System and method for measuring hemodynamic parameters using a wearable cardiovascular sensor

Abstract

An exemplary embodiment of the present disclosure provides a system for assessing cardiac health, comprising: a first sensor, a second sensor, a processor, and a memory. The first sensor can be configured to measure at least one electrical characteristic of a user's heart. The second sensor can be configured to measure cardiogenic oscillations of the user. The memory can include instructions that, when executed by the processor, cause the processor to generate an assessment of the user's cardiac health, the assessment including data indicative of a filling characteristic of the heart based at least in part on measurements from the first sensor and the second sensor.

Description

[Technical Field]

[0001] (Cross-reference of related applications) This application claims priority under U.S. Provisional Application No. 63 / 117,766, filed on 24 November 2020, which is incorporated herein by reference in its entirety as it is shown below.

[0002] (Government licensing rights) This disclosure was made with government support under authorization number 1R01HL130619-A1 granted by the National Institutes of Health. The U.S. government has certain rights in this disclosure.

[0003] (Technical field) This disclosure relates, in general, to medical systems and methods, and more specifically to wearable systems and methods for assessing cardiac health. [Background technology]

[0004] Heart failure (HF) is a debilitating disease in the United States, causing approximately 300,000 deaths and more than 800,000 hospitalizations annually. HF-related costs exceed $30 billion per year and are projected to rise to $70 billion by 2030. One factor contributing to the high cost and mortality rates of HF is the high likelihood of readmission after the initial hospitalization. Consequently, there is a need for improved approaches to HF care that reduce costs while improving outcomes by enabling patients to be better managed at home and reducing readmissions.

[0005] Clinical strategies that have demonstrated favorable outcomes in the home management of HF patients include measuring pulmonary artery (PA) pressure using an implantable device and performing titration therapy in response to elevated PA pressure (indicating congestion or impending decompensation). While detecting blood flow congestion using PA pressure is a sound and effective method in large randomized clinical trials, the costs and complications associated with surgery make this method suitable for only a small number of HF patients.

[0006] Therefore, there is a need for non-invasive and inexpensive technologies that can detect elevated PA pressure without using implantable devices that allow observation of HF patients, as well as personalized treatment methods that enable effective care and improved treatment outcomes, such as reducing hospitalizations and improving quality of life. [Overview of the project]

[0007] This disclosure relates to medical systems and methods. The disclosed technology includes a system for evaluating the cardiac health of a user. This cardiac health evaluation system may include a first sensor, a second sensor, a processor, and memory. The first sensor may be configured to measure at least one electrical characteristic of the user's heart. The second sensor may be configured to measure the user's cardiogenic oscillation. The memory may include instructions that, when executed by the processor, cause the processor to generate an evaluation of the user's cardiac health, which includes data indicating the filling characteristics of the heart, based at least in part on measurements from the first and second sensors.

[0008] The first sensor can be configured to measure the user's electrocardiogram signal.

[0009] The second sensor can be configured to measure the user's electrocardiogram signal.

[0010] The memory may include instructions that, when executed by the processor, cause the processor to generate an assessment of the user's cardiac health, which includes data showing the filling characteristics of the heart based at least partially on the horizontal axis of the user's electrocardiogram signal.

[0011] The memory can include an instruction that, when executed by a processor, causes the processor to generate an evaluation of the user's heart health state that includes data indicative of the filling characteristics of the heart based at least in part on the head-to-foot axis of the user's mental electrocardiogram signal.

[0012] The memory can include an instruction that, when executed by a processor, causes the processor to generate an evaluation of the user's heart health state that includes data indicative of the filling characteristics of the heart based at least in part on the back-to-front axis of the user's mental electrocardiogram signal.

[0013] The memory can include an instruction that, when executed by a processor, causes the processor to generate an evaluation of the user's heart health state that includes data indicative of the filling characteristics of the heart based at least in part on the user's mental electrocardiogram signal during the diastolic phase of the heartbeat.

[0014] The evaluation of the heart health state can include data indicating the classification of the user's clinical state of heart failure.

[0015] The evaluation of the heart health state can include data indicating signs of changes in the user's hemodynamics.

[0016] The evaluation of the heart health state can include data indicating signs of changes in the user's filling pressure.

[0017] The evaluation of the heart health state can include data indicating signs of changes in the user's pulmonary artery pressure.

[0018] The evaluation of the heart health state can include data indicating signs of changes in the user's pulmonary capillary wedge pressure.

[0019] The memory can include an instruction that, when executed by a processor, causes the processor to perform a calibration step to generate reference values for one or more parameters associated with the user's heart health state.

[0020] The standard value could be the standard filling pressure.

[0021] The calibration step may include generating baseline values ​​using a population-level regression model.

[0022] The calibration step may include generating baseline values ​​using the user's personal data.

[0023] Personalized data can include data from right heart catheterization.

[0024] Personalized data can include data from clinical trials.

[0025] The first sensor may include a first wearable sensor positioned in close proximity to the heart. The second sensor may include a second wearable sensor positioned in close proximity to the heart.

[0026] A system for evaluating cardiac health may include a third sensor.

[0027] A third sensor can be configured to measure environmental parameters.

[0028] The third sensor can be configured to measure the user's photoelectric volume pulse wave signal.

[0029] The second sensor can be configured to measure the user's gyrocardiogram signal.

[0030] A system for evaluating cardiac health can include an output indicating the user's cardiac health status.

[0031] A system for assessing cardiac health may include a wireless transmitter. The wireless transmitter may be configured to wirelessly transmit the assessment of the user's cardiac health to a remote device.

[0032] A system for evaluating cardiac health may be wearable by the user.

[0033] The disclosed technology includes a wearable system for assessing a user's cardiac health. This wearable system for assessing cardiac health may include a first sensor, a second sensor, and a controller. The first sensor may be configured to measure at least one electrical characteristic of the user's heart. The second sensor may be configured to measure the user's cardiogenic oscillation. The controller may be configured to perform a calibration step and generate reference values ​​for one or more parameters associated with the user's cardiac health. The controller may be configured to create an assessment of the user's cardiac health, including data showing cardiac filling characteristics, which are at least partially based on the reference values ​​and measurements from the first and second sensors.

[0034] The first sensor can be configured to measure the user's electrocardiogram signal.

[0035] The second sensor can be configured to measure the user's electrocardiogram signal.

[0036] An assessment of a user's cardiac health, including data indicating cardiac filling characteristics, can be based at least partially on the horizontal axis of the user's psychocardiogram signal.

[0037] An assessment of a user's cardiac health, including data indicating cardiac filling characteristics, can be based at least partially on the head-foot axis of the user's psychocardiogram signal.

[0038] An assessment of a user's cardiac health, including data indicating cardiac filling characteristics, can be based at least partially on the dorsal-ventral axis of the user's psychocardiogram signal.

[0039] Assessing a user's cardiac health, including data indicating cardiac filling characteristics, can be based at least partially on the user's psychocardiogram signals during the diastolic phase of the heartbeat.

[0040] The assessment of cardiac health can include data that classifies the user's clinical state of heart failure.

[0041] The assessment of cardiac health can include data indicating signs of changes in the user's hemodynamics.

[0042] The assessment of cardiac health can include data indicating signs of changes in the user's filling pressure.

[0043] The assessment of cardiac health can include data indicating changes in the user's pulmonary artery pressure.

[0044] The assessment of cardiac health can include data indicating changes in the user's pulmonary capillary wedge pressure.

[0045] The standard value could be the standard filling pressure.

[0046] The calibration step may include generating baseline values ​​using a population-level regression model.

[0047] The calibration step may include generating baseline values ​​using the user's personal data.

[0048] Personalized data can include data from right heart catheterization.

[0049] Personalized data can include data from clinical trials.

[0050] The first sensor may include a first wearable sensor positioned in close proximity to the heart. The second sensor may include a second wearable sensor positioned in close proximity to the heart.

[0051] A wearable system for assessing cardiac health may include a third sensor.

[0052] A third sensor can be configured to measure environmental parameters.

[0053] The third sensor can be configured to measure the user's photoelectric volume pulse wave signal.

[0054] The second sensor can be configured to measure the user's gyrocardiogram signal.

[0055] A wearable system for assessing cardiac health can include an output indicating the user's cardiac health status.

[0056] A wearable system for assessing cardiac health may include a wireless transmitter. The wireless transmitter may be configured to wirelessly transmit the user's cardiac health assessment to a remote device.

[0057] A wearable system for assessing cardiac health may be wearable by the user.

[0058] The disclosed technology includes a method for non-invasively observing a user's cardiac health. The method may include the step of receiving a first signal, which indicates at least one electrical characteristic of the user's heart, via a wearable device. The method may include the step of receiving a second signal, which indicates the user's cardiogenic oscillation, via a wearable device. The method may include the step of creating an assessment of the user's cardiac health, which includes data indicating the filling characteristics of the heart, based at least in part on the first and second signals. The method may include the step of providing an output indicating the assessment of the user's cardiac health.

[0059] The first signal can represent the user's electrocardiogram signal.

[0060] The second signal can represent the user's electrocardiogram signal.

[0061] An assessment of a user's cardiac health, including data indicating cardiac filling characteristics, can be based at least partially on the horizontal axis of the user's psychocardiogram signal.

[0062] An assessment of a user's cardiac health, including data indicating cardiac filling characteristics, can be based at least partially on the head-foot axis of the user's psychocardiogram signal.

[0063] An assessment of a user's cardiac health, including data indicating cardiac filling characteristics, can be based at least partially on the dorsal-ventral axis of the user's psychocardiogram signal.

[0064] Assessing a user's cardiac health, including data indicating cardiac filling characteristics, can be based at least partially on the user's psychocardiogram signals during the diastolic phase of the heartbeat.

[0065] The assessment of cardiac health can include data that classifies the user's clinical state of heart failure.

[0066] The assessment of cardiac health can include data indicating signs of changes in the user's hemodynamics.

[0067] The assessment of cardiac health can include data indicating signs of changes in the user's filling pressure.

[0068] The assessment of cardiac health can include data indicating changes in the user's pulmonary artery pressure.

[0069] The assessment of cardiac health can include data indicating changes in the user's pulmonary capillary wedge pressure.

[0070] This method may further include a calibration step that generates reference values ​​for one or more parameters associated with the user's cardiac health.

[0071] The standard value could be the standard filling pressure.

[0072] The calibration step may include generating baseline values ​​using a population-level regression model.

[0073] The calibration step may include generating baseline values ​​using the user's personal data.

[0074] Personalized data can include data from right heart catheterization.

[0075] Personalized data can include data from clinical trials.

[0076] Wearable devices can be positioned close to the heart.

[0077] This method may include the step of receiving a third signal.

[0078] The third signal can indicate at least one environmental parameter.

[0079] The third signal can represent the user's photoplethysmography signal.

[0080] The second signal can represent the user's gyroscopic electrocardiogram signal.

[0081] These and other aspects of the Disclosure are described below in the detailed description and accompanying drawings. Other aspects and features of the embodiments will become apparent to those skilled in the art by referring to the following description of specific exemplary embodiments together with the drawings. Features of the Disclosure may be described in relation to specific embodiments and drawings, but all embodiments of the Disclosure may include one or more of the features described herein. Furthermore, one or more embodiments may be described as having a particular effective feature, and one or more such features may be used together with the various embodiments described herein. Similarly, exemplary embodiments may describe embodiments of apparatus, systems, or methods below, but it should be understood that such exemplary embodiments are implementable in various apparatus, systems, and methods of the Disclosure. [Brief explanation of the drawing]

