Autoencoder in Quantitative Seismocardiography

Abstract

A technique for determining signs of heart failure in a subject (18) is proposed. It includes obtaining (100a) a first signal (36) interval from a source signal recorded using an accelerometer (14) arranged on the chest of the subject (18), the first signal interval (36) corresponding to a first sub-interval of the cardiac cycle, inputting (200a) the first signal interval (36) into a first autoencoder, the first autoencoder being trained with corresponding first signal intervals obtained from healthy subjects and outputting a reconstructed first signal interval (44), determining (300a) a first correlation between the first signal interval (36) and the reconstructed first signal interval (44), and determining (400) signs of heart failure based on the first correlation.

Description

Technical Field

[0001] The proposed technology is related to seismocardiography and technology for assisting in the diagnosis of heart failure.

Background Art

[0002] Seismocardiography (SCG) is the analysis of sub-audible low-frequency vibrations in the chest wall caused by the beating heart. More generally, SCG is related to non-invasively measuring the acceleration in the chest wall caused by myocardial movement. Heart sounds are the audible components of chest wall vibrations, typically above 40 - 60 Hz, while the frequencies of SCG are typically below 5 Hz.

[0003] SCG is typically measured using an accelerometer. However, when an accelerometer is used, both the low-frequency SCG component and the audible component are sampled simultaneously. The SCG component and the audible component reveal different cardiovascular functions and thus enable different approaches to the diagnosis of cardiovascular function. For example, SCG is typically suitable for estimating the time intervals between features in the cardiac cycle, while heart sounds are suitable for detecting heart murmurs caused by blood flow disorders.

[0004] Heart failure does not have a single origin and can be caused by various underlying diseases. Thus, heart failure can change the SCG signal in various ways. For example, the measured amplitude can potentially increase and decrease due to heart failure. This is a challenge in identifying heart failure from the SCG signal of a subject.

Summary of the Invention

Problems to be Solved by the Invention

[0005] An object of the present invention is to provide an improved technique for identifying heart failure. A further object is to provide a reliable and easy-to-use technique.

Means for Solving the Problems

[0006] According to a first configuration of the proposed technology, a method for determining signs of heart failure in a subject or a person is proposed. The method includes obtaining a first signal interval from a source signal recorded using an accelerometer arranged on the chest of the subject, inputting the first signal interval into a first autoencoder, where the first autoencoder is trained with corresponding first signal intervals obtained from healthy subjects and outputs a reconstructed first signal interval, determining a first correlation or a first error between the first signal interval and the reconstructed first signal interval, and determining the signs of heart failure based on the first correlation. In other words, a method for determining signs of heart failure in a subject or a person based on a first signal interval from a source signal recorded using an accelerometer arranged on the chest of the subject is proposed, the method including inputting the first signal interval into a first autoencoder, where the first autoencoder is trained with corresponding first signal intervals obtained from healthy subjects and outputs a reconstructed first signal interval, determining a first correlation or a first error between the first signal interval and the reconstructed first signal interval, and determining the signs of heart failure based on the first correlation.

[0007] In other words, the last two steps of the method determine or estimate the signs of heart failure based on the correlation or error between the first signal interval and the reconstructed first signal interval. The first signal interval corresponds to or is temporally correlated with a first sub-interval of the cardiac cycle. Alternatively, it may correspond to or be temporally correlated with the complete cardiac cycle.

[0008] According to a second configuration of the proposed technology, a system for determining signs of heart failure in a subject is proposed, the system including (A) an accelerometer configured to be arranged on the chest of the subject and (B) a processor operably connected to the accelerometer and configured to perform the method according to the first configuration of the proposed technology.

[0009] According to a third configuration of the proposed technology, a system for determining signs of heart failure in a subject is proposed. The system includes a plurality of modules. It is understood that the modules jointly perform, or are configured to jointly perform, the method according to the first configuration of the proposed technology, and each module performs one of the steps identified in relation to the first configuration of the proposed technology.

