Prediction device, prediction method, and program

Abstract

To provide a prediction device, a prediction method, and a program which predict the occurrence of arrhythmia.SOLUTION: The prediction device includes: a classification model training unit which uses first electrocardiogram data to which a first label representing normal sinus rhythm or arrhythmia is given, to train a classification model which classifies unit electrocardiogram data representing electrocardiogram data in a preliminarily determined prescribed time interval ΔT into normal sinus rhythm and arrhythmia; a data generation unit which uses second electrocardiogram data in a time interval equal to or longer than the time interval ΔT and the trained classification model to generate third electrocardiogram data which is represented by a series of predetermined L pieces of unit electrocardiogram data and to which a second label representing normal sinus rhythm or arrhythmia after a preliminarily determined prescribed time Δt is given; and a prediction model training unit which uses the third electrocardiogram data to train a prediction model which predicts normal sinus rhythm or arrhythmia after the time Δt in response to receiving the series of L pieces of unit electrocardiogram data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] This disclosure relates to a prediction device, a prediction method, and a program. [Background technology]

[0002] For example, patients who undergo cardiac and major vascular surgery and receive acute postoperative treatment in the intensive care unit are susceptible to arrhythmias due to stress on the heart and cardiac conduction system during surgery, as well as various environmental factors during the perioperative period. In particular, atrial fibrillation, a type of arrhythmia, is prone to occurring postoperatively and, in addition to destabilizing hemodynamics, can cause, for example, intracardiac thrombi to form, leading to cardiogenic stroke. Thus, the occurrence of atrial fibrillation after surgery can trigger other complications, resulting in longer stays in the intensive care unit and hospitalization, increased medical costs, and adverse long-term prognosis. For this reason, techniques for detecting the occurrence of atrial fibrillation have been known for some time (for example, Patent Document 1). [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] U.S. Patent No. 7,020,514 [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] However, conventional technology has not been able to predict the onset of atrial fibrillation. For example, if it were possible to predict the onset of atrial fibrillation in the future, such as a few minutes later, it would be possible to address it early, which would be extremely useful in acute care.

[0005] This disclosure is made in view of the above points and provides a technology that can predict the occurrence of arrhythmias. [Means for solving the problem]

[0006] A prediction device according to one aspect of the present disclosure includes a classification model learning unit configured to learn a classification model that classifies unit electrocardiogram data representing electrocardiogram data in a predetermined time interval ΔT into at least a class including normal and arrhythmia, using first electrocardiogram data with a first label that can represent at least normal and arrhythmia; a data creation unit configured to create third electrocardiogram data represented by a series of L (where L is one or more predetermined integers) unit electrocardiogram data, and having a second label representing normal or arrhythmia at a predetermined time Δt after, using the second electrocardiogram data in a time interval of ΔT or more and the learned classification model; and a prediction model learning unit configured to learn a prediction model that predicts normal or arrhythmia at the time Δt after when a series of L unit electrocardiogram data is given, using the third electrocardiogram data.

Advantages of the Invention

[0007] A technique capable of predicting the occurrence of arrhythmia is provided.

Brief Description of the Drawings

[0008] [Figure 1] It is a diagram showing an outline of a method for predicting the occurrence of atrial fibrillation. [Figure 2] It is a diagram showing an example of the hardware configuration of the prediction device according to the present embodiment. [Figure 3] It is a diagram showing an example of the functional configuration of the prediction device according to the present embodiment. [Figure 4] It is a flowchart showing an example of the processing in the model learning phase and the processing in the prediction phase. [Figure 5] It is a diagram schematically showing an example of the configuration of the classification model. [Figure 6] It is a diagram schematically showing an example of labeling for electrocardiogram data without a label representing a long-term electrocardiogram waveform. [Figure 7] It is a diagram schematically showing an example of the configuration of the prediction model. [Modes for carrying out the invention]

[0009] The following describes one embodiment of the present invention.

