ECG information analysis device

Abstract

To provide an electrocardiographic information analysis device capable of accurately detecting an abnormality of a subject from an electrocardiographic waveform while suppressing an arithmetic amount.SOLUTION: An electrocardiographic information analysis device includes: a time frequency conversion unit 21 for calculating a frequency spectrum of an electrocardiographic waveform of each frame from the electrocardiographic waveform indicating a behavior of a subject's heart, shown in electrocardiographic information obtained from an electrocardiograph 11 attached to the subject; a peak interval detection unit 22 for detecting at least one of an interval between peaks of the same kind of the electrocardiographic waveform and an interval between peaks of different kinds for each frame; a selection unit 23 for selecting a set of weighting coefficients corresponding to the attribute of the subject out of the sets of the respective weighting coefficients of a plurality of attributes registered beforehand; and an abnormality determination unit 24 for determining presence or absence of an abnormality of a behavior of the subject's heart using a discriminator on the basis of a feature vector obtained by weighting the frequency spectrum and the peak interval with the set of weighting coefficients selected for the frequency spectrum of the electrocardiographic waveform and the detected peak interval.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to, for example, an electrocardiogram information analysis device that analyzes electrocardiogram information. [Background technology]

[0002] Conventionally, electrocardiogram waveforms, which are measured by an electrocardiograph and indicate the behavior of a subject's heart, have been used to diagnose the subject's heart disease, etc. Therefore, a technique for accurately detecting abnormalities in the behavior of the heart based on the electrocardiogram waveform has been proposed (see Patent Document 1).

[0003] The anomaly detection device described in Patent Document 1 performs learning using an autoencoder using waveform data of time changes obtained from the inherent motion of an object, and executes persistent homology transformation on the waveform data to calculate a change in the number of connected components corresponding to a change in a threshold in the value direction. This anomaly detection device then inputs the output of the autoencoder based on the waveform data and the output of the persistent homology transformation, and detects anomalies based on the waveform data and the discrimination result of a learning device that has undergone machine learning on anomalies. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2020-36633 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, in the above technique, not only calculations by a learning device but also calculations by an autoencoder and persistent homology transform are performed, resulting in a very large amount of calculations.

[0006] Therefore, an object of the present invention is to provide an electrocardiogram information analyzer that can accurately detect abnormalities in a subject from an electrocardiogram waveform while suppressing the amount of calculation. [Means for solving the problem]

[0007] According to one embodiment, an electrocardiogram information analysis device is provided, which includes a storage unit that stores, for each of a plurality of preregistered attributes, a set of weighting factors for the frequency spectrum of an electrocardiogram waveform and at least one of the intervals between peaks of the same type and the intervals between peaks of different types, a time-frequency conversion unit that calculates a frequency spectrum of an electrocardiogram waveform for each frame having a predetermined time length by performing a time-frequency conversion on an electrocardiogram waveform that represents the behavior of the subject's heart and is shown in electrocardiogram information obtained by an electrocardiograph attached to the subject, a peak interval detection unit that detects at least one of the intervals between peaks of the same type and the intervals between peaks of different types for each frame, a selection unit that selects a set of weighting factors corresponding to the subject's attribute from the sets of weighting factors for the plurality of preregistered attributes based on attribute information representing the subject's attribute, and an abnormality determination unit that determines the presence or absence of an abnormality in the behavior of the subject's heart by inputting the frequency spectrum of the electrocardiogram waveform for each frame and at least one peak interval to a classifier that has been trained in advance to determine the presence or absence of an abnormality in the behavior of the heart. The classifier then calculates a feature vector weighted by a set of selected weighting coefficients for the frequency spectrum of the electrocardiogram waveform for each frame and at least one peak interval, and determines whether or not there is an abnormality in the subject's cardiac behavior based on the calculated feature vector. Effect of the Invention

[0008] The electrocardiogram information analysis device according to the present disclosure has the effect of being able to accurately detect abnormalities in a subject from an electrocardiogram waveform while reducing the amount of calculation. [Brief description of the drawings]

