Biosignal processing device and method for generating data for electrocardiogram analysis

Abstract

To provide a biosignal processing device and a method for generating electrocardiogram analysis data that can efficiently detect arrhythmias through automated electrocardiogram analysis using deep learning. [Solution] This is a biosignal processing device that generates electrocardiogram analysis data used in electrocardiogram analysis using deep learning. The biosignal processing device has an acquisition means for acquiring electrocardiogram data relating to multiple types of leads, and a generation means for generating electrocardiogram analysis data based on the electrocardiogram data. The generation means generates data representing a two-dimensional image having a waveform image region representing waveform information relating to multiple types of leads, and a rhythm image region representing rhythm information relating to a predetermined lead among the multiple types of leads, as electrocardiogram analysis data.

Description

Technical Field

[0001] The present invention relates to a biological signal processing device and a method for generating electrocardiogram analysis data.

Background Art

[0002] In recent years, the practical application of AI technology has advanced rapidly, and methods have been proposed for automatically analyzing electrocardiogram waveforms by treating them as time-series data or images and using machine learning or deep learning models. Also, the applicant previously proposed a method for generating electrocardiogram analysis data that enables efficient deep learning of the relationship between multiple types of waveform information and leads in an electrocardiogram (Patent Document 1).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In Patent Document 1, an efficient deep learning regarding waveform information and the mutual relationship between leads was realized by generating a two-dimensional image in which images of multiple types of lead waveforms were arranged as electrocardiogram analysis data. However, the electrocardiogram analysis data proposed in Cited Document 1 does not include information in the time axis direction of the electrocardiogram, more specifically, information regarding the rhythm of the heartbeat. Therefore, it cannot be used for automatic analysis that requires information regarding the rhythm of the heartbeat, such as arrhythmia.

[0005] On the other hand, when attempting to perform deep learning on electrocardiogram waveforms as time-series data rather than images, in order to learn data of a certain length (for example, 10 seconds), it is necessary to construct a model with a considerably large number of layers. Therefore, not only a large amount of memory is required, but the computational load is extremely large, and it is not practical to implement it in medical devices such as electrocardiographs.

[0006] This invention has been made in view of the problems of the prior art, and aims to provide a biosignal processing device and a method for generating electrocardiogram analysis data that can efficiently detect arrhythmias by automatically analyzing electrocardiograms using deep learning. [Means for solving the problem]

[0007] The above objective is achieved by a biosignal processing device that generates electrocardiogram analysis data for use in electrocardiogram analysis using deep learning, comprising: acquisition means for acquiring electrocardiogram data relating to multiple types of leads; and generation means for generating electrocardiogram analysis data based on the electrocardiogram data, wherein the generation means generates data representing a two-dimensional image having a waveform image region representing waveform information relating to multiple types of leads and a rhythm image region representing rhythm information relating to a predetermined lead among the multiple types of leads, as electrocardiogram analysis data. [Effects of the Invention]

[0008] The electrocardiogram analysis data obtained by the biosignal processing device and electrocardiogram analysis data generation method according to the present invention can efficiently represent not only the characteristics of the electrocardiogram waveform but also the characteristics of rhythmic changes over a predetermined period of time. Therefore, it becomes possible to efficiently detect arrhythmias by automated electrocardiogram analysis using deep learning. [Brief explanation of the drawing]

[0009] [Figure 1] This is a block diagram showing an example of the functional configuration of a biosignal processing device according to an embodiment. [Figure 2] This is a flowchart relating to the data generation process for electrocardiogram analysis according to the embodiment. [Figure 3] This is a flowchart relating to the data generation process for electrocardiogram analysis according to the embodiment. [Figure 4] This is a flowchart relating to the data generation process for electrocardiogram analysis according to the embodiment. [Figure 5]This figure shows a specific example of the pretreatment according to the embodiment. [Figure 6] This figure shows a specific example of electrocardiogram analysis data according to the embodiment. [Figure 7] This figure shows a specific example of electrocardiogram analysis data according to the embodiment. [Figure 8] This figure shows an example of the evaluation results of a learning model using electrocardiogram analysis data according to the embodiment. [Modes for carrying out the invention]

[0010] The present invention will now be described in detail based on exemplary embodiments with reference to the accompanying drawings. The embodiments described below do not limit the present invention in any sense. Furthermore, not all configurations described in the embodiments are essential to the present invention. Also, configurations included in different embodiments may be combined or replaced, unless it is clearly impossible or denied. In order to avoid redundant explanations, identical or similar components are given the same reference numerals throughout the accompanying drawings.

