Non-invasive cardiac monitor and method for inferring physiological characteristics of a patient using recorded cardiac data

Abstract

The present invention relates to a non-invasive cardiac monitoring device that records cardiac data to infer physiological characteristics of a person, such as arrhythmias. Some embodiments of the present invention allow for long-term monitoring of physiological signals. Other embodiments allow detected cardiac rhythm signals to be processed partially on the wearable cardiac monitoring device and partially on a remote computing system. Some embodiments include a wearable cardiac monitoring device for long-term attachment to a mammal to extend the detection of cardiac rhythm signals.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Provisional Patent Application No. 62 / 975,626, filed on February 12, 2020, and U.S. Provisional Patent Application No. 63 / 090,951, filed on October 13, 2020, the entire contents of which are incorporated herein by reference. Background Art

[0003] For the purposes of this disclosure, certain aspects, advantages, and novel features of various embodiments are described herein. It should be understood that not necessarily all of these advantages can be realized according to any particular embodiment. Therefore, various embodiments can be implemented or performed in a manner that realizes one advantage or a group of advantages as taught herein without necessarily realizing other advantages as taught or suggested herein. Summary of the invention

[0004] Embodiments described herein relate to a physiological monitoring device that can be worn continuously and comfortably by a human or animal subject for at least one week or more, more typically two to three weeks or more. Although the device can sense and record one or more additional physiological parameters in various optional embodiments, in one embodiment, the device is specifically designed to sense and record heart rhythm (e.g., electrocardiogram (ECG)) data. Such a physiological monitoring device can include many features to facilitate and / or improve the patient experience and make the diagnosis of arrhythmias more accurate and timely.

[0005] Some embodiments include a wearable cardiac monitoring device for monitoring cardiac rhythm signal data of a user, the wearable cardiac monitoring device comprising: a waterproof housing; a surface on the housing configured to be sealably engaged to a mammal; an adhesive on the surface configured to be continuously attached to the mammal for at least 7 days and not removed until monitoring is completed; at least two electrodes permanently disposed within the housing and positioned to detect continuous cardiac rhythm signals of the mammal when the surface is sealably engaged to the mammal; a hardware processor configured to process the detected cardiac rhythm signals through a first subset of computer-executable instructions; and a transmitter configured to transmit the output of the first subset to a computing system configured to process the output through a second subset of computer-executable instructions to infer the possibility of an arrhythmia.

[0006] In some embodiments, the arrhythmia comprises at least one of: ventricular tachycardia, supraventricular tachycardia, ectopy, ventricular fibrillation, or prolonged pause.

[0007] In some embodiments, inferring the likelihood of arrhythmia occurrence includes processing an output of the first subset by the second subset, wherein the first subset processes stored cardiac rhythm signals detected continuously for at least 24 hours.

[0008] In some embodiments, the housing includes a patient trigger configured to be depressed and initiate recording of the instance upon sensing a cardiac event.

[0009] In some embodiments, the electrodes are disposed entirely within the housing.

[0010] In some embodiments, the first subset of computer executable instructions comprises a first subset of layers of a neural network, wherein the second subset of computer executable instructions comprises a second subset of layers of the neural network.

[0011] Some embodiments include a wearable cardiac monitoring device for monitoring biosignal data of a user, the wearable device comprising: an adhesive assembly including a housing and a wing, the wing including electrodes configured to detect cardiac signals from the user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of a neural network; and a transmitter configured to transmit the output of the first subset to a computing system configured to process the output through a second subset of layers of the neural network to infer the likelihood of an arrhythmia occurring.

[0012] In some embodiments, the computing system is further configured to determine an atrial fibrillation burden based on the detected cardiac signals.

[0013] In some embodiments, the atrial fibrillation burden includes the amount of time a user spends in atrial fibrillation over a period of time.

[0014] In some embodiments, the atrial fibrillation burden includes the amount of time the user spends in atrial fibrillation during sleep and during wakefulness.

[0015] Some embodiments include a wearable device for monitoring biosignal data of a user, the wearable device comprising: a component including electrodes configured to detect cardiac signals from the user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of a neural network; and a transmitter configured to transmit outputs of the first subset to a computing system configured to process the outputs through a second subset of layers of the neural network to infer a likelihood of an arrhythmia occurring.

[0016] In some embodiments, the hardware processor is configured to continuously process the detected cardiac signals through a first subset of layers of the neural network while detecting the ECG signals in substantially real time.

[0017] In some embodiments, the output of the first subset of layers of the neural network includes an indication of an R-peak.

[0018] In some embodiments, the computing system is further configured to reconstruct the cardiac signal via a second subset of the layers.

[0019] In some embodiments, outputs of the first subset of layers of the neural network have smaller dimensions than inputs to the first subset of layers of the neural network.

[0020] In some embodiments, the hardware processor is configured to select a neural network from a plurality of neural networks based on the characteristic.

[0021] In some embodiments, the characteristic includes at least one of: a characteristic of the wearable device, a characteristic of a network between the wearable device and the computing system.

[0022] 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, wherein the computing system is configured to decompress the transmitted data; and wherein processing the outputs by a second subset of layers includes processing the decompressed data.

[0023] In some embodiments, the hardware processor is further configured to quantize outputs of a first subset of the layers of the neural network, wherein transmitting the outputs of the first subset to the computing system includes transmitting the outputs.

[0024] In some embodiments, the hardware processor is further configured to compress the quantized output, wherein transmitting the first subset of the output to the computing system includes transmitting the compressed output, wherein the computing system is configured to decompress the transmitted data; and wherein processing the output through the second subset of layers includes processing the decompressed data.

[0025] In some embodiments, transmitting the output of the first subset to the computing system includes transmitting the quantized output.

[0026] Some embodiments include a monitor for inferring the likelihood of a user characteristic, the monitor comprising: a housing; at least two electrodes permanently disposed within the housing and positioned to detect a continuous signal from a surface; a printed circuit board assembly including a hardware processor configured to process the detected signal via a first subset of computer-executable instructions; and a transmitter configured to transmit data output of the first subset of computer-executable instructions to a computing system configured to process the data output via a second subset of computer-executable instructions to infer the likelihood of an event.

[0027] In some embodiments, the dimensions of the data output of the first subset of computer executable instructions transmitted to the computing system are smaller than the dimensions of the data of the detection signals from the at least two electrodes.

[0028] In some embodiments, a data output of the first subset of the computer executable instructions is encrypted, wherein the transmitter transmits the encrypted data output, and wherein the computing system processes the encrypted data output with a second subset of the computer executable instructions.

[0029] In some embodiments, the transmitter transmits the data output to the computing system via a wireless communication channel.

[0030] In some embodiments, the monitor further includes a receiver configured to receive a first updated subset of computer-executable instructions from the computing system and update the first subset of computer-executable instructions to a first updated subset of computer-executable instructions, wherein the hardware processor is further configured to process the signal with the first updated subset of computer-executable instructions.

[0031] In some embodiments, the monitor further includes an accelerometer configured to collect movement data, wherein the computing system is configured to match the movement data with the detected signal to infer the likelihood of the event.

[0032] In some embodiments, the monitor is a cardiac monitor, wherein the continuous signal is a cardiac signal, and wherein the event is an arrhythmic event.

[0033] In some embodiments, the first subset of computer executable instructions comprises a first subset of layers of a neural network and the second subset of computer executable instructions comprises a second subset of layers of a neural network.

[0034] Some embodiments include a monitor for inferring the likelihood of a user characteristic, the monitor comprising: a waterproof housing; a surface on the housing configured to sealably engage with a mammal; an adhesive on the surface configured to remain continuously attached to the mammal for at least seven days and not be removed until monitoring is completed; at least two electrodes permanently disposed within the housing and positioned to detect a continuous cardiac rhythm signal of the mammal when the surface is sealably engaged with the mammal; a printed circuit board assembly including a hardware processor configured to process the detected cardiac rhythm signal via a first subset of computer-executable instructions; and a transmitter configured to transmit data output of the first subset of computer-executable instructions to a computing system configured to process the data output via a second subset of computer-executable instructions to infer the likelihood of an arrhythmia occurring.

[0035] In some embodiments, the arrhythmia comprises at least one of: ventricular tachycardia, supraventricular tachycardia, ectopy, ventricular fibrillation, or prolonged pause.

[0036] In some embodiments, inferring the likelihood of arrhythmia occurrence includes processing an output of the first subset by the second subset, wherein the first subset processes stored cardiac rhythm signals detected continuously for at least 24 hours.

[0037] In some embodiments, the housing includes a patient trigger configured to be depressed and initiate recording of the instance upon sensing a cardiac event.

[0038] Some embodiments include a system for training a neural network to infer a likelihood of a user characteristic, the system including a wearable device, the wearable device including: a component including electrodes configured to detect a cardiac signal from a user; a hardware processor configured to process the detected cardiac signal through a first subset of layers of the neural network; and a transmitter configured to transmit outputs of the first subset to a computing system configured to process the outputs through a second subset of layers of the neural network to infer a likelihood of an arrhythmia occurring, wherein the system is configured to train the neural network by: training the first neural network to recognize a first characteristic by processing first training data for a first time period by the first neural network; freezing weights of the first neural network; training the second neural network to recognize a second characteristic by processing second training data for a second time period by the first and second neural networks, wherein the second time period is longer than the first time period; unfreezing weights of the first neural network; and simultaneously training the first and second neural networks to recognize the second characteristic by processing third training data for a third time period by the first and second neural networks, wherein the third time period is longer than the first time period.

[0039] In some embodiments, outputs of the first subset of layers of the neural network have smaller dimensions than inputs to the first subset of layers of the neural network.

[0040] In some embodiments, the hardware processor is configured to select a neural network from a plurality of neural networks based on the characteristic.

[0041] In some embodiments, the characteristic includes at least one of: a characteristic of the wearable device, a characteristic of a network between the wearable device and the computing system.

[0042] 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, wherein the computing system is configured to decompress the transmitted data; and wherein processing the outputs by a second subset of layers includes processing the decompressed data.

[0043] In some embodiments, the hardware processor is further configured to quantize outputs of a first subset of the layers of the neural network, wherein transmitting the outputs of the first subset to the computing system includes transmitting the outputs.

[0044] Some embodiments include a method for training a neural network to infer the likelihood of a user characteristic, the method comprising: training the first neural network to recognize a first feature by processing first training data for a first time period by the first neural network; freezing weights of the first neural network; training the second neural network to recognize a second feature by processing second training data for a second time period by first and second neural networks, wherein the second time period is longer than the first time period; unfreezing the weights of the first neural network; and simultaneously training the first and second neural networks to recognize the second feature by processing third training data for a third time period by the first and second neural networks, wherein the third time period is longer than the first time period.

[0045] In some embodiments, the second training data is the third training data, and the second time period is the third time period.

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

[0047] In some embodiments, the training is based on available memory of the wearable device.

[0048] In some embodiments, training is based on network availability between the wearable device and the external computing system.

[0049] In some embodiments, the training is based on types of user characteristics.

[0050] In some embodiments, the type of user characteristic includes the occurrence of cardiac arrhythmia.

[0051] In certain embodiments, the computing system is further configured to provide a report that includes the likelihood of an arrhythmia occurring.

[0052] In certain embodiments, the report includes a graph of atrial fibrillation burden over time.

[0053] In certain embodiments, the report includes an indication that atrial fibrillation is present.

[0054] In certain embodiments, the report includes a monitoring period of at least 14 days.

[0055] In some embodiments, the computing system is a server or a gateway.

[0056] In certain embodiments, the computing system is a smart phone.

[0057] In certain embodiments, the computing system communicates with the transmitter through a smartphone intermediary.

[0058] In certain embodiments, the wearable device is configured to collect the secondary signal, and the hardware processor is configured to process the secondary signal through a first subset of layers of the neural network.

[0059] In some embodiments, the wearable device further includes an accelerometer configured to measure movement of the user, and wherein the secondary signal is accelerometer data.

[0060] In certain embodiments, the computing system is further configured to determine an atrial fibrillation burden, and the atrial fibrillation burden comprises an amount of time spent in atrial fibrillation during movement of the user.

[0061] In some embodiments, the movement of the user includes a first degree of movement and a second degree of movement.

[0062] In some embodiments, the secondary signal is electrode contact quality data.

[0063] In some embodiments, the electrodes and hardware processor are contained within a chest strap.

[0064] In some embodiments, the electrodes and hardware processor are contained within a watch configured to be worn on a person's wrist.

[0065] In some embodiments, the electrodes and hardware processor are contained within a wearable fitness band.

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

[0067] In certain embodiments, the arrhythmia comprises supraventricular tachycardia.

[0068] In certain embodiments, the arrhythmia comprises an ectopic rhythm.

[0069] In certain embodiments, the arrhythmia comprises ventricular fibrillation.

[0070] In certain embodiments, the arrhythmia comprises a prolonged pause.

[0071] In certain embodiments, the hardware processor is configured to be removed from the wearable device and to be modified while removed from the wearable device.

[0072] In some embodiments, quantizing the output includes rounding or truncating values ​​in the output.

[0073] In some embodiments, the quantization amount of the output is based on lossless compression performance.

[0074] In some embodiments, the quantization amount of the output is based on efficiency of at least one of: processing power, storage or network usage, and accuracy of the neural network.

[0075] In certain embodiments, the neural network is trained by applying quantization to outputs of a first subset of layers of the neural network.

[0076] In certain embodiments, a neural network is trained by applying lossless compression to outputs of a first subset of layers of the neural network.

[0077] In certain embodiments, a neural network is trained by applying quantization and lossless compression to outputs of a first subset of layers of the neural network.

[0078] In some embodiments, a neural network is trained without quantization and lossless compression to generate a first neural network, and the first neural network is trained by applying quantization and lossless compression to outputs of a first subset of layers of the neural network to generate a second neural network.

[0079] Some embodiments include a wearable device for monitoring data, the wearable device comprising: a sensor configured to detect a cardiac signal from a user; a hardware processor configured to process the detected cardiac signal through a first subset of layers of a neural network; and a transmitter configured to transmit an output of the first subset to a computing system configured to process the output through a second subset of layers of the neural network to infer a likelihood of an arrhythmia occurring.

[0080] Some embodiments include a computing system for estimating the likelihood of an arrhythmia occurring, the computing system comprising: one or more first hardware processors configured to receive data from a wearable device, wherein the wearable device includes a sensor configured to detect cardiac signals from a user; and one or more second hardware processors configured to process the detected cardiac signals through a first subset of layers of a neural network, process the received data through a second subset of layers of the neural network, and receive an inference of the likelihood of an arrhythmia occurring from an output of the second subset of layers.

