Non-invasive cardiac monitor and methods for inferring or predicting patient physiological characteristics

Abstract

The present disclosure relates to a non-invasive cardiac monitoring device that records cardiac data to infer the onset of a human physiological characteristic, such as the onset of cardiac arrhythmia, and recommends treatment / intervention. Some embodiments include clustering patient data history to determine efficacy of treatment based on cluster groups corresponding to the patient data. In further embodiments, processing of detected cardiac heart rate signals can be partially processed in the wearable cardiac monitoring device and partially processed in a remote computing system. Some embodiments include a wearable cardiac monitoring device that is attached to a patient for extended periods of time to enable long-term detection of cardiac heart rate signals.

Description

Background Art

[0001] Cross - reference to Related Applications This application claims the benefit of the filing date of U.S. Provisional Patent Application No. 63 / 265,335, filed on December 13, 2021, which is hereby incorporated by reference in its entirety.

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

Summary of the Invention

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

[0004] Some embodiments described herein may include a computing system for inferring an onset of a user's cardiac arrhythmia, the computing system including: a sensor configured to detect cardiac signals from the user; and a hardware processor configured to apply the detected cardiac signals to a first machine learning process, the first machine learning process configured to infer an onset of a cardiac arrhythmia based on the detected cardiac signals from the user, the first machine learning process comprising: accessing historical cardiac signal patient data for multiple patients; accessing cardiac arrhythmia situations for multiple patients, the cardiac arrhythmia situations for multiple patients being determined after the time the historical cardiac signal patient data was recorded; and training the first machine learning process based on the historical cardiac signal patient data and the cardiac arrhythmia situations in order to infer an onset of a cardiac arrhythmia, the user being trained by the first machine learning process, the user being different from the multiple patients.

[0005] In some embodiments, the hardware processor is further configured to apply a second machine learning process, which is trained to remove or detect noise segments or artifacts from the user's detected cardiac signal, and the first machine learning process receives the cardiac signal from which the noise segments or artifacts have been removed as input. In some embodiments, the hardware processor is further configured to apply a second machine learning process, which is trained to reduce noise from the user's detected cardiac signal, and the first machine learning process receives the cardiac signal from which the noise has been reduced as input. In some embodiments, the onset of cardiac arrhythmia includes the probability of confirming cardiac arrhythmia at a later point in time. In some embodiments, at least a portion of the historical cardiac signal patient data does not show cardiac arrhythmia at the time the historical cardiac signal patient data was recorded. In some embodiments, the first machine learning process is further configured to process photoplethysmography (PPG) data, heart rate, or acceleration data to infer the onset of cardiac arrhythmia. In some embodiments, cardiac arrhythmias include at least one of ventricular tachycardia, supraventricular tachycardia, abduction, or ventricular fibrillation. In some embodiments, the computing system includes a wearable smartwatch. In some embodiments, the sensors are electrodes.

[0006] Some embodiments include an electronic device for monitoring a user's physiological signals, the electronic device comprising: an adhesive assembly comprising a housing and a flexible wing, the housing comprising a circuit board, the flexible wing extending from the housing and configured to conform to the user's surface; a sensor coupled to the flexible wing, the sensor electrically communicating with the circuit board and configured to be positioned to conform to and contact the user's surface in order to detect physiological signals; and one or more hardware processors configured to apply the detected physiological signals to a machine learning process, the machine learning process being configured to infer the onset of cardiac arrhythmias based on the detected physiological signals, the machine learning process being trained by: accessing historical cardiac signal patient data for multiple patients; accessing cardiac arrhythmia situations for multiple patients, the cardiac arrhythmia situations for multiple patients being determined since the time the historical cardiac signal patient data was recorded; and training the machine learning process based on the historical cardiac signal patient data and cardiac arrhythmia situations to infer the onset of cardiac arrhythmias, the user being different from the multiple patients.

[0007] In some embodiments, inferring the onset of cardiac arrhythmia includes generating a cardiac arrhythmia risk score at a point in time after a physiological signal has been detected by a sensor. In some embodiments, one or more hardware processors are further configured to determine that cardiac arrhythmia is not currently present in the detected physiological signal before inferring the onset of cardiac arrhythmia. In some embodiments, inferring the onset of cardiac arrhythmia is in response to determining that cardiac arrhythmia is not currently present in the detected physiological signal. In some embodiments, one or more hardware processors are further configured to adjust the function of the electronic device based on the inferred onset of cardiac arrhythmia. In some embodiments, adjusting the function of the electronic device includes adjusting the window of physiological signals processed by the electronic device. In some embodiments, one or more hardware processors are further configured to send an alert to the user or physician's device in response to the risk score meeting a threshold score. In some embodiments, one or more hardware processors are further configured to recommend treatment or intervention for the onset of cardiac arrhythmia. In some embodiments, recommending treatment or intervention is based on clustering of previous patient data. In some embodiments, a cluster includes multiple centroids, and one or more hardware processors apply a distance process from the centroids of the corresponding clusters to the user's records to recommend treatment. In some embodiments, the clustering process clusters patients based on the status of cardiac arrhythmias, the type of intervention, and the outcomes of the application of the intervention type.

[0008] Some embodiments include a method for training a machine learning process to infer the onset of cardiac arrhythmias, the method including: accessing past cardiac signal patient data of multiple patients, wherein at least a portion of the past cardiac signal patient data does not show cardiac arrhythmias at the time the cardiac signals were recorded; accessing cardiac arrhythmia circumstances for multiple patients, wherein the circumstances for multiple patients are determined after the time the cardiac signals were recorded; and training a first machine learning process based on the past cardiac signal patient data and cardiac arrhythmia circumstances to infer the onset of cardiac arrhythmias, wherein the user differs from that of the multiple patients.

[0009] In some embodiments, training a first machine learning process further includes the step of training the first machine learning process to infer hospitalization, heart failure onset, stroke onset, or death condition. In some embodiments, the method further includes training a second machine learning process to infer hospitalization, heart failure onset, stroke onset, or death condition, wherein the input to the second machine learning process includes cardiac signals over a longer period than those input to the first machine learning process. In some embodiments, the method further includes determining an atrial fibrillation burden from the detected cardiac signals, wherein the atrial fibrillation burden includes the amount of time the user spent in atrial fibrillation during a given period.

[0010] Some embodiments include a method comprising the steps of: detecting a cardiac signal from a user; applying the detected cardiac signal to a first machine learning process, the first machine learning process being configured to infer the onset of cardiac arrhythmias based on input data, the first machine learning process comprising: accessing historical cardiac signal patient data for multiple patients; accessing cardiac arrhythmia situations for multiple patients, the situations for multiple patients being determined after the time the historical cardiac signal patient data was recorded; and training the first machine learning process based on the historical cardiac signal patient data and cardiac arrhythmia situations to infer the onset of cardiac arrhythmias in a user, the user being trained by the steps of:

[0011] These and other aspects and embodiments of the present invention will be described in further detail below with reference to the drawings. [Brief explanation of the drawing]

[0012] [Figure 1A] Figure 1A is a perspective view of an example of a physiological monitoring device. [Figure 1B] Figure 1B is an exploded view of an example of a physiological monitoring device. [Figure 2A] Figure 2A is a top perspective view of an example of a physiological monitoring device. [Figure 2B] Figure 2B is a bottom view of an example of a physiological monitoring device. [Figure 2C] Figure 2C is a top perspective view including the liner of an example of a physiological monitoring device. [Figure 2D] Figure 2D is a bottom view including the liner of an example of a physiological monitoring device. [Figure 3A] Figure 3A is a perspective view of an example of a physiological monitoring device. [Figure 3B] Figure 3B is a top view of an example of a physiological monitoring device. [Figure 3C]FIG. 3C is a bottom view of an example of a physiological monitoring device. [Figure 3D1] FIG. 3D1 is a side view of an example of a physiological monitoring device. [Figure 3D2] FIG. 3D2 is a side view of a ridge for sealing the upper and lower parts of the housing of an example of a physiological monitoring device. [Figure 3E] FIG. 3E is a bottom view of an example of a physiological monitoring device with layers made transparent to show a perspective view of the device. [Figure 3F] FIG. 3F is a top view of an example of a physiological monitoring device with layers made transparent to show a perspective view of the device. [Figure 3G] FIG. 3G is an exploded view of various components of an example of a physiological monitoring device. [Figure 3H] FIG. 3H is an exploded view of various components of an example of a physiological monitoring device. [Figure 4A] FIG. 4A is a schematic diagram of an embodiment of a system for predicting rhythm annotation using neural network encoding. [Figure 4B] FIG. 4B is a schematic diagram of an embodiment of the first and second subsets of layers within a single neural network. <第 [Figure 5] [[ID=第27]]FIG. 5 is a flowchart diagram of an embodiment of a system and an external monitor. [Figure 6] FIG. 6 shows a schematic diagram of an embodiment of an autoencoder. [Figure 7A] FIG. 7A is a schematic diagram of an embodiment applying an LSTM transducer block to the output of a decoder. [Figure 7B] FIG. 7B shows an autoencoder architecture of an embodiment applying an LSTM transducer block to the output of an encoder. [Figure 7C] FIG. 7C shows a self-attention architecture for inferring the occurrence of cardiac arrhythmia. [Figure 7D] It should be noted that there is an incorrect "第0000081" and "第27" in the original text which are likely errors in numbering or tagging. I have translated them as is while pointing out the potential issue.Figure 7D shows a network training architecture for applying an LSTM transducer block to the output of a decoder. [Figure 7E] Figure 7E shows a network inference architecture for applying an LSTM transducer block to the output of a decoder to generate hazards and / or survival probabilities. [Figure 7F] Figure 7F shows a network training architecture for applying a self-attention architecture using a loss function. [Figure 7G] Figure 7G shows a network inference architecture for applying a self-attention architecture to generate hazards and / or survival probabilities. [Figure 8] Figure 8 shows an embodiment of grouping individual signatures (represented as projections on a 2D plane) from individuals that may have different progressions for an event time frame or that respond well to a particular intervention. [Figure 9] Figure 9 shows a schematic of a risk score alert for patients in which markers indicating a high confidence of the onset / detection of atrial fibrillation (AF) during a pre-specified follow-up period are included in the record, although no event occurred during the current patch wear time. [Figure 10] Figure 10 shows a graph according to an embodiment of a risk curve stratified by demographic or clinical characteristics. [Figure 11] Figure 11 shows an example of a report for a patient in whom atrial fibrillation has not been confirmed. [Figure 12A] Figure 12A shows an example of a report for a patient having atrial fibrillation identified in a heart monitor record. [Figure 12B] Figure 12B shows an example of a report for a patient having atrial fibrillation identified in a heart monitor record. [Figure 13A] Figure 13A shows an example of a flowchart for training an encoder and a decoder and applying the encoder to a wearable device to infer the onset of cardiac arrhythmia. [Figure 13B]Figure 13B shows an example of a flowchart for training an encoder and a machine learning model to infer the onset of cardiac arrhythmias, applying the encoder on a wearable device, and applying the machine learning model on a remote computing device. [Figure 13C] Figure 13C shows an example of a flowchart for training an encoder, decoder, and machine learning model using historical patient cardiac data to infer the onset of cardiac arrhythmias. [Figure 14] Figure 14 is a schematic diagram of one embodiment of a computer network system. [Figure 15] Figure 15 is a schematic diagram of one embodiment of a programming and distribution module. [Modes for carrying out the invention]

[0013] The following description is directed towards numerous different embodiments. However, the embodiments described can be implemented and / or modified in many different ways. For example, the embodiments described can be implemented in any suitable device, instrument, or system for monitoring any of numerous physiological parameters. For example, the following discussion will focus primarily on long-term patch-based heart rate monitoring devices. In one alternative embodiment, the physiological monitoring device may be used, for example, for the diagnosis of pulse oximetry and obstructive sleep apnea. The methods of using the physiological monitoring device also vary. In some embodiments, the device can be worn for less than a week, while in other embodiments, the device can be worn for at least 7 days and / or more than 7 days, for example, 14 to 21 days or more.

[0014] Abnormal heart rhythms and arrhythmias are often caused by other less serious factors, making it a crucial challenge to determine which of these symptoms is due to arrhythmia. Often, arrhythmias are infrequent and / or paroxysmal, making rapid and reliable diagnosis difficult. Currently, heart rate monitoring is primarily performed using devices such as Holter monitors, which involve attaching electrodes to the chest for short periods (less than one day). The electrodes are connected to a recording device via wires and are usually worn on a belt. Electrodes need to be replaced daily, and the wires are cumbersome. Furthermore, the device has limitations in memory and recording time. Wearing the device restricts patient movement, often preventing certain activities during monitoring, such as bathing. Additionally, Holter monitors are capital equipment with limited availability, frequently leading to supply constraints and resulting delays in testing. Such limitations significantly hinder the diagnostic usefulness of the device, patient compliance, and the possibility of capturing all crucial information. Lack of compliance and device shortcomings often lead to the need for additional devices, follow-up monitoring, or other tests to make a correct diagnosis.