[0011] Figure 1 is a block diagram showing an example of the functional configuration of the biosignal processing device 100 according to this embodiment. The biosignal processing device 100 can be implemented, for example, by executing an application program that performs the operations described later using a programmable processor. Therefore, the biosignal processing device can be implemented in electronic devices in general that have a programmable processor.

[0012] The biosignal processing device 100 generates data suitable for deep learning of a neural network and automated analysis using a trained neural network (ECG analysis data) from time-series ECG data containing multiple leads (hereinafter referred to as ECG data). In particular, the biosignal processing device 100 of this embodiment generates ECG analysis data that enables the detection of arrhythmias through automated analysis using a neural network. The ECG analysis data generated by the biosignal processing device is used during neural network training and automated analysis using a trained neural network.

[0013] The control unit 110 has a programmable processor, ROM, and RAM, and by reading programs stored in the ROM and recording unit 130 into the RAM and executing them, it realizes the processing of the biosignal processing device 100, including the electrocardiogram analysis data generation process described later.

[0014] The external I / F 120 is an interface for the biosignal processing device 100 to communicate with an external device via wired and / or wireless communication. Through the external I / F 120, the biosignal processing device 100 can output analysis results to an external device or acquire data such as electrocardiogram data from an external device. The external I / F 120 can have a configuration conforming to one or more standards for inter-device communication, such as USB, wireless LAN, wired LAN, or Bluetooth®.

[0015] The recording unit 130 is a device for storing electrocardiogram data to be analyzed, the results of the electrocardiogram data analysis, etc., and may be an internal storage device such as an SSD or HDD, and / or a removable storage device such as a USB memory or memory card. The control unit 110 records data to the recording unit 130 and reads data recorded to the recording unit 130.

[0016] The signal processing unit 140 applies predetermined processing to biological signals including electrocardiogram data. The signal processing unit 140 can be composed of, for example, a programmable processor, a ROM that stores a program executable by the processor, and a RAM used for program execution. The programmable processor used by the signal processing unit 140 may be a general CPU, but in order to execute signal processing and neural network-related processing (application of deep learning and trained neural networks) at high speed, a GPU, a DSP, etc. may be used at least as auxiliary.

[0017] The signal processing unit 140 executes deep learning processing of a neural network using electrocardiogram analysis data generated from electrocardiogram data for learning, and automatic analysis processing that applies electrocardiogram analysis data generated from electrocardiogram data to be analyzed to a neural network that has been trained in deep learning. In this specification, deep learning means learning of a neural network having multiple layers.

[0018] The display unit 150 is a display device such as a liquid crystal display device (LCD), and displays a user interface (GUI) of the biological signal processing device 100, biological information such as an electrocardiogram waveform, an automatic analysis result of electrocardiogram data, etc. The display unit 150 may be an external display device.

[0019] The operation unit 160 is a general term for input devices such as a keyboard, a pointing device, a touch panel, a switch, a button, etc. for the user to input instructions to the biological signal processing device 100. When the display unit 150 is a touch display, the software keys displayed on the display unit 150 constitute a part of the operation unit 160.

[0020] Figure 2 is a flowchart of electrocardiogram analysis data generation processing executed by the biological signal processing device 100. In S210, the control unit 110 acquires electrocardiogram data for multiple leads (for example, 12-lead electrocardiogram data) from, for example, the recording unit 130. Here, it is assumed that electrocardiogram data that has been measured in advance and stored in the recording unit 130 is acquired, but if the biosignal processing device 100 is incorporated into the electrocardiograph, electrocardiogram data may be acquired in real time.