[0082] The following detailed descriptions of specific embodiments of this disclosure will be better understood when read in conjunction with the accompanying drawings. Specific embodiments are illustrated in the drawings for illustrative purposes of this disclosure. However, it should be understood that this disclosure is not limited to the exact arrangement and means of the embodiments shown in the drawings. [Figure 1A] This figure shows an example of a system for evaluating cardiac health related to this disclosure. [Figure 1B] This figure shows a photograph of an example of a system for evaluating heart health related to this disclosure. [Figure 1C] This figure shows a photograph of an example of a system for evaluating heart health related to this disclosure. [Figure 1D] This figure shows an example of signal data from a system for evaluating cardiac health related to this disclosure. [Figure 1E] This figure shows an example of a system for evaluating cardiac health and an example of signal data related to this disclosure. [Figure 2] a. is a diagram showing the data processing method relating to this disclosure. b. is a diagram showing the steps used in the data processing method. c. is a diagram visualizing the data processing method. [Figure 3] a) is a graph of experimentally measured SCG data relating to this disclosure. b) is a graph of experimentally measured SCG data relating to this disclosure. c) is a graph of experimentally measured SCG data relating to this disclosure. d) is a graph of experimentally measured SCG data relating to this disclosure. [Figure 4] This is a graph showing the results of the classification experiment related to this disclosure. [Figure 5] This is a graph showing the results of the windowing experiment related to this disclosure. [Figure 6A] This figure shows an example of signal data from a system for evaluating cardiac health related to this disclosure. [Figure 6B] This figure shows an example of a signal processing method related to this disclosure. [Figure 6C] This figure shows an example of a signal processing method related to this disclosure. [Figure 7A] This is a graph of pulmonary artery pressure measured by experiment related to this disclosure. [Figure 7B] This is a graph of pulmonary capillary wedge pressure measured experimentally, relating to this disclosure. [Figure 7C] This is a graph of SCG data measured through experiments related to this disclosure. [Figure 8A] This is a graph of the correlation analysis related to this disclosure. [Figure 8B] This is a graph of the correlation analysis related to this disclosure. [Figure 8C] This is a graph of the correlation analysis related to this disclosure. [Figure 8D] This is a graph of the correlation analysis related to this disclosure. [Figure 9A]This is a chart of the correlation analysis related to this disclosure. [Figure 9B] This is a chart of the correlation analysis related to this disclosure. [Figure 10A] This is a graph showing the relative weighting of the algorithm features related to this disclosure. [Figure 10B] This is a graph showing the relative weighting of the algorithm features related to this disclosure. [Figure 11] This flowchart shows an example of a method for evaluating cardiac health related to this disclosure. [Figure 12] This flowchart shows an example of a method for evaluating cardiac health related to this disclosure. [Figure 13A] This figure shows the receiver operating characteristic (ROC) curve of the best classifier using features with a narrowed selection range set in the training set, as disclosed herein. [Figure 13B] This figure shows the ROC curve of a model trained on a training set and evaluated on a validation set (not shown) related to this disclosure. [Figure 14A] This figure shows three examples of pulse rates from a single decompensated subject related to this disclosure, along with the calculation of the best feature (srPower). [Figure 14B] This figure shows three examples of pulse rates from a single compensatory subject related to this disclosure, along with the calculation of the best feature (srPower). [Figure 15] (A) is a figure showing the plots of power spectral density (PSD) based on the average of each PSD for each SCG pulse for the decompensation curve from randomly selected decompensated subjects and the compensatory curve from randomly selected compensated subjects, as relating to this disclosure. (B) is a supplementary explanation of the low frequency band (5-40 Hz) of Figure 15A as relating to this disclosure. (C) is a supplementary explanation of the high frequency band (200-250 Hz) of Figure 15A as relating to this disclosure. [Figure 16](A) is a bar graph showing the number of times SFS is used to select features when performing 5-fold cross-validation as per the Disclosure. (B) is a bar graph showing the top 3 most important features calculated using linear SVM feature weighting as per the Disclosure. (C) is a bar graph showing the top 3 most important features as evaluated by permutation feature importance as per the Disclosure. [Modes for carrying out the invention]

[0083] This disclosure describes systems and methods for assessing cardiac health, including, for example, wearable systems and methods for assessing a user's cardiac health. For example, one embodiment of this disclosure provides a virtually inconspicuous wearable system for remotely observing a HF patient using a wearable seismocardiography (SCG) monitor. This system can assess the user's cardiac health and communicate the results to the user and / or caregiver. For example, this system can estimate the internal filling pressure of the heart, which may indicate congestion (preload).

[0084] While the disclosed technology describes a system and method for assessing cardiac health in this disclosure, those skilled in the art will understand that the disclosed technology is not limited thereto and is applicable to other scenarios and uses. For example, the disclosed technology is also envisioned to be applicable to any use related to quantifying preload, such as observing bleeding or detecting cardiovascular collapse from decreased blood volume.

[0085] Several implementations of the disclosed technology will be described in more detail with reference to the accompanying drawings. However, the technology of this disclosure may be implemented in many different forms and should not be construed as being limited to the implementations described below. The components described below as constituting various elements of the disclosed technology are illustrative and not intended to be limiting. It will be understood that other examples are conceivable in practice. Many suitable components that perform the same or similar functions as those described herein are intended to be included within the scope of the disclosed electronic devices and methods. Such other components not described herein may include, but are not limited to, components developed after the development of the disclosed technology.

[0086] Here, the use of terms such as “having,” “possessing,” “containing,” or “including” is intended to be unrestrictive and equivalent in meaning to terms such as “equipped,” or “equipped,” and does not preclude the existence of other structures, materials, or actions. Similarly, the use of terms such as “can” or “may” is unrestrictive and reflects that the structure, material, or action is not essential, but the absence of such terms is not intended to reflect that the structure, material, or action is essential. To the extent that the structure, material, or action is considered essential at that time, they are identified as such.

[0087] It will be understood that the reference to one or more method steps does not preclude the existence of further method steps or method steps intervening between explicitly identified steps. Similarly, it will be understood that the reference to one or more components in an apparatus or system does not preclude the existence of further components or components intervening between explicitly identified components. Furthermore, it is assumed that the disclosed methods and processes may, but do not necessarily, include all the steps described herein. In other words, the methods and processes relating to the disclosed technology may include some of the other methods and processes that are disclosed but not described.

[0088] Throughout the specification and claims, the following terms have at least the meanings expressly associated herein, unless otherwise stated. The term “or” is intended to mean comprehensive “or.” “A,” “an,” and “the” are intended to mean one or plural, unless otherwise stated or unless the context makes it clear that they relate to a singular form. “To comprise,” “to contain,” or “to include” means that at least the specified element or method step is present in the part or method, but does not exclude the presence of other elements or method steps, even if such other elements or method steps have the same function as the specified one.

[0089] As used herein, unless otherwise specified, the use of order-determining adjectives such as “first,” “second,” “third,” etc., to describe a general subject simply indicates that another example of a similar subject is being cited, and does not imply that these subjects described in this manner must be in a predetermined order in time, space, order, or any other way.

[0090] While the disclosed technology is described herein in relation to various systems and methods, it is assumed that alternative embodiments or implementations of the disclosed technology having the same or substantially similar characteristics as such methods or systems may be implemented. For example, any aspect, element, or feature described herein in relation to a method may similarly apply to a system. As another example, any aspect, element, or feature described herein in relation to a system may similarly apply to a method.

[0091] Next, examples of the disclosed technology shown in the attached drawings and disclosed herein will be described in detail. Wherever it is convenient, the same reference numerals will be used throughout the drawings to reference identical or similar parts.

[0092] Next, refer to the drawings. Here, similar numbers represent similar elements, and examples of the present disclosure are provided. As will be described in more detail, the present disclosure may include systems and methods for assessing cardiac health. To provide background to the systems described in the present disclosure, the components of a system for assessing cardiac health are shown in Figure 1A, which will be described first.

[0093] To facilitate understanding of the principles and features of this disclosure, various examples of the disclosed technology are described herein. The components, steps, and materials described herein as constituting various elements of the disclosed technology are illustrative and not intended to be limiting. Many suitable components, steps, and materials that would perform the same or similar functions as those described herein are intended to be included within the scope of this disclosure. Such other components, steps, and materials not described herein may include, but are not limited to, similar components or steps developed after the development of the embodiments disclosed herein.

[0094] As used herein, unless otherwise specified, the term “heart health” refers to the overall health of the heart and cardiovascular system.

[0095] As used herein, unless otherwise specified, the term "signal" refers to one or more signals.

[0096] As shown in Figures 1A and 1E, the disclosed technology includes a system 100 for assessing cardiac health. The system 100 may include a wearable device 110. The wearable device 110 may be a device worn on a person (e.g., a user 130). Alternatively or additionally, the wearable device 110 may be configured to be positioned in close proximity to the user's heart 120. For example, as shown in Figures 1A and 1E, the wearable device 110 may be worn on the user's chest.

[0097] The wearable device 110 may include one or more sensors. For example, the wearable device 110 may include a first sensor 112. Alternatively or additionally, the wearable device 110 may include a second sensor 114.

[0098] The first sensor 112 can be configured to measure at least one electrical characteristic of the heart. For example, the first sensor 112 can measure the electrocardiogram (ECG) signal of user 130. The first sensor 112 can include one or more electrodes that can be placed on the user 130's body. For example, one or more electrodes can be attached to user 130's skin. Alternatively or additionally, one or more electrodes attached to user 130's skin can further secure the wearable device 110 to the user.

[0099] The second sensor 114 may be configured to measure the user's cardiogenic oscillating. For example, the second sensor may be configured to measure the user's 130 psychocardiogram (SCG) signal. The second sensor 114 may be configured to measure a three-axis SCG signal. For example, the three-axis SCG signal may include the dorsoventral axis, the transverse axis, and / or the cephalofoot axis. Alternatively or additionally, the second sensor 114 may be configured to measure the user's gyrocardiogram signal.

[0100] System 100 may optionally or additionally include a catheter 140. The catheter 140 can be inserted into the user's body to measure hemodynamic parameters. For example, the catheter 140 can be inserted into a user 130 undergoing right heart catheterization (RHC) treatment.

[0101] As shown in Figures 1B and 1C, the wearable device 110 may include a first surface 150 and a second surface 160. For example, the wearable device may have an external structure including the first surface 150 and the second surface 160. The first surface 150 and the second surface 160 may be connectable and detachable structures. The first surface 150 and the second surface 160 are generally circular and may be connectable to generally produce a flat disc shape. Alternatively or additionally, the wearable device 110 may include electronic equipment 170 for performing various operations of the wearable device 110. For example, the electronic equipment 170 may be located inside the wearable device 110 between the first surface 150 and the second surface 160.

[0102] The first surface 150 can be configured to face away from the user's body (for example, to be positioned distal to the heart 120). The first surface 150 can include an alignment marker 152. For example, the alignment marker 152 may be an arrow indicating the direction in which the wearable device 110 should be pointed when the user puts it on (for example, the arrow must be pointed towards the user's head).

[0103] The second surface 160 can be configured to face the user's body (for example, to be positioned proximal to the heart 120). The second surface 160 can include a connector 162. The connector 162 can be connected to the first sensor 112. For example, the first sensor 112 can include one or more electrodes that can be attached to the user's body, and the connector 162 can connect the second surface 160 to one or more electrodes. In this way, one or more electrodes of the first sensor 112 can be attached to the user 130, and the other parts of the wearable device 110 (for example, the second sensor 114, the first surface 150, the second surface 160, and the electronics 170) can be connected to one or more electrodes by the connector 162, thereby securing the wearable device 110 to the user. Connector 162 may be any technically known connector, including but not limited to buttons, snap buttons, push buttons, adhesives, hooks and loops, etc., or combinations thereof.

[0104] The electronic device 170 may include electronic components of system 100. The electronic device 170 may include a processor and memory. For example, the electronic device 170 may include a CPU, a microprocessor, and the like. The memory may include logical instructions that, when implemented by the processor, cause the processor to perform one or more of the functions disclosed herein. Alternatively or additionally, the memory may include a removable memory card. For example, a microsecure digital (microSD) card capable of storing data collected (e.g., from one or more sensors). Alternatively or additionally, the electronic device 170 may include a transceiver. For example, the transceiver may receive data from one or more sensors (e.g., a first sensor 112, a second sensor 114) and transmit the data to a remote device. Alternatively or additionally, the electronic device may include a plug 172. The plug 172 may be configured to supply power to the wearable device 110. For example, plug 172 can be connected to a power source to directly power the wearable device 110 and / or charge the battery of the wearable device 110. Alternatively or additionally, plug 172 can be configured to connect to an external device (e.g., to transmit and receive data). For example, the plug allows the wearable device to connect to an external computer, tablet, mobile phone, other processor, and similar device to transmit and receive data. Plug 172 may be a USB connector. The electronic device 170 may include a power source. For example, the power source may be a battery to power the components of the wearable device (e.g., the first sensor 112, the second sensor 114, the processor, the transceiver). Alternatively or additionally, the electronic device may include one or more further sensors (in addition to the first sensor 112 and the second sensor 114). For example, the electronic device may include a third sensor.

[0105] The third sensor may be configured to measure environmental parameters. For example, the third sensor may be configured to measure one or more of the following, or a combination thereof: temperature, humidity, altitude, etc. Alternatively or additionally, the third sensor may be configured to measure the user's photoplethysmography signal.

[0106] The disclosed technology includes methods for assessing cardiac health, such as method 1100 shown in Figure 11. Method 1100 and / or other methods described herein can be implemented by a controller or computer. For example, method 1100 can be implemented by a wearable device 110 including a controller or computer. Alternatively or additionally, method 1100 can be implemented by a remote controller or computer. For example, the wearable device 110 can be connected to a remote controller or computer (e.g., via a transceiver, via a memory card, via plug 172 and wires).

[0107] Method 1100 may include step 1102 of receiving data from a first sensor. The data from the first sensor may be data relating to at least one electrical characteristic of the user's heart. For example, the first sensor may be capable of measuring the user's electrocardiogram signal.

[0108] Method 1100 may include step 1104 of receiving data from a second sensor. The data from the first sensor may be data related to the user's cardiogenic oscillations. For example, the second sensor may measure the user's psychocardiogram signal. Alternatively or additionally, the second sensor may measure the user's gyrocardiogram signal.

[0109] Method 1100 may include step 1106 of determining cardiac filling characteristics based on data from first and second sensors. For example, cardiac filling characteristics may be at least partially based on one or more axes of the psychocardiogram signal (e.g., transverse axis, cephalofoot axis, dorsoventral axis). Alternatively or additionally, filling characteristics may be at least partially based on the psychocardiogram signal during the diastolic phase of the heartbeat.

[0110] Method 1100 may include step 1108, which assesses cardiac health based on cardiac filling characteristics. For example, the assessment of cardiac health may include classifying the user's clinical state of heart failure. Alternatively or additionally, the assessment of cardiac health may include data showing changes in the user's cardiac filling characteristics. For example, it may include data showing changes in hemodynamics, filling pressure, pulmonary artery pressure, pulmonary capillary wedge pressure, etc., or a combination thereof.