[0015] Current methods for linking symptoms to arrhythmias, including the use of heart rate monitoring devices such as Holter monitors and cardiac event recorders, are often insufficient for making accurate diagnoses. In fact, studies have shown that Holter monitors are misleading in 90% of cases (DE Ward et al., "Assessment of the Diagnostic Value of 24-Hour Ambulatory Electrocardiographic Monitoring," Biotelemetry Patient Monitoring, vol. 7, published 1980).

[0016] Furthermore, the medical process from obtaining a heart rate monitoring device to actually starting monitoring is generally very complex. Typically, ordering, tracking, monitoring, searching for, and analyzing data from such monitoring devices involves numerous steps. Currently, cardiac monitoring devices are almost always ordered by a cardiologist or electrophysiologist (EP), not the patient's primary care physician (PCP). This is important because the PCP is often the first doctor to see the patient and determine if their symptoms may be due to arrhythmia. After the patient sees the PCP, the PCP schedules an appointment for the patient to see a cardiologist or EP. This appointment is usually made several weeks after the initial PCP visit, which itself can lead to delays in diagnosis and increase the likelihood that arrhythmia episodes may go undiagnosed. When the patient finally sees a cardiologist or EP, a heart rate monitoring device is usually prescribed. Monitoring periods are typically 24-48 hours (Holter monitor) or up to one month (cardiac event monitor or mobile telemetry device). Once monitoring is complete, the patient usually has to return the device to the clinic. After the data is processed by a monitoring company or technicians at the hospital or office, it is finally sent to a cardiologist or EP for analysis. As a result of this complex process, the number of patients receiving cardiac rhythm monitoring is lower than the ideal number.

[0017] To address some of these issues relating to cardiac monitoring, the assignee of this application has developed various embodiments of a small, long-lasting physiological monitoring device. One embodiment of this device is the Zio® patch and the Zio® monitoring platform / system. Various embodiments are also described, for example, in U.S. Patent Nos. 8,150,502, 8,160,682, 8,244,335, 8,560,046, 8,538,503, 9,597,004, and 11,246,524, the full disclosures of which are incorporated herein by reference. Generally, the physiological patch monitors described in the above documents are designed to be comfortably worn on the patient's chest and to be worn for at least one week, typically two to three weeks. The monitor continuously detects and records heart rate signal data while the device is worn, and this heart rate data is then available for processing and analysis.

[0018] Physiological monitoring device Figures 1A–1B show several figures of another example of the physiological monitoring device 100, such as the features described in U.S. Patents No. 11,350,864 and No. 11,337,632, which are incorporated in whole by reference. The physiological monitoring device 100 may include one or more of the components described elsewhere in this specification. The physiological monitoring device 100 may include a housing 115 comprising an upper housing 140 and a lower housing 145 configured to fit together with a flexible body 110 sandwiched between them. The flexible body 110 may include a trace layer 109 and one or more substrate layers that form the wings of the physiological monitoring device 100. The wings may include an adhesive layer 140 and electrodes 150, as described elsewhere in this specification. The rigid body and / or housing 115 can enclose the PCBA 120, the flexible upper frame 142, the battery 160, the battery terminal connector 150, a portion of the trace layer 109, the spring contact spacer 132, and the spring 165.

[0019] Figure 1A is a perspective view of an example of the physiological monitoring device 100. Figure 1B is an exploded view of the physiological monitoring device 600.

[0020] The upper housing 140 and the lower housing 145 can sandwich the flexible body 110 as described elsewhere in this specification. In some embodiments, the flexible body 110 may include one or more openings 132 extending through one or more substrate layers to provide ventilation and moisture management and / or to facilitate drug delivery to the surface skin, as described elsewhere in this specification. An upper gasket layer 160 and / or a lower gasket layer 170 (not shown) may be provided on opposing sides of the flexible body 110 (not shown). The gasket layers 160, 170 may be adhesives for bonding to the flexible body 110. A compressible seal may be formed above and / or below the flexible body 110. In some embodiments, the compressible seal may be formed together with the flexible upper frame 142. The battery 160 may be located below the flexible body 110, including the trace layer 109. PCBA120 may be positioned above the flexible body 110, which includes the trace layer 109. The battery terminal connector 150 may be bonded or otherwise attached to the battery 160 such that the first and second battery traces 152, 153 are exposed on the outer surface of the battery terminal connector 150 above the battery 160. The first and second battery traces 152, 153 may be exposed to the internal volume of the upper housing 140 through a large central opening in the housing region of the trace layer 109.

[0021] Electrical contact between PCBA120 and the first and second battery traces 152, 153, and / or between PCBA120 and the electrocardiogram interface portions 113 of the electrical traces 111, 112, can be established by spring contacts 137. The spring contacts 137 may be coupled to the bottom surface of PCBA120. The housing 115 may include a spring contact spacer 132 positioned below PCBA120. In some examples, the spring contact spacer 132 may be rigidly attached (e.g., glued) to the bottom surface of PCBA120. In some examples, the spring contact spacer is attached to or integrated with a flexible body 110. In some examples, the spring contact spacer may be integrated with the battery terminal connector. The spring contact spacer 132 may include a flat body and a plurality of downwardly extending legs 133. The legs 133 may be configured to seat against the top and / or side surfaces of the battery 160 so that the spring contact spacers 132 maintain a minimum separation distance between the battery 160 and the PCBA 120 and provide sufficient space for the spring contacts 137. The spring contact spacers 132 may include one or more holes 134 through which the spring contacts 137 can extend downward from the bottom surface of the PCBA 120. The lower housing 145 may include a spring 165 positioned below the battery 160, as described elsewhere in this specification. The spring 165 may bias the battery 160 upward and bias the first and second battery traces 152, 153 to make physical and electrical contact with the corresponding spring contacts 137. The electrocardiogram interface portions 113 of traces 111 and 112 can also be seated on the top surface of the battery 160 by biasing the battery 160 upward so that the electrocardiogram interface portions 113 of traces 111 and 112 also make physical and electrical contact with the corresponding spring contacts 137. The substantially constant spacing between the traces and the PCBA 120 provided by the springs 165 and spring contact spacers 132 can reduce, minimize, or eliminate electrical signal noise caused by variations in the degree of electrical contact between the spring contacts 137 and the traces.The assembly may include at least one spring contact 137 for each of the first battery trace 152, the second battery trace 153, the first electrical trace 111, and the second electrical trace 112. The assembly may include multiple spring contacts 137 for some or all of the traces. The spring contacts 137 may be configured to establish an electrical path between each trace and the PCBA 120 under compression induced by the arrangement of various components, including the spring 165. Since the spring contacts 137 extend further downward as the separation distance increases and the biasing force decreases accordingly, the compression contact between the spring contacts 137 and the traces can be maintained even under small changes in the separation distance between the traces and the PCBA 120 (e.g., caused by movement). In some embodiments, the first and second battery traces 152, 153 may be configured to be located on the opposite side of the housing 115 from the first and second electrical traces 111, 112. In some embodiments, spring contacts may be configured to transmit electrical signals or electrocardiogram signals from a battery by contacting electrical traces applied to the upper housing 140 or bottom housing 145. These electrical traces can be applied to the housing using laser direct structuring, plating on a platable substrate applied in a secondary molding process, or printing of conductive material by aerosol jet, inkjet, or screen printing. In some examples, RF antennas for wireless communication (such as Bluetooth®) may be configured through the use of such electrical traces in the upper housing 140 or bottom housing 145.

[0022] Figures 2A-2D depict several diagrams of an example of a physiological monitoring device 200 similar to the physiological monitoring device depicted in Figures 1A-1B. Here, the physiological monitoring device includes a central housing 202, which includes an upper housing 202 and a lower housing 206 sandwiched on a flexible substrate 210. Those skilled in the art will understand that the housing may be made of any suitable material disclosed herein, e.g., a rigid polymer or a flexible polymer. In some examples, the housing may include an indicator 208, which may be any suitable shape such as elliptical, circular, square, or rectangular. The indicator may include an LED light source (not shown) or any suitable light source and may be overlaid by a transparent or translucent viewing layer positioned against the inner surface of the upper housing. The viewing layer may be made of thermoplastic polyurethane or any suitable material. The indicator can be used to indicate the status of the physiological monitoring device, such as the battery life of the physiological monitoring device. In some examples, the indicator can show whether the physiological monitoring device is collecting data, transmitting data, paused, experiencing an error, or analyzing data. The indicator can display any appropriate color, such as red, amber, or green.

[0023] Extending outward from the housing are several wings 212. Those skilled in the art will understand that although two wings are depicted here, some embodiments of the physiological monitoring device 200 may include more than two wings. As described elsewhere in this specification, the wings may be shaped to improve adhesion to the skin and retention of the physiological monitoring device to the skin. In this example, the wings may be asymmetrical, with the majority of one wing (upper lobe) 214 lying above the longitudinal line and the majority of another wing (lower lobe) 216 lying below the longitudinal line, thereby allowing the physiological monitoring device to be positioned diagonally above the heart, such that the lower lobe is below the heart when the patient is in a standing position.

[0024] Extending outward from the housing and contained on or within the wings are electrode traces 218 similar to those described elsewhere in this specification, such as in Figure 1A. As described elsewhere in this specification, the electrode traces may be printed directly onto a flexible substrate, which may be part of a multilayer flexible assembly 220. For visual enhancement of the physiological monitoring device, additional printed lines 222 may surround the electrode traces 218, but these printed lines 222 may be printed on a different layer from the flexible substrate on which the electrode traces are printed. The printed lines may be printed to blend with the shape of the electrode traces. As described elsewhere in this specification, the electrode traces may surround a series of breathing holes 224 that allow air to pass through to the underlying hydrogel layer. In the example, there may be one, two, three, four, or more breathing holes. As described elsewhere in this specification, openings 226 may extend through one or more layers of the physiological monitoring device to provide air permeability and humidity control. In the embodiment, the adhesive boundary layer 228 may extend outward from the wings, thereby improving adhesion. Figure 2B depicts the underside of the physiological monitoring device 200 as shown in Figure 2A. Here, the lower housing 206 is clearly visible, as are the electrode traces 218 and printed lines 222 extending outward from the housing. Figures 2C and 2D depict the physiological monitoring device 200 of Figures 2A-2B, here including an outward-facing upper liner 226 and a skin-facing patient release liner 228 surrounding the housing 202 above the wings. Such a release liner serves to protect the physiological monitoring device 200 during storage, particularly the adhesive surfaces of the physiological monitoring device. In the embodiment, the liner may be shaped such that two sides meet, forming an opening for the housing to extend vertically beyond the liner.

[0025] In some embodiments, an abrader can be used to abrade the patient's skin before adhering the physiological monitoring device 200 (as described elsewhere in this specification) to the patient. The abrader can be used to remove the surface layer of skin from the patient in order to improve the long-term adhesion of the physiological monitoring device and / or the signal quality of the physiological monitoring device.

[0026] In various alternatives, the shape of a particular physiological monitoring device varies. The shape, footprint, perimeter, or boundary of the device may be, for example, circular, elliptical, triangular, or a compound curve. In some examples, a compound curve may include one or more concave curves and one or more convex curves. There may be concave portions between convex shapes. The concave portions may be between convex portions on the housing and convex portions on the electrodes. In some examples, the concave portions may at least partially correspond to hinges, hinge regions, or areas of reduced thickness between the body and wings.

[0027] Although described in the context of cardiac monitors, the improvements to the devices described herein are not limited to this. The improvements described in this application can be applied to any of the wide variety of physiological data monitoring, recording, and / or transmission devices. Furthermore, the features of the improved adhesive design can also be applied to devices useful for the electronic control and / or time-release delivery of pharmacological drugs and blood tests, such as glucose monitors and other blood testing devices. Thus, the descriptions, characteristics, and functions of the components described herein can be modified as necessary to include specific components for a particular application, such as electronics, antennas, power or charging connections, data ports or connections for downloading or offloading information from the device, adding or offloading fluids from the device, monitoring or sensing elements such as electrodes, probes or sensors or other components, or components required for device-specific functions. In addition, or alternatively, the devices described herein can be used to detect, record, or transmit signals or information related to signals generated by the body, including but not limited to one or more of ECG, EEG, and / or EMG. In certain examples, additional data channels may be included, for example, to detect the movement of the device, the bending of the device, or bending In addition, additional data such as heart rate and / or ambient electrical or acoustic noise can be collected.

[0028] The physiological monitors described above and elsewhere in this specification can be further combined with data processing and transmission methods and systems that improve data collection from the monitors. Furthermore, the methods and systems described below can improve the performance of the monitors by enabling timely transmission of clinical information while maintaining the high patient compliance and ease of use of the monitors described above. For example, the data processing and transmission methods and systems described in this section or elsewhere in this specification may help extend the battery life of the monitors, improve the accuracy of the monitors, and / or provide other improvements and benefits as described in this section or elsewhere in this specification.

