of low amplitude signals) so that low amplitude signals are taken into account for identifying characteristics in the signal, such as by the classifier 2236 of the remote server 2228. For example, the neural network can exaggerate low amplitude signals so that the signal reaches a threshold. Thus, the remote server 2228 can register low amplitude signals even though they are processed by quantification. Other systems without such training of the neural network can simply clip low amplitude signals via quantification, and the classifier 2238 may not be able to identify features in such signals.

[0258] In some embodiments, both lossless compression and quantification can be applied in an optimized manner. In some embodiments, the optimal lossless compression instead of the neural network and quantification can be determined by modeling the performance of different lossless compression algorithms or features thereof. For example, the output of the neural network trait can be set to have a distribution centered at 0 with a standard deviation of 1. The system can then model how well a particular lossless compression algorithm performs under various quantification schemes. For example, if the output of a first subset of layers of the neural network is 128 bits / sec, the quantifier can quantize the data and output 100 bits / sec. The optimization can be governed by at least the available network and / or network constraints (such as a setting of bits per second or a bit size of the data). A lossless compression model (e.g., as described herein) can then be applied to find the optimal performance for the identified trait of the neural network in the remote server.

[0259] In some embodiments, the optimization may be governed by at least the encoder architecture. For example, a first encoder design may output 8 bits and a second encoder design may output 4 bits but at twice the sample rate of the first encoder design. In this example, the network constraints may be the same for the first and second encoders. The first encoder design may output fewer features and have more aggressive quantification. However, in some circumstances, the second encoder design may outperform due to its higher rate of feature generation despite its higher quantification rate.

[0260] In some embodiments, the implementation of quantization in training can be applied to the tensorflow implementation. During training, tensorflow can introduce quantization in the forward pass, but the gradients are calculated without quantization. This is done so that quantization errors are propagated through the network, but the gradients are still smooth, allowing weight updates to respond appropriately to the errors.

[0261] 27 illustrates an embodiment having both a digitizer 2602 and lossless compression via a lossless compressor 2502 and a lossless decompressor 2504. In this embodiment, the output of the ECG encoder 2230 is input to the digitizer 2602 after the encoder passes the data through a first subset of layers 2222 of the neural network. The digitizer 2602 digitizes the encoded output data. The lossless compressor 2502 can then compress and transmit 2234 the digitized data to a remote server 2228. The remote server 2228 can then decompress the data 2234 transmitted via the lossless decompressor 2504, and a classifier 2238 can process the decompressed data through a second subset of layers 2224 of the neural network.

[0262] Analysis of QT interval algorithm In some embodiments of the present disclosure, the system can quantify the QT interval, which can include quantifying the time between the Q and T morphological features present in each heart beat, where an increase in the time between these features can indicate a problem with the heart.

[0263] In conventional systems, quantification of the QT interval is typically accomplished by first identifying distinct morphological features and then simply measuring the difference between specific points on the features. However, in systems that do not have access to complete raw ECG data, conventional methods of quantifying the QT interval may not be applicable.

[0264] Described herein are several embodiments that may quantify the QT interval without having access to the complete raw ECG signal. In some embodiments, a remote server may receive the encoded features from the encoder and reconstruct the ECG signal from the encoded features. The remote server may then use signal processing and algorithmic techniques to localize the morphological features.

[0265] In some embodiments, the remote server can train a machine learning algorithm to directly predict the QT interval from a segment of encoded traits that include the QT interval.

[0266] In some embodiments, the remote server can train a machine learning algorithm that predicts the mean QT interval directly from a window of encoded traits that includes one or more QT intervals.

[0267] In some embodiments, the system can generate a template beat representative of several beats. The template beat can be generated by averaging the raw ECG of a fixed window of samples surrounding the detected beat location. The QT interval can be derived from the template beat using algorithmic techniques or machine learning algorithms described herein. Figures 28A and 28B show a process for generating a template beat. For example, Figure 28A illustrates a window of samples over four cycles. Each cycle has a fixed window of samples. The system can determine the number of cycles over a period of time or can set the number of cycles used to generate the template beat. Figure 28B shows generating a template beat from four cycles. In some embodiments, the system determines the samples relative to each other. For example, the system can identify corresponding points in each cycle and average each of the samples together. Thus, the four cycles illustrated on the left side of Figure 28B can be averaged to generate the template beat on the right side.

[0268] In some embodiments, the patch can generate the template beat. In other embodiments, a remote server can generate the template beat from received data from the patch. In some embodiments, the template beat can be processed and identified by processing data through layers of a neural network. For example, the patch can generate the template beat and input the template beat into a neural network described herein.

