Observation target selection device, method for selecting observation target, and program

Abstract

To select an observation target shown by a missing value according to individual data.SOLUTION: The present disclosure relates to an observation target selection device for selecting an observation target to observe in a prediction phase, and the device includes: an input unit for inputting data including a predetermined missing value due to no observations having been done; a prediction unit for selecting a specific observation target on the basis of the data including the predetermined missing value due to no observations having been done, by using a selection model trained with machine learning including data showing features of an observation target having been done, the selection model being for selecting a predetermined observation target of a plurality of observation target candidates shown by the predetermined missing value on the basis of the data including the predetermined missing value due to no observations having been done; and an output unit for outputting data of a selection result showing the specific observation target selected by the prediction unit.SELECTED DRAWING: Figure 5

Description

Technical Field

[0001] The present disclosure relates to an observation target selection device, an observation target selection method, and a program.

Background Art

[0002] Generally, when there are many missing values in data to which labels are assigned in the field of machine learning, the accuracy of predicting labels for that data becomes low. To improve the accuracy, it is necessary to observe more observation targets and reduce the missing values, but it may be costly to observe. For example, in the case of medical examinations, it is costly to examine many examination items (observation targets), and the burden on patients also increases. In such a case, a selection method for selecting a predetermined observation target from among unexamined items (a plurality of observation target candidates indicated by missing values) has already been proposed (Non-Patent Document 1).

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the conventional method, among a plurality of observation target candidates indicated by missing values, since a fixed observation target is selected in advance, there has been a problem that the observation target indicated by the missing value cannot be selected according to individual data (for example, data related to individual patients).

[0005] The present invention has been made in view of the above points, and an object thereof is to select an observation target indicated by a missing value according to individual data.

Means for Solving the Problems

[0006] In order to solve the above problems, the invention according to claim 1 is an observation target selection device that selects an observation target to be observed in a prediction phase, and includes an input unit that inputs data including a predetermined missing value due to non-observation, and a machine learning-trained selection model that includes data indicating the characteristics of the observed observation target. Using the machine learning-trained selection model for selecting a predetermined observation target from among a plurality of observation target candidates indicated by the predetermined missing value based on the data including the predetermined missing value due to non-observation, a prediction unit that selects a specific observation target based on the data including a specific missing value due to non-observation, and an output unit that outputs data of a selection result indicating the specific observation target selected by the prediction unit. The prediction unit uses a machine learning-trained prediction model for predicting the label of the data including the predetermined missing value, and includes the characteristics of the specific observation target obtained by newly observing based on the data of the selection result output by the output unit, and based on the data including the remaining missing values that are not observed, predicts the label of the data including the missing values other than the specific observation target, and the output unit outputs data of a prediction result indicating the label predicted by the prediction unit. It is an observation target selection device.

Effects of the Invention

[0007] As described above, according to the present invention, there is an effect that an observation target indicated by a missing value can be selected according to individual data.

Brief Description of the Drawings

[0008]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Mode for Carrying Out the Invention

[0009] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0010] 〔System Configuration of Embodiment〕 First, with reference to FIG. 1, an outline of the configuration of the communication system 1 of the present embodiment will be described. FIG. 1 is a schematic diagram of a communication system according to an embodiment of the present invention.

[0011] As shown in FIG. 1, the communication system 1 of the present embodiment is constructed by an observation target selection device 3 and a communication terminal 5. The communication terminal 5 is managed and used by the user Y.

[0012] In addition, the observation target selection device 3 and the communication terminal 5 can communicate via a communication network 100 such as the Internet. The connection form of the communication network 100 may be either wireless or wired.

[0013] The observation target selection device 3 is composed of one or more computers. When the observation target selection device 3 is composed of a plurality of computers, it may be referred to as an "observation target selection device" or a "label prediction system".

[0014] The observation target selection device 3 is a computer, and by machine learning, even if the input data is data including missing values, it selects a new observation target according to each piece of data and predicts the label of the data including missing values. Therefore, the observation target selection device 3 can also be referred to as a label prediction device.

[0015] For example, when the observation target selection device 3 is used for medical examinations, the observation target selection device 3 uses a machine-learned prediction model g2 to predict the label (e.g., the disease name of the patient) of data including missing values due to unexamined items. Further, the observation target selection device 3 uses a machine-learned selection model f2 including data indicating the already observed features (e.g., the "value" of height) of the patient to select an unexamined item (e.g., weight) that is a specific observation target among a plurality of observation target candidates (unexamined items) indicated by the missing values. Thereby, the doctor performs the next examination of the patient with reference to this selected unexamined item. Then, the observation target selection device 3 predicts again the label (such as the disease name of the patient) of the data including the remaining missing values (e.g., blood pressure, blood, etc.) including the features (e.g., the "value" of weight) of the examination item examined next. By repeating such processing and operations, the observation target selection device 3 can select examination items according to individual patients and predict the disease name of the patient, so that finally the disease name of each individual patient can be determined with as few examinations as possible.