[0009] [Figure 1] 1 is a hardware configuration diagram of an electrocardiogram information analysis device according to one embodiment. [Diagram 2] FIG. 2 is a functional block diagram of a processor related to electrocardiogram information analysis processing. [Diagram 3] 13A and 13B are diagrams showing examples of frequency spectra of electrocardiogram waveforms. [Figure 4] FIG. 2 is a diagram showing an example of an electrocardiogram waveform. [Diagram 5] 13(a) and 13(b) are diagrams showing an example of the interval between two peaks in an electrocardiogram waveform. [Figure 6] FIG. 2 is a schematic diagram illustrating the configuration of a classifier. [Figure 7] 13 is an operational flowchart of an electrocardiogram information analysis process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an electrocardiogram information analysis device, an electrocardiogram information analysis method, and an electrocardiogram information analysis computer program executed by the electrocardiogram information analysis device will be described with reference to the drawings. The electrocardiogram information analysis device detects abnormalities in the behavior of the subject's heart by analyzing electrocardiogram information representing the electrocardiogram waveform of the subject obtained by an electrocardiograph. To this end, the electrocardiogram information analysis device converts the electrocardiogram waveform from the time domain to the frequency domain to calculate the frequency spectrum of the electrocardiogram waveform and calculates the occurrence interval of a predetermined peak indicated by the electrocardiogram waveform. The electrocardiogram information analysis device then inputs the frequency spectrum and peak interval of the electrocardiogram waveform to a classifier that has been trained in advance to determine the presence or absence of abnormalities in the behavior of the heart, thereby detecting abnormalities in the behavior of the subject's heart. The classifier has a weighting layer that performs weighting calculations on each element included in the frequency spectrum and peak interval input to the classifier, and a set of weighting coefficients used in the weighting layer is selected according to the attributes of the subject.

[0011] 1 is a hardware configuration diagram of an electrocardiogram information analysis device according to one embodiment. The electrocardiogram information analysis device 1 may be a device worn by a subject together with an electrocardiograph 11, or may be configured as a server. The electrocardiogram information analysis device 1 has a communication interface 2, a memory 3, and a processor 4. The electrocardiogram information analysis device 1 may further have a user interface device (not shown) such as a touch panel, a keyboard, or a display device. Furthermore, the electrocardiogram information analysis device 1 may have a storage device (not shown) such as a hard disk device or an optical recording medium and an access device therefor.

[0012] The communication interface 2 has a communication interface conforming to a predetermined wired or wireless communication standard and its control circuit. The communication interface 2 receives electrocardiogram information showing an electrocardiogram waveform showing the behavior of the subject's heart from the electrocardiograph 11 connected via a communication line capable of communication according to the predetermined communication standard, and passes the electrocardiogram information to the processor 4. When the communication interface 2 receives attribute information showing the attributes of the subject from a user interface device or another device, it passes the attribute information to the processor 4. Furthermore, the communication interface 2 may output information showing the determination result of the presence or absence of abnormality in the behavior of the subject's heart obtained as a result of execution of the electrocardiogram information analysis process received from the processor 4 to the user interface device or another device.

[0013] The memory 3 is an example of a storage unit, and includes, for example, a readable / writable semiconductor memory and a read-only semiconductor memory. The memory 3 stores, for example, a computer program for electrocardiogram information analysis processing executed by the processor 4, various data used in the electrocardiogram information analysis processing, for example, a parameter set that specifies a classifier used in the electrocardiogram information analysis processing. Such a parameter set includes a set of weighting coefficients used in a weighting layer of the classifier corresponding to each of a plurality of attributes registered in advance. Furthermore, the memory 3 stores attribute information of the subject inputted via a user interface device or other device. Furthermore, the memory 3 may store electrocardiogram information to be subjected to electrocardiogram information analysis processing, various data generated during execution of the electrocardiogram information analysis processing, and information representing the result of the electrocardiogram information analysis processing.

[0014] The processor 4 is an example of a processing unit, and includes, for example, one or more CPUs and their peripheral circuits. The processor 4 may further include an arithmetic circuit for numerical calculation or an arithmetic circuit for graphic processing. The processor 4 controls the entire electrocardiogram information analysis device 1. The processor 4 also executes electrocardiogram information analysis processing on the electrocardiogram information.