[0029] Figures 3A–3H depict an example of a physiological monitoring device 300 similar to the physiological monitoring device described in U.S. Patent No. 11,350,864, which is incorporated by reference in whole. Figure 3A shows a perspective view of the physiological monitoring device. The physiological monitoring device 300 may include wings 330, 331, each asymmetrical with respect to the longitudinal axis, extending substantially between electrode interface portions 302 that cover electrodes located on the underside of the wings. Electrode traces 304 extend from the housing to the electrodes and provide electrical communication between the electrodes and the central housing. One wing 330 includes a body disproportionately positioned above the longitudinal axis, and the other wing 331 includes a body disproportionately positioned below the longitudinal axis. Accordingly, the wings 330, 331 can be made asymmetrical with respect to a transverse axis perpendicular to the longitudinal axis and extending through the housing 306, the transverse axis which may include a patient trigger 307, as with other patient triggers disclosed in this section of the Spec. As described elsewhere in the Spec., in certain embodiments, the patient trigger is about 10–30% of the total upper area, for example, about 20% of the total upper area, or about 23%, such as about 22.8% of the total upper area. In certain embodiments, the patient trigger may cover about 20% or more, about 30% or more, about 40% or more, about 50% or more, or about 75% or more. In some embodiments, the patient trigger may cover the entire upper surface of the housing. The wings 330, 331 may consist of the same shape inverted or inverted with respect to both the longitudinal and transverse axes, as shown in Figures 3A–3C. In some embodiments, the wings may be asymmetrical in size and shape; for example, the upper wing 330 may be larger than the lower wing 331, or vice versa. The relative shape of the upper wing 330 may differ from the relative shape of the lower wing 331, so that the shape of the wings 330, 331 differs from the relative shape of the lower wing 331.In some embodiments, the upper wing 330 may experience greater tension than the lower wing 331, or vice versa; therefore, different sizes and shapes between the two wings may help address inherent force vectors during the use of the physiological monitoring device. The wing configuration is particularly well-suited for positioning electrodes diagonally relative to the subject's height, which may reduce gravity-induced detachment. Those skilled in the art will understand that the orientation of the wings may be modified so that the wings are mirrored, rather than being unbalanced above or below the longitudinal axis. Furthermore, those skilled in the art will understand that the wing shapes described herein may differ from the generally rounded shapes depicted in Figures 3A–3H. For example, the wings may be angular, such as square, rectangular, triangular, pentagonal, or any suitable polygon. These polygons may have rounded corners to reduce the possibility of detachment from the corners. Liner 308, as depicted elsewhere in this specification, can be used to cover and protect any adhesives before applying the physiological monitoring device to a patient or user. In the example, the liner is separated into two parts, one on each wing.

[0030] In certain embodiments, an additional visualization pattern 310 may extend through the wing. The visualization pattern 310 may be of any appropriate size or shape to outline the electrode trace and border the shape of the wing. For example, the visualization pattern 310 may be in the form of lines, such as rounded lines, to reflect the contour of the electrode trace and the shape of the wing. In certain embodiments, there may be one, two, three, four, or more lines. In some embodiments, the visualization pattern may be formed from a pattern of dots, shapes, or other combinations, such as a pattern that maintains the visual cleanliness of the device even if the otherwise transparent adhesive layer becomes visually unacceptable to the user over time (e.g., if the adhesive layer picks up foreign matter and / or becomes cloudy due to moisture absorption). In certain embodiments, the visualization pattern may have another functional purpose, such as warning the user of the period of time the device is being worn by changing color over time or due to wear. Such a change in appearance can warn the user to remove the device at an appropriate time. Figure 3B shows a top view of an example of the physiological monitoring device 300, Figure 3C shows a bottom view, and Figure 3D1 shows a side view. In Figure 3C, the flexible electrode 312 is visible. As shown in Figure 3D1, the upper 314 and bottom 316 portions of the housing may be positioned above and below the flexible body 318. Figures 3E and 3F show the bottom and top of the physiological monitoring device 300, with each layer being transparent so that all layers are visible. Each layer is described in more detail below in the exploded view of the physiological monitoring device 300. An opening 320 may be located in a substrate layer positioned above the adhesive layer. As described in more detail above, such an opening may provide ventilation through one or more layers and may facilitate the evaporation of moisture from below the adhesive layer through the layers or multiple layers constituting the opening. As shown in Figure 3D2, in the embodiment, a gasket 319 may be positioned between a co-molded upper housing cover 314 and a lower housing 316 on one or more housings.The gasket may compress the adhesive assembly and raised interface (shown below Figure 3D2) or another gasket on the opposite housing to provide waterproofing to the internal electronic hardware. Figure 3. D As depicted in Figure 2, the ridge 321 may be positioned on the upper edge of the lower housing 316, and the ridge 321 is configured to be pressed against the adhesive layer 319. Those skilled in the art will understand that the ridge 321 may be any suitable shape, such as a fringed ridge as depicted in Figure 3D1. In some embodiments, the ridge may be rounded, square, and / or polygonal. In certain examples, the height of the ridge may be about 0.15 mm, such as about 0.01 mm to 0.5 mm, about 0.05 mm to 0.4 mm, about 0.1 mm to 0.3 mm, about 0.1 mm to 0.2 mm, or about 0.13 mm.

[0031] Figure 3G shows an exploded view of an example of the flexible body 301 of the physiological monitoring device 300 as described herein and elsewhere in the specification. The housing 306 is not shown. As will be understood by those skilled in the art, the image in Figure 3G is oriented upside down with respect to positioning on the skin. According to the numbering in Figure 3G, #7 depicts the release liner, which protects the adhesive layer (340 and hydrogel electrode 350). Immediately above the adhesive layer are a perforated layer 304 (including openings as described herein) and a flap layer 303. In certain embodiments, the perforated layer and flap layer may be composed of any suitable material such as polyethylene terephthalate (PET) and / or polyurethane. Immediately above the perforated layer may be a lower substrate layer #1, which may be composed of polyurethane. In embodiments, the lower substrate layer may have at least one textured side, the textured side of which is the flap layer 303 It may be positioned to face the. In the embodiment, the flap layer 303It may also include at least a textured side. This textured side may be configured to face the lower substrate layer #1. The conductive electrode trace may be printed on an additional separate substrate (311, 312). Alternatively, in some examples, the conductive electrode trace may be printed directly on substrate layer #1. Located above the conductive electrode trace may be an upper substrate layer 300. Located above the upper substrate layer may be an additional carrier layer #10, followed by an adhesive layer #11 and an uppermost rigid liner #9. Those skilled in the art will understand that such layer arrangements are applicable to any example of the physiological monitor described herein, such as the examples in Figures 3A–3F.

[0032] Figure 3H shows an exploded view of an example of the housing 306 of the physiological monitoring device 300, through which the flexible body 301 described in detail above passes. The upper housing cover 314 may include a patient trigger 307. The upper housing cover may enclose a circuit board 322. A spacer 323 located below the circuit board is configured to maintain a constant distance between a conductive contact spring located below the circuit board and the battery terminals / ECG trace contacts. The spacer can further provide electrical insulation between the circuit board and the battery. The spacer may have a hole for the conductive contact spring, which is connected to the circuit board. The battery terminals 325 may be located below the flexible body 301 and the circuit board 322, thereby being positioned above the wave spring 326. In some embodiments, the battery terminals 325 may be wrapped around and bonded to a coin-cell battery 328. The battery terminal 325 can be configured as a flexible circuit having conductive vias 327, thereby connecting the positive terminal bottom surface of the coin-type battery 328 to the negative terminal top surface of the battery, so that both the negative and positive terminals are presented on the top surface of the battery and can contact the contact springs of the circuit board. Alternatively, the battery contacts of the bottom housing can pull the positive side bottom surface of the coin-type battery up to the negative side top surface and bring it into contact with the circuit board. The ventilation layer 329 can be positioned above the ventilation holes 332 of the lower housing relative to the lower housing portion 316. In embodiments, the ventilation layer can be made of a material that allows the passage of gases while blocking the passage of liquids, such as ePTFE or any other suitable material. The combination of the ventilation holes 332 and the ventilation layer allows for the normalization of air pressure between the outside and inside of the housing. In the embodiment, the ventilation holes 332 combined with the ventilation layer prevent the buttons and / or triggers 307 from being blown out or sucked in in response to external air pressure, for example, when the patient is at a different altitude, such as on an airplane. The ventilation layer may be thin and round with a ring-shaped adhesive on its bottom surface. The area of ​​the ventilation layer covered with adhesive may be gas-impermeable, while the central portion may be gas-permeable and liquid-impermeable.The central portion of the ventilation layer can be positioned above the ventilation holes, thereby allowing gas to enter and exit the housing while restricting the entry and exit of liquid. In some embodiments, the ventilation layer can be integrated with the bottom housing by molding, or it can be ultrasonically welded to the bottom housing or bonded by any suitable means.

[0033] A system that uses neural network encoding to estimate load and / or predict rhythm annotation. Some embodiments disclose a wearable device capable of processing detected biosignals via an encoder that includes a first subset of layers of a neural network, such as the feature described in U.S. Patent No. 11,246,524, which is incorporated in whole by reference. Figure 4A is a schematic diagram of an embodiment of a system that predicts rhythm annotations using neural network encoding. In some embodiments, the system can make one or more predictions, such as predicting rhythm annotations of cardiac rhythms including atrial fibrillation and / or atrial flutter, or estimating load. A wearable device, such as a cardiac monitor patch 402, can process an ECG input 404 via a first subset of layers 406 of a neural network, such as an encoder. The wearable device receives the output of a first subset 406 of the neural network layers and transmits that output to a computing device 408 (for example, an external system such as a smartphone or server) which can further process the data through a decoder containing a second subset 410 of the same neural network layers to derive user 412 characteristics, such as indications or predictions of past cardiac arrhythmias and / or predictions of future arrhythmia occurrences.

[0034] In some embodiments, a first subset and a second subset of layers reside within a single neural network. The neural network may be designed so that the output of the first subset of layers has a lower dimensionality than the input to the neural network, and the output of the second subset of layers may be designed to provide indicators of user characteristics, such as past or future predictions of atrial fibrillation. Figure 4B is a schematic diagram of an embodiment of the first and second subsets of layers within a single neural network. The neural network may be trained simultaneously on both the first subset 422 of layers and the second subset 424 of layers. For example, if the neural network has 10 hidden layers, the first four layers are processed on a wearable device such as a cardiac monitor patch 426, and the output of the fourth layer, which has a lower dimensionality than the input to the first layer (e.g., has good data compression capabilities), is sent to an external computing system such as a server 428. The external computing system processes the outputs of layers 4 through 5-10. The dimensionality of each layer of a neural network can be designed to output a certain data size (e.g., the output dimension of each convolutional or pooling layer). For example, patch 426 may include an ECG encoder 430 that receives ECG data 432 at 2400 bits per second (bps). Patch 426 can process the ECG data 432 through a first subset of layers 422 of the neural network (such as within the ECG encoder 430) and output data 434 with a smaller dimensionality, such as data at 128 bps. The output data 434 can be sent to an external server 428. The external server 428 can process the output data 434 through a second subset of layers 424 of the neural network and may include a classifier 438 that outputs patient signs or predictions 436 that it has been trained on. The entire neural network, including the first subset 422 and the second subset 424 of layers, can be designed and trained as a single neural network.

[0035] In some embodiments, the output of a first subset of layers can have a smaller number of dimensions than the input to the neural network. Therefore, instead of sending the entire ECG signal (e.g., the input to the neural network) from the wearable device to an external computing device, the wearable device can send a smaller amount of data to the external computing device, such as the 128 bps output data 434 of the first subset of layers instead of the entire 2400 bps ECG signal 432. Advantageously, this reduces the network throughput required to derive signs of past cardiac arrhythmias and / or predict future arrhythmia occurrences.

[0036] Furthermore, instead of processing the ECG signal through all layers of the neural network in the wearable device to derive signs of past cardiac arrhythmias and / or predict future onset of arrhythmias, the wearable device can process the ECG signal through only a first subset of the neural network layers (e.g., through the ECG encoder 430) and send the output of the first subset to an external device 428 that processes a second subset of the layers (e.g., through the decoder or classifier 438).

[0037] Embodiments of prognosis and prediction Several embodiments described herein are improvements on conventional systems for detecting atrial fibrillation using long-term continuous monitoring techniques. Some conventional monitors typically involve implantable devices with lead wires embedded in the heart. However, such approaches are highly invasive, lack adequate communication technology, and are extremely difficult to update hardware / software.

[0038] Furthermore, conventional monitors fail to provide treatment recommendations such as interventions or hospitalization recommendations, nor do they offer predictive outcomes based on current cardiac signals or treatment methods.