[0016] The communication terminal 5 is a computer. In FIG. 1, a notebook personal computer is shown as an example, but it is not limited to a notebook type and may be a desktop personal computer. Further, the communication terminal may be a smartphone or a tablet terminal. In FIG. 1, the user Y is operating the communication terminal 5.

[0017] 〔Hardware Configuration of Observation Target Selection Device and Communication Terminal〕 Next, the hardware configurations of the observation target selection device 3 and the communication terminal 5 will be described with reference to FIG. 2. FIG. 2 is a hardware configuration diagram of the observation target selection device and the communication terminal.

[0018] As shown in FIG. 2, the observation target selection device 3 includes a processor 301, a memory 302, an auxiliary storage device 303, a connection device 304, a communication device 305, and a drive device 306. Each piece of hardware constituting the observation target selection device 3 is interconnected via a bus 307.

[0019] The processor 301 serves as a control unit that controls the entire observation target selection device 3 and has various arithmetic devices such as a CPU (Central Processing Unit). The processor 301 reads various programs onto the memory 302 and executes them. Note that the processor 301 may include a GPGPU (General-purpose computing on graphics processing units).

[0020] The memory 302 has main memory devices such as a ROM (Read Only Memory) and a RAM (Random Access Memory). The processor 301 and the memory 302 form a so-called computer, and by the processor 301 executing various programs read onto the memory 302, the computer realizes various functions.

[0021] The auxiliary storage device 303 stores various programs and various information used when the various programs are executed by the processor 301.

[0022] The connection device 304 is a connection device that connects an external device (for example, a display device 310 and an operation device 311) to the observation target selection device 3.

[0023] The communication device 305 is a communication device for transmitting and receiving various information to and from other devices.

[0024] The drive device 306 is a device for setting the recording medium 330. The recording medium 330 here includes media that optically, electrically, or magnetically record information, such as a CD-ROM (Compact Disc Read-Only Memory), a flexible disk, and a magneto-optical disk. The recording medium 330 may also include semiconductor memories that electrically record information, such as a ROM (Read Only Memory) and a flash memory.

[0025] Note that various programs installed in the auxiliary storage device 303 are installed, for example, when the distributed recording medium 330 is set in the drive device 306 and various programs recorded on the recording medium 330 are read by the drive device 306. Alternatively, various programs installed in the auxiliary storage device 303 may be installed by being downloaded from a network via the communication device 305.

[0026] Also, although FIG. 2 shows the hardware configuration of the communication terminal 5, since each configuration is the same except that the reference numerals change from the 300s to the 500s, these descriptions are omitted.

[0027] 〔Functional Configuration of Observation Target Selection Device〕 Next, the functional configuration of the observation target selection device will be described with reference to FIG. 3. FIG. 3 is a functional configuration diagram of the observation target selection device according to the present embodiment.

[0028] In FIG. 3, the observation target selection device 3 includes an input unit 31, a prediction model learning unit 32, a selection model learning unit 33, a prediction unit 34, and an output unit 39. Each of these units is a function realized by an instruction of the processor 301 in FIG. 2 based on a program.

[0029] The input unit 31 receives input of learning data D indicating characteristics (such as an inspection result that the height is 170 cm, etc.) of an observed observation target (such as height) from the communication terminal 5 of the user Y and data including missing values.

[0030] The prediction model learning unit 32 machine-learns a prediction model g1 based on the learning data D input by the input unit 31.

[0031] The selection model learning unit 33 machine-learns a selection model f1 based on the data including missing values.

[0032] The prediction unit 34 sequentially selects an observation target from among the observation target candidates indicating missing values using the learned selection model f2, and predicts the label of the data including the missing values using the learned prediction model g2. In the case of medical examinations, the observation targets include unexamined items (such as blood tests) indicating missing values and examined features (such as height and weight).

[0033] The output unit 39 outputs the result predicted by the prediction unit 34 as result data to the outside of the observation target selection device 3 (communication terminal 5, display device 310, etc.).

[0034] Note that the prediction model learning unit 32, the selection model learning unit 33, and the prediction unit 34 will be described in detail later.

[0035] 〔Processing or operation of the embodiment〕 Subsequently, with reference to FIGS. 4 to 5, the processing or operation of this embodiment will be described in detail.