[0021] As will be described in detail later, in this embodiment, two-dimensional image data is generated as electrocardiogram analysis data, having an image region representing the waveform for each lead (waveform image region) and an image region representing rhythm information for a predetermined lead (rhythm image region). In S210, electrocardiogram data for generating the waveform image is acquired.

[0022] In order to generate a waveform image for one heartbeat for each lead, the control unit 110 acquires electrocardiogram data for one heartbeat in S210. Alternatively, the control unit 110 may generate electrocardiogram data for one heartbeat by dividing the electrocardiogram data for a predetermined time into one-heartbeat segments and averaging them. The control unit 110 may also extract the electrocardiogram data for one heartbeat that is judged to be of the best quality from the electrocardiogram data for the predetermined time. The electrocardiogram data for each lead obtained in S210 is called representative electrocardiogram data. Note that the representative electrocardiogram data for each lead may be longer than one heartbeat.

[0023] Furthermore, the division points for dividing electrocardiogram data into individual heartbeats or into multiple segments within a single heartbeat are determined using one of several types of leads and can be used commonly across all leads. The lead used to determine the division points may be predetermined, or the lead with the largest maximum amplitude may be used. Electrocardiogram data can be divided into individual heartbeat units using any known method.

[0024] In S220, the control unit 110 applies preprocessing to the electrocardiogram data for one heartbeat in each lead. Preprocessing is a process that reduces the amount of electrocardiogram data while preserving the waveform features. Details of the preprocessing will be described later. Preprocessing is not mandatory, but performing it can reduce the computational load of deep learning and automated analysis, and improve the efficiency of extracting waveform features.

[0025] In S230, the control unit 110 generates a composite waveform image by arranging the pre-processed electrocardiogram data images (lead images) of each lead in a two-dimensional manner. The control unit 110 also acquires electrocardiogram data for a predetermined time for one of the multiple leads and extracts predetermined rhythm information. It then generates a rhythm image representing the extracted rhythm information. The rhythm information includes the length of a specific interval within one heartbeat in the electrocardiogram, and the interval between adjacent heartbeats of feature points. Details of the extraction of rhythm information and the rhythm image will be described later.

[0026] The control unit 110 then generates two-dimensional image data in which the composite waveform image is placed in the waveform image area and the rhythm image is placed in the rhythm image area. Details of the image data generation process in S230 will be described later.

[0027] In S240, the control unit 110 applies post-processing to the electrocardiogram analysis data generated in S230, as necessary. The post-processing may be, for example, pixel compression using a known method such as the bicubic method. However, if the processing capacity of the signal processing unit 140 is sufficiently high, post-processing may not necessarily be performed. Therefore, the 2D image data generated in S230, or the 2D image data to which post-processing has been applied in S240, becomes the electrocardiogram analysis data. The control unit 110 records the generated electrocardiogram analysis data, for example, in the recording unit 130. At least some of the above-described processes may be performed by the signal processing unit 140 instead of the control unit 110.

[0028] When the signal processing unit 140 performs deep learning (or validation) of a neural network, the control unit 110 supplies the signal processing unit 140 with electrocardiogram analysis data generated from training (or validation) electrocardiogram data (known electrocardiogram data). Furthermore, when the signal processing unit 140 performs automated analysis using a trained neural network, the control unit 110 supplies the signal processing unit 140 with electrocardiogram analysis data generated from analysis electrocardiogram data (unknown electrocardiogram data).

[0029] Next, we will explain the details of the preprocessing performed in S220 of Figure 2 using the flowchart shown in Figure 3. The preprocessing is applied to the representative electrocardiogram data for each lead obtained in S210.

[0030] In S222, the control unit 110 applies nonlinear amplification processing to the electrocardiogram data. Nonlinear amplification processing applies a large gain to the portion of the electrocardiogram data where the amplitude value (difference from the baseline level) is small, and a small gain to the portion where the amplitude value is large. There are no particular restrictions on the characteristics of nonlinear amplification, but a logarithmic gain curve can be used. Nonlinear amplification can also be described as nonlinear compression of the dynamic range of the electrocardiogram data.