[0111] Method 1100 may include step 1110, which outputs an assessment of cardiac health to the user. For example, the assessment of cardiac health may be transmitted to a connected device (e.g., a smartphone, tablet, or computer). Alternatively or additionally, the assessment of cardiac health may be displayed on a cardiac health assessment device (e.g., a wearable device). The assessment of cardiac health may include alerts to the user. For example, Method 1100 may be performed repeatedly and in real time, and the assessment of cardiac health may include alerts when the assessment of cardiac health changes. Real-time observation and alerting enable treatment of patients with heart failure by the user and / or clinicians. For example, the user and / or clinicians can interrupt the body's compensatory loops (e.g., by titration of drugs).

[0112] The disclosed technology includes a method 1200 for assessing cardiac health, as shown in Figure 12. Method 1200 and / or other methods described herein can be implemented by a controller or computer. For example, method 1200 can be implemented by a wearable device 110 including a controller or computer. Alternatively or additionally, method 1200 can be implemented by a remote controller or computer. For example, the wearable device 110 can be connected to a remote controller or computer (e.g., via a transceiver, via a memory card, via plug 172 and wires).

[0113] Method 1200 may include a step 1202 for determining a reference value. For example, step 1202 for determining a reference value may include calibrating system 100 to a reference value. The reference value may be a reference filling pressure. The reference value may be determined based on population-level reference values. For example, a reference pulmonary artery pressure value can be tracked using a population-level regression algorithm with features from non-invasive SCG and ECG signals. Alternatively or additionally, the reference value may be determined based on an individual's reference value. An individual's reference value can be determined based on data collected from RHC treatment. For example, an individual's reference value may be determined at least in part on pressure values ​​obtained from an RHC device. Alternatively or additionally, an individual's reference value may be determined based on data from a clinician. For example, an individual's reference value may be determined at least in part on pressure values ​​obtained in a clinical trial.

[0114] Method 1200 may include step 1204 of receiving data from a first sensor. The data from the first sensor may be data relating to at least one electrical characteristic of the user's heart. For example, the first sensor may be capable of measuring the user's electrocardiogram signal.

[0115] Method 1200 may include step 1206 of receiving data from a second sensor. The data from the first sensor may be data related to the user's cardiogenic oscillations. For example, the second sensor may measure the user's psychocardiogram signal. Alternatively or additionally, the second sensor may measure the user's gyrocardiogram signal.

[0116] Method 1200 may include step 1208 of determining cardiac filling characteristics based on data from first and second sensors. For example, cardiac filling characteristics may be at least partially based on one or more axes of the psychocardiogram signal (e.g., transverse axis, cephalofoot axis, dorsoventral axis). Alternatively or additionally, filling characteristics may be at least partially based on the psychocardiogram signal during the diastolic phase of the heartbeat. In addition, filling characteristics may be at least partially based on a determined baseline value.

[0117] Method 1200 may include step 1210, which assesses cardiac health based on cardiac filling characteristics. In addition, the assessment of cardiac health may be at least partially based on determined reference values. The assessment of cardiac health may include a classification of the user's clinical state of heart failure. Alternatively or additionally, the assessment of cardiac health may include data showing changes in the user's cardiac filling characteristics. For example, it may include data showing changes in hemodynamics, filling pressure, pulmonary artery pressure, pulmonary capillary wedge pressure, etc., or a combination thereof.

[0118] Method 1200 may include step 1212 of outputting an assessment of cardiac health to the user. For example, the assessment of cardiac health may be transmitted to a connected device (e.g., a smartphone, tablet, or computer). Alternatively or additionally, the assessment of cardiac health may be displayed on a cardiac health assessment device (e.g., a wearable device). The assessment of cardiac health may include alerts to the user. For example, Method 1200 may be performed repeatedly and in real time, and the assessment of cardiac health may include alerts when the assessment of cardiac health changes. Real-time observation and alerting enable the user and / or clinician to treat the patient with heart failure. For example, the user and / or clinician can interrupt the body's compensatory loop (e.g., by titration of a drug).

[0119] Methods 1100 and 1200 may further include the step of receiving data from one or more additional sensors. For example, Methods 1100 and 1200 may include receiving data from a third sensor. The third sensor may be configured to measure environmental parameters. For example, the third sensor may be configured to measure one or more of the following, or a combination thereof: temperature, humidity, altitude, etc. Alternatively or additionally, the third sensor may be configured to measure the user's photoplethysmography signal. Methods 1100 and 1200 may also include receiving data from a third sensor and a fourth sensor. For example, the third sensor may be configured to measure environmental parameters and the fourth sensor may be configured to measure the user's photoplethysmography signal.

[0120] The following embodiments further illustrate aspects of this disclosure. However, these are not limited to the teachings or disclosures of this disclosure described herein. (Examples) [Examples]

[0121] The following example demonstrates how signal processing and machine learning algorithms are used to extract relevant information from non-invasive cardiovascular electromechanical signals measured by our proprietary wearable patch hardware to estimate intracardiac and pulmonary artery pressure.

[0122] This embodiment includes a wearable device that receives signals representing biological functions (e.g., cardiac filling pressure), and a method for processing the information contained in these signals to estimate, for example, filling pressure from ECG and SCG signals.

[0123] In this example, we developed signal processing and machine learning algorithms to extract relevant information from non-invasive cardiovascular electromechanical signals measured by proprietary wearable patch hardware and estimate intracardiac and pulmonary artery pressure. Simultaneous single-lead electrocardiogram (ECG) and triaxial psychocardiogram (SCG) signals were collected from HF patients undergoing right heart catheterization (RHC), a representative clinical diagnostic method for measuring intracardiac and pulmonary artery pressure, using a custom wearable patch. Reference pressure values ​​(obtained from the RHC device) were estimated using features from the simultaneously recorded SCG and ECG signals, and changes in these pressures in some subjects / patients who underwent pharmacological stress during RHC treatment.

[0124] Using population-level regression algorithms, features from non-invasive SCG and ECG signals can be used to track baseline pulmonary artery pressure values ​​and pressure changes. A discreet, wearable patch-based device that tracks cardiac filling pressure by measuring pulmonary artery pressure, along with corresponding signal processing and machine learning algorithms, enables remote home monitoring for patients with HF and other cardiovascular diseases. Remote monitoring of HF patients allows for tailoring of home care, ultimately leading to shorter hospital stays and improved quality of life for affected patients.

[0125] The wearable patch hardware possesses a unique combination of electrocardiogram (ECG), psychocardiogram (SCG), and environmental sensing capabilities. This enables context-aware determination of hemodynamic parameters in unsupervised situations. Compared to other technologies that measure only electrophysiology, or other wearable systems that can also measure psychocardiogram signals, measuring environmental parameters such as altitude, humidity, and temperature allows for an understanding of estimated hemodynamic parameters, such as pulmonary capillary wedge pressure (PCWP) or pulmonary artery pressure (PAP), within the context of activity and / or the environment. For example, vasodilation within the heart leads to a relative decrease in blood volume in the cardiovascular system, thereby reducing preload (decreasing PCWP and PAP). In other words, by adding these environmental features along with SCG and ECG features to machine learning-based models, the estimation of hemodynamic parameters becomes more accurate.

[0126] Furthermore, in part of our research, we focus on estimating PAP and PCWP changes caused by perturbations. Pharmacological perturbations (such as vasodilator delivery) are used as an example, but such perturbations may also be caused by exercise, changes in ambient temperature (e.g., heat-induced vasodilation), changes in posture (e.g., changes in preload due to standing up from a seated position), or other modulation techniques based on electrical stimulation of peripheral nerves or other changes in the baseline physiological state. Such techniques may be used, for example, to calibrate hemodynamic parameters. Measuring signals before and after such perturbations is possible, rather than only measuring signals at rest, and the direction or magnitude of the physiological changes caused by the perturbations may be incorporated into algorithms to improve the accuracy of the estimated PAP, PCWP, and other intracardiac hemodynamic parameters.

[0127] We developed a signal processing and machine learning algorithm to track cardiac filling pressure (preload) by extracting relevant features from wearable ECG and SCG signals. Data (single-lead ECG and 3-axis SCG) were collected using a custom wearable chest patch. Tracking cardiac filling pressure and taking proactive measures (drug titration, continuous outpatient visits, etc.) has proven effective in shortening hospital stays for HF-related complications and for all causes in HF patients. However, the high cost of devices used in clinics and home settings to track cardiac filling pressure makes them unsuitable for the large HF population. Our developed wearable sensor and corresponding algorithm offer a low-cost alternative to expensive hemodynamic monitoring systems.

[0128] We have demonstrated for the first time that cardiac filling pressure can be tracked using SCG and ECG signals. We were able to estimate baseline pressure values ​​in different ventricular and pulmonary cavities (pulmonary artery, pulmonary capillaries, right atrium, and right ventricle). We also developed an algorithm to track changes in these pressures due to pharmacological perturbations (e.g., infusion of vasodilators). To estimate baseline pressure, patients were asked to remain as still as possible in a supine position for at least 10 minutes. During this period of recording baseline values, patients may perform unconscious movements that distort the SCG signal. We developed an algorithm to detect and reject SCG signals degraded by movement. Signal quality indexing was performed to stratify high-quality SCG heart rates from patients. From high-quality SCG heart rates, relevant features from each axis of the SCG, including the dorsal-ventral axis, transverse axis, and cephalopodal axis, and the magnitude of all three axes of the SCG signal were extracted. Population regression models were trained on the top of these features to predict mean pressure values. The models were validated using a one-miss cross-validation method.

[0129] To track pressure changes, we developed a dynamic time-stretching (DTW) based distance metric that estimates changes in SCG from a baseline state to a vasodilator infusion state, and uses these DTW distances from different parts of the SCG signal (systolic and diastolic portions) and different axes (cephalofoot axis, transverse axis, dorsoventricular axis, and overall scale) to estimate changes in pressure (pulmonary artery, pulmonary capillaries, right atrium, and right ventricle). We developed an algorithm to remove outliers and low-quality signals (due to motion artifacts and other reasons) from the heart rate signal using dimensionality reduction methods (principal component analysis) and Gaussian mixture models. We developed global regression models (e.g., Ridge, Lasso, Random Forest, etc.) to estimate changes in pressure values ​​from the corresponding DTW distances from the SCG signal, and validated our models using a single-point cross-validation method.

[0130] Most studies using SCG focus on the contractile portion and the dorsal-ventral axis of the signal. However, our study demonstrates that features from the dilating portion of the signal, as well as the transverse axis and overall magnitude of the signal, provide significant / important information regarding changes in intracardiac and pulmonary artery pressure, in addition to the dorsal-ventral systolic direction of the SCG signal.

[0131] The algorithm may include other metrics for estimating changes in the SCG signal from one state to another due to fundamental changes in hemodynamics, namely Euclidean distance, Mahalanobis distance, clustering-based unsupervised learning approaches, deep learning models, etc. Different other methods, namely manifold mapping, clustering, etc., can be used to improve signal quality and have reliable feature extraction. The algorithm may also include fluctuations in the SCG signal to track changes in hemodynamics during activities in daily life. This can provide information about exercise intolerance, which can be used to stratify the risk associated with HF. In this disclosure, changes in pressure on the right side of the heart (right atrium, right ventricle), pulmonary artery, and pulmonary capillaries were estimated. The algorithm can be translated into changes in pressure on the left side of the heart (left atrium, left ventricle) and aorta.

[0132] This embodiment provides a wearable, inexpensive, and virtually invisible system for remotely monitoring HF patients using a wearable psychocardiogram (SCG) monitor. The SCG signal captures chest wall vibrations in response to blood movement and cardiac motion during ventricular ejection. We studied the use of the SCG signal to classify the clinical state and SCG feature ability of HF patients in a baseline state and to accurately estimate filling pressure. Specifically, in this study, we demonstrated accurate classification of the clinical state of HF patients in a resting state with an AUC of 0.8. Furthermore, we disclose a machine learning-based regression model for measuring pulmonary capillary wedge pressure (PCWP), i.e., indirectly measuring left atrial pressure similar to PA pressure. The ability to measure PCWP provides caregivers with important information, similar to implantable devices that measure PA pressure. All data in this study were collected in the cardiac catheterization laboratory to obtain both SCG signals and hemodynamic parameters while the patient was undergoing right heart catheterization (RHC). (Subject demographics, experimental protocol, and sensing hardware) A. Data acquisition and experimental protocol

[0133] The study included a total of 50 subjects diagnosed with heart failure. Exclusion criteria included patients with implanted ventricular assist devices and patients who had previously undergone a heart transplant. The demographics of the study population are shown in Table I. Each subject submitted written informed consent prior to data collection. [Table 1]

[0134] The objective of the protocol was to investigate the identification features of SCG in relation to the clinical state (i.e., decompensated and compensated states) of HF patients in a baseline state (i.e., at rest), and the correlation between SCG and hemodynamic parameters. RHC was performed using the Mac-Lab hemodynamic recording system to determine the clinical state and obtain hemodynamic parameters. Based on RHC treatment, the following hemodynamic parameters were measured: cardiac index (CI), pulmonary capillary wedge pressure (PCWP), pulmonary artery pressure (PAP), right atrial pressure (RAP), and right ventricular pressure (RVP). Based on the hemodynamic parameters, the clinical state was determined as follows: the patient's mean PCWP was ≥20 mmHg and the cardiac index (CI) was 2.2 L / min / m². 2 The patient is considered decompensated if the following conditions are met. Otherwise, the patient is considered compensated. In some rare cases, if either the PCWP or CI value is significantly high / low, the caregiver's judgment takes precedence over this rule. For example, a borderline PCWP of 16 mmHg and an extremely low CI of 1.3 L / min / m². 2 In that case, the patient was considered decompensated.