[0036] <Learning phase> First, with reference to FIG. 4, the processing or operation in the learning phase of the observation target selection device 3 will be described. FIG. 4 is a flowchart showing the processing or operation in the learning phase of the observation target selection device according to this embodiment.

[0037] First, the input unit 31 inputs learning data D indicating the features of the observed observation targets from the communication terminal 5 of the user Y, the operation device 311, etc. (S11).

[0038] This learning data D is

Number

Number

[0039] Next, the prediction model learning unit 32 learns a prediction model g1 based on the learning data D. The prediction model g1 is a model for predicting the labels of data that may (or may not) contain missing values.

Number

Number

[0040] The prediction model g1 is learned by the prediction model learning unit 32 to fit the learning data D. For example, in the case of a classification problem where the label is a discrete value, the prediction model learning unit 32 can learn by minimizing the cross-entropy error between the predicted value and the true value. The prediction model learning unit 32 can improve the prediction performance in the case of missing values by artificially changing the learning data D into data containing missing values using a random mask vector and learning to reduce the expected error regarding the changed data. Note that the prediction model learning unit 32 can also change the data into data containing artificial missing values not randomly but by the selection model learning unit 33 sequentially observing the missing values using the selection model f1 described later.

[0041] Next, the input unit 31 inputs data containing missing values from the communication terminal 5 of the user Y, the operation device 311, etc. (S13).

[0042] Next, the selection model learning unit 33 learns a selection model f1 based on the data containing missing values. which is the selection model f1

Number

Number

[0043] As the selection model f1, any model such as a neural network or a linear model can be used. When the prediction performance improves when the d-th observation target is observed and does not improve when the d'-th observation target is observed by the prediction model learning unit 32, the score (S nd > S nd' ) is output so as to learn the selection model f1.

[0044] For example, such learning is possible by maximizing the objective functions represented by the following (Equation 1) and (Equation 2).

Number

Number

Number

Number

[0045] When the costs of inspections and the like required for observations are different, it is also possible to use an objective function weighted by the cost. For example, when the cost of the d-th feature is cd, [Number] is used as the objective function. As a result, for example, the higher the inspection cost, the less likely the inspection item or the like is to be selected as the observation target.

[0046] Thus, the processing or operation in the learning phase ends.

[0047] [Prediction Phase] Subsequently, with reference to FIG. 5, the processing or operation in the prediction phase of the observation target selection device 3 will be described. FIG. 5 is a flowchart showing the processing or operation in the prediction phase of the observation target selection device according to the present embodiment.

[0048] First, the input unit 31 inputs data including specific missing values from the communication terminal 5, the operation device 311, etc. (S21).

[0049] Next, the prediction unit 34 uses the machine-learned prediction model g2 for predicting the label of the data including the predetermined missing value, and based on the data including the specific missing value input by the input unit 31, predicts the label of the data including the missing value (for example, the disease name of the patient) (S22). Then, the output unit 39 outputs the data of the prediction result predicted by the prediction unit 34 to the outside of the observation target selection device 3 (communication terminal 5, display device 310, etc.) (S23).

[0050] Furthermore, the prediction unit 34 uses the machine-learned selection model f2 including data indicating the characteristics of the observed observation target (for example, the "value" of height), and from among the specific missing values in the data including the missing value input by the input unit 31, selects a specific observation target to be observed (for example, weight) (S24). Then, the output unit 39 outputs the data of the selection result (for example, weight) selected by the prediction unit 34 to the outside of the observation target selection device 3 (communication terminal 5, display device 310, etc.) (S25). As a result, the user Y (for example, a doctor) can predict the label of the data including the missing value (such as the disease name of the patient) and can grasp the observation target to be observed next (for example, weight).

[0051] Next, the prediction unit 34 determines whether or not the end condition is satisfied (S24). Examples of the end condition include the number of selections reaching a certain value or more, the number of observation targets reaching a certain value or more, the number of unobserved observation targets becoming a certain value or less, and the fluctuation of the predicted value becoming a certain value or less (for example, when the predicted disease name does not change even if the prediction is repeated, further prediction is meaningless). When the end condition is not satisfied (S26: NO), user Y (for example, a doctor) observes a new specific observation target (for example, weight) with reference to the selection result (for example, weight) by the above processing (S25) (for example, weight measurement), and the input unit 31 includes the characteristics (for example, the "value" of weight) of this specific observation target and data including the remaining missing values of unobserved observation targets (for example, blood pressure, blood) other than this specific observation target, and returns to the processing (S21) to input it. As a result, in the processing (S22) again, the prediction unit 34 predicts a label (for example, the disease name of the patient) using the learned prediction model g2, and in the processing (S23) again, outputs the data of the prediction result. In this case, since the number of observation targets has increased, there is a possibility that new prediction result data will be output.

