JP6843109B2 - Medical data structure - Google Patents

Description

The present invention relates to medical data structures.

A graph structure learning technique capable of deep learning of graph structure data (hereinafter, a form of a device for performing such graph structure learning is referred to as a "DeepTensor (DT)") is known. .. The DT uses a graph structure as an input, and treats the graph structure as tensor data (hereinafter, may be referred to as a tensor). Then, in DT, highly accurate prediction is realized by extracting the partial structure of the graph (partial pattern of the tensor) that contributes to the prediction as a core tensor.

Japanese Unexamined Patent Publication No. 08-329196

However, in the above DT, although the core tensor can process a partial common pattern, data that are similar on the partial tensor but have different characteristics as a whole are actually processed as the same common pattern. In some cases, the prediction accuracy is reduced.

For example, when learning attendance record data by DT, the attendance record data of a person with poor physical condition and the attendance record data of a normal person are input to DT to learn a prediction model. Then, the attendance record data of the discriminant target person is input to the trained prediction model to predict the possibility that the discrimination target person will be treated (leave work).

In such learning, the attendance record data of a person who has developed a mental illness in the past and has undergone medical treatment is disturbed, such as frequent vacations and late arrivals even if he / she returns to work and is working normally. There may be. At this time, the attendance record data with disordered attendance, which is a sign before new medical treatment, may be in a similar state on a partial tensor. However, medical treatment does not always occur after that. Therefore, the attendance record data of the past person becomes noise, and the prediction accuracy of the new medical treatment person is lowered. The same applies not only to medical treatment predictions but also to hospitalization predictions.

One aspect is to provide a medical data structure that can suppress the deterioration of prediction accuracy.

In the first proposal, the medical data structure is composed of a patient's medical data composed of a plurality of elements, and whether the patient is in poor physical condition who has been hospitalized or a normal person who has never been hospitalized. Includes disease information to identify any. The learning device includes a storage unit that stores a target core tensor, a storage unit that stores a weighting rule for medical data of a recovery patient who has been hospitalized in the past but has been discharged from the hospital, and a generation unit that generates tensor data from the medical data. Have. In the data structure, when the patient corresponds to the recovery patient, the learning device changes the weight of any element of the medical care data to change the tensor data according to the weighting rule, and the target core tensor. A core tensor is generated from the modified tensor data so as to be similar to the above, and the core tensor is input to a learner using a deep tensor and used in a process of executing supervised learning.

According to one embodiment, deterioration of prediction accuracy can be suppressed.

FIG. 1 is a diagram illustrating an overall example of machine learning according to the first embodiment. FIG. 2 is a diagram illustrating an example of learning data. FIG. 3 is a diagram showing an example of the relationship between the graph structure and the tensor. FIG. 4 is a diagram showing an example of extraction of a subgraph structure. FIG. 5 is a diagram illustrating a learning example of a deep tensor. FIG. 6 is a functional block diagram showing a functional configuration of the learning device according to the first embodiment. FIG. 7 is a diagram showing an example of weight information stored in the weight information DB. FIG. 8 is a diagram showing an example of information stored in the attendance record data DB. FIG. 9 is a diagram showing an example of information stored in the learning data DB. FIG. 10 is a diagram for explaining the determination of the medical treatment period and the weight setting. FIG. 11 is a diagram illustrating tensorization. FIG. 12 is a diagram showing a comparative example of tensor data. FIG. 13 is a diagram for explaining the weight change on the tensor data of the past person. FIG. 14 is a diagram illustrating learning data that is not subject to weight change. FIG. 15 is a diagram for explaining the learning data that is the target of the weight change. FIG. 16 is a diagram illustrating the effect of the weight change. FIG. 17 is a flowchart showing the flow of the learning process. FIG. 18 is a flowchart showing a processing flow at the time of prediction. FIG. 19 is a diagram illustrating an example of learning data according to the second embodiment. FIG. 20 is a diagram illustrating determination of hospitalization period and weight setting. FIG. 21 is a diagram illustrating the tensorization according to the second embodiment. FIG. 22 is a diagram illustrating a hardware configuration example.

Hereinafter, examples of the medical data structure disclosed in the present application will be described in detail with reference to the drawings. The present invention is not limited to this embodiment. In addition, each embodiment can be appropriately combined within a consistent range.

[Overall example]
In recent years, employee physical condition management has been positioned as an important issue for companies to tackle, and from employee attendance record data, it is possible to predict mental illness several months in the future and take early measures such as counseling. It is being done. In general, full-time staff browses the attendance record data of a huge number of employees and has characteristic patterns of work conditions such as frequent business trips, long overtime, continuous absenteeism, absenteeism, and combinations of these. We are visually looking for employees who fall under. It is difficult to clearly define such a characteristic pattern because the standards may differ depending on each dedicated staff.