[0135] All SCG data were collected while the patient was receiving RHC treatment. First, the patient was positioned supine on an examination table. A wearable patch hardware was attached to the patient's central sternum to acquire the SCG signal. During this procedure, the patient was instructed to remain as still as possible. Subsequently, a catheter was inserted into the patient's body to measure hemodynamic parameters. The patient's supine and immobile state is referred to below as the baseline state. All signals collected during a single RHC treatment are referred to as a recording, which consists of ECG and SCG signals from the wearable patch and pressure waveforms from the Mac-Lab hemodynamic recording system. Furthermore, compensatory recordings are defined as recordings obtained from patients diagnosed with compensatory disease, and decompensated recordings are defined similarly. Table I shows the number of compensatory and decompensated recordings in the collected datasets. Note that since one patient received RHC treatment twice, the number of recordings is one more than the number of patients. The data collection setup, along with exemplary signal extraction, is shown in Figure 1E. Figure 1E provides an illustration of the data acquisition setup. The heart failure patient is in a supine position while undergoing right heart catheterization (RHC) treatment. During the treatment, a wearable patch is attached to the patient's central sternum to collect ECG and SCG signals. On the right side, an example of a representative set of ECG and SCG signals collected from the wearable SCG patch is shown, along with the RHC waveform collected from the Mac-Lab hemodynamic recording system. The waveforms are unfiltered and in their raw form. B. Sensing Hardware

[0136] The wearable patch samples the ECG signal at 1 kHz, the accelerometer signal at 500 Hz, and the ambient signal at 20 Hz, and saves this data to a micro-secure digital (microSD) card within the patch. A custom graphical user interface transfers all the data from the microSD card to a computer, where the accelerometer and ambient signals are sampled again at 1 kHz, ensuring all signals share the same sampling rate for easier processing. Subsequently, for further analysis, our processing algorithm decimates all signals to 500 Hz. (Signal processing and machine learning methods) A. Signal preprocessing

[0137] The first preprocessing step is to remove the first 5 minutes from the start and the last 5 minutes from the end of each recording. This is because, during recording, the sensor may start recording before the device is attached to the subject, and the recording may end after the device is removed. As a second step, the signals are subjected to band-pass filtering with the following digital filter specifications: finite impulse response (FIR) filtering with passbands of 4–25 Hz and 1–40 Hz for the ECG and SCG signals, respectively, and a stopband attenuation of 80 dB. The resulting equiripple FIR filters have orders of 332 and 1407 for the ECG and SCG signals, respectively. The passband for the ECG signal was chosen to isolate the R peak and facilitate its detection in the next processing step. For the SCG signal, the passband was chosen to suppress out-of-band noise and maintain the SCG signal characteristics.

[0138] After filtering both the ECG and SCG signals, a further channel of the SCG, called the magnitude channel, was formed.

number

[0139] As the final preprocessing step, R peaks were detected in the ECG signal, followed by pulse segmentation in the SCG signal. Two algorithms were used for R peak detection: Pan-Tompkins, performed by Physionet, and Phasor Transform, implemented using a custom method. To reduce false positives in R peak detection, only R peaks detected by both algorithms were selected. Twelve second windows were used for detection to identify R peaks.

[0140] Using the R peak, pulse segmentation of the SCG signal was performed as follows: 200 ms before the R peak and 700 ms after the R peak were defined as the start and end of the pulse, respectively. As a result, an SCG pulse array was constructed for each channel of the SCG. Unlike other conventional studies, typical pulse segmentation was performed from the R peak (i.e., 0 ms before / after the R peak) to approximately 700 ms after the R peak. This study intentionally included ventricular diastolic timing because the features observed at this time were predicted to be highly relevant to estimating filling pressure. B. Motion artifact rejection and signal quality indexing

[0141] SCG signals are susceptible to motion artifacts; that is, when a subject moves, SCG oscillations are degraded by higher amplitude motion artifacts. In the collected dataset, patients were instructed to remain as still as possible, but motion artifacts were still present in the recordings. Figure 3 shows an example of such artifacts in a single recording. Figure 3 provides an illustration of motion artifact rejection for the dorsal-ventral channel of the SCG. Here, Figure 3a) shows a complete recording obtained from one representative subject, Figure 3b) is a magnified view of the visualization of the recording of the segment degraded by motion artifacts in this recording, Figure 3c) shows the two most similar consecutive pulses in the recording, and Figure 3d) shows a pulse corrupted by motion artifacts in the complete recording, where pulse 310 indicates a sample where the magnitude channel exceeded the threshold.

[0142] By utilizing the observation that the amplitude of motion artifacts is higher than that of SCG oscillations, we devised a simple algorithm to detect SCG pulses damaged by motion and then discard those degraded by motion. This algorithm takes a segmented pulse array as input and outputs an index of motion-degraded pulses. Pseudocode is shown in Algorithm 1. The detection clue is to look for the two most similar consecutive pulses. If two consecutive pulses are not degraded by motion, they should have similar morphologies, as it is unlikely that a large change in SCG morphology would occur in such a short time. Therefore, the two most similar consecutive pulses should not have motion artifacts (see Figure 3c), which shows an example of the two most similar consecutive pulses in this recording). Motion artifacts are detected as outlined in Algorithm 1 by calculating a simple threshold using the portion of the recording that does not have motion artifacts. The results of the motion artifact detection algorithm are shown in Figure 3. [Table 2]

[0143] The following formula was used to measure the similarity between two pulses.

number

[0144] After detecting and rejecting pulses degraded by motion artifacts, Signal Quality Indexing (SQI) was applied separately to each channel of the SCG to extract high-quality SCG pulses. This differed from methods where SQI was already applied in the following ways: 1) there was no set of templates, but rather one template per channel for recording, and 2) an ensemble average of pulses using Woody's method was used as the template. This change allowed the template to be fitted to specific recordings without motion artifacts. The quality score for each pulse was returned as the output of the SQI. Feature extraction was performed using the top 5% of pulses. C. Feature Extraction

[0145] The output from SQI was used in two ways: 1) the ensemble mean of pulses was calculated, and 2) statistical features derived from individual pulses were used. Basic features were extracted from a single pulse and listed in Table II. When extracting the features listed in Table II using the ensemble mean pulse, the obtained features were largely related to the baseline SCG value. Subsequently, baseline features output from SQI were extracted from each pulse. Then, the following statistical features were calculated: mean, median, standard deviation, maximum, and minimum values. This method was used to obtain the change from pulse to pulse. The time window in the table was selected to reflect the ventricular diastolic (-200ms to 0ms and 300ms to 600ms) and systolic (-50ms to 250ms) phases of the cardiac cycle. [Table 3]

[0146] To obtain possible relationships between expansion and contraction intervals, we combined features from both windows to form new features. Specifically, for all power and amplitude features, we calculated the ratio between the contraction interval (-50ms to 250ms) and the expansion interval (300ms to 600ms). In other words, we calculated the absolute difference between the contraction delay and the expansion delay for all delay-related features. We also aimed to obtain features with cross-channel information. For this purpose, for all channel pairs (cephalofoot, dorsoventral, and transverse channels), we calculated the ratio (amplitude and power features) and absolute difference (delay features) of the basic features only when they span the contraction and expansion time windows, and formed new cross-channel features. Finally, a total of 690 features were extracted from the recording.

[0147] Figure 2 provides an example of processing wearable patch data. Figure 2a is a block diagram of this processing. Figure 2b is a brief mathematical explanation of the steps used in the processing. Figure 2c provides a visualization of the processing from a 10-second window of data from a representative participant. For ease of visualization, the visualization shows only the dorsal-ventral channels of the SCG; that is, the same processing is performed on other SCG channels. C. Classifier and Regression Algorithm Design

[0148] In this study, we used Support Vector Machines (SVMs) for both classification and regression tasks. SVMs search for a separating hyperplane to distinguish between two classes. Through kernels, SVMs can model nonlinear relationships in the data. In this study, we considered polynomial and radial basis function (rbf) kernels. In addition, for regression tasks, we considered Lasso linear regression. E. Experimental Design

[0149] The primary goal of the experiment was to analyze the function of SCG features to identify HF clinical states. For this purpose, we treated the task as a binary classification, where the input was an SCG feature and the output was whether the patient was in a decompensated or compensated state. To evaluate the classification model, we performed one-miss cross-validation (LOSOCV) to calculate classification accuracy and the area under the concentration curve (AUC) of the receiver operating characteristic (ROC) curve. After evaluating the model using all extracted features, we performed successive forward feature selection (SFFS). SFFS helps address the curse of dimensionality because there are more features (690) than data points (51). Furthermore, SFFS is more beneficial because features are more important in classification than in other methods.

[0150] In the two previous experiments, each recording differed in length from 25 to 200 minutes and was treated as a single data point for classification. To ensure accurate classification of recording time, each recording was divided into non-overlapping windows of 1, 5, 10, 15, 20, and 25 minutes. Within each window, features were extracted using the same method outlined in those examples, because this feature extraction is invariant to the recording length. Using the best-performing features and hyperparameters from the two previous experiments, LOSOCV was again used to evaluate performance in the different windows.

[0151] In the final experiment, the mean PCWP from SCG features was estimated using a regression algorithm. Here, the input included features extracted from the recordings, and the output was the mean PCWP value from RHC. Again, the regression model was evaluated using LOSOCV. The root mean square error (rMSE) between the regression output and the ground truth value measured during RHC treatment was used as the performance metric. SFFS was performed, as in the classification experiment. (Experimental results) A. Classification experiment

[0152] The results of the classification experiments are shown in Table III and Figure 4. Figure 4 shows the receiver operating characteristic (ROC) curves for different classifiers trained and tested (using the best features 410; using all features 420; using three features 430; using four features 440; using five features 450; and using random chance 460). The 420 curve is the ROC curve for the RBF kernel SVM used in successive forward feature selection. These results show that using fewer features improved performance as expected due to the curse of dimensionality. [Table 4]

[0153] The experimental results of windowing are shown in Figure 5. Figure 5 shows the best classifier performance at different window lengths, with AUC 510 and precision 520. These results show a decrease in classifier performance with short window lengths up to 10 minutes. After 10 minutes of recording, performance levels off. B. Regression experiment

[0154] The results of the regression experiment are shown in Table IV. Similar to the classification experiment, the model trained using SFFS outperforms the other models. [Table 5] (Consideration)

[0155] The results in Figure 4 and Table III demonstrate that accurate classification of HF patients is achievable using SCG features. Furthermore, performance improves by using only a fraction of the SCG features. This secondary result is predictable because there are more features than data points. Currently, the clinical status of physiologically decompensated HF patients requires expensive and invasive catheterization procedures. This wearable device can accurately classify clinical status and can be used as a pre-screening tool to reduce the number of RHC treatments. Reducing the number of RHC treatments can lower the cost of HF care and improve quality of life.

[0156] The results in Figure 5 show that performance decreases with shorter recording times. Performance improves with longer recording times until it plateaus. Based on these results, at least 10 minutes of SCG data is required to obtain meaningful performance. This information is important for determining the appropriate time to implant the necessary devices in the patient before assessing their clinical condition.

[0157] The regression experiment results are promising for the SCG signal. The ability to predict mean PCWP from the SCG signal could significantly impact home-based HF care. Measuring PCWP daily or more frequently as needed with this wearable device would enable pressure-inducing therapy, similar to the techniques used with implantable pressure monitors. Importantly, by providing this parameter rather than a black-box output that makes a judgment, the information can be shared with physicians, allowing them to tailor care as desired based on existing flowcharts and guidelines.

[0158] Furthermore, the features selected in the experiment provide important scientific insights into SCG signals, which are not as well understood as ECG signals. We observed that the top five features among the best-performing classifiers originated from both dorsoventral (1) and transverse (4) channels. This demonstrates the importance of analyzing all axes of SCG signal data rather than just dorsoventral channels. Moreover, for the regression task, the top two features also originated from the transverse channels of the SCG.

[0159] These examples further include low-cost systems that can track changes in hemodynamic congestion and have the potential to help many people suffering from HF. Elevated intracardiac filling pressure is an early and active sign of the onset of congestion in heart failure (HF). Hemodynamic changes indicate that the progression of chronic compensated HF occurs several weeks earlier than that of acute decompensated HF (ADHF). Recent studies have also shown that the generation of slight changes in pulmonary artery pressure over a long period is closely related to the transition to ADHF. Therefore, tracking hemodynamics using an implantable hemodynamic monitoring system and subsequent proactive HF management treatments (e.g., drug titration, early follow-up clinic visits, etc.) to reduce asymptomatic congestion has proven effective in reducing HF-related readmissions. Compared to hemodynamically induced HF management, conventional HF management treatments, including tracking daily weight, telescopic vital signs, and clinical symptoms to detect ADHF, have not shown effectiveness in reducing HF-related readmissions in randomized controlled trials. This is because these changes progress relatively late from compensatory HF to ADHF.