[0010] According to a fourth configuration of the proposed technology, a computer program product for use in a system for determining signs of heart failure in a subject is proposed. The system includes (A) an accelerometer arranged on the chest of the subject and (B) a processor operably connected to the accelerometer. The computer program product includes program code instructions configured to cause the processor of the system to perform a method according to the first configuration of the proposed technology when executed by the processor of the system.

[0011] According to a fifth configuration of the proposed technology, a non-transitory memory storing a computer program product according to the fourth configuration of the proposed technology is proposed.

[0012] The first autoencoder is trained on healthy subjects. This means that it is efficient in reconstructing the first signal intervals from healthy subjects and less efficient in reconstructing the first signal intervals, which are often more complex, from subjects with heart failure. Thus, for subjects with heart failure, the first correlation is low. This difference in correlation can be used when determining signs of heart failure.

[0013] It is understood that the accelerometer is configured to measure the acceleration and vibration of the subject's chest wall caused by the movement of the myocardium. That the first signal interval corresponds to a first sub-interval of the cardiac cycle means that it covers or corresponds to the first part of one cardiac cycle.

[0014] In the method of the first configuration, an accelerometer can be arranged on the chest of the subject and attached to the skin of the subject with an adhesive in order to measure acceleration and vibration. The systems of the second, third, and fourth configurations can further include an adhesive patch configured to support the accelerometer or a housing to be described later and to adhere to the skin of the subject. By attaching the accelerometer or the housing to the skin, the quality of the recorded signal is improved.

[0015] The accelerometer can be arranged on the front side of the chest of the subject. The fact that the accelerometer is arranged on the chest of the subject means that it is arranged outside rather than inside the body. This has the advantage of being easily applicable without the need for trained staff. This also has the advantages of not requiring invasive procedures and being able to be used in a non-sterile environment.

[0016] The accelerometer can include a piezoelectric element. The signal can represent the voltage generated by the piezoelectric element. Accordingly, the signal strength or amplitude of the source signal can represent a voltage value.

[0017] The various configurations described above can be modified as follows.

[0018] The first signal interval or the first sub-interval can cover the systolic phase or a part of the systolic phase of the cardiac cycle. Alternatively, it can cover the diastolic phase or a part of the diastolic phase of the cardiac cycle.

[0019] The first autoencoder can compress the first signal interval into several nodes or variables, and the number of nodes is less than 15, less than 12, or less than 10, or in the range of 5 to 15, 6 to 12, or 7 to 9. The first autoencoder can be a single-layer autoencoder. Preferably, the first autoencoder is an Undercomplete Autoencoder, which means that the number of layers of the first autoencoder is less than the number of samples in the signal interval. Alternatively, the first autoencoder is a Sparse Autoencoder, a Convolutional Autoencoder, a Variational Autoencoder, a Contractive Autoencoder, or a Deep Autoencoder.

[0020] The first correlation can be a correlation measurement value or an error measurement value between the first signal interval and the reconstructed first signal interval, for example, based on the Pearson correlation coefficient, the mean squared error (MSE), or the root mean squared error (RMSE). In other words, determining the first correlation can include determining a correlation measurement value or an error measurement value between the first signal interval and the reconstructed first signal interval. For example, the correlation measurement value can be the Pearson correlation coefficient, the Spearman correlation coefficient, the MSE, or the RMSE between the first signal interval and the reconstructed first signal interval.

[0021] Signs of heart failure can be measurement values such as measurement values indicating the possibility that a subject has heart failure. Signs of heart failure can be, for example, a probability score or a risk score regarding heart failure based on a logistic regression model and a first correlation. In other words, determining signs of heart failure can include, for example, determining a probability score or a risk score regarding heart failure based on a logistic regression model and a first correlation. Determining a probability score or determining a probability score can further be based on demographic data such as gender and age.