[0039] Furthermore, conventional implantable technologies cannot continuously capture ECGs. These implantable monitors measure ECG data and capture events detected within the ECG data. Because these implantable monitors capture event data based on small data windows, such as 10 seconds of ECG data, they can provide predictive results regarding current cardiac signals. patient Results of current treatment provided to prediction of offer It may be insufficient for determining treatment recommendations or making decisions regarding treatment recommendations.

[0040] Figure 5 is a flowchart 500 of embodiments of the system and monitors such as an ambulatory monitor. The monitor can be an implantable, wearable, and / or wearable recording device. In block 502, the monitor can capture continuous and / or discrete biosignal data. Some embodiments described herein include monitors that can continuously monitor cardiac signals over long periods of time, such as 7, 14, 21, 30 days, or longer, using electrodes and monitoring circuits. Some embodiments can provide outcome predictions based on the patient's currently measured cardiac signals over periods longer than seconds, minutes, or hours. Some embodiments can provide predictive results regarding the treatment currently being provided to the patient. Some embodiments can provide treatment recommendations for the patient based on the currently measured cardiac signals and / or current treatment.

[0041] In some embodiments, the monitor may measure one or more physiological signals, such as atrial fibrillation, ECG, PPG, accelerometer data, heart rate, pulse transit time, impedance measurement, acoustic measurement, RR interval, blood pressure measurement, temperature measurement, and / or glucose level.

[0042] In some embodiments, the system can continuously measure, record, and / or analyze biosignals via a device implanted, attached, or fitted to a person in an external environment. The signals may include raw measurements or features extracted based on signal processing, machine learning (ML), or deep learning (DL) methods.

[0043] In block 503, the monitor can analyze the received continuous or discrete biosignal data to remove noisy segments and unwanted artifacts. The monitor can apply signal processing machine learning algorithms to classify noisy segments and segments below a noise threshold level, for example, by using a pre-trained model. The monitor can apply signal processing machine learning algorithms to determine whether the received signal is noise, an artifact (such as whether the user is sleeping, engaged in high-intensity activity, or has a pacemaker), and / or similar. This technique can be used to remove or reduce noise and improve the signal-to-noise ratio. Noise can be generated by internal or external sources, such as environmental noise, noise resulting from improper contact with electrodes, and / or similar.

[0044] In some embodiments, signal processing techniques for removing noise and / or unwanted artifacts may include machine learning models such as neural networks. These machine learning models may differ from neural networks used for inference on physiological data (as further discussed herein).

[0045] In some embodiments, a noise machine learning algorithm can be trained to determine whether or not artifacts are present in the data. The noise machine learning algorithm can also reduce noise in the recorded data. In some embodiments, the same trained algorithm can perform both or either artifact identification and noise reduction. In other embodiments, separate machine learning algorithms are used. In some embodiments, noise and / or artifact prediction can be performed on a sample-by-sample basis or on a sample group.

[0046] In some embodiments, artifact segment identification, noise segment identification, and / or noise reduction can be performed using one or more of the following architectures: a) Fully connected neural networks (NNs): For example, a fully connected NN can learn the complex relationship between signal morphology and the actual label of whether the signal is an artifact or noise. By training with a diverse set of signals representing artifacts / noise or useful signals, and combining this with dropout, the network's ability to learn important relationships between signals and labels (useful signals, noise, artifacts, etc.) can be enhanced while reducing the possibility of overfitting. For example, the system may randomly disable layers or specific nodes of the neural network during training to avoid overfitting and / or to train the neural network more restrictively to learn more robust and predictive weights for the training purpose. b) Encoder / decoder architecture and / or fully connected NN following a convolutional layer: The use of a convolutional layer allows for the use of longer signal sequences as inputs, providing better context (e.g., useful ECG signal-to-noise transitions) for identifying onsets and / or offsets of noise / artifact regions. The encoder / decoder section can advantageously identify latent variables that can effectively learn transitions to and from noise / artifact regions, as well as other signal / noise characteristics, using a smaller set of features. These latent variables can be applied to the fully connected layer to generate a complex prediction function. Due to the small network size, noise prediction can be performed on the device itself, or the encoder output can be transmitted efficiently and continuously in real time via a wireless connection, with predictions performed on a server. c) A convolutional layer may be followed by an LSTM layer (e.g., a bidirectional LSTM layer in some embodiments), and then many fully connected layers: Adding LSTM layers enhances the predictive power of the network by enabling it to recognize parts of the signal that are highly correlated with other parts of the signal. For example, in the transition between a noiseless signal and a noisy signal, the LSTM can identify more correlations between the waveforms before and after the transition point, and the sequence of the system's output weights can be adjusted to better identify the transition point between the noisy segment and the noiseless segment. d) Transformer architectures that can provide accurate classifications that take into account the context in which noise / artifacts may be present by utilizing the similarity / dependence of characteristics of continuous signals: Transformer networks can advantageously offer a more flexible structure when encoding the similarity / difference between signal sections. For example, a transformer architecture can adjust its processing based on user selections such as the temporal distance at which similarity is observed, or the sequence of signal frequencies to be considered. For instance, a transformer architecture can adjust the length of consecutive segments at specific intervals, such as looking at 1-minute intervals, 1-second intervals, 1-hour intervals, 6-hour intervals, and / or similar. e) Self-supervised learning to extract latent features from a large number of unlabeled signals, including noisy and non-noisy regions with diverse types of rhythms, and train a model to classify signals in the noisy region: The pre-trained features can be combined as input to smaller fully connected layers to generate final predictions of noise or artifacts. Since these features are rich but the final classification network needs to be smaller, implementation can be performed efficiently on both the device itself and / or the server. For example, a first pre-trained model can extract latent features, an additional layer of a second pre-trained model can receive the latent features as input and apply it after the first pre-trained features to classify noise or artifacts, and / or an additional layer can take pre-trained features from a self-supervised model and apply the features to make predictions. f) Siamese network: In this approach, the system can train a network to identify specific waveforms selected by the user as noise / artifact segments. This allows the system to perform classifications that are better suited to the application at hand. Furthermore, if new versions of a signal are to be considered as segments containing artifacts and / or noise, these segments can be added to the process by adding a newly trained network for the new waveforms of interest. For example, if there is a model already trained to identify noise, the model can be trained to also identify artifacts without interfering with the part of the model already trained to identify noise.

[0047] In block 504, the monitor can perform analysis on the captured biosignals to determine whether or not an arrhythmia is present and / or predict the likelihood of future arrhythmia development. Some embodiments described herein include a monitor that can determine that a patient does not currently have atrial fibrillation. The monitor can predict the future onset of a potential atrial fibrillation state. Conventional monitoring devices have been limited to providing information on the current atrial fibrillation state.

[0048] In block 506, the monitor determines whether an arrhythmia is present in the cardiac signal and / or whether the onset of an arrhythmia is detected. If the determination is yes, the flowchart proceeds to block 508, where a communication of the presence or onset of the arrhythmia is sent to a physician or a user on a device such as a mobile device, and / or the monitor can modify its functions, such as analyzing data over a longer period or collecting more data over a longer period, after which the flowchart can proceed to block 509. In some embodiments, the monitor can update its functions based on the arrhythmia determined in block 509.

[0049] Advantageously, alerts can be sent from the patient to a physician located remotely, and these alerts can be sent in real time or virtually real time when the arrhythmia or onset is detected. The physician's computing device and / or applications on the computing device can be activated by the system to display the alerts. Thus, even if the physician's computing device does not have the application for the alert open, the system described herein can alert the physician by sending an alert to activate the application. In some embodiments, if no arrhythmia or onset is detected, the flow chart proceeds to block 510.

[0050] Some embodiments described herein include outpatient monitors that can monitor cardiac signals and determine their association with potential future strokes, deaths, atrial fibrillation, heart failures, and the need for hospitalization. Such outpatient monitors can make these predictions (such as the onset of atrial fibrillation or a potential stroke or hospitalization) based on cardiac signals in the patient's normal state, such as while the patient is going about their daily life without hospitalization. For example, in predicting stroke, the outpatient monitor can make such predictions by applying cardiac monitoring data over several weeks, such as two weeks. Predicting the onset of atrial fibrillation can be done with the same and / or less amount of data as predicting the onset of stroke. The amount of data required to predict the onset of atrial fibrillation may depend on the patient's age, as the prevalence of atrial fibrillation may increase more rapidly in patients aged 70 years and older.

[0051] Some embodiments described herein include outpatient monitors that may include an encoder on a patch. For example, in the case of outpatient monitors that process less data (e.g., less than a week), an encoder may be included on the patch, and the encoder may process the data to output data with a smaller number of dimensions for transmission to an external computing system such as a server. The server may process the output of the encoder to determine physiological characteristics such as whether atrial fibrillation can be detected from the cardiac monitoring data or whether the onset of atrial fibrillation can be predicted. Advantageously, the signal may be partially processed on the patch and partially processed by a server, such as a cloud server, during the time the patch is worn. The cloud server may make predictions about the onset of atrial fibrillation and / or provide treatment recommendations, etc., while the user is placing the patch on his or her body.

[0052] In some embodiments, the cloud server can send alerts to computing devices such as the user's mobile phone and / or a physician's computing device to notify of the onset of a condition. Because the encoder has a small number of dimensions, it can send less data to the server side and therefore can offer technical advantages such as automatic encryption, extended battery life, reduced processing power required for patching while the server further processes the data, and reduced network data throughput used, as further described herein. Furthermore, such immediate processing and reduction in the use of critical hospital resources are critical, such as in the case of stroke onset, where response time largely correlates with survival rate, especially if the patient is already hospitalized. Outpatient monitoring can provide predictions of physiological characteristics such as stroke and / or provide treatment recommendations, such as sending the patient to an immediate treatment unit (ICU). The encoder can be trained to output signal features derived from monitored cardiac signals. For example, the encoder can output cardiac signal features such as boundaries, edges, and / or analogues.

[0053] In block 510, the monitor can upload recorded continuous and / or discrete biosignals and / or associated derived signal features output from the encoder. The encoder output and / or a portion of the measured biosignals can be continuously uploaded to an external computer, such as a cloud. The encoder output can be uploaded by connecting a patch to the device, which then sends it to an external computer, for example, via a USB connector to a mobile phone, which then sends the data to a server. Advantageously, if only the encoder output without the measured biosignals, or only a portion of the measured biosignals, is sent to the cloud, the monitor can conserve battery life by storing less data, transmitting less data, requiring less network connectivity, requiring less data transmission, and / or requiring less data transmission frequency.

[0054] In block 512, an external computing device such as a server can analyze recorded and / or associated derived features to identify a set of cardiac features. For example, features extracted from machine learning and / or deep learning techniques may include one or more of the following: atrial fibrillation load, total atrial fibrillation time, longest atrial fibrillation episode, heart rate quantiles, rhythm counts such as supraventricular tachycardia (SVT), ventricular tachycardia (VT), atrial fibrillation, ventricular premature contractions (PVCs), atrial premature contractions (PACs), and atrioventricular block (AVB), event duration and density (day, night, week), other arrhythmia load, number and duration of episodes at low, moderate, and high activity levels, transition pattern features / feature clustering, transition time features, rhythm categories, arrhythmia density and density changes, ectopic density and density changes, activity load, rest load, and / or sleep arrhythmia / load.

[0055] In some embodiments, if the onset of atrial fibrillation is detected, the monitor may adjust its settings, and / or an external computing system may instruct the monitor to adjust its settings, such as observing a longer monitoring window. In some embodiments, if the monitor is measuring a longer monitoring window, the monitor may modify the data sent to the external computing system. For example, by lengthening the monitoring window, the monitor may send only the encoder output and not the measured ECG signal, or the monitor may shorten or lengthen the window for the measured ECG signal and send the encoder output.

[0056] In block 514, a system such as an external computing device can calculate the associated risk for each predictive endpoint based on a set of biosignal recording cardiac features. Some embodiments include a system that can determine a risk score based on measured cardiac signals indicating the onset of atrial fibrillation. The system can determine the risk score based on a specific intervention and a specific outcome, such as by clustering previous patient data, as further disclosed herein. For example, a clustering algorithm can cluster patients based on their current state, the type of intervention, and the outcome of that type of intervention, and then identify the current patient as similar to one of the clusters.

[0057] In some embodiments, based on identified related clusters, the system can recommend interventions for physiological conditions such as the onset of atrial fibrillation. The clustering algorithm can be based on long-term clinical trials in which patients are monitored for their current and / or onset conditions, treated with various types of interventions, or not treated at all, and the outcomes of these interventions are recorded. Such data is processed by the clustering algorithm to create patient clusters based on current / onset condition, intervention type, and outcomes.

[0058] In block 516, the system can determine whether a particular intervention is associated with a predictive endpoint. If yes, the flowchart can proceed to block 518, where the system can calculate the predictive impact for each predictive endpoint associated with the relevant intervention, based on a set of biosignal recording cardiac features and a predictive endpoint risk assessment. In block 520, the system can incorporate the risks into a report and display the report on the patient and / or physician's device, along with the predictive endpoint and, where applicable, the predictive impact associated with the intervention option. For example, an intervention may be associated with a long-term change in the course of the disease. An intervention may be associated with a short-term treatment, such as when the patient is already hospitalized.