Therefore, in this embodiment, as an example of deep learning using the deep tensor, the attendance record data of employees and the like is targeted for learning, and the attendance record data of the poorly ill person and the attendance record data of the normal person are input to the deep tensor. Learn a predictive model that predicts employee mental health problems.

FIG. 1 is a diagram illustrating an overall example of machine learning according to the first embodiment. As shown in FIG. 1, the learning device 100 according to the first embodiment machine-learns attendance record data including situations such as daily commuting, leaving time, vacation acquisition, and business trip of an employee to generate a prediction model. The learning device 100 is an example of a computer device that predicts whether or not the employee will be treated (leave) or not from the attendance record data of the employee to be predicted by using the prediction model after learning. Although the example in which the learning device 100 executes the learning process and the prediction process will be described here, they can also be executed by separate devices.

For example, the learning device 100 includes employee attendance record data composed of a plurality of elements, and employee information that identifies whether the employee is a poorly ill person who has been treated or a normal person who has never been treated. (Label) and the learning data including. Further, the learning device 100 stores the target core tensor and the weighting rule for the attendance record data of the past person indicating the employee who has received medical treatment in the past but has returned to work.

In such a state, the learning device 100 executes tensorization on the learning data that does not correspond to the past person without changing the weight of the attendance record data to be tensorized (for example, it remains 1). Then, the learning device 100 decomposes the tensorized tensor data into tensors to generate a core tensor so as to resemble the target core tensor. After that, the learning device 100 inputs the core tensor into the learning device using the deep tensor and executes the supervised learning.

On the other hand, the learning device 100 changes the tensor data by changing the weight of any element of the attendance record data to be tensorized according to the weighting rule stored in advance for the learning data corresponding to the past person. To do. Then, the learning device 100 decomposes the weight-changed data into tensors to generate a core tensor similar to the target core tensor. After that, the learning device 100 inputs the core tensor into the learning device using the deep tensor and executes the supervised learning.

Here, the learning data to be input to the deep tensor will be described. FIG. 2 is a diagram illustrating an example of learning data. The learning data is composed of attendance record data every 6 months and a label indicating whether or not there is a medical treatment record within 3 months from the 6th month onward. FIG. 2 (a) is attendance record data of a person with poor physical condition with a label (with medical treatment), and FIG. 2 (b) is a normal person with a label without medical treatment (without medical treatment). Attendance record data. As shown in FIG. 2, the learning device 100 according to the first embodiment has "6 months'worth of attendance record data and label (with medical treatment)" and "6 months' worth of attendance record data and label (without medical treatment)". The prediction model is trained using "and" as training data. The learning device 100 predicts whether or not to receive medical treatment within 3 months from the attendance record data for 6 months of a certain person after learning. The shaded area in FIG. 2 indicates a vacation.

Next, the deep tensor will be described. Deep tensor is deep learning that inputs tensor (graph information), and automatically extracts a subgraph structure that contributes to discrimination along with learning of a neural network. This extraction process is realized by learning the parameters of the tensor decomposition of the input tensor data together with the learning of the neural network.

Next, the graph structure will be described with reference to FIGS. 3 and 4. FIG. 3 is a diagram showing an example of the relationship between the graph structure and the tensor. In the graph 20 shown in FIG. 3, four nodes are connected by edges showing the relationship between the nodes (for example, “correlation coefficient is equal to or higher than a predetermined value”). It is shown that there is no such relationship between the nodes that are not connected by the edge. When the graph 20 is represented by a second-order tensor, that is, a matrix, for example, the matrix representation based on the number on the left side of the node is represented by "matrix A", and the matrix based on the number on the right side of the node (the number surrounded by the box). The representation is represented by "matrix B". Each component of these matrices is represented by "1" when the nodes are connected (connected), and is represented by "0" when the nodes are not connected (not connected). .. In the following description, such a matrix is also referred to as a connection matrix. Here, the "matrix B" can be generated by simultaneously replacing the second and third rows and the second and third columns of the "matrix A". In the deep tensor, by using such a replacement process, the process is performed ignoring the difference in order. That is, "matrix A" and "matrix B" are treated as the same graph, ignoring the order in the deep tensor. The same process applies to tensors on the third floor and above.

