Method, system and computer program for detecting heart health status

Abstract

A method for detecting a cardiac health state of a user comprises the steps of: receiving, in a processing device, a measurement signal from an IMU placed on a body of the user; determining, in the processing device, a plurality of features based on the measurement signal; and determining, in a machine learning engine of the processing device, a heart health state based on the plurality of features.

Description

Technical Field

[0001] The present invention relates to a method, system and computer program for detecting a user's heart health status, preferably a method, system and computer program for detecting congestive heart failure (CHF). Prior art

[0002] The health of a human heart is crucial to their health. Early detection of heart problems can avoid serious complications of human health. These problems may be CHF, coronary artery disease, atrial fibrillation, hypertension, etc. Traditionally, heart monitoring is done by electrocardiogram (ECG). Some heart monitoring technologies are based on ballistocardiography (BCG) or shockcardiography (SCG). SCG is usually measured on a person's chest. BCG can be measured in many different ways. While many SCG or BCG only measure or use one or a few dimensions, WO201736887 proposes to base heart monitoring on a combination of ECG and BCG measured at the center of gravity of the body, and ultimately combined with SCG. It is proposed to use multidimensional BCG and SCG measurements, measuring three linear dimensions with accelerometers to obtain the linear kinetic energy of heart motion, and measuring three rotational dimensions with gyroscopes to obtain the rotational kinetic energy of heart motion. However, these solutions require all complex and specific clinical measurement equipment, such as ECG, SCG and / or BCG equipment, which increases the cost of such heart monitoring and makes such heart monitoring unsuitable for home heart examinations that can be performed without medical assistance.

[0003] Some existing technologies propose to use daily life technology to measure the SCG of the heart and use machine learning (ML) technology to analyze the SCG so that everyone can perform simple heart monitoring.

[0004] US20210057101 discloses measuring BCG using a remote sensor installed in an apartment and analyzing BCG using an ML engine to detect CHF. However, remote detection of BCG is not very reliable, and features for ML analysis are not disclosed.

[0005] EP3267886 proposes to input 6-dimensional SCG signals into an ML engine to determine cardiac dysfunction such as atrial fibrillation. In order to reduce the dimensionality of the 6-dimensional SCG signal, a principal component analysis is proposed. Principal component analysis can reduce the dimensionality of the signal, but it cannot solve the problem of a large number of samples in the time dimension, which makes the ML analysis computationally complex and the output results unreliable.

[0006] EP3897355 proposes cardiac monitoring to detect internal bleeding. It discloses the use of an ML engine to analyze BCG and ECG to determine internal bleeding. BCG is measured when the body eats. Therefore, this solution also requires specific cardiac monitoring equipment.

[0007] EP3895605 discloses monitoring the heart using cardiac data such as ECG, electromyography, BCG or any other type. The data is subdivided into multiple sub-periods of about 30 seconds to 2 minutes. In each sub-period, a feature vector is calculated and fed into an ML engine. The ML engine detects sub-periods that are worth further analysis by humans, but cannot detect the health status of the heart.

[0008] EP3731746 discloses the use of a smartphone equipped with capacitive electrodes to measure the SCG and ECG of a user. The ECG is used to detect the heartbeat cycle in the measurement signal and to divide each heartbeat cycle into sub-segments such as systole and diastole, or even finer sub-segments. Features are extracted from the different sub-segments of the heartbeat cycle to feed into an ML engine to detect cardiac anomalies. As features, many potential time or frequency features are proposed. It is mentioned that cardiac anomalies can be detected without ECG, but it is not explained how to identify the heartbeat cycle and sub-segments without ECG.

[0009] However, most of the above solutions require specific measurement equipment, which makes it unrealistic to perform such heart monitoring in daily life. Most solutions require ECG measurements, which are difficult to obtain without ECG equipment. Solutions based solely on SCG measurements still remain vague regarding the feature vectors used to feed the ML engine. Therefore, there is currently no known solution that can provide simple and robust heart health status detection. Summary of the invention

[0010] The object of the present invention is to provide a method, system and computer program for detecting the heart health status of a user, which overcomes the problems of the prior art.

[0011] According to the invention, this object is solved by a method, a system and a computer program according to the independent claims.

[0012] According to the present invention, the object is solved by a method for detecting the heart health status of a user, the method comprising the following steps: receiving a measurement signal from an IMU placed on the user's body in a processing device; determining a plurality of features based on the measurement signal in the processing device; and determining the heart health status based on the plurality of features in an ML engine of the processing device.

[0013] The use of ML allows for the detection of cardiac health status independent of ECG measurements and complex clinical devices, for example using a simple smartphone.

[0014] The dependent claims relate to further advantageous embodiments.

[0015] In one embodiment, the IMU is arranged in a smartphone or a wearable device connected to the smartphone, wherein the measurement signal comprises a six-dimensional signal having three linear dimensions representing the linear acceleration of the heart measured by the accelerometer of the IMU and three rotational dimensions representing the angular velocity of the heart measured by the gyroscope of the IMU, wherein a heartbeat cycle of the user's heart is identified based on the measurement signal, wherein a plurality of sub-cycles in the heartbeat cycle are determined, wherein a plurality of features comprises linear features and rotational features, wherein the linear features comprise the value of at least one linear feature variable of the heart motion in each sub-cycle calculated based on the three linear dimensions of the measurement signal, wherein the rotational features comprise the value of at least one rotational feature variable of the heart motion in each sub-cycle calculated based on the three rotational dimensions of the measurement signal. The combination of ML, six-dimensional IMU measurement by smartphone, ECG independence and robustness of rotational and linear parameters with respect to the orientation of the smartphone allows good heart state detection using a standard smartphone, simply by running software on the smartphone, and measuring the six-dimensional measurement signal with the IMU of the smartphone. Even if this combination of features is particularly effective, several features may be used alone or in other sub-combinations. Preferably, the plurality of features are determined independently of the ECG measurements.However, it is also possible to determine the plurality of features based on additional ECG measurements.

[0016] In one embodiment, the measurement signal comprises a six-dimensional signal, wherein three linear dimensions represent the linear acceleration of the heart measured by an accelerometer of the IMU, and three rotational dimensions represent the angular velocity of the heart measured by a gyroscope of the IMU, wherein a heartbeat cycle of the user's heart is identified based on the measurement signal, wherein a plurality of sub-cycles in the heartbeat cycle are determined, wherein a plurality of features comprise linear features and rotational features, wherein the linear features comprise the value of at least one linear feature variable of the heart motion in each sub-cycle calculated based on the three linear dimensions of the measurement signal, wherein the rotational features comprise the value of at least one rotational feature variable of the heart motion in each sub-cycle calculated based on the three rotational dimensions of the measurement signal. Having a combined / scalar linear value and a combined / scalar rotational value allows to eliminate the dependency of the measurement signal on the orientation of the IMU. The dependency of the measured components on the orientation of the IMU has been shown to increase the error rate of the ML engine.

[0017] In one embodiment, the plurality of features include values ​​of a linear feature variable at different times in a heartbeat and values ​​of a second feature variable at different times in a heartbeat, wherein the first feature variable is determined based on three linear dimensions, and wherein the second feature variable is determined based on three rotational dimensions. Having combined / scalar linear values ​​and combined / scalar rotational values ​​allows for elimination of dependency of the measured signal on the orientation of the IMU. Dependence of components of the measurements on the orientation of the IMU has been shown to increase the error rate of the ML engine.

[0018] In one embodiment, the (at least one) linear characteristic variable is one or more of the following: linear (cardiac) power, linear (cardiac) kinetic energy, and linear (cardiac) work. In one embodiment, the (at least one) rotational characteristic variable is one or more of the following: rotational (cardiac) power, rotational (cardiac) kinetic energy, and rotational (cardiac) work. These kinetic parameters are independent of the orientation of the sensor and have been shown to be very robust in detecting cardiac health states.