[0059] Therefore, the implementation of this system and / or outpatient monitoring offers significant advantages to cardiac monitoring technology. This system and / or outpatient monitoring provides prognosis, risk stratification, and treatment guidance for the numerous consequences and associated costs resulting from, or related to, the presence of symptomatic and asymptomatic cardiac arrhythmias, such as atrial fibrillation (AF). As a result, the risk stratification system performs better than conventional risk factors. While conventional system diagnostics are limited to providing information on the need for potential treatment or continuous monitoring at the population level, the system and / or outpatient monitoring described herein provides personalized guidance for continuous monitoring (timing and intensity) and treatment initiation and dosage setting (e.g., depending on the current level of arrhythmia severity and the predicted future severity of consequences if left untreated). In particular, systems and / or outpatient monitors can use partial or complete recorded signatures created from wearable devices, such as signal sequences of multiple modalities and / or different time horizon settings, to predict the risk of adverse outcomes, the need for intervention, estimated intensity, treatment costs, hospitalization, and / or similar events.

[0060] In some embodiments, the system and / or outpatient monitor may use latent multimodal features (encoded markers) that encapsulate an individual's temporal and physiological characteristics, which can be used as predictors of multiple future health-related endpoints. The system and / or outpatient monitor may inform the user or physician about specific combinations of explainable features that can be used to access the role of asymptomatic and symptomatic arrhythmias in the risk of unfortunate outcomes. The system and / or outpatient monitor may assess the probability of the presence of undiagnosed (and often asymptomatic) atrial fibrillation, often referred to as silent atrial fibrillation, even if the diagnostic device does not detect atrial fibrillation during the wear time. Onset It is possible to predict the occurrence of atrial fibrillation. The detection of atrial fibrillation is a function of the incidence of atrial fibrillation and the probability that the occurrence of an event coincides with the time the device was worn. By identifying markers that indicate active but undiagnosed atrial fibrillation based on a single-lead ECG or PPG sensor worn during routine activities rather than during clinical practice, it is possible to guide follow-up examination strategies for patients with high confidence scores based on predictive models, leading to earlier diagnosis of the disease. This system and / or outpatient monitor can provide a timeframe for follow-up examinations of asymptomatic atrial fibrillation and indicate the recommended timing for these examinations.

[0061] While this prediction method is described for atrial fibrillation, the same method can be applied to other arrhythmias, such as tachycardia (including supraventricular and ventricular tachycardia) and bradycardia events (including pauses and atrioventricular block). Prediction of ectopic beats (e.g., premature atrial and ventricular beats) may also be developed. In fact, rather than specializing in a single arrhythmia type, predicting future combinations or categories of arrhythmias may offer advantages in predictive accuracy.

[0062] In some embodiments, using a single-lead continuous monitoring device offers the advantage of detecting and / or predicting the presence of asymptomatic atrial fibrillation from signals associated with the patient's normal activity, an advantage less likely to be provided by a 12-lead ECG in a clinic, as it only measures for a short time (typically 10 seconds) under controlled conditions. Conventional systems cannot predict the future onset of atrial fibrillation. For example, the system described here may be effective as a multi-device solution, for instance, to notify the patient whether they should wear another outpatient monitoring patch, or to engage in an extended monitoring paradigm (up to 2 years) with a watch / ring or implantable monitor.

[0063] In some embodiments, the system and / or outpatient monitor can evaluate cluster patient records within specific disease and / or risk groups, assess clinical trajectory / journey based on a number of signatures obtained over time, and evaluate the probability of response to a specific or intervention family. The advantage of this is that it differs from predicting response to intervention, as it informs the group of combinations of disease trajectories and the history of using specific intervention families that improve outcomes. This further personalizes healthcare by linking disease characteristics with responses to specific interventions. The system and / or outpatient monitor can predict the potential course of disease progression in patients and the effects of interventions. The system and / or outpatient monitor can use risk scores to select individuals to include in clinical trials for a given intervention or treatment. The system and / or outpatient monitor can determine the detectability of future events during the monitoring period, which can be used by clinicians to access the need to continue / extend / terminate monitoring.

[0064] In some embodiments, after the external computing device determines AF, the onset of AF, and / or intervention, the external computing system can send data back to the monitor. Advantageously, the transmitted data can initiate changes to the monitor, such as how much data is being measured, the algorithm being applied on the monitor, the amount of data being processed on the monitor and / or the amount of data being sent to the external computing device (e.g., window length), and / or similar. In some embodiments, the external computing device notifies a physician, such as a cardiologist or attending clinician, that an atrial fibrillation state has been detected, that there is a possibility of onset atrial fibrillation, and / or recommendations regarding intervention (e.g., provision of ablation or medication based on the clustering algorithm described herein).

[0065] In some embodiments, the system can connect to two wearable devices. The patient wears one device and can upload data to be analyzed in a cloud environment. The system can connect to a second device. In some embodiments, the system connects to the second device simultaneously with the first device. The system can acquire similar data to correlate the two and / or receive different data, such as different types of data, to further enhance processing on the system side. In some embodiments, the output of an autoencoder architecture recommends that the patient wear the second device, including what type of device and / or when to wear the second device.

[0066] In some embodiments, the system can connect to a second device at a later point in time and correlate the recordings with previous recordings of the same patient. Thus, the system can create longitudinal datasets of a single modality (such as ECG) and / or a set of simultaneously measured modalities (such as heart rate, PPG, blood pressure, acceleration, and acoustics). Combining various modalities can effectively create more progressive scores. Advantageously, different modalities can provide a much better picture and prediction of atrial fibrillation onset and other physiological characteristics, as well as the progression of interventional treatment. For example, distinguishing signals under stress from signals while the patient is sleeping can provide a more accurate picture of atrial fibrillation onset. Data collection by wearable devices can include continuous signals from long-term continuous monitoring devices such as cardiac monitor patches, smartwatches, smart fabrics, smart eyewear, and / or similar devices. Wearable devices can post uploads to the system, and the system can perform analysis of the recorded signals in a mobile computing or cloud environment. The algorithms, data, and results of such analyses can be stored in the cloud for long-term preservation.

[0067] In some embodiments, data received from different devices can yield separate risk scores that indicate the progression of physiological characteristics. For example, based on measurements from a smartwatch-type wearable device (e.g., score 20 at time 1, score 50 at time 2), the system can indicate that atrial fibrillation is progressing rapidly and recommend patch application.

[0068] Risk prediction implementation In some embodiments, a risk prediction model may require variable / feature inputs and an algorithm capable of appropriately weighting these inputs to generate a risk score for a given outcome over a specific time frame. The system can select inputs from any number of features, as described herein. For machine learning, statistical modeling, and deep learning models, discrete features from the aforementioned features can be combined to generate risk estimates using the methods described herein. For continuously measured signals or signals with a time dimension (e.g., rhythm labels derived from an ECG through an annotation algorithm based on waveforms within a given time interval), the system can use a number of numerical methods to generate an appropriate form of input for calculating a risk score through a trained algorithm. The system can use these methods for individual signal measurements or combine them to convey information simultaneously.

[0069] In some embodiments, these signal combinations can be temporally aligned and sampled (upsampled, downsampled, or maintained) at the same sampling rate to generate a simultaneous set of measurements. Each type of one-dimensional signal may be scaled and combined to form a data strip. The X-axis of the strip is the time axis, and the Y-axis represents the stacking of one-dimensional signal modalities. Advantageously, structuring the data in this way has the advantage of simultaneously creating features based on different signal modalities. This data structure can be further extended by utilizing signal transformations and interactions of the individual signals. Such transformations include Fourier transforms and wavelet transforms of the original waveforms. Interactions include simple multiplication of two or more signals, powers of these signals, or time-lagged convolutions of the signals.

[0070] In some embodiments, the system can form the data structure as an image and apply deep learning image analysis techniques to generate a reduced representation of the signal. Specifically, the system can use an autoencoder model that utilizes two-dimensional convolution of the signal to create a complete latent representation of the signal. Such latent features can be of various dimensions and can encapsulate relationships between different signal information in a low-dimensionality. In some embodiments, these latent variables are extracted from the bottleneck layer of the autoencoder.

[0071] In some embodiments, the system may apply an autoencoder architecture representing a dimensionality reduction technique that, when combined with an appropriate weight and layer architecture, can extract a set of features that can reconstruct the original signal with high fidelity. This dimensionality reduction allows the autoencoder to create a “recording signature” that can uniquely represent a recording, such as a partial or complete recording of a signal, or unique features of the signal measured simultaneously. Representations corresponding to “similar” recordings may have similar “recording signatures” at an appropriate distance metric. In other embodiments, dimensionality reduction techniques such as principal component analysis or random projection can be used instead of an autoencoder. The system can perform feature selection and / or extraction in a supervised or unsupervised manner.

[0072] Supervised Feature Extraction and Model Development Implementation Figure 6 schematically shows an embodiment of the autoencoder architecture 600. The autoencoder architecture may include a noise machine learning model 603 that removes unwanted artifacts and reduces noise (as further described herein), an encoder 604, and inputs to the encoder such as rhythm data 602A, ECG data 602B, PPG 602C, Q95 heart rate 602D, and acceleration 602E. The autoencoder architecture may also include the output of the encoder 606, a decoder 608 that receives the output of the encoder 606, and output features that represent the same inputs to the encoder such as rhythm data 602A, ECG data 602B, PPG 602C, Q95 heart rate 602D, and acceleration 602E.

[0073] In some embodiments, encoder 604 can be trained together with decoder 608 so that decoder 608 outputs an approximation of the same data (often with reduced noise levels) as the data input to encoder 604. Encoder 604 and decoder 608 can be trained with multiple modalities so that encoder 604 can be trained to analyze each of the data input to encoder 606 in order to extract features from the output of encoder 606. Inputs to encoder 604, such as rhythm data 602A, ECG data 602B, PPG 602C, Q95 heart rate 602D, and acceleration 602E, can be temporally aligned. Since the encoder decoder (604, 606, 608) is trained, encoder 604 can, based on the training, automatically determine that for a particular patient dataset, the end point of the ECG aligns with a low point in the PPG where the heart rate is at the 5th percentile (above the lowest 5%) and the acceleration for that frame is high, corresponding to the start of AF. Advantageously, encoders can be trained to evaluate multiple modalities simultaneously to feature and / or predict the onset of atrial fibrillation.

[0074] In some embodiments, the desired result can guide the feature extraction process. The autoencoder architecture can be trained to produce a compressed representation of a large number of samples. The encoder 604 can be connected to another deep learning network architecture trained for classification or risk prediction using an appropriate loss function. The encoder is implemented on a device and can send the encoded data to a server, where the rest of the deep learning network uses the encoded features to predict the relevant outcome (classification probability, survival probability, outcome label, e.g., AF).

[0075] In some embodiments, the encoder can be pre-trained and connected to other deep learning network architectures that are trained together with the decoder to generate approximations of the original data, and then simultaneously trained to predict a number of desired outcomes. Each architecture may predict a single outcome based on the same output 606 of the encoder, or it may use a multi-task architecture to predict multiple outcomes simultaneously (e.g., stroke, heart failure, hospitalization risk). During this training, the encoder weights may remain frozen or may be allowed to be continuously adjusted. The advantage of using a pre-trained encoder with frozen weights is that the output can be the features of layer 606. This single output can then be sent and used from multiple deep learning networks trained with the encoder to predict one outcome per network. Such a solution reduces the amount of data transmitted because one output can be used from multiple deep learning networks for the prediction of multiple desired outcomes or a single multi-task deep learning network. Furthermore, the encoder's bottleneck layer, or the dimensionality of the output of encoder 606, can be adjusted to satisfy conflicting operational requirements and performance. For example, the number of dimensions can be reduced to reduce data transmission, less processing of input data by the server, to require less storage, and / or similar reasons. The number of dimensions of the encoder can be selected based on the desired algorithm on the server side, such as having multiple algorithms for different scenarios and / or desired results.

[0076] In some embodiments, the number of dimensions of the encoder can be determined based on the desired accuracy of the output. For example, a smaller number of dimensions results in lower accuracy but less network bandwidth usage. In some embodiments, the number of dimensions can be frozen to predict a specific set of results. However, if different sets of results are desired, the system can adjust the number of dimensions to be suitable for predicting different sets of results, such as predicting hospitalizations and predicting strokes.

[0077] In some embodiments, the training labels include one or more of the following: disease / disease classification for different outcomes (e.g., stroke, MI), a set of time-to-event values ​​with failure (death / morbidity) or censoring indices, a set of economic outcomes related to medical costs, a set of length of hospital stay, a set of time to the next device prescription, and / or a set of time to the onset of arrhythmia (e.g., atrial fibrillation) with failure or censoring indices.