[0022] It has been specified above that the first signal corresponds to a first sub-interval of the cardiac cycle. The method includes obtaining a second signal interval from a source signal recorded using an accelerometer arranged on the chest of a subject, the second signal interval corresponding to or being temporally correlated with a second sub-interval of the cardiac cycle, inputting the second signal interval into a second autoencoder, the second autoencoder being trained with corresponding second signal intervals obtained from healthy subjects and outputting a reconstructed second signal interval, determining a second correlation or a second error between the second signal interval and the reconstructed second signal interval, and determining the signs of heart failure based on the first correlation. In other words, the first configuration of the proposed technique, the method can further be based on the second signal interval from the source signal, the second signal interval corresponding to or being temporally correlated with a second sub-interval of the cardiac cycle, the method including inputting the second signal interval into a second autoencoder, the second autoencoder being trained with corresponding second signal intervals obtained from healthy subjects and outputting a reconstructed second signal interval, and further including determining a second correlation or a second error between the second signal interval and the reconstructed second signal interval.

[0023] Determining signs of heart failure can thus be further based on the second correlation. In other words, determining the second correlation and the signs, the method can include determining signs of heart failure based on the correlation or error between the first signal interval and the reconstructed first signal interval and between the second signal interval and the reconstructed second signal interval. Alternatively, the signs based on the first signal interval can be the first signs, and the signs based on the second signal interval can be second signs independent of the first signs.

[0024] The first sub - interval and the second sub - interval of the cardiac cycle can overlap. The first signal interval or the first sub - interval can have a center within or cover the systolic phase of the cardiac cycle, and the second signal interval or the second sub - interval can have a center within or cover the diastolic phase of the cardiac cycle. It has been found that this contributes to improving the indication of heart failure.

[0025] The second auto - encoder can compress the first signal interval into several nodes or variables, where the number of nodes is less than 15, less than 12, or less than 10, or in the range of 5 to 15, 6 to 12, or 7 to 9. The second auto - encoder can be a single - layer type auto - encoder. The second auto - encoder can be of any of the types of auto - encoders described above in relation to the first auto - encoder. For example, it can be an under - complete auto - encoder. The first auto - encoder and the second auto - encoder can be of the same type.

[0026] The second correlation can be, for example, a correlation measurement value or an error measurement value between the second signal interval and the reconstructed second signal interval based on the Pearson correlation coefficient, MSE, or RMSE. In other words, determining the second correlation can include determining a correlation measurement value or an error measurement value between the second signal interval and the reconstructed second signal interval. For example, the correlation measurement value can be the Pearson correlation coefficient, Spearman correlation coefficient, MSE, or RMSE between the second signal interval and the reconstructed second signal interval.

[0027] The symptoms of heart failure can be, for example, a probability score or a risk score for heart failure based on a logistic regression model, a first correlation, and a second correlation. In other words, determining the symptoms of heart failure can include determining a probability score or a risk score for heart failure based on a logistic regression model, a first correlation, and a second correlation.

[0028] It is understood that the source signal can extend or be recorded over a period covering a plurality of cardiac cycles. Further, it is understood that the first signal interval can be determined based on or from an average segment from or among a plurality of signal segments of the source signal, where each signal element covers one cardiac cycle.

[0029] Obtaining the first signal interval can include recording the source signal using an accelerometer disposed on the chest of the subject, where the source signal is recorded over a period covering a plurality of cardiac cycles of the subject, dividing the source signal into a plurality of signal segments, where each signal segment covers one cardiac cycle, aligning the plurality of signal segments, determining an average segment based on or from the plurality of signal segments, and determining the first signal interval in the average segment.

[0030] Stated another way, the first configuration of the proposed technology is that the method can further include downloading a source signal from a digital storage such as a computer server system or cloud storage. The method can further include dividing the source signal into a plurality of signal segments, each signal segment covering one cardiac cycle, aligning the plurality of signal segments, determining an average segment based on or from the plurality of signal segments, and determining a first signal interval in the average segment. It is understood that the digital storage can be located at a position far from the accelerometer that recorded the source signal. The source signal can be obtained for a purpose different from determining the risk regarding heart failure. For example, it can be recorded for research related to seismocardiography (SCG) or ballistocardiography (BCG). This means that it can be used in many statistical studies on the data stored by the proposed technology.