FIG. 4 is a diagram showing an example of extraction of a subgraph structure. In the graph 21 shown in FIG. 4, six nodes are connected by an edge. The graph 21 can be expressed as shown in the matrix 22 when expressed by a matrix (tensor). A subgraph structure can be extracted by combining the operation of exchanging specific rows and columns, the operation of extracting specific rows and columns, and the operation of replacing non-zero elements in the connection matrix with zero for the matrix 22. .. For example, if the matrix corresponding to "nodes 1, 4, 5" of the matrix 22 is extracted, the matrix 23 is obtained. Next, when the value between "nodes 4 and 5" of the matrix 23 is replaced with zero, the matrix 24 is obtained. The subgraph structure corresponding to the matrix 24 is the graph 25.

The extraction process of such a subgraph structure is realized by a mathematical operation called tensor decomposition. The tensor decomposition is an operation that approximates the input n-th order tensor by the product of the n-th order and lower tensors. For example, when the input n-th order tensor is one n-th order tensor (called a core tensor) and n lower-order tensors (n> 2, the second-order tensor, that is, a matrix is usually used. ) Is approximated by the product. This decomposition is not unique and the core tensor can contain any subgraph structure in the graph structure represented by the input data.

The attendance record data can be composed of graph data including a plurality of nodes and an edge connecting the plurality of nodes. Here, the plurality of nodes are composed of nodes of date, month, and attendance classification. There are as many date, month, and attendance nodes as there are dates, months, and attendance categories, respectively. Each node stores values according to the date, month, and attendance classification status. For example, if the date is 1, a value "1" is set, if the attendance category is "vacation", a value "2" is set, and if "attendance", a value "1" is set. The edge is connected to the date node, the monthly node, and the related node of the attendance division node.

Next, learning of the deep tensor will be described. FIG. 5 is a diagram illustrating a learning example of a deep tensor. As shown in FIG. 5, the learning device 100 generates an input tensor from attendance record data with a teacher label (label A) such as with medical treatment. Then, the learning device 100 performs tensor decomposition on the input tensor to generate a core tensor so as to resemble the target core tensor randomly generated at the first time. Then, the learning device 100 inputs the core tensor into the neural network (NN: Neural Network) and obtains the classification result (label A: 70%, label B: 30%). After that, the learning device 100 calculates a classification error between the classification result (label A: 70%, label B: 30%) and the teacher label (label A: 100%, label B: 0%).

Here, the learning device 100 executes learning of the prediction model and learning of the tensor decomposition method by using the extended error propagation method which is an extension of the error back propagation method. That is, the learning device 100 modifies various parameters of the NN so as to reduce the classification error by propagating the classification error to the lower layer with respect to the input layer, the intermediate layer, and the output layer of the NN. Further, the learning device 100 propagates the classification error to the target core tensor so as to approach the partial structure of the graph that contributes to the prediction, that is, the feature pattern showing the characteristics of the poorly ill person or the feature pattern showing the characteristics of the normal person. Modify the target core tensor.

At the time of prediction after training, the input tensor is converted to a core tensor (partial pattern of the input tensor) by tensor decomposition so that it resembles the target core tensor, and the core tensor is input to the neural net. Can be obtained.

Then, the learning device 100 is used for a period (for example, 6 months + label) cut out as one learning data when learning using the tensors generated from the attendance record data of the normal person and the poor physical condition. Regardless of whether or not medical treatment is included within (3 months), the entire period of attendance record data is referred to to identify persons who have medical treatment experience (previous persons) in the past. Then, when the learning device 100 is converted into a core tensor (partial pattern extraction), the learning device 100 may have data partially similar to the disordered attendance record pattern, which is a sign before medical treatment. Later) Reduce the weight of the tensor for the period data.

By doing so, in the learning device 100, the attendance record data of the past person is partially similar to the attendance record data of the person who is in poor physical condition before starting new medical treatment, such as frequent vacations and late arrivals. However, a person with a history and a person with poor physical condition (new medical treatment person) can be extracted as separate core tensors. Therefore, since the learning device 100 can accurately learn the characteristics of the person who is in poor physical condition, it is possible to suppress the deterioration of the prediction accuracy of the new medical treatment.

[Functional configuration]
FIG. 6 is a functional block diagram showing a functional configuration of the learning device 100 according to the first embodiment. As shown in FIG. 6, the learning device 100 includes a communication unit 101, a storage unit 102, and a control unit 110.

The communication unit 101 is a processing unit that controls communication with other devices, and is, for example, a communication interface. For example, the communication unit 101 receives a processing start instruction, learning data, an input tensor in which the learning data is converted into a tensor, and the like from the terminal of the administrator. Further, the communication unit 101 outputs a learning result, a prediction result, and the like to the terminal of the administrator.