[0019] In one embodiment, the measurement signal includes a six-dimensional signal having three linear dimensions representing the linear acceleration of the heart measured by an accelerometer of the IMU, and three rotational dimensions representing the angular velocity of the heart measured by a gyroscope of the IMU, wherein a heartbeat cycle of the user's heart is identified, wherein multiple sub-cycles in the heartbeat cycle are determined, wherein multiple features include linear features and rotational features, wherein the linear features include values ​​of linear kinetic energy of the heart motion in each sub-cycle calculated based on the three linear dimensions of the measurement signal, and wherein the rotational features include values ​​of rotational kinetic energy of the heart motion in each sub-cycle calculated based on the three rotational dimensions of the measurement signal.

[0020] In one embodiment, the measurement signal comprises a six-dimensional signal having three linear dimensions representing the linear acceleration of the heart measured by the accelerometer of the IMU, and three rotational dimensions representing the angular velocity of the heart measured by the gyroscope of the IMU. The inventors have found that measurement signals in all six dimensions bring significantly better results, because information cannot be hidden in some unmeasured dimensions. Prior art usually measures only certain dimensions, or at least only considers certain dimensions for analysis.

[0021] In one embodiment, a linear velocity signal is calculated based on three linear dimensions, and the plurality of features include at least one linear feature calculated based on the linear velocity signal; and / or a rotational acceleration signal is calculated based on three rotational dimensions, and the plurality of features include at least one rotational feature calculated based on the rotational acceleration signal. It has been found that the linear velocity-based and / or rotational acceleration-based dynamic features that are not directly measured contain a lot of valuable information that cannot be obtained from the measured signals. The features of the ML engine using these features are proven to be more robust than the features that only use the signal characteristics of the signals measured in the IMU. In particular, the linear kinetic energy calculated from the linear velocity signal is of particular interest as a feature of the ML engine.

[0022] In one embodiment, the method comprises the step of determining in the processing device a velocity signal based on the measurement signal, the velocity signal being representative of the velocity of the cardiodynamic movement, wherein the processing device determines a plurality of features based on the velocity signal.

[0023] In one embodiment, the processing device determines from the velocity signal a processed velocity signal representing the velocity squared of the cardiac dynamic motion or the kinetic energy of the cardiac motion, wherein the processing device determines a plurality of features based on the processed velocity signal.

[0024] In one embodiment, the plurality of features comprises values ​​of at least one feature variable at different times in a heartbeat.

[0025] In one embodiment, a heartbeat cycle of a user's heart is identified, wherein a plurality of sub-cycles in the heartbeat cycle are determined, wherein a plurality of features include the value of at least one feature variable in each determined sub-cycle. The definition of sub-cycles based (only) on measurement signals from an IMU (and not on an ECG) allows such segmentation to be created also in the absence of an ECG signal. This allows the feature extraction to be performed also in a simple smartphone.

[0026] In one embodiment, a reference point in a heartbeat cycle is determined, and a sub-cycle is defined based on the reference point and based on a fixed time width of the sub-cycle. Instead of calculating the time width (or start and end) of the sub-cycle for each heartbeat, the time width of each sub-cycle is selected to be fixed (from one heartbeat to another heartbeat), so it can be easily and reliably determined relative to the reference point. The use of a fixed time width of a sub-cycle allows a simple, robust and repeatable calculation of the sub-cycle, and therefore allows a repeatable and robust calculation of the value of the characteristic variable of the sub-cycle. Preferably, the fixed time width is defined independently of a physiological parameter or a time point (e.g., certain peaks in the signal, or the end or beginning of the systolic or diastolic phase). If only one physiological parameter is not correctly detected, the sub-cycle defined based on the physiological parameter is computationally complex and prone to error. The fixed time width is preferably an absolute fixed time width or a relative fixed time width. The absolute fixed time width has an absolute value of the time width that does not change from one heartbeat to another heartbeat. The relatively fixed time width has a relative value (percentage value) with respect to, for example, the length of a heartbeat cycle, and the relative value does not change from one heartbeat cycle to another.

[0027] In one embodiment, a heartbeat cycle in the measured signal is identified based on a matrix profile motif algorithm applied to the signal based on the measured signal, wherein the matrix profile motif algorithm identifies one or more reference heartbeat cycles as motifs in the signal based on the matrix profile. It is difficult to detect heartbeats in IMU signals without measuring ECG signals. The use of matrix profiles allows robust detection of heartbeat cycles at low computational cost.

[0028] In one embodiment, breathing information of the user is detected based on the measurement signal.

[0029] In one embodiment, the multiple features include breathing information of the user. Since the heartbeat depends largely on the breathing phase / state of the user, it is found that using this feature as an input to the ML engine can significantly improve the results of the ML engine. This is especially true when the breathing information is combined with multiple feature variables related to the linear and rotational motion of the heart in different sub-cycles. This allows the ML engine to distinguish between different motions of the heart in different breathing states, thus avoiding erroneous conclusions due to differences in the breathing states of the users.

[0030] In one embodiment, multiple features are determined independently of the ECG measurement. This allows the method to be implemented in any common smartphone with an IMU.

[0031] In one embodiment, the ML engine determines whether the heart exhibits congestive heart failure.

[0032] In one embodiment, the method includes the steps of placing the IMU on the chest of the user and measuring the measurement signal with the IMU when placed on the chest of the user.

[0033] In one embodiment, the IMU is included in a smartphone or a wearable device connected to a smartphone, wherein the processing device requests user input about what type of smartphone or wearable device or about the type of smartphone or wearable device from an operating system running on the smartphone, wherein a plurality of features depend on the type of smartphone received, in particular the weight of the type of smartphone received. The type of smartphone / wearable device is preferably the model. This allows any dynamic parameters to be calibrated by the correct mass or moment of inertia of the smartphone. Calibration using the sensor type makes the measurement signals more comparable. This embodiment is particularly advantageous for ML engines that determine cardiac health status, since different measurements in different smartphones / wearable devices should be comparable to provide good results. However, in the prior art, this point is always overlooked. However, this embodiment may also be advantageous when detecting the health status of a user without using an ML engine.

[0034] In one embodiment, the user lies down while a smartphone including an IMU is placed on the user's chest. Although the prior art mainly proposes sleeves or fixtures for fixing the measuring device to the chest or other parts of the body, the inventors have found that measuring the IMU signal while lying down has several advantages for the measurement quality. The user does not move, so the movement caused by the user's movement does not interfere with the measurement. The user relaxes when lying down, and the measurement signal is always measured in a relaxed state, which produces a more repeatable signal. Finally, the sleeve that fixes the device with the IMU suppresses movement, thereby reducing the signal quality. Other embodiments according to the invention are mentioned in the attached claims and the subsequent description of the embodiments of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 An embodiment of a system for detecting a user's heart health status according to the present invention is shown.

[0036] Figure 2 An embodiment of a method for detecting a user's heart health status according to the present invention is shown.

[0037] Figure 3 An exemplary embodiment of feature extraction is shown.

[0038] Figure 4 A first embodiment of the system is shown.

[0039] Figure 5 A second embodiment of the system is shown.

[0040] In the drawings, the same reference numerals are assigned to the same or similar elements.

[0041] Figure 6 Examples of measurement signals of a breathing signal and a pre-characteristic signal are shown.

[0042] Figure 7 An example illustration of some steps of feature extraction is shown. DETAILED DESCRIPTION

[0043] Other characteristics and advantages of the invention will emerge from the following non-limiting description, made with reference to the accompanying drawings and examples.