[0031] Dividing the source signal into a plurality of signal segments includes identifying a plurality of heart sounds in the source signal. For example, the heart sound can be the S1 heart sound or the S2 heart sound. Each heart sound is associated with one cardiac cycle. The recorded source signal is divided into a plurality of segments based on the identified plurality of heart sounds. This can include extracting an audio signal from the source signal and determining the heart sounds in the audible signal. For example, the audio signal can be extracted by filtering the source signal using a high-pass filter having a low cut-off frequency in the range of 40 - 60 Hz or about 50 Hz.

[0032] Alternatively, the method may further include recording an audio signal using a microphone disposed on the chest of the subject while recording the source signal using an accelerometer, and dividing the source signal into a plurality of signal segments includes identifying a plurality of heart sounds in the audio signal, each heart sound being associated with one cardiac cycle, and dividing the source signal into a plurality of segments based on the identified plurality of heart sounds.

[0033] The first signal interval may cover the aortic valve opening (AO) of the cardiac cycle. Further, or alternatively, it may cover the aortic valve closure (AC) of the cardiac cycle.

[0034] Determining the first signal interval in the average segment may include identifying a first fiducial point in the average segment and positioning the first signal interval relative to the first fiducial point. For example, the first fiducial point can be the Gs point in systole, which is defined, for example, in "Definition of Fiducial Points in the Normal Seismocardiogram" by K. Sorensen, S. E. Schmidt, A. S. Jensen et al. in Sci Rep 8, 15455 (2018) (https: / / doi.org / 10.1038 / s41598-018-33675-6). The first fiducial point may be located at the aortic valve opening (AO) of the cardiac cycle. The first signal interval may have a fixed length. For example, it may have a length of 750 ms. The first signal interval may extend from -250 to +500 ms relative to the Gs point.

[0035] Similar to the first signal interval, the second signal interval can be determined from or within the average segment. Determining the second signal interval may include determining the second signal interval in the average segment. In other words, the first configuration of the proposed technique involving downloading the above-described source signal, the method may include determining the second signal interval in the average segment. The second signal interval may cover the aortic valve closure (AC) of the cardiac cycle. Thus, the second signal interval includes information regarding the effeminacy of diastolic relaxation, which is considered an imported parameter.

[0036] Determining the second signal interval in the average segment may include identifying a second reference point in the average segment and positioning the second signal interval relative to the second reference point. For example, the second reference point can be the Dd point in diastole, which is defined, for example, in "Definition of Fiducial Points in the Normal Seismocardiogram" by K. Sorensen, S. E. Schmidt, A. S. Jensen et al. in Sci Rep 8, 15455 (2018) (https: / / doi.org / 10.1038 / s41598-018-33675-6). The second reference point may be located at the aortic valve closure (AC) of the cardiac cycle. The first signal interval may have a fixed length. For example, it may have a length of 750 ms. The first signal interval may extend from -300 to +500 ms relative to the Ds point.

[0037] The method may further include outputting or displaying signs of heart failure. For example, the signs of heart failure can be displayed as numbers in the range from 0 to 100. The system may include a display for displaying the signs of heart failure.

[0038] The method can further include filtering the source signal, the plurality of segments, or the average segment using a high-pass filter having a cut-off frequency below 1. In other words, the source signal can be an SCG signal. The source signal can span frequencies outside the audible range and audible frequencies.

[0039] The method further includes discarding noisy signal segments. This can be done as shown in WO2017 / 216374A1.

[0040] The system of the above configuration can include a non-transitory memory storing program code instructions that configure the processor to perform the method when executed by the (C) processor.

[0041] The system can include a smartphone. The processor and / or the non-transitory memory can be components incorporated in the smartphone. Also, the system can include a case or holder for supporting the smartphone, and the case or holder can include an adhesive patch configured to attach the case or holder to the subject's skin. Instead of making the accelerometer a component incorporated in the smartphone, the accelerometer can form part of an auxiliary unit configured to communicate with the smartphone, wired or wirelessly, such as a band that can be strapped around the subject's chest.

[0042] The processor can further be configured to operate the accelerometer to record the source signal, for example, using an accelerometer arranged on the subject's chest, and the signal is recorded over a time period covering a plurality of cardiac cycles of the subject. The processor can further be configured to divide the recorded signal into a plurality of signal segments.