[0010] <Outline of Atrial Fibrillation Prediction> The embodiments described below propose a method for predicting the occurrence of atrial fibrillation from electrocardiogram data using machine learning techniques. Generally, in order to construct (train) a machine learning model that can accurately predict the occurrence of atrial fibrillation (hereinafter referred to as the prediction model), a large amount of training data is required, consisting of electrocardiogram data and labels (training data) indicating whether the electrocardiogram data represents normal (SR: sinus rhythm) or atrial fibrillation (AF: atrial fibrillation). Moreover, since electrocardiogram data is time-series data of potential difference and its waveform is meaningful, it is thought that a certain length of at least one to several minutes is necessary for the prediction model to learn the characteristics of when atrial fibrillation occurs.

[0011] However, it is difficult to prepare a large amount of labeled electrocardiogram (ECG) data of a certain length. On the other hand, for ECG data of relatively short lengths, such as tens of seconds to over a hundred seconds, labeled ECG data such as CPSC2018 (China Physiological Signal Challenge 2018) is publicly available and easily obtainable. Also, for ECG data of relatively long lengths, such as 2 to 3 days, even if not labeled, ECG data such as MIMIC-III is publicly available and easily obtainable.

[0012] Therefore, training data for training the prediction model is created according to (1) to (3) below. Here, ΔT is the length (time interval) of the electrocardiogram data to be used as input for the classification model. Hereinafter, we assume ΔT = 1 [minute], but this is just an example and is not limited to this; it is possible to set any value from several tens of seconds to several minutes. Also, L is the length of the electrocardiogram data to be used as input for the prediction model (sequence length when the electrocardiogram data is considered as time series data of potential difference at each ΔT interval). The sequence length L is a predetermined parameter, for example, L ∈ {1, ..., 5}.

[0013] (1) Using a dataset of labeled electrocardiogram data such as CPSC2018 (hereinafter referred to as the first dataset), a machine learning model (hereinafter referred to as the classification model) is constructed (trained) to infer the labels of electrocardiogram data over a time interval ΔT.

[0014] (2) Using an electrocardiogram data dataset such as MIMIC-III (hereinafter referred to as the second dataset), the electrocardiogram data is divided into time intervals ΔT, and a pre-trained classification model is used to label the electrocardiogram data for each time interval ΔT. Hereafter, the electrocardiogram data for each time interval ΔT will be referred to as "unit electrocardiogram data".

[0015] Furthermore, in the following, as an example, we will assume that the electrocardiogram data is sampled at 50 Hz, and that the unit electrocardiogram data contains 3000 data points (potential difference values). However, the sampling of electrocardiogram data at 50 Hz is just an example and is not limited to this. Also, when a certain electrocardiogram data is divided into unit electrocardiogram data for each time interval ΔT, the data points contained in the i-th unit electrocardiogram data are represented as a column vector x i =[x i,1 ,···,x i,3000 ] τ Let's label it y i Let's assume that τ represents transpose.

[0016] (3) For each time interval ΔT, extract a unit electrocardiogram data series of length L and its label from the electrocardiogram data labeled by time, and use it as learning data for the prediction model. At this time, when it is desired to predict the occurrence of atrial fibrillation after Δt = mΔT [minutes], where m is a non-negative integer determined in advance, as the label for the unit electrocardiogram data series {x i+1 , ···, x i+L}, extract the label y i+L+m+1 of x i+L+m+1 . More specifically, as the unit electrocardiogram data series {x i+1 , ···, x i+L}, extract those for which all labels y i+1 , ···, y i+L+m are normal (SR), and the label y i+L+m+1 is either normal (SR) or atrial fibrillation (AF). Note that predicting the occurrence of atrial fibrillation after Δt means predicting whether atrial fibrillation occurs during the ΔT after Δt.

[0017] As a result, a learning data set for the prediction model, which is composed of the learning data for the prediction model ({x i+1 , ···, x i+L [[ID=第26]]}, y i+L+m+1 ), is obtained.

[0018] Hereinafter, the outline of the method for predicting the occurrence of atrial fibrillation from electrocardiogram data will be described with reference to FIG. 1.

[0019] Step 1) First, as described in (1) above, learn a classification model using a first data set, which is a data set composed of electrocardiogram data with labels for a relatively short time.