[0044] The system, method and computer program according to the present invention detect the cardiac health state (cardiac health state) of the user 10. The terms body and heart refer to the body and heart of the user 10 whose cardiac health state is detected. The user is usually a human being. However, the user can also be an animal, preferably a mammal. The user 10 is called a user, not a patient, because the system, method and computer program can be performed by anyone and does not require medical monitoring. Obviously, this does not exclude that medical personnel can also perform the system, method and computer program according to the present invention on the user 10 (in this case, the patient). The user 10 should always be the person whose cardiac health state needs to be detected. The person who controls the system, method and computer program according to the present invention is preferably the same user, but can also be another person.

[0045] Figure 1 An embodiment of a system for detecting a user's heart health status according to the present invention is shown. The system comprises an inertial measurement unit (IMU) 1 and a processing device 2 .

[0046] IMU 1 is configured to measure a measurement signal of cardiac motion / movement. The measurement signal preferably includes at least six dimensions. The six dimensions include three translational dimensions (also referred to as linear dimensions) and three rotational dimensions (also referred to as angular dimensions). IMU 1 typically includes an accelerometer for measuring cardiac motion acceleration (three linear dimensions) and a gyroscope for measuring angular velocity (three rotational dimensions). IMU 1 includes an accelerometer and a gyroscope typically in a common chip or device. However, the gyroscope and accelerometer may also be implemented as separate chips or devices and will still be considered as IMU 1. IMU 1 is preferably placed on the chest of user 10 to measure the measurement signal, which is typically referred to as SCG. Typically SCG refers only to three linear acceleration measurements, or even only to one linear and / or rotational dimension. However, in this article, SCG should refer to a six-dimensional measurement signal including three linear dimensions and three rotational dimensions. Typically, one, two, or three of the three linear dimensions are referred to as SCG, and one, two, or three of the three rotational dimensions are referred to as gyrocardiograms (GCG). For simplicity, the six-dimensional measurement signal with SCG and GCG is referred to as SCG in this article.

[0047] In a preferred embodiment, the IMU 1 is arranged in a smartphone 5, such as Figure 4 Thus, the standard IMU 1 available in almost every smartphone 4 can be used to detect the cardiac measurement signals required by the present invention. No other sensors or additional sensors are required to perform cardiac measurements.

[0048] However, it is also possible that the IMU 1 is not arranged in the smartphone 5. For example, the IMU 1 can be arranged in a wearable device 7 worn on the user's body, such as a smart watch, a heart monitor fixed on the chest with a sleeve around the torso, or other wearable devices. Such a wearable device can preferably be connected to the smartphone 5 via a wireless connection (e.g., a Bluetooth connection). For example, the IMU 1 can be arranged in a heart monitor that is fixed on the chest of the user 10 with a sleeve.

[0049] It is also possible that the IMU 1 is arranged in a dedicated device which is not associated with the smartphone 5 .

[0050] The smartphone 5 preferably comprises a smartphone processor, storage means, communication means, input means (preferably a touch screen) and output means (preferably a (touch) screen). The storage means of the smartphone 5 is configured to store application programs so that when the application programs are executed on the smartphone processor, specific functions are performed on the smartphone 5. Such application programs are also generally referred to as applications (apps). Figure 4 and Figure 5The smartphone 5 in the embodiment of the present invention preferably includes an application (hereinafter referred to as a monitoring application) for executing or requesting at least some steps of the method of the present invention described later. The monitoring application is preferably installed on the smartphone 5, i.e. stored in a storage device of the smartphone 5. However, the term "monitoring application" should also refer to a front-end program that is downloaded each time a website is accessed using a browser of the smartphone. Preferably, the monitoring application is configured to request measurement of a measurement signal from the IMU 1 and to obtain the measurement signal from the IMU 1 (arranged in a manner similar to the IMU 1). Figure 4 in the same smartphone 5 or connect to Figure 5 The IMU 1 of the smartphone 5 in the embodiment of the present invention receives the measurement signal. Preferably, the monitoring application is configured to communicate with the server 6. The communication device is configured to establish a communication connection to the server 6. The communication device may be a WLAN and / or cellular telephone network communication module. The cellular telephone network communication module preferably includes a user identification module (virtual user identification module (eSIM), SIM card or holder of such a SIM card) for receiving data required for communication with the cellular telephone network.

[0051] The processing device 2 is configured to perform all or at least some / most of the steps (S2 to S4) of the present invention described in more detail below.

[0052] The processing device 2 may be a single processor or may be a combination of multiple processors. For example, different processors may implement different steps. Preferably, the processing device 2 may include a local processor and a remote processor. The local processor may be arranged in a smartphone 5 (smartphone processor) and / or a wearable device 7 or the above-mentioned dedicated device. The remote processor may be arranged in a server 6 connected to the local processor (or a device including a local processor) via the Internet, for example. The remote processor should not be understood as only a physical processor, but anything that implements the functions of the remote processor or server 6, such as a server center or a cloud. The steps from S2 to S4 described below may all be implemented on a local processor, or all on a remote processor, or distributed on both. Preferably, the processing device includes a general-purpose processor (or multiple general-purpose processors), on which software is run to implement the functions of the method of the present invention described later. The software may implement all functions or only some functions. If there is a local processor, preferably there is local software running on the local processor, such as a monitoring application. If there is a remote processor, preferably there is remote software running on the remote software, such as a server or backend software.

[0053] The processing device 2 includes a feature extractor 3 and an ML engine 4 .

[0054] The feature extractor 3 is configured to implement step S3 described later to extract features from the measurement signal for the ML engine 4. The feature extractor 3 may be implemented in a local processor / smartphone 5 or a remote processor / server 6 or both. Preferably, the processing means comprises a general purpose processor (or multiple general purpose processors) on which software is run to implement the functionality of the feature extractor 3, such as a monitoring application and / or server software.

[0055] The ML engine 4 is configured to implement step S4 described later. The ML engine 4 can be implemented on a general purpose processor or a dedicated ML processor. The general purpose processor can be the same processor as the one running the feature extractor 3. In this case, the ML engine 4 is implemented as software. However, the ML engine 4 can also be implemented on a dedicated chip or processor dedicated to ML applications. Such a ML dedicated chip would then be programmed for the present ML application as described below. The ML engine 4 is preferably arranged on a remote processor / server 6. However, it is also possible to have the ML engine 4 on a local processor / smartphone 5.

[0056] Figure 2 An embodiment of a method for detecting a user's heart health status according to the present invention is shown.

[0057] In a first step S1, a measurement signal is measured using an IMU 1 placed on the body of a user 10. The measurement signal is preferably an SCG. The IMU 1 is placed on the user's body so that the IMU 1 can detect the movement of the heart. Preferably, the IMU 1 is placed on the torso, preferably on the upper part of the torso and / or preferably on the chest side of the torso, most preferably on the chest of the user 10. Preferably, the user 10 lies down for the measurement and the IMU 1 (preferably in the form of a smartphone 5) is placed on the torso. Since the user 10 lies down, the IMU 1 or the smartphone 5 does not have to be fixed to the torso and can freely measure the movement generated by the heart. The sleeve or any other device for pushing the IMU 1 against the torso may suppress the movement of the IMU 1. In addition, in this way, the user 10 is in a relaxed state and the measurement results are not tampered with by other body movements that may tamper with the measurement results. The user 10 can start the measurement by providing an input command to the monitoring application / smartphone 5 / IMU 1 and perform the measurement himself by placing the smartphone 5 / IMU 1 on his torso / chest (in any order). Therefore, the measurement can be performed with any ordinary smartphone 5 including an IMU 1. The user only needs to download a specific monitoring application on the smartphone 5. Preferably, the measurement lasts less than a maximum length, preferably less than 10 minutes, preferably less than 5 minutes, preferably less than 2 minutes, preferably less than 1 minute. Therefore, the user 10 can quickly implement the measurement step. The measurement lasts preferably longer than a minimum length, which is preferably longer than one heartbeat, preferably longer than 1 second, preferably longer than 5 seconds, preferably longer than 10 seconds, preferably longer than 20 seconds. Therefore, the IMU 1 records the measurement signal SCG of a certain measurement window. The measurement window is preferably between the minimum length and the maximum length. However, other lengths are obviously possible. Since the IMU 1 / smartphone 5 is placed on the torso of the user 10, it moves with the movement of the heart or heartbeat. The measurement signal SCG therefore represents the heart movement in a measurement window of six dimensions (degrees of freedom). The measurement signal SCG is measured at a certain sampling rate. The sampling rate is preferably higher than 10 Hz, preferably higher than 25 Hz, preferably higher than 50 Hz, preferably higher than 60 Hz, preferably higher than 70 Hz, preferably higher than 80 Hz, preferably higher than 90 Hz, preferably higher than 100 Hz.