[0043] The system of the above configuration can include a microphone configured to be arranged on the chest of a subject for measuring the sound generated by the beating heart. The processor is further operably connected to the microphone and is configured to, for example, operate the accelerometer to record the source signal with an accelerometer arranged on the chest of the subject, and, simultaneously with the recording of the source signal using the accelerometer, operate the microphone to record the audio signal using the microphone arranged on the chest of the subject, and, in the audio signal, identify a plurality of heart sounds where each heart sound is associated with one cardiac cycle, and, based on the identified plurality of heart sounds, divide the recorded source signal into a plurality of segments to obtain a plurality of signal segments. Alternatively, the processor is further configured to, for example, operate the accelerometer to record the source signal with an accelerometer arranged on the chest of the subject, and filter the source signal to obtain the audio signal, and, in the audio signal, identify a plurality of heart sounds where each heart sound is associated with one cardiac cycle, and, based on the identified plurality of heart sounds, divide the recorded signal into a plurality of segments. Filtering can include a high-pass filter having a low cut-off frequency in the range of 40 - 60 Hz or about 50 Hz.

[0044] Here, the plurality of heart sounds can be the first heart sound (S1). Alternatively, the plurality of heart sounds can be the second heart sound (S2).

[0045] The system can include a housing or cover that supports and encloses or covers the processor. The housing or cover can further enclose or cover at least a part or the whole of the accelerometer and / or at least a part or the whole of the microphone.

[0046] Obtaining the first signal interval and, optionally, the second signal interval can further include storing a plurality of signal segments and / or averages in a non-transitory memory or an auxiliary non-transitory memory. The auxiliary non-transitory memory can form part of a computer server system that can be at a remote location.

[0047] Determining signs of heart failure can include storing signs of heart failure, or first and second signs of heart failure, in a non-transitory memory or an auxiliary non-transitory memory. The auxiliary non-transitory memory can form part of a computer server system that can be at a remote location.

[0048] Outputting or displaying signs of heart failure can further include obtaining previously determined signs of heart failure of the subject, and the output information of the signs of heart failure can further be based on previously obtained signs of heart failure. The previously obtained measurements can be stored in a non-transitory memory or an auxiliary non-transitory memory. Outputting or displaying the information can further include outputting or displaying previously determined signs of heart failure.

[0049] Previously determined signs of heart failure can be those determined in the same manner as the signs of heart failure. Previously determined signs of heart failure can be those determined at an earlier time, such as five days or ten days before the determination of the measurements or average segments.

[0050] Further advantages and features of the various configurations will become apparent from the following description of the drawings.

[0051] A more complete understanding of the above and other features and advantages of the present invention will be apparent from the following detailed description of the drawings.

Brief Description of the Drawings

[0052]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7a

Figure 7b

Figure 8a

Figure 8b

Figure 9

[0053] DETAILED DESCRIPTION OF THE DRAWINGS Figure 1 schematically shows an embodiment of a system 12 for determining signs of heart failure in a subject 18. The system 12 has an accelerometer 14 in the form of a piezoelectric element, which can be placed on the chest of the subject 18 and is for measuring the vibrations of the chest wall caused by the movement of the heart. A processor 20 is connected to the accelerometer 14. The processor 20 has a temporary memory 22 that can store the signals received from the accelerometer 14, whereby program code instructions can be executed. The system 12 includes a support 26 that supports the accelerometer 14 and a housing 28 that houses the processor 20. Also, the system 12 has a non - temporary memory 24 that stores program code instructions for the processor 20. For example, the system 12 can be made up of components that are entirely incorporated into a smartphone, or all components except the accelerometer 14 and the support 26 can form part of a smartphone. In one embodiment, the accelerometer is an accelerometer incorporated into a smartphone. In one embodiment of the system 12, it further has a display 25 that can display output information from the processor 20 such as numbers and the like.

[0054] The program code instructions in the non - temporary memory 24 act to cause the processor 20 to perform the method schematically shown in FIG. 2. A source signal is recorded 110 using an accelerometer placed on the chest of the subject 18. The source signal is recorded over a period that covers a plurality of cardiac cycles of the subject 18.

