Biological signal analysis system

Abstract

The biological signal analysis system according to the present application is a series of biological signal analysis systems for determining arrhythmia based on electrocardiographic information, and includes: an input unit that receives input of electrocardiographic information; a data conversion unit that generates determination data by performing time-frequency conversion on the electrocardiographic information; a data restoration unit that generates restored data obtained by restoring the determination data using a learned model composed of a self-encoder-decoder; a difference calculation unit that calculates a difference between the restored data and the determination data; and a determination unit that determines which of normal waveform data and arrhythmia waveform data the determination data is based on the difference.

Description

TECHNICAL FIELD

[0001] The present application relates to a biological signal analysis system that determines normal / arrhythmia, for example, based on electrocardiographic signals (hereinafter, electrocardiographic information). BACKGROUND

[0002] The heart repeatedly contracts and expands to function as a pump that delivers blood to the whole body. This activity is kept regular by a weak electric stimulus of myocardial cells, but if this electric stimulus appears abnormal, the activity of the heart also becomes abnormal. This situation is called arrhythmia. Arrhythmia has various kinds, for example, fatal ventricular fibrillation that causes sudden cardiac arrest, ventricular tachycardia, atrial fibrillation that causes cerebral infarction. Thus, it is very important to determine arrhythmia, and a process of determining arrhythmia using electrocardiographic information acquired from a determination object by a biological signal measuring device is performed.

[0003] However, for the determination of arrhythmia, a doctor views long-term and continuous electrocardiographic information to make a diagnosis, and the burden on the doctor is large. Therefore, a biological signal analysis device and a method for automatically determining arrhythmia are used (for example, refer to Non-Patent Literature 1, 2).

[0004] The method described in Non-Patent Literature 1 is a method of determining atrial fibrillation, which is one of arrhythmias. Atrial fibrillation is an arrhythmia in which a site called the atrium is slightly shaken due to an electric stimulus, and features for determining from electrocardiographic information are listed as the generation of F waves, which are small-amplitude vibration components, and the abnormality of a heartbeat cycle (hereinafter, RR ratio). Non-Patent Literature 1 focuses on the latter, and in a histogram in which the RR ratio is taken as the horizontal axis, the deviation of atrial fibrillation is larger than that of normal sinus rhythm (hereinafter, normal waveform), so atrial fibrillation is determined from this situation.

[0005] In addition, the method described in Non-Patent Literature 2, like Non-Patent Literature 1, focuses on atrial fibrillation, but performs automatic analysis based on machine learning. Non-Patent Literature 2 segments the acquired electrocardiographic information every 15 seconds, makes the waveform features appear by performing time-frequency transformation thereon, and classifies into "normal waveform, noise, atrial fibrillation, and others" that have been learned to determine atrial fibrillation.

[0006] PRIOR ART DOCUMENTS

[0007] NON-PATENT LITERATURE

[0008] Non-Patent Literature 1: B. Logan and J. Healey, "Robust detection of atrial fibrillation for a long term telemonitoring system," Computers in Cardiology, 2005, Lyon, 2005, PP. 619-622, doi: 10.1109 / CIC.2005.1588177.

[0009] Non-Patent Literature 2: A. Qayyum, F. Meriaudeau and G. C. Y. Chan, "Classification of Atrial Fibrillation with Pre-Trained Convolutional Neural Network Models," 2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES), Sarawak, Malaysia, 2018, PP. 594-599, doi: 10.1109 / IECBES.2018.8626624. SUMMARY

[0010] PROBLEMS TO BE SOLVED BY THE INVENTION

[0011] However, in Non-Patent Literature 1 and Non-Patent Literature 2, the following problems exist.

[0012] In Non-Patent Literature 1, only arrhythmia in which abnormality of RR ratio has a characteristic can be determined. As described above, there are various kinds of arrhythmia, and since there are also arrhythmia in which RR ratio does not have a characteristic such as ventricular tachycardia, there is a case where such arrhythmia is missed. In addition, since abnormality of RR ratio is evaluated based on deviation, in order to determine atrial fibrillation, arrhythmia (deviation) needs to be continued for a long time, and it is difficult to determine arrhythmia that occurs suddenly.

[0013] In Non-Patent Literature 2, since learning is required in advance, determination of arrhythmia that has not been learned cannot be performed.

[0014] In the present application, the above problems of the related art are solved, and a biological signal analysis system capable of determining whether or not it is a series of arrhythmia based on electrocardio information regardless of the kind of arrhythmia is provided.

[0015] MEANS FOR SOLVING THE PROBLEMS

[0016] A series of biological signal analysis systems of the present application for achieving the above object are constructed of the following structures.

[0017] (1) A biological signal analysis system is a series of biological signal analysis systems that makes an arrhythmia determination based on electrocardiographic information, and includes:

[0018] an input unit that accepts input of the electrocardiographic information;

[0019] a data conversion unit that generates determination data by performing time-frequency conversion on the electrocardiographic information;

[0020] a data restoration unit that generates restoration data obtained by restoring the determination data using a completed model composed of a self-encoder-decoder;

[0021] a difference calculation unit that calculates a difference between the restoration data and the determination data; and

[0022] a determination unit that determines which of normal waveform data and arrhythmia waveform data the determination data is based on the difference.

[0023] (2) The biological signal analysis system according to (1), wherein

[0024] the completed model is a model generated by learning using 95% or more of known learning data generated based on normal waveform data.

[0025] (3) The biological signal analysis system according to (1), wherein

[0026] the completed model is a model generated by learning using 95% or more of known learning data generated based on normal waveform data and learning data acquired from a determination target.

[0027] (4) The biological signal analysis system according to (1), wherein

[0028] the completed model is a model generated by learning using learning data acquired from a determination target.