[0020] Step 2) Next, as described in (2) - (3) above, use a second data set, which is a data set composed of electrocardiogram data without labels for a relatively long time, to create a learning data set for the prediction model using the learned classification model.

[0021] Step 3) Next, train the predictive model using the training dataset created in Step 2 above.

[0022] Step 4) Then, using the electrocardiogram data to be predicted, the predictive model trained in Step 3 above is used to predict the occurrence of atrial fibrillation after Δt.

[0023] The following describes a prediction device 10 that predicts the occurrence of atrial fibrillation after Δt using the method shown in steps 1 to 4 above. Steps 1 to 3 above may also be called the model learning phase, and step 4 above may be called the prediction phase. The model learning phase is performed in advance of the prediction phase. On the other hand, the prediction phase is performed each time the electrocardiogram data to be predicted is provided.

[0024] <Example of hardware configuration for prediction device 10> Figure 2 shows an example of the hardware configuration of the prediction device 10 according to this embodiment. As shown in Figure 2, the prediction device 10 according to this embodiment includes an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these hardware components is connected to the others via a bus 109 so as to be able to communicate with each other.

[0025] The input device 101 is, for example, a keyboard, mouse, touch panel, or physical button. The display device 102 is, for example, a display or display panel. The prediction device 10 does not necessarily have to have at least one of the input device 101 and the display device 102.

[0026] External I / F 103 is an interface to external devices such as recording media 103a. The prediction device 10 can read from and write to the recording media 103a via the external I / F 103. Examples of recording media 103a include flexible disks, CDs (Compact Discs), DVDs (Digital Versatile Disks), SD memory cards (Secure Digital memory cards), and USB (Universal Serial Bus) memory cards.

[0027] The communication interface 104 is an interface for connecting the prediction device 10 to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily holds programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. The auxiliary storage device 107 is a storage device (storage device) such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or flash memory. The processor 108 is a computing device such as a CPU (Central Processing Unit) or GPU (Graphics Processing Unit).

[0028] The prediction device 10 according to this embodiment has the hardware configuration shown in Figure 2, which enables the various processes described later (processing in the model learning phase and processing in the prediction phase). Note that the hardware configuration shown in Figure 2 is just one example, and the hardware configuration of the prediction device 10 is not limited to this. For example, the prediction device 10 may have multiple auxiliary storage devices 107 or multiple processors 108, it may not have some of the hardware shown, or it may have various hardware other than the hardware shown.

[0029] <Example of functional configuration of prediction device 10> Figure 3 shows an example of the functional configuration of the prediction device 10 according to this embodiment. As shown in Figure 3, the prediction device 10 according to this embodiment includes a classification model learning unit 201, a learning data creation unit 202, a prediction model learning unit 203, and a prediction unit 204. Each of these functional units is realized, for example, by a process in which one or more programs installed in the prediction device 10 are executed by a processor 108 or the like. The prediction device 10 according to this embodiment also includes a first dataset storage unit 205, a classification model storage unit 206, a second dataset storage unit 207, a learning dataset storage unit for the prediction model 208, and a prediction model storage unit 209. Each of these storage units is realized, for example, by an auxiliary storage device 107. However, at least one of the above storage units may be realized by a storage device such as a database server connected to the prediction device 10 via a communication network.

[0030] The classification model learning unit 201 learns a classification model using the first dataset stored in the first dataset storage unit 205. The classification model learning unit 201 then stores the parameters of the learned classification model in the classification model storage unit 206. The classification model is, for example, a model including a neural network and has learnable parameters. Furthermore, the learning of the classification model can be performed using a known supervised learning method, such as backpropagation.

[0031] The training data creation unit 202 uses the second dataset stored in the second dataset storage unit 207 to create a training dataset for the prediction model using the trained classification model. Then, the training data creation unit 202 saves the created training dataset for the prediction model to the training dataset storage unit 208.

[0032] The prediction model learning unit 203 learns the prediction model using the prediction model learning dataset stored in the prediction model learning dataset storage unit 208. The prediction model learning unit 203 then saves the parameters of the learned prediction model to the prediction model storage unit 209. The prediction model is, for example, a model including a neural network and has learnable parameters. Furthermore, the prediction model can be learned using a known supervised learning method, such as backpropagation.