[0078] In some embodiments, the encoder 604 and decoder 608 are trained together. After training, the encoder 604 is applied to patient data, and the output of the encoder 606 is processed through a machine learning (ML) time-to-event model 612 that predicts the probability of survival / failure (occurrence or non-occurrence of an event) within a given time frame. The ML model is trained to predict probabilities using the output 606 and a combination of failure time 614A (time until an event occurs) and censor 614B (whether a failure was observed at that point).

[0079] Two options for developing a risk-based model are highlighted in Figure 5. The autoencoder network can reduce its dimensional representation (bottleneck layer) as the output of encoder 606 through a process of receiving a 2D tensor of time-aligned measured or transformed values ​​(e.g., heart rate extracted from an ECG or PPG signal) at encoder 604 and reconstructing the signal. A 2D tensor has time as its first dimension and any combination of the aforementioned quantities as its second dimension. By using multiple of these inputs from the same or multiple devices, the autoencoder architecture can record and reconstruct a low-dimensional representation of a partial or complete device signal tensor. In the first version, the bottleneck layer can be extracted and introduced as its input to a machine learning model 612 (if the bottleneck layer is 2D or greater, it can be flattened into a vector). Standard parametric and nonparametric time-to-event statistical models (e.g., Cox proportional hazards models, parametric models (e.g., Weibull, log-normal, etc.), or survival tree models) can be used to appropriately weight and integrate the most informative parameters of the bottleneck layer to fit the survival model.

[0080] In some embodiments, the transformer architecture can be used to incorporate demographic data (such as age, race, and place of residence), clinical data (current measurements), and historical health data and features (either explainable or latent) to predict a sequence of outcomes, interventions, and associated costs that can be used to predict the course of a patient's life. These architectures can also incorporate longitudinal recordings to identify the recordings or segments of recordings that best predict the desired outcome. This architecture can be used with longitudinal monitoring to predict a set of events of interest (such as the occurrence of future atrial fibrillation episodes, strokes, hospitalizations, and treatments). The context may include events within a recording, events between longitudinally obtained recordings, and / or future events related to the hand-held measurements (such as stroke or death). As an example, a cardiac monitor patch may contain features that point to atrial fibrillation that has not yet been detected by the patient.

[0081] In some embodiments, subsequent application of monitoring patches may diagnose atrial fibrillation and corresponding stress based on information from the initial patch. A series of patches may be used to estimate the probability of adverse events such as stroke. Thus, a converter-based model trained with appropriate data can be used to automatically select the most relevant records and predict the next event of interest (e.g., patient unaware of atrial fibrillation → asymptomatic atrial fibrillation → symptomatic atrial fibrillation → stroke) or to incorporate treatment decisions that reduce the probability of such events. Furthermore, the system can also include predictions of hospitalization procedures and corresponding costs associated with treatment and hospitalization.

[0082] Advantageously, encoder architectures can apply encoders to generate features, and these features can then be processed through other algorithms, such as neural networks, to determine desired outcomes, such as the conditions for atrial fibrillation. Traditional algorithms have limited ability to incorporate the temporal relationships of data within long time frames. However, architectures like those described here (LSTM, transformers, etc.) can naturally incorporate these relationships to make such predictions.

[0083] In some embodiments, a machine learning model for cleaning noise (referred to herein as the noise machine learning model) can be trained together with a machine learning model for inference (referred to herein as the inference machine learning model). Training data can be input to the noise machine learning model, and the output of the noise machine learning model (or its derived signal) can be input to the inference machine learning model. It is understood that additional computational or signal processing techniques can be performed before and / or after either of the trained models.

[0084] In other embodiments, the machine learning model may be the same model as the neural network performing the inference. A single machine learning model can be trained to both clean up the data from unwanted artifacts and / or noise, and to perform inferences on the data.

[0085] In other embodiments, training of machine learning models can be performed remotely from the monitor. However, one or more machine learning models can be applied on the monitor side. In some embodiments, one or more machine learning models (or all of them) can be applied on the server side. For example, for short-term forecasting of data, the monitor can run one or more machine learning models. Advantageously, alerts can be immediately displayed to the user from either the monitor or another computing device such as a mobile phone. Furthermore, if inference has been made, the device can trigger further processes, such as applying long-term forecasting models or different algorithms that perform different inferences (such as those that examine different characteristics of the user). Such long-term forecasting models can be run on the server side.

[0086] In some embodiments, data cleaning can be performed at specific time intervals, such as every second, every minute, every hour, and / or other times. For example, artifacts can be discovered and / or noise reduced based on a 30-minute window. In other embodiments, a sliding window of a certain time period can be used, and data cleaning can be performed based on the amount the sliding window slides. For example, a 30-minute data window can be used to make predictions every 5 minutes.

[0087] In some embodiments, one or more machine learning models (such as all of the machine learning models) can be run on the server side for long-term forecasts that require data over longer periods. For example, the monitor can record data for days, weeks, months, and / or such long periods. The data on the monitor can be downloaded to the server side, and a remote server can perform long-term forecasts on the data.

[0088] In some embodiments, a subset of a pre-trained model and / or a portion of the pre-trained model can be run on the monitor side, while the rest of the architecture is run on the server side. For example, a noise machine learning model may run on the monitor, and an inference machine learning model may run on a remote server. In some embodiments, a portion of an inference machine learning model (e.g., a first subset of layers) may be run on the monitor, while the rest of the model (e.g., a second subset of layers) is run on the server side.

[0089] Advantageously, not all data needs to be sent from the monitor to the server, reducing the amount of data transmitted. This results in lower network bandwidth requirements. Furthermore, since the entire dataset doesn't need to be stored before transmission, the monitor requires less memory. The data is processed through one or more machine learning models, which can output specific data such as vectors. These vectors are then stored and sent to a remote server.

[0090] Figure 7A shows an example network architecture 700 for applying one or more LSTM converter blocks at the decoder output. The network architecture 700 may include input features to encoder 604, encoder 604, encoder output 606, decoder 608, and RNN and / or LSTM blocks 702, 704, 706, 708. This architecture will be trained using data including failure time 710A and censored output 710B to generate probabilities of outcomes of interest within a pre-specified time frame. The autoencoder may be pre-trained separately or as part of the full model training. It is included as a feature extraction / compression and noise reduction tool, particularly for long (time-direction) recordings of coupled inputs, where an applicable method for image recognition is desired; otherwise, it is removed and the input is connected to the remaining layers.

[0091] Figure 7B shows an embodiment of a deep learning architecture 750 for applying a series of LSTM blocks and / or converter blocks to the output of an encoder. In some embodiments, the autoencoder can consist of a combination of LSTM layers, RNNs, and / or converter blocks that capture unidirectional or bidirectional relationships. Advantageously, these architectures can incorporate signal features, their interaction with other signal features in time, and physical distances that indicate a person's health status and the risk of developing health problems in different timeframes. Thus, the encoder architecture can naturally encode information about correlated features and / or timing between features (groups of patterns and timing characteristics), which can then be processed by other layers incorporating temporal relationships, making it more powerful in predicting time to events.

[0092] In Figure 7B, inputs 602A, 602B, 602C, 602D, 602E, encoder 604, and encoder output 606 (e.g., bottleneck) may be attached to a number of other layers (possibly comprising iterative blocks of such layers incorporated into the model architecture), followed by one or more dense layers of the model, where demographic or clinical variables may also be incorporated as inputs, and output layers are connected to represent right-censored data for any of AF, stroke, or other clinical outcomes of interest.

[0093] In some embodiments, the autoencoder architecture 750 can be trained front to end. The autoencoder architecture 750 can be trained using a patient's historical cardiac data (such as PPG, ECG, heart rate, and / or analogues) and the same patient's resulting state. For example, the training data could include patient data from the past several years, with the current patient state and / or subsequent patient states and / or subsequent states, from recordings of cardiac data. The training data can be used to train the system by adjusting the weights in the encoder and transducer blocks. Pre-trained weights are adjusted in the encoder. The transducer following the encoder can be trained with random weights. During training, the encoder weights are expected to change more slowly than the transducer weights and the rest of the network, including other layers. Throughout training, the encoder weights are adjusted to highlight specific features suitable for detecting atrial fibrillation, the onset of atrial fibrillation, and / or other arrhythmia elements.

[0094] Figure 7C shows a self-attention architecture for inferring the onset of cardiac arrhythmias. The self-attention architecture 770 may include multiple input nodes 772A, 772B, 772C, 772D (collectively referred to as input nodes 772), encoders 774A, 774B, 774C, 774D (collectively referred to as encoders 774), latent features 776A, 776B, 776C, 776D (collectively referred to as latent features 776), self-attention modules 778A, 778B, 778C, 778D (collectively referred to as self-attention modules 778), pooling layers and / or flattening layers 780A, 780B, 780C, 780D (collectively referred to as pooling layers and / or flattening layers 780), a concatenation module 782 that concatenates outputs, and output nodes such as time loss functions to events 784, failure times 786, censorship 788, and / or analogous functions.

[0095] In some embodiments, the self-attention architecture can accept multiple different types of data as input, such as heart rate data 772A, ECG data 772B, PPG data 772C, acceleration data 772D, and / or similar. The architecture may include multiple encoders 774, each encoder capable of taking in one or more of the input data. For example, each encoder may be trained to output latent features for individual input data types (e.g., a first encoder for rhythm data, a second encoder for ECG data).

[0096] In some embodiments, the self-attention module 778 can receive latent features output by an encoder as input. The self-attention module 778 can find similarities among these latent features based on a distance metric or a loss function. Based on these similarities and the input latent features, the self-attention module 778 can compute a weighted output vector. For example, the self-attention module 778A can determine the output by weighting the 2-minute, 5-minute, and 10-minute rhythm data more highly than the rest, and the self-attention module 778B can determine the output by weighting the 1-minute, 4-minute, and 6-minute ECG data more highly than the rest.

[0097] In some embodiments, the outputs of the self-attention module 778 can be pooled and / or flattened into a single value or vector. For example, the maximum or average value can be taken from the outputs of the self-attention module. The self-attention architecture can concatenate the outputs into a single vector. In some cases, the outputs 782 are temporally aligned, so the lengths of the outputs 782 do not need to be the same size.

[0098] In some embodiments, the self-attention architecture may be trained using patient outcome information, including failure time and censoring variables, via a time loss function to event (e.g., negative log partial likelihood of Cox proportional hazards)784, failure time786 (time until the event occurs), and censoring788 (whether the failure was observed at that time), which predicts the probability of survival / failure (occurrence or non-occurrence of the event) within a given time frame. In some embodiments, the output may include a single-value survival or death probability, or a probability vector of survival or death (e.g., a probability vector over different time periods).

[0099] Figure 7D shows a network training architecture for applying an LSTM transformer block to the decoder output. In this example, the autoencoder and LSTM transformer blocks from Figure 7A are shown. However, it is understood that other neural networks and / or architectures can be applied. During training, input data (such as historical patient data including rhythms, ECGs, PPGs, and / or analogues) is input to encoder 604. The data is processed through encoder 604, decoder 608, and LSTM transformer blocks 702, 704, 706, and 708. The output of the last block 708 of the LSTM transformer is input to loss function 722. True values ​​(e.g., from failure time 701A, censorship 710B) are also input to loss function 722. Based on the output of loss function 722, the system adjusts the weights of the neural network. For example, if the output of the last block 708 of the LSTM converter is close to the true value, the loss function 722 can output the minimum loss, thereby allowing the system to strengthen the neural network's weights (for example, by increasing the weights so that the neural network is strengthened to produce an output closer to the true value).

[0100] Figure 7E shows a network inference architecture for applying LSTM transformer blocks to the decoder output to generate hazard probabilities and / or survival probabilities. In this example, the autoencoder and LSTM transformer blocks from Figure 7A are shown. However, it is understood that other neural networks and / or architectures can be applied. During inference, input data (such as rhythm, ECG, PPG, and / or current patient data containing these) is input to encoder 604. The data is processed through encoder 604, decoder 608, and LSTM transformer blocks 702, 704, 706, and 708. The output of the last block 708 of the LSTM transformer is output through the output node or layer 730.

[0101] In some embodiments, the potential output may include patient or user probabilities such as hazard probabilities or survival probabilities. Hazard probabilities are calculated over time t. i This may include the probability of failure (e.g., patient death) in time t. i Under the condition that the patient survived until time t i The probability of dying at time t can be expressed as follows. Therefore, the probability can be determined by filtering out users who have survived up to that point and determining the user's probability based on a subset of past users who have survived up to that point. The probability of survival is given by time t i It can include the probability of users remaining alive after a certain period.