[0081] Some embodiments include a method for monitoring data, the method comprising: detecting a cardiac signal from a user; processing the detected cardiac signal through a first subset of layers of a neural network; and transmitting an output of the first subset to a computing system, which processes the output through a second subset of layers of the neural network to infer a likelihood of an arrhythmia occurring.

[0082] Some embodiments include a non-transitory computer storage medium storing computer executable instructions that, when executed by a processor, cause the processor to perform a method of detecting a cardiac signal from a user; processing the detected cardiac signal through a first subset of layers of a neural network; and transmitting an output of the first subset to a computing system that processes the output through a second subset of layers of the neural network to infer a likelihood of an arrhythmia occurring.

[0083] Some embodiments include a method for training a neural network to infer the likelihood of a user characteristic, the method comprising: training the first neural network to recognize a first feature by processing first training data for a first time period by the first neural network; freezing weights of the first neural network; training the second neural network to recognize a second feature by processing second training data for a second time period by first and second neural networks, wherein the second time period is longer than the first time period; unfreezing the weights of the first neural network; and simultaneously training the first and second neural networks to recognize the second feature by processing third training data for a third time period by the first and second neural networks, wherein the third time period is longer than the first time period.

[0084] In some embodiments, the second training data is the third training data, and the second time period is the third time period.

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

[0086] In certain embodiments, the training is based on the available memory of the wearable device.

[0087] In certain embodiments, training is based on network availability between the wearable device and the external computing system.

[0088] In some embodiments, the training is based on types of user characteristics.

[0089] In some embodiments, the type of user characteristic includes the occurrence of cardiac arrhythmia.

[0090] Some embodiments include a system for training a neural network to infer the likelihood of a user characteristic, the system comprising a wearable cardiac monitoring device, comprising: a waterproof housing; a surface on the housing configured to sealably engage with a mammal; an adhesive on the surface configured to continuously adhere to the mammal for at least 7 days and not be removed until monitoring is completed; at least two electrodes permanently disposed within the housing and positioned to detect a continuous cardiac rhythm signal of the mammal when the surface is sealably engaged with the mammal; a hardware processor configured to process the detected cardiac rhythm signal through a first subset of layers of the neural network; and a transmitter configured to transmit an output of the first subset to a computing system, the computing system The system is configured to process the outputs by a second subset of the layers of the neural network to infer the likelihood of an arrhythmia occurring, wherein the system is configured to train the neural networks by: training the first neural network to recognize a first feature by processing first training data of a first time period by the first neural network; freezing weights of the first neural network; training the second neural network to recognize a second feature by processing second training data of a second time period by the first and second neural networks, wherein the second time period is longer than the first time period; unfreezing weights of the first neural network; and simultaneously training the first and second neural networks to recognize a second feature by processing third training data of a third time period by the first and second neural networks, wherein the third time period is longer than the first time period.

[0091] Some embodiments include a wearable cardiac monitoring device for monitoring biosignal data of a user, the wearable device comprising: an adhesive assembly including a housing and a wing, the wing including electrodes configured to detect cardiac signals from the user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of a neural network; and a transmitter configured to transmit the output of the first subset to a computing system configured to process the output through a second subset of layers of the neural network to infer a QT interval.

[0092] In certain embodiments, the computing system is further configured to determine an atrial fibrillation burden based on the detected cardiac signals.

[0093] In certain embodiments, the atrial fibrillation burden includes the amount of time a user spends in atrial fibrillation over a period of time.

[0094] In certain embodiments, the atrial fibrillation burden includes the amount of time the user spends in atrial fibrillation during sleep and during wakefulness.

[0095] In certain embodiments, the hardware processor is configured to continuously process the detected cardiac signals through a first subset of layers of the neural network while detecting the ECG signals in substantially real time.

[0096] In some embodiments, the output of the first subset of the layers of the neural network includes an indication of an R-peak.

[0097] In certain embodiments, the computing system is further configured to reconstruct the cardiac signal via a second subset of the layers.

[0098] Some embodiments include a system for training a neural network to infer a likelihood of a user characteristic, the system including a wearable device, the wearable device including: a component including electrodes configured to detect a cardiac signal from a user; a hardware processor configured to process the detected cardiac signal through a first subset of layers of the neural network; and a transmitter configured to transmit outputs of the first subset to a computing system configured to process the outputs through a second subset of layers of the neural network to infer a likelihood of an arrhythmia occurring, wherein the system is configured to train the neural network by: training the first neural network to recognize a first characteristic by processing first training data for a first time period by the first neural network; freezing weights of the first neural network; training the second neural network to recognize a second characteristic by processing second training data for a second time period by the first and second neural networks, wherein the second time period is longer than the first time period; unfreezing weights of the first neural network; and simultaneously training the first and second neural networks to recognize the second characteristic by processing third training data for a third time period by the first and second neural networks, wherein the third time period is longer than the first time period.

[0099] In some embodiments, the outputs of the first subset of the layers of the neural network have smaller dimensions than the inputs of the first subset of the layers of the neural network.

[0100] In some embodiments, the hardware processor is configured to select a neural network from a plurality of neural networks based on the characteristic.

[0101] In some embodiments, the characteristic includes at least one of: a characteristic of the wearable device, a characteristic of a network between the wearable device and the computing system.

[0102] 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, wherein the computing system is configured to decompress the transmitted data; and wherein processing the outputs through a second subset of layers includes processing the decompressed data.

[0103] In some embodiments, the hardware processor is further configured to quantize 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 outputs.

[0104] In certain embodiments, the computing system infers the QT interval by reconstructing the ECG signal from the outputs of the second subset of layers of the neural network.

[0105] In certain embodiments, the computing system infers the QT interval based on an average QT interval from a window of encoding features that includes a plurality of QT intervals.

[0106] Some embodiments include a wearable cardiac monitoring device for monitoring biosignal data of a user, the wearable device comprising: an adhesive assembly including a housing and a wing, the wing including electrodes configured to detect cardiac signals from the user; a hardware processor configured to process the detected cardiac signals through a first subset of layers of a neural network; and a transmitter configured to transmit an output of the first subset to a computing system configured to generate a template beat based on a plurality of beats over a period of time.

[0107] Some embodiments include a wearable device for monitoring biosignal data of a user, the wearable device comprising: a component including electrodes configured to detect a cardiac signal from the user; a hardware processor configured to process the detected cardiac signal through a first subset of layers of a neural network, the output of the first subset of layers of the neural network including an RR peak interval, wherein the RR peak interval includes a duration between two consecutive R peaks; and a transmitter configured to transmit the output of the first subset to a computing system, the computing system configured to process the RR peak interval in the output of the first subset of layers of the neural network through a second subset of layers of the neural network to infer the possibility of an arrhythmia.

[0108] In certain embodiments, a first subset of layers of the neural network generates a sequence of RR peak intervals by extracting a subsequence of RR intervals using overlapping sliding windows and moving the sliding windows over the detected cardiac signal.

[0109] These and other aspects and embodiments of the invention are described in more detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] Figure 1A and Figure 1B 1 and 2 are respectively a perspective view and an exploded cross-sectional view of a physiological monitoring device according to one embodiment.

[0111] Figure 2A and Figure 2B 2 are top and bottom perspective views, respectively, of a printed circuit board assembly of a physiological monitoring device according to one embodiment.

[0112] Figure 3A , Figure 3B , Figure 3C , Figure 3D and Figure 3E1 is a perspective view and an exploded view of a flexible body and a gasket of a physiological monitoring device according to one embodiment.

[0113] Figure 4 is an exploded view of a rigid housing of a physiological monitoring device according to one embodiment.

[0114] Figure 5A and Figure 5B A perspective view of a battery holder for a physiological monitoring device according to one embodiment is provided.

[0115] Fig. 6A and Figure 6B is a cross-sectional view of a physiological monitoring device according to one embodiment

[0116] Figure 7 is an exploded view of a physiological monitoring device including a number of optional features according to one embodiment.

[0117] Fig. 8A and Figure 8B is a perspective view of two people wearing a physiological monitoring device according to one embodiment, showing how the device flexes to accommodate body movement and position.

[0118] Fig.9A , Fig. 9B , Fig. 9C , Fig.9D , Fig.9E and Fig.9F Various steps are shown for applying a physiological monitor to a patient's body according to one embodiment.

[0119] Fig.10 A schematic diagram showing an embodiment of a heart rhythm inference service is shown.

[0120] Fig.11 is a schematic diagram of an embodiment of a system for extracting and transmitting data features from a physiological monitor.

[0121] Fig.12 is a schematic diagram of an embodiment of a system for extracting and transmitting data features from a physiological monitor using a transmission device.

[0122] Fig.13 is a schematic diagram of an embodiment of a physiological monitoring system utilizing additional data channels.

[0123] Fig.14 is a schematic diagram of an embodiment of a physiological monitoring system including a data filter.

[0124] Fig.15 is a schematic diagram of an embodiment of a wearable device system.

[0125] Fig.16 is a schematic diagram of an embodiment of a symptom propagation system.

[0126] Fig.17 is a schematic diagram of an embodiment of an asymptomatic transmission system.

[0127] Fig.18A is a graph showing an embodiment of readings from an ECG strip corresponding to a NSR heart rhythm.

[0128] Fig.18B is a graph showing an embodiment of readings from an ECG strip corresponding to a NSR heart rhythm.

[0129] Fig.19 It is a high-level architecture of the burden prediction model.

[0130] Fig. 20 yes Fig.19 An embodiment of a high-level architecture.

[0131] Fig.21 yes Fig.19 Embodiment of the feature extraction model.

[0132] Fig.22A is a schematic diagram of an embodiment of a system for predicting rhythm annotation using neural network encoding.

[0133] Fig. 22B is a schematic diagram of an embodiment of a first and a second subset of layers within a single neural network.

[0134] Fig. 22C is a schematic diagram of an embodiment of a system for processing a subset of layers of a neural network on a tile.

[0135] Fig.22D is a schematic diagram of an embodiment for transmitting outputs of a first subset of layers to a server.

[0136] Fig.22E is a schematic diagram of an embodiment of designing (and / or training) a neural network.

[0137] Fig.22F is a schematic diagram of an embodiment of designing, training, and / or selecting a neural network based on a user's current activity.

[0138] Fig.23 is a schematic diagram of an embodiment of a computer network system.

[0139] Fig.24 is a schematic diagram of an embodiment of a programming and distribution module.

[0140] Fig.25 An implementation using an ECG encoder for lossless compression is shown.

[0141] Fig.26 An embodiment of a quantizer that performs quantization on output data of an ECG encoder after processing a first subset of layers in a neural network is shown.

[0142] Fig. 27 An embodiment is shown having both a quantizer and lossless compression via a lossless compressor and a lossless decompressor.

[0143] Fig.28A and Fig.28B An embodiment of a process for generating template jumps is shown. DETAILED DESCRIPTION

[0144] The following description is directed to a number of different embodiments. However, the described embodiments may be implemented and / or varied in many different ways. For example, the described embodiments may be implemented in any suitable device, apparatus, or system to monitor any of a number of physiological parameters. For example, the following discussion focuses primarily on long-term, patch-based cardiac rhythm monitoring devices. In an optional embodiment, the physiological monitoring device may be used for, for example, pulse oximetry and diagnosis of obstructive sleep apnea. The method of using the physiological monitoring device may also vary. In some cases, the device may be worn for a week or less, while in other cases, the device may be worn for at least 7 days and / or more than 7 days, such as between 14 and 21 days or even longer.

[0145] Since abnormal heart rhythms or arrhythmias are often caused by other, less serious causes, a key challenge is to determine when any of these symptoms is caused by an arrhythmia. Typically, arrhythmias occur infrequently and / or sporadically, making rapid and reliable diagnosis difficult. Currently, heart rhythm monitoring is primarily accomplished using devices (e.g., Holter monitors) that use short-term (less than 1 day) electrodes fixed to the chest. Wires connect the electrodes to a recording device, which is typically worn on a belt. The electrodes need to be changed daily, and the wires are cumbersome. The device also has limited memory and recording time. Wearing the device interferes with the patient's activities and often prevents the performance of certain activities, such as bathing, while being monitored. In addition, Holter monitors are capital equipment with limited availability, a situation that often results in supply restrictions and corresponding testing delays. These limitations severely hinder the diagnostic utility of the device, the compliance of patients using the device, and the possibility of capturing all important information. Lack of compliance and device shortcomings often result in the need for additional equipment, follow-up monitoring, or other tests to make a correct diagnosis.

[0146] Current methods of correlating symptoms with the onset of arrhythmias, including the use of cardiac rhythm monitoring devices such as Holter monitors and cardiac event recorders, are often inadequate for accurate diagnosis. In fact, Holter monitors have been shown to fail to lead to a diagnosis in up to 90% of cases (“Assessment of the Diagnostic Value of 24-Hour Ambulatory Electrocardiographic Monitoring”, DE Ward et al., Biotelemetry Patient Monitoring, Vol. 7, 1980).

[0147] Furthermore, the medical process of actually obtaining a cardiac rhythm monitoring device and beginning monitoring is often very complex. Ordering, tracking, monitoring, retrieving, and analyzing data from such a monitoring device often involves many steps. In most cases, the cardiac monitoring devices used today are ordered by a cardiologist or cardiac electrophysiologist (EP) rather than by the patient's primary care physician (PCP). This is important because the PCP is often the first physician to see the patient and determine that the patient's symptoms may be caused by an arrhythmia. After the patient meets with the PCP, the PCP will schedule an appointment for the patient to see a cardiologist or EP. This appointment is often weeks away from the initial PCP visit, which in itself can lead to delays in potential diagnosis and increase the likelihood that an arrhythmia will develop and go undiagnosed. When the patient finally sees a cardiologist or EP, a cardiac rhythm monitoring device is often ordered. Monitoring can last from 24 to 48 hours (Holter monitors) or up to a month (cardiac event monitors or mobile telemetry devices). Once monitoring is complete, the patient often has to return the device to the clinic, which in itself can be an inconvenience. After the data is processed by the monitoring company or a field technician at the hospital or office, a report is ultimately sent to the cardiologist or EP for analysis. This complex process results in fewer patients receiving heart rhythm monitoring than would be ideal.