[0015] Fig. 2 is a functional block diagram of the processor 4 related to electrocardiogram information analysis processing. As shown in Fig. 2, the processor 4 has a time-frequency conversion unit 21, a peak interval detection unit 22, a selection unit 23, and an abnormality determination unit 24. Each of these units in the processor 4 is a functional module realized by, for example, a computer program executed on the processor 4. Alternatively, each of these units in the processor 4 may be a dedicated arithmetic circuit provided in the processor 4.

[0016] The time-frequency conversion unit 21 divides the electrocardiogram waveform represented in the electrocardiogram information into frames each having a predetermined time length, and performs a time-frequency conversion on the electrocardiogram waveform for each frame to calculate a frequency spectrum of the electrocardiogram waveform. The time-frequency conversion unit 21 can use, for example, a fast Fourier transform (FFT) as the time-frequency conversion. The time-frequency conversion unit 21 can perform the time-frequency conversion after multiplying each frame by a window function such as a Hamming window or a Hanning window. Alternatively, the time-frequency conversion unit 21 can use a discrete cosine transform or a wavelet transform as the time-frequency conversion. The time length of a frame is preferably, for example, from several seconds to about one minute, and more preferably from several seconds to about 10 seconds, so that one frame contains a plurality of peaks of a predetermined type (e.g., R peaks) of the electrocardiogram waveform.

[0017] Figures 3(a) and 3(b) are diagrams showing examples of frequency spectra of electrocardiogram waveforms. In Figures 3(a) and 3(b), the horizontal axis represents frequency, and the vertical axis represents spectrum intensity. Waveform 301 shown in Figure 3(a) represents the frequency spectrum of an electrocardiogram waveform when there is no abnormality in the behavior of the heart, and waveform 302 shown in Figure 3(b) represents the frequency spectrum of an electrocardiogram waveform when there is an abnormality in the behavior of the heart, such as atrial fibrillation.

[0018] As shown in waveform 301 in Fig. 3(a), when there is no abnormality in the behavior of the heart, clear peaks appear at regular frequency intervals in the frequency spectrum of the electrocardiogram waveform. In contrast, as shown in waveform 302 in Fig. 3(b), when there is an abnormality in the behavior of the heart, the frequency intervals at which peaks appear in the frequency spectrum of the electrocardiogram waveform become unclear. In this way, the frequency spectrum of the electrocardiogram waveform has a different waveform shape depending on whether there is an abnormality in the behavior of the heart, and therefore provides useful information for determining whether there is an abnormality in the behavior of the heart.

[0019] The time-frequency conversion unit 21 outputs the frequency spectrum of the electrocardiogram waveform for each frame to the abnormality determination unit 24 .

[0020] The peak interval detection unit 22 detects at least one of the intervals between peaks of the same type and the intervals between peaks of different types in the electrocardiogram waveform represented in the electrocardiogram information.

[0021] FIG. 4 is a diagram showing an example of an electrocardiographic waveform. In FIG. 4, the horizontal axis represents time, and the vertical axis represents the intensity of the electrocardiographic signal. Waveform 400 shows an example of the electrocardiographic waveform. In the electrocardiographic waveform 400, peaks called P, Q, R, S, and T, where the electrocardiographic signal has a maximum value or a minimum value, generally appear for each beat of the heart. The P peak is a peak that appears during the excitation process of the atrium. The Q peak, R peak, and S peak are peaks that appear during the excitation process of the ventricle, respectively. And the T peak is a peak that appears during the process in which the excitation of the ventricular muscle disappears.

[0022] In this embodiment, the peak interval detection section 22 detects at least one of the intervals between two successive peaks of different types and the intervals between two successive peaks of the same type, and preferably two or more of these peaks.

[0023] 5(a) and 5(b) are diagrams showing an example of an interval between two peaks detected by the peak interval detection unit 22. In FIG. 5(a) and FIG. 5(b), the horizontal axis represents time, and the vertical axis represents the intensity of an electrocardiogram signal. A waveform 500 represents an example of an electrocardiogram waveform. As shown in FIG. 5(a), for example, at least one of the following intervals is detected by the peak interval detection unit 22: a PQ interval between a P peak and the next Q peak appearing in a waveform representing the same heartbeat, a QR interval between a Q peak and the next R peak appearing, an RS interval between an R peak and the next S peak appearing, an ST interval between an S peak and the next T peak appearing, and a TP interval between a T peak and the next P peak appearing in the next heartbeat. When only one of these peaks is detected, one or more intervals between two consecutive peaks of the same type are also detected, as described later. More preferably, two or more or all of these intervals may be detected by the peak interval detection unit 22.