[0102] Figure 7F shows a network training architecture for applying a self-attention architecture using a loss function. In this example, the self-attention architecture of Figure 7C is shown. However, it is understood that other neural networks and / or architectures can be applied. During training, input data (such as historical patient data including rhythms, ECGs, PPGs, and / or analogues) is input to encoder 774. The data is processed through encoder 774, self-attention architecture 778, pooling / flattening layer 780, and concatenation 782. The output of concatenation 782 is input to time to event loss function 784. The true value (e.g., failure time 786, censored 788) is also input to loss function 784. Based on the output of loss function 784, the system adjusts the weights of the neural network.

[0103] Figure 7G shows a network inference architecture for applying a self-attention architecture to generate hazard probabilities and / or survival probabilities. In this example, the self-attention architecture of Figure 7C is shown. However, it is understood that other neural networks and / or architectures can be applied. During inference, input data (such as current patient data including rhythm, ECG, PPG, and / or analogues) is input to encoder 774. The data is processed through encoder 774, self-attention architecture 778, pooling / flattening layer 780, and concatenation 782. The output of concatenation 782 is output through output node or layer 796. Potential outcomes such as hazard probabilities and survival probabilities 798 are determined.

[0104] Examples of unsupervised feature extraction, model development, and patient clustering In some embodiments, the model is referred to as a “feature engineering” engine. Specifically, the model can provide a representation of the complexity and interaction structure of features that largely distinguish one record from others. For example, autoencoders, principal component, and / or random projection architectures may be able to extract features for a complete or partial record that represent the overall informational content of the record.

[0105] In some embodiments, all recordings can be passed through a model to obtain a “feature representation summary” (or “signature”), which can then be used in conjunction with many other modeling techniques to identify the feature summary that best relates to or predicts a particular set of results described earlier. The “signature” may include a compressed mathematical representation of the original signal. The signature can consist of a set of numbers representing the original signal in a lower-dimensional space. The signature can be represented as a vector or tensor, where each number in the representation can be thought of as a coordinate of the original signal in a new k-dimensional space. As an example, in Figure 8 (further described herein), a set of individual signatures performed through the model is projected into a two-dimensional space and represented as points clustered based on their similarity. Higher-dimensional representations, including but not limited to the original signal, are given for each cluster centroid in the form of line plots and color strips representing the magnitude of the projection of the original signal into a greater-than-two-dimensional coordinate space represented by the points. The advantages of these models are that given the results and event The goal is to extract informational content that is not specific to the modeling approach (e.g., classification versus survival rate).

[0106] In some embodiments, every new record can be run through the model to create a new "signature" for that record. These signatures are stored in a database and clustered based on appropriate similarity metrics. Such clusters can be created using conventional machine learning clustering approaches or based on deep learning representations such as sparse autoencoders. This database, along with demographic and clinical data, can be used to develop supervised models for outcomes of interest.

[0107] Alternatively, based on the creation of a universe of patient signatures, which may include demographic and clinical data, the system can assess the proximity of new records to existing ones, and, if the latter are annotated with specific class labels and / or risk groups, assign clinically useful features to the new records. Furthermore, sequential results can generate a single patient signature, providing information about the patient's clinical course over time and allowing assessment of the rate and severity of transitions from healthy to diseased states and to new states after intervention.

[0108] Risk prediction In some embodiments, as described herein, many categories of algorithms are available for use with the extracted features. Examples of such algorithms include: a time-to-event approach (parametric / nonparametric survival analysis) by combining demographic and clinical factors as predictors to the time-to-event approach (combination of multiple methods); classification of event / no-event probabilities obtained from ML / DL classification by combining demographic and clinical factors as predictors to the time-to-event approach (combination of multiple methods); and / or deep learning approaches for event risk prediction, including methods that use loss functions as used in conventional statistical models for survival, LSTM, RNN, and transformer models.

[0109] Prediction report In some embodiments, the system can provide a digital representation of risk prediction. For example, the system can provide a risk score out of 100, where 0 represents no risk and 100 guarantees the occurrence of a future event, and the corresponding risk for a given individual of a specific event within a predetermined follow-up time (see, e.g., Figure 9). The system can include corresponding risk curves derived from validation studies as supporting evidence (see, e.g., Figure 10). The curves may be stratified based on demographic characteristics, symptoms for monitoring, clinical characteristics such as CHA2DS2-VASc scores and HAS-BLED scores, treatment regimens, and / or analogues. The system can include corresponding treatment benefit curves as a function of the event risk score and other demographic / clinical characteristics based on retrospective or randomized prospective studies as supporting evidence of treatment effectiveness. The system can visually display "record signatures" and measures of proximity to neighboring records. The system can report on treatment choices (specific interventions, medications, and percentages of such information) and outcomes for patients near records of interest.

[0110] Examples of patient data clustering Figure 8 shows an embodiment 800 in which individual signatures from individuals that may progress differently over an event timeframe or respond well to a particular intervention (e.g., a signature indicating the probability of responding to a given intervention) are grouped together (represented as projections on a 2D plane). For example, various group signatures 802A, 802B, 802C can be projected onto a 2D plane as circles 804A, 804B, 804C representing various groupings. Stars 806A, 806B, 806C represent representative signatures (centroids) of groups with similar characteristics, and circles represent individual observations that are "closer" to a particular centroid in terms of mathematical distance. Among other dimensionality reduction methods, encoders can create dimensionally reduced representations of bottleneck information, such as 2D vectors, from large amounts of data. If the goal is to reduce the dimensionality to 2D, the system can determine predictions for these 2D dimensions. For example, a two-dimensional model could include two distinct groups of people, such as a first group that benefited from an intervention and a second group that did not. If a patient is close to the centroid of the group of people who benefited from a certain treatment, the system can predict that the patient may benefit from a certain type of intervention.

[0111] In some embodiments, the system can create clusters from multiple modalities. The system can use one or more neural networks (such as deep learning algorithms) to generate representations of signals in latent space. These neural networks can be trained to incorporate contrastive losses for assigning distances between signals. The contrastive loss may include training examples with values ​​that are "close" (e.g., measurement vectors corresponding to individuals with similar characteristics and identical or similar clinical outcomes) along with other characteristics that are "far" from the reference sample.

[0112] In some embodiments, the system can use a sample as a reference point and compare it to a large number of other "near" and "far" samples. The neural network is trained to generate and apply a function that can compute a distance matrix between samples. This distance metric is based on the latent representation of each sample (e.g., a projection onto a coordinate system that places more similar elements closer together and more dissimilar elements further apart).

[0113] In some embodiments, the system can apply these latent vectors as the basis for a clustering algorithm (e.g., k-means, hierarchical clustering, BIRCH, etc.) to create clusters of patients with similar outcomes and establish their representative samples ("centroids"). Advantageously, the system can make better predictions based on a holistic view of the data (neighbors of patient samples with similar outcomes) rather than corresponding to any single patient sample. New samples can then be clustered using a combination of latent projections obtained from a contrast-loss-based network and a clustering algorithm. In another embodiment, the same latent features can be used to create classifications of outcome classes by using a classification neural network. The trained network can then be used to predict any new observations.

[0114] Figure 9 shows a schematic diagram of a risk score alert for a patient whose records include markers indicating a high confidence in the onset / detection of atrial fibrillation within a pre-specified follow-up period, even though no events occurred within the current patch wearing time. The risk score can be a score between 1 and 100, such as 75; other scores within a score range, such as 1 to 5 or 1 to 10; or other scores, such as letter grading. The atrial fibrillation portion and the other (non-atrial fibrillation) portion can represent the population of people. For example, 80% of the population does not develop atrial fibrillation, while 20% does. The risk score can be applied to people who do not have atrial fibrillation and can help develop monitoring strategies to manage future risk. The risk score can also be applied to people with atrial fibrillation to manage the condition, such as pharmacological, surgical, or lifestyle interventions. The risk score can be applied to the entire population, regardless of whether the patient has atrial fibrillation or not, and can optimize the public health of the population.

[0115] Figure 10 shows Graph 1000 with embodiments of risk curves stratified by demographic or clinical characteristics. This graph can show risk score curves based on cumulative hazards over time, such as time / day / year. Even with the same risk score, the probability of outcomes may differ based on such characteristics. For example, a low-risk curve is suggestive for patients aged 30-50 years. Thus, risk scores can have different meanings for patients of different ages. In some embodiments, risk curves can be stratified by clinical characteristics such as the patient's current measurements (e.g., blood pressure, PPG). In some embodiments, risk curves can be stratified by past health history.

[0116] Examples of reports Figure 11 shows an example of a report 1100 for a patient who has not been diagnosed with atrial fibrillation. The report may include a display of an atrial fibrillation risk score 1102 indicating the patient's likelihood of developing atrial fibrillation in the future. The number 78 may represent a high score (and / or be displayed within a circle). The report may also include a display of relative risk 1104. 6X may represent the relative risk as a fold change for this patient compared to a person with the same age and sex characteristics but without the abnormal characteristics. The relative risk may be derived from a model and / or may be another representation of risk determined based on morphological, heart rate, demographic, and / or other findings.

[0117] In some embodiments, report 1100 may include a graph 1106 of the predicted atrial fibrillation risk over the next five years. In some embodiments, the graph may span different time periods such as days, weeks, or months. Graph 1106 may display the risk curve as a function of the score (after five years). Report 1100 may include another graph 1108 that provides a prediction of the absolute risk progression for this person as a function of time. The absolute risk may be displayed as a percentage, such as an absolute risk of 18% after five years. Report 1100 may include a table 1110 that provides the risk of onset over time, such as months, along with the atrial fibrillation risk score and / or the progression of the risk.

[0118] Figures 12A and 12B illustrate embodiments of reports 1200, 3500 for patients with atrial fibrillation identified in cardiac monitor recordings. Reports 1200, 3500 can show the patient's risk of adverse outcomes over time. Outcomes may correspond to risk scores indicating the likelihood of events occurring within a period such as 36 months. Scores and / or risk curves can be generated by deep learning algorithms that can take into account the patient's ECG morphology, findings, heart rate, and / or demographic conditions. Reports 1200, 3500 may include risk curve graphs representing the risks associated with outcomes of interest, such as stroke 1202, heart failure 1204, hospitalization 1206, and / or mortality 1208. Each risk curve graph may include risk curves for no intervention (solid line), with atrial fibrillation ablation (first dotted line), and with anticoagulant therapy (second dotted line). For example, a no-intervention risk curve can represent the risk as a function of time for a cohort with the same characteristics but no intervention. Risk circles 1203, 1205, 1207, and 1209 can show levels of risk, such as a high risk score of 78. Risk curves can show risk reduction based on specific treatments. For example, in stroke curve 1202, the patient's risk is reduced by 28% with atrial fibrillation ablation and by 23% with anticoagulation therapy.

[0119] In some embodiments, the system can generate risk scores and / or curves for one or more other physiological signals. For example, the system can generate curve graphs of future risk for stroke, heart failure, cardiovascular hospitalization, structural heart disease (such as left / right ventricular hypertrophy, reduced infusion rate, atrial / mitral regurgitation, stenosis, tricuspid regurgitation, etc.), and / or similar conditions.

[0120] Example of a flowchart Figure 13A shows an embodiment of flow chart 1300 for training an encoder and decoder and applying the encoder to a wearable device to infer the onset of cardiac arrhythmias. In block 1302, the system can train the encoder and decoder to receive and output cardiac data. The encoder can output features of the cardiac data. The encoder and decoder can be trained to output data that represents the cardiac data with sufficient fidelity, such as an input similarity threshold level.

[0121] In block 1304, the system can generate a software package containing an encoder for the wearable patch so that the wearable patch can apply the encoder to new patient cardiac data. In block 1306, the wearable patch can record patient cardiac data, for example, by using electrodes on the wearable patch. In block 1308, the wearable device can apply the encoder to the patient's cardiac data and generate features of the patient's cardiac data. The encoder can output features such as boundaries, edges, single-line gradients, offsets, signal magnitudes, and / or analogs thereof. The encoder output can have a smaller number of dimensions than the data input to the encoder.

[0122] In block 1310, the wearable device can transmit the features output from the encoder to a remote computing device. Advantageously, the wearable device may have many technical advantages, such as longer battery life, smaller memory capacity requirements, and / or other advantages further described herein. In block 1312, the remote computing device can apply the features to a machine learning process for inferring the onset of cardiac arrhythmias. The machine learning process may be a separate process from the encoder and decoder. The machine learning process may be trained separately from the encoder and decoder, as further described herein.

[0123] Figure 13B shows an embodiment of flow chart 1330 for training an encoder and a machine learning model to infer the onset of cardiac arrhythmias, applying the encoder on a wearable device, and applying the machine learning model on a remote computing device. In block 1332, the system can receive cardiac data and train an encoder and a separate machine learning model to output an inference of the onset of cardiac arrhythmias. The training can apply historical patient cardiac data and the cardiac arrhythmia status of the same patient at later time points, such as one month, five months, one year, and five years later. Advantageously, the encoder can be specifically trained to output the cardiac data features most relevant to infer the onset of cardiac arrhythmias.

