Classification device, classification method, and program

Abstract

A classification device according to one aspect of the present disclosure is provided with: a creation unit that uses first series data composed of a plurality of feature amounts to create second series data composed of only one feature amount in each prescribed time zone; a model training unit that uses the second series data and correct answer labels for the first series data to train a classification model that predicts a label by using arbitrary length series data as an input; an order estimation unit that uses third series data composed of a plurality of feature amounts, correct answer labels for the third series data, and the trained classification model to estimate an optimal order of the types of the feature amounts for which the highest classification performance can be obtained when series data composed of only one feature amount in each of the time zones is input to the trained classification model; and a prediction unit that predicts a label for fourth series data by using the trained classification model and the fourth series data in which only one feature amount is measured in each of the time zones in accordance with the optimal order.

Description

Classification device, classification method, and program

[0001] The present disclosure relates to a classification device, a classification method, and a program.

[0002] Methods for classifying diseases from electrocardiogram data using machine learning techniques have been proposed. For example, Non-Patent Document 1 describes a method for classifying 12-lead electrocardiogram data into nine classes (NSR (normal sinus rhythm), AF (atrial fibrillation), IAVB (first-degree atrioventricular block), LBBB (left bundle branch block), RBBB (right bundle branch block), PAC (supraventricular extrasystole), PVC (premature ventricular extrasystole), STD (ST segment depression), and STE (ST segment elevation)) using a machine learning technique.

[0003] Zhang, D., Yang, S., Yuan, X., & Zhang, P. (2021). Interpretable deep learning for automatic diagnosis of 12-lead electrocardiogram. iScience, 24(4), 102373.

[0004] In general, more accurate classification results can be obtained by using 12-lead electrocardiogram data measured simultaneously at 12 locations (i.e., series data of 12 types of feature amounts) than by using electrocardiogram data measured at one location (i.e., series data of one type of feature amount). However, measuring 12-lead electrocardiogram data requires expensive equipment and requires time for measurement preparation.

[0005] The present disclosure has been made in consideration of the above points, and provides a technique that can accurately classify sequence data measured while sequentially changing the type of feature.

[0006] A classification device according to one aspect of the present disclosure includes: a creation unit that uses first sequence data composed of multiple features to create second sequence data composed of only one feature for each predetermined time period; a model training unit that uses the second sequence data and a correct label for the first sequence data to train a classification model that predicts labels for input sequence data of an arbitrary length; an order estimation unit that uses third sequence data composed of multiple features, the correct label for the third sequence data, and the trained classification model to estimate an optimal order of feature types that will achieve the highest classification performance when the sequence data composed of only one feature for each time period is input to the trained classification model; and a prediction unit that uses fourth sequence data in which only one feature is measured for each time period according to the optimal order and the trained classification model to predict a label for the fourth sequence data.

[0007] A technique is provided that can accurately classify measured sequence data while sequentially changing the type of feature.

[0008] 1 is a diagram illustrating an example of a classification model; FIG. 2 is a diagram illustrating an example of a hardware configuration of a classification device according to an embodiment; FIG. 3 is a diagram illustrating an example of a functional configuration of a classification device according to an embodiment; FIG. 4 is a flowchart illustrating an example of a model learning process according to an embodiment; FIG. 5 is a flowchart illustrating an example of an order estimation process according to an embodiment; and FIG. 6 is a flowchart illustrating an example of a classification process according to an embodiment.

[0009] An embodiment of the present invention will be described below. In the following embodiment, a classification device 10 will be described that can accurately classify sequence data measured while sequentially changing the type of feature (i.e., one-dimensional sequence data in which the type of feature changes). In the following, electrocardiogram data is mainly assumed as the sequence data, and the types of feature are assumed to be electrocardiogram measurement locations (leads 1, 2, 3, aVR, aVL, aVL, V1 to V6). Note that "measurement" may also be referred to as, for example, "observation" or "measurement."

[0010] However, electrocardiogram data is just one example, and the classification device 10 according to the present embodiment can be applied to classifying any one-dimensional series data in which the type of feature changes. For example, in addition to electrocardiogram data, the classification device 10 according to the present embodiment can also be applied to classifying one-dimensional series data composed of multiple types of data, such as one-dimensional series data composed of electrocardiogram data and cardiac sound data.