[0029] (5) The biological signal analysis system according to any one of (1) to (4), wherein

[0030] the input unit divides waveform data possessed by the electrocardiographic information into unit waveform data obtained by extracting waveform data in a prescribed time range,

[0031] the data conversion unit generates the determination data by performing time-frequency conversion on the unit waveform data.

[0032] (6) The biological signal analysis system according to any one of (1) to (5), wherein

[0033] The data transformation section generates weighted data obtained by weighting a time-frequency characteristic region of arrhythmia in the determination data.

[0034] (7) The biological signal analysis system according to any one of (1) to (6), wherein

[0035] The difference calculation section generates difference data indicating a difference between the recovered data and the determination data, and calculates the difference amount on the basis of the difference data,

[0036] The determination section determines a main cause of arrhythmia on the basis of a time and frequency characteristic of the difference data.

[0037] (8) The biological signal analysis system according to any one of (1) to (7), wherein

[0038] The learned model is a model generated by learning using learning data known to be the normal waveform data.

[0039] (9) The biological signal analysis system according to any one of (1) to (8), wherein

[0040] Further comprising a learning section that generates the learned model by learning of the autoencoder using learning data.

[0041] (10) The biological signal analysis system according to (2), wherein

[0042] Further comprising a learning section that generates the learned model by learning of the autoencoder using the learning data,

[0043] The learning section has a plurality of autoencoders corresponding to characteristics of the known learning data,

[0044] One of the autoencoders is selected from the plurality of autoencoders on the basis of a characteristic of normal waveform data of a determination object.

[0045] (11) The biological signal analysis system according to (3), wherein

[0046] Further comprising a learning section that generates the learned model by learning of the autoencoder using the learning data,

[0047] The learning section has a plurality of groups of learning data or a plurality of autoencoders corresponding to characteristics of the known learning data,

[0048] selecting one of the plurality of learning data groups or the plurality of autoencoders based on a feature of normal waveform data of the determination target.

[0049] Effects of the Invention

[0050] According to the present invention, it is possible to provide a series of biological signal analysis system capable of determining whether or not it is arrhythmia based on electrocardiographic information regardless of the type of arrhythmia. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a diagram showing a configuration example of the series of biological signal analysis system according to Embodiment 1.

[0052] Figure 2 is a diagram showing a detailed example of the electrocardiographic information used for analysis.

[0053] Figure 3 is a diagram showing a processing example of the input unit of the electrocardiographic information according to Embodiment 1.

[0054] Figure 4 is a flowchart showing a learning process according to Embodiment 1.

[0055] Figure 5 is a flowchart showing a determination process according to Embodiment 1.

[0056] Figure 6A is a diagram showing a processing example of the data conversion unit according to Embodiment 1 (one of).

[0057] Figure 6B is a diagram showing a processing example of the data conversion unit according to Embodiment 1 (two of).

[0058] Figure 7 is a diagram showing a structure of the autoencoder according to Embodiment 1.

[0059] Figure 8 is a diagram showing a processing example of the learning unit according to Embodiment 1.

[0060] Figure 9A is a diagram showing a processing example of the data restoration unit according to Embodiment 1 (one of).

[0061] Figure 9B is a diagram showing a processing example of the data restoration unit according to Embodiment 1 (two of).

[0062] Figure 10A is a diagram showing a processing example of the difference calculation unit according to Embodiment 1 (one of).

[0063] Figure 10B is a diagram showing a processing example of the difference calculation section according to Embodiment 1 (two).

[0064] Figure 11A is a diagram showing a processing example of the determination section according to Embodiment 1 (one).

[0065] Figure 11B is a diagram showing a processing example of the determination section according to Embodiment 1 (two).

[0066] Figure 12 is a diagram showing an analysis example based on the processing according to Embodiment 1.

[0067] Figure 13A is a diagram showing a weighting example of the arrhythmia-specific feature in the three-dimensional feature data according to Embodiment 6 (one).

[0068] Figure 13B is a diagram showing a weighting example of the arrhythmia-specific feature in the three-dimensional feature data according to Embodiment 6 (two).

[0069] Figure 14 is a diagram showing a specific example of the arrhythmia main cause based on the time-frequency coordinates according to Embodiment 7. DETAILED DESCRIPTION

[0070] Hereinafter, the embodiments of the series of biological signal analysis systems according to the present application will be described in detail based on the drawings. Furthermore, the present application is not limited to the embodiments. In addition, the respective embodiments of the present application are not independent, and can be combined as appropriate to be implemented.

[0071] [Embodiment 1]

[0072] Figure 1 is a diagram showing a structure example of the series of biological signal analysis systems according to Embodiment 1 of the present application. According to Figure 1 , the electrocardiographic information 10 as an analysis object is acquired from the determination object 1 via the biological signal measurer 2. The determination object 1 is a person, an animal, or the like, and is not particularly limited. As an example of the biological signal measurer 2, an electrocardiograph scanner is cited.

[0073] The electrocardiographic information 10 acquired from the determination object includes information of the determination object 1, an acquisition date and time, a place, and a measurement result. The measurement result is waveform data 11 obtained by plotting with a time as a horizontal axis and a voltage as a vertical axis (refer to Figure 1 ).

[0074] Figure 2 is a diagram showing a detailed example of the electrocardiographic information 10 used for analysis. Figure 2(a) is an example of waveform data 11A showing a normal pulse, Figure 2 (b) is an example of waveform data 11B showing atrial fibrillation in which F waves are generated. Each waveform data can contain components called P wave 11a, Q wave 11b, R wave 11c, S wave 11d, T wave 11e, and F wave 11f, and these shapes are used as a determination reference for arrhythmia. In addition, the shape of the waveform data possessed by the electrocardiographic information 10 also sometimes changes depending on the determination object and the biological signal measuring device.