[0160] Cardiovascular health can be observed using psychocardiograms (SCGs), local mechanical vibrations of the chest wall associated with cardiac movement, and blood in vascular structures. By using SCG timing to estimate the cardiac prodromal phase due to exercise and physiological perturbations, changes in myocardial contractility can be calculated. Importantly, SCGs can be used to assess the clinical status of patients with decompensated HF. In addition to assessing the clinical status of HF patients, SCGs have shown effectiveness in tracking instantaneous oxygen consumption during cardiopulmonary exercise testing in HF patients and uncontrolled daily living activities in healthy individuals. Based on these SCG tracking results in both healthy individuals and HF patients, changes in hemodynamic congestion can be tracked, and SCG signals can be recorded by simultaneously estimating changes in PAP and pulmonary capillary wedge pressure (PCWP). (method)

[0161] SCG and ECG signals from HF patients were recorded using a custom wearable patch during right heart catheterization (RHC), a typical method for measuring hemodynamic congestion using PAP and PCWP. During RHC treatment, PAP and PCWP were modulated with intravenously administered systemic vasodilators, and changes in mean pressure values ​​were estimated by tracking changes in the simultaneously recorded SCG signals. Various segments of the SCG signal were analyzed to understand key segments that provide significant information regarding changes in PAP and PCWP. By tracking rapid changes in hemodynamic congestion using SCG, this novel wearable technology—a discreet and low-cost alternative to current observation systems—can be demonstrated for long-term monitoring of intracardiac filling pressure in remote HF management, and ultimately, potentially shorten HF-related readmission periods. A. Experimental Protocol

[0162] RHC treatment was administered to a total of 20 HF patients (8 inpatients and 12 outpatients) referenced for hemodynamic evaluation of HF. Patients with cardiac shock were excluded. The dataset was randomly divided into two groups: a group of 15 HF patients for the training test set and a group of 5 HF patients for a separate, proprietary validation set (details are described in the "Supplementary Materials" section below).

[0163] Figures 1A–1C show the experimental setups and placements of different sensors for each patient. Figure 1A provides an experimental setup of a wearable patch placed on a subject undergoing right heart catheterization (RHC) treatment. Here, the axis (upper right) represents the axis of the psychocardiogram (SCG) signal. Figure 1B shows the top, bottom, and inside of the wearable patch. Figure 1C shows the front (left) and side (right) of the wearable patch placed on a representative subject. Before initiating RHC treatment, the custom wearable patch was placed directly below the jutting incision, and the cath lab recording system was time-synchronized with the wearable patch.

[0164] RHC treatment was performed in a quiet, environmentally controlled cardiac catheterization laboratory with an ambient temperature of 25°C or lower. The right neck or right elbow fossa region was disinfected and prepared using a chlorhexidine cotton ball with a standard sterilization method. Local anesthesia was administered with 2% lidocaine. Under ultrasound guidance, venous access was made, and a 5-French (F) introducer sheath was placed in the right internal jugular or right brachial vein. After resting in the supine position for at least 20 minutes, a 6F pulmonary artery wedge catheter with a balloon at its tip was advanced into the right atrium, right ventricle, pulmonary artery, and pulmonary capillary wedge positions, guided by fluoroscopy. At each position, pressure was measured for 60 seconds while slowly exhaling and holding the breath (end-expiratory), and this was repeated three times according to the standard right heart catheterization protocol, with the average taken. Cardiac output was obtained using Fick's principle and thermodilution. After measuring baseline hemodynamics and cardiac output, and after a 10-minute rest in the supine position, medications were administered at the discretion of the heart failure specialist handling this case. Nitroglycerin was administered as a sublingual spray (400 or 800 mcg), and nitroprusside was started as an intravenous (IV) infusion at a rate of 0.3 mcg / kg / min (and titrated at 0.3 mcg / kg / min every 5 minutes until a hemodynamic effect was achieved). Hemodynamic monitoring was repeated according to the standard protocol until the peak hemodynamic effect was determined by the heart failure specialist. Subsequently, the pulmonary artery wedge catheter with a balloon tip and the venous sheath were removed.

[0165] Wearable ECG and SCG signals were continuously recorded throughout RHC treatment, and specific portions of the wearable signals were extracted in subsequent analysis using timestamps from both the RHC and wearable systems. Changes in PAP and PCWP were estimated from changes in the wearable signals. Figure 1D shows the wearable signals with corresponding PAP signals from the RHC computer during baseline RHC recordings from representative subjects. Figure 1D shows representative cardiogenic signals: electrocardiogram (ECG), triaxial SCG (cephalofoot axis [HtoF], transverse axis [Lat], and dorsal-ventral axis [DV]), and RHC pulmonary artery pressure (PAP) signals. SCG is a mechanical signal associated with myocardial contraction, heart valve movement, and blood movement from the left ventricle to the aorta. B. Sensing Hardware

[0166] RHC pressure values ​​were extracted from the cath lab Mac-Lab system. Wearable ECG and 3-axis SCG (axes: cephalopod axis (HtoF), dorsoventricular axis (DV), and transverse axis (Lat)) were collected using a custom wearable patch as shown in Figure 1B. The patch in this embodiment had a diameter of 7 cm and a weight of 39 gm. All wearable ECG and SCG signals were sampled at 1 kHz. Figure 1D shows representative ECG and 3-axis SCG signals from the wearable patch. C. Signal Processing and Feature Extraction

[0167] Figure 6 shows the signal processing and feature extraction used for wearable and pressure signals from the Cath Lab Mac Lab system. Figure 6 provides an overview of the method. Figure 6A shows that wearable ECG and SCG (only one-axis signals are shown for simplification) signals were synchronized with right heart catheter pressure (RHCP) signals. If the catheter was recording pulmonary artery (PA) pressure and pulmonary capillary wedge pressure (PCW) signals, 20 seconds of signals from both baseline values ​​(BL) during vasodilator infusion (VI) were extracted. Figure 6B shows that the R peak of the ECG signal was detected and later used to segment the corresponding SCG signal into individual heartbeats. Anomaly removal and noise reduction steps were performed on the SCG heartbeats, and features were extracted and used in a regression algorithm to estimate changes in RHC mean pressure (MP) values ​​(e.g., changes in pulmonary artery mean pressure [ΔPAM] and changes in pulmonary capillary wedge mean pressure [ΔPCWP]). MP BL and MP VI The values ​​were extracted from the RHC Mac-Lab computer and used to calculate the target variables (ΔPAM and ΔPCWP). Figure 6C shows the details of the wearable signal processing. Specifically, the R peak of the ECG signal was first detected, and the SCG signal was segmented into individual heartbeats. Next, the SCG BL and SCG VIThe heart rates were passed through an outlier removal algorithm (using principal component analysis [PCA] and a Gaussian mixed model [GMM]) to obtain an ensemble mean with two mean SCG heart rates for each axis (one axis for BL and one axis for VI). Thirdly, the dynamic time-stretching (DTW) distance between the BL heart rate and the VI heart rate for each axis was calculated, which is used as a feature (f) in the regression algorithm.

[0168] Before starting this procedure, both systems were time-synchronized. For both baseline values ​​(BL) and during vasodilator infusion (VI), the mean pulmonary pressure (PAM) and PCWP values ​​were extracted by a heart failure specialist (LK). These were then used to calculate the change in PAM (ΔPAM) and the change in PCWP (ΔPCWP), respectively, by subtracting the mean pressure value in BL from the mean pressure value in VI.

[0169] Due to a technical issue with the Mac-Lab system, one subject's PCWP VI Since the values ​​were not recorded, these values ​​are missing from this analysis. Overall, ΔPAM values ​​were obtained for 20 subjects (15 subjects for the training test set and 5 subjects for the validation set), and ΔPCWP values ​​were obtained for 19 subjects (14 subjects for the training test set and 5 subjects for the validation set).

[0170] When the catheter reached the wedge position in the pulmonary artery and pulmonary capillaries, wearable signals (ECG and SCG) of 20 seconds in length were extracted from the protocol's BL and VI states using synchronized timestamps. Changes in ΔPAM and ΔPCWP values ​​in the wearable signals were analyzed and later used in cross-validation in a population regression model. Details of wearable signal processing, feature extraction, and the regression model are described below.

[0171] Pretreatment and noise reduction: The BL and VI wearable signals were processed separately (filtered, outliers removed, and ensemble average calculated) and later used to calculate the dynamic time warping (DTW) distance between the two states (BL and VI). The DTW distance between different parts from different axes of the SCG signal was used in a regression algorithm to perform leave-one-subject-out (LOSO) cross-validation for the training test set and later verified in the independent validation set to estimate ΔPAM and ΔPCWP.

[0172] The raw ECG and SCG signals from the wearable patch were digitally filtered to calculate a fourth SCG signal representing the accelerometer magnitude (SCG Mag ) using the vector sum of the three already obtained SCG axes (SCG HtoF , SCG Lat , SCG DV ). After the filtering step, the R-peaks of the ECG signal were detected and the 20-second-long signals of the four-axis SCG were segmented into individualized heartbeats. The SCG heartbeats were clipped to a duration of 500 ms before and after the R-peak, which roughly indicates most of the relevant expansion and contraction cardiac events of the subject (e.g., rapid inflow, atrial systole, isovolumetric contraction, ventricular ejection, etc.).

[0173] Following the heartbeat segmentation of the wearable SCG signal, an automated unmonitored algorithm (details are described in the "Supplementary Materials" chapter below) was used to separately remove the outlier heartbeats from the SCG for each axis and each part (expansion and contraction) of the SCG signal for the two distributions from the two states (BL and VI). The actual SCG heartbeats corresponding to the outliers of the distribution were removed, and two separate distributions (SCG BL and SCG VI ) were obtained for each axis. The ensemble average of the remaining heartbeats was taken to generate two ensemble-averaged heartbeats for BL and VI for a specific axis and part. These were later used to calculate the DTW distance. The ensemble averaging step reduced the inherent variability and residual noise in the SCG heartbeats. Figure 7 shows the ensemble-averaged SCG from the BL and VI statesDV This figure shows the heart rate and the corresponding PAP and PCWP heart rates. Figure 8 shows the changes in pulmonary artery pressure (PAP) in Figure 7A, and PAP BL 710 and PAP VI It shows 720, and Figure 7B shows the change in pulmonary capillary wedge pressure (PCWP), PCWP BL 730 and PCWP VI It shows 740, and in Figure 7C, the dorsoventral psychocardiogram (SCG) DV ) provides changes in SCG in representative subjects who received intravenous infusion of vasodilators. BL 750 and SCG VI The value is 760, and the arrows indicate the change in each signal. Time "0" indicates the position of the corresponding ECG R peak.

[0174] Dynamic Time Warping and Feature Extraction: To calculate the change in SCG from BL to VI, DTW was used, and the DTW distances from different parts of the SCG heart rate to ΔPAM and ΔPCWP were compared with the correlation analysis shown in Figure 9. Figure 9 shows the correlation analysis of the target variable. For different DTW distances of the corresponding SCG signals for the training test set, Figure 9A shows ΔPAM and Figure 9B shows ΔPCWP, with the color bars indicating R 2The values ​​are shown, and the dotted line 910 indicates the boundary between ventricular diastole and systole (i.e., the R peak of the corresponding ECG). That is, total diastole (-500ms: R peak), early diastole (-500ms:-200ms), late diastole (-200ms: R peak), total systole (R peak: 500ms), early systole (25ms: 150ms), and late systole (200ms: 500ms). DTW is a time series analysis method that aligns signals to find similarities between them. The DTW distance between the signal from BL and the signal from VI was calculated from different parts of the SCG heartbeat, namely, total diastole (-500ms: R peak), early diastole (-500ms:-200ms), late diastole (-200ms: R peak), total systole (R peak: 500ms), early systole (25:150ms), and late systole (200:500ms). Here, negative times represent the period before the R peak, and positive times represent the period after the R peak. Early diastole corresponds to passive ventricular filling, late diastole corresponds to atrial systole, early systole corresponds to isovolumetric contraction (IVC), and late systole corresponds to the ventricular ejection phase of the cardiac cycle.

[0175] The preprocessing and feature extraction steps described above were performed identically for both the training test and validation datasets. After the feature extraction step, a regression algorithm using LOSO cross-validation was developed using only data from the training test set. In this step, the model's hyperparameters were tuned to maximize the performance of the developed model on the training test set (coefficient of determination R). 2 (We maximized the error and minimized the root mean square error RMSE). The resulting trained model was then validated against a proprietary validation set to demonstrate the generalizability of the developed model. This step is described in detail in the following chapters. D. Regression

[0176] Before developing a regression algorithm using data from the training test set, features from different parts of the SCG heart rate (i.e., DTW distance) were compared to the target variable using a simple correlation analysis (shown in Figure 9), and the coefficient of determination (R) between them was determined. 2We calculated the relevant SCG segments and tracked changes in PAM and PCWP. Later, we created and tuned a regression algorithm using DTW distance as a feature to estimate changes in PAM and PCWP relative to the training test set, and then validated it against our own validation set.

[0177] Following a simple correlation analysis, a population-level regression model using LOSO cross-validation was applied to the training and test sets to estimate ΔPAM and ΔPCWP from the DTW distance. To this end, different regression algorithms were investigated, and a linear kernel support vector regression (SVR) model was selected as the regression model for our initial analysis.