[0031] In S224, the control unit 110 normalizes the nonlinearly amplified electrocardiogram data. For example, the electrocardiogram data for each lead may be normalized using the maximum value of all leads, or each lead may be normalized using the maximum value of that lead. After normalization, the electrocardiogram data for each lead will have values ​​ranging from -1 to 1.

[0032] In S226, the control unit 110 applies downsampling to the normalized electrocardiogram data for each lead to reduce the number of data points (samples) in the time axis direction. Downsampling is not mandatory, but it has the effect of reducing the load on learning and analysis.

[0033] When performing downsampling, it is preferable to suppress the reduction in the number of samples by setting a higher sampling rate in sections with large waveform changes than in sections with small changes, rather than keeping the sampling rate constant. For example, the control unit 110 divides the duration of one heartbeat into a QRS interval (for example, a first interval of a predetermined number of samples (time) before and after the peak of the R wave), a P interval (for example, a second interval of a predetermined number of samples (time) before the QRS interval), and a third interval. The sampling rates for downsampling can then be varied so that the proportion of samples to be downsampled is greater in the third interval > second interval > first interval. Furthermore, within an interval, instead of downsampling samples at equal intervals, the samples to be downsampled may be adjusted so that, for example, adjacent samples with large differences in value are not downsampled. Although a simple sample downsampling configuration is used here, subsampling with interpolation may also be performed.

[0034] In S228, the control unit 110 trims (deletes) data from unimportant sections of the electrocardiogram data as needed. For example, the control unit 110 can delete data from the section between the end of the T wave and the start of the P wave within a single heartbeat period. While trimming is not mandatory, it reduces the amount of data, thereby lowering the computational load during training and automatic analysis. Trimming may also be performed before downsampling.

[0035] The subsampling rate and the number of samples to trim during preprocessing can be determined by the size of the waveform image region in the ECG analysis data and the size of the waveform image per lead, which is determined by the number of leads whose waveform images are included in the waveform image region. This allows for efficient generation of waveform images.

[0036] After this preprocessing is performed, the details of the image data generation process carried out in S230 will be explained using the flowchart in Figure 4. In S232, the control unit 110 generates image data (lead image data) for each lead from pre-processed representative electrocardiogram data. In this embodiment, rectangular lead image data having a predetermined common size (number of pixels in the horizontal and vertical directions) is generated from the electrocardiogram data of each lead.

[0037] There are no particular restrictions on how the values ​​of each sample that make up the electrocardiogram data (i.e., information about the waveform of the leads) are represented as lead image data, and various methods are possible. Here, we assume that lead image data in which the waveform is plotted in a Cartesian coordinate space is generated by setting the horizontal direction of the rectangular region as the time axis and the vertical direction as the amplitude axis, and making the pixel values ​​of the coordinates determined according to each sample value different from the background pixel values.

[0038] For simplicity, we assume that the horizontal size of the lead image is equal to the number of pixels corresponding to the number of samples in the electrocardiogram data. If the electrocardiogram data has been normalized in preprocessing, the vertical size of the lead image is assigned a value range from +1 to -1 to determine the vertical coordinate corresponding to each sample value. The horizontal coordinate can be increased by 1 for each sample. By setting the pixel value corresponding to the coordinate of each sample value determined in this way to 1 and the pixel value corresponding to the other coordinates to 0, lead image data is generated with the sample values ​​plotted in white on a black background. Note that the lead image data may also be a multi-level image.

[0039] If the number of ECG data samples is less than the horizontal size (number of pixels) of the generated image, the horizontal size of the image may be reduced to match the number of samples, or areas where no samples are plotted may be left as background. When performing preprocessing, it is preferable to adjust the subsampling rate and the number of samples to be trimmed to match the number of ECG data samples to the horizontal size (number of pixels) of the generated image.

[0040] As described above, once lead image data of the same image size has been generated from the electrocardiogram data of each lead, in S234 the control unit 110 generates data for a composite waveform image by arranging the lead images in two dimensions. The size of the composite waveform image is determined so that the data for automatic analysis, combined with the rhythm image, becomes a square image. This is because many commonly available neural network program libraries are designed to handle square images.