[0075] Returning to Figure 1 The biological signal analysis system 3 according to the present application includes an input unit 31, a data conversion unit 32, a learning unit 33, a data restoration unit 34, a difference calculation unit 35, a determination unit 36, a control unit 37, and a storage unit 38.

[0076] The input unit 31 receives input of the electrocardiographic information output from the biological signal measuring device 2. In addition, the input unit 31 extracts unit waveform data from the waveform data 11 of the acquired electrocardiographic information. The input unit 31 is constituted by a connector electrically connected to the biological signal measuring device 2, or a communication device provided with a communication unit with the biological signal measuring device 2, or an input port of a medium in which data is stored, and can also include a user interface such as a keyboard, a mouse, a microphone, and the like.

[0077] The data conversion unit 32 performs conversion processing on the electrocardiographic information 10. Specifically, frequency analysis is performed on the waveform data representing the relationship between time and intensity (voltage in this case) to generate three-dimensional feature data of time-frequency-intensity.

[0078] Figure 3 is a diagram showing a processing example of the input unit 31 of the electrocardiographic information according to Embodiment 1. The input unit 31 generates unit waveform data 12 extracted in units of a specified unit from the waveform data 11 shown in (a) of Figure 3 (see (b) of Figure 3 ), and registers it to the storage unit 38. For example, it is considered that the extraction of the unit waveform data 12 is performed in units of a specified period based on time axis information, or in units of a specified number of waveforms with the peak position of the R wave 11c as the extraction reference position, but is not limited thereto. In the example shown in Figure 3 , the peak position 11cp of the R wave 11c is detected, and the specified number of seconds before and after the peak position 11cp of the R wave 11c is taken as one unit, whereby the unit waveform data 12 represents a single waveform based on the waveform data 11. Subsequently, the data conversion unit 32 applies processing, and generates three-dimensional feature data 13 from the unit waveform data 12 (see (c) of Figure 3 ).

[0079] Here, regarding the waveform data 11, the interval of each unit waveform data 12 is not fixed due to the variation of the heartbeat, but when segmented by unit waveforms, the learning accuracy of the autoencoder 41 in the learning section 33 described later is improved, and the noise mixing into the difference data and the difference in the difference calculation section 35 is suppressed, so the accuracy of the arrhythmia determination is improved. In addition, as Figure 3 The interval that does not have information meaningful for determination such as the unit external waveform data 11g in (b) is excluded, so the processing amount can be suppressed.

[0080] In addition, in the example shown in Figure 3 In order to segment by unit waveforms, the specified number of seconds before and after the peak position 11cp of the R wave 11c detected from the maximum value of a certain interval is set as one unit waveform, but it is not limited thereto. In the case where the waveform data does not have the R wave 11c and thus the peak position 11cp of the R wave 11c cannot be detected, the input section 31 finds the maximum value of the waveform data and extracts the specified number of seconds before and after the time as one unit with the time as the reference position.

[0081] The learning section 33 generates a learned model by performing learning that optimizes the network parameters of the neural network. The learning section 33 has an autoencoder, and performs learning of the autoencoder using three-dimensional feature data generated based on electrocardiographic information as learning data. As the learning, for example, learning using the error backpropagation method and the stochastic gradient descent method, etc. is exemplified. The learning section 33 stores the learned model (hereinafter, also simply referred to as a learning model) after learning in the storage section 38.

[0082] Here, the learning processing of the learning section 33 will be described. Figure 4 is a flowchart showing the learning processing of the biological signal analysis system 3. In the learning processing, the three-dimensional feature data 13 is acquired as learning data from the unit waveform data generated by the data conversion section 32 applying each processing based on the electrocardiographic information 10 input to the input section 31 (step S101). The learning section 33 performs learning of the autoencoder using the generated learning data, and outputs a learned learning model. At this time, a plurality of learning data is prepared, and it is known whether it is normal or arrhythmia, and data in which the majority is derived from normal waveforms is used. The "majority" here means 95% or more. In addition, "normal" adopted as the learning data means "normal" determined by a doctor, a test technician, etc. In the present embodiment 1, the learning section 33 performs learning using only learning data of waveform data based on normal pulse (normal waveform data). In addition, it is preferable that the learning section 33 use only normal waveform data in terms of improving the accuracy of determination, but it is not limited thereto.

[0083] Returning to Figure 1The data restoration unit 34 performs restoration processing of the three-dimensional feature data obtained from the determination target 1 generated by the data transformation unit 32 using the learning model generated by the learning unit 33.

[0084] The difference calculation unit 35 calculates the difference between the data before and after the restoration processing performed by the data restoration unit 34. Specifically, the difference calculation unit 35 generates difference data between the three-dimensional feature data before the restoration and the three-dimensional feature data after the restoration, and calculates the difference based on the difference data.

[0085] The determination unit 36 determines which one of the normal waveform data and the arrhythmic waveform data the pulse of the determination target 1 is based on the difference calculated by the difference calculation unit 35.

[0086] The control unit 37 comprehensively controls the operation of the biological signal analysis system 3. In addition, the control unit 37 causes the determination result of the determination unit 36 to be displayed on a display or output to the outside.

[0087] The storage unit 38 stores various programs for causing the biological signal analysis system 3 to operate, various data including data generated by each unit, and the like. The determination program executed using the learning model is also included in the various programs. The storage unit 38 is constituted using a ROM (Read Only Memory) on which various programs and the like are installed in advance, a RAM (Random Access Memory) that stores operation parameters, data, and the like of each process, an HDD (Hard Disk Drive), an SSD (Solid State Drive), and the like.

[0088] The various programs can also be recorded on a recording medium that can be read by a computer, such as an HDD, a flash memory, a CD-ROM, a DVD-ROM, a Blu-ray (registered trademark), and the like, and widely distributed. In addition, the input unit 31 can also acquire the various programs via a communication network. The communication network described here is constituted using an existing public line network, a LAN (Local Area Network), a WAN (Wide Area Network), and the like, and can be wired or wireless.