[0052] Furthermore, in the processing (S24) again, the prediction unit 34 selects an observation target (for example, blood pressure) to be observed from among the remaining missing values using the learned selection model f2. Then, in the processing (S25) again, the output unit 39 outputs the data of the selection result (for example, blood pressure). In this way, by repeating the processing (S22, S24) by the prediction unit 34, sequential processing is performed, and the observation target can be selected according to the data (for example, disease name) of individual patients and the like. Also, thereby, according to individual patients and the like, the disease name and the like can be grasped with as few observation times as possible.

[0053] On the other hand, when the end condition is satisfied (S26: YES), the processing shown in FIG. 5 ends.

[0054] Note that the observation target selection device 3 repeats predicting labels for data including a specific observation target selected by the prediction unit 34 in the above-described processes (S22, S23), but it is not limited to this. For example, the observation target selection device 3 may select all observation targets in the process (S22) and then predict labels for data including all the observation targets selected in the process (S22) in the process (S23).

[0055] Thus, the processing or operation in the prediction phase ends.

[0056] 〔Evaluation Results〕 Subsequently, the evaluation results of the results predicted by the observation target selection device of the present embodiment will be described. FIG. 6 is a diagram showing the evaluation results.

[0057] FIG. 6 shows the results of evaluating the present embodiment using handwritten digit data. RL represents a method based on reinforcement learning, FI represents a method based on feature estimation, Var represents a method of selecting the one with the maximum variance of features, and Random represents a method of randomly selecting features. As shown in FIG. 6, the present embodiment achieves a higher correct answer rate compared to other methods.

[0058] 〔Main Effects of the Embodiment〕 As described above, according to the present embodiment, the observation target selection device 3 can select a specific observation target indicated by a missing value according to each data, which has the effect of being able to do so.

[0059] In addition, since the observation target selection device 3 does not need to observe all observation targets by the above-described process (S24), it has the effect of reducing the cost of observation and reducing the burden of observing observation targets such as the examination items of patients.

[0060] 〔Supplementary Explanation〕 The present invention is not limited to the above-described embodiment, and may have the following configurations or processes (operations).

[0061] The observation target selection device 3 can be realized by a computer and a program. However, it is also possible to record this program on a (non-transitory) recording medium or provide it via the communication network 100.

Explanation of Signs

[0062] 1 Communication system 3 Observation target selection device 5 Communication terminal 31 Input unit 32 Prediction model learning unit 33 Selection model learning unit 34 Prediction unit 39 Output unit

Claims

1. An observation target selection device for selecting an observation target to be observed in a prediction phase, an input unit for inputting data including predetermined missing values ​​due to unobserved data; a prediction unit that selects a specific observation target based on data including a specific missing value due to unobservation, using a machine-learned selection model that includes data indicating characteristics of observed observation targets, the machine-learned selection model being for selecting a specific observation target from among a plurality of observation target candidates indicated by the specific missing value based on data including the specific missing value due to unobservation; an output unit that outputs data of a selection result indicating the specific observation target selected by the prediction unit; having the prediction unit uses a machine learning-based prediction model for predicting a label of data including the predetermined missing value, and predicts a label of data including missing values ​​other than the specific observation target, based on data including remaining missing values ​​that have not been observed and that include features of the specific observation target obtained by newly observing the data of the selection result output by the output unit; The output unit outputs prediction result data indicating the label predicted by the prediction unit. Observation object selection device.

2. An observation target selection method executed by an observation target selection device for selecting an observation target to be observed in a prediction phase, comprising: The observation object selection device includes: An input process for inputting data including predetermined missing values ​​due to unobserved data; a prediction process for selecting a specific observation target based on data including a specific missing value due to unobservation, using a machine-learned selection model including data indicating characteristics of observed observation targets, the machine-learned selection model being for selecting a specific observation target from among a plurality of observation target candidates indicated by the specific missing value based on data including the specific missing value due to unobservation; an output process for outputting data of a selection result indicating the specific observation target selected by the prediction process; Run The prediction process includes a process of predicting a label of data including missing values ​​other than the specific observation target, based on data including remaining missing values ​​that have not been observed and including features of the specific observation target obtained by newly observing the data of the selection result output by the output process, using a machine learning prediction model for predicting a label of data including the specific missing value, The output process includes a process of outputting prediction result data indicating a label predicted by the prediction process. Method of selecting observation subjects.

3. A program for causing a computer to execute the method according to claim 2.

Images