The storage unit 102 is an example of a storage device that stores programs and data, such as a memory and a hard disk. The storage unit 102 stores the weight information DB 103, the attendance record data DB 104, the learning data DB 105, the tensor DB 106, the learning result DB 107, and the prediction target DB 108.

The weight information DB 103 is a database that stores a weighting rule indicating the setting content of the weight to be set in the tensor data. FIG. 7 is a diagram showing an example of weight information stored in the weight information DB 103. As shown in FIG. 7, the weight information DB 103 stores the "type and set value (weight)" in association with each other. The "type" stored here indicates the type of data, and the "set value (weight)" indicates the value to be set.

In the example of FIG. 7, it is shown that the weight "0.5" is set for the tensor data after the medical treatment period of the past person, and the weight "1.0" is set for the other tensor data. Show that. Since the weight "1.0" can be interpreted as not changing the tensor data, a default value or the like can be used. Further, since the weight "0.5" can be interpreted as changing a part of the weight of the tensor data to lower the importance, a value smaller than the default value can be adopted. The weight can be set for each element such as monthly and attendance classification.

The attendance record data DB 104 is a database that stores attendance record data related to attendance of employees and the like. The attendance record data stored here is a data of the attendance record used by each company, and can be obtained from various known attendance management systems and the like. FIG. 8 is a diagram showing an example of information stored in the attendance record data DB 104. As shown in FIG. 8, the attendance record data is stored in association with "employee No., date, day of the week, attendance classification, attendance time, leaving time, overtime hours, business trip". In the attendance classification, types such as attendance, medical treatment, and vacation are stored. In addition, dates and days of the week are examples of elements that make up attendance record data.

In the example of FIG. 8, the employee No. It shows attendance record data for 100 employees. For example, the second line in Fig. 8 is the attendance record data for "Thursday, April 2, 2015". On this day, there was no business trip, and I arrived at "8:49" and left at "0:00". However, an example is shown in which overtime hours are "360 minutes". In addition, the 7th line of FIG. 8 is the attendance record data of "Wednesday, August 24, 2015", and indicates that "recuperation" was performed from this day to "Tuesday, October 4, 2015". The unit of attendance record data is not limited to daily, but may be weekly or monthly.

The learning data DB 105 is a database that stores learning data to be tensorized. Specifically, the learning data DB 105 stores data obtained by cutting out attendance record data in a period of 6 months and each learning data as a set of labels.

For example, if 6 months of attendance record data is used as one learning data and there is a medical treatment period within the next 3 months, "with medical treatment" is set as a label, and medical treatment is performed within the following 3 months. If there is no period, "no medical treatment" is set as the label. If the 6-month attendance record data includes a medical treatment period, that data will not be adopted as learning data. This is because it is known that those who already have "medical treatment" in the attendance record data for 6 months, which is the data (input) of the prediction source at the time of prediction, are clearly receiving medical treatment most recently. This is because it is not included in the monthly medical treatment prediction.

FIG. 9 is a diagram showing an example of information stored in the learning data DB 105. As shown in FIG. 9, the learning data DB 105 stores "employees, data (explanatory variables), labels (objective variables)" in association with each other. The "employee" stored here is an employee corresponding to the attendance record data that is the source of the learning data, and the learning data with the data as the explanatory variable and the label as the objective variable is stored.

In the example of FIG. 9, it is shown that "no medical treatment" is set as a label in the attendance record data of employee A from January to June, and employee A attends work from February to July. The book data indicates that "no medical treatment" is set as the label. In addition, it indicates that "with medical treatment" is set as a label in the attendance record data of employee A from March to August.

The tensor DB 106 is a database that stores each tensor (tensor data) generated from the learning data of each employee. The tensor DB 106 stores tensor data in which each tensor and a label are associated with each other. For example, the tensor DB 106 stores "tensor No. 1, label (without medical treatment)", "tensor No. 2, label (with medical treatment)" and the like as "data No., label". Each tensor stored here may be generated by a device other than the learning device 100, or may be generated by the learning device 100.

The learning result DB 107 is a database that stores the learning results. For example, the learning result DB 107 stores the discrimination result (classification result) of the learning data by the control unit 110, various parameters of the NN and various parameters of the deep tensor learned by machine learning or deep learning.

The prediction target DB 108 is a database that stores attendance record data of a target that newly predicts the presence or absence of medical treatment using a learned prediction model. For example, the forecast target DB 108 stores the forecast target attendance record data, the tensor generated from the forecast target attendance record data, and the like.