[0089] The biological signal analysis system 3 having the functional configuration described above is a computer configured with one or more hard disks such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an ASIC (Application Specific Integrated Circuit), and an FPGA (Field Programmable Gate Array).

[0090] Next, the normal / arrhythmia determination processing based on the electrocardiographic information obtained from the determination target 1 will be described. Figure 5 is a flowchart showing the determination processing according to Embodiment 1. In the determination processing, first, the input section 31 acquires the electrocardiographic information 10 obtained from the determination target 1 via the biological signal measurer 2 (step S201). The input section 31 generates unit waveform data 12 obtained by partially extracting waveform data 11 of the acquired electrocardiographic information 10.

[0091] The data conversion section 32 converts the unit waveform data to generate three-dimensional feature data (step S202). The three-dimensional feature data is determination data or learning data.

[0092] Figure 6A and Figure 6B is a diagram showing a processing example of the data conversion section 32 according to Embodiment 1. Figure 6A shows an example of various data in a normal waveform, Figure 6B shows an example of various data in an arrhythmic waveform. In addition, (a) of each diagram shows unit waveform data, and (b) shows learning data or determination data. The data conversion section 32 applies a time-frequency conversion to the unit waveform data 12 obtained by the input section 31. At this time, known learning data 14X is obtained as three-dimensional feature data for learning processing, and determination data 14A, 14B is obtained as three-dimensional feature data for determination processing. Here, for convenience, a distinction is made between the determination data 14A (normal waveform: Figure 6A ) and the determination data 14B (arrhythmic waveform: Figure 6B ), but it is unknown whether it is a normal waveform or an arrhythmic waveform. Furthermore, in the time-frequency conversion, a general frequency conversion method such as a short-time Fourier transform or a wavelet transform can be used, but it is not limited thereto. In addition, the known learning data 14X and the determination data 14A, 14B are preferably processed as numerical data, but can be processed as other forms such as image data.

[0093] Here, Figure 6AThe four unit waveform data (normal unit waveform data) 12A to 12D in FIG. 10 are acquired from different subjects, and the shapes of the waveforms differ depending on the subjects. As examples of the individual differences, the amount of descent of the S wave, the height of the T wave, and the like are given. Therefore, when the time-frequency conversion of the data conversion section 32 is applied to these unit waveform data 12A to 12D, three-dimensional feature data 13A to 13D are obtained. These three-dimensional feature data 13A to 13D are used as known learning data 14X and determination data 14A derived from normal waveforms, and the individual differences are reduced compared to the unit waveform data 12A to 12D. Thus, in the learning section 33 described later, the accuracy of learning by using known learning data 14X (three-dimensional feature data 13A to 13D) obtained from different subjects can be improved.

[0094] In addition, Figure 6B The four unit waveform data (arrhythmia unit waveform data) 12E to 12H in FIG. 11 are examples of main arrhythmias or precursors thereof, respectively. The purpose of the biological signal analysis system 3 is to distinguish the unit waveform data 12A to 12D from the unit waveform data 12E to 12H. However, it is difficult to capture the characteristics of the data in the state of the unit waveform data, and it is difficult to distinguish them. Therefore, when the time-frequency conversion of the data conversion section 32 is applied as well, three-dimensional feature data 13E to 13H are obtained. These three-dimensional feature data 13E to 13H are used as determination data 14B derived from arrhythmia waveforms, and exhibit characteristics unique to arrhythmias in the regions R1 to R5 shown in (b) of FIG. 11, and are easily distinguished from normal waveforms. Figure 6B

[0095] After the data conversion, the data restoration section 34 acquires the learning model generated by the learning section 33 (step S203). At this time, the data restoration section 34 acquires the learning model from the storage section 38 or the learning section 33.

[0096] Here, the learning data generated by the learning section 33 is described with reference to Figure 7 and Figure 8 Figure 7 is a diagram showing the structure of the autoencoder according to Embodiment 1. As shown in Figure 7 ​​As shown, the autoencoder 41 is a neural network model composed of an encoder 411 and a decoder 412. In the encoder 411, the features possessed by each node 45 are input to the nodes 45 of the next layer while being weighted with the weights inherent to each edge 46 and being reduced in dimension. This enables compression of the features of the input data 43 into low dimensions. In the decoder 412, processing inverse to that of the encoder 411 is performed to generate the restored data 44. That is, the restored data 44 is data obtained by reconstructing the input data 43 from the features compressed by the encoder 411. Hereinafter, the compression of features by the encoder 411 will be referred to as feature extraction, and the reconstruction of features by the decoder 412 will be referred to as restoration. Note that, regarding the input and output of the autoencoder 41, the input data 43 refers to the learning data 14X in the learning processing, and refers to the determination data in the determination processing. Further, in this embodiment 1, the number of layers of the autoencoder 41 and the number of nodes 45 possessed by each layer are not limited.