[0055] In an alternative embodiment (not shown), instead of recording the source signal, the source signal is downloaded from a computer server system (not shown), and that computer server system is, for example, a general - purpose storage for source signals obtained by an accelerometer placed on the chest of a subject for purposes such as SCG or BCG. This means that the source signal is not specifically intended for studying heart failure.

[0056] The source signal is filtered 120 using a high-pass filter having a cut-off frequency below 1. Next, the source signal is divided into a plurality of signal segments 130, which is done by extracting an audio signal from the source signal using a high-pass filter having a low cut-off frequency of 50 Hz 152. A portion of the audio signal 32 is shown in FIG. 3. The horizontal axis represents the signal intensity X (unitless), and the vertical axis represents time in milliseconds (ms). The S1 heart sound in the audio signal is identified 134, which is used to divide the source signal into a plurality of signal segments 130, and each signal segment covers one cardiac cycle. Noisy signal segments are discarded as shown in O2017 / 216374A1.

[0057] Next, by aligning those signal segments using the S1 heart sound associated with the plurality of signal segments 152, an average segment 34 is determined 150 based on the plurality of signal segments. An example of the average segment 34 is shown in FIG. 4. The horizontal axis represents acceleration in g (ms -2 ) and the vertical axis represents time in milliseconds (ms). Here, g is proportional to the voltage from the accelerometer 14.

[0058] By identifying the Gs point in the systolic phase as the first reference point 38 in the average segment 34 172, a first signal interval 36 is obtained 100a. The first reference point 38 is located within the systolic phase. The first signal interval 36 can extend from -250 to +500 ms relative to the first reference point 38. This means that the first signal interval 36 essentially covers the diastolic phase of the average segment 34 and corresponds to the first sub-interval of the cardiac cycle. Similarly, by identifying the Dd point in the diastolic phase as the second reference point 42 182, a second signal interval 40 is obtained 100a. The second signal interval 40 extends from -300 to +500 ms relative to the first reference point 38. Only the start of the second signal interval 40 is shown in FIG. 4. This means that the second signal interval 40 essentially covers the systolic phase of the average segment 34.

[0059] Autoencoders 46 and 48 are schematically shown in FIG. 5. Autoencoders 46 and 48 have an encoder 50 that maps an input to a code, and a decoder 54 that maps the code to a reconstructed input. The input is signal intervals 36 and 40 having a sample length i. Encoder 50 maps the input to lower it to j nodes. In the current embodiment, the number of nodes is 8 and the sample length i is 401. Next, decoder 54 maps the nodes to outputs that are reconstructed signal intervals 52 and 56 having the same sample length i as the input. Weight matrices W and W' and bias vectors b and b' are determined by training autoencoders 46 and 48 on healthy subjects.

[0060] A first signal interval 36 is input 200a to a first autoencoder 46 trained with a corresponding first signal interval from a healthy subject. The first autoencoder 46 outputs a reconstructed first signal interval 52. Similarly, a second signal interval 40 is input 200b to a second autoencoder 48. The second autoencoder 48 is trained with a corresponding second signal interval from a healthy subject. The second autoencoder 48 outputs a reconstructed second signal interval 56.

[0061] A first correlation between the first signal interval 36 and the reconstructed first signal interval 52 is determined 300a by calculating a first Pearson correlation coefficient for the two signals. Similarly, a second correlation between the second signal interval 40 and the reconstructed second signal interval 56 is determined 300b by calculating a second Pearson correlation coefficient for the two signals. In an alternative embodiment, the MSE or RMSE between the signal interval and the reconstructed signal interval is calculated.

[0062] Signs of heart failure are determined 400 as a probability score for heart failure by a logistic regression model for the first and second Pearson correlation coefficients. Thus, the signs are based on the first correlation and the second correlation. Next, the signs of heart failure are displayed 500 on display 25.

[0063] In an alternative embodiment of the system 12, the second signal interval is not obtained and is not input to the second autoencoder. The second correlation is not determined, and the signs of heart failure are based only on the first correlation.