[0058] Step S1 is in fact optional, and the method or computer program of the present invention may start with step 2 of receiving the measurement signal SCG described in step S1. If step S1 is included in the method or computer program, the monitoring application / smartphone 5 / dedicated device preferably receives a user input to start measuring with the IMU 1. Once this user input is received, the monitoring application / smartphone 5 / dedicated device sends an instruction to the IMU 1 to start measuring the signal SCG. The duration of the measurement is preferably predetermined or can be configured in the monitoring application. When the duration of the measurement has ended, the monitoring application / smartphone 5 / dedicated device preferably gives an indication that the measurement has ended, such as an audio signal.

[0059] In step S2 , the system / method / computer program / monitoring application / smartphone 5 / processing device 2 receives a measurement signal SCG from the IMU 1 .

[0060] In step S3, the processing device 2 determines a plurality of features based on the measurement signal SCG. These features are used in the next step as input features (also called feature vectors) for the ML engine 4. These features preferably include data features derived from the measurement signal SCG and subject features associated with the user 10.

[0061] The subject characteristics are typically independent of the measurement signal SCG. The subject characteristics preferably include one or more of the age, sex, weight, height, body mass index (BMI), body surface area (BSA) of the user 10. The subject characteristics are preferably derived from input information, preferably from user input. The user input can be received, for example, at a monitoring application in the smartphone 5. However, the input information can also be stored in a database on the server 6 for use by the user and retrieved therefrom. For example, the information can be received from a parallel application that already has the information and allows transmission. BMI and BSA can be calculated based on the weight and height of the user 10. Even if the subject characteristics are preferably present in the characteristics of the ML engine 4, step S4 can be performed without the subject characteristics or with different subject characteristics.

[0062] In a preferred embodiment, the data features are derived (only) from the measurement signal SCG measured with the IMU 1. Preferably, the data features are derived only from the measurement signal SCG measured with the IMU 1, i.e. the IMU 1 is independent of other sensors arranged on the body, in particular independent of sensors not included in a standard smartphone, in particular independent of ECG measurements which usually require specific equipment. However, the present invention can also be applied to more complex embodiments comprising ECG sensors and / or other sensors, such as a specific wearable device 7 having an IMU sensor 1 and an ECG sensor. The wearable device can be connected to a smartphone 5. The wearable device can be worn by means of a chest strap, or can simply be placed or fixed on the chest of the user 10 for measurements.

[0063] Figure 3 Preferred embodiments of determining data features based on a measurement signal SCG (hereinafter referred to as SCG for short) are described.The data features may be obtained for a complete measurement window of the SCG signal or only for a sub-window of the measurement window.

[0064] In step S31, the breathing information is retrieved from the SCG. Since breathing causes an increase and decrease in the volume of the trunk or chest / thoracic cavity, the SCG also reflects the breathing information. Preferably, a breathing signal is extracted from the SCG. This can be achieved by independent component analysis, a low-pass filter or other measures. The breathing signal is preferably determined based on a complete measurement window. The breathing information is preferably based on the breathing signal. Preferably, the breathing signal is discretized into multiple breathing states. The breathing state preferably includes at least four breathing states. The breathing state preferably includes two or more of the following: a high volume state (HV), a medium volume state during exhalation (MV-EX), a low volume state (LV) and a medium volume state during inspiration (MV-IN). There may also be more than four breathing states. The high volume state HV can be divided into two states (high volume HV-EX during exhalation and high volume LV-IN during inspiration) or three states (HV-EX, HV, HV-IN). This can also be done for the low volume state LV. Similarly, the medium volume state MV can be divided into smaller states. Thus, one respiratory cycle / period comprises a sequence of HV, a subsequent MV-EX, a subsequent LV and a subsequent MV-IN (or other complete sequence of respiratory states of a respiratory cycle / period). Discretization of the respiratory signal results in a respiratory state signal showing that the respiratory state changes over time (in the measurement window). Discretization of the respiratory signal can be performed only on the complete measurement window or only on the subwindow of interest. These features preferably include at least one respiratory information, preferably the respiratory state at at least one time point of the heartbeat. The respiratory signal / information can be determined from all six dimensions of the SCG, from subsets of the SCG dimensions or from one or more composite signals, each composite signal being determined from one or more (preferably different) subsets of the SCG signal. Figure 6 An example of a breathing signal 23 extracted from an example SCG is shown.

[0065] In step S32, preprocessing of the SCG signal is performed. This mainly includes removing noise, that is, removing signal components that are not related to cardiac motion. This can be achieved, for example, by a frequency filter (such as a high-pass filter or a low-pass filter or a band-pass filter). The preprocessing step preferably removes frequencies above 100 Hz, preferably above 80 Hz, preferably above 70 Hz, preferably above 60 Hz. The preprocessing step preferably removes frequencies below 1 Hz, preferably below 2 Hz, preferably below 3 Hz. Preferably, the preprocessing also removes any trends in the SCG. Preprocessing is a preferred step, but is not essential to the present invention. The respiratory signal / information is preferably determined based on the SCG before preprocessing or using a different preprocessing that excludes the removal of frequencies below the normal heart rate. Figure 6 shows a having three linear dimensions a x (t), a y(t), a z (t) and the three rotation dimensions ω x (t),ω y (t),ω z (t) An example of a preprocessed six-dimensional SCG signal. Signal a x (t) shows the acceleration in the x direction, signal a y (t) shows the acceleration in the y direction, signal a z (t) shows the acceleration in the z direction. Signal ω x (t) shows the rotation speed around the x direction, signal ω y (t) shows the rotation speed around the y direction, signal ω z (t) shows the rotation speed around the z direction. The x direction, y direction and z direction refer to the three vertical dimensions of the Cartesian coordinate system of the IMU 1. Preferably, the coordinate system of the IMU 1 is aligned with the coordinate system of the user 10. The user's coordinate system preferably has a y direction in the direction from the user's feet to the head, an x ​​direction in the direction from one arm to the other arm (preferably from the right arm to the left arm), and a z direction from the back of the torso to the chest. This alignment can be achieved by giving the user instructions on how to place the smartphone 5 on the body. For example, the instructions may display an arrow pointing to the head of the user 10. However, it is also possible that the arrangement of the IMU 1 relative to the user 10 is not known.