[0097] Figure 8 Fig. 2 is a diagram showing an example of the processing of the learning unit 33 according to the embodiment 1. The learning unit 33 performs learning of the autoencoder 41 using the learning data. The autoencoder performs feature extraction of the input data 43 input as the learning data and restoration based on the extracted features to generate the restored data. Figure 8 (a) of Fig. 2 (a) shows an initial stage of learning, Figure 8 (b) of Fig. 2 (b) shows a stage where learning has progressed to some extent. As Figure 8 As shown in (a) of Fig. 2, when learning is insufficient, the accuracy of feature extraction is low, and the restored data (during learning) 44A does not restore the input data 43. Through learning of the autoencoder 41a (see (a) of Fig. 2), the amount of weighting of each edge 46 is adjusted in such a manner that the difference between the input data 43 and the restored data (during learning) 44A becomes small. As a result of this learning, (a) of Fig. 2 changes to the state of (b) of Fig. 2. Figure 8 As shown in (a) of Fig. 2, when learning is insufficient, the accuracy of feature extraction is low, and the restored data (during learning) 44A does not restore the input data 43. Through learning of the autoencoder 41a (see (a) of Fig. 2), the amount of weighting of each edge 46 is adjusted in such a manner that the difference between the input data 43 and the restored data (during learning) 44A becomes small. As a result of this learning, (a) of Fig. 2 changes to the state of (b) of Fig. 2. Figure 8 (b) of Fig. 2. Specifically, the autoencoder 41b that has undergone learning performs feature extraction with respect to the input data 43, and thus can generate restored data 44B that is restored with high accuracy. Figure 8 (b) of Fig. 2. Specifically, the autoencoder 41b that has undergone learning performs feature extraction with respect to the input data 43, and thus can generate restored data 44B that is restored with high accuracy.

[0098] In the embodiment 1, the learning model is learned using only the learning data 14X known to be normal, and thus can perform feature extraction and restoration of normal waveforms, but the known learning data 14X required for learning of the learning model is obtained, for example, from the unit waveform data 12A to 12D that can be easily and abundantly obtained. Thus, it is not necessary to learn the arrhythmia that is the object of detection as in the past, and the learning data is easily obtained, and thus it is easy to improve the accuracy of learning.

[0099] When the learning model is acquired, the data restoration unit 34 performs restoration processing using the determination data generated in step S202 and the learning model generated by the learning unit 33 (step S204). By the restoration processing, the three-dimensional feature data in which the determination data generated in step S202 is restored can be obtained.

[0100] Figure 9A and Figure 9B is a diagram illustrating an example of the processing of the data restoration unit 34 according to Embodiment 1. Further, for convenience, the determination data 13I (normal waveform: Figure 9A ) and the determination data 13J (arrhythmia waveform: Figure 9B ) are distinguished, but it is unknown whether the actual waveform is a normal waveform or an arrhythmia waveform. The data restoration unit 34 restores the determination data 13I, 13J generated by the data conversion unit 32 using the learning model 42 learned by the learning unit 33, and generates restored data 15A, 15B.

[0101] Here, since the learning model 42 performs feature extraction and restoration of the normal waveform, in a case where the determination data 13I originating from the normal waveform is input, the learning model 42 can correctly perform feature extraction and restoration. On the other hand, in a case where the determination data 13J originating from the arrhythmia is input, the learning model 42 cannot extract the feature of the arrhythmia, and thus the restoration accuracy is low.

[0102] After the restoration processing, the difference calculation unit 35 generates difference data between the determination data (three-dimensional feature data) generated in step S202 and the three-dimensional feature data restored in step S204 (step S205). Thereafter, the difference calculation unit 35 calculates the difference amount from the difference data (step S206).

[0103] Figure 10A and Figure 10B is a diagram illustrating an example of the processing of the difference calculation unit 35 according to Embodiment 1. The difference calculation unit 35 compares the determination data and the restored data which are input and output of the data restoration unit 34, generates difference data in which the difference is obtained, and calculates the difference amount on the basis of the difference data. Further, in the calculation of the difference amount, a general difference calculation method such as RMSE (Root Mean Square Error), MAE (Mean Absolute Error), or the like can be used, but is not limited thereto. In a case where the determination data 13I is input to the data restoration unit 34, the learning model 42 can correctly restore the feature of the normal waveform, and thus the difference data 16A is data of a flat intensity distribution (black in this example), and the value of the difference amount 17A becomes small (refer to Figure 10A). On the other hand, in a case where the determination data 13J is input to the data restoration section 34, the learning model 42 cannot correctly restore the features of the arrhythmia, and the difference data 16B is data of a non-uniform intensity distribution, and the value of the difference 17B becomes large (refer to Figure 1 OB).

[0104] The determination section 36 determines whether the waveform data obtained from the determination target 1 is normal or arrhythmic based on the difference calculated in step S206 (step S207). Thereafter, the determination section 36 outputs the determination result (step S208).

[0105] Specifically, the determination section 36 determines whether it is normal or arrhythmic by whether the difference (for example, the differences 17A, 17B) calculated by the difference calculation section 35 is smaller than a threshold value set in advance, and outputs the determination result. Figure 11A and Figure 11B is a diagram showing a processing example of the determination section according to Embodiment 1. Here, it is assumed that the difference 17A is smaller than the threshold value, and the difference 17B is larger than the threshold value. As shown in Figure 11A , the difference 17A is determined to be "normal" because the difference is small. On the other hand, as shown in Figure 11B , the difference 17B is determined to be "arrhythmic" because the difference is large.

[0106] Figure 12 is a diagram showing an analysis example based on the processing according to Embodiment 1. Specifically, in the determination processing, 10 samples of the determination data 14A and the determination data 14B (refer to Figure 6A , Figure 6B ) are each provided as determination data, and the differences calculated by the RMSE described above are shown. At this time, as described above, the difference from the arrhythmic waveform is larger than the difference from the normal waveform, and by setting the threshold value Th based on the difference, it is possible to perform determination of arrhythmia.

[0107] In principle, the determination processing is performed after the learning processing. However, in a case where the learning model has been output by the learning processing, by using the existing learning model, it is also possible to omit the learning processing and determine normal / arrhythmic.

[0108] According to Embodiment 1 described above, the determination section 36 is able to perform determination of arrhythmia, which is waveform data different from normal waveform data. In addition, it is possible to perform determination even in a case where if the unit of performing determination is shortened, data cannot be processed as a sufficient continuous amount due to the influence of noise or the like, and it is possible to perform determination even of arrhythmia that occurs only once.