[0178] A simple correlation analysis between the DTW distance from different parts of the SCG heart rate and the corresponding target variables (ΔPAM and ΔPCWP) for the training test set revealed, as shown in Figure 9, that not all changes from different parts of the SCG (i.e., DTW distance) are related to changes in mean pressure (MP). Therefore, a feature selection technique was performed using the Sequential Forward Selection (SFS) algorithm, and the top five features were selected as estimates in the regression model.

[0179] A linear SVR regressor was trained on the DTW distance from 14 selected subjects (excluding one subject each time) using LOSO cross-validation on the 15 subjects in the training test set. The target variables (ΔPAM and ΔPCWP) were predicted for the excluded subjects, and this process was repeated 14 or more times by excluding a different subject each time. As a result, predictions were obtained for all 15 subjects in the training test set. Using this cross-validation method, a global regression model with optimized hyperparameters for the data only in the training test set was developed. To validate this global model, the regression model (with optimized hyperparameters) was trained on the entire training test set (data from 15 subjects) and tested on a separate validation set (data from 5 subjects). As a result, all target variables were predicted for all 20 subjects.

[0180] The regression model and method were evaluated using a diagram illustrating two advantages. First, the estimated target variable (ΔMP) was evaluated. Pred ) and the Mac-Lab system (ΔMP Act The RMSE between the ground truth target variable and the true value was calculated. Next, a simple correlation analysis (Pearson) was performed between the true value and the predicted ΔMP to obtain the statistical importance of the prediction, and the R between the true value and the predicted value was calculated. 2 The following was calculated. In this study, p-values ​​of 0.05 or less were considered statistically significant. RMSE and R 2 Both were calculated separately for the training test set and the validation set. (result)

[0181] Patient demographics and clinical characteristics are shown in detail in Table V, and RHC characteristics are shown in Table VI. [Table 6] [Table 7]

[0182] The subjects were 15 HFrEF patients and 5 HFpEF patients. Four of the patients were female, and the mean age was 54.6 ± 13.2 years. The mean weight was 94.2 ± 19.4 kg, the mean height was 175.3 ± 9.9 cm, and the mean ejection fraction (EF) was 32 ± 19.5.

[0183] Intravenous administration of vasodilators reduced PAM and PCWP by 8.9±5.5 mmHg and 8±6.3 mmHg, respectively, from their corresponding BL values. Figure 7 shows the changes in PAP and PCWP signals, as well as SCG, for one representative subject receiving intravenous vasodilators. DV This shows the changes in [location]. Note that all signals shown are synchronized with their corresponding R peaks. The overall average of PAP and PCWP signals decreased with vasodilator infusion, but SCG DV The contraction portion of the signal later shifted relative to the ECG R peak after vasodilator infusion. Figure 9 shows the correlation analysis between different portions of the SCG signal and the DTW distance from the axis (changes in the SCG signal due to vasodilators) (i.e., R), with changes in mean pulmonary artery pressure (ΔPAM) and mean wedge pressure (ΔPCWP) of the SCG signal (i.e., changes in the SCG signal due to vasodilators). 2 Figure 10 shows the values ​​of the five DTW distances selected using the feature selection techniques described in the Methods chapter, and their relative importance (weights) in the regression model for the training test set. Figure 10 shows the ranking (i.e., relative weighting) of the relative feature importance of the features in the ΔPAM (Figure 10A) and ΔPCWP (Figure 10B) regression algorithms for the training test set. Dias: Total vasoconstriction, ED: Early vasoconstriction, LD: Late vasoconstriction, Sys: Contraction, ES: Early contraction, and LS: Late contraction. The length of time for each segment is explained in Figure 9.

[0184] Figure 8 shows the correlation analysis between the actual (measured) and estimated values ​​of ΔPAM and ΔPCWP for both the training test and validation sets. Figure 8 shows the correlation analysis between predicted and measured ΔPAM for the training test set (Figure 8A), and the correlation analysis between predicted and measured ΔPAM for the validation set (Figure 8B). Correlation analyses between predicted and measured ΔPCWP for the training test set (Figure 8C), and the correlation analysis between predicted and measured ΔPCWP for the validation set (Figure 8D) are also shown. The correlation results show that for ΔPAM, the RMSE was 3.2 mmHg for the training test set. 2 The value is 0.8, and in the validation set the RMSE was 3 mmHg. 2 The value is 0.77, and for ΔPCWP, the RMSE in the training test set was 3.6 mmHg and R 2 The value was 0.87, and in the validation set the RMSE was 6 mmHg. 2 The value is 0.78. E. Supplementary Materials

[0185] Separation of training test and validation sets: The dataset was randomly divided into two groups in a 75:25 ratio: 15 heart failure (HF) patients for the training test set and 5 HF patients for a separate, independent validation set. The only constraint used to separate the dataset was to maintain an equal proportion of HF with reduced ejection fraction (HFrEF) and HF with preserved ejection fraction (HFpEF) in both the training test and validation sets. As a result of this randomization, the training test set consisted of 4 HFpEF patients and the validation set consisted of 1 HFpEF patient.

[0186] Filtering and Heart Rate Segmentation: Digital filtering (cutoff frequencies: 0.5-40.0 Hz for ECG signals, 1-40 Hz for SCG signals) was applied to the raw ECG and SCG signals from the wearable patch to remove out-of-band noise. These cutoff frequencies were used to remove out-of-band noise without altering the signal shape. After the filtering step, accelerometer magnitude (SCG) was segmented. Mag The fourth SCG signal, which represents ) is obtained from the three SCG axes (SCG HtoF SCG Lat SCG DV It was calculated using the vector sum of ).

[0187] The ECG signal (in a 20-second frame) was amplitude-normalized, and the R peak of the ECG signal was detected using the Pan-Tompkins method. The SCG signal (4 axes of the SCG) was segmented into individual heartbeats using the R peak of the ECG signal. Each heartbeat was clipped to a duration of 500 ms before and after the R peak. The 500 ms SCG frame before the R peak roughly represents the ventricular diastolic phase of the cardiac cycle, and the 500 ms SCG frame after the R peak roughly represents the ventricular systolic phase of the cardiac cycle. The 500 ms duration before and after the R peak was selected because most of the relevant diastolic and systolic cardiac events of the subject (e.g., rapid inflow, atrial contraction, isovolumetric contraction, ventricular ejection, etc.) occur within this time frame for the corresponding R peak of the ECG. A fixed time window was selected to clip the ECG and SCG signals to obtain a repeatable and comprehensive feature extraction process.

[0188] Removal of abnormal SCG heart rates: Following heart rate segmentation of wearable psychocardiogram (SCG) signals, abnormal heart rates were separately removed from the SCG for each axis and segment (dilation and contraction) of the SCG signal for two distributions from two states (reference value, BL and vasodilator infusion, VI). Principal component analysis (PCA) was used to remove abnormalities from specific distributions, taking the first three principal components (PCs) to reduce the dimensionality of 500-sample length SCG heart rates (frames 500 ms long at a sampling frequency of 1 kHz) to three dimensions. This low-level representation of SCG heart rates was used in a Gaussian mixed model (GMM) to determine the likelihood that each sample belonged to a specific distribution (BL or VI) for a specific segment and axis of the SCG. For a specific distribution, the point with the lowest 20% probability was detected as an abnormality in the distribution. The decision to exclude 20% was based on an initial analysis that removed 10%, 20%, and 30% of pulses as outliers. The number of principal components (e.g., three in this case) that generate the GMM for a particular distribution was determined based on an analysis of the proportion of variance explained by the number of PCs and the overall estimation accuracy. Since most of the power in the SCG signal remains in the contraction portion of the signal, this ultimately governs the removal of outliers in the expansion portion of the signal. For this reason, outlier removal was performed separately for the expansion and contraction portions of the SCG.

[0189] Correlation analysis between SCG DTW distances due to ΔPAM and ΔPCWP: Figure 9 shows the R-value of the dynamic time-stretching (DTW) distances from different parts and axes of the SCG signal for the training test set, due to the change in mean pulmonary artery pressure (ΔPAM) and the change in mean wedge pressure of pulmonary capillaries (ΔPCWP). 2 The values ​​are shown. In the case of ΔPAM, the change in SCG during the initial systolic phase (isovolumetric contraction, IVC phase) provides the most relevant information regarding the change in PAM, and the dorsal-ventral direction (SCG) during the IVC phase. DV The change in SCG in ) is the highest R for ΔPAM. 2This shows 0.8. In the case of ΔPCWP, the change in SCG during the late diastolic (atrial systolic) phase provides the most relevant information related to the change in PCWP. Here, the SCG magnitude signal (SCG) during late diastolic (atrial systolic) Mag The change in ) is the highest R for ΔPCWP. 2 This shows a value of 0.87. Overall, the figure shows that ΔPAM is more associated with changes in the contractile portion (more precisely, the IVC phase) of the SCG signal, while ΔPCWP is more associated with changes in the later ventricular dilation (i.e., atrial contraction) portion of the SCG. This may be explained by physiological reasons, as the pulmonary artery is directly connected to the right ventricle, and the ventricular contraction (reduction) phase dominates the changes in PAM. On the other hand, the pulmonary capillaries are connected to the left atrium and are more associated with the atrial contraction phase. These preliminary results can be verified with simultaneous imaging modalities in large-scale studies of diverse subjects with various cardiovascular diseases.

[0190] Relative Feature Importance: Figure 10 shows the relative weighting of features in support vector regression with a linear kernel for estimating ΔPAM and ΔPCWP in the training test set. Here, the top-ranking features related to both ΔPAM and ΔPCWP are the changes in SCG during the IVC period. Lat As with the results obtained from the individual correlation analyses (Figure 9) between the target variables of DTW distance (ΔPAM and ΔPCWP), all of the top five features for ΔPAM originate from the contraction portion of the SCG. In the case of ΔPCWP, three of the top five features originate from the contraction portion of the SCG, and two originate from the expansion portion. Both Figures 10 and 11 illustrate the importance of the expansion portion of the SCG in estimating ΔPCWP. Most SCG research studies have focused on the contraction portion of the signal. These results suggest that the expansion portion of the SCG signal also has the potential to provide relevant information about pulmonary congestion. (Consideration)

[0191] This embodiment demonstrates that changes in SCG can track rapid changes in PAM and PCWP caused by systemic vasodilators administered intravenously to HF patients. SCG signals obtained using a wearable patch can track changes in hemodynamic congestion. This, in turn, provides a way to track hemodynamic changes in HF patients' daily lives and activities using wearable sensors.

[0192] Two findings in this embodiment are the correlation and ranking of the importance of individual features (DTW distance) corresponding to changes in PAM and PCWP due to systemic vasodilator infusion. The results from this embodiment indicate that changes in SCG during the ventricular systole (more specifically the IVC phase) of the cardiac cycle provide significant information regarding changes in PAM, while changes in SCG signals during ventricular systole and atrial systole (active ventricular diastole) provide the most appropriate information related to changes in PCWP. With vasodilator infusion, PAM and PCWP decrease as well as preload, and the IVC time interval (i.e., PEP) is inversely correlated with preload. Therefore, as expected, we observed that PEP (as shown in Figure 7) increases as preload decreases. These scientific findings disclose a method for elucidating the source of the SCG signal itself and for using these signals to extract physiologically meaningful information beyond vital signs and cardiac timing intervals. These results also demonstrate the importance of the ventricular diastole phase of the SCG signal. This aspect is often ignored in most research studies focusing on SCG. It is possible to incorporate image modalities to understand how changes in SCG are related to fundamental physiological changes caused by physiological or pharmacological perturbations.

[0193] Another finding from this embodiment is the use of a simple linear model (linear SVR) rather than a complex nonlinear model to estimate changes in hemodynamic congestion. Simple linear models can provide more insight into the model and the corresponding features used to form it than complex nonlinear models, which are often a "black box" in practice. In this study, due to the small number of subjects, using a simpler regression model allowed for a better understanding of key features (SCG segments and axes) associated with abrupt changes in hemodynamic congestion, improving the reliability of the generalizability of this method. These methods make the model more physiologically insightful and interpretable.

[0194] Another finding of this study is the use of noise reduction and outlier removal, which significantly improved the overall accuracy of the estimations. The analysis revealed that 20–30 seconds of high-quality wearable ECG and SCG recordings were sufficient to track changes in hemodynamic congestion.

[0195] The wearable patch of this embodiment has been shown to track adaptive parameters of cardiopulmonary function (i.e., the patient's instantaneous oxygen uptake and clinical state) from both HF patients in a controlled clinical setting (cardiopulmonary exercise testing) and healthy subjects in an uncontrolled daily life setting. The method of tracking changes in hemodynamic congestion using SCG described in this embodiment can be incorporated into the method of tracking cardiopulmonary parameters using SCG of this disclosure.