[0066] In step S33, at least one pre-characteristic signal is calculated based on the (preferably pre-processed) SCG signal. In a preferred embodiment, based on the three linear dimensions a of the SCG x (t), a y (t), a z (t)(independent of the three rotation dimensions ω of the SCG x (t),ω y (t),ω z (t)) determines at least one (scalar) linear pre-signal and / or three rotational dimensions ω based on SCG x (t),ω y (t),ω z (t)(Independent of the three linear dimensions of SCG a x (t), a y (t), a z (t)) determine at least one (scalar) rotation pre-signature signal. The (linear and / or rotational) pre-signature preferably corresponds to a physical variable describing the heart motion, such as kinetic energy, kinetic power, work, force / torque, etc. The physical value describing the linear motion of the heart is, for example,

[0067]

[0068]

[0069]

[0070]

[0071] Vector a is a three-dimensional vector describing the linear acceleration at each sampling time, m is the weight (mass) of the device including the IMU 1, and vector v is a three-dimensional vector describing the linear velocity of the heart at the sampling time. The arrows above the variables represent vectors. As shown in equation (1), the force vector F(t) at each sampling time t is obtained by multiplying the mass m by the acceleration vector a(t) at each sampling time t = (F x (t), F y (t), F z (t))=(ma x (t), ma y (t), ma z The acceleration vector a(t) corresponds to the acceleration vector a(t) measured by the IMU 1 (and preferably preprocessed in S32) = (a x (t), a y (t), a z (t)). Based on the mass multiplied by the square of the linear velocity v at sampling time t 2 (t) to calculate the linear kinetic energy K at sampling time t lin (t), as shown in equation (2). The square of the linear velocity v 2 (t) is preferably based on the velocity vector v at sampling time t = (v x (t), v y (t), v z The distance / length of the velocity vector is preferably the square root of the sum of the squares of the three linear components of the velocity. That is, the Euclidean distance metric of the velocity vector v. However, other distance metrics can be used. Velocity vector v = (v x (t), v y (t),u z (t)) can be obtained by calculating the acceleration vector a(t)=(a(t)) from t0 (e.g. the start of the measurement signal) until sampling time t. x (t), a y (t), a z (t)) is integrated to obtain the linear velocity signal. Preferably, the linear velocity signal is calculated based on the linear acceleration vector signal (by integration). The linear velocity signal may be a signal having three linear velocity components (v x (t), v y (t), vz (t)) or the length of the linear velocity vector or a linear velocity vector signal of a function thereof. The linear velocity signal can be used to calculate a linear pre-characteristic signal and / or (data) feature, such as linear kinetic energy. The linear kinetic energy K lin (t) is distinguished from the signal energy often used in the prior art. For example, the signal energy of the SCG in the x direction is defined as the acceleration on the measured signal The sum of and the linear kinetic energy K of the heart at sampling time t lin (t) is irrelevant. The linear work of the heart during the period between t1 and t2 is W lin (t1, t2) is the force vector F(t) at time t and the path change vector d at time t s The linear kinetic power dP at sampling time t can be calculated by the scalar product of the velocity vector v(t) at sampling time t and the force vector F(t) at sampling time t: lin (t)(or P for short lin (t))P lin (t) = v x (t)F x (t)+u y (t)F y (t)+u z (t)F z (t). As is known, these physical variables can be obtained in different ways without departing from the invention.

[0072] Physical values ​​describing the rotational motion of the heart are, for example,

[0073]

[0074]

[0075]

[0076]

[0077] The vector α(t) is a 3D vector describing the rotation / angular acceleration at each sampling time, I is the moment of inertia of the device containing the IMU 1, and the vector ω(t) is the 3D vector of the rotation / angular velocity of the heart at sampling time t. More precisely, the moment of inertia I is a 3×3 tensor / matrix, and the rotational kinetic energy is calculated as 0.5ω(t)Iω(t) T The superscript T denotes the transpose of a vector. The moment of inertia tensor I can be simplified to the three principal moments of inertia I xx ,I yy ,I zzis the diagonal matrix of the diagonal, I xx ,I yy ,I zz Moments of inertia along the three principal axes of the coordinate system of the device containing the IMU 1, in the x-, y- and z-directions. The formula is written in the form of vectors and tensors, but can also be written in different forms without changing the protection range, for example, in the case of a tensor in diagonal form, such as K rot =0.5*(I xx *ω x (t) 2 +I yy *ω y (t) 2 +I zz *ω z (t) 2 ). If the device has the same moment of inertia in all three directions, the formula can be simplified to I being the scalar value of the moment of inertia. Obviously, the tensor can also be more complex, with non-diagonal elements not equal to zero. As shown in equation (5), the torque vector τ(t) at each sampling time t is obtained by multiplying the moment of inertia I by the rotational acceleration vector α(t) at the corresponding sampling time t. The angular acceleration vector α(t) corresponds to the first-order time derivative of the measured (and preferably preprocessed) angular velocity vector ω(t), where the vector α(t) = (d x ω(t) / dt, d y ω(t) / dt, d z ω(t) / dt). The angular / rotational acceleration signal is a signal in which the acceleration vector, one of its components, or a combination of its components (e.g., its length) changes with time. The rotational acceleration signal can be used to calculate the rotational pre-signature signal. The rotational kinetic energy K at sampling time t rot (t) is based on the moment of inertia (of the device containing the IMU 1) multiplied by the square of the rotational velocity at sampling time t, ω 2 (t) is calculated as shown in equation (6). The square of the rotation speed v 2 (t) is preferably based on the rotational velocity vector ω=(ω at sampling time t x (t),ω y (t),ω z The heart rotation work W in the time period between t1 and t2 is calculated as the square of the distance / length of the heart. rot (t1, t2) is the integral of the scalar product of the torque vector τ(t) at time t and the angle change vector dθ(t) at time t over the time t between t1 and t2 (see equation (3)). rot (t)(or P for short rot(t)) can be calculated by the scalar product of the angular velocity vector ω(t) at sampling time t and the linear force vector F(t) at sampling time t, P rot (t) = ω x (t)F x (t)+ω y (t)F y (t)+ω z (t)F z (t). As is known, these physical variables can be obtained in different ways without departing from the invention.

[0078] The measurement or preprocessing step S32 of the IMU 1 provides the following two measurement sub-signals: a three-dimensional linear acceleration vector signal a(t) and a three-dimensional angular velocity vector signal ω(t). Preferably, in step S33, at least one composite linear dynamic pre-signal signal is calculated based on the three-dimensional linear acceleration vector signal a(t) and / or at least one composite rotational dynamic pre-signal signal is calculated based on the three-dimensional rotational velocity vector signal ω(t). The at least one composite linear dynamic pre-signal signal is preferably a scalar value combining the three linear dimensions of the measurement signal into one scalar value, and / or the at least one composite rotational dynamic pre-signal signal is preferably a scalar value combining the three rotational dimensions of the measurement signal into one scalar value, so that when the measurement signal is measured, the composite value is independent of the direction of the IMU 1. The at least one composite linear pre-signal signal preferably includes a linear kinetic energy signal K lin (t) and / or linear dynamic power signal P lin (t). A composite linear pre-characteristic signal and / or linear kinetic energy signal K lin (t) is preferably calculated based on the linear velocity vector v(t). The at least one composite rotation pre-signature signal preferably includes a rotational kinetic energy signal K rot (t) and / or rotational power signal P rot (t). A composite linear pre-characteristic signal and / or a rotary power signal P rot (t) is preferably calculated based on the rotational acceleration vector α(t).

[0079] The movement of the heart is approximated in this article as the movement of the IMU 1. Therefore, the physical variables of the heart movement ultimately correspond to the (preferably preprocessed) movement of the IMU 1 or the device containing the IMU 1. The value / signal of the linear or rotational kinetic energy and the value of the linear or rotational kinetic power depend on the device in which the IMU 1 is arranged, for example, on its moment of inertia I and / or its mass m. If this is a dedicated device that is always the same, this is just a constant and does not play a big role. However, when using a smartphone 5 or other wearable device, the correct mass and / or moment of inertia must be retrieved to calculate the pre-characteristic signal, characteristic, lin / rot kinetic energy and / or lin / rot kinetic power. In this case, the processing device, preferably a monitoring application, retrieves the type of device including the IMU 1 from the operating system of the smartphone 5 or the wearable device. The information of the smartphone 5 about the type of smartphone 5 or wearable device 7 can also be input by the user, instead of automatically obtaining the type of smartphone 5 / wearable device 7 from the operating system of the smartphone 5. The processing device is capable of storing relevant information for calculating / calibrating characteristics for each type of smartphone 5 or wearable device 7, preferably the moment of inertia and / or mass of the smartphone 5 / wearable device 7. The type of smartphone 5 / wearable device 7 preferably refers to the model of smartphone 5 / wearable device 7. Based on the type of device, the mass and / or moment of inertia of the device can be retrieved. Thus, the kinetic energy and power are calibrated using the device characteristics for measuring the cardiac motion, thereby avoiding erroneous results based on the use of different measuring devices. Subsequently, the term kinetic energy signal K(t) refers to the two-dimensional signal K(t)=(K lin (t), K rot (t)), which contains as a first dimension or component the linear kinetic energy signal K lin (t) and the rotational kinetic energy signal K as the second dimension or component rot (t), or refers to one of the two components alone. Subsequently, the term dynamic power signal P(t) refers to the two-dimensional signal P(t) = (P lin (t), P rot (t)), which contains as a first dimension or component the linear dynamic power signal P lin (t) and the rotational power signal P as the second dimension or component rot (t), or refers to one of the two components alone.