[0064] FIG. 6 schematically shows an alternative embodiment of a system for determining signs of heart failure in a subject. System 12 is similar to the system described in connection with FIG. 1, and features having the same or related functions have the same numeral designations. Further, system 12 includes a microphone 30 in the form of a transducer that can convert sound into an electrical signal. Microphone 30 is supported by support 26.

[0065] The program code instructions in the non-transitory memory 24 correspond to those described in connection with FIG. 2, except with respect to 130 dividing the source signal into a plurality of signal segments.

[0066] An accelerometer 14 and a microphone 30 disposed on the subject's chest are used, and the processor 20 operates the microphone 30 according to the program code instructions such that an audio signal is recorded using the microphone 30 simultaneously with the source signal being recorded using the accelerometer 14. The S1 heart sound is identified within the audio signal. The source signal is divided into a plurality of signal segments based on the temporal correlation between the source signal and the S1 heart sound identified as the audible signal. The subsequent alignment of the plurality of signal segments is thus based on the S1 heart sound in each signal segment.

[0067] FIG. 7a shows an alternative embodiment of the system 12 described in connection with FIG. 1, the only difference being that the support 26 forms part of the housing 28 and the housing 28 covers at least part of the accelerometer 14. In this embodiment, the housing 28 is positioned on the subject's chest, which means that the accelerometer 14 is also positioned on the subject's chest. Similarly, FIG. 7b shows an alternative embodiment of the system 12 described in connection with FIG. 6, the only difference being that the support 26 forms part of the housing 28 and the housing 28 covers at least part of the accelerometer 14 and the microphone 30. In this embodiment, the housing 28 is positioned on the subject's chest, which means that the accelerometer 14 and the microphone 30 are both positioned on the subject's chest simultaneously.

[0068] Proof of concept The system was used based on the system described in connection with FIGS. 1 and 2. A method was developed to estimate the risk of having heart failure (HF) with reduced systolic function (HFrEF). Hereinafter, the corresponding source signal is referred to as the SCG signal.

[0069] For development and validation, SCG signals from 200 subjects were used. The dataset was split into a training set and a test set, with each set related to 100 subjects.

[0070] [Table 1]

[0071] Table 1 shows the baseline demographics and known HF diagnoses for the training set and the test set.

[0072] The autoencoder was trained using only 75 patients in the training set for whom a diagnosis of no heart failure (NoHF) was made. Since the autoencoder is not trained to detect HFrEF, the autoencoder can be considered an unsupervised machine learning model.

[0073] The first and second signal intervals of 200 subjects were input into the trained first and second autoencoders respectively, and the first and second reconstructed signal intervals were output. The first signal interval 34 solid line and the reconstructed first signal interval 44 for the NoHF subjects are shown in FIG. 8a. The corresponding signals for the subjects with known HFrEF are shown in FIG. 8b. The Gs point 58 is shown in FIGS. 8a and 8b. In FIGS. 8a and 8b, it can be understood that the signal intervals and the reconstructed signal intervals differ more for the HFrEF subjects than for the NoHF subjects. The first and second correlations were determined as previously explained in connection with FIG. 2. Hereinafter, the first correlation is denoted as rSys and the second correlation is denoted as rDia. It was discovered that the correlations rSys and rDia were higher for the NoHF subjects than for the HFrEF subjects.

[0074] To provide a risk prediction score based on rSys and rDia, a logistic regression model was constructed. Hereinafter, the risk prediction score is called the HF score and ranges from 0 to 100. The HF scores for the training set, test set, and combined set (all) are shown in the box plot of FIG. 9. For HFrEF, a risk threshold of 5% was defined as high risk, which is indicated by the horizontal dashed line.

[0075] [Table 2]

[0076] Table 3 shows the performance of the first autoencoder (rSys) and the second autoencoder (rDia) measured as the area under the receiver characteristic operator curve (AUC).

[0077]

Table 3

[0078] Table 4 shows the classification performance of the HF score.