[0024] 5(b), for example, the peak interval detection unit 22 may detect at least one of the intervals between each of the P, Q, R, S, and T peaks appearing in a waveform representing one heartbeat and the same type of peak appearing in a waveform representing the next heartbeat, that is, the PP interval, the QQ interval, the RR interval, the SS interval, and the TT interval. When only one of these peaks is detected, one or more intervals between two consecutive peaks of different types as described above are also detected. More preferably, two or more or all of these intervals may be detected by the peak interval detection unit 22.

[0025] The peak interval detection unit 22 divides the electrocardiographic waveform represented in the electrocardiographic information into frames each having a predetermined time length, similarly to the time-frequency conversion unit 21. The peak interval detection unit 22 then detects the maximum and minimum values ​​of the electrocardiographic signal by examining the time change in the intensity of the electrocardiographic signal in the electrocardiographic waveform represented in the electrocardiographic information for each frame, thereby detecting each type of peak. For example, when the absolute value of the difference between the intensity of the electrocardiographic signal that is the maximum or minimum value and the intensity of the electrocardiographic signal before and after the maximum or minimum value is equal to or greater than a predetermined threshold, the peak interval detection unit 22 detects the maximum or minimum value as a peak. Alternatively, the peak interval detection unit 22 may detect each peak according to any of various methods for detecting each type of peak from the electrocardiographic waveform. The peak interval detection unit 22 may detect the peak interval by calculating the difference in appearance timing of two peaks that are the targets of peak interval detection.

[0026] In this case, the peak interval detection unit 22 may apply a bandpass filter to the electrocardiogram waveform that attenuates frequency components other than a predetermined frequency band in order to reduce the influence of noise superimposed on the electrocardiogram waveform and accurately detect each peak. The peak interval detection unit 22 may then detect each peak and peak interval from the electrocardiogram waveform in which noise has been reduced by applying the bandpass filter. The predetermined frequency band may be, for example, a band including frequency components corresponding to each peak, for example, a band of 1-100 Hz. More preferably, the predetermined frequency band may be a band of 2-40 Hz. In addition, when a single frame contains multiple peak intervals of the same type, the peak interval detection unit 22 may set the average value of those peak intervals as the new peak interval. For example, when a single frame contains multiple RR intervals, the peak interval detection unit 22 may set the average value of those multiple RR intervals as the RR interval in that frame.

[0027] The peak interval detection section 22 outputs two or more detected peak intervals to the abnormality determination section 24 for each frame.

[0028] The selection unit 23 selects a set of weighting coefficients corresponding to the subject's attributes indicated in the input subject's attribute information from among the sets of weighting coefficients for a plurality of attributes registered in advance. Then, the selection unit 23 reads the selected set of weighting coefficients from the memory 3.

[0029] The attribute of a subject is specified by, for example, the age group to which the subject belongs, and each age group is treated as one registered attribute. In this case, each age group is divided in 10-year age increments, such as under 10 years old, 10 to under 20 years old, 20 to under 30 years old, etc. Note that the division of each age group is not limited to this example. For example, ages 60 and over and ages 70 and over may be divided into the same age group.

[0030] The subject's attributes may be specified by the gender, the type of electrocardiograph 11 worn by the subject, vital information other than the electrocardiogram information of the subject, or the subject's surrounding environment. The vital information other than the electrocardiogram information includes at least one of body temperature, blood pressure, and heart rate. The subject's vital information may be acquired via the communication interface 2 from a sensor (not shown) attached to the subject, such as a thermometer, a blood pressure monitor, or a heart rate sensor. The range of possible values ​​of the vital information is also divided into a plurality of intervals, and each interval is treated as one registered attribute.