[0033] The prediction unit 204 uses the electrocardiogram data to be predicted to predict the occurrence of atrial fibrillation after Δt using a trained prediction model. This prediction result (SR / AF) is output to a predetermined output destination. Examples of such output destinations include a display device 102 such as a display, an auxiliary storage device 107, and other devices or terminals connected via a communication network.

[0034] The first dataset storage unit 205 stores the first dataset. The classification model storage unit 206 stores the parameters of the trained classification model. The second dataset storage unit 207 stores the second dataset. The training dataset storage unit 208 for the prediction model stores the training dataset for the prediction model. The prediction model storage unit 209 stores the parameters of the trained prediction model.

[0035] In the example shown in Figure 3, the prediction device 10 has all the functional units and all the storage units, but these functional units and storage units may be distributed among multiple devices. For example, there may be a learning device that performs the model learning phase and a prediction device that performs the prediction phase, and the learning device may have a classification model learning unit 201, a learning data creation unit 202, a prediction model learning unit 203, a first dataset storage unit 205, a classification model storage unit 206, a second dataset storage unit 207, and a learning dataset storage unit 208 for the prediction model, and the prediction device may have a prediction unit 204, a learning dataset storage unit 208 for the prediction model, and a prediction model storage unit 209.

[0036] <Processing in the model learning phase and the prediction phase> The following describes the processing in the model learning phase and the prediction phase, referring to Figure 4. Steps S101 to S103 are the processing in the model learning phase, and step S104 is the processing in the prediction phase.

[0037] Step S101: First, the classification model learning unit 201 learns a classification model using the first dataset stored in the first dataset storage unit 205. Then, the classification model learning unit 201 saves the parameters of the learned classification model to the classification model storage unit 206.

[0038] In the following, we will assume the use of CPSC2018 as the first dataset as an example. In CPSC2018, 12-lead electrocardiogram data from 10 to 140 seconds are assigned one of nine labels (Normal, AF, I-AVB (First-degree atrioventricular block), LBBB (Left bundle brunch block), RBBB (Right bundle brunch block), PAC (Premature atrial contraction), PVC (Premature ventricular contraction), STD (ST-segment depression), STE (ST-segment elevated)). Therefore, we will consolidate these nine labels into three and perform a three-class classification in the classification model. Specifically, the six types of labels—Normal, I-AVB, LBBB, RBBB, STD, and STE—will be classified as "SR," the single type of label, AF, will be classified as "AF," and the two types, PAC and PVC, will be classified as "Other," resulting in a three-class classification system.

[0039] Furthermore, since CPSC2018 consists of electrocardiogram data ranging from 10 to 140 seconds, the data is processed so that its length (time interval) becomes ΔT = 1 minute (= 60 seconds) in order to be input into the classification model (i.e., processed to become unit electrocardiogram data). Specifically, for electrocardiogram data less than 60 seconds in length, the data is zero-padded until the number of data points reaches 3000 and used as training data. On the other hand, for electrocardiogram data longer than 60 seconds, the portion exceeding 3000 data points is truncated and used as training data.

[0040] Furthermore, to obtain a more accurate classification model, data augmentation techniques may be used to increase the number of electrocardiogram (ECG) data used as training data. Specifically, for example, multiple 60-second ECG data may be randomly sampled from ECG data exceeding 60 seconds, or new ECG data may be generated by adding white noise, square waves, sine waves, etc., to the ECG data.

[0041] As a classification model, any machine learning model used for a 3-class classification task can be employed, but below, as an example, we will assume the classification model shown in Figure 5. The classification model shown in Figure 5 consists of a 1D convolutional layer (Conv1d), a batch normalization layer (Batch Norm), a normalized linear unit layer (ReLU), a dropout layer (Dropout), a maximum pooling layer (Max Pooling), a fully connected layer (Dense), and a softmax layer (Softmax). The input dimension of the first 1D convolutional layer is 3000 dimensions, and the output dimension of the last softmax layer is 3 dimensions. However, the configuration of the classification model shown in Figure 5 is just an example and is not limited to this. Also, 3-class classification is just an example, and multi-class classification of 4 or more classes is also possible.