[0080] In step S34, at least one heartbeat is identified. Preferably, multiple heartbeats (e.g., all) are identified in the SCG. If multiple heartbeats have been identified, at least one heartbeat is selected for further analysis. A heartbeat is a time window associated with one heartbeat of the heart. A heartbeat is a periodic cardiac cycle of the heart. A heartbeat is usually defined relative to a characteristic point in the heartbeat (usually R or Q in the ECG signal). Because the present invention is preferably implemented without any ECG signal, at least one heartbeat is identified based on the SCG, preferably based on at least one pre-characteristic signal calculated in step S33, preferably based on a kinetic energy signal and / or a kinetic power signal, or a combination thereof. In a preferred embodiment, the (data) feature is calculated based on the SCG (or pre-characteristic signal) of one heartbeat (i.e., one cardiac cycle). This will also include the case where a representative heartbeat is calculated by averaging multiple heartbeats (preferably multiple heartbeats of the same respiratory state). In a preferred embodiment, the heart health state is determined based on multiple heartbeats by selecting multiple heartbeats and detecting the health state of each heartbeat in steps S34, S35, and S4. The composite heart health state can then be calculated by combining the heart health states from a single heartbeat. The prior art generally does not calculate feature vectors based on a single heartbeat, which reduces the robustness of the method. Some prior art documents propose to calculate feature vectors based on data from one heartbeat, but these methods always require ECG signals to detect heartbeats. This is one of the innovations of the present invention, that is, heartbeats can be identified from SCG (without using ECG signals). The heartbeat cycle is determined based on heartbeat characteristics in SCG and / or pre-feature signals (e.g., kinetic energy or power) obtained from SCG. Although the heartbeats in SCG and pre-feature signals lack such clear characteristics in ECG signals, robust detection of heartbeat cycles is difficult. A preferred embodiment for detecting heartbeats in SCG is subsequently described.

[0081] It is proposed to detect the heartbeat cycle in SCG (also in the pre-feature signal determined from SCG) based on a matrix profile. The matrix profile m(t1) of the signal s(t) at the sampling time t1 shows the distance from s(t1) to its nearest neighbor in the signal s(t) in the complete time window of the signal s(t). The distance is preferably the Euclidean distance metric, but other distance metrics can obviously also be used. The matrix profile m(t1) allows similar motifs to be found in the signal s(t). Since each person's heartbeat looks a little different, this allows repeated motifs of the heartbeat to be found in the signal used to determine the matrix profile m(t). Preferably, the matrix profile m(t) is calculated based on SCG (including calculating the matrix profile m(t) based on the (pre-feature) signal calculated from SCG), and the heartbeat cycle is determined very quickly and with high robustness. The matrix profile m(t) provides not only the location of the motif, i.e., the regular heartbeat, but also the data anomalies called conflicts and the quality of each motif. The quality of each motif is a similarity measure of the most common motif in the signal s(t). Preferably, the multidimensional signal s(t) is used to detect the motif / heartbeat cycle. More details of examples of algorithms based on motifs in multidimensional signals based on matrix profiles can be found in the article "Meaningful Multidimensional Motif Discovery" written by Chin-ChiaMichael Yeh, NickolasKavantzas, Eamonn Keogh, published in ICDM 2017, which is incorporated herein by reference as exemplary details of matrix profiles. The motif length is preferably set to a value less than 1 second, preferably less than 800 milliseconds (ms), preferably less than 700ms. The motif length is preferably set to a value greater than 200ms, preferably greater than 400ms, preferably greater than 500ms. The motif length is preferably set to a value less than 1800ms, preferably less than 1500ms, preferably less than 1000ms, preferably less than 800ms. Figure 7 It shows that in step S34, the linear kinetic energy K is used lin (t) and rotational kinetic energy K rot (t) is an example of a motif 22 detected by the matrix profile. In the embodiment, one motif 22 is detected. However, as Figure 7 As shown, two or more motifs 22 can also be detected in the basic signal.

[0082] The heartbeat cycle is preferably identified based on a pre-signature signal, preferably based on a kinetic energy and / or power signal.

[0083] In step S35, a sub-period 24 of a heartbeat is identified in the identified heartbeat cycle. Preferably, the sub-period 24 is determined based on a characteristic point of the SCG in the heartbeat cycle (also including the meaning of the pre-characteristic signal determined based on the SCG). Preferably, the characteristic point is the maximum rotational kinetic energy or the maximum rotational kinetic power in the heartbeat. Preferably, the sub-period 24 is determined based on the time of the characteristic or reference point and the fixed time relationship / width between the sub-periods 24 of the heartbeat cycle. The fixed time relationship / width of the sub-period 24 means that the time width of each sub-period 24 does not change from one heartbeat to the next. This is not the case if the start and end points of the sub-period are defined by the physiological time points that need to be detected for each heartbeat. The fixed time relationship / width can be an absolute fixed time relationship / width or a relatively fixed time relationship / width. The absolutely fixed time relationship means that the width of the sub-period 24 and the time distance between them have a fixed absolute time relationship (fixed absolute time value), and only the position of the sub-period 24 and the fixed absolute time relationship between them are correctly located based on the time of the characteristic point. Relatively fixed time relationship refers to that the width of sub-period 24 and / or the time distance between them all have fixed relative time relationship (fixed percentage). Therefore, in addition to the correct positioning of sub-period 24, the ratio must also be adjusted. For example, the ratio can depend on the heart rate or the length of the selected heartbeat. Sub-period 24 preferably has the same (absolute or relative) width. The width of each sub-period 24 is preferably less than 200 milliseconds, preferably less than 150 milliseconds, preferably less than 140 milliseconds, preferably less than 130 milliseconds, preferably less than 120 milliseconds, preferably less than 110 milliseconds, preferably less than 105 milliseconds, and / or preferably greater than 20 milliseconds, preferably greater than 40 milliseconds, preferably greater than 50 milliseconds, preferably greater than 60 milliseconds, preferably greater than 70 milliseconds, preferably greater than 80 milliseconds, preferably greater than 90 milliseconds, preferably greater than 95 milliseconds. Preferably, the width of each sub-period 24 is 100 milliseconds. Preferably, the width of each sub-period 24 is preferably less than 30%, preferably less than 25%, preferably less than 20%, preferably less than 15%, preferably less than 10%, and / or preferably greater than 2%, preferably greater than 5%, preferably greater than 7%, preferably greater than 8%, preferably greater than 9% of the heartbeat period / length. The sub-periods 24 preferably do not overlap, i.e. there are no sampling times that occur in two sub-periods 24. The sub-periods 24 are preferably arranged one after another, so that there are no sampling times between two sub-periods 24 that do not belong to any sub-period 24. Preferably, more than 5, preferably more than 6, preferably more than 7, preferably more than 8, preferably more than 9, preferably more than 10 sub-periods 24 are identified. Preferably, less than 20, preferably less than 15, preferably less than 14, preferably less than 13, preferably less than 12 sub-periods 24 are identified. Preferably, 11 sub-periods 24 are identified. Preferably, the first sub-period 24 is arranged to cover the characteristic point.A first group of sub-periods 24 is arranged before a first sub-period 24 and a (different) second group of sub-periods 24 is arranged after the first sub-period 24. The first group preferably comprises less than half of the sub-periods 24 and / or the second group preferably comprises more than half of the sub-periods 24. Figure 7 An example of step S35 is shown, for an arrangement of 11 sub-cycles 24 per 100 ms, wherein four sub-cycles 24 are before the characteristic point, one is at the characteristic point, and six are after the characteristic point. The cumulative period of all sub-cycles 24 may be longer or shorter than the heartbeat cycle, i.e. may have extended to the previous and / or next heartbeat cycle, or may not cover the entire heartbeat cycle.