[0196] In this embodiment, we estimated the changes in pulmonary artery mean pressure and pulmonary capillary wedge pressure in HF patients due to vasodilator infusion, i.e., changes in simultaneously recorded SCG signals. We developed a global regression model for estimating ΔPAM and ΔPCWP using a machine learning algorithm validated by one-out cross-validation. We demonstrated that changes in asymptomatic congestion can be tracked by tracking changes in SCG. This has the potential to be used for remote home management of HF patients. In summary, this study demonstrates that a discreet wearable patch can track hemodynamic congestion. [Examples]

[0197] This embodiment provides a wearable, inexpensive, and virtually invisible system for remotely monitoring HF patients using a wearable psychocardiogram (SCG) monitor. The SCG signal captures chest wall vibrations in response to ventricular ejection of blood and cardiac motion. The use of the SCG signal was investigated to classify the clinical state of HF patients in a resting state. Specifically, this embodiment demonstrated that the clinical state of HF patients in a resting state could be accurately classified. Since the patient's clinical state is determined using pulmonary capillary wedge pressure (PCWP) and cardiac index (CI), this classification indicates elevated filling pressure and decreased cardiac index. Because PCWP better represents left ventricular filling pressure, the ability to detect elevated PCWP provides healthcare providers with important information as well as, or better than, that of implantable devices that measure PA pressure. All data in this embodiment were collected in a cardiac catheterization laboratory to obtain both SCG signals and hemodynamic parameters while the patient was undergoing standard right heart catheterization (RHC) treatment. Although all data collected in this embodiment were obtained from measurements taken in a hospital, these models are predicted to be generalizable to home environments. This is because they only require data from non-invasive wearable patches to estimate clinical conditions.

[0198] (result)

[0199] The clinical status of HF patients was accurately classified using SCG characteristics.

[0200] We conducted an experiment to evaluate the discriminability of SCG features in HF clinical states, clinically determined from elevated filling pressure and low cardiac index. A binary classification task was defined to build a machine learning model. Here, the input was a set of 83 extracted features, and the output was a decompensated hemodynamic congestion state in patients (PCWP 20 mmHg and CI 2.2 L / min / m²). 2 ) or was expressed as compensatory. This task involved developing all models on data from 50 patients (52 records). This data is referred to as the training set. After building the classification model for the training set, we quantified the generalized performance of the developed classification model using unknown data from 13 other patients. This data is referred to as the validation set. Tables 2-1 and 2-2 show the demographic information of the subjects in the training set and validation set, respectively. [Table 8] [Table 9]

[0201] For the classification model, a Support Vector Machine (SVM) for this task was employed. Through kernels, the SVM can model the nonlinear relationship between the input, SCG features, and the output, clinical state. In this study, we assumed that the possible nonlinear relationships could be obtained using polynomial and radial basis function (rbf) kernels. To evaluate the classifier's performance, five-fold cross-validation was performed, and the classification accuracy and AUC for the training set were calculated. In each cross-validation split, the training and validation data consisted of different subjects; therefore, the cross-validation performance can reflect the generalization performance to unknown, new patient data. The results of this experiment are summarized in Table 2-3 and Figures 13A-B. The first three columns of Table 2-3 summarize the results of classifying HF clinical states based on the full set of features (82 SCG features). The last column shows the results using a portion of this feature set. The receiver operating characteristic (ROC) curve of the optimal classifier using the extracted features is shown in Figure 13A.

[0202] [Table 10]

[0203] In this example, since more features (83 features) are extracted than the number of data points (52 records from 50 patients), sequential forward selection (SFS) was employed to improve the generalization performance of the classifier. From Table 2-3, it can be seen that the performance of the classifier is improved by using SFS. Furthermore, SFS is useful in identifying which features are useful for performing the task, which in turn provides insight into which aspects of the SCG signal are useful. Nested cross-validation was performed to use SFS in 5-fold cross-validation. The outer cross-validation loop was used for an intermediate evaluation of model performance, and the inner cross-validation loop was used for model selection (i.e., SFS and hyperparameter optimization using grid search).

[0204] To further quantify the generalization performance of the classification framework, performance metrics for the validation set were calculated. A new classifier was trained using the entire training set and tested against the validation set. Although the classification distributions of the training set and the validation set are significantly different, the performance on the unknown validation set is remarkably similar, as shown in Figure 13B.

[0205] Useful SCG features for evaluating hemodynamic congestion include features from the horizontal axis and dilated portion of the recording.

[0206] The classification model learned in this embodiment relies on SCG features and provides insights into SCG signal characteristics relevant to hemodynamic assessment in HF. Since SCG signals are less well understood than other more commonly measured cardiovascular signals (e.g., electrocardiogram (ECG), photovoltameter (PPG), and electroimpedance electrocardiogram (ICG) signals), this dataset provides an opportunity to advance knowledge of SCG signal characteristics relevant to assessing congestion in HF patients. In this embodiment, we identified SCG features useful for distinguishing the clinical state of HF patients. These features are extracted from processed SCG pulses. Details of this processing are described in the "Materials and Methods" chapter. A key scientific problem addressed is the most effective use of SCG pulse clusters in recordings for extracting distinguishable and predictable information. The features examined here are based on fundamental previous studies on SCG signal analysis in HF patients, namely: (1) ensemble-averaged pulse characteristics from signal expansion and contraction time intervals, (2) variability characteristics across multiple pulses, and (3) frequency domain characteristics.

[0207] Ensemble-averaged pulse characteristics: Several relevant properties of a recording, such as clinical state and mean filling pressure, remain largely unchanged from one pulse to the next. These properties can be captured from ensemble-averaged pulses. The following features were calculated from ensemble-averaged pulses: the square root of the signal power (srPower) and the delay between the R peak and the minimum amplitude (delayMinAmp).

[0208] The selection of these features is motivated by numerous previous studies, suggesting that these methods are relevant to several key cardiac events, such as the timing of aortic valve opening and important events within the cardiac cycle. For a given pulse, these features can be calculated over physiologically defined time intervals corresponding to the systolic and diastolic phases. This can be indexed by defining the position of the ECG R peak at time zero, based on the ECG signal. Three intervals were defined, as shown in Figures 14A (decompensated signal) and 14B (compensated signal): ventricular dilation 1 (-150ms to 0ms, dias1), sys (0ms to 300ms, sys), and ventricular dilation 2 (300ms to 650ms, dias2).

[0209] Conventional methods for SCG analysis focus only on systolic timing intervals. However, since we are interested in finding the relationship between the SCG signal and filling characteristics (i.e., PCWP pressure), we included these typically excluded intervals in the analysis. From each interval, we obtained six reference level features for each channel by calculating these two features. Reference level features can be calculated for individually segmented pulses and ensemble-averaged pulses.

[0210] Variability Features: In addition to the features from these ensemble-averaged pulses, other features we examined specifically focus on signal variability within pulse groups. We hypothesize that HF ​​patients with hemodynamic congestion (decompensated) have higher SCG signal variability than compensated patients. This is because the cardiac and vascular structures operate in a suboptimal state and therefore cannot always produce pulses with the same intensity and timing characteristics. Capturing this variability would therefore be important as it may suggest a suboptimal clinical state for the cardiovascular system. Specifically, we extract baseline features from each individual pulse. Then, we calculate the standard deviation of the features using the set of feature values ​​from each pulse. An example of variability in features is shown in Figures 14A-B.

[0211] Descriptive statistics such as standard deviation provide a single fixed feature representation across different length records (which will have different numbers of pulses). Since the standard deviation is calculated for each of the six baseline features, six statistical features were calculated from the pulse group.

[0212] Frequency Domain Features: The third set of features examined is frequency domain features. Figures 15(A)-(C) visualize the normalized power spectral density of randomly selected compensatory and decompensatory recordings. In these visualizations, it can be seen that the frequency characteristics differ for decompensatory and compensatory recordings around 200-250 Hz and 0-50 Hz. Therefore, two frequency domain features are extracted from the SCG recordings: the ratio of signal power between 205-250 Hz and 5-40 Hz, and the ratio of signal power between 0-5 Hz and 5-40 Hz.

[0213] After processing the records, a total of 83 features (one of which included the RR interval) were extracted. Using SFS, a greedy approach was employed to determine which features were suboptimally optimized for the classification task. Figures 16(A)-(C) show the results from feature voting as an estimate of feature importance, and from two further methods more commonly used to assess feature importance. In Figure 16(A), the bar graph shows the number of times a feature was selected from five cross-validation splits. The more features selected, the higher their importance. In Figure 16(B), feature importance was calculated based on the separating hyperplane weights of a linear SVM. The higher the weight, the more strongly the change in that particular feature influences the determination of its importance. Finally, permutation feature importance analysis was performed, observing the performance degradation obtained by swapping each feature value. The greater the performance degradation due to permutation, the higher the importance of that feature. The results of permutation feature importance are shown in Figure 16(C). In particular, in the analysis of the importance of all three features, the top two features are retained. Both features are extracted from the horizontal axis of the SCG, and the second feature is calculated directly from the extended portion of the SCG pulse.

[0214] (Consideration)

[0215] The results in Figures 14A-B and Table 2-3 demonstrate that accurate classification of the clinical status of HF patients is possible using SCG features. Currently, determining the clinical status of HF patients for physiological decompensation requires catheterization, which is expensive and invasive. If this wearable device can provide data that facilitates accurate classification of clinical status, it could be used as a pre-screening tool to reduce the number of RHC treatments, thereby lowering the cost of HF care and improving the quality of life. These research results suggest that such pre-screening using SCG is promising.

[0216] Accurate classification of the clinical state of HF is crucial because it represents the potential detection of elevated filling pressure from HF patients. Accurate classification of the clinical state at home using wearable devices can significantly improve HF care through reduced hospital stays. Daily or more frequent assessment of the clinical state with wearable devices enables filling pressure-inducing therapy, similar to the methods used in conventional studies with implantable hemodynamic monitors. Importantly, by providing indications of elevated filling pressure rather than a black-box output that makes decisions as in conventional studies, physicians can better engage in the process with explainable and understandable results, and existing flowcharts and guidelines can be directly utilized.

[0217] The features selected in this experiment, along with feature importance analysis, provide important scientific insights into the characteristics of SCG signals (see Figure 16 (A)-(C)). Although SCG signals have been measured and studied since the 1960s, the sources of these signals and their relationship to the underlying hemodynamic events are not well understood. This study directly measures SCG signals along with RHC waveforms from HF patients, thereby enabling an examination of how SCG characteristics represent underlying hemodynamics. The optimal set of features selected by the best-performing classifier, SFS, consists of transverse and magnitude channels, thus highlighting the importance of analyzing all axes of SCG signal data, not just the dorsal-ventral axis as some previous studies have suggested. In all feature importance methods, the top two features always consist of SCG transverse channels. This finding that SCG transverse channels provide primary hemodynamic information has been supported by recent studies. The top features originate from the frequency domain and had not been previously studied. In other words, the ratio of higher frequency components (205–250 Hz) to lower frequency components (4–50 Hz) is a discriminant feature. This may be due to higher filling pressure in decompensated patients. Higher filling pressure can lead to louder or more rapid valve closure. This can be reflected as higher frequency components in the acceleration signal captured in the chest. Future research should study lateral SCG measurements to better understand the physiological causes of these vibrations and provide mechanistic insights into why they make an important contribution to estimating clinical conditions.

[0218] This study is the first to demonstrate the usefulness of SCG signals in classifying the clinical state of HF in a stationary state and detecting elevated filling pressure. The results presented in this study provide strong scientific evidence to support the investigation of SCG signals in HF care in home settings, potentially reducing HF-related hospital stays. Furthermore, this study not only predicts the risk of hospital stay (i.e., acute decompensation) but also provides feedback to healthcare providers by offering an indication of elevated filling pressure.

[0219] (Materials and Methods)

[0220] research design

[0221] The objective of this study was to investigate the discriminative features of SCG to identify the clinical state (i.e., decompensated and compensated states) of quiescent HF patients and to examine the correlation between SCG and hemodynamic parameters. Data were collected from a group of patients prescribed RHC (using the Mac-Lab hemodynamic recording system) to determine clinical state and capture hemodynamic parameters. This study was conducted based on a protocol reviewed and approved by the University of California, San Francisco (UCSF) and Georgia Institute of Technology (GIT) IRS. A total of 63 subjects diagnosed with HF participated in the study. Exclusion criteria included patients with cardiac shock, patients with implanted ventricular assist devices, or patients who had previously undergone a heart transplant. Demographics of the study population are shown in Table 2-1 (training set) and Table 2-2 (validation set). Each subject submitted written informed consent prior to data collection.

[0222] All SCG data were collected while the patient was receiving RHC treatment. First, the patient was positioned supine on an examination table. A wearable patch hardware was attached directly below the patient's cephalic incision to acquire the SCG signal. During this procedure, the patient was instructed to remain as still as possible. Subsequently, a catheter was inserted and hemodynamic parameters were measured. The patient's supine position and immobility are referred to as the resting state. All signals collected during a single RHC treatment are referred to as a recording, which consists of ECG and SCG signals from the wearable patch and pressure waveforms from the catheterization method using the Mac-Lab hemodynamic recording system (65 recordings in total). Furthermore, compensatory recordings are defined as recordings obtained from patients diagnosed with hemodynamically compensatory conditions, and decompensatory recordings are defined similarly. The data acquisition setup, along with exemplary signal extraction, is shown in Figure 1E.