[0041] Next, in S236, the control unit 110 generates rhythm image data. The control unit 110 acquires electrocardiogram data for a predetermined lead from the recording unit 130 for a predetermined time and extracts rhythm information. The electrocardiogram data acquired here is the electrocardiogram data for one of the leads acquired in S210 for generating the lead image. The electrocardiogram data for any lead may be acquired, or the lead with the dominant heartbeat may be selected from among multiple leads, or the lead to be used may be set in advance.

[0042] For example, suppose that in S210, 10 seconds of electrocardiogram data measured simultaneously for each lead is acquired, and representative electrocardiogram data is generated. In this case, in S236, the control unit 110 acquires the same 10 seconds of electrocardiogram data for one lead as acquired in S210, and extracts rhythm information.

[0043] Rhythm information is information about the timing of an electrocardiogram, such as the length of a specific interval within a single heartbeat or the interval between feature points in adjacent heartbeats. In this embodiment, rhythm information particularly useful for detecting arrhythmias is extracted. Such rhythm information includes, but is not limited to, the RR interval, PR interval, QRS interval length, and QT interval length. The RR interval is the interval between R wave peaks in adjacent heartbeats. The PR interval is the interval between the peaks of the P wave and R wave within a single heartbeat. The QRS interval length is the length from the Q wave to the S wave within a single heartbeat. The QT interval length is the length from the Q wave to the T wave peak within a single heartbeat. Since known methods can be used for dividing the electrocardiogram data into individual beats and for detecting feature points, a detailed explanation is omitted.

[0044] The control unit 110 generates rhythm image data representing the extracted rhythm information. The size of the rhythm image is determined so that when combined with the waveform image, it becomes an image of a predetermined size. Specific examples of rhythm images will be described later. Once the rhythm image data is generated, the control unit 110 proceeds to processing S238.

[0045] In S238, the control unit 110 combines the waveform image data generated in S234 and the rhythm image data generated in S236 to generate composite image data. The position and distribution of the waveform image and rhythm image within the composite image can be freely adjusted. As described above, the composite image data generated in S238 may be used as is for electrocardiogram analysis, or it may be used as electrocardiogram analysis data after post-processing has been applied.

[0046] Figure 5 shows a specific example of preprocessing. Figure 5(a) shows the electrocardiogram data for one heartbeat, with a sample size of 500. Figure 5(b) shows the result after applying nonlinear amplification to this electrocardiogram data and normalizing it by the maximum value of all leads. In this state, the sample size remains unchanged.

[0047] For downsampling, the ECG is divided into P-wave, QRS, T-wave, and other segments, with sampling rates set to DSR2, DSR3, DSR4, and DSR1, respectively. Here, the proportion of samples removed by downsampling follows the relationship DSR1 > DSR4 >= DSR2 > DSR3. In this case, the segments with downsampling rate DSR1 are removed by trimming, while DSR3 is reduced to 1 / 2, DSR2 to 1 / 6, and DSR4 to 1 / 6. Here, a downsampling rate of 1 / n indicates that the number of samples is reduced to 1 / n of the original number by downsampling. Figure 5(c) shows the ECG data after subsampling and trimming. The number of samples has been reduced from 500 to 84, but the waveform characteristics related to the leads are preserved.

[0048] Figure 6(a) shows an example of a two-dimensional image represented by electrocardiogram analysis data generated by the biosignal processing device 100. The two-dimensional image 400 has a waveform image region 401 and a rhythm image region 402. In the example shown in Figure 6(a), the two-dimensional image 400 is square in shape, but it may also be rectangular. However, a square shape offers higher applicability to two-dimensional neural networks.

[0049] First, the waveform image region 401 will be explained further using Figure 7. Figure 7 shows an example of a waveform image region obtained by applying the preprocessing shown in Figure 5 to representative electrocardiogram data of a standard 12-lead ECG, and then synthesizing the waveform images generated for each lead. Here, the size of the 2D image 400 shown in Figure 6(a) before post-processing is 360 × 360 pixels, the size of the waveform image region 401 is 252 vertical × 360 horizontal pixels, and the size of the rhythm image region 402 is 108 vertical × 360 horizontal pixels. The size of the waveform image for each lead in the waveform image region 401 is 63 vertical × 120 horizontal pixels, which is obtained by dividing the 252 vertical × 360 horizontal pixels horizontally into 3 horizontal divisions and 4 vertical divisions.