The control unit 110 is a processing unit that controls the processing of the entire learning device 100, and is, for example, a processor. The control unit 110 includes a learning data generation unit 111, a history determination unit 112, a weight setting unit 113, a tensor generation unit 114, a learning unit 115, and a prediction unit 116. The learning data generation unit 111, the history determination unit 112, the weight setting unit 113, the tensor generation unit 114, the learning unit 115, and the prediction unit 116 are examples of processes executed by an electronic circuit such as a processor or a processor. ..

The learning data generation unit 111 is a processing unit that generates learning data that is a set of labels corresponding to the start period and data for a certain period having different start dates from each attendance record data stored in the attendance record data DB 104. Specifically, the learning data generation unit 111 allows duplication from the attendance record data of one person and samples the data for a specified period. The learning data generation unit 111 extracts a plurality of data having different start (start) of the period from each attendance record data, and if there is a medical treatment period within 3 months from the end of the data for each data, the label "There is medical treatment". , And if there is no medical treatment period within 3 months from the end of the data, set the label "No medical treatment". After that, the learning data generation unit 111 stores the learning data in which the extracted data and the set label are associated with each other in the learning data DB 105.

For example, the learning data generation unit 111 extracts the attendance record data from January to June from the attendance record data from January to December. Then, the learning data generation unit 111 generates learning data by adding the label "no medical treatment" to the extracted attendance record data when there is no medical treatment period in the three months from July to September. Subsequently, the learning data generation unit 111 extracts the attendance record data from February to July from the attendance record data from January to December. Then, the learning data generation unit 111 generates learning data by adding the label "with medical treatment" to the extracted attendance record data when there is a medical treatment period for three months from August to October.

The past person determination unit 112 is a processing unit that determines whether or not the corresponding employee is a past person based on the attendance record data that is the source of each learning data. For example, the past person determination unit 112 refers to the attendance record data, which is the entire data section of the relevant employee, instead of cutting out the section such as “6 months” used for prediction, and if there is a “medical treatment period” in the past, It is judged as a past person, and if there is no "medical treatment period" in the past, it is judged as a normal person. Then, the past person determination unit 112 notifies the weight setting unit 113 of the determination result for each learning data.

The past person is not limited to the period used as one learning data, but refers to an employee who has a medical treatment period in the entire past attendance record data. For example, if the medical treatment period is two years before the time of study, the data for the last 6 months may not be a "sick person" but may be a "previous person".

The weight setting unit 113 is a processing unit that determines whether or not each learning data corresponds to the medical treatment period of a past person, and changes a part of the weight of the tensor data according to the determination result. Specifically, the weight setting unit 113 sets weights for each learning data stored in the learning data DB 105 according to the weighting rules stored in the weight information DB 103. Then, the weight setting unit 113 outputs the result of the weight setting to the tensor generation unit 114.

For example, the weight setting unit 113 sets the weight "0.5" for the learning data generated from the attendance record data that is determined to be a past person by the past person determination unit 112 and corresponds to the attendance record data after the medical treatment period. .. Further, the weight setting unit 113 sets the weight "1.0" for each learning data other than "previous person and after the medical treatment period". That is, the weight setting unit 113 sets weights for the learning data of a past person, which is an example of a certain condition, in order to reduce the importance of the data corresponding to the end of the medical treatment period after returning to work after the medical treatment.

Here, the setting of the medical treatment section and the weight will be described with reference to FIG. FIG. 10 is a diagram for explaining the determination of the medical treatment period and the weight setting. As shown in FIG. 10, the weight setting unit 113 determines whether each learning data corresponds to before or after the “medical treatment period” at the time of tensorization. In the example of FIG. 10, when the weight setting unit 113 detects the “medical treatment period” from “August 24, 2015” to “October 4, 2015”, it starts from “August 24, 2015” when the medical treatment starts. The weight is set to "1" for the training data that corresponds to the front, and the weight is set to "0.5" for the training data after "October 4, 2015" after the end of medical treatment.

That is, the pre-treatment period before the medical treatment period is determined to be important for extracting as a core tensor (partial pattern that affects the prediction) which is a partial pattern that becomes a medical treatment factor, and "1" is set as the weight. On the other hand, it is determined that the post-treatment period after the medical treatment period is not important as a target to be extracted as a core tensor which is a partial pattern that becomes a medical treatment factor, and "0.5" is set as a weight. In this way, a part of the weight of the tensor data is changed for the learning data corresponding to the medical treatment period of the past person.