[0148] To address some of these issues 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 the device is Patches. Various embodiments are also described, for example, in U.S. Pat. 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. Typically, the physiological patch-based monitors described in the above references fit comfortably on the patient's chest and are designed to be worn for at least one week, typically two to three weeks. While the device is worn, the monitor continuously detects and records heart rhythm signal data, which can then be used for processing and analysis.

[0149] Physiological monitoring device

[0150] refer to Figure 1A and Figure 1B , provides a stereoscopic view and exploded cross-sectional view of an embodiment of a physiological monitoring device 100. Figure 1A As shown, the physiological monitoring device 100 may include a flexible body 110 coupled to a waterproof rigid housing 115, and a hinge portion 132. The flexible body 110 (which may be referred to as a "flexible substrate" or "flexible structure") generally includes two wings 130, 131 extending laterally from the rigid housing 115 and two flexible electrode traces 311, 312, each of which is embedded in one of the wings 130, 131. Each electrode trace 311, 312 is coupled to a flexible electrode (at the bottom surface of the flexible body 110) on the bottom surface of the flexible body 110. Figure 1A The electrodes are configured to sense heart rhythm signals from a patient to whom the monitoring device 100 is attached. The electrode traces 311, 312 then transmit these signals to an electronic device (in the rigid housing 115) mounted therein. Figure 1A The rigid housing 115 also typically contains a power source, such as one or more batteries.

[0151] Reference now Figure 1B , a partially exploded view of the physiological monitoring device 100 shows in more detail the components that make up and are contained within the rigid housing 115. In this embodiment, the rigid housing 115 includes an upper housing member 140 that is detachably coupled to a lower housing member 145. An upper gasket 370 and a lower gasket 360 are sandwiched between the upper housing member 140 and the lower housing member 145 (in Figure 1B370, 360). The gaskets 370, 360 help to waterproof the rigid housing member 115 when assembled. Multiple components of the monitoring device 100 can be housed between the upper housing member 140 and the lower housing member 145. For example, in one embodiment, the housing 115 can include 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 located within the housing 115 to contact the electrode traces 311, 312 and the batteries 160. In various embodiments, one or more additional components can be contained within or attached to the rigid housing 115. Some of these optional components are further described below with reference to the additional figures.

[0152] According to various optional embodiments, the battery holder 150 can accommodate two batteries (as in the illustrated embodiment), one battery, or more than two batteries. In other optional embodiments, other power sources can be used. In the illustrated embodiment, the battery holder 150 includes a plurality of retaining tabs 153 to hold the battery 160 in the battery holder 150. In addition, the battery holder 150 includes a plurality of legs 152 to establish the correct spacing of the battery 160 from the surface of the PCBA 120 and to ensure proper contact with the spring contacts 235 and 236. In this embodiment, spring contacts 235 and 236 are used instead of welding the battery 160 to the PCBA 120. Although welding can be used in optional embodiments, one advantage of the spring contacts 235 and 236 is that they allow the battery 160 to be removed from the PCBA 120 and the holder 150 without damaging any of these components, thereby allowing multiple reuses of both. Eliminating the welded connection also simplifies and speeds up the assembly and disassembly of the monitoring device 100. In some embodiments, the upper housing member 140 can act as a patient event trigger.

[0153] Reference now Figure 2A and Figure 2B In the embodiment of FIG. 1 , the printed circuit board assembly 120 (or PCBA) can 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 can be configured to relay a signal from a patient trigger (e.g., the upper housing member 140 described above) to the PCBA 120. For example, the patient trigger input 210 can be a PCB switch or button that responds to pressure from a patient trigger (e.g., the upper surface of the upper housing portion 140).

[0154] Reference now Figure 3A and Figure 3B The embodiment of FIG. 1 shows the flexible body 110 in more detail. Figure 3AAs shown, the flexible body 110 may include wings 130, 131, a thin border 133 (or "rim" or "edge") surrounding 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 the rigid housing 115.

[0155] like Figure 3C and Figure 3D As shown in the embodiment of the present invention, when the device 100 or the enlarged device portion 101 is assembled, the ECG circuit interface portion 313 is in physical contact with the spring contact 237 and provides electrical communication with the PCBA 120. The electrode interface portion 310 contacts the hydrogel electrode 350. Therefore, the electrode traces 311, 312 transmit the heart rhythm signal (and / or other physiological data in various embodiments) from the electrode 350 to the PCBA 120.

[0156] Figure 3E Another embodiment is depicted where the upper gasket 370 includes a tab 371 that protrudes from the contour of the top housing 140 but remains adhered to the upper base plate 300. The tab 371 covers a portion of the electrode traces 311, 312 and provides strain relief for the traces at the highest stress point where the flexible body meets the rigid housing.

[0157] Reference now Figure 4 , the upper housing member 140 and the lower housing member 145 of the rigid housing 115 are shown in more detail. The upper housing member 140 and the lower housing member 145 can be configured to form a waterproof housing for accommodating the PCBA 120, the battery holder 150, the battery 160, and any other components contained within the rigid housing 115 when coupled together with the gaskets 360, 370 therebetween. The housing members 140, 145 can be made of any suitable material (e.g., waterproof plastic) to protect the internal components.

[0158] Reference now Figure 5A 1, the battery holder 150 is shown in more detail. The battery holder 150 can be made of plastic or other suitable material, is configured to be mounted to the PCBA 120 and then attached to the rigid housing 115, and is capable of holding two batteries 160 ( Figure 1B The plurality of protrusions 152 provide a stable platform for the battery 160 to position the battery 160 at a fixed distance above the surface of the PCBA 120, avoiding unwanted contact with sensitive electronic components, while providing sufficient compression of the spring contacts 235 ( Figure 5B ).

[0159] Reference now Fig. 6A and Figure 6B In an embodiment of the present invention, the physiological monitoring device 100 is shown in a side view cross-section. Fig. 6A 6 , 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 electrodes 350. The 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 they are not shown in FIG. 6 . The flexible body 110 forms two wings 130, 131 extending to either side of the housing 115, and a border 133 surrounding 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 such that it sandwiches a portion of the flexible body 110 and provides a waterproof sealed compartment for the PCBA 120. The upper housing member 140 may include an internal trigger member 430, and the PCBA may include a patient trigger member 210. As previously described, the lower housing member 145 may include a plurality of dimples 450 or recesses to improve the comfort of the monitoring device 100. In certain embodiments, additional mechanisms may be used to reduce and prevent unwanted bending of the PCBA 120. Such mechanisms may include: Figure 6B shown.

[0160] refer to Figure 7 In some embodiments, the physiological monitoring device 100 may include one or more additional optional features. For example, in one embodiment, the monitoring device 100 may include a removable liner 810, a top label 820, a device identifier 830, and a bottom label 840.

[0161] Reference now Fig. 8A and Figure 8B In an embodiment of the present invention, the physiological monitoring device 100 generally includes a hinge portion 132 located at or near the junction of each wing 130, 131 with the rigid housing 115. In addition, each wing 130, 131 is generally adhered to the patient via an adhesive layer 340, while 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 movement and change of patient position. In other words, when the patient's chest contracts, the rigid housing bounces or floats on the skin, thereby minimizing stress on the device 100, improving comfort, and reducing the tendency of the wings 130, 131 to peel off the skin.

[0162] Reference now Figure 9A-9F , describes one embodiment of a method of applying a physiological monitoring device 100 to the skin of a human subject. In this embodiment, Fig.9A Prior to the first step shown, 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 then cleaning and / or scrubbing the shaved portion. Fig.9A As shown, once the patient's skin is 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, thereby exposing the adhesive layer 340. Fig. 9B As shown, the next step may be to apply the device 100 to the skin so that the adhesive layer 340 adheres to the skin at the desired location. Fig. 9C , after the device 100 has been applied to the skin, pressure may be applied to the flexible body 110 to press it down onto the chest to help ensure that the device 100 adheres to the skin.

[0163] In the next step, see Fig.9D , remove (e.g., peel off) the liner 810 from the top surface of the flexible body 110. Fig.9E As shown, once the liner 810 is removed, pressure can again be applied to the flexible body 110 to help ensure that it adheres to the skin. Fig.9F As shown, the upper housing member 140 can be pressed to turn on the physiological monitoring device 140. The method described is only one embodiment. In alternative embodiments, one or more steps can be skipped and / or one or more additional steps can be added.

[0164] The physiological monitors described in the above sections and elsewhere in the specification can be further combined with methods and systems for improving data processing and transmission of data collected from the monitors. In addition, the methods and systems described below can improve the performance of the monitors by enabling timely transmission of clinical information while maintaining high patient compliance and ease of use of the above-mentioned monitors. For example, the methods and systems for data processing and transmission described in this section or elsewhere in the specification can be used to extend the battery life of the monitors, improve the accuracy of the monitors, and / or provide other improvements and advantages as described in this section or elsewhere in the specification.

[0165] Device Monitoring and Clinical Analysis Platform

[0166] Reference below Figures 10 to 17The systems and methods described in detail in the embodiments can selectively extract, transmit and analyze electrocardiogram signal data and other physiological data from a wearable physiological monitor, as described above in connection with Figures 1 to 9. The systems and methods described below can improve the performance of a wearable physiological monitor that simultaneously records and transmits data in a variety of ways. For example, the selective transmission of extracted data allows for reduced power consumption because the wearable patch does not need to transmit all recorded data. By sending the extracted data, most of the analysis can be performed outside the wearable device without the need for a complete onboard heart rhythm analysis, which can also consume a lot of power, thereby shortening battery life. In addition, remote analysis without the inherent power limitations of the wearable device can improve the sensitivity and accuracy of the data analysis. Reducing power consumption helps improve patient compliance because it extends the time interval between device replacement, battery replacement, or battery charging during the monitoring cycle, or even eliminates the need for such. By reducing battery consumption, longer monitoring times can be achieved without replacing the device, for example, at least one week, at least two weeks, at least three weeks, or more than three weeks.

[0167] Fig.10 A general overview of an embodiment of a system 900 for inferring cardiac rhythm information from an RR interval time series 902 is depicted, which can 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 the time intervals between consecutive heartbeats. The RR interval time series 902 data may be extracted or received from a dedicated heart rate monitor (e.g., a heart rate chest strap or a heart rate watch or a wearable health or fitness device 906, 908 containing heart rate sensing functionality). Alternatively, the RR interval time series 902 may be derived from a wearable patch designed to measure an ECG signal 904 (e.g., by locating the R peak in the ECG using a QRS detection algorithm). In addition, the RR interval time series 902 may be estimated from an alternative physiological signal (e.g., a signal obtained from a photoplethysmogram (PPG)). In this case, 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.

[0168] In some embodiments, the heart rhythm inference system may accept multiple RR interval time series 918 measured from a given user's device in addition to the individual RR interval time series 902. In certain embodiments, in addition to the RR interval time series data, the heart rhythm inference system 910 may also accept additional data sources, generally described as alternative sensor channels, to improve the accuracy and / or value of the inference results.

[0169] Current wearable sensors (e.g., iRhythm ZioPatch TM904) and the further description above with respect to FIGS. 1 to 9 is capable of recording single-lead electrocardiogram (ECG) signals and lasting up to two weeks on a single charge.

[0170] Extraction, transfer and processing systems

[0171] Fig.11 is a schematic diagram of an embodiment of a system and method 1000 for a wearable medical sensor 1002 having transmission capabilities, which is similar to the above description of Fig.10 In some embodiments, the sensor 1002 can be any type of sensor or monitor described in this section or elsewhere in the specification that continuously senses an ECG or similar biosignal 1004 and continuously records an ECG or similar biosignal 1004. The collected signal 1004 can then be continuously extracted into one or more features 1006 representing example features A, B, and C. The features of the ECG or similar biosignal are extracted to facilitate remote analysis of the signal 1004.

[0172] Once feature extraction as described above is complete, the various features 1008 may then be transmitted 1010 to a processing device / server 1012. Features 1008 (as well as alternative sensor channel data and / or features as described below) are transmitted 1010 at regular intervals to a processor 1012 that is not a physical part of the sensor 1002.

[0173] In some embodiments, the transmitted features 1014 are processed by a remote processor using the data features 1014 to perform analysis via a rhythm inference system 1016, which analyzes and identifies segments / locations 1018 that may include arrhythmias. For example, the types of arrhythmias and ectopics that may be identified may include: pauses, 2nd or 3rd degree AVB, complete heart block, SVT, AF, VT, VF, bigrams, triplets, and / or ectopics. The determined confidence level may be included in the identification of the rhythm.

[0174] The identified arrhythmia location 1018 is then transmitted 1020 back to the sensor 1002. The transmission 1020 back to the sensor can be accomplished by any communication protocol / technology described in this section or elsewhere in the specification, such as via Bluetooth. The sensor then reads the transmitted identification location 1022 and accesses 1024 a memory area corresponding to the ECG that transmitted the identification location 1022. In some embodiments, the sensor applies additional analysis of the identified segment to further establish the confidence of the arrhythmia identification. This further rhythm confidence determination step 1026 allows for increased positive predictability prior to the power-consuming transmission step. In an embodiment, if the confidence exceeds a defined threshold, the data segment is transmitted. If the confidence exceeds a threshold as described above, the sensor 1002 can transmit the requested ECG segment 1028 to the processing device via any transmission method described in this section or elsewhere in the specification.

[0175] Fig.12 is a schematic diagram of an embodiment of a system and method 2000 for a wearable ECG and / or medical sensor 2002 having transmission capabilities, which is very similar to the above description of Fig.11 System and / or method 1000 is described. Fig.12 system and Fig.11 The system differs in that it includes an auxiliary transmission device 2004.

[0176] Fig.13 is a schematic diagram of an embodiment of a system and method 3000 for a wearable ECG and / or medical sensor 3002 having transmission capabilities, which is very similar to the above description of Fig.11 and Fig.12 The systems and / or methods described herein are: Fig.13 and Fig.11 and Fig.12 The difference is Fig.13 Alternative sensor channels 3004, 3006 are shown that produce alternative outputs and / or extraction 3008 of features 3010. Collection of other data channels can be used to further enhance the features extracted by the ECG. Data from alternative sensor channels can be sent in their entirety, or specific features 3010 of the data channels can be extracted 3008.

[0177] Fig.14 is a schematic diagram of an embodiment of a system and method 4000 for a wearable ECG and / or medical sensor 4002 having transmission capabilities, which is very similar to the above description of Figures 11 to 13 The systems and / or methods described herein are: Fig.14 and Figures 11 to 13 Different because Fig.14Embodiments of include additional data filters. In some embodiments, processing device 4004 may also filter cardiac rhythms 4006 identified by cardiac rhythm inference system 4008 by applying filtering criteria that may be derived from a plurality of sources.