[0223] Based on RHC treatment, the following hemodynamic parameters were measured for each record: right atrial pressure (RAP), right ventricular pressure (RVP), pulmonary artery pressure (PAP), cardiac output (CO), and PCWP. Based on the hemodynamic parameters, the clinical status was determined as follows: the patient's mean PCWP was ≥20 mmHg and CI was 2.2 L / min / m². 2 The patient is considered decompensated if the following conditions are met. Otherwise, the patient is considered compensated. In some rare cases, if either the PCWP or CI value is significantly high / low, the caregiver's judgment takes precedence over this rule. For example, a borderline PCWP of 16 mmHg and an extremely low CI of 1.3 L / min / m². 2 In this case, the patient was considered decompensated. Note that since two patients received RHC treatment twice, the number of records is two more than the number of patients in the training set.

[0224] In this embodiment, a wearable patch device was used to capture SCG and ECG signals. This device samples the ECG signal at 1 kHz and the accelerometer signal at 500 Hz, and saves the data to a microSecure Digital (microSD) card within the patch. Custom software transfers all data from the microSD card to a computer, samples the accelerometer again at 1 kHz, and ensures that all signals share the same sampling rate to facilitate processing. Subsequently, for further analysis, all signals are decimated to 500 Hz in our processing algorithm.

[0225] The SCG signal is processed to extract features that may be useful in identifying the clinical condition of HF patients. Figures 2a-c illustrate how the SCG signal is processed. The processing results in high-quality SCG pulses for feature extraction, which can be seen in Figure 2c.

[0226] The first preprocessing step is to remove the first 5 minutes from the start and the last 5 minutes from the end of each recording. This is because, during recording, the sensor may start recording before the device is attached to the subject, and the recording may end after the device is removed. As a second step, the ECG signal was subjected to band-pass filtering with the following digital filter specifications: finite impulse response (FIR) filtering with a passband of 1–30 Hz. For the SCG signal, high-pass filtering was performed with a cutoff frequency of 1 Hz. In both filters, the stopband was attenuated by 80 dB. The passband of the ECG signal was selected with respect to the separation of R peaks to facilitate their detection in the next processing step. For the SCG signal, the passband was selected to suppress out-of-band noise, preserve SCG signal characteristics, and investigate high-frequency components.

[0227] After filtering both the ECG and SCG signals, two further channels were formed in the SCG. These are called the magnitude channel and the XY magnitude channel, and are represented by the following equation.

number

[0228] As a final preprocessing step, R peaks were detected in the ECG signal, followed by pulse segmentation in the SCG signal. Two algorithms were used for R peak detection: Pan Tompkins performed by PhysioNet and phasor transformation using a custom implementation. To reduce false positives in R peak detection, R peaks detected by both algorithms were selected. Twelve second windows were used for detection of R peaks.

[0229] Using the R peak, the pulse segmentation of the SCG signal was performed as follows. That is, 150 ms before the R peak (including the ventricular diastole) and the 15th percentile of the RR interval after the R peak (including only the current pulse) were respectively segmented at the start and end of the pulse. As a result, an SCG pulse array was constructed for each channel of the SCG. Note that, different from the conventional research that performs pulse segmentation from the R peak (i.e., 0 ms before and after the R peak) to about 700 ms after the R peak, in this embodiment, the SCG features observed at the ventricular diastolic timing may be highly relevant to including information on the filling pressure, so the ventricular diastolic timing was intentionally included.

[0230] Motion artifact rejection and signal quality indexing

[0231] The SCG signal is susceptible to the influence of motion artifacts. That is, when the subject moves, the SCG vibration deteriorates due to higher-amplitude motion artifacts. In the collected dataset, the patients were instructed to keep as quiet as possible, but still motion artifacts were present in the recordings. Figures 3 a)-d) show examples of such artifacts in one recording.

[0232] By utilizing the observation that the amplitude of the movement artifact is higher than the SCG vibration, we considered a simple algorithm that detects SCG pulses damaged by movement and then discards SCG pulses deteriorated by movement. This algorithm takes a segmented pulse array as input and outputs an index of the pulse deteriorated by movement. Pseudo-code is shown for Algorithm 1 (described above). The clue for detection is to find the two most similar consecutive pulses. If two consecutive pulses are not deteriorated by movement, it is unlikely that there will be a large change in the SCG morphology in a short time, so they should be in a similar form. Therefore, the two most similar consecutive pulses should not have movement artifacts (see c) of Figure 3, which shows an example of the two most similar consecutive pulses in this recording). By calculating a simple threshold using the part of the recording without movement artifacts, movement artifacts as outlined in Algorithm 1 (described above) are detected. The results of the movement artifact detection algorithm are shown in a)-d) of Figure 3.

[0233] To measure the similarity between two pulses, the following formula was used.

Number

[0234] The range of values obtained from the output of this formula is 0 to 1. The closer the output is to 1, the more similar the inputs are, and the closer the output is to 0, the less similar the inputs are.

[0235] After detecting and rejecting pulses degraded by motion artifacts, signal quality indexing (SQI) was applied separately to each channel of the SCG to extract high-quality SCG pulses. One method can yield many templates that can be extracted from a set of subjects and then used to analyze SCG recordings from subsequent subjects. This method attempts to generalize to subjects, and it has been found that it generates only insufficient templates for this application. The use of healthy subjects in previous studies made it relatively easy to extract a generalized template set, but in this case, it is assumed that the HF population exhibits significantly greater heterogeneity. The solution is to use a specific set of templates from each recording (i.e., each patient). Specifically, one template is used for each channel of the SCG from one subject. By modifying it in this way, the templates are fitted to specific subjects without motion artifacts. The output of the SQI is a set of pulses with a quality score associated with each pulse. Based on the quality scale, the top 5% of pulses were used to extract features.

[0236] statistical analysis

[0237] Standard analyses were used to evaluate the classification quality of the clinical state estimation model. Ground truth data was collected from catheter measurements. Specifically, five-fold cross-validation was performed against the training set to estimate performance against another unknown validation set.

[0238] It will be understood that the embodiments and claims disclosed herein are not limited in their intended use to the structure and arrangement of components described in the specification and illustrated in the drawings. Rather, the specification and drawings provide examples of the assumed embodiments. The embodiments and claims disclosed herein are also possible in other embodiments and can be implemented and performed in various ways. It will also be understood that the text and terminology used herein are for illustrative purposes only and should not be considered to limit the scope of the claims.

[0239] Therefore, it will be apparent to those skilled in the art that the concepts on which the present application and claims are based can be readily used as a basis for designing other structures, methods, and systems to carry out the embodiments and some of the objectives of the claims presented herein. It is therefore important to consider the claims as including such equivalent structures.

[0240] Furthermore, the purpose of the abstract is to enable the United States Patent and Trademark Office and the general public (in particular, technical trainees unfamiliar with patent and legal terminology and language) to quickly determine the essence and principles of the technical disclosure of this application upon reading it. The abstract is not intended to define the scope of the claims of this application, nor to limit the scope of the claims in any way.

Claims

1. A housing including a first surface configured to face away from the user's chest and a second surface configured to face the chest, It is configured to measure the electrocardiogram signal of the user's heart, and includes electrodes attached to the chest, An accelerometer disposed within the housing, configured to measure one or more psychocardiogram (SCG) signals of the user, A connector disposed on the second surface of the housing, configured to connect to the electrodes such that the electrodes are attached to the chest, thereby fixing the housing in a predetermined position on the user's body; The memory located within the aforementioned enclosure, The system comprises a memory, an electrode, and a processing system including one or more processors operably connected to the accelerometer, The processing system is Processing of the electrocardiogram signal, Based on the timing of one or more features of the electrocardiogram signal, one or more of the psychocardiogram signals are segmented into a set of psychocardiogram segments associated with a set of individual heartbeats. The set of psychoelectrocardiogram segments is input into a motion artifact detection algorithm that identifies pulses degraded by motion artifacts in the set of psychoelectrocardiogram segments based on changes in morphology between different psychoelectrocardiogram segments in the set of psychoelectrocardiogram segments. Rejecting a portion of the set of psychocardiogram segments based at least partially on the identification of pulses degraded by the aforementioned motion artifacts, A device configured to generate an assessment of the user's cardiac health by rejecting a portion of the set of psychoelectrocardiogram segments and then determining one or more hemodynamic parameters of cardiac filling characteristics based on the characteristics of the remaining psychoelectrocardiogram segments.

2. The apparatus according to claim 1, wherein the one or more psychocardiogram signals include one or more psychocardiogram signals from two or more axes.

3. The apparatus according to claim 2, wherein the axis is selected from a group consisting of the transverse axis, the cephalopedus axis, the dorsoventral axis, and combinations thereof.

4. The processing system further includes: Before generating the assessment of cardiac health, a combined psychoelectrocardiogram signal is generated using one or more psychoelectrocardiogram signals from two or more axes. The apparatus according to claim 2, configured to determine one or more hemodynamic parameters based on the combined psychoelectrocardiogram signals.

5. The apparatus according to claim 1, wherein the processing system is further configured to determine one or more hemodynamic parameters based on features extracted from the remaining psychoelectrocardiogram segments in at least the expanded portion of the heartbeat.

6. The apparatus according to claim 1, wherein the processing system is further configured to use a classification model to process the electrocardiogram signal and the remaining psychoelectrocardiogram segment to obtain a classification of the user's clinical state of heart failure.

7. The apparatus according to claim 1, wherein the one or more hemodynamic parameters include the user's filling pressure or a change in the user's filling pressure.

8. The processing system is Furthermore, it is configured to determine reference values ​​for one or more hemodynamic parameters for the user before generating the assessment of the cardiac health status, and The apparatus according to claim 1, further configured to generate an evaluation of the heart's health status based on the aforementioned reference values.

9. The apparatus according to claim 8, wherein the processing system is further configured to determine the reference value using a population-level regression algorithm.

10. The apparatus according to claim 8, wherein the processing system is further configured to determine the reference value using user data collected during right heart catheterization or clinical trials.

11. The apparatus according to claim 1, wherein the housing is configured to be attached to the user's chest below the neck incision.

12. It further includes sensors configured to measure environmental parameters, The apparatus according to claim 1, wherein the processing system is configured to determine one or more hemodynamic parameters based on the environmental parameters.

13. The housing further comprises a photoplethysmography sensor located within the housing, The photoplethysmography sensor is configured to measure the user's photoplethysmography signal. The apparatus according to claim 1, wherein the processing system is configured to determine one or more hemodynamic parameters based on the photoplethysmography signal.

14. The enclosure further comprises a gyroscope located within the aforementioned housing, The gyroscope is configured to measure the user's gyrocardiogram signal. The apparatus according to claim 1, wherein the processing system is configured to determine one or more hemodynamic parameters based on the gyrocardiogram signal.

15. A housing configured to be worn on the user's chest below a neck incision, the housing includes a first surface configured to face away from the chest when worn on the chest, and a second surface configured to face the chest, An electrode is attached in contact with the skin on the user's sternum and configured to measure the electrocardiogram signal of the user's heart, A sensor disposed within the housing and configured to measure the user's psychocardiogram (SCG) signal on two or more axes, A connector disposed on the second surface of the housing, configured to connect to the electrode such that the electrode is attached to the skin and the housing is fixed to the user on the sternum, The enclosure includes a controller located within the aforementioned housing, The aforementioned controller, Processing of the electrocardiogram signal from the electrodes, Based on the timing of one or more features of the electrocardiogram signal, the psychocardiogram signal from the sensor is segmented into a set of psychocardiogram segments associated with a set of individual heartbeats. The set of psychoelectrocardiogram segments is input to a motion artifact detection algorithm that identifies pulses degraded by motion artifacts within the set of psychoelectrocardiogram segments. Based on the identification of pulses degraded by the aforementioned motion artifact, a first portion of the set of psychoelectrocardiogram segments is rejected, After rejecting the first portion of the electrocardiogram segment, signal quality indexing is applied to the remaining electrocardiogram segment to identify a second portion of the remaining electrocardiogram segment that has a quality score at a predetermined percentage higher than the quality scores of the other electrocardiogram segments. Determination of reference values ​​for the aforementioned user, A device configured to generate an assessment of cardiac health by determining one or more hemodynamic parameters of cardiac filling characteristics based on the second portion and the reference value.

16. The aforementioned sensor is a first sensor, The apparatus further comprises a second sensor located within the housing and configured to measure environmental parameters. The apparatus according to claim 15, wherein the controller is configured to determine one or more hemodynamic parameters based on the environmental parameters.

17. The apparatus according to claim 16, wherein the environmental parameters include at least one of temperature, humidity, and altitude.

18. The aforementioned sensor is a first sensor, The apparatus further comprises a second sensor located within the housing and configured to measure the user's photoplethysmography signal. The apparatus according to claim 15, wherein the controller is configured to determine one or more hemodynamic parameters based on the photoplethysmography signal.

19. The apparatus according to claim 15, wherein the one or more hemodynamic parameters include the user's filling pressure or a change in the user's filling pressure.

20. The apparatus according to claim 15, wherein the controller is further configured to determine the reference value by determining a reference filling pressure for the user using a collective-level regression algorithm.

21. The apparatus according to claim 15, wherein the controller is further configured to use a classification model to process the electrocardiogram signal and the second portion of the psychoelectrocardiogram segment to obtain a classification of the user's clinical state of heart failure.

22. The sensor is the first sensor configured to measure psychocardiogram signals in three axes, The device further comprises a second sensor located within the housing and configured to measure the user's gyrocardiogram signal. The apparatus according to claim 15, wherein the controller is configured to determine one or more hemodynamic parameters based on the gyrocardiogram signal.

Images