[0084] In step S36, (data) features are calculated (used as input in the ML engine 4). The (data) features preferably include at least one feature K for each sub-period 24 i, where i represents the index of the corresponding sub-cycle 24 from sub-cycle i=1 to sub-cycle i=n, ​​where n is the number of sub-cycles in the cardiac cycle. Preferably, the same physical (dynamic) variable is calculated for each sub-cycle. Preferably, the data features include one or more of linear kinetic energy, rotational kinetic energy, linear kinetic power, rotational kinetic power, respiratory information or state, rotational cardiac work, linear cardiac work for each corresponding sub-cycle. The linear / rotational kinetic energy of a sub-cycle can be one or more of the cumulative linear / rotational kinetic energy in the sub-cycle, the linear / rotational kinetic energy density in the sub-cycle, the maximum value of the linear / rotational kinetic energy in the sub-cycle, the time sample of the maximum value of the linear / rotational kinetic energy in the sub-cycle, the minimum value of the linear / rotational kinetic energy in the sub-cycle, the time sample of the minimum value of the linear / rotational kinetic energy in the sub-cycle, or other values ​​(such as the average value, median, variance, etc. of the linear / rotational kinetic energy in the sub-cycle). The cumulative linear / rotational energy of a sub-cycle means the integral of the linear / rotational energy over the sub-cycle. The linear / rotational kinetic energy density in a sub-cycle is the cumulative linear / rotational kinetic energy in the sub-cycle divided by the width of the sub-cycle (its duration). The linear / rotational kinetic power of a sub-cycle can be one or more of the cumulative linear / rotational kinetic power in the sub-cycle, the maximum value of the linear / rotational kinetic power in the sub-cycle, the time sample of the maximum value of the linear / rotational kinetic power in the sub-cycle, the minimum value of the linear / rotational kinetic power in the sub-cycle, the time sample of the minimum value of the linear / rotational kinetic power in the sub-cycle, or other values ​​(such as the average, median, variance, etc. of the linear / rotational kinetic power in the sub-cycle). The cumulative linear / rotational kinetic power of a sub-cycle means the integral of the linear / rotational power over the sub-cycle. The linear / rotational kinetic power density in a sub-cycle is the cumulative linear / rotational kinetic power in the sub-cycle divided by the width of the sub-cycle (its duration). For simplicity, linear / rotational in this paragraph means that this applies to linear variables such as linear kinetic energy, linear kinetic power, etc. and / or rotational variables such as rotational kinetic energy, rotational kinetic power, etc. These (data) features are calculated based on the SCG signal in the cardiac cycle (or the corresponding sub-cycle), preferably based on the pre-feature signal in the cardiac cycle (or the corresponding sub-cycle), preferably based on the linear / rotational kinetic energy signal in the cardiac cycle (or the corresponding sub-cycle) and / or based on the linear / rotational kinetic power signal in the cardiac cycle (or the corresponding sub-cycle). Step S33 for calculating the pre-feature signal can also be implemented later, i.e. after steps S33 and S35, if these steps are not needed. Step S33 can also be omitted and directly included in step S36 for calculating the (data) features.

[0085] Based on the respiratory signal or respiratory state in the sub-cycle determined in step S31, the respiratory state of each sub-cycle is calculated. The respiratory information / state can also be input only once for the entire heartbeat, rather than for each sub-cycle. In this case, it is preferred to select a heartbeat cycle that completely corresponds to (only) one respiratory state.

[0086] The (data) features may also include the motif quality obtained in step S34. This provides a deviation from the normal heartbeat pattern of the user 10 and may be used as another input feature of interest to the ML engine 4.

[0087] Figure 3 Preferred embodiments for determining features of the data fed to the ML engine 4 are described. Obviously, alternative embodiments of (data) feature extraction are possible. For example, a sub-period may extend over multiple heartbeats (two or more). For example, different feature variables are calculated for different sub-periods. The start and end points of a sub-period may be detected based on characteristic points in the data. One, some or all features may be calculated for a complete heartbeat cycle instead of a sub-period. Many alternatives are possible.

[0088] Back to Figure 2 , the features determined in step S3 are fed into the ML engine 4 of the processing device 2. The ML engine 4 processes the feature input and outputs the heart health state of the user 10. The heart health state preferably refers to a medical indication / problem of the heart. In a preferred embodiment, the heart health state refers to CHF. However, the method discussed can also be used to detect other medical indications / problems of the heart, such as atrial fibrillation, valvular disease, arrhythmia, heart murmur, myocardial infarction, drug-induced cardiotoxicity. Preferably, the heart health state is binary, that is, the heart health state corresponds to a first heart health state indicating that the heart of the user 10 is not suffering from a medical indication / problem (such as CHF), or corresponds to a second heart health state indicating that the heart of the user 10 is actually suffering from or has a risk of suffering from a medical indication / problem (such as CHF). In addition to the binary output of the ML engine 4, a heart health state with more than two states can also be output. The second heart health state can be distinguished, for example, into different stages of the medical indication / problem or different risk ranges of the medical indication / problem and / or also distinguished into different medical indications / problems. For CHF, the second heart health state may be subdivided into different second sub-states indicating the stage of CHF, such as stage A, stage B, stage C, stage D. Thus, the output of the ML engine 4 will indicate not only the presence of CHF, but also its stage. It is also possible that the second heart health state may additionally give a score indicating the probability of having the medical indication / problem.

[0089] ML Engine 4 can follow simple ML methods like random forest models or more complex ML methods.

[0090] The ML engine 4 is preferably trained by a supervised training model. The ML engine 4 can operate in a detection mode and a training mode.

[0091] In the detection mode, the plurality of features calculated for the user 10 are input to the ML engine 4 (without prior knowledge of the heart health status of the user 10), and the heart health status of the user 10 is output by the trained ML engine 4. At a higher level, the IMU measurement signals of the user 10 (whose heart health status is unknown and needs to be estimated) are used to calculate the plurality of features of the user 10, which are input to the ML engine 4. Preferably, the plurality of features are calculated based on (only) the IMU measurement signals and are input to the trained ML engine 4.

[0092] In the training mode, a plurality of features calculated for the user 10 (the same features as used in the detection mode) are input into the ML engine 4 together with the heart health state of the user 10, i.e., a plurality of features are input into the ML engine 4, which are annotated with the actual health state of the heart of the user 10. The heart health state of the training user 10 annotated with the plurality of features is used for training and can be determined before the doctor's diagnosis. The diagnosis can be based only on the same plurality of features, but preferably on additional measurements, such as ECG, echocardiogram, cardiovascular magnetic resonance imaging, phonocardiogram, electrocardiogram, MRT, CT, etc. Preferably, the additional measurements also depend on the health state of the heart to be assessed. For example, if the health state of the cardiac output is about the health state of the user's congestive heart failure, the diagnosis of the user 1 can be based on ECG, blood pressure measurement and SCG measurement. Other heart problems may require other additional measurements. Obviously, the heart state of the user used for training can also be estimated based on a computerized estimate (rather than a doctor), but based on a more complete set of data that provides a very high certainty of the estimated heart state input by the user in the ML engine 4 for training. Training is performed on a plurality of users 10 of at least two types. A first type of user has a first health state, and a second type of user has a second health state. To train the ML engine 4, the same features are typically used as inputs to the ML engine 4 for a known output of the user 10, i.e., a heart health state. This known heart health state is also input into the ML engine 4 along with the features of the user 10 to train the ML engine 4. This is done for many users 10 to obtain a well-trained ML engine 4.