[0178] Fig.15 is a schematic diagram of an embodiment of a system 5000 for a consumer wearable device without full ECG detection, which is similar to Figures 10 to 14 Sensor 5002 need not be a medical grade ECG sensor, but rather only allows for the detection of a heartbeat. In an embodiment, sensor 5002 may continuously sense a channel of data from which the heartbeat location can be derived. Possible data sources include: PPG (optionally with multiple channels for increased accuracy), bioimpedance, and Figures 10 to 14 ECG is not fully implemented due to insufficient signal quality compared to the sensor. Figures 10 to 14 A device that can extract features from a signal, such as: R-peak position, R-peak overflow flag, saturation flag, respiratory rate, P / T wave position, R-peak amplitude (or proxy) or ECG signal amplitude proxy.

[0179] Consumer device system 5000 without full ECG sensing advantageously uses consumer available heart rate sensors to enable arrhythmia analysis, thereby reducing costs and increasing device availability. This may therefore make arrhythmia screening available to a greater population, including via over-the-counter screening.

[0180] Fig.16 6000 is a schematic diagram of an embodiment of an ECG monitoring system 6000 with symptom transmission. Such a system would involve a wearable ECG sensor, similar to that described with respect to FIGS. 1 to 60 Fig.14 As described above, this sensor continuously senses and records the ECG. Each symptom triggered by the patient can initiate the transmission of an ECG data strip.

[0181] Fig.17 is a schematic diagram of an embodiment of an ECG monitoring system 7000 having both symptomatic and asymptomatic transmissions. In some embodiments, each asymptomatic trigger initiates the transmission of an ECG data strip, as described above.

[0182] A system for estimating burden from R-peak sequences using neural networks

[0183] Some embodiments disclose a wearable device. The wearable device may include a sensor for detecting a biosignal (e.g., an ECG signal) of a user. The wearable device may process the detected biosignal through a first subset of layers of a neural network. The wearable device may obtain 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 smart phone or a server) to further process the data through a second subset of layers of the same neural network to derive characteristics of the user, such as an indication of past 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, as described herein, at least some of the calculations may occur on one or more processors external to the wearable device and / or one or more processors within the wearable device.

[0184] In some embodiments, the output of the first subset of layers is an indication of an R-peak sequence. In other embodiments, the output of the first subset of layers is an indication of RR intervals. The wearable device may process the ECG signal through the first subset of layers of the neural network to receive the R-peak sequence from the output of the first subset of layers. The wearable device may transmit the R-peak sequence to an external computing device. The external computing device may process the R-peak sequence through a second subset of layers of the neural network and receive a derived feature of the user as an output of the second subset of layers (e.g., an indication of past arrhythmia and / or a prediction of future onset of arrhythmia, burden of atrial fibrillation, etc.).

[0185] There has been much recent exploration into the use of wearable sensors to detect activity in cardiac arrhythmias, such as atrial fibrillation (AFib). The wearable sensors used for such research fall broadly into two categories. The first category of devices are electrocardiogram (ECG) monitoring devices. The second category of devices are photoplethysmography (PPG)-based devices. When PPG-based devices are used for AFib detection, the results have been mixed. For example, in one study, a deep neural network-based algorithm produced good results using heart rate and step count data from a PPG-based device on a dataset obtained from a patient population before and after cardioversion of AFib in a clinical setting, but the same approach and algorithm performed more mediocrely in an outpatient setting with a broader population of persistent AFib patients. There may be many reasons for the poor performance in outpatient settings, but the most influential factor is likely the poor quality of the PPG signal that estimates heart rate in uncontrolled conditions. This raises the need to more accurately detect AFib using heart rate data derived from ECG signals recorded using wearable heart monitors such as ECG monitoring devices.

[0186] AFib burden is the fraction of time a patient is in the AFib state. The concept of AFib burden has gained a lot of attention recently due to studies showing an association between AFib burden and ischemic stroke risk in adults. Therefore, a cardiac monitoring solution that can quickly and accurately estimate a patient's AFib burden could lead to meaningful medical interventions for patients who have been identified as being at higher risk for stroke.

[0187] For clinical analysis, a patient's ECG signal recorded using a cardiac monitoring device is typically segmented into episodes of different durations and different rhythm types. Normal, healthy patients are typically in a normal heart rhythm state known as normal sinus rhythm (NSR). Various cardiac conditions can manifest as irregular / abnormal heartbeat patterns, which are known as arrhythmias, which in turn translate into irregular / abnormal patterns of ECG signals. There are various types of arrhythmias, such as atrial fibrillation. Patients can switch back and forth between normal heart rhythm and one / more different types of arrhythmias, or even be in one of these abnormal rhythms, such as AFib, all the time. The AFib burden is the fraction of time that a patient is in atrial fibrillation. Atrial fibrillation is another arrhythmia that is closely related to AFib and increases the risk of stroke. AFib and atrial fibrillation are sometimes difficult to distinguish from each other, and due to similar clinical significance, they are often combined together for load calculations. Therefore, for the purposes of this disclosure, embodiments will be described with reference to AFib, but it should be understood that some embodiments may also (or only) be applied to atrial fibrillation.

[0188] The QRS complex is the main feature of the ECG signal. The peak within the QRS complex is usually called the R peak. Some embodiments detect the R peak signal 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 consecutive R peaks (called the RR interval). A key feature of an Afib attack is high RR interval variability.

[0189] Fig.18A is a graph showing an embodiment of readings from an ECG strip corresponding to a NSR heart rhythm. Fig.18B is a graph showing an example of readings from an ECG strip corresponding to an NSR rhythm. The dots indicate the locations of the R-peaks.

[0190] The ECG signal recorded from a patient can be segmented into episodes of different rhythm types, and the AFib burden can then be calculated by calculating the fraction of time in the AFib state. However, if the calculation is based only on the extracted sequence of R-peak positions (rather than the entire ECG signal), then estimating the AFib burden becomes a different algorithmic problem. Scenarios where such an algorithm may be useful are when we may want to avoid doing too much calculation on the monitoring device itself (using the entire ECG signal), or when we may want to limit the amount of data transmitted from the monitoring device using wireless communication to a computer server that calculates the AFib burden.

[0191] Described herein are embodiments that use heart rate data obtained from a wearable heart monitor to reliably detect AFib. Some embodiments may also reliably estimate a patient's daily AFib burden.

[0192] In some embodiments, an algorithm (eg, a neural network) may estimate AFib burden over a shorter analysis window (eg, approximately half an hour or an hour in duration).

[0193] In some embodiments, the burden for longer time periods (e.g., longer than shorter analysis windows) can be determined by accumulating the burden predictions within the analysis windows. For example, the burden can be accumulated within non-overlapping analysis windows of half-hour duration. Given a sequence of R peaks within a half-hour analysis window, a sequence of RR intervals within the analysis window is derived. The RR interval sequence is then converted to a RR interval subsequence by extracting RR interval subsequences from a set of possibly overlapping sliding windows of duration w (contained in the analysis window), and the RR interval subsequences are shifted by a certain amount 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 N subsequences within the analysis window w Such sliding windows, the dimension of the time series is (N w ,M). The time series is input into the neural network model for burden prediction. To explain this concept, w is selected as 30 seconds, s is selected as 25 seconds, and M is selected 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 a second portion of the neural network (e.g., a second subset of the layers of the neural network)) to make inferences, such as inferring the possibility of an arrhythmia. In some embodiments, the neural network can be trained to make determinations about the use of the wearable device, such as indicating that the user should wear the device longer due to a higher risk of potential negative health events, or indicating that the user is at a low level of risk and can end their wearing period early.

[0194] Advantageously, this separate processing of data can have many technical benefits. For example, this selective transmission of partially processed data can reduce power consumption in the wearable path. Instead of storing the entire data processing algorithm (e.g., the entire neural network), the wearable patch may only need to retrieve a subset of the layers of the neural network for processing. As a result, data processing by the wearable device can be faster, require less computation, and use less battery via processing. Analysis on the wearable patch side may not require full onboard heart rhythm analysis, which is typically very power-hungry, thereby shortening battery life. Reduced battery consumption can extend the use time of the wearable device, such as the time period between device replacement, battery charging, etc. In addition, the wearable device can have a longer monitoring time to gain a more comprehensive understanding of the patient's heart rhythm.

[0195] Another advantage of remote analysis based on RR interval data can allow for increased sensitivity and precision in the analysis of the pathway. Analysis on a wearable patch can be limited to memory, processing, and battery limitations of the wearable patch. Remote analysis can process the data through a variety of algorithms, even those based on deterministic and eventual decision trees.

[0196] Additionally, the wearable patch can send data that is smaller than the entire detected cardiac signal dimension. Advantageously, less network throughput is required. Even in situations where wireless network conditions may not be optimal, the wearable patch can send enough data for remote data processing. Furthermore, because only a subset of the neural network needs to be uploaded to the wearable patch, the wearable patch can be updated with updated neural network computer executable instructions or downloadable executable files more quickly under weaker network conditions and can occur more frequently.

[0197] Furthermore, due to the lower battery and power requirements, a smaller CPU or smaller battery can be used, allowing the wearable patch to have a much smaller form factor. Furthermore, the reduction in battery and power requirements required for data processing frees up the battery and power budget of the wearable device, allowing the wearable device to perform other critical tasks, such as running other critical software processes, or powering other hardware devices (e.g., more electrodes, other sensors such as accelerometers), or running the same hardware and software for longer periods of time.

[0198] The neural network encoding method advantageously allows very complex algorithms to be run based on data on a battery-constrained device. One application is to aggregate the encoded data from the device in real time and run an algorithm that assesses the subject's overall risk of developing a specific clinical outcome. Thus, the system can dynamically notify the user of the wear period. For example, if after a certain amount of data has been aggregated by the wearable device, the remote computing device may determine that the subject's risk of developing a specific clinical outcome is very low. The system can notify the user of the wear period in advance. Conversely, if the computing device has determined that the user is at high risk for a certain clinical outcome, the system can notify the user to wear the device longer.

[0199] Another technical advantage of having the first subset of layers of the neural network output RR interval data to be sent for remote processing is that the algorithm used to perform inference remotely can be updated 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. The technical advantage is that the second subset of layers can be selected based on a decision tree, updated or swapped with other algorithms, etc., without updating the wearable patch, thereby making the system more compatible with other algorithms.

[0200] Example high-level architecture of a burden prediction model

[0201] Fig.19 is the high-level architecture of the burden prediction model. The burden prediction model takes the RR interval sequence as input and produces an AFib burden prediction. Here, N w is the number of sliding windows, and M is the size of the RR interval subsequence extracted from each sliding window.

[0202] The first layer 1902 is a time distribution feature extraction layer, which extracts a 32-dimensional feature vector for each RR interval subsequence obtained from the sliding window. The output of this layer is a dimensional (N w ,32). This transformed sequence is then input to two recurrent neural network (RNN) layers 1904, 1906 and a long short-term memory (LSTM) unit. The first RNN layer 1904 returns another dimension (N w ,32), and the second RNN layer 1906 returns an output of dimension 32. This is followed by two dense (fully connected) layers 1908, 1910, which produce a scalar value representing the AFib burden of the analysis window.

[0203] Fig. 20 yes Fig.19 An example of a high-level architecture of , which shows the LSTM architecture in more detail.

[0204] Fig.21 yes Fig.19 Embodiment of the feature extraction model of. As shown in the figure, the model for feature extraction can be distributed over a sliding window in the first layer of the model in time. The feature extraction model distributed over the sliding window in time can itself be another neural network with multiple layers.

[0205] exist Fig.19 The beginning and Fig.21 In step 2102, the feature extraction model takes as input a RR interval subsequence of length M. Fig.19 Steps 1902, 1904, 1906 and Fig.21 Steps 2104, 2106, 2108, 2110, 2112, 2114 of the network are processed by a set of one-dimensional convolutional layers and a maximum pooling layer. This is followed by a dense (fully connected) layer, which is Fig.19 Steps 1908, 1910 and Fig.21 A feature vector of dimension 32 is output at 2120, 2122. The purpose of this feature extraction model is to learn features that indicate the presence / absence of AFib within each sliding window of duration w.

[0206] For neural networks, a loss function that can be used is the average (over all samples) of the absolute differences between the burden predictions from the model and the true burden values:

[0207]

[0208] Here, is the load forecast output by the model for the i-th window in the data set, b i is the true loading of the window and N is the total number of samples in the dataset.

[0209] In some embodiments, a feature extraction network (e.g., Fig.21 This was done by adding a dense (fully connected) layer with scalar output and sigmoid activation to the feature extraction network and then training the resulting model for binary classification to detect the presence of AFib in 30-second ECG strips.

[0210] The parameters of the feature extraction network thus obtained are then frozen, and the obtained feature extraction network is inserted into the load forecasting model ( Fig. 20 ). The subsequent layers of the burden prediction model (recurrent network layers and fully connected layers) are then trained on the half-hour windows described in Section 4.2.

[0211] The feature extractor network is trained on a larger and different set of ECG strips and directly learns features that indicate the presence / absence of AFib. When training the remaining layers of the load prediction model, the number of parameters that must be learned is reduced, so the search for optimal parameters may be faster. In contrast, if the entire network were to be trained at the same time, the number of parameters that must be learned simultaneously would be larger and the model would be exposed to a smaller set of ECG strips. This increases the risk of overfitting and may explain the larger gap between validation loss and training loss.

[0212] System for estimating burden and / or predicting cardiac rhythm annotation using neural network encoding

[0213] Some embodiments disclose a wearable device that can process a detected biological signal via an encoder that includes a first subset of neural network layers. Fig.22A 22 is a schematic diagram of an embodiment of a system for predicting cardiac rhythm annotations using neural network encoding. In some embodiments, the system can make one or more predictions, such as predicting cardiac rhythm annotations or estimating the burden of cardiac rhythms including atrial fibrillation and / or atrial flutter. A wearable device, such as a cardiac monitor patch 2202, can process an ECG input 2204 through a first subset 2206 of neural network layers (e.g., an encoder). The wearable device can receive the output of the first subset 2206 of layers of the neural network and transmit the output to a computing device 2208 (e.g., an external system such as a smartphone or server) to further process the data through a decoder to derive a characteristic of the user 2212, such as an indication or prediction of past arrhythmias and / or predict future onset of arrhythmias, wherein the decoder includes a second subset 2210 of layers of the same neural network.