[0109] [Embodiment 2]

[0110] Next, Embodiment 2 of the present invention will be described. This Embodiment 2 is subordinate to Embodiment 1. The structure of the biosignal analysis system according to Embodiment 2 is the same as that of the biosignal analysis system 3 according to Embodiment 1, so the description is omitted. Hereinafter, the parts that are different from Embodiment 1 will be described.

[0111] In this embodiment 2, the learning unit 33 classifies the waveform shape features of the unit waveform data that forms the basis of the known learning data 14X used for learning into learning data groups, and learns from the learning data groups according to the classification type (feature) to generate multiple learning models. Examples of waveform shape features mentioned here include the descent of the S-wave and the height of the T-wave. In embodiment 2, the learning unit 33 has multiple self-coders set according to classification type. Learning models different according to classification type are generated by each coder. Furthermore, the number of learning models (self-coders) prepared by the learning unit 33 in this embodiment 2 can be freely set, but is preferably four or more.

[0112] The decision-making process in this embodiment 2 only involves the learning model acquisition process. Figure 5 The step S203 shown is different, except that it follows the procedure. Figure 5 The flowchart shown illustrates the learning model acquisition process. In this process, based on the waveform shape characteristics of the unit waveform data of the judgment object 1, one model is selected from multiple learning data sets or learning models (self-encoded decoders) possessed by the learning unit 33. The selected learning model depends on the objective, but is preferably the best model in the judgment process. Furthermore, as a method for selecting the best learning model, one could consider, for example, applying each process to the multiple learning models and the waveform data obtained from the judgment object 1 through the data restoration unit 34 and the difference calculation unit 35, and selecting the learning model that minimizes the difference in the output; however, this method is not limited to this.

[0113] In Embodiment 1, by using time-frequency transformation of unit waveform data by the data transformation unit 32 as a feature for learning processing, individual differences can be reduced. However, if the learning model in Embodiment 2 is used, individual differences can be further reduced, and the judgment accuracy can be improved.

[0114] In addition, the normal waveform obtained from the determination target in which the symptoms of cardiac hypertrophy and cardiac enlargement have occurred is different in the characteristics of the waveform shape from the waveform obtained from a healthy person in which the symptoms of cardiac hypertrophy and cardiac enlargement have not occurred. Therefore, the learning model that has learned from the electrocardiographic information 10 obtained from the healthy person cannot correctly reproduce the characteristics of the normal waveform based on cardiac hypertrophy and cardiac enlargement, and the intensity distribution of the difference data is uneven, and the value of the difference becomes large, so there is a possibility that it will be erroneously determined as arrhythmia. At this time, the learning unit 33 in Embodiment 2 can accurately determine arrhythmia with high precision by preparing a learning model of cardiac hypertrophy and a model of cardiac enlargement with respect to the determination target in which the symptoms of cardiac hypertrophy and cardiac enlargement have occurred.

[0115] [Embodiment 3]

[0116] Next, Embodiment 3 of the present application will be described. The biological signal analysis system according to Embodiment 3 has the same structure as the biological signal analysis system 3 according to Embodiment 1, and therefore the description will be omitted. Hereinafter, the portions different from Embodiment 1 will be described.

[0117] In Embodiment 3, the learning unit 33 simultaneously uses known learning data including determination targets other than the determination target itself (hereinafter, referred to as learning data 14X) and learning data based on the normal waveform obtained from the determination target itself (hereinafter, referred to as learning data 14Y) to perform learning. In the learning using the learning data 14X and the learning data 14Y, for example, a method of performing learning using the learning data 14X and the learning data 14Y as one data group is considered, but is not limited thereto. By using the learning model thus obtained as the learning model of the data reproduction unit 34 in the determination processing, it is possible to make the reproduction accuracy higher and improve the determination accuracy.

[0118] In Embodiments 1 and 2, by performing time-frequency conversion of the unit waveform data by the data conversion unit 32 as the characteristics of the learning processing, it is possible to reduce the individual differences. However, if the learning processing of the learning unit 33 according to Embodiment 3 is used, it is possible to further exclude the individual differences by the learning data obtained from the determination target itself, and it is possible to improve the determination accuracy. The learning model thus learned can also be used as the learning model in Embodiment 1. The learning unit 33 in Embodiment 3 can perform learning because only the normal waveform data that can be easily and abundantly obtained is required for the learning of the learning model, and the degree of difficulty of the learning is very low.

[0119] [Embodiment 4]

[0120] Next, Embodiment 4 of the present application will be described. Embodiment 4 is based on Embodiment 3. The biological signal analysis system according to Embodiment 4 has the same configuration as the biological signal analysis system 3 according to Embodiment 1, and thus the description thereof will be omitted. Hereinafter, the different parts from Embodiment 3 will be described.

[0121] In Embodiment 4, the learning unit 33 has a plurality of learning models that are classified into learning data groups based on features of waveform shapes of unit waveform data that become a basis in known learning data 14X used for learning, and are learned based on the plurality of learning data groups after the classification. As the features of the waveform shapes, the amount of descent of S wave, the height of T wave, and the like are exemplified. In Embodiment 4, the learning unit 33 has a plurality of autoencoders provided for each classification type. The learning models that are different for each classification type are generated by the respective autoencoders. In addition, the number of learning data groups or learning models (autoencoders) prepared by the learning unit 33 in the present embodiment can be freely set, but is preferably four or more.

[0122] The learning process or the determination process in Embodiment 4 basically follows the flowchart shown in Figure 4 , Figure 5 However, in the learning model generation process (step S102 shown in Figure 4 ) and the learning model acquisition process (step S203 shown in Figure 5 ), one, preferably the best model, is selected from the plurality of learning data groups or the plurality of learning models possessed by the learning unit 33 based on the features of the waveform shapes of the unit waveform data of the determination target 1, and is used in the learning or the determination. Further, as a method of selecting the best learning data group or the learning model, for example, a method of applying each process of the data restoration unit 34 and the difference calculation unit 35 to the plurality of learning models learned based on the plurality of learning data groups and the normal waveform data obtained from the determination target 1, and selecting the model that minimizes the amount of output difference is considered, but is not limited thereto.