[0050] Figure 6(b) is a magnified view of the rhythm image region 402 in Figure 6(a). For clarity, the rhythm image is shown in gray. The rhythm image shown in Figure 6 is an example of imaging four rhythm information components from 10 seconds of electrocardiogram data: RR interval, PR interval, QRS interval length, and QT interval length.

[0051] Here, four types of rhythm information are visualized using three types of rectangular or linear image patterns 403-405. What image patterns 403-405 have in common is that one side of the rhythm image (in this case, the side extending horizontally) is used as the time axis, and the position is determined along that side.

[0052] The vertically elongated rectangular image pattern 403 represents the length of the QRS interval by its length in the time axis direction, and the timing of the R wave by its position in the time axis direction. The reason why there are 10 image patterns 403 lined up in the time axis direction is that the 10 seconds of electrocardiogram data from which rhythm information was extracted contained 10 heartbeat-length QRS intervals. Therefore, the interval between adjacent image patterns 403 corresponds to the RR interval. Note that the image patterns 403 may be arranged so that the distance between centers in the time axis direction represents the RR interval, or so that the distance between one end in the time axis direction (the left end in the example of Figure 6(b)) represents the RR interval. Note that the vertical size of image pattern 403 does not have any particular significance, but since the length of the RR interval and QRS interval is considered to be more important than the length of the PR interval and QT interval in arrhythmia detection, the vertical size is larger than that of image patterns 404 and 405.

[0053] The rectangular image pattern 404, which is elongated in the time axis direction, represents the length of the QT interval in terms of its length in the time axis direction. The starting position of image pattern 404 is aligned with that of image pattern 403 for the same heartbeat in the time axis direction. This is because the starting positions of image pattern 403 and image pattern 404 correspond to the same Q wave start timing. The vertical size of image pattern 404 is fixed and does not represent any particular meaning.

[0054] The image pattern 405, which is long vertically and has a fixed length in the time axis direction, represents the timing of the P wave based on its position in the time axis direction. The position of image pattern 405 in the time axis direction is determined so that the interval in the time axis direction between it and image pattern 403 relating to the same heartbeat represents the PR interval. The vertical size of image pattern 405 is fixed and does not represent any particular meaning.

[0055] Here, since image pattern 405 representing the timing of the P wave and image pattern 404 representing the length of the QT interval are placed adjacent to each other in the time axis direction, one image pattern is made vertically elongated and the other horizontally elongated to prevent them from being mistaken for different image patterns. However, other methods may be used to suppress misidentification of image patterns. For example, as shown in Figure 6(c), image patterns 404 and 405 may be placed vertically spaced apart, like image patterns 403 and 404.

[0056] Note that the method for visualizing rhythm information is not limited to the method described using Figure 6. If rhythm information (numerical values) for multiple heartbeats can be represented as different images depending on the notification, rhythm images can be generated by other methods.

[0057] As described above, according to this embodiment, a single two-dimensional image is generated as electrocardiogram analysis data, which combines waveform images for multiple leads with rhythm images that visualize rhythm information related to the electrocardiogram for a predetermined period of time. By using the electrocardiogram analysis data generated by this embodiment, it becomes possible to efficiently detect arrhythmias based on lead waveforms and rhythm information through automated analysis using deep learning.

[0058] The following describes an example of applying the electrocardiogram analysis data from this embodiment to a two-dimensional convolutional neural network (2D CNN). The parameters for the 2D CNN used for evaluation are as follows. Number of layers: 10~16 layers Kernel size: 3x3 Activation function: ReLU Batch size: 64 Number of Epochs: 20-60 Dropout rate: 0.5

[0059] Figure 8 shows that 767 12-lead electrocardiogram (ECG) data points each for sinus rhythm and arrhythmia were prepared, and data for automatic analysis was generated using the method of the embodiment described above. After training with 537 of these data points for automatic analysis, the results of validation were performed on 230 of these data points for ECG analysis.