[0031] The surrounding environment of the subject is classified into, for example, whether the subject is stationary, walking, or riding a vehicle. In this case, the processor 4 acquires the measured value of the moving speed of the subject from a speed sensor (not shown) built into the device worn by the subject via the communication interface 2 as attribute information of the subject. Then, the selection unit 23 determines whether the subject is stationary, walking, or riding a vehicle based on the average or maximum moving speed in a recent fixed period. In this case, a speed range corresponding to each of stationary, walking, and riding is set in advance. Then, the selection unit 23 may determine whether the subject is stationary, walking, or riding a vehicle by identifying the speed range to which the average or maximum moving speed in a recent fixed period belongs.

[0032] Alternatively, the attribute of the subject may be determined by a combination of two or more of the above-mentioned items. For example, the selection unit 23 may determine the attribute of the subject based on the age and sex of the subject.

[0033] The attribute information of the subject is input via a user interface device, for example, when the subject is registered in the electrocardiogram information analysis device 1. When the attribute information of the subject is specified by the type of electrocardiograph 11 worn by the subject, the attribute information may be acquired from the electrocardiograph 11 when the electrocardiograph 11 is turned on or together with the electrocardiogram information. When the attribute information of the subject is specified by the subject's surrounding environment, the attribute information may be acquired at regular intervals from a device that specifies the surrounding environment (for example, the above-mentioned speed sensor).

[0034] The selection unit 23 passes the set of weighting factors corresponding to the attributes of the subject, which are read from the memory 3, to the abnormality determination unit 24.

[0035] The abnormality determination unit 24 determines whether or not there is an abnormality in the behavior of the subject's heart by inputting the frequency spectrum of the electrocardiogram waveform and one or more peak intervals for each frame into a classifier that has been trained in advance to determine the presence or absence of abnormality in the behavior of the heart.

[0036] In this embodiment, the abnormality determination unit 24 can use a multi-layer perceptron type neural network as a classifier.

[0037] FIG. 6 is a schematic diagram of a classifier used in this embodiment. The classifier 600 can be a deep neural network (DNN) having a so-called Self Attention Networks (SAN) type architecture. That is, the classifier 600 has a weighting layer 601, an output layer 603 that outputs a result of determining whether or not there is an abnormality, and one or more hidden layers 602 that are connected in order from the weighting layer 601 to the output layer 603 between the weighting layer 601 and the output layer 603. A frequency spectrum 611 and one or more peak intervals 612 of an electrocardiogram waveform 610 are input to the weighting layer 601. The weighting layer 601 calculates a feature vector 613 by weighting each element of the input frequency spectrum and peak interval using a selected set of weighting coefficients. Each hidden layer 602 executes a predetermined operation, such as a full connection operation or an activation operation, on the output from the upper layer. The feature vector 613 output from the weighting layer 601 is input to the most upstream hidden layer 602. Moreover, the output layer 603 calculates an output value indicating the presence or absence of an abnormality by performing an operation using, for example, a softmax function or a sigmoid function on the output from the immediately preceding hidden layer.

[0038] The weighting layer 601 may be, for example, a self-attention (SA) layer that weights each element of the input frequency spectrum and peak interval by SA. Each element of the frequency spectrum may be, for example, a value of the frequency spectrum for each predetermined sampling interval in the frequency domain. Each element of the peak interval may be a value of each peak interval input to the classifier.

[0039] In this case, the weighting layer 601 generates three types of vectors, a query, a key, and a value, from a vector having each element of the frequency spectrum as an element value. At this time, the weighting layer 601 generates a query vector by multiplying a vector of the frequency spectrum by a matrix for generating a query vector obtained by learning. Similarly, the weighting layer 601 generates a key vector and a value vector by multiplying a vector of the frequency spectrum by a matrix for generating a key vector and a matrix for generating a value vector obtained by learning, respectively. In this embodiment, the matrix for generating a query vector, the matrix for generating a key vector, and the matrix for generating a value vector are a set of weighting coefficients. Then, the weighting layer 601 calculates an inner product of the query vector and the key vector, and applies a softmax function to the calculated inner product to generate a weighting matrix. Then, the weighting layer 601 calculates a product of the generated weighting matrix and the value vector to calculate a feature vector. By configuring the weighting layer 601 as an SA layer, it is possible to learn the characteristics of the arrangement order of each element and obtain a feature vector such that the contribution of an element useful for determining the behavior of the subject's heart is increased among the elements. The weighting layer 601 performs the same process on one or more input peak intervals to calculate a feature vector for the peak interval. The weighting layer 601 then integrates the feature vectors into one feature vector so that each element included in the feature vector calculated from the frequency spectrum and each element included in the feature vector calculated from the peak interval are included as elements. According to a modified example, the weighting layer 601 may multiply one vector including each element of the frequency spectrum and each element of the peak interval by each of the above matrices to calculate a query vector, a key vector, and a value vector. The weighting layer 601 may then perform the above process on these calculated vectors to calculate a feature vector. According to another modified example, the weighting layer 601 may receive frequency spectra and peak intervals of multiple frames as input. The weighting layer 601 may then perform the above process on the frequency spectra and peak intervals to calculate a feature vector.