[0042] At this time, the classification model learning unit 201 learns the classification model so that when the electrocardiogram data used as training data is input to the classification model, the classification result accurately estimates the label assigned to that electrocardiogram data. The classification model can be trained using a known supervised learning method, such as backpropagation.

[0043] Step S102: Next, the training data creation unit 202 uses the second dataset stored in the second dataset storage unit 207 to create a training dataset for the prediction model using the trained classification model. Then, the training data creation unit 202 saves the created training dataset for the prediction model to the training dataset storage unit 208.

[0044] In the following, we will assume, as an example, that MIMIC-III is used as the second dataset. MIMIC-III is a collection of unlabeled electrocardiogram data with lengths such as 2 to 3 days.

[0045] First, the training data creation unit 202 divides these electrocardiogram data into time intervals ΔT, as explained in (2) above, and then assigns labels to the electrocardiogram data of each time interval ΔT using the trained classification model. For example, if a certain electrocardiogram data is divided into time intervals ΔT and each time interval is represented as a vector, then, as shown in Figure 6, the unit electrocardiogram data sequence {x1,···,x n} and their label sequence {y1,···,y n The result is obtained. Note that n is the sequence length when the electrocardiogram data is divided into time intervals ΔT and represented as a sequence of unit electrocardiogram data.

[0046] Next, the training data creation unit 202 generates the label y as described in (3) above. i+1 ,···,y i+L+m All are normal (SR), and label y i+L+m+1 Unit electrocardiogram data series {x i+1 ,···,x i+L} and label yi+L+m+1 Extract and and use them as training data for the prediction model ({x i+1 ,···,x i+L},y i+L+m+1 ) where m is a predetermined non-negative integer and satisfies Δt = mΔT [minutes].

[0047] This allows us to use training data ({x i+1 ,···,x i+L},y i+L+m+1 The set of (X) is obtained as the training dataset for the prediction model. For simplicity, below, a certain training data for the prediction model is (X) i ,Y i ) is expressed as D={(X i ,Y i Let |i=1,···,|D|}. Here, X i ={x1 (i) ,···,x L (i)}, Y i is X i These are the labels (training data). Furthermore, any training data included in the training dataset D for the predictive model will simply be represented as (X,Y).

[0048] Step S103: Next, the prediction model learning unit 203 learns the prediction model using the prediction model learning dataset stored in the prediction model learning dataset storage unit 208. Then, the prediction model learning unit 203 saves the parameters of the learned prediction model to the prediction model storage unit 209.

[0049] As a prediction model, any machine learning model used for time series data prediction tasks can be employed, but below, as an example, we will assume the prediction model shown in Figure 7. The prediction model shown in Figure 7 consists of a one-dimensional ResNet (1d-ResNet), a GRU (gated recurrent unit), a fully connected layer (Dense), and a sigmoid layer (Sigmoid). In this prediction model configuration, a unit electrocardiogram data sequence of sequence length L {x1,···,xL When} is entered, the i-th unit electrocardiogram data x will be entered sequentially for i=1,···,L. i In a one-dimensional ResNet, z i It is converted to h s =GRU(z s ,h s-1 ) is hidden by h s This is calculated recursively, and the final hidden state h L The expression is converted to a one-dimensional scalar value in a fully connected layer, and the prediction result ^y is obtained using a sigmoid function. The hidden state h0 can be, for example, an appropriate initial vector.

[0050] The above prediction result ^y is a unit electrocardiogram data series {x1,···,x} of series length L. L This value represents the probability that atrial fibrillation will occur during the ΔT period following the Δt obtained, and takes a value between 0 and 1. For convenience, in the text of this specification, the symbol "^" representing the predicted value is placed immediately before the variable.

[0051] At this time, the prediction model learning unit 203 learns the prediction model so that the prediction result ^y when a unit electrocardiogram data sequence X ∈ D is input to the prediction model accurately estimates the label Y assigned to that unit electrocardiogram data sequence X. The prediction model can be learned using a known supervised learning method, such as backpropagation.