[0123] In Embodiment 3, the learning data 14Y obtained from the determination target himself or herself as a feature of the learning process can further exclude individual differences. However, if the learning process of the learning unit 33 according to Embodiment 4 is used, individual differences can be further reduced, and the determination accuracy can be improved.

[0124] Furthermore, the waveforms obtained from subjects exhibiting symptoms of cardiomegaly and enlargement differ in shape from those obtained from healthy individuals. Therefore, the learning model, trained using ECG information 10 obtained from healthy individuals, cannot accurately reproduce the characteristics of normal waveforms based on cardiomegaly and enlargement, resulting in uneven intensity distribution of the difference data and larger difference values, potentially leading to misdiagnosis as arrhythmia. In this case, the learning unit 33 in this embodiment, by preparing a learning dataset or learning model for cardiomegaly and enlargement, can perform high-precision arrhythmia determination even for subjects exhibiting symptoms of cardiomegaly and enlargement. The learning model 42 trained in this way can also be used as the learning model in embodiments 1 to 3.

[0125] [Implementation Method 5]

[0126] Next, Embodiment 5 of the present invention will be described. The structure of the biosignal analysis system according to Embodiment 5 is the same as that of the biosignal analysis system 3 according to Embodiment 1, so the description is omitted. Hereinafter, the parts that are different from Embodiment 1 will be described.

[0127] The learning process in this embodiment 5 basically follows... Figure 4 The flowchart shown, but in the learning model generation process ( Figure 4 In step S102 shown, the learning unit 33 performs learning using only the learning data 14Y obtained from the judgment subject. By using the learning model obtained in this way as the learning model of the data restoration unit 34 in the judgment process, the restoration accuracy can be improved, and the judgment accuracy can be enhanced.

[0128] In Embodiment 3, as a feature of the learning process, individual differences can be further eliminated by including known learning data 14X of the judgment object (excluding the judgment object itself) and learning data 14Y obtained from the judgment object itself. However, if the learning process of the learning unit 33 according to Embodiment 5 is used, individual differences can be completely eliminated using only the learning data 14Y obtained from the judgment object itself, thereby improving the judgment accuracy. In addition, Embodiments 2 and 4 are effective even when there is no corresponding learning data set or learning model. The learning model learned in this way can also be used as the learning model in Embodiments 1 to 4. The learning unit 33 in Embodiment 5 can use the learning data 14Y obtained from the judgment object itself for learning because the learning model only requires normal waveform data of the judgment object itself, which can be easily and abundantly obtained, making the learning difficulty very low.

[0129] [Implementation Method 6]

[0130] Next, Embodiment 6 of the present invention will be described. The structure of the biosignal analysis system according to Embodiment 6 is the same as that of the biosignal analysis system 3 according to Embodiment 1, so the description is omitted. Hereinafter, the parts that are different from Embodiment 1 will be described.

[0131] exist Figure 6B In (b), the three-dimensional feature data 13E derived from atrial fibrillation is set to have a pulse irregularity-specific feature in regions R1 and R2. However, this feature is weak and may disappear through the restoration process of the data restoration unit 34, or become smaller as a difference in the overall data during the pulse irregularity determination in the determination unit 36, making it difficult to separate from the normal waveform data.

[0132] Therefore, the data transformation unit 32 in this embodiment 6 performs weighting on the time-frequency feature region in the determination data that exhibits arrhythmia. Figure 13A and Figure 13B This is a graph showing a weighted example of the features specific to irregular pulses in the three-dimensional feature data involved in Embodiment 6. Figure 13A and Figure 13B In this example, atrial fibrillation characteristics were weighted based on region R6, which has separate ranges for time and frequency, but this is not the only method used. For instance, in the weighted data 13K′ obtained by weighting region R6 using data 13K based on the determination that it originates from a normal waveform, the effect of the weighting is very small, and the resulting difference data 16K is... Figure 10A The example is also a flat intensity distribution (see...) Figure 13A On the other hand, in the weighted data 13L′ obtained by weighting the data 13L in region R6 based on the determination that it is from atrial fibrillation, the resulting difference data 16L is a non-uniform intensity distribution due to the large features in region R6 (see reference). Figure 13B ).

[0133] Thus, according to the data transformation unit 32 in this embodiment 6, the F wave 11f (for example, referring to) is used to transform the data in atrial fibrillation and atrial flutter. Figure 2 By pre-weighting the characteristic features of weak arrhythmias such as (b)), the loss of features due to restoration can be prevented. As a result, it becomes easier to separate normal waveform data from arrhythmic waveform data.

[0134] [Implementation Method 7]

[0135] Next, Embodiment 7 of the present invention will be described. The structure of the biosignal analysis system according to Embodiment 7 is the same as that of the biosignal analysis system 3 according to Embodiment 1, so the description is omitted. Hereinafter, the parts that are different from Embodiment 1 will be described.

[0136] Figure 14 is a graph showing a specific example of the time-frequency coordinate-based main cause of arrhythmia according to Embodiment 7. Figure 14 (Sa) shows an example of electrocardio information related to atrial fibrillation, (Sb) shows an example of electrocardio information related to ventricular premature contraction, and (Sc) is an example of electrocardio information related to cardiac hypertrophy. Further, the respective corresponding unit waveform data is Figure 14 (a) of FIG. 18, the determination data obtained by frequency transforming the unit waveform data is Figure 14 (b) of FIG. 19, and the difference data based on the determination data and the restored determination data is Figure 14 (c) of FIG. 20. In Embodiments 1 to 3, the determination unit 36 performs normal / arrhythmia determination based on the normal waveform all the time, and can perform arrhythmia determination regardless of the kind, but cannot determine the main cause of arrhythmia. The determination of the main cause of arrhythmia here means determination of the arrhythmia or the disease name thereof, or determination of the main cause of arrhythmia narrowed down to a roughly framed "R-wave origin" or the like.