The tensor generation unit 114 is a processing unit that converts each learning data into a tensor. Specifically, the tensor generation unit 114 is composed of elements included in each learning data for each learning data stored in the training data DB 105 and for which the weight setting by the weight setting unit 113 has been completed. The tensor is generated and stored in the tensor DB106. For example, the tensor generation unit 114 generates a fourth-order tensor composed of four elements included in each learning data for each learning data and stores it in the tensor DB 106. At this time, the tensor generation unit 114 stores the label (with medical treatment) or the label (without medical treatment) added to the learning data in association with the tensor.

Specifically, the tensor generation unit 114 generates a tensor from the learning data with each attribute that is supposed to characterize the tendency to recuperate as each dimension. For example, the tensor generation unit 114 generates a four-dimensional fourth-order tensor using four elements of month, date, attendance classification, and presence / absence of business trip. In the case of data for 6 months, the number of elements for each month is "6", and since the maximum number of dates for each month is 31, the number of elements for dates is "31", and the type of attendance is attendance. -For vacations and holidays, the number of attendance classification elements is "3", and since there are no business trips, the number of elements with or without business trips is "2". Therefore, the tensor generated from the learning data is a "6 x 31 x 3 x 2" tensor, and the value of the element corresponding to each month of the learning data, the attendance classification on the date, and the presence or absence of a business trip is 1, and the value of the element corresponding to that is not. The value becomes 0.

FIG. 11 is a diagram illustrating tensorization. As shown in FIG. 11, the tensor generated by the tensor generation unit 114 has monthly data in the horizontal direction, date in the vertical direction, attendance classification in the depth direction, business trips on the left side, and no business trips on the right side. The dates indicate the first day from the top, and the attendance classification indicates attendance, vacation, and holidays from the front. For example, FIG. 11 (a) shows an element of going to work on the first day of the month 1 and going on a business trip, and FIG. 11 (b) shows the element of taking a vacation on the second day of the month 1 and not going on a business trip. Indicates an element.

In this embodiment, the above-mentioned tensor will be simplified and described as shown in FIG. 11 (c). In other words, the monthly, date, attendance classification, and presence / absence of business trips are expressed in a cube shape, and the presence / absence of business trips for each month and date is expressed separately, and the attendance classification for each month and date is expressed separately. I will do it.

The learning unit 115 is a processing unit that receives each tensor and label generated from each training data as an input and executes learning of a prediction model by a deep tensor and learning of a method of tensor decomposition. Specifically, the learning unit 115 executes learning by utilizing the property of the deep tensor that "the partial structure of the graph (tensor) can be recognized". For example, the learning unit 115 extracts the core tensor from the input target tensor (input tensor) and inputs it to the NN, as in the method described with reference to FIG. 5, the classification result from the NN and the label given to the input tensor. And the error (classification error) is calculated. Then, the learning unit 115 uses the classification error to learn the parameters of the NN and optimize the target core tensor. After that, when the learning is completed, the learning unit 115 stores various parameters as learning results in the learning result DB 107.

In the tensor decomposition described in Example 1, the core tensor is calculated so as to be similar to the target core tensor, so that the structure important for classification is arranged at a similar position of the core tensor. Then, by learning the NN using this core tensor, highly accurate classification is realized.

The prediction unit 116 is a processing unit that predicts the label of the data to be discriminated by using the learning result. Specifically, the prediction unit 116 reads various parameters from the learning result DB 107 and constructs a neural network in which various parameters are set. Then, the prediction unit 116 reads the attendance record data of the prediction target from the prediction target DB 108, and generates a tensor to be input to the neural network from the attendance record data of the prediction target by the same method as at the time of learning.

After that, the prediction unit 116 acquires the prediction result of whether or not to be treated from the trained neural network (prediction model). Then, the prediction unit 116 displays the prediction result on the display or transmits the prediction result to the administrator terminal.

In addition, the prediction unit 116 refers to the attendance record data of the employee to be predicted, determines whether or not the medical treatment period is included before the attendance record data period of the prediction target, and includes the medical treatment period. It is also possible to execute a prediction in the case of a normal person who does not have a medical treatment period and output an alert in the case of a past person who includes a medical treatment period.

[Effect of weight change]
Next, with reference to FIGS. 12 to 16, the effect of changing the weight of the attendance record data of the past person after the medical treatment period to 0.5 on the learning of NN will be described. FIG. 12 is a diagram showing a comparative example of tensor data. FIG. 13 is a diagram for explaining the weight change on the tensor data of the past person. FIG. 14 is a diagram illustrating learning data that is not subject to weight change. FIG. 15 is a diagram for explaining the learning data that is the target of the weight change. FIG. 16 is a diagram illustrating the effect of the weight change.