[0052] Step S104: When the electrocardiogram data to be predicted is provided, the prediction unit 204 uses this electrocardiogram data to predict the occurrence of atrial fibrillation after Δt using a trained prediction model.

[0053] For example, a unit electrocardiogram data sequence of sequence length L {x1,···,x L Let} be given. Here, x L This is the latest unit electrocardiogram data for ΔT=1 [min], and x L-i (i=1,···,L-1) represents the unit electrocardiogram data from (i+1)ΔT minutes before to iΔT minutes before. That is, x L-i(i=0,···,L-1) represents the unit electrocardiogram data from (i+1)ΔT minutes before to iΔT minutes before.

[0054] At this time, the prediction unit 204 generates a unit electrocardiogram data sequence of sequence length L {x1,···,x L Using}, the trained prediction model calculates the prediction result ^y. This gives us the prediction result ^y which indicates whether or not atrial fibrillation will occur during the interval ΔT after Δt.

[0055] Note that ^y takes a value between 0 and 1 and represents the probability of atrial fibrillation occurring during the interval ΔT after Δt. Therefore, the prediction unit 204 may, for example, use 0.5 as a threshold to determine whether ^y is greater than or equal to that threshold. In this case, if the prediction unit 204 determines that ^y is greater than the threshold, it may provide information indicating that "atrial fibrillation will occur after Δt" or "there is a high probability that atrial fibrillation will occur after Δt" as the prediction result. Otherwise, it may provide information indicating that "atrial fibrillation will not occur after Δt" or "there is a low probability that atrial fibrillation will occur after Δt".

[0056] <Rating> Using CPSC2018 as the first dataset and MIMIC-III as the second dataset, the data (X,Y) included in the training dataset D for the prediction model created in step S102 above was divided into 80% training data, 10% validation data, and 10% test data. The sequence length L was varied from 1 to 5, and Δt from 0 to 5 minutes. After training the prediction model with the training data, its prediction performance was evaluated using the test data with an F1 score. Five tests were conducted using the same test data, and the F1 score was the average of these tests. The evaluation results are shown in Table 1 below.

[0057] [Table 1]

[0058] As shown in Table 1 above, even in the case of L=1, Δt=5 [minutes], which is considered the most difficult to predict within the range of sequence length L=1~5 and Δt=0~5, an F1 score of over 0.7 was obtained, indicating that high prediction performance was achieved.

[0059] <Summary> As described above, the prediction device 10 according to this embodiment creates a training dataset for the prediction model by (1) to (3) above. This allows for obtaining a large amount of training data necessary to obtain an accurate prediction model, and as a result, the trained prediction model can accurately predict whether or not atrial fibrillation will occur in the future.

[0060] Therefore, by using the prediction device 10 according to this embodiment, for example, doctors can predict the onset of atrial fibrillation in a patient, enabling early intervention. This makes it possible not only to prevent the onset of atrial fibrillation itself, but also to prevent the onset of other complications, and as a result, it is expected that the length of stay in the intensive care unit or hospitalization period, and the increase in medical expenses will be prevented.

[0061] In the above embodiment, the occurrence of atrial fibrillation was predicted from electrocardiogram data. However, in addition to electrocardiogram data, time-series data representing various information obtainable from the human body may also be used to predict the occurrence of atrial fibrillation. For example, in addition to electrocardiogram data, time-series data representing information such as arterial pressure, pulmonary artery pressure, central venous pressure, and blood oxygen saturation may be used to predict the occurrence of atrial fibrillation.

[0062] The present invention is not limited to the embodiments specifically disclosed above, and various modifications, changes, and combinations with known technologies are possible without departing from the scope of the claims. [Explanation of Symbols]

[0063] 10 Prediction device 101 Input Device 102 Display device 103 External I / F 103a Recording medium 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage 108 processors 109 Bus 201 Classification Model Learning Unit 202 Training Data Creation Department 203 Predictive Model Learning Department 204 Prediction Section 205 First Dataset Storage Unit 206 Classification Model Memory Unit 207 Second Dataset Storage Unit 208 Storage for training datasets for predictive models 209 Predictive Model Memory Unit