[0214] In some embodiments, the first subset of layers and the second subset of layers are within a neural network. The neural network can be designed such that the output of the first subset of layers can output data with a smaller dimension than the input of the neural network, and the output of the second subset of layers can be designed to provide an indication of a user characteristic, such as a past or future prediction of AFib. Fig. 22B2224. The neural network can be trained on the first subset of layers 2222 and the second subset of layers 2224 simultaneously. For example, if the neural network has 10 hidden layers, the first 4 layers are processed on a wearable device (e.g., a heart monitor patch 2226), and the output of the 4th layer (which has a lower dimension than the input of the 1st layer (e.g., has good data compression characteristics) is transmitted to an external computing system, such as a server 2228. The external computing system processes the output of the 4th layer through layers 5 to 10. The dimension of each layer of the neural network can be designed to output a certain data size (e.g., the output dimension of each convolution 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 include an ECG encoder 2230 that receives ECG data 2232 at 2400 bits per second (bps). 2226 can process the ECG data 2232 through a first subset of layers 2222 of the neural network (e.g., within the ECG encoder 2230) and output data 2234 of a smaller dimension, such as 128 bps data. The output data 2234 can be transmitted to an external server 2228. The external server 2228 can include a classifier 2238, which processes the output data 2234 through a second subset of layers 2224 of the neural network and outputs an indication or prediction 2236 of the patient, which is also the purpose of neural network training. The entire neural network including the first subset of layers 2222 and the second subset 2224 can be designed and trained as a single neural network.

[0215] In some embodiments, the output of the first subset of layers may have a smaller dimension than the input of the neural network. Thus, rather than transmitting the entire ECG signal (e.g., the input of the neural network) from the wearable device to the external computing device, the wearable device may transmit a smaller amount of data to the external computing device, such as transmitting 128 bps of output data 2234 of the first subset of layers instead of transmitting the full 2400 bps ECG signal 2232. Advantageously, less network throughput is required to obtain an indication of a past arrhythmia and / or a prediction of a future onset of an arrhythmia.

[0216] Furthermore, rather than processing the ECG signal on the wearable device through all layers of the neural network to derive an indication of a past arrhythmia and / or predict a future onset of an arrhythmia, the wearable device may process the ECG signal through only a first subset of the layers of the neural network (e.g., through an encoder) via an 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., through a decoder or classifier 2238). Fig. 22C2246 may process a first subset 2248 of the layers of the neural network and an external server 2250 may process a second subset of the layers. Advantageously, the wearable device performs a smaller amount of computation and requires less memory to store data between layers.

[0217] In addition, the size of the wearable device 2246 can be smaller because the battery, processor, and other electronic devices required to process only a subset of the layers of the neural network are smaller. The wearable device can have more resources to perform other tasks. For example, in one case, the wearable device 2242 may have enough computing power and battery to process ECG signals internally and identify AFib for 5 days. However, if the wearable device 2246 can calculate a first subset of the neural network layers on the wearable device 2246, and the external computing device 2250 processes the second subset, the system as a whole may be able to identify AFib and atrial fibrillation over a span of 20 days. Some embodiments disclosed herein can reduce network throughput, reduce computing requirements and / or memory storage of the wearable device. Advantageously, the wearable device can maintain battery life by using fewer processors.

[0218] Certain embodiments disclosed herein improve the memory requirements of a wearable device. If the wearable device processes a first subset of neural network layers and an external computing system processes a second subset, the wearable device must store less data (e.g., software executable instructions for the first subset of layers rather than the entire neural network). As a result, the size of the electronic file is much smaller. Advantageously, not only does the wearable device have less memory requirements, but the neural network is also loaded and processed faster than a complete neural network. In addition, under weaker network connection conditions, the software embedded in the wearable device can be updated faster and can be programmed to be updated more frequently. In addition, more neural networks can be stored in the wearable device. For example, instead of storing a single complete neural network in memory, a wearable patch can store a subset of the layers of 5 neural networks in local memory.

[0219] In some embodiments, because the purpose of the external computing device is not to reconstruct the ECG signal, but to derive some user characteristics (e.g., predict a heart rhythm classification such as AFib), the data transmitted from the wearable device does not have to be sufficient to restore the full fidelity of the ECG signal. Therefore, the neural network can be trained to produce an output in one of its layers that is smaller than the ECG signal input dimension, and transmit much less information to the external computing system to execute 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. Patch 2262 may transmit the full ECG signal (e.g., at 2400 bps) or other features of the data to server 2264. However, such data may be hacked and sensitive data may be retrieved by an unexpected third party. Some embodiments disclose processing a subset of neural network layers on a wearable device, which provides technical advantages over remote processing of the entire ECG signal. Transmitting the entire ECG signal, or a combination of R peaks, R intervals, and other features at 2400 bps from a wearable device to an external computing system may require more network bandwidth than the wearable device transmitting the output of a subset of layers that are smaller in dimension than the ECG signal. In contrast, some embodiments of the present disclosure include a patch 2266 that processes a subset of layers 2268 and outputs the first subset of layers to a server 2270, which executes the remaining layers 2272 of the neural network. Advantageously, the data transmitted may be the output of a layer in the neural network, which provides enhanced data security and encryption.

[0220] Some embodiments improve the security of health data. For example, if the entire ECG signal, R peak, RR interval, or the final determination of the possibility of AFib is transmitted from a wearable device to an external computing system, a third party may intercept such data and identify the user's sensitive health data. In contrast, some embodiments disclosed herein transmit the output of an intermediate layer of a neural network (e.g., the output of an encoder) to be processed externally by subsequent layers (e.g., a decoder). The transmitted data can be the output of a layer in a neural network that inherently provides data encryption for data transmission. Therefore, even if a third party intercepts such data, they cannot decode the data or reverse engineer the original input to the neural network.

[0221] In some embodiments, the first subset of layers may include less computationally intensive layers, while the second subset of layers may include more computationally intensive layers. Advantageously, the wearable device may process the less computationally intensive layers, and an external computing system independent of the processing of the wearable device may process the more computationally intensive layers (e.g., in a server with greater processing power and memory capacity).

[0222] In some embodiments, the wearable patch may perform preprocessing on the detected cardiac signal before processing the data through a first subset of the layers of the neural network. For example, the wearable patch may perform preprocessing such as downsampling, normalization, numerical filtering, etc. In addition, the wearable patch may perform more complex preprocessing such as discrete wavelet transform, continuous wavelet transform, discrete Fourier transform, and discrete cosine transform. These transforms may be beneficial in lossy compression schemes because the transform transforms the data in a way that causes many values ​​in the data to be very close to zero. An encoding may then be created to simply store the value with the highest amplitude, which will limit errors when reconstructing the signal.

[0223] The wearable patch can apply a discrete wavelet transform or a discrete cosine transform and simply apply an algorithm to make inferences on the detected cardiac data. However, this approach may not capture more subtle features that are critical to the end application (e.g., ECG rhythm classification) because higher amplitude features are favored. In some embodiments, the system can train a neural network encoder on this transformed dataset (e.g., already preprocessed cardiac data) to more intelligently encode the data required by the end application. Advantageously, this may be beneficial for certain applications because the transform may already perform certain actions on the signal required by the end application, which may result in a more powerful or simpler neural network encoder design with fewer parameters compared to the encoder for the original signal.

[0224] In some embodiments, a neural network may be trained and / or designed to transfer optimal data dimensions from a wearable device to an external computing system. Fig.22E 2284, the second neural network 2286, and / or the third neural network 2288. The patch 2281 can 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 can transmit the output of the first subset of layers to the server 2282. The server can 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.

[0225] The neural network can be trained and / or designed based on one or more factors (e.g., available processing power and / or memory of the wearable device (for purposes of this disclosure, the neural network will be described as being trained, but features can be applied to designing or processing data through the neural network and vice versa). For example, if the wearable device has lower processing power, the neural network can be trained to perform less computation on the wearable device and / or process through fewer neural network layers, and perform more processing on the external computing system side. The patch 2281 can apply a third neural network 2288 that includes fewer computations and / or neural network layers.

[0226] In some embodiments, the neural network can be trained based on the network availability between the wearable device and the computing system. For example, when the network throughput is low, the neural network can be trained to process more layers and / or transmit smaller data dimensions from the wearable device to the external computing system. When the network connection is weak (or below a threshold), the patch 2281 can apply a second neural network 2286 including smaller data dimensions that are transmitted to the server 2282.

[0227] In some embodiments, the neural network can be trained based on the available battery life of the wearable device. For example, when the battery life is short, the neural network can be trained to process fewer layers on the wearable device. The patch 2281 can apply a third neural network 2288 that includes fewer computations and / or neural network layers.

[0228] In some embodiments, the neural network may be trained according to the desired results. For example, patch 2281 may apply 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 in intensive care or has signs of AFib, patch 2281 may apply first neural network 2284 and process more layers and nodes through a first subset of layers 2284A so that server 2282 can output more information via the output of a second subset of layers 2284B.

[0229] In some embodiments, the software in patch 2281 may include instructions for a particular neural network designed (and / or trained) for a particular application. For example, a neural network may be selected for a patch 2281 that has certain technical limitations (e.g., memory or processor) and serves a particular purpose (e.g., predicting AFib), while another neural network may be selected for a different patch with different characteristics and a different purpose. For example, accurately detecting AFib may require more patch data than another intended purpose. Thus, a patch used to detect AFib may have software that computes more layers than another patch.

[0230] In some embodiments, the software in patch 2281 can be designed (and / or trained) based on performance metrics. For example, the software can select a neural network that is designed to produce a certain detection accuracy and / or a certain false alarm rate.

[0231] In some embodiments, the software in patch 2281 may include multiple neural networks. Patch 2281 may select a particular neural network from the multiple networks. For example, the selection may be based on a neural network designed for low battery power or network connectivity (e.g., switching to a neural network that processes less on the patch or transmits less data over the network).

[0232] In some embodiments, the neural network can be trained to reconstruct the original input signal of the neural network, such as ECG signals, RR interval information, R peak information, metrics derived from ECG signals, accelerometer data, impedance, temperature, audio (e.g., snoring), ambient light, etc. Patch 2281 can transmit data of smaller data dimensions to server 2282 for server 2282 to reconstruct the original input signal or its derivative signal, such as the original input signal sampled at a lower frequency, a lower quality signal, a signal with a certain reconstruction error threshold, etc.

[0233] In some embodiments, the neural network may be trained based on the user's current activity. Fig.22F It is a schematic diagram of an embodiment of 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 receiving certain treatments that may require different types of peak sensitivity. Certain activities may require higher accuracy or precision to effectively detect heartbeat abnormalities. If the activity requires higher R peak accuracy, for example, for the sleeping user 2296, the signal level 2298 is less affected by movement artifacts, so the ECG signal can be relied on more confidently than the running user 2292, so for the running user 2292, higher R peak accuracy is preferred to obtain more accurate predictions. For the running user 2292, a more robust encoding process can be used to ensure data accuracy, and the neural network can be trained so that the wearable device can transmit higher resolution data. On the contrary, for the running user 2292, the R signal level 2294 can be clear (e.g., clearly showing the R peak), and therefore, a lower R peak sensitivity may be required. In this case, the patch can select a neural network designed to not require a higher R peak sensitivity, which will achieve the advantages described herein, such as reducing processing power or network throughput.

[0234] The training of models (e.g., training of artificial intelligence models) is necessarily rooted in computer technology and is improved by using training data to train such models and thereafter applying the models to new user data sets. Such training involves complex processing using large training data sets, typically requiring a large amount of processor computation and a long time, and these processes are typically performed by large server systems. The training of the model may require logistic regression or forward / backward propagation of training data, which may include input data and expected output values ​​for adjusting model parameters. Such training is a framework for machine learning algorithms that enables the model to be applied to new data (e.g., new biometric data) and make predictions based on weights or scores adjusted during the training process. Such training reduces false positives and improves the performance of AFib detection.

[0235] In some embodiments, the wearable device can store multiple neural networks and can apply a neural network based on one or more features such as characteristics of the wearable device, the network, the user's current activity, etc. For example, if the battery life is short, the wearable device can process a first neural network, or if the network throughput is high, the wearable device can process a second neural network.

[0236] In some embodiments, when the wearable device detects the ECG signal in real time, the wearable device can continuously process the ECG signal through a first subset of layers of the neural network. The wearable device can continuously transmit the output of the first subset of layers to an external computing device, and the external computing device can process the received output through a second subset of layers of the neural network to derive the characteristics of the user (e.g., arrhythmia) in substantially real time. Advantageously, the external computing device can identify arrhythmias in substantially real time when the user has an arrhythmia, while reducing network throughput, memory requirements, and processing power requirements. In addition, more powerful algorithms and neural networks that cannot be executed on the wearable device itself due to device limitations, and cannot be executed entirely on external computing systems due to input requirements, are now feasible.

[0237] 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 one or more types of data (e.g., a user's ECG signal, R peak data, RR interval subsequences (e.g., RR intervals in a sliding window of 30 seconds), encoded features, and / or other similar data as further described herein) as input. The first neural network may process the input RR interval subsequences and generate atrial fibrillation (AF) features for each sliding window. The output of the first neural network may be fed into a second neural network, which includes a recurrent layer having a long short-term memory (LSTM) unit. The LSTM unit may process the AF features of the current sliding window and the past sliding windows to process the data using a transfer learning approach, wherein over the course of multiple sliding windows, the LSTM may advantageously predict indications of arrhythmias that take into account data longer than a 30 second window.

[0238] 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 at 30 second intervals can be input into 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 longer data intervals fed into the first neural network in a 30 second window, and the output is fed into the second neural network through an LSTM unit. The second neural network can make predictions for AFib on a longer data set, and the predicted output can be used to adjust the weights of the second neural network. Once the second neural network is trained, the weights of the first neural network can be unfrozen, and the first and second neural networks can be trained simultaneously using other training data.

[0239] As wearable biosignal processing applications grow, there is an increasing need to increase the complexity and practicality of these applications. A common scenario is to have a small, battery-powered wearable device with an array of sensors that measure processed and optionally transmit biosignals in real time or near real time. Since device size affects comfort, which in turn affects compliance and therefore practicality, it is advantageous to have a device that is as small as possible. This places some constraints on the system (primarily battery life and processing power), which in turn can negatively impact the user experience by requiring the user to recharge the battery, or by causing the analysis to degrade in quality and reduce compliance. These constraints typically limit the complexity of the processing that occurs on the device, so the option used by many applications is to offload the processing to less constrained systems, such as smartphones, cloud servers, or base stations.