Here, the attendance record data of employee A corresponding to a person with poor physical condition to which the label "with medical treatment" is given, and the attendance record data of employee B corresponding to a past person to which the label "without medical treatment" is given. Let's take an example. In addition, the attendance record data of employee B is the attendance record data after the medical treatment period.

Further, in FIG. 6 and the like, an example in which the weight "0.5" is uniformly set for all the elements for the attendance record data of the past person and after the medical treatment period has been described, but here, one element is used. An example of changing the weight will be described. That is, an example will be described in which the weight of the attendance category among the date, month, and attendance category of the attendance record data is changed to 0.5, and the weight other than the attendance category is set to 1.0. It should be noted that which element to change the weight can be set by the weighting rule.

As shown in FIG. 12, tensor data generated from the attendance record data of employee A (person with poor physical condition) to which the label "with medical treatment" is given, and employee B (past) to which the label "without medical treatment" is given. The tensor data generated from the attendance record data of the person) is different data when compared over a long period such as one year or two years. However, when focusing on the range of short-term P such as 6 months, the data are similar or the same, and the characteristics cannot be distinguished. That is, the data that should be treated separately is treated as similar data having the same feature amount in the data for 6 months of the extraction target extracted as learning data. Therefore, in the learning of the prediction model, since these are processed as the same case, the data of the past person becomes noise, which leads to the optimization of the target core tensor and the deterioration of the learning accuracy of the NN, resulting in the deterioration of the accuracy of the prediction model. To do.

Therefore, as shown in FIG. 13, the medical treatment of a former person (employee B) who may have partially similar data when core tensorized (partial pattern extraction) by weight setting by the weight setting unit 113. The dawn data changes some of the tensor weights. For example, in the case of a person who has a medical treatment period, as a part of the weight of the data section after medical treatment, the value of the record corresponding to the vacation such as "annual leave" or "quasi-absence" in the "attendance classification" is set. Set to "0.5". That is, the length of the edge is changed, and the value set in the node is changed.

In this way, by changing the weight of the data after the medical treatment period of the past person, the edge of the graph structure is changed, and as a result, the tensor data is changed. As a result, since the core tensor extracted from the attendance record data can be differentiated, the characteristics of the attendance record data of the past person and after the medical treatment period can be distinguished from the characteristics of the other attendance record data.

Specifically, as shown in FIG. 14, the learning device 100 sets the annual leave in the attendance classification for the attendance record data of the employee A who corresponds to the unwell person to which the label “with medical treatment” is given. Even if vacations such as are included, the weight "1.0" is set for all records. Therefore, the attendance record data has a graph structure as it is, and is converted into a tensor.

On the other hand, as shown in FIG. 15, for the attendance record data of employee B to which the label "no medical treatment" is given, the weight "1.0" is basically set for all the records. , The weight is changed to "0.5" for records (data) whose attendance classification includes annual leave or semi-absence. Therefore, since the shape of the graph structure representing the attendance record data is different from the case of the weight "1.0", the tensor data after tensorization is also different from the case of the weight "1.0". ..

By changing the weight in this way, it is possible to generate different tensor data even if the data is similar on the attendance record data. Therefore, even if the attendance record data that is the extraction source of the core tensor is similar data, different tensor data can be generated from each data, so that NN can be learned as a separate feature.

Specifically, as shown in FIG. 16, since the shape of each graph structure that is the input data changes before and after the weight change, each input tensor (tensor) that is generated from each and is the source of the core tensor. The shape of the data) also changes. At this time, since the principal component direction of the input tensor is in an uncertain state, the principal component directions may match before and after the weight change.

However, if the tensor decomposition that extracts the core tensor from the input tensor is performed after that, the input tensor that is the decomposition source is different before and after the weight change, so that different core tensors are generated. Here, since the principal component direction is determined in the core tensor indicating the feature amount of the input tensor, different core tensors having different principal component directions are extracted before and after the weight change. That is, even if the attendance record data is similar, different core tensors can be extracted by changing the weight. As a result, even if the attendance record data of the past person and after the medical treatment period is used as the learning data, the deterioration of the accuracy of the prediction model can be suppressed.

[Flow of learning process]
FIG. 17 is a flowchart showing the flow of the learning process. As shown in FIG. 17, the learning data generation unit 111 reads the attendance record data from the attendance record data DB 104 (S101) and selects one employee to be learned (S102).