Claims

1. A classification model learning unit is configured to learn a classification model that classifies unit electrocardiogram data representing electrocardiogram data for a predetermined time interval ΔT into classes that include at least normal and arrhythmias, using first electrocardiogram data to which first labels that can represent at least normal and arrhythmias are assigned; A data creation unit is configured to create a third electrocardiogram data set, which is represented by a sequence of L (where L is a predetermined integer of 1 or more) unit electrocardiogram data, using a second electrocardiogram data set for a time interval of ΔT or longer, and the learned classification model, and which is assigned a second label indicating normal or arrhythmia after a predetermined time Δt. A prediction model learning unit is configured to learn a prediction model that uses the third electrocardiogram data to predict whether the time after time Δt is normal or arrhythmia, given a sequence of L unit electrocardiogram data. A prediction device having the following features.

2. The aforementioned data creation unit, Each of the unit electrocardiogram data obtained by dividing the second electrocardiogram data into the time interval ΔT is classified into a class including normal and arrhythmia using the learned classification model. The prediction device according to claim 1, configured to create the third electrocardiogram data by assigning a label representing the class of the unit electrocardiograms at a time Δt after the sequence of L unit electrocardiogram data as the second label.

3. The aforementioned data creation unit, The prediction device according to claim 2, wherein the sequence of L unit electrocardiogram data, in which the unit electrocardiogram data from the beginning up to L-1 are classified into a class representing normal, and the L-th unit electrocardiogram data is classified into a class representing either normal or arrhythmia, is configured to create the third electrocardiogram data by assigning a label representing the class of the unit electrocardiogram after time Δt as the second label to the sequence of L unit electrocardiogram data in which the class of the unit electrocardiogram data classified after time Δt is either normal or arrhythmia.

4. The prediction device according to any one of claims 3, wherein the time Δt is expressed as Δt = mΔT, where m is a predetermined integer of 0 or more.

5. A prediction device according to any one of claims 1 to 4, comprising a prediction unit configured to predict whether a time Δt later is normal or arrhythmia, given a sequence of L unit electrocardiogram data, using a trained prediction model.

6. The prediction device according to claim 5, wherein the arrhythmia is atrial fibrillation.

7. The aforementioned classification model learning unit, The prediction device according to claim 6, which is configured to learn a classification model that classifies electrocardiogram data for a time interval ΔT into classes including normal, atrial fibrillation, and arrhythmias other than atrial fibrillation, using first electrocardiogram data to which first labels that can represent at least normal, atrial fibrillation, and arrhythmias other than atrial fibrillation are assigned.

8. A classification model learning procedure for learning a classification model that uses first electrocardiogram data to which first labels that can represent at least normal and arrhythmias are assigned, to classify unit electrocardiogram data representing electrocardiogram data for a predetermined time interval ΔT into classes that include at least normal and arrhythmias, A data creation procedure for creating a third electrocardiogram data set, which is represented by a sequence of L (where L is a predetermined integer of 1 or more) unit electrocardiogram data, and which is assigned a second label indicating normal or arrhythmia after a predetermined time Δt, using a second electrocardiogram data set for a time interval of ΔT or longer and the learned classification model, A prediction model learning procedure for learning a prediction model that uses the third electrocardiogram data to predict whether the time after time Δt is normal or arrhythmia, given a sequence of L unit electrocardiogram data, A prediction method performed by a computer.

9. A classification model learning procedure for learning a classification model that uses first electrocardiogram data to which first labels that can represent at least normal and arrhythmias are assigned, to classify unit electrocardiogram data representing electrocardiogram data for a predetermined time interval ΔT into classes that include at least normal and arrhythmias, A data creation procedure for creating a third electrocardiogram data set, which is represented by a sequence of L (where L is a predetermined integer of 1 or more) unit electrocardiogram data, and which is assigned a second label indicating normal or arrhythmia after a predetermined time Δt, using a second electrocardiogram data set for a time interval of ΔT or longer and the learned classification model, A prediction model learning procedure for learning a prediction model that uses the third electrocardiogram data to predict whether the time after time Δt is normal or arrhythmia, given a sequence of L unit electrocardiogram data, A program that causes a computer to execute something.

Images