[0240] In this paradigm, there is a tradeoff between the amount and type of data that is offloaded for additional processing. Ideally, the full-fidelity raw biosignal is offloaded in real time to maximize processing complexity in the less constrained parts of the system. Unfortunately, with current technology limitations, this is often not feasible due to battery costs and technical limitations on transmitting the full-fidelity signal in real time or even in a timely manner. One way to overcome this limitation is to send a number of smaller representations of the signal, thereby minimizing power requirements and maximizing the time between battery charging or replacement, or maximizing the potential monitoring duration of the device if powered by a single battery charge.

[0241] There are many strategies for creating this alternative representation. Depending on the application, a custom designed algorithm can be used to select uploaded features for further processing. An example of such a strategy is a mobile cardiac telemetry (MCT) device, or a wearable activity tracker that is able to detect and upload heartbeat locations derived by an algorithm running on the device, which can then be further analyzed by more complex algorithms on a cloud server to detect arrhythmias. Another example is a wearable device that uses an accelerometer to analyze gait, assess activity level, and fall risk by uploading the results of a pedometer algorithm on the device. A limitation of this set of approaches is that the features selected may not be the best choice for the end application, and the computational burden of the device-side algorithm may also be greater than necessary.

[0242] Another common approach is to compress the raw biosignal data in a way that allows the signal to be reconstructed. Commonly used techniques include lossless and lossy compression schemes. Lossless compression, while allowing for optimized processing on cloud-based servers, generally does not provide compression ratios high enough to sufficiently reduce the data transmission burden to make it feasible. Lossy compression techniques, such as wavelet-based compression, can achieve high levels of compression relative to the amount of error introduced during reconstruction. While such strategies may be suitable for some applications, there is no guarantee that the reconstructed signal will retain the necessary information for the end application, as their goal is simply to minimize signal reconstruction errors. This type of goal is primarily driven by signal features with high amplitudes and may ignore or distort features with small amplitude characteristics. Depending on the application, signal features with relatively small amplitudes may be critical to the end goal, such as the subtle P waves associated with atrial depolarization in the electrocardiogram, which are critical to the detected heart rhythms (e.g., atrial fibrillation and complete heart block).

[0243] The systems and methods disclosed herein achieve an optimal balance of all elements in the system: device processing burden, system transmission burden, and end application utility. This is achieved using neural network encoding, which can be used to generate a scheme for compressing raw biosignal data in a computationally friendly manner, which retains the information required for a specific end application. The neural network can encode the signal into a compressed format and then reconstruct the signal. More generally, the neural network architecture can be built into encoder and decoder parts.

[0244] For this application, the encoder part will be responsible for compressing the raw bio-signal, will run on the target device, and will produce an output with smaller dimensions than the input.

[0245] The decoder portion will be responsible for generating predictions for the target application starting from the encoding layer of the neural network and will be used somewhere other than the target device, such as on a smartphone, cloud server, or communication hub / gateway. The encoder and decoder portions are trained with an end-to-end neural network, which will effectively create the best encoder and decoder at the same time because the encoder and decoder portions are being optimized for the same target. This approach maximizes flexibility in system design, allowing for the most optimized implementation. For example, the encoder portion can be designed for a specific computational complexity and compression, and the neural network output can be selected to exactly meet the needs of the end application. This technology is ideal for any real-time and near-real-time signal processing application that requires high computational complexity with mechanisms for offloading processing.

[0246] Devices that utilize electrocardiogram (ECG), photoplethysmography (PPG), or audio signals to detect electrical and structural heart conditions in real time or near real time face the above challenges. Certain arrhythmias appear in subtle ways, such as low amplitude P wave patterns in ECG recordings, often requiring a high degree of sophistication and sensitivity in the detection algorithms. To maximize clinical utility, the rhythms can be classified into up to 15 different categories (including noise), relying not only on beat-to-beat heart rate patterns but also on often elusive morphological differences, especially in single-lead ECG applications. Existing devices may simply offload the entire raw signal, requiring frequent charging or battery replacement, or use compression techniques to offload processing, or detect commonly understood features of the ECG signal on the device and upload these features to a separate system for further processing. These features can include the location of ECG morphological features (e.g., QRS complexes, P waves, etc.). Due to the limited computing power and battery life of these devices, the accuracy of the algorithms in their ability to detect these morphological features is limited. Neural network encoding can be used to transmit a compressed representation of the ECG signal. This representation can be optimized for a specific end application, such as identifying specific arrhythmias, identifying ECG morphology features (e.g., QRS complexes), classifying beat types, determining heart rate, etc.

[0247] In an example implementation of this strategy for detecting arrhythmias in real time or near real time from ECG recordings, a neural network can be first trained using a set of 30s of ECG strips (sampled at 200Hz, e.g., with 14 different rhythm labels per sample). As a preprocessing step, the 200Hz signal is downsampled by a factor of 3 using a moving average filter and then conditioned using an IIR high pass filter with a 2Hz cutoff frequency. The input is also scaled to have unit variance. This will leave information in the range of about 2-33Hz, which will contain most of the ECG morphology information. This makes the input dimension of the signal in the neural network 2000x1. As mentioned above, the goal of the first part of the neural network is to compress the signal, so the dimensionality of the input signal needs to be reduced.

[0248] In the example embodiment, the encoding portion of the neural network will have a series of convolutional and pooling layers, where each pooling layer reduces the time dimension by a factor of 2. With the sequence of convolutional and pooling layers repeated 3 times, this will result in a reduction in the time dimension by a factor of 8. Also in this example embodiment, the final convolutional layer of the encoder portion will have 2 filters, which will give an output dimension of 250x2 and is referred to as the encoded output.

[0249] Subsequent layers of the network adopt a strategy of returning the size of the time dimension to its original size at the network input. This can be done using several transposed convolutional layers (sometimes called deconvolutional layers). In one embodiment, the layers of the network so far can be pre-trained to predict the original input signal. In one embodiment, the remaining layers are composed of residual convolutional filter blocks. In this example, 12 sequential blocks are used, each containing 2 convolutional layers with a residual connection from input to output. These blocks also reduce the time dimension by a factor of 2 every other block. After the last of these blocks, a dense layer of time distributions is evaluated at each time point to predict a unique rhythm class. The above example describes an embodiment of a framework for training an algorithm to encode ECG information in a way that preserves the diagnostic information in a compressed format.

[0250] In some embodiments, the encoding process can be implemented to generate ECG features permanently. Using the example neural network above, an ECG strip of 6000 samples in length will create an example of 250 two-dimensional feature vectors. This can be thought of as two encoded features produced for every 24 separate ECG samples. Since the nature of this process is causal, these features can be created permanently without any concept of a 30s ECG segment. These features can be scaled to be stored in a compact integer representation, and lossless compression can be applied to further reduce feature memory usage. With a continuous stream of encoded features, server-side algorithms can use the encoded features to detect target events using a variety of different strategies.

[0251] In one embodiment for detecting segments containing arrhythmias (e.g., CHB, VT, AFIB / AFL, SVT, AVB, etc.), the original multi-class algorithm used to train the ECG encoding can be evaluated using a 30-second context of the encoding features generated on the patch. The algorithm can be evaluated on a rolling window of the 30-second context, with or without overlap. Segments with a higher probability of containing arrhythmias can be used to determine whether a target event is present.

[0252] In one embodiment, the encoding portion of the neural network may use a combination of convolutional layers, pooling layers, and recurrent layers (e.g., LSTM) to reduce the dimensionality of the signal.

[0253] In one embodiment for detecting VT events, a neural network can be trained using encoded features to predict VT sequences in ECG segments that are approximately 1 second in duration. The benefit of this is that more or less severe segments can be distinguished to determine which events are of greater concern. For example, longer segments can be prioritized over shorter segments. The algorithm can be used to determine the burden of VT episodes that meet more specific length criteria. Traditional VT detection methods (especially on battery-constrained embedded devices) rely on beating patterns that show sudden increases in rate.

[0254] The reliance on beat detection (e.g., the widely used Pan-Tompkins algorithm, which generally performs poorly on wider QRS beats) inevitably limits the ability to reliably discern beat patterns in arrhythmias such as VT. Even if a full ECG is present and the morphological information can be studied via signal processing techniques, the computational burden on battery-operated devices is high, resulting in the need to recharge the device one or more times per day and / or replace batteries on the sensor and gateway or mobile phone unit. For example, many current mobile cardiac telemetry (MCT) modalities are affected by this requirement, which shifts its burden onto the patient, thereby resulting in reduced compliance, analyzable time, and clinical value of the diagnostic device.

[0255] In one embodiment for assessing AF burden, the original neural network trained to create the ECG encoding can be evaluated in the context of 30 seconds to determine whether the segment contains atrial fibrillation (AFIB) or atrial flutter (AFL), and the sum of these times can be used to calculate the burden. With the exception of atrial fibrillation, existing models generally have difficulty distinguishing AFL from other regular rhythms (e.g., normal sinus rhythm) because the irregular beat pattern characteristic of atrial fibrillation is not present. This requires a closer examination of the signal between beats when the encoding features allow. Of course, the full ECG also allows this, but requires more computing power to process the information, which is usually in the form of frequency domain analysis, which in turn can result in a less patient-friendly device, such as charging or battery replacement discussed in the MCT embodiment above.

[0256] In another embodiment for assessing AF burden, the encoded features can be fed into a recurrent neural network within a larger context (e.g., 30 minutes) to predict AF burden in the same manner as described above.

[0257] In another embodiment for characterizing VT, the encoded features can be fed into the network within a larger context (e.g., 30 minutes) to predict the total duration of VT and / or the number of discrete VT episodes present.

[0258] In one embodiment for determining heart rate, a neural network can be trained to use the encoded features to predict the number of heart beats contained in a time period of, for example, 3 seconds. The neural network can be used to predict the number of beats within a sliding window of time periods, which can then be used to estimate heart rate at different times. This heart rate information can be used to further refine or prioritize target events detected using other algorithms. A similar strategy would be able to estimate heart rate variability instead of or in addition to heart rate.

[0259] In one embodiment for determining an ectopic beat count, a neural network can be trained to use the encoded features to predict the number of ectopic beats contained in a time period of, for example, 3 seconds. The neural network can be used to estimate the total number of ectopic beats within a time window (e.g., within 1 hour, 12 hours, or 24 hours). Optionally, in conjunction with the above heartbeat embodiment, this embodiment can be used to estimate the ratio of ectopic beats to normal beats, for example, expressed as a percentage over a 1 hour, 12 hour, or 24 hour period. This embodiment described above agnostic can be applied to ventricular or supraventricular ectopic beats, and / or couplets and triplets of each.

[0260] In one embodiment, the R peak locations can be approximated from the reconstructed time series from the encoded features. These R peak locations can be used alone to calculate heart rate, improve the accuracy of heart rate estimation through alternative methods, or detect cardiac arrest events defined as a clinically significant time interval (e.g., greater than three seconds) between two ventricular contractions. As with heart rate, this information can be used to further filter or prioritize target events for transmission and reporting of ECGs. In another embodiment, the reconstructed signal itself can be viewed along with portions of the full ECG to provide more context, such as whether the heart rate is low, high, or irregular starting at the end of the ECG segment. This contextual information can help manually request a more full resolution ECG from the patch based on visual inspection, thereby achieving an overall higher diagnostic acuity with the support of manual quality assurance.

[0261] In an alternative embodiment, the heart rate based on these R-peak locations can be plotted in the clinical report along with the selected ECG strip in a manner that provides greater temporal context (e.g., 8 or 16 second strips) for the events shown in the presentation. The heart rate can be plotted as an instantaneous heart rate (e.g., based on the interval between two R-peaks), or as a trend of heart rate over time, where each data point is from multiple R-peaks. An alternative embodiment can plot the average, maximum, and minimum heart rate trends simultaneously to provide greater clinical context.

[0262] In some embodiments, a wearable device utilizing PPG and an accelerometer can be used to detect atrial fibrillation (AFIB) episodes and estimate the AFIB burden level. Sensors in such a configuration often suffer from signal degradation due to environmental factors (e.g., improper wearing, movement of the device, or adverse physiological conditions). For this reason, utilizing full fidelity raw sensor readings to detect AFIB or estimate AFIB burden is advantageous because the noise can be understood and the algorithm is prevented from predicting false positives.

[0263] Existing strategies can detect heart rate intervals on the device and use that information to infer an AFIB episode. Signal quality metrics can be calculated and uploaded separately. The strategy will be affected when periods of poor signal quality produce erroneous heart rate interval readings. Using the strategy outlined above, a neural network can be developed to predict AFIB episodes or AFIB burden from raw PPG / accelerometer signals, which also encode the signal in a compressed representation. The network can learn to encode the information in a way that is least prone to false predictions due to poor signal quality. The encoding can be run on the device and the encoded data can be uploaded so that the decoding portion can be evaluated on a more powerful server. In addition to identifying AFIB episodes and AFIB burden, the encoded sensor data from the wearable PPG device can also be used to characterize the pulse rate or detect other arrhythmias such as ventricular tachycardia, pauses, atrioventricular block, etc. They can also be used to detect health conditions such as diabetes, hypertension, depression, congestive heart failure, sleep apnea, etc.

[0264] In another embodiment, the encoded features can be optimized to describe waveform components of the ECG heartbeat other than the R peak, such as the P wave, T wave, and / or segments between points in the waveform. A neural network can be trained on the encoded features to estimate the PR interval, with patterns that can distinguish between subtle classifications of AV blocks (1st degree, 2nd degree Mobitz IeWenckebach, 2nd degree Mobitz II) or characterize Wolff-Parkins-White syndrome while inferring QRS width. Some of these less life-threatening asymptomatic arrhythmias may require transmission, especially if the device does not continuously record and store all signals like event recorders and ILRs or some MCTs, where not all signals are fully analyzed.

[0265] In a similar manner, a neural network can be trained on the encoded features to estimate the QT interval, which can reveal clinically significant findings (e.g., long QT syndrome), which can be monitored over the course of wear time. For example, this QT measurement method can be used for drug titration, where the patient will adjust the drug dosage to determine the best combination of individualized drug efficacy and safety. The neural network can also be trained to detect changes in the ST segment, which can be monitored to look for elevation or depression of the ST segment, which may be a warning sign of myocardial infarction.