[0060] For electrocardiogram analysis, 2D images of 360x360 pixels were created, and then post-processed to reduce them to 120x120 pixels and 90x90 pixels. In Figure 8, the learning model was evaluated using precision, recall, and the F1 score. The F1 score is calculated as 2 × precision × recall / (precision + recall).

[0061] The evaluation results shown in Figure 8 demonstrate that even at a stage where the learning model has not been fully optimized (loss = 0.1 to 0.2), it is possible to accurately detect arrhythmias in electrocardiogram data. Furthermore, no performance degradation was observed even when the size of the electrocardiogram analysis data was reduced, and discrimination can be performed at a speed that is sufficiently practical even in execution environments with limited computing power and memory capacity. In a typical personal computer test using 90x90 pixel electrocardiogram analysis data, the model size was 4.5MB, and the time required to discriminate 230 electrocardiogram analysis data points was approximately 0.13 seconds. Therefore, the discrimination time per electrocardiogram analysis data point was approximately 0.57 milliseconds, and even considering the time required to generate the electrocardiogram analysis data, real-time or near-real-time processing is achievable.

[0062] As described above, according to this embodiment, two-dimensional image data is generated as electrocardiogram analysis data using deep learning, which includes waveform images for multiple types of leads and rhythm images based on rhythm information for a predetermined lead. This makes it possible to detect arrhythmias that cannot be identified from waveform information alone accurately and quickly through automated analysis using a deep learning model.

[0063] The invention is not limited to the embodiments described above, and various modifications and changes are possible within the scope of the gist of the invention. For example, although the above embodiments described the case using 12-lead electrocardiogram data, the present invention is applicable to electrocardiogram data relating to any multiple types of leads. For example, waveform images may be generated from electrocardiogram data relating to three types of leads, such as lead II, lead V5, and lead V2. It is also possible to generate a composite image by arranging waveform images from single-lead or few-lead electrocardiograms, such as Holter or monitor electrocardiograms, with multiple rhythm images obtained by dividing a long recording time.

[0064] Furthermore, the biosignal processing device according to the present invention can also be realized by running a program (application software) that executes the operations of the flowcharts shown in Figures 2 to 4 using a generally available electronic device capable of executing programs, such as a personal computer, smartphone, or tablet terminal. Accordingly, such a program and a storage medium that stores the program (such as an optical recording medium like a CD-ROM or DVD-ROM, a magnetic recording medium like a magnetic disk, or a semiconductor memory card) also constitute the present invention. [Explanation of symbols]

[0065] 100...Biosignal processing unit, 110...Control unit, 140...Signal processing unit

Claims

1. A biosignal processing device that generates electrocardiogram analysis data for use in electrocardiogram analysis using deep learning, A means for acquiring electrocardiogram data related to multiple types of leads, It includes a generation means for generating electrocardiogram analysis data based on the electrocardiogram data, The biosignal processing device is characterized in that the generation means generates data representing a two-dimensional image having a waveform image region representing waveform information relating to multiple types of leads and a rhythm image region representing rhythm information relating to a predetermined lead among the multiple types of leads, as data for electrocardiogram analysis.

2. The biosignal processing apparatus according to claim 1, characterized in that the generation means generates data representing a two-dimensional image in which an image pattern having a position and length in the direction of the time axis determined according to the rhythm information is arranged in the rhythm image region, with one side of the two-dimensional image as the time axis, as electrocardiogram analysis data.

3. The aforementioned rhythm information includes the time interval between predetermined characteristic points in the electrocardiogram of one heartbeat. The biosignal processing apparatus according to claim 1 or 2, characterized in that the generation means generates data representing a two-dimensional image in which an image pattern showing the feature points is arranged in the rhythm image region at intervals corresponding to the time interval, as data for electrocardiogram analysis.

4. The biosignal processing apparatus according to claim 3, characterized in that the aforementioned feature point is at least one of the peaks of the P wave and the R wave.