[0093] Preferably, steps S3 and S4 are repeated for multiple heartbeats so as to estimate the cardiac health state for multiple heartbeats. The results of the cardiac health state for multiple heartbeats can be combined to obtain a combined cardiac health state. The combination can, for example, count the number of first cardiac health states and the number of second cardiac health states, and if the percentage of the second cardiac health state is higher than a threshold, give the second cardiac health state as the combined cardiac health state. In the case of outputting a score for the second cardiac health state, the score can be used to improve the combination of different cardiac cycles, for example by weighted averaging. Step S3 may not be completely repeated for all cardiac cycles. Steps S31 to S34 can be performed once for a complete measurement window, and only steps S35 and S36 are repeated for each cardiac cycle that needs to be analyzed with the ML engine 4.

[0094] In a preferred embodiment, the feature extractor 3 and the ML engine 4 are arranged in a server 6 or a remote processor. Preferably, the local processor or smartphone 5 sends the measurement signal SCG received from the IMU 1 (via a communication connection like the Internet) to the server 6 or remote processor for further processing (e.g., steps S31 to S36 and S4). The output of the ML engine 4, i.e. the heart health state, is preferably sent back to the local processor or smartphone 5 and output to the user 10 through the output device of the smartphone 5. Obviously, it is also possible to send the heart health state elsewhere, such as to a healthcare professional (who has requested the measurement), to the user's account or elsewhere. However, some or all of these steps may also be performed in the local processor or smartphone 5. If all steps are performed on the smartphone 5 or local processor, there is no problem of data privacy, and the output from the ML engine 4 does not have to be sent back to the smartphone 5 or elsewhere.

[0095] It is understood that the invention is not limited to the embodiments described and that variations can be applied without exceeding the scope of the claims.

Claims

1. A method for detecting the heart health status of a user (10), comprising the following steps: receiving (S1) in a processing device (2) measurement signals (17, 18) from an IMU (1) placed on the body of the user (10), In the processing device (2), a plurality of features (K) are determined (S3) based on the measurement signals (17, 18). lin,i , K rot,i ), In the machine learning engine (4) of the processing device (2), based on the plurality of features (K lin,i , K rot,i ) determine (S4) the heart health status.

2. The method according to claim 1, wherein: The IMU (1) is arranged in a smartphone (5) or a wearable device, wherein the measurement signal comprises a six-dimensional signal having three linear dimensions (17) representing the linear acceleration of the heart measured by an accelerometer of the IMU (1) and three rotational dimensions (18) representing the angular velocity of the heart measured by a gyroscope of the IMU (1), wherein a heartbeat cycle of the heart of the user (10) is identified based on the measurement signal, wherein a plurality of sub-cycles (24) in the heartbeat cycle are determined, wherein the plurality of features (K lin,i , K rot,i ) includes linear features (K lin,i ) and rotation features (K rot,i ), where the linear feature (K lin,i ) comprises the value of at least one linear parameter of the heart motion in each sub-period (24) calculated based on the three linear dimensions (17) of the measurement signal, wherein the rotation characteristic (K rot,i ) comprises the value of at least one rotational parameter of the heart movement in each sub-cycle (24) calculated based on the three rotational dimensions (18) of the measurement signal.

3. The method according to the preceding claim, wherein: At least one linear characteristic variable (K lin,i , K rot,i ) is one or more of the following: linear power, linear kinetic energy and linear work, and at least one rotational characteristic variable is one or more of the following: rotational power, rotational kinetic energy and rotational work.

4. The method according to claim 2 or 3, wherein Calculating a linear velocity signal based on the three linear dimensions, and calculating the at least one linear characteristic variable based on the linear velocity signal, and / or A rotational acceleration signal is calculated based on the three rotational dimensions, and the at least one rotational parameter includes at least one rotational feature calculated based on the rotational acceleration signal.

5. The method according to claim 2, wherein: A reference point in the heartbeat cycle is determined, and the sub-cycle is defined based on the reference point and based on a fixed time width of the sub-cycle (24).

6. The method according to claim 1, wherein: Identifying a heartbeat cycle of the user's (10) heart, wherein a plurality of sub-cycles (24) in the heartbeat cycle are determined, wherein the plurality of features (K lin,i , K rot,i ) comprises the value of at least one characteristic variable in each determined sub-cycle (24), wherein a reference point in the heartbeat cycle is determined and the sub-cycle (24) is defined based on the reference point and based on a fixed time width of the sub-cycle.

7. The method according to claim 5 or 6, wherein: The fixed time width can be an absolute fixed time width or a relative fixed time width, wherein the absolute fixed time width is a fixed absolute time value that does not change from one heartbeat cycle to another, and wherein the relative fixed time width of the sub-cycle (24) is a relative value relative to the heartbeat cycle that does not change from one heartbeat cycle to another.

8. The method according to one of claims 5 to 7, wherein: The definition of the fixed time width is independent of the physiological time point of the heartbeat.

9. The method according to claim 2, wherein: Heartbeat cycles in the measurement signal are identified based on a matrix profile motif algorithm applied to a signal based on the measurement signal, wherein the matrix profile motif algorithm identifies one or more reference heartbeat cycles as a motif in the signal based on a matrix profile.

10. The method according to claim 5, wherein: The machine learning engine (4) determines whether the heart exhibits congestive heart failure.

11. The method according to claim 2, wherein: Respiration information of the user (10) is detected based on the measurement signal, wherein the plurality of features include the respiration information of the user (10).

12. The method according to claim 2, wherein: Respiration information of the user (10) is detected based on the measurement signal, wherein the plurality of features include the respiration information of the user (10).

13. The method according to one of claims 1 to 4 or claim 12, wherein: The machine learning engine (4) determines whether the heart exhibits congestive heart failure.

14. The method according to one of the preceding claims, wherein: The plurality of features are determined independently of the ECG measurement.

15. The method according to one of the preceding claims, comprising the step of placing the IMU (1) on the chest of the user (10) and measuring the measurement signal with the IMU (1) when placed on the chest of the user (10).

16. A method according to the preceding claim, wherein: The IMU (1) is included in a smartphone (5) or a wearable device (7) connected to the smartphone (5), wherein the processing device requests from an operating system running on the smartphone (5) what type the smartphone (5) or the wearable device (7) is or requests user input regarding the type of the smartphone (5) or the wearable device (7), wherein the plurality of characteristics depends on the type of smartphone (5) or wearable device (7) received.

17. A method according to the preceding claim, wherein: The user (10) lies down, and the smartphone including the IMU (1) is placed on the chest of the user (10).

18. A system for detecting the cardiac health status of a user (10), comprising: an IMU (1) configured to measure measurement signals (17, 18) when placed on a body of a user (10), A processing device (2) having a machine learning engine (4) configured to perform the steps of the method according to one of claims 1 to 17 based on measurement signals (17, 18) received from the IMU (1).

19. A system according to the preceding claim, wherein: The system comprises a smartphone (5) and a server (6), wherein the IMU (1) is included in the smartphone (5) or a wearable device (7) connected to the smartphone (5), wherein an application is run on a processor of the smartphone (5), wherein the application is configured to connect to the server (6), wherein the processor of the smartphone (5) and / or the processor of the server (6) operates as the processing device.

20. A computer program comprising instructions configured to, when executed by a processing device, perform the steps of the method according to one of claims 1 to 17.

Images