[0266] In another embodiment, the encoding method can be dynamically adjusted and adapted as the system receives and learns more about the characteristics of the signal. It can also be swapped out with a different scheme if the target rhythm for a particular patient changes. For example, for certain neurological applications (e.g., post-stroke monitoring), rather than a general method for detecting the type of arrhythmia, the primary focus may be on detecting AF, without regard to heart rate or onset time. In this case, a pair of encoding and decoding networks pre-trained and optimized for high sensitivity to AF can be used. These different encoders can be learned offline and updated to the firmware over the air when changes are needed, or pre-programmed as options on the embedded device to be activated when required. Optional embodiments can implement algorithms optimized for sensitivity or specificity performance in a similar manner.

[0267] In some embodiments, the input signal may include channels from accelerometer and / or gyroscope axes, which tend to take up a lot of space due to the number of axes multiplied by the required sampling rate (e.g., 20 Hz to be able to infer the number of steps). Features in 3-axis accelerometer and / or 3-axis gyroscope signals can be reduced to smaller dimensions, which may contain valuable information that is not limited to predefined physical indicators, such as amplitude. In fact, features that can only be obtained through higher frequency sampling can be encoded to better distinguish activities. The encoded accelerometer and / or gyroscope features themselves can be used to determine sleep stages and activity levels, steps, direction, activity type, etc. The encoded accelerometer and / or gyroscope features themselves, or features encoded from these signals together with the ECG, can be used to further distinguish signals generated by physiological events, such as the distinction between arrhythmias and movement artifacts. The same data is also helpful as an input (trend activity level) to assess the health status of patients with heart failure, for example, in the form of dynamic monitoring of decompensation events after discharge.

[0268] In other embodiments, impedance information or some other measurement indicative of skin contact (e.g., galvanic skin response) may be encoded with the ECG or independently. A neural network may be trained to detect lead-off areas and combined with ECG features that are used to filter out segments of unreliable signals. Since the same impedance value may mean different levels of signal quality in different patients and conditions, a simple state machine type algorithm targeting only impedance values ​​may result in a lead condition where the signal is actually a non-physiological condition, or conversely, a lead detected when the signal is actually an ECG and therefore analyzable. The former is particularly a problem for mobile devices that are susceptible to noise, which can sometimes mimic arrhythmias and therefore cause false positive transmissions, although some of these issues can be mitigated through hardware design and flexible materials (e.g., patch form). Decoder algorithms using ECG encoded features, together with raw or encoded features of secondary signals such as impedance or accelerometers, can further eliminate these false positives while also ensuring that the true ECG signal is not lost due to poor impedance.

[0269] In certain embodiments, suspected arrhythmias detected using coded features can benefit from further analysis using the full-resolution ECG on the original device. The energy savings achieved by the implementation of the encoder will allow decoder-based algorithms to require transmission or onboard analysis of portions of the full-resolution ECG stored on the device. A separate algorithm (possibly derived from a convolutional neural network trained on a labeled full-resolution ECG) running on a restricted area of ​​the original ECG is computationally feasible.

[0270] In certain embodiments, the encoded features may be used to learn and predict cardiac arrhythmias (eg, atrial fibrillation or other health conditions that may occur in the near future), thereby enabling early intervention.

[0271] In some embodiments, encoding or subsequent layers of processing using encoded features (decoding) can occur on an intermediate device, such as a smartphone or gateway device, prior to transmission over a remote connection (e.g., cellular or satellite). In some use cases (e.g., the original signal does not need to be written to device memory, which is often a significant portion of battery consumption in these limited power modes), both encoding and decoding can occur on the device itself (e.g., a patch, smartwatch, or other wearable device capable of continuously monitoring ECG or other signals), and only the encoded signal is written. This approach may be useful in situations where storage space is limited, but the algorithm may need to operate on data stored in memory (e.g., ECG data that occurred in the past).

[0272] In certain embodiments, the technology can be applied to other surface or transcutaneous biosignals in addition to ECG, such as PPG from the wrist, earlobe, or chest. Wearable devices (typically consumer grade) have PPG signal capture capabilities to be able to obtain heart rate trends and, in some cases, screen for atrial fibrillation. These health insight capabilities are typically limited to detecting highly limited durations (e.g., 30 seconds) and require relatively high-power optical sensing (e.g., with green light) to reliably capture waveforms across a variety of skin types and movements. The full content of the PPG cannot be stored or transmitted for richer contextual analysis, beyond simple rhythm detection, to prevent the morphology of the PPG wave from containing unique features of arrhythmic blood flow, or aggregated into more long-term and clinically useful indicators (e.g., AF burden).

[0273] In some embodiments, PPG and / or ECG waveforms are encoded to infer blood pressure and other cardiac indices, such as blood perfusion, blood volume, cardiac output, ejection fraction, valve function and / or health, etc., where conventionally calculated beat-to-beat or even fiducial intervals may not be the most useful features. Although certain embodiments are described using PPG, it should be understood that ECG may also be used where applicable, and vice versa.

[0274] In some embodiments, the technique can be applied to even more data-heavy signals, such as EEG. Wearable applications using EEG are typically limited to laboratory settings due to the number of channels and the size of the device. The encoded features of EEG training can offload high-computational power decoder algorithms to servers and enable mobility of the device, for example, to convert brain activity into communication.

[0275] In some embodiments, the neural network encoding can be implemented in dedicated hardware (e.g., FPGA IC) to further reduce battery consumption of the target device. The neural network encoder can be further optimized for embedded applications by utilizing integer math or binary operations.

[0276] In some embodiments, image or video streams resulting in detection of markers or objects or physiological measurements may be encoded using a neural network to offload processing to a more powerful system.

[0277] The example application would be

[0278] Object detection in vehicles, safety monitoring systems

[0279] Emotional responses from facial videos

[0280] Physiological measurements such as heart rate, respiration based on video streams of the face or chest

[0281] Detection of epileptic seizures

[0282] Fall Detection

[0283] Augmented Reality

[0284] Infant monitoring for significant events (crying, breathing problems)

[0285] In some embodiments, the audio data may be encoded for purposes of identifying speech events or translating speech in real time, respiratory monitoring, diagnosis of respiratory disease, detection of snoring or apnea.

[0286] In some embodiments, ultrasound data may be encoded for purposes of identifying contractions during pregnancy, detecting and characterizing fetal heartbeats, identifying abnormal physiology (eg, tumors, strokes, blocked arteries, infections, etc.).

[0287] Computing systems and methods

[0288] In some embodiments, the above-described systems, tools, and methods of using them enable interactivity and data collection performed by computing system 13000 . Fig.23 1 is a block diagram illustrating an embodiment of a computing system 13000 with a network 13002 and various external computing systems 13004 (e.g., a wearable system 13005, a gateway device 13006 that also communicates with the network 13002). The computing system 13000 can be used to implement the systems and methods described herein. Although the external systems 13004 are shown as grouped, it is recognized that each system can be external to each other and / or located remotely.

[0289] In some embodiments, the computing system 13000 includes one or more computing devices, such as a server, a laptop computer, a mobile device (e.g., a smart phone, a smart watch, a tablet computer, a personal digital assistant), an information kiosk, a car 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 (e.g., a hard drive, a floppy disk, a solid-state drive, or an optical media storage device). In some 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 a computer using a standard-based bus system. In different embodiments, for example, the standard-based bus system may be implemented in a peripheral component interconnect (PCI), a microchannel, a small computer computing system interface (SCSI), an industry standard architecture (ISA), and an extended ISA (EISA) architecture. In addition, the functionality provided in the components and modules of computing device 13000 may be combined into fewer components and modules, or further separated into additional components and modules.

[0290] The computing device 13000 can be controlled and coordinated by operating system software (e.g., 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 can be any available operating system, such as MAC OS X. In other embodiments, the computing device 13000 can be controlled by a proprietary operating system. Conventional operating systems control and schedule computer processes for execution, perform memory management, provide file systems, networks, I / O services, and provide user interfaces, such as graphical user interfaces (GUIs), etc.

[0291] The exemplary computing device 13000 may include one or more I / O interfaces and devices 13110, such as 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 (e.g., a touch screen or a monitor) that allow data to be visually presented to a user. More specifically, for example, a display device can provide presentation of a GUI, application software data, and a multimedia presentation. The computing system 13000 may also include one or more multimedia devices 13140, such as a camera, a speaker, a video card, a graphics accelerator, and a microphone.

[0292] In one embodiment of the computing system and application tool, I / O interfaces and devices 13110 can provide communication interfaces to various external devices. In one embodiment, the computing device 13000 is electrically coupled to a network 13002, such as via a wired, wireless, or combination of wired and wireless communication links 13115, and the network 13002 includes one or more of a local area network, a wide area network, and / or the Internet. The network 13002 can communicate with various sensors, computing devices, and / or other electronic devices via wired or wireless communication links.

[0293] In some embodiments, filtering criteria, signals, and data are processed by a cardiac rhythm inference module of an application tool according to the methods and systems described herein, and may be provided to a computing system 13000 from one or more data sources 13010 via a network 13002. The data source may include one or more internal and / or external databases, data sources, and physical data storage. The data source 13010, the external computing system 13004, and the cardiac rhythm interface module 13190 may include a database for storing data (e.g., feature data, raw signal data, patient data) according to the above-described systems and methods, and a database for storing data that has been processed according to the above-described systems and methods (e.g., data to be transmitted to a sensor, data to be sent to a clinician). Fig.24 In one embodiment, in some embodiments, sensor data 14050 can store data received from sensors, data received from clinicians, etc. In some embodiments, rule database 14060 can store data (e.g., instructions, preferences, profiles) that establish parameters for analyzing thresholds for feature data. In some embodiments, one or more databases or data sources can use relational databases (e.g., Sybase, Oracle, CodeBase, MySQL, SQLite, and SQL Server) and other types of databases (e.g., flat file databases, entity-relational databases, object-oriented databases, NoSQL databases, and / or record-based databases).

[0294] In one embodiment, the computing system includes a cardiac rhythm interface module 13190, which can be stored in a mass storage device 13120 as executable software code executed by a CPU 13105. The cardiac rhythm interface module 13190 can have a feature module 14010, a substitute data module 14020, a reasoning module 14030, a feedback module 14040, a sensor data database 14050, and a rule database 14060. These modules can include, for example, components, such as software components, object-oriented software components, subroutines, program code segments, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. These modules are also configured to perform the processes disclosed herein, which in some embodiments include the following: Figures 10 to 17 Describe the process.

[0295] Lossless compression and / or quantization

[0296] In some embodiments of the present disclosure, the system may perform lossless compression on the output of a first subset of the layers of the neural network before transmitting the data to a remote server to process the data through a second subset of the layers of the neural network. The lossless compression algorithm may allow the original data to be perfectly reconstructed from the compressed data. In some embodiments, the system may perform compression where the remote server can reconstruct the compressed data with a certain threshold of accuracy, thereby reducing losses.

[0297] Fig.25 An example of an embodiment of lossless compression using an ECG encoder is shown. Patch 2226 can receive ECG data 2232 of a user. The monitor (e.g., patch 2226) can encode the data through a first subset 2222 of layers of a neural network (e.g., via ECG encoder 2230) and transmit the encoded output to a lossless compressor 2502. The lossless compressor 2502 can compress the data to a smaller dimension for transmission to a remote server 2228. The remote server 2228 can decompress the data via a lossy decompressor 2504. The remote server 2228 can then process the decompressed data through a second subset 2224 of layers of a neural network via a classifier 2238.

[0298] In one example, a lossless compression scheme can be used to map between output coded features and shorter length representations. For an 8-bit integer, the value 0 as a binary representation "00000000" can be an output coded feature of the first subset of the layers of the neural network of the ECG encoder 2230, but the lossless compression scheme can change the "00000000" 8-bit digital representation to something else. The lossless compression scheme can have a mapping between the values ​​output from the ECG encoder 2230 and the compressed values. For example, the "0000000" 8-bit digital representation from the ECG encoder 2230 can be represented by "10", which only requires 2 bits instead of 8 bits. The lossless compression scheme can use shorter bit length codes for more common output coded features, and longer bit length codes for rarer coded features. Advantageously, less data is transmitted to the remote server. For example, if there are many more 0 values ​​output from the ECG encoder 2230, the lossless compression scheme can map the 0 value to "10". However, if a value of 433 is rarely output from the ECG encoder 2230, a lossless compression scheme may map the value of 433 having a binary representation of "110110001" to "5523".

[0299] In some embodiments, the lossless decompressor 2504 can retrieve a predetermined mapping value. In some embodiments, the lossless decompressor 2504 can retrieve a static and unchanging mapping value. The lossless decompressor 2504 can receive the mapping value from the lossless compressor 2502. The remote server 2228 can map the compressed data (e.g., the mapping value) to the actual value corresponding to the ECG encoder 2230. For example, if the lossless decompressor 2504 receives a "10" value and a "5523" value, the lossless decompressor 2504 can map these values ​​to "0" and "443" and output these values ​​accordingly. These values ​​can then be processed by a second subset of the layers in the neural network via the classifier 2238.

[0300] Advantageously, for a system that (1) encodes data via a neural network and (2) compresses the encoded data, the dimensionality of the data transmitted from the patch to the server can be smaller than for a system that (1) only has a neural network encoder (no lossless compression). In addition, the use of lossless compression enables more data to be transmitted to the remote server 2228 while still maintaining the integrity of the data itself because the data can be reconstructed via a lossless decompressor. To illustrate, if there are zero-value maps that are very close to each other, the system can transmit a representative indication of the zero-value map without having to transmit all zero values ​​separately. Advantageously, the data transmitted from the patch to the server can use less network throughput and can work under tighter network constraints without sacrificing the accuracy and performance of the neural network. In addition, this compression can be another form of encryption, which can further enhance the privacy and security of the data when the data is transmitted from the patch 2226 to the remote server 2228 over various networks.

[0301] As described herein, some embodiments of the present disclosure train a neural network to compress data to reduce the dimensionality of the output, for example, via an ECG encoder 2230, using a first subset of layers of the neural network. The first subset of layers of the neural network is trained to output a certain dimensionality. The compressed data is transmitted to a remote device, which applies the output of the first subset of layers to a second subset of layers of the neural network, which may include the steps described herein, such as decoding the data or determining a physiological characteristic of the data.

[0302] In traditional neural networks, the systems use large representations, such as floating point numbers (e.g., 32-bit data). Floating point numbers can represent very large or very small values ​​with many decimal places of accuracy. These traditional systems use such large representations to train neural networks and process data very accurately. However, using such large number representations leads to inefficiencies in the storage, processing, and transmission of data.