[0269] In general, the term "module" as used herein refers to logic embodied in hardware or firmware, or a collection of software instructions, possibly having entry and exit points, written in a programming language, such as, for example, Python, Java, Lua, C and / or C++. A software module may be compiled, linked into an executable program, installed in a dynamic link library, or written in an interpreted programming language, such as, for example, BASIC, Perl, or Python. It will be understood that a software module may be called from other modules or from itself, and / or in response to a detected event or interrupt. A software module configured to execute on a computing device may be provided on a computer readable medium, such as a compact disc, digital video disc, flash drive, or any other tangible medium. Such software code may be stored, in part or in whole, on a memory device of an executing computing device, such as the computing system 13000, for execution by the computing device. The software instructions may be embedded in firmware, such as an EPROM. It will be further understood that a hardware module may be composed of connected logic units such as gates and flip-flops, and / or may be composed of programmable units such as programmable gate arrays and processors. The block diagrams disclosed herein may be implemented as modules. The modules described herein may be implemented as software modules, but may also be represented in hardware or firmware. In general, a module as described herein refers to a module of logic that can be combined with other modules or divided into sub-modules, regardless of its physical organization or storage.

[0270] Each of the processes, methods, and algorithms described in the previous sections may be embodied in code modules executed by one or more computer systems or computer processors constituting computer hardware, and may be fully or partially automated. The code modules may be stored in any type of non-transitory computer-readable medium or computer storage device, such as a hard drive, solid-state memory, optical disk, and / or the like. The systems and modules may also be transmitted as a generated data signal (e.g., as part of a carrier wave or other analog or digital propagating signal), in various computer-readable transmission media, including wireless-based and wire / cable-based media, and in various forms (e.g., as part of a single or multiplexed analog signal, or as multiple individual digital packets or frames). The processes and algorithms may be partially or wholly implemented in application-specific circuitry. The results of the disclosed processes and process steps may be stored, persistently or otherwise, in any type of non-transitory computer storage, such as, for example, volatile or non-volatile storage.

[0271] The various features and processes described above may be used independently of each other or may be combined in various ways. All possible combinations and subcombinations are intended to fall within the scope of the present disclosure. In addition, certain method or process blocks may be omitted in some embodiments. The methods and steps described herein are also not limited to any particular order, and blocks or states thereof may be performed in other orders as appropriate. For example, described blocks or states may be performed in orders other than those specifically disclosed, or multiple blocks or states may be combined in a single block or state. Illustrated blocks or states may be performed in serial, parallel, or other ways. Blocks or states may be added or deleted in the disclosed exemplary embodiments. The exemplary systems and components described herein may be configured differently than described. For example, elements may be added, deleted, or rearranged as compared to the disclosed exemplary embodiments.

[0272] In particular, conditional language such as "can," "could," "may," or "might," unless otherwise stated or understood within the context in which it is used, is generally intended to convey that a particular embodiment includes certain features, elements, and / or steps that other embodiments do not include. Thus, such conditional language is generally not intended to imply that the features, elements, and / or steps are somehow required for one or more embodiments, or that one or more embodiments necessarily include logic that, in conjunction with user input or activation, determines whether those features, elements, and / or steps are included in or performed in any particular embodiment. The term "including" means "including but not limited to." The term "or" means "and / or."

[0273] It should be understood that any process descriptions, elements, or blocks in the flow or block diagrams described herein and depicted in the accompanying figures may represent modules, segments, or portions of code that include one or more executable instructions for implementing a particular logical function or step in the process. Alternative implementations are within the scope of the embodiments described herein in which elements or functions may be omitted or may be performed in a different order than shown or discussed, including substantially simultaneously or in reverse order, depending on the functionality involved, as will be appreciated by those skilled in the art.

[0274] All of the methods and processes described above may be at least partially embodied in software code modules executed by one or more computers, and may be partially or fully automated. For example, the methods described herein may be executed by a computing system and / or any other suitable computing device. The methods may be executed on a computing device in response to execution of software instructions or other executable code read from a tangible computer readable medium. A tangible computer readable medium is a data storage device that can contain data readable by a computer system. Examples of computer readable media include read-only memory, random access memory, other volatile or non-volatile memory devices, CD-ROMs, magnetic tapes, flash drives, and optical data storage devices.

[0275] It should be emphasized that many variations and modifications can be made to the above-described embodiments, and that elements thereof are understood to be among other acceptable examples. All such modifications and variations are intended to be included herein within the scope of the present disclosure. The foregoing description details certain embodiments. However, no matter how detailed the foregoing may be in textual description, it will be understood that the systems and methods can be implemented in many ways. For example, features of one embodiment can be used with features of a different embodiment. Also, as described above, it should be noted that the use of a particular term when describing a particular feature or aspect of the systems and methods should not be construed as meaning that the term is redefined herein to be limited to include any particular feature of the feature or aspect of the systems and methods with which the term is associated.