[0011] <Classification Model> A classification model for classifying one-dimensional sequential data with varying feature types can be any machine learning model that takes one-dimensional sequential data of any length as input and outputs a label indicating the class to which the sequential data belongs. Below, the classification model shown in Figure 1 is mainly assumed as an example. The classification model shown in Figure 1 is composed of a convolutional neural network (CNN) that takes sequential data over a unit period as input, an average pooling layer that averages the output of the CNN over the entire period (K unit periods in the example shown in Figure 1), and a forward propagation neural network (FNN) that takes the output of the average pooling layer as input and outputs a label. Here, the unit period refers to a period of a predetermined length. Hereinafter, the classification model will be represented by f, and the length of the unit period (i.e., the time width of the period) will be represented by T.

[0012] However, the classification model shown in FIG. 1 is just an example and is not limited to this. Any machine learning model can be used as long as it takes one-dimensional sequence data of any length as input and outputs a label indicating the class to which the sequence data belongs.

[0013] <Example of Hardware Configuration of Classification Device 10> An example of the hardware configuration of the classification device 10 according to this embodiment is shown in Fig. 2. As shown in Fig. 2, the classification 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 pieces of hardware is connected to each other via a bus 109 so as to be able to communicate with each other.

[0014] The input device 101 is, for example, a keyboard, a mouse, a touch panel, a physical button, etc. The display device 102 is, for example, a display, a display panel, etc. Note that the classification device 10 does not necessarily have to include at least one of the input device 101 and the display device 102, for example.

[0015] The external I / F 103 is an interface with an external device such as a recording medium 103a. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

[0016] The communication I / F 104 is an interface for connecting the classification device 10 to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily stores programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can store programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), or a flash memory. The processor 108 is a computing device such as a CPU (Central Processing Unit) or a GPU (Graphic Processing Unit).

[0017] 2 is an example, and the hardware configuration of the classification device 10 is not limited to this. For example, the classification device 10 may have multiple auxiliary storage devices 107 or multiple processors 108, may not have some of the hardware shown in the figure, or may have various hardware other than the hardware shown in the figure.

[0018] <Example of Functional Configuration of Classification Device 10> Fig. 3 shows an example of the functional configuration of the classification device 10 according to this embodiment. As shown in Fig. 3, the classification device 10 according to this embodiment includes a learning unit 201, an order estimation unit 202, and a classification unit 203. These units are realized, for example, by processing in which one or more programs installed in the classification device 10 are executed by the processor 108 or the like. The classification device 10 according to this embodiment also includes a memory unit 204. The memory unit 204 is realized, for example, by a memory area of ​​the auxiliary storage device 107 or the like. However, the memory unit 204 may also be realized, for example, by a memory area of ​​a storage device connected to the classification device 10 via a communication network.

[0019] The learning unit 201 receives a learning data set D １ is used to learn a classification model f for classifying one-dimensional sequence data in which the types of features change.

[0020] Here, when training a classification model f, a training dataset D １ = {(x ｎ , y ｎ )|n=1,...,N １} is given. N １ is the training dataset D １ is the number of training data included in x ｎ is the n-th D-dimensional sequence data, and is expressed as follows:

[0021] However, x ｎｔｄ is the dth feature at time t, D is the number of feature types, M ｎ is the sequence length of the n-th D-dimensional sequence data. ｎ is the nth D-dimensional sequence data x ｎFor example, the label (correct label) for the sequence data x ｎ is a vector whose elements are the probability that each of the classes belongs to a predetermined class. ｎ is 12-lead electrocardiogram data, then D=12.

[0022] At this time, the learning unit 201 performs the following steps to learn the classification model f (i.e., each iteration when updating the parameters of the classification model f): ｎ Using this, we can generate sequence data x consisting of only one type of feature in each time period. ｎ For example, the period in which the same type of feature is measured is defined as unit period Δ, and the kth unit period is defined as Δ ｋ The learning unit 201 performs the learning for n=1, . . . , N １ For each of the sequence data x ｎ From series data x ｎ ', where Δ ｋ refers to the period from t=(k-1)T+1 to t=kT.

[0023] Step 11: The learning unit 201 ｎ T≦M ｎ An integer m of 1 or more that satisfies ｎ Select randomly.

[0024] Step 12: The learning unit 201 performs the learning process for the unit period Δ ｋ (k=1,...,m ｎ ) for each of the feature types d ｋ Select randomly.

[0025] Step 13: The learning unit 201 performs the learning process for the unit period Δ ｋ (k=1,...,m ｎ ) for each unit period Δ ｋ Series data x ｎ From d ｋ th feature x ｎｔｄ_ｋ (However, "d_k" is "d ｋ That is, the learning unit 201 extracts the unit period Δ ｋ (k=1,...,m ｎ ) for each of dｋ th feature x ｎｔｄ_ｋ (t∈Δ ｋ ) to extract.