[0124] In block 1334, the system can generate a software package including an encoder for the wearable patch so that the wearable patch can apply the encoder to new patient cardiac data, such as in block 1304 and / or as further described herein. In block 1336, the wearable patch can record patient cardiac data, such as by using electrodes on the wearable patch, such as in block 1306 and / or as further described herein. In block 1338, the wearable device can use the encoder to apply patient cardiac data to generate features of the patient cardiac data, such as in block 1308 and / or as further described herein.

[0125] In block 1340, the wearable device can transmit the features output from the encoder to a remote computing device, such as in block 1310, and / or as further described herein. In block 1342, the remote computing device can apply the features to a machine learning process to infer the onset of cardiac arrhythmias. The machine learning process is trained using the encoder. However, the encoder is applied on the wearable device, and the machine learning process is applied on the remote computing device.

[0126] Figure 13C shows an embodiment of flow chart 1360 for training an encoder, decoder, and machine learning model using historical patient cardiac data to infer the onset of cardiac arrhythmias. In block 1362, the system can access a record of historical patient cardiac data. For example, historical patient cardiac data may include data and corresponding historical cardiac data from hundreds or thousands of past patients. In block 1364, the system can train the encoder and decoder using historical patient cardiac data, such as in block 1302 and / or as further described herein.

[0127] In block 1366, the system can access the patient's status regarding cardiac arrhythmias at a later time in the cardiac data measurement record. For example, a patient's past cardiac data may be from five years ago, and that data did not show cardiac arrhythmias five years ago. However, the patient's status five years later was that of a patient who later developed cardiac arrhythmias.

[0128] In block 1368, the system can train a machine learning process to infer the onset of cardiac arrhythmias based on the patient's condition. The data that can be used may include features output by an encoder trained using historical patient cardiac data. Thus, the condition and historical patient cardiac data can correspond to the same individual.

[0129] In block 1370, the wearable patch can record patient cardiac data, for example, by using electrodes on the wearable patch, as described in block 1306 and / or further herein. In block 1372, the wearable device can apply the patient cardiac data using an encoder to generate features of the patient cardiac data, as described in block 1308 and / or further herein. In block 1374, the remote computing device can apply the features to a machine learning process to infer the onset of cardiac arrhythmias, as described in block 1312 and / or further herein.

[0130] Computing systems and methods In some embodiments, the systems, tools, and methods of using them described above enable bidirectional interaction and data collection performed by the computing system 1400. Figure 14 shows the computing system 1400 as a network 1402 , and wearable systems Hmm, Gateway device Sna This block diagram shows an embodiment communicating with various external computing systems 1404. Computing system 1400 may be used to carry out the systems and methods described herein. The external systems 1404 are shown grouped together, but it is recognized that each system may be external to and / or located remotely from one another.

[0132] In some embodiments, the computing system 1400 includes one or more computing devices, such as a server, laptop computer, mobile device (e.g., smartphone, smartwatch, tablet, personal digital assistant), kiosk, car console, or media player. In one embodiment, the computing device 1400 includes one or more central processing units (CPUs) 1405, each of which may include a conventional or proprietary microprocessor. The computing device 1400 further includes one or more memories 1430, such as 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 1420, such as a hard drive, diskette, solid-state drive, or optical media storage device. In certain embodiments, the processing device, cloud server, server, or gateway device is part of the computing system 1400 It may be implemented as follows. In one embodiment, modules of the computing system 1400 are connected to a computer using a standards-based bus system. In different embodiments, the standards-based bus system may be implemented using, for example, PCI (Peripheral Component Interconnect), Microchannel, SCSI (Small Computer computing system Interface), ISA (Industrial Standard Architecture), and EISA (Extended ISA) architectures. Furthermore, the functions provided by the components and modules of the computing device 1400 may be combined into fewer components and modules, or separated into even more additional components and modules.

[0133] The computing device 1400 may be controlled and tuned by operating system software, such as 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 Macintosh systems, the operating system may be an available operating system such as MAC OS X. In other embodiments, the computing device 1400 may be controlled by a proprietary operating system. Conventional operating systems, in particular, control and schedule computer processes for execution, perform memory management, provide file systems, networking, I / O services, and provide user interfaces such as graphical user interfaces (GUIs).

[0134] An exemplary computing device 1400 may include one or more I / O interfaces and devices 1410, such as a touchpad or touchscreen, but may also include a keyboard, mouse, and printer. In one embodiment, the I / O interfaces and devices 1410 include one or more display devices (such as a touchscreen or monitor) that enable the visual presentation of data to the user. More specifically, the display devices may provide, for example, a GUI, application software data, and a multimedia presentation. The computing system 1400 may also include one or more multimedia devices 1440, such as a camera, speaker, video card, graphics accelerator, and microphone.

[0135] In one embodiment of a computing system and application tool, the I / O interface and device 1410 may provide a communication interface to various external devices. In one embodiment, the computing device 1400 communicates, for example, via a wired, wireless, or wired and wireless communication link 1415 to a network including one or more of a local area network, a wide area network, and / or the internet. 1402 It is electronically coupled to the network. 1402 It can communicate with various sensors, computing devices, and / or other electronic devices via wired or wireless communication links.

[0136] In some embodiments, filter extraction conditions, signals, and data may be processed by a rhythm inference module application tool according to the methods and systems described herein and provided from one or more data sources 1410 to a computing system 1400 via a network 14002. The data sources may include one or more internal and / or external databases, data sources, and physical data stores. The data source 1410, the external computing system 1404, and the rhythm interface module 1490 may include databases for storing data (e.g., feature data, raw signal data, patient data) according to the systems and methods described above, and databases for storing data processed according to the systems and methods described above (e.g., data sent to sensors, data sent to clinicians). [Figure] 15In one embodiment, the sensor data 1550 can store data received from sensors, data received from clinicians, etc., in some embodiments. The rule database 1560 can store data (e.g., instructions, preferences, profiles) that establish threshold parameters for analyzing feature data, in some embodiments. In some embodiments, one or more of the databases or data sources are Sybase, Oracle, CodeBase, MySQL (Registered trademark) It may be implemented using relational databases such as SQLite and Microsoft® SQL Server, as well as other types of databases such as flat file databases, entity-relational databases, object-oriented databases, NoSQL databases, and / or record-based databases.

[0137] In one embodiment, the computing system includes a rhythm interface module 1490 which can be stored in a mass storage device 1420 as executable software code executed by a CPU 1405. The rhythm interface module 1490 may have a feature module 1510, an alternative data module 1520, an inference module 1530, a feedback module 1540, a sensor data database 1550, and a rule database 1560. These modules may include, as an example, components such as software components, object-oriented software components, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. These modules are also configured to perform the processing disclosed herein.

[0138] In some embodiments, various components illustrated or described herein can be implemented as software and / or firmware on a processor, controller, ASIC, FPGA, and / or dedicated hardware. The software or firmware may include instructions stored in non-transient computer-readable memory. The instructions can be executed by the processor, controller, ASIC, FPGA, or dedicated hardware. Hardware components such as controllers, processors, ASICs, and FPGAs may include logic circuits.

Claims

1. A computing system for inferring the onset of cardiac arrhythmias in a user, The computing system is A receiver configured to receive the patient's cardiac signals detected by a sensor applied to the patient, A hardware processor configured to apply cardiac signals to a first machine learning model, The first machine learning model is configured to infer the onset of cardiac arrhythmias based on the user's heart signals. The first machine learning model is generated using the machine learning training process. The machine learning training process is A step of accessing past patient data for multiple patients, wherein the past patient data includes past cardiac signal data for multiple patients. A step of accessing the cardiac arrhythmia status for each of several patients, wherein the cardiac arrhythmia status for each of the several patients is determined after the time when past cardiac signal data was recorded for each of the several patients, A step of training a first machine learning model using patient outcome information for multiple patients, including failure time (time until a cardiac event occurs) and a censoring variable via a time-to-event loss function that predicts the probability of survival / failure within a specified time frame, in order to infer the onset of cardiac arrhythmias, and at least partially based on past cardiac signal data of multiple patients and the cardiac arrhythmia status of each of the multiple patients, wherein the user has different steps for multiple patients. including, Computing system.

2. The hardware processor is further configured to apply the user's heart signal to a second machine learning model. The second machine learning model is trained to remove artifacts from the user's heart signal to obtain a filtered heart signal. The step of applying cardiac signals to the first machine learning model is: This includes applying the filtered cardiac signal as input to a first machine learning model. The computing system according to claim 1.

3. The hardware processor is further configured to apply the user's heart signal to a second machine learning model. The second machine learning model is trained to reduce noise from the user's heart signal and obtain a filtered heart signal. The step of applying cardiac signals to the first machine learning model is: This includes applying the filtered cardiac signal as input to a first machine learning model. The computing system according to claim 1.

4. Inferring the onset of cardiac arrhythmias involves determining the probability of cardiac arrhythmias occurring at a point later than the monitoring period. The computing system according to claim 1.

5. At least some of the past patient data did not show cardiac arrhythmias at the time the past cardiac signal data was recorded. The computing system according to claim 1.

6. The first machine learning model is further configured to process photoelectric pulse wave (PPG) data, heart rate, or heart rate acceleration data to infer the onset of cardiac arrhythmias. The computing system according to claim 1.

7. Cardiac arrhythmias include at least one of the following: ventricular tachycardia, supraventricular tachycardia, abduction, or ventricular fibrillation. The computing system according to claim 1.

8. Past patient data further includes the cardiac arrhythmia status of multiple patients. The computing system according to claim 1.

9. The recovery includes electrodes. The computing system according to claim 1.

10. A method for training a machine learning model to infer the onset of cardiac arrhythmias, The method is, A hardware processor configured to execute instructions performed by a computer, This includes steps to access past patient data for multiple patients, Past patient data includes past cardiac signal data from multiple patients. Past cardiac signal data includes cardiac signals, At least some of the past cardiac signal data did not show cardiac arrhythmias at the time the cardiac signals were recorded. The method is also done by the hardware processor, A step of accessing the cardiac arrhythmia status for each of several patients, wherein the cardiac arrhythmia status for each of the several patients is determined after the time when cardiac signals were recorded for that patient, Steps include training a first machine learning model based on patient outcome information for multiple patients, including failure time (time until a cardiac event occurs) and a censoring variable via a time-to-event loss function that predicts the probability of survival / failure within a specified time frame, in order to infer the onset of cardiac arrhythmias in the user, wherein the user is not included in the multiple patients, and Step 3. including, method.

11. This further includes applying the user's heart signal to a second machine learning model trained to remove artifacts from the user's heart signal and obtain a filtered heart signal. The method according to claim 10.

12. The process further includes applying the filtered cardiac signals to a first machine learning model to infer the onset of the user's cardiac arrhythmia. The method according to claim 11.

13. The process further includes applying the user's heart signal to a second machine learning model that has been trained to reduce noise from the user's heart signal and obtain a filtered heart signal. The method according to claim 10.

14. The process further includes applying the filtered cardiac signals to a first machine learning model to infer the onset of the user's cardiac arrhythmia. The method according to claim 13.

15. The step of inferring the onset of cardiac arrhythmias includes determining the probability of cardiac arrhythmias occurring within a specific period. The method according to claim 10.

16. The first step of training the machine learning model further includes training the first machine learning model to infer the occurrence of a user's hospitalization, onset of heart failure, onset of stroke, or death. The method according to claim 10.

17. This further includes training a second machine learning model to infer the occurrence of a user's hospitalization, onset of heart failure, onset of stroke, or death. The input to the second machine learning model includes cardiac signals over a longer period than those provided to the first machine learning model. The method according to claim 10.

18. The process further includes a step of determining the atrial fibrillation load from the user's cardiac signals, Atrial fibrillation load includes the amount of time a user spends experiencing atrial fibrillation during a given period. The method according to claim 10.

19. When performed by the processor, A step of accessing past patient data of multiple patients, wherein the past patient data includes past cardiac signal data of multiple patients, and the past cardiac signal data includes cardiac signals, At least some of the past cardiac signal data shows that there are steps that do not show cardiac arrhythmias when the cardiac signals are recorded, A step of accessing the cardiac arrhythmia status for each of several patients, wherein the cardiac arrhythmia status for each of the several patients is determined after the time when cardiac signals were recorded for that patient, Steps include training a first machine learning model based on patient outcome information for multiple patients, including failure time (time until a cardiac event occurs) and a censoring variable via a time-to-event loss function that predicts the probability of survival / failure within a specified time frame, in order to infer the onset of cardiac arrhythmias in the user, wherein the user is not included in the multiple patients, and Step 3. A non-transient computer storage medium that stores computer-executable commands that cause a processor to perform an operation including [a specific operation].

20. The operation is The steps include applying the user's heart signal to a second machine learning model trained to remove artifacts from the user's heart signal or reduce noise to obtain a filtered heart signal, The process further includes the step of applying filtered cardiac signals to a first machine learning model to infer the onset of cardiac arrhythmias in the user, The non-transient computer storage medium according to claim 19.

Images