[0303] Some embodiments of the present disclosure alleviate these shortcomings by quantizing the data. Fig.26 A quantizer 2602 is shown performing quantization on the output data of the ECG encoder 2230 after processing a first subset 2222 of layers in the neural network. The patch 2226 can receive the ECG data 2232 of the user. The monitor (e.g., patch 2226) can encode the data through the first subset 2222 of layers of the neural network (e.g., via the ECG encoder 2230) and transmit the encoded output to the quantizer 2602. The quantizer 2602 can quantize the data to be transmitted to the remote server 2228 (e.g., truncate or round the data). The remote server 2228 can then process the quantized data through the second subset 2224 of layers of the neural network.

[0304] This quantization can be optimized to have better lossless compression performance while balancing the overall performance of the neural network. The system can quantize the outputs of a first subset of the layers of the neural network before transmitting the data to an external server. In some embodiments, quantizing the data can include rounding (e.g., rounding to the nearest integer), truncating, or reducing the number of bits represented by the data.

[0305] In some embodiments, the amount of quantization can be determined to optimize the balance between efficiency and accuracy. More aggressive quantization (e.g., truncating more bits from the data) can reduce the size of the data representation to fewer bits, but can negatively impact the accuracy of the neural network predictions. However, less aggressive quantization (e.g., truncating fewer bits from the data) can improve accuracy, but at the expense of increased storage and transmission requirements for the on-person monitoring device, increased storage requirements for having to store larger data representations on the patch 2226 and the remote server 2228, and increased network transmission of data between the patch 2226 and the remote server 2228.

[0306] As noted herein, more aggressive quantization can introduce errors into the quantized data, which can result in reduced performance of the algorithm using the quantization. While a neural network can work to a certain extent based on the errors introduced by quantization, at a certain point the errors can significantly reduce the performance of the algorithm. Therefore, the amount of quantization can be optimized based on the reduction in hardware efficiency and the accuracy of the neural network as it is classified by the classifier 2238.

[0307] In some embodiments, the neural network is trained to allow for more aggressive quantization schemes. For example, the neural network can be trained to maximize lossless compression performance. The neural network can be trained by introducing quantization during training to reduce degradation associated with quantization. In some embodiments, the neural network is initially trained without quantization. Then, the trained neural network is trained again using quantization. Advantageously, the training of the neural network takes quantization into account. This is particularly relevant here, where the neural network is divided into a first subset of layers running on the patch 2226 and a second subset of layers running on the remote server 2228. Advantageously, the results of such training can include a neural network and / or quantization that can better represent low-amplitude signals from quantization. Conversely, the neural network can adjust the numerical value of its output (e.g., increase the amplitude of the low-amplitude signal) to ensure that the low-amplitude signal is considered, such as by the classifier 2236 of the remote server 2228 to identify features in the signal. For example, the neural network can amplify the low-amplitude signal so that the signal reaches a threshold. Therefore, even if the low-amplitude signal is quantized, the remote server 2228 can still register the low-amplitude signal. Other systems without such neural network training may simply clip low amplitude signals via quantization, and classifier 2238 may fail to recognize features in these signals.

[0308] In some embodiments, lossless compression and quantization can be optimized and applied. In some embodiments, the best lossless compression instead of neural network and quantization can be determined by modeling the performance of different lossless compression algorithms or their characteristics. For example, the feature output of the neural network can be set to have a distribution centered at 0 and with a standard deviation of 1. The system can then model the performance of certain lossless compression algorithms under various quantization schemes. For example, if the output of the first subset of the layers of the neural network is 128 bits per second, the quantizer can quantize the data to output 100 bits per second. The optimization can be controlled at least by the available network and / or network constraints (e.g., the number of bits per second or the bit size of the data). The lossless compression model (e.g., as described herein) can then be applied to find the best performance for the recognition features of the neural network at the remote server.

[0309] In some embodiments, the optimization may be dictated at least by the architecture of the encoder. For example, a first encoder design may output 8 bits and a second encoder design may output 4 bits, but the output is sampled at twice the 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 has less aggressive quantization and fewer output features. However, in some cases, the second encoder design may perform better despite the higher quantization due to a higher feature generation rate.

[0310] In some embodiments, the quantization-to-train implementation can apply a tensorflow implementation. During training, tensorflow can introduce quantization in the forward pass, but the gradients will be calculated without quantization. This is done so that quantization errors are propagated through the network, but the gradients are still smooth and allow weight updates to respond appropriately to the errors.

[0311] Fig. 27 An embodiment is shown having both a quantizer 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 quantizer 2602 after the encoder has passed the data through a first subset 2222 of layers of a first neural network. The quantizer 2602 quantizes the encoded output data. The lossless compressor 2502 can then compress the quantized data to be transmitted 2234 to a remote server 2228. The remote server 2228 can then decompress the transmitted data 2234 via the lossless decompressor 2504, and the classifier 2238 can process the decompressed data by applying a second subset 2224 of layers of the network.

[0312] Analysis of QT interval algorithm

[0313] In some embodiments of the present disclosure, the system can quantify the QT interval, which can include quantifying the time between Q and T morphology features present in an individual's heartbeat. The extended time between these features can indicate a heart condition.

[0314] In traditional systems, quantifying the QT interval is usually accomplished by first identifying different morphological features and simply measuring the difference between specific points on the features. However, systems that do not have access to the complete raw ECG data may not be able to apply traditional methods of quantifying the QT interval.

[0315] Some embodiments are described herein that can quantify the QT interval without access to the complete raw ECG signal. In some embodiments, a remote server can receive encoded features from an encoder and can reconstruct the ECG signal from the encoded features. The remote server can then use signal processing and algorithmic techniques to locate the morphological features.

[0316] In some embodiments, the remote server may train a machine learning algorithm to predict the QT interval directly from the encoded feature segments containing the QT interval.

[0317] In some embodiments, the remote server may train a machine learning algorithm that predicts the average QT interval directly from a window containing encoding features of one or more QT intervals.

[0318] In some embodiments, the system can generate a template jump representing several jumps. The template jump can be generated by averaging samples of a fixed window around the detected jump position. The QT interval can be derived from the template jump using the algorithmic techniques described herein or a machine learning algorithm. Fig.28A and Fig.28B The process of generating template jumps is shown. For example, Fig.28A The sample window is shown over 4 cycles. Each cycle has a fixed sample window. The system can determine the number of cycles in a time period or set the number of cycles to be used to generate the template jump. Fig.28B The template jitter is generated from 4 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 sample together. Thus, Fig.28B The 4 cycles shown on the left are averaged to generate the template jump on the right.

[0319] In some embodiments, the patch may generate a template beating. In other embodiments, the remote server may generate a template beating based on data received from the patch. In some embodiments, the template beating may be processed and identified via data processing by layers of a neural network. For example, the patch may generate a template beating and input the template beating to a neural network described herein.

[0320] Generally, as used herein, the term "module" refers to logic implemented in hardware or firmware, or to a collection of software instructions written in a programming language (e.g., Python, Java, Lua, C, and / or C++) that may have entry and exit points. Software modules can be compiled and linked into an executable program, installed in a dynamic link library, or can be written in an interpretive programming language such as BASIC, Perl, or Python. It is understood that software modules can be called from other modules or themselves, and / or can be called in response to detected events or interrupts. Software modules configured to run on a computing device can be set on a computer-readable medium (e.g., a compact disc, a digital video disc, a flash drive, or any other tangible medium). Such software code can be stored in part or in whole on a memory device that runs a computing device (e.g., computing system 13000) for the computing device to run. Software instructions can be embedded in firmware (e.g., EPROM). It will also be understood that a hardware module can include connected logic units, such as gates and triggers, and / or can include programmable units, such as programmable gate arrays or processors. The block diagram disclosed herein can be implemented as a module. The modules described herein may be implemented as software modules, but may be represented by hardware or firmware. Generally, the modules described herein refer to logical modules that can be combined with other modules or divided into sub-modules, regardless of their physical organization or storage method.

[0321] Each of the processes, methods, and algorithms described in the foregoing sections may be implemented in a code module executed by one or more computer systems or computer processors including computer hardware, and implemented in whole or in part by it. The code module may be stored on any type of non-transitory computer-readable medium or computer storage device, such as a hard drive, solid-state memory, optical disk, etc. The system and module may also be transmitted on various computer-readable transmission media (including wireless-based and wired / cable-based media) as a generated data signal (e.g., as part of a carrier or other analog or digital propagation signal), and may take a variety of forms (e.g., as part of a single or multiplexed analog signal, or as multiple discrete digital packets or frames). The process and algorithm may be implemented in part or in whole in a dedicated circuit. The results of the disclosed processes and process steps may be stored persistently or otherwise in any type of non-transitory computer memory (e.g., volatile or non-volatile memory).

[0322] The various features and processes described above can be used independently of each other, or can be used in combination in various ways. All possible combinations and sub-combinations are intended to fall within the scope of the present disclosure. In addition, in some embodiments, certain method or process blocks can be omitted. The methods and processes described herein are also not limited to any particular order, and the blocks or states associated therewith can be performed in other appropriate orders. For example, the blocks or states described can be performed in a different order than that specifically disclosed, or multiple blocks or states can be combined in a single block or state. The example blocks or states can be performed in series, in parallel, or in some other way. Blocks or states can be added to or removed from the disclosed example embodiments. The example systems and components described herein can be configured differently from those described. For example, elements can be added, removed, or rearranged compared to the disclosed example embodiments.

[0323] Unless explicitly stated otherwise, or otherwise understood in the context in which it is used, conditional language (e.g., "may," "could," "might," or "might," etc.) is generally intended to convey that some embodiments include, while other embodiments do not include, certain features, elements, and / or steps. Thus, such conditional language is generally not intended to imply that features, elements, and / or steps are in any way necessary for one or more embodiments, or that one or more embodiments necessarily include logic for determining whether such features, elements, and / or steps are included or performed in any particular embodiment with or without user input or prompting. The term "includes" means "including but not limited to." The term "or" means "and / or."

[0324] Any process description, element or block in the flowchart or block diagram described herein and / or depicted in the accompanying drawings should be understood to potentially represent a module, segment or code portion that includes one or more executable instructions for implementing specific logical functions or steps in the process. Alternative implementations are included within the scope of the embodiments described herein, in which elements or functions may be deleted, executed in an order other than that shown or discussed, including substantially simultaneously or in reverse order, depending on the functions involved, as will be understood by those skilled in the art.

[0325] All of the above methods and processes may be implemented at least in part in a software code module run by one or more computers, and may be implemented partially or completely via a software code module. For example, the method described herein may be performed by a computing system and / or any other suitable computing device. The method may be run on a computing device in response to the execution of software instructions or other executable codes read from a tangible computer-readable medium. A tangible computer-readable medium is a data storage device that can store 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.

[0326] It should be emphasized that many variations and modifications may be made to the above-described embodiments, and elements thereof should be understood to be other acceptable examples. All such modifications and variations are intended to be included within the scope of the present disclosure. The foregoing description describes certain embodiments in detail. However, it should be understood that no matter how detailed the foregoing appears in the text, the systems and methods can be implemented in a variety of ways. For example, features of one embodiment can be used together with features in different embodiments. As described above, it should be noted that the use of a specific term in describing certain features or aspects of the systems and methods should not be taken as implying that the term is redefined herein to be limited to including any specific characteristics of the features or aspects of the systems and methods associated with the term.

[0327] 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 various features of the embodiments may be made without departing from the scope of the invention. For example, in some cases, the order of the various method steps may be changed, and / or one or more optional features may be added to or deleted from the described devices. Therefore, the description of the embodiments provided above should not be interpreted as improperly limiting the scope of the invention as set forth in the claims.

[0328] Various modifications may be made to the embodiments described in the present disclosure, and the general principles defined herein may be applied to other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the scope of the present disclosure is not intended to be limited to the embodiments shown herein, but is to be consistent with the widest scope consistent with the present disclosure, the principles and novel features disclosed herein.

[0329] Certain features described in this specification in the context of a single embodiment may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually in multiple embodiments, or in any applicable sub-combination. In addition, although features may be described above as functioning in certain combinations and even initially claimed as such, in some cases one or more features of the claimed combination may be excluded from the combination, and the claimed combination may be directed to a sub-combination or a variation of the sub-combination.

[0330] Similarly, although operations are depicted in a particular order in the accompanying drawings, these operations do not need to be performed in the particular order shown or in sequence, or all the operations shown need to be performed to obtain the desired results. In addition, the accompanying drawings may schematically depict another example process in the form of a flow chart. However, other operations not depicted may be incorporated into the schematically illustrated example process. For example, one or more additional operations may be performed before, after, simultaneously, or between any of the illustrated operations. In addition, the separation of various system components in the above-described embodiments should not be interpreted as requiring such separation in all embodiments. In addition, other embodiments are within the scope of the claims. In some cases, the actions described in the claims may be performed in different orders and still achieve the desired results.

Claims

1. A monitor for inferring the likelihood of a user feature, the monitor comprising: a sensor positioned to detect continuous physiological signals of a user when the monitor is engaged to the user; a hardware processor configured to process the detected physiological signal via an encoder to produce a data output, wherein the encoder includes a first subset of layers of a neural network; as well as A transmitter is configured to transmit the data output of the encoder to an external computing system for processing the data output by a decoder to infer the likelihood of an arrhythmia occurring, wherein the decoder includes a second subset of the layers of the neural network.

2. The monitor of claim 1, wherein the transmitter transmits the data output to the external computing system via a wireless communication channel.

3. The monitor of claim 1, wherein the monitor is a wearable patch.

4. The monitor of claim 1, wherein the monitor is a patch that is attached to the user's chest.

5. The monitor of claim 1, wherein the monitor is an implantable device.

6. A monitor according to claim 1, wherein the monitor includes another sensor, the other sensor is configured as an accelerometer, the accelerometer is configured to collect movement data, and the external computing system is configured to match the movement data with the detected physiological signals to infer the possibility of an arrhythmia.

7. The monitor of claim 1, wherein the sensor comprises one or more electrodes and the physiological signal comprises electrocardiogram (ECG) data.

8. The monitor of claim 1, wherein the sensor comprises an optical sensor and the physiological signal comprises photoplethysmogram (PPG) data.

9. The monitor of claim 1, wherein a first subset of layers of the neural network and a second subset of layers of the neural network are trained together.

10. The monitor of claim 9, wherein the monitor comprises a waterproof or water resistant housing.

11. The monitor of claim 9, wherein the monitor includes a surface on the housing configured to engage the user.

12. The monitor of claim 9, wherein the sensor is permanently disposed within the housing.

Images