[0137] Therefore, the determination unit 36 in Embodiment 7 does not simply determine arrhythmia, but determines the main cause of arrhythmia with reference to the coordinates at which the features of the main cause of arrhythmia appear. The known learning data 14X and the determination data 19A to 19C can distinguish the waveform components (P-wave 11a, Q-wave 11b, R-wave 11c, S-wave 11d, T-wave 11e, F-wave 11f) according to the time coordinates. Figure 2 Further, it is possible to read what kind of waveform is depicted from the change state of the frequency band. In Figure 14 In Embodiment 7, as an example, an example in which the time-frequency coordinates are divided into a region including the P-wave or F-wave (P or F-wave), a region including the Q-wave, R-wave, and S-wave (QRS-wave), and a region including the T-wave or F-wave (T or F-wave) is shown. For example, in the time coordinates in which the R-wave 11c appears as in the ventricular premature contraction (Sb), in the case where the frequency band generated by the intensity moves to the low frequency direction, the amplitude of the R-wave 11c is enlarged. Therefore, based on the difference data obtained from the determination data and the restored data, the difference is calculated with respect to each of the specified coordinate ranges (for example, regions R 11 , R 12 , R 13 , R 14 ) and, according to the size thereof or their correlation, the main cause of arrhythmia can be determined. At this time, the specification of the coordinate range can be performed automatically by image processing such as edge detection or the like, or can be input by the user via the input unit 31.

[0138] In addition, as another effect obtained by the determination section 36 in Embodiment 7, the difference is calculated for each of the specified coordinate ranges, and thus it is easy to determine even a weak feature such as an F wave that occurs in atrial fibrillation.

[0139] (Other Embodiments)

[0140] The above describes a mode for implementing the present application, but the present application should not be limited to the above-described embodiments. For example, the biological signal analysis system 3 is described as being provided with the learning section 33, but the biological signal analysis system 3 can be configured to acquire a learning model via a communication network without having the learning section 33. In addition, the information to be analyzed is not limited to electrocardiographic information, and can be applied as long as the above-described waveform data is used.

[0141] Industrial Applicability

[0142] With the series of biological signal analysis systems according to the present application, it is possible to determine arrhythmia based on electrocardiographic information.

[0143] Explanation of Reference Numerals

[0144] 1: determination target; 2: biological signal measuring device; 3: biological signal analysis system; 10: electrocardiographic information; 11: waveform data; 12, 12A to 12H: unit waveform data; 13, 13A to 13H: three-dimensional feature data; 13I to 13L, 14A, 14B, 19A to 19C: determination data; 13K', 13L': weighted data; 14X: learning data; 15A, 15B: restored data; 16A, 16B, 16K, 16L, 20A to 20C: difference data; 17A, 17B: difference; 31: input section; 32: data conversion section; 33: learning section; 34: data restoration section; 35: difference calculation section; 36: determination section; 37: control section; 38: storage section; 41a, 41b: self-encoder / decoder; 42: learning model; 43: input data.

Claims

1. A biological signal analysis system that is a series of biological signal analysis systems that make a determination of arrhythmia based on electrocardiographic information, comprising: an input section that accepts input of the electrocardiographic information; a data transformation section that generates determination data by performing time-frequency transformation on the electrocardiographic information; a data restoration section that generates restoration data obtained by restoring the determination data using a learned model composed of a self-encoder-decoder; a difference calculation section that calculates a difference between the restoration data and the determination data; and a determination section that determines which of normal waveform data and arrhythmia waveform data the determination data is based on the difference, wherein the learned model is a model generated by learning that uses 95% or more of known learning data generated based on normal waveform data.

2. The biological signal analysis system according to claim 1, wherein the learned model is a model generated by learning that uses learning data acquired from a subject of determination.

3. The biological signal analysis system according to any one of claims 1 to 2, wherein the input section segments waveform data possessed by the electrocardiographic information into unit waveform data obtained by extracting waveform data of a prescribed time range, and the data transformation section generates the determination data by performing time-frequency transformation on the unit waveform data.

4. The biological signal analysis system according to any one of claims 1 to 2, wherein the data transformation section generates weighted data obtained by weighting a time-frequency characteristic region of arrhythmia in the determination data.

5. The biological signal analysis system according to any one of claims 1 to 2, wherein the difference calculation section generates difference data that indicates a difference between the restoration data and the determination data, and calculates the difference based on the difference data, and the determination section determines a main cause of arrhythmia based on a time and frequency characteristic of the difference data.

6. The biological signal analysis system according to any one of claims 1 to 2, wherein the learned model is a model generated by learning that uses learning data that is known to be the normal waveform data.

7. The biological signal analysis system according to any one of claims 1 to 2, further comprising a learning section that generates the learned model using learning of the self-encoder-decoder using learning data.

8. The biological signal analysis system according to claim 1, further comprising a learning section that generates the learned model using learning of the self-encoder-decoder using the learning data, the learning section having a plurality of self-encoder-decoders corresponding to characteristics of the known learning data, and selecting one self-encoder-decoder from the plurality of self-encoder-decoders based on a characteristic of normal waveform data of a subject of determination.

9. The biological signal analysis system according to claim 2, further comprising a learning section that generates the learned model using learning of the self-encoder-decoder using the learning data. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ The learning unit has a plurality of learning data groups or a plurality of autoencoders corresponding to features of the known learning data, One learning data group or autoencoder is selected from the plurality of learning data groups or the plurality of autoencoders based on features of normal waveform data of the determination target.

Images