[0040] In this embodiment, a set of weighting coefficients according to the attributes of the subject is used as a set of weighting coefficients used in the weighting layer 601. Therefore, the classifier 600 can appropriately weight each element according to the attributes of the subject.

[0041] Note that the weighting layer 601 is not limited to the above configuration. For example, the weighting layer 601 may be configured as a convolution layer that performs a convolution operation having a predetermined kernel size. In this case, the set of weighting coefficients includes individual weighting coefficients of the kernel that performs the convolution operation.

[0042] Furthermore, the output layer 603 of the classifier 600 may be configured to output different values ​​depending on the type of anomaly detected.

[0043] The classifier 600 is trained in advance using training data according to a predetermined supervised training method such as backpropagation. In this embodiment, the weighting layer 601 of the classifier 600 is trained using multiple training data representing the frequency spectrum and peak interval of a subject having each attribute and annotated with the presence or absence of abnormality in cardiac behavior. Meanwhile, each hidden layer 602 and output layer 603 are trained in common for all attributes using each training data of each attribute. This allows the classifier 600 to accurately determine whether or not there is an abnormality in the cardiac behavior of the subject.

[0044] When the classifier outputs a determination result that there is an abnormality in the behavior of the subject's heart for any frame, the abnormality determination unit 24 detects an abnormality in the behavior of the subject's heart. On the other hand, when the classifier outputs a determination result that there is no abnormality in the behavior of the subject's heart for any frame of the electrocardiogram information, the abnormality determination unit 24 determines that there is no abnormality in the behavior of the subject's heart. Note that the abnormality determination unit 24 may determine that there is an abnormality in the behavior of the subject's heart only when the classifier outputs a determination result that there is an abnormality in the behavior of the subject's heart for a predetermined number of frames (e.g., two or more) or more in the most recent predetermined period. Then, the abnormality determination unit 24 outputs information representing the detection result regarding the presence or absence of an abnormality in the behavior of the subject's heart to a user interface device or another device via the communication interface 2. Note that, when the classifier also outputs information representing the type of abnormality, the abnormality determination unit 24 may include the information representing the type of abnormality obtained by the classifier in the information representing the above detection result.

[0045] 7 is an operational flowchart of electrocardiogram information analysis processing. Each time the processor 4 receives electrocardiogram information of a subject, it may execute electrocardiogram information analysis processing on the received electrocardiogram information in accordance with the operational flowchart below.

[0046] The time-frequency conversion unit 21 of the processor 4 performs a time-frequency conversion on the electrocardiogram waveform represented in the electrocardiogram information of the subject for each frame, and calculates the frequency spectrum of the electrocardiogram waveform (step S101).

[0047] The peak interval detection unit 22 of the processor 4 detects a predetermined type of peak for each frame from the electrocardiogram waveform represented in the electrocardiogram information (step S102).Then, the peak interval detection unit 22 detects at least two of the interval between two consecutive peaks of different types and the interval between two consecutive peaks of the same type (step S103).

[0048] The selection unit 23 of the processor 4 selects a set of weighting coefficients corresponding to the attributes of the subject represented in the input attribute information from among the sets of weighting coefficients for each of a plurality of attributes registered in advance, and reads the selected set of weighting coefficients from the memory 3 (step S104).