[0026] Step 14: The learning unit 201 ｋ th feature x ｎｔｄ_ｋ (t∈Δ ｋ ,k=1,...,m ｎ ) ｎ That is, the learning unit 201 sets x ｎ '=(x ｎｔｄ_ｋ ), (t∈Δ ｋ ,k=1,...,m ｎ ) As a result, sequence data consisting of only one random feature value in each unit period and of random length is obtained.

[0027] The learning unit 201 also uses the sequence data x ｎ ' and its corresponding label y ｎ The parameters of the classification model f are updated (learned) using the above so that the classification performance is improved. For example, the learning unit 201 updates (learns) the parameters of the classification model f for n=1, . . . , N １ f(x) ｎ ') and y ｎ The parameters of the classification model f are updated (learned) using a known optimization method such as a stochastic gradient method so as to minimize errors such as the cross-entropy loss of the classification model f.

[0028] The order estimation unit 202 estimates the order of the given development dataset D ２ and the classification model f learned by the learning unit 201 (hereinafter also referred to as the learned classification model f), to estimate the optimal order of the types of feature values ​​when measuring while sequentially changing the types of feature values.

[0029] Here, when estimating the optimal order of feature types, the development dataset D ２ = {(x ｎ , y ｎ )|n=1,...,N ２} is given. N ２ is the development dataset D ２ The number of development data included in the development dataset D ２ is the training dataset D１ or the training data set D １ It may be different from.

[0030] The following is the development dataset D ２ x included in ｎ is the n-th D-dimensional sequence data, and the training dataset D １ x included in ｎ Similarly, it is expressed as follows:

[0031] In addition, x ｎｔｄ is the dth feature at time t, D is the number of feature types, M ｎ is the sequence length of the n-th D-dimensional sequence data.

[0032] At this time, the order estimation unit 202 estimates the optimal order of the types of feature quantities when measuring while sequentially changing the types of feature quantities so that the classification performance of the trained classification model f is maximized. For example, the order estimation unit 202 uses a greedy algorithm to estimate the optimal order of the types of feature quantities when measuring while sequentially changing the types of feature quantities. １ Specifically, the order estimation unit 202 determines the optimal order o of the feature types by the following steps 21 to 24. Note that hereinafter, the optimal order from 1st to kth is referred to as o ｋ = (d １ , ..., d ｋ )

[0033] Step 21: The order estimation unit 202 performs the order estimation for n=1, . . . , N ２ For each of the first unit period Δ １ Sequence data x in ｎ (That is, feature x ｎｔｄ (t∈Δ １ )) to obtain D pieces of sequence data x ｎｄ （１） = (x ｎｔｄ ), (t∈Δ １ , d=1,...,D).

[0034] Step 22: The order estimation unit 202 calculates the order of n=1, . . . , N for each of d=1, . . . , D. ２ f(x)ｎｄ （１） ) and y ｎ The error (e.g., cross-entropy loss) of １ This gives the optimal order o １ = (d １ ) is determined.

[0035] Step 23: The order estimation unit 202 determines the optimal order o ｋ－１ (where k≧2) is determined, the optimal order o ｋ Determine.

[0036] Step 23-1: The order estimation unit 202 performs the order estimation for n=1, . . . , N ２ For each of the k-th unit period, the sequence data x ｎ (That is, feature x ｎｔｄ (t∈Δ ｋ ) and the sth period Δ ｓ (s=1,...,k-1) is the feature x ｎｔｄ_ｓ (However, "d_s" is "d ｓ ") is the sequence data x ｎｄ （ｋ－１） = (x ｎｔｄ_ｓ ), (t∈Δ ｓ , s=1, . . . , k−1), D pieces of sequence data x ｎｄ （ｋ） (d=1, ..., D) is created. Here, the dth sequence data x ｎｄ （ｋ） = (x ｎｔｄ' ) is t∈Δ ｓ For (s=1,...,k-1), d'=d ｓ , t∈Δ ｋ Regarding the feature x ｎｔｄ' This is series data.

[0037] Step 23-2: The order estimation unit 202 calculates n=1, . . . , N for each of d=1, . . . , D. ２ f(x) ｎｄ （ｋ） ) and y ｎ The error (e.g., cross-entropy loss) ofｋ This gives the optimal order o ｋ = (d １ , ..., d ｋ ) is determined.

[0038] Step 24: The order estimation unit 202 sets k←k+1 and determines whether kT≦minM ｎ If so, repeat step 23 above and check if kT>minM ｎ If , the optimal order is o←o ｋ－１ Let's say.