5. The aforementioned rhythm information includes the length of a predetermined interval in the electrocardiogram within one heartbeat. The biosignal processing apparatus according to any one of claims 1 to 4, characterized in that the generation means generates data representing a two-dimensional image in which an image pattern having a length corresponding to the length is arranged in the rhythm image region, as data for electrocardiogram analysis.

6. The biosignal processing apparatus according to claim 5, characterized in that the predetermined interval is at least one of a QRS interval and a QT interval.

7. The biosignal processing device according to any one of claims 1 to 6, characterized in that the generation means generates data representing a two-dimensional image in which an image pattern having a length corresponding to the length of the QRS interval in each heartbeat included in the electrocardiogram data relating to the predetermined lead is arranged in the rhythm image region at intervals corresponding to the R-R interval between adjacent heartbeats, the data representing the electrocardiogram analysis data.

8. The biosignal processing device according to claim 7, characterized in that the generation means generates data representing a two-dimensional image in which an image pattern indicating the timing of the P wave is arranged in the rhythm image region at intervals corresponding to the P-R interval, starting from a position indicating the timing of the R wave in an image pattern having a length corresponding to the length of the QRS interval, with intervals corresponding to the P-R interval, as data for electrocardiogram analysis.

9. The biosignal processing apparatus according to claim 7 or 8, characterized in that the generation means generates data representing a two-dimensional image in which an image pattern having a length corresponding to the QT interval is arranged in the rhythm image region such that it represents the timing of the Q wave at the same position as the image pattern having a length corresponding to the length of the QRS interval, and does not overlap with the image pattern having a length corresponding to the length of the QRS interval, as data for electrocardiogram analysis.

10. The biosignal processing apparatus according to any one of claims 1 to 9, characterized in that the generation means generates data representing a two-dimensional image in which images representing the values ​​of the electrocardiogram data as waveforms are arranged in the waveform image area for each of the plurality of leads, as data for electrocardiogram analysis.

11. The biosignal processing apparatus according to claim 10, characterized in that the generation means generates an image in which the values ​​of the electrocardiogram data are represented as waveforms based on values ​​obtained by nonlinear compression of the values ​​of the electrocardiogram data.

12. The biosignal processing device according to claim 11, characterized in that the generation means normalizes the nonlinearly compressed values ​​for each lead and then generates an image in which the values ​​of the electrocardiogram data are represented as waveforms.

13. The biosignal processing apparatus according to claim 12, characterized in that the generation means applies downsampling to the electrocardiogram data having the normalized values ​​and then generates an image in which the values ​​of the electrocardiogram data are represented as waveforms.

14. The biosignal processing device according to claim 13, characterized in that the downsampling has a sampling rate corresponding to the interval obtained by dividing one heartbeat period of the electrocardiogram.

15. The biosignal processing apparatus according to any one of claims 10 to 14, characterized in that the generation means reduces the number of pixels by post-processing the two-dimensional image to obtain data for electrocardiogram analysis.

16. A method for generating electrocardiogram analysis data used in electrocardiogram analysis using deep learning, The acquisition process involves obtaining electrocardiogram data for multiple types of leads, The process includes a generation step of generating electrocardiogram analysis data based on the electrocardiogram data, A method for generating electrocardiogram analysis data, characterized in that the generation step generates data representing a two-dimensional image having a waveform image region representing waveform information relating to multiple types of leads and a rhythm image region representing rhythm information relating to a predetermined lead among the multiple types of leads, as electrocardiogram analysis data.

17. A program for causing a computer to function as one of the means of a biosignal processing device according to any one of claims 1 to 15.

18. ECG analysis data used for ECG analysis using deep learning, ECG analysis data characterized by representing a two-dimensional image having a waveform image region that represents waveform information for multiple types of leads and a rhythm image region that represents rhythm information for a predetermined lead among the multiple types of leads.

19. The electrocardiogram analysis data according to claim 18, characterized in that each of the aforementioned multiple leads includes a two-dimensional image representing the time-series data values ​​as waveforms.

20. A method for learning a neural network, characterized by using electrocardiogram analysis data as described in claim 18 or 19.

21. An electrocardiogram analysis device that performs automatic electrocardiogram analysis using a neural network trained with electrocardiogram analysis data as described in claim 18 or 19.

Images