[0049] The abnormality determination unit 24 of the processor 4 inputs the frequency spectrum of the electrocardiogram waveform and one or more peak intervals to the classifier for each frame. The classifier then weights the frequency spectrum and each peak interval using the selected set of weighting coefficients, and determines whether or not there is an abnormality in the behavior of the subject's heart based on the feature vector obtained by weighting (step S105). The abnormality determination unit 24 then outputs information representing the detection result regarding the presence or absence of an abnormality in the behavior of the subject's heart to a user interface device or other device via the communication interface 2 (step S106). The processor 4 then ends the electrocardiogram information analysis process. The processor 4 may change the order of the process of step S101, the processes of steps S102 and S103, and the process of step S104, or may execute these processes in parallel.

[0050] As described above, this electrocardiogram information analysis device inputs the frequency spectrum of the electrocardiogram waveform of the subject and one or more types of peak intervals of the electrocardiogram waveform to a classifier that has been trained in advance to determine the presence or absence of abnormality in the behavior of the heart, and determines the presence or absence of abnormality in the behavior of the heart of the subject. The frequency spectrum of the electrocardiogram waveform and the peak intervals of the electrocardiogram waveform vary depending on the state of the behavior of the heart. Therefore, this electrocardiogram information analysis device uses the frequency spectrum of the electrocardiogram waveform and the peak intervals of the electrocardiogram waveform as inputs to the classifier. Furthermore, this electrocardiogram information analysis device selects a set of weighting coefficients for weighting the frequency spectrum and each peak interval input to the classifier according to the attributes of the subject. With these, this electrocardiogram information analysis device can accurately detect abnormality in the behavior of the heart of the subject. Furthermore, this electrocardiogram information analysis device inputs data, such as the frequency spectrum and peak intervals of the electrocardiogram waveform, which are not very large in data volume, to the classifier, and therefore can reduce the amount of calculation of the classifier. Therefore, this electrocardiogram information analysis device can reduce the amount of calculation of the entire electrocardiogram information analysis process.

[0051] As described above, those skilled in the art can make various modifications to the embodiments within the scope of the present invention. [Explanation of symbols]

[0052] 1 Electrocardiogram information analysis device 2. Communication Interface 3. Memory 4 processors 11 Electrocardiograph 21 Time-frequency conversion section 22 Peak interval detector 23 Selection section 24 Abnormality determination section

Claims

[Claim 1] a storage unit that stores, for each of a plurality of preregistered attributes, a set of weighting coefficients for at least one of a frequency spectrum of an electrocardiogram waveform and an interval between peaks of the same type and an interval between peaks of different types of the electrocardiogram waveform; a time-frequency conversion unit that performs time-frequency conversion on an electrocardiogram waveform, which is represented in electrocardiogram information obtained by an electrocardiograph attached to a subject and indicates the behavior of the subject's heart, for each frame having a predetermined time length, to calculate the frequency spectrum of the electrocardiogram waveform for each frame; a peak interval detection unit that detects, for each frame, at least one of an interval between peaks of the same type and an interval between peaks of different types in the electrocardiogram waveform; a selection unit that selects a set of weighting coefficients corresponding to the attributes of the subject from among the sets of weighting coefficients for the plurality of preregistered attributes based on attribute information representing the attributes of the subject; an abnormality determination unit that determines the presence or absence of an abnormality in the behavior of the heart of the subject by inputting the frequency spectrum of the electrocardiogram waveform for each frame and the at least one peak interval to a classifier that has been trained in advance to determine the presence or absence of an abnormality in the behavior of the heart; having the classifier has a weighting layer that calculates a feature vector weighted by the set of weighting coefficients selected for the frequency spectrum of the electrocardiogram waveform for each frame and the at least one peak interval, at least one hidden layer that performs a full-connection operation on the feature vector calculated by the weighting layer, and an output layer that calculates an output value indicating the presence or absence of an abnormality in the behavior of the heart of the subject based on an output from the at least one hidden layer, the weighting layer is an attention layer that generates a key vector, a query vector, and a value vector by weighting the frequency spectrum of the electrocardiogram waveform for each frame and the at least one peak interval with the selected set of weighting coefficients, and calculates the feature vector from the generated key vector, query vector, and value vector. Electrocardiogram information analysis device.

Images