[0039] The classification unit 203 classifies the given test data x using the trained classification model f (i.e., predicts the label y using f(x)). Here, the test data x is ｋ The test data x is a sequence data of an arbitrary length in which the feature quantity is measured in the optimal order o for each test data. ｔｄ' ) is, for example, the type of the kth feature included in the optimal order o as d ｋ For k=1,...,L (where L is the number of unit periods included in the test data x), t∈Δ ｋ When d' = d ｋ The feature x ｔｄ' In other words, the test data x is sequence data measured for each unit period while sequentially changing the type of feature according to the optimal order o (i.e., one-dimensional sequence data in which the type of feature changes for each unit period according to the optimal order o).

[0040] The storage unit 204 stores various data. Examples of the various data stored in the storage unit 204 include parameters of the classification model f, a learning data set D, and the like. １ , Development Dataset D ２ , test data x, etc.

[0041] 3, the learning unit 201, the order estimation unit 202, and the classification unit 203 are included in one classification device 10, but this is merely an example, and these units may be distributed across multiple devices. For example, one device may have the learning unit 201 and the order estimation unit 202, and another device may have the classification unit 203. Alternatively, for example, three devices may each have the learning unit 201, the order estimation unit 202, and the classification unit 203.

[0042] <Model Learning Process> The model learning process for learning the classification model f will be described below with reference to FIG.

[0043] The learning unit 201 uses the learning data set D stored in the storage unit 204. １ is input (step S101).

[0044] Next, the learning unit 201 uses the learning data set D １ Each series data x included in ｎ Using this, we can generate sequence data x consisting of only one type of feature in each time period. ｎ For example, the learning unit 201 randomly generates each of the sequence data x ｎ From series data x ｎ ' respectively.

[0045] Next, the learning unit 201 calculates each sequence data x ｎ ' and its corresponding label y ｎ The parameters of the classification model f are updated (learned) using the above so that the classification performance is improved (step S103). For example, the learning unit 201 updates (learns) the parameters of the classification model f for n=1, . . . , N １ f(x) ｎ ') and y ｎ The parameters of the classification model f are updated (learned) using a known optimization method such as a stochastic gradient method so as to minimize errors such as the cross-entropy loss of the classification model f.

[0046] Next, the learning unit 201 determines whether to terminate the model learning process (step S104). Here, the learning unit 201 may determine to terminate the model learning process if a predetermined termination condition is met, and may determine not to terminate the model learning process if a predetermined termination condition is not met. Examples of the termination condition include the number of repetitions of steps S102 to S103 exceeding a predetermined number, or the parameters of the classification model f converging.

[0047] If it is determined in step S104 above that the model learning process should be terminated, the learning unit 201 terminates the model learning process. On the other hand, if it is determined in step S104 above that the model learning process should not be terminated, the learning unit 201 returns to step S102 above. As a result, steps S102 to S103 above are repeatedly executed until it is determined in step S104 above that the model learning process should be terminated. Note that one repetition of steps S102 to S103 above is also called a learning step, an iteration, or the like.

[0048] <Order Estimation Process> The order estimation process for estimating the optimal order o will be described below with reference to FIG.

[0049] The order estimation unit 202 uses the development dataset D stored in the storage unit 204 ２ is input (step S201).

[0050] The order estimation unit 202 uses the development dataset D ２ Each series data x included in ｎ and its label y ｎ and estimate an optimal order o of the types of feature quantities to be measured while sequentially changing the types of feature quantities so that the classification performance of the trained classification model f is maximized (step S202). For example, the order estimation unit 202 determines the optimal order o by the above steps 21 to 24.

[0051] <Classification Process> Hereinafter, a classification process for classifying sequence data x measured for each unit period while sequentially changing the type of feature amount according to the optimal order o will be described with reference to FIG.

[0052] The classification unit 203 receives the sequence data x stored in the storage unit 204 as test data x (step S301).

[0053] The classification unit 203 predicts the label y by f(x) using the trained classification model f and the test data x (step S302). As a result, the test data x is classified by the label y.

[0054] <Evaluation> To evaluate the classification device 10 according to this embodiment (hereinafter also referred to as the proposed method), an experiment was conducted using 12-lead electrocardiogram data with nine types of labels. In this experiment, 12-lead electrocardiogram data from 6,877 people was used, with 80% of the data used as a training data set, 10% as a development data set, and 10% as test data. Furthermore, one-dimensional electrocardiogram data was randomly generated with a unit period of 2 seconds, and a classification model f was trained using this data. Furthermore, AUC (Area Under the ROC Curve) was used as an evaluation index.