[0276] Various embodiments of physiological monitoring devices, methods, and systems are disclosed herein. These various embodiments may be used alone or in combination, and various changes to the individual features of the embodiments may be modified without departing from the scope of the invention. For example, the order of various method steps may be changed in some cases, and / or one or more optional features may be added to or deleted from the described devices. Thus, the description of the embodiments provided above should not be construed as unnecessarily limiting the scope of the invention as defined in the claims.

[0277] Various modifications may be made to the embodiments described in this disclosure, and the general principles defined herein may be applied to other embodiments without departing from the spirit or scope of the disclosure. Thus, the scope of the disclosure is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the disclosure, the principles and novel features disclosed herein.

[0278] Certain features that are described in the context of separate embodiments herein can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although various features may be described above as acting in a particular combination, and may even be initially claimed as such, one or more features from a claimed combination may in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or a variation of the subcombination.

[0279] Similarly, although operations are depicted in the figures in a particular order, such operations need not be performed in the particular order depicted, or sequentially, to achieve desirable results, nor need all depicted operations be performed. Additionally, the figures may depict one more exemplary process in the form of a flow diagram. However, other operations not shown can be incorporated into the exemplary process depicted in the schematic. For example, one or more additional operations can be performed before, after, simultaneously, or between any of the depicted operations. Additionally, the separation of various system components in the embodiments described above should not be construed as requiring such separation in all embodiments. Additionally, other embodiments are within the scope of the following claims. In some cases, the operations recited in the claims can be performed in a different order and still achieve desirable results.

Claims

1. 1. A monitoring device for a system for inferring a likelihood of an occurrence of arrhythmia in a user, comprising: The monitoring device is a sensor positioned to detect a first set of continuous physiological signals of the user while the monitoring device is engaged with the user; a hardware processor configured to select a neural network from among the plurality of neural networks to generate a data output; the neural network includes a first subset of layers and a second subset of layers; Physiological data corresponding to a first set of continuous physiological signals is processed through an encoder implemented in a hardware processor and having a first subset of layers; The output of the encoder is transmitted to an external computing system and processed through a decoder hosted by the external computing system and having a second subset of layers, and a likelihood of arrhythmia occurrence is inferred. Monitoring equipment.

2. Neural networks are trained to maintain information in a compressed form. The monitoring device of claim 1 .

3. the hardware processor is further configured to obtain a compressed representation of the input to the first subset of layers; The input corresponds to a first set of continuous physiological signals. The monitoring device of claim 1 .

4. A compressed representation of an input is associated with a small size compared to the input, A second subset of layers is configured to resize the input to match the size of the input. The monitoring device of claim 3.

5. the input comprises a filtered representation of a first set of continuous physiological signals; The monitoring device of claim 3.

6. the first subset of layers and the second subset of layers are trained together; The monitoring device of claim 1 .

7. Training the neural network comprises: training a first subset of layers with a first set of training data, the training step including adjusting weights of the first subset of layers; Training a neural network also maintaining weights of a first subset of layers after training on a first set of training data; training a second subset of layers with outputs of the first subset of layers generated in response to a second set of training data provided to the first subset of layers, wherein training the second subset of layers includes adjusting weights of the second subset of layers; Training a neural network also allowing the weights of the first subset of layers to be modified after being trained on the second set of training data; and jointly training the first subset of layers and the second subset of layers using a third set of training data.

7. The monitoring device of claim 6.

8. the second set of training data includes a longer data interval than the first set of training data; 8. The monitoring device of claim 7.

9. the hardware processor is further configured to process data through the first subset of layers. The monitoring device of claim 1 .

10. the number of layers in the first subset of layers is less than the number of layers in the second subset of layers; The monitoring device of claim 1 .

11. The encoder and decoder are trained together as a single neural network. The monitoring device of claim 1 .

12. The monitoring device is implemented as a wearable watch configured to be attached to a user; The monitoring device of claim 1 .

13. The monitoring device is implemented as a wearable device; The monitoring device of claim 1 .

14. the first set of continuous physiological signals includes electrocardiogram data (ECG); The monitoring device of claim 1 .

15. The decoder is hosted by an external computing system configured to infer the likelihood of an arrhythmia occurring. The monitoring device of claim 1 .

16. a second sensor positioned to detect a second set of continuous physiological signals of the user; A likelihood of an arrhythmia occurrence is inferred based on the first set and the second set of continuous physiological signals. The monitoring device of claim 1 .

17. The hardware processor is further configured to provide an output of the first subset of layers to a lossless compressor that compresses the output into a compressed form; The compressed form is transmitted to an external computing system configured to host a second subset of the layers. The monitoring device of claim 1 .

Images