[0079] In conclusion, the AUC values obtained using both autoencoders (rSys and rDia) and the classification performance of the HF score guarantee the potential ability of the proposed technique to strongly agree with the risk of HFrEF.

Claims

1. A method for determining signs of heart failure in a subject (18), comprising: - obtaining a first signal interval (36) from a source signal recorded using an accelerometer (14) arranged on the chest of the subject (18) (100a), wherein the first signal interval (36) corresponds to a first sub-interval of the cardiac cycle; - inputting the first signal interval (36) into a first autoencoder (200a), the first autoencoder being trained with corresponding first signal intervals obtained from healthy subjects and outputting a reconstructed first signal interval (44); - determining a first correlation between the first signal interval (36) and the reconstructed first signal interval (44) (300a); - determining signs of heart failure based on the first correlation (400). A method comprising the above steps.

2. The method according to claim 1, wherein the first signal interval (36) covers a part of the systolic phase of the cardiac cycle.

3. The method according to claim 1 or 2, wherein the first autoencoder compresses the first signal interval (36) into a plurality of nodes, and the number of the plurality of nodes ranges from 5 to 15, from 6 to 12, or from 7 to 9.

4. The method according to any one of claims 1 - 3, wherein the first autoencoder is a single-layer autoencoder.

5. The method according to any one of claims 1 - 4, wherein the first correlation is a correlation measurement value between the first signal interval (36) and the reconstructed first signal interval (44), and the first correlation is based on a Pearson correlation coefficient, a mean squared error (MSE), or a root mean squared error (RMSE).

6. The method according to any one of claims 1 - 5, wherein the signs of heart failure are a probability score related to heart failure, and the probability score is based on a logistic regression model and the first correlation.

7. The method according to any one of claims 1 - 6, further comprising: - obtaining a second signal interval (40) from the source signal recorded using the accelerometer (14) arranged on the chest of the subject (18) (100b), wherein the second signal interval (40) corresponds to a second sub-interval of the cardiac cycle; - inputting the second signal interval (40) into a second autoencoder (200b), the second autoencoder being trained with corresponding second signal intervals obtained from healthy subjects and outputting a reconstructed second signal interval (56). - determining a second correlation between the second signal interval (40) and the reconstructed second signal interval (56) (300b); - determining the signs of heart failure (400) is further based on the second correlation and a method further comprising. **Claim 8** A method according to any one of claims 1-7, wherein obtaining the first signal interval (36) (100a) comprises: - recording a source signal using an accelerometer (14) arranged on the chest of a subject (18) (110), the source signal being recorded over a period covering a plurality of cardiac cycles of the subject (18); - dividing the source signal into a plurality of signal segments (130), each signal segment covering one cardiac cycle; - aligning the plurality of signal segments (152); - determining an average segment (34) based on or from the plurality of signal segments (150); - determining the first signal interval (36) in the average segment (34) (160) and a method comprising. **Claim 9** A method according to claim 8, wherein determining the first signal interval (36) in the average segment (34) (160) comprises: - identifying a first reference point (38) in the average segment (34) (162); - placing the first signal interval (36) relative to the first reference point (38) (164) and a method comprising. **Claim 10** A method according to any one of claims 1-9, further comprising displaying signs of heart failure. **Claim 11** A system (12) for determining signs of heart failure in a subject (18), comprising: (A) an accelerometer (14) configured to be arranged on the chest of a subject (18); (B) a processor (20) operatively connected to the accelerometer (14) and configured to perform the method according to any one of claims 1-10 and a system comprising. **Claim 12** A system (12) according to claim 11, comprising: (C) a non-transitory memory (24) storing program code instructions, the program code instructions configuring the processor (20) to perform the method when executed by the processor (20); and a system. **Claim 13** A computer program product for use in a system (12) for determining signs of heart failure in a subject (18), the system (12) including: (A) an accelerometer (14) disposed on the chest of the subject (18); and (B) a processor (20) operably connected to the accelerometer (14), the computer program product including program code instructions that, when executed by the processor (20) of the system (12), cause the method of any one of claims 1-10 to be performed by the processor (20).

14. A non-transitory memory (24) storing the computer program product of claim 13.

Images