[0055] Table 1 below shows the AUCs for the proposed method and for the case where electrocardiogram data measured at one fixed location (one-lead electrocardiogram data) is used.

[0056] As shown in Table 1 above, the proposed method is able to predict correct labels with higher accuracy than when using electrocardiogram data measured at a single fixed location.

[0057] <Summary> As described above, in the classification device 10 according to this embodiment, the sequence data x consisting of only one type of feature value in each time period is ｎ ' are randomly created, and then these series data x ｎ ' and its corresponding label y ｎThe classification model f is trained from the feature types. The classification device 10 according to this embodiment also estimates the order of feature types that will provide the highest classification performance in the trained classification model f. This allows the classification device 10 according to this embodiment to accurately classify sequential data in which feature values ​​are measured for each time period according to this order. Therefore, even if it is not possible to measure multiple feature values ​​simultaneously, the classification device 10 according to this embodiment can accurately classify sequential data obtained by this measurement by measuring feature values ​​while changing the type of feature value for each time period according to the optimal order.

[0058] The present invention is not limited to the above-described specifically disclosed embodiments, and various modifications, changes, and combinations with known technologies are possible without departing from the scope of the claims.

[0059] 10 Classification 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 device 108 Processor 109 Bus 201 Learning unit 202 Order estimation unit 203 Classification unit 204 Storage unit

Claims

1. A classification device comprising: a creation unit that uses first sequence data consisting of a plurality of features to create second sequence data consisting of only one feature for each predetermined time period; a model learning unit that uses the second sequence data and a correct label for the first sequence data to learn a classification model that predicts a label by inputting sequence data of an arbitrary length; an order estimation unit that uses third sequence data consisting of a plurality of features, the correct label for the third sequence data, and the trained classification model to estimate an optimal order of feature types that will provide the highest classification performance when the sequence data consisting of only one feature for each time period is input to the trained classification model; and a prediction unit that uses fourth sequence data in which only one feature is measured for each time period according to the optimal order and the trained classification model to predict a label for the fourth sequence data.

2. The classification device of claim 1, wherein the creation unit randomly selects the number of time periods, randomly selects only one feature from the plurality of features for each of the time periods, and creates one-dimensional sequence data consisting of only the one selected feature for each of the time periods as the second sequence data.

3. The classification device according to claim 1 or 2, wherein the order estimation unit estimates the optimal order by using the third sequence data to create multiple sequence data consisting of only one feature for the time periods, starting from the first sequence data, and determining the type of feature that provides the highest classification performance when each of the created multiple sequence data is input to the trained classification model.

4. The classification device of claim 3, wherein the order estimation unit estimates the optimal order by repeating the following steps: when the order from the 1st time slot to the s-1th time slot has been determined, where D is the number of types of the plurality of feature quantities, creating D pieces of sequence data in the sth time slot consisting of only a d (d=1, ..., D)th feature quantity, using the third sequence data; and determining, as the type of the sth feature quantity, the type of feature quantity that provides the highest classification performance when each of the D pieces of sequence data consisting of the 1st time slot to the s-1th time slot and each of the D pieces of sequence data in the sth time slot is input to the trained classification model.

5. A classification method executed by a computer, comprising: a creation step of using first sequence data consisting of a plurality of features to create second sequence data consisting of only one feature for each predetermined time period; a model learning step of using the second sequence data and a correct label for the first sequence data to learn a classification model that predicts a label using input sequence data of any length; an order estimation step of using third sequence data consisting of a plurality of features, the correct label for the third sequence data, and the trained classification model to estimate an optimal order of feature types that will provide the highest classification performance when sequence data consisting of only one feature for each time period is input to the trained classification model; and a prediction step of predicting a label for the fourth sequence data using fourth sequence data in which only one feature is measured for each time period according to the optimal order and the trained classification model.

6. A program for causing a computer to execute the following steps: a creation procedure for using first sequence data consisting of multiple features to create second sequence data consisting of only one feature for each specified time period; a model learning procedure for using the second sequence data and a correct label for the first sequence data to learn a classification model that predicts a label by inputting sequence data of any length, the model learning procedure for using third sequence data consisting of multiple features, the correct label for the third sequence data, and the trained classification model to estimate an optimal order of feature types that will provide the highest classification performance when sequence data consisting of only one feature for each time period is input to the trained classification model; and a prediction procedure for predicting a label for the fourth sequence data using fourth sequence data in which only one feature is measured for each time period according to the optimal order, and the trained classification model.