Subsequently, the learning data generation unit 111 cuts out the data for the 6-month period from the attendance record data, and labels each of the cut out data depending on whether or not there is a medical treatment period in the attendance record data for the following 3 months. It is added to generate training data (S103). If medical treatment is included in the 6-month data, it will not be adopted as learning data.

After that, the learning data generation unit 111 selects one learning data (S104), refers to all the past attendance record data of the employees corresponding to the learning data, and determines whether or not the medical treatment period is included. (S105). Then, when the learning data generation unit 111 includes a medical treatment period (S105: Yes), the learning data label sets “with medical treatment” (S106). On the other hand, when the learning data generation unit 111 does not include the medical treatment period (S105: No), the learning data generation unit 111 sets "no medical treatment" in the label of the learning data (S107).

Then, when there is unprocessed learning data (S108: Yes), S104 and subsequent steps are repeated, and when there is no unprocessed learning data (S108: No), S109 and subsequent steps are executed.

Specifically, the past person determination unit 112 selects one learning data (S109) and determines whether or not it corresponds to "condition: past person and after the medical treatment period" (S110). Here, the weight setting unit 113 changes the weight to "0.5" (S111) according to the weighting rule in the case of the learning data corresponding to the condition (S110: Yes), and in the case of the learning data not satisfying the condition. (S110: No), the weight "1.0" is set as it is (S112).

Then, the tensor generation unit 114 executes tensorization of the weighted learning data to generate tensor data (S113). After that, when there is unprocessed learning data (S114: Yes), S109 and subsequent steps are repeated, and when there is no unprocessed learning data (S114: No), S115 and subsequent steps are executed.

Specifically, when an unprocessed employee to be learned remains (S115: Yes), S102 and subsequent steps are repeated. On the other hand, when the processing is completed for all the employees (S115: No), the learning process by the learning unit 115 using the learning data is executed (S116).

[Forecast flow]
FIG. 18 is a flowchart showing a processing flow at the time of prediction. As for the attendance record data to be predicted, the period of 6 months is input as in the case of learning, and it is predicted that medical treatment will be performed or not within 3 months thereafter. As shown in FIG. 18, the forecasting unit 116 reads attendance record data from the forecasting target DB 108 (S201) and selects one employee to be predicted (S202).

Subsequently, the prediction unit 116 refers to the attendance record data of the target employee and determines whether or not it corresponds to a past person (S203).

Here, when it does not correspond to a past person (S203: No), the prediction unit 116 executes tensorization of attendance record data (S204), executes prediction using a learned prediction model (S205), and corresponds. A label of a prediction result indicating that the employee will be treated or not treated within 3 months is stored in the storage unit 102 or the like (S206).

After that, when there is a next target employee (S207: Yes), S202 or later is executed, and when there is no next target employee (S207: No), the process ends. Further, in S203, when the prediction unit 116 determines that the person corresponds to a past person (S203: Yes), the prediction unit 116 stores a past person alert indicating that the person is a past person in the prediction result (S208). In addition, in S203, it corresponds to a past person except when the target employee is a past person whose attendance record data for the period before the forecast target attendance record data (for 6 months) includes medical treatment. , Medical treatment is included in the attendance record data (for 6 months) to be predicted, and it may not be predicted.

[effect]
As described above, when it is desired to learn the characteristics of a person who is in poor physical condition, the attendance record data of the person who is in poor physical condition and the attendance record data of a person who has been in poor physical condition after the medical treatment period are similar. Accuracy is reduced. Therefore, the learning device 100 according to the first embodiment can change the graph structure of the attendance record data by changing the weight of the attendance record data after the medical treatment period of the past person, and thus the core input to the NN. The tensor can be differentiated.

Specifically, the learning device 100 corresponds to learning data (label: no medical treatment) corresponding to a normal person, learning data corresponding to a poor physical condition (with medical treatment), a past person and before the medical treatment period according to a weighting rule. "1.0" is set as a weight for the learning data (label: with medical treatment) to be performed. On the other hand, the learning device 100 changes the weight to "0.5" for the learning data (label: with medical treatment) corresponding to the past person and after the medical treatment period.

By doing so, the learning device 100 clearly differentiates the feature amount (core tensor) of each learning data. As a result, the learning device 100 can effectively utilize the property of the deep tensor that "the partial structure of the graph (partial pattern of the tensor) that contributes to the prediction can be extracted as the core tensor", and the amount of training data is small. However, it is possible to make accurate predictions.

By the way, in the first embodiment, the attendance record data has been described as an example of the data structure, but the present invention is not limited to this, and for example, medical care data for medical use may be used. Therefore, in Example 2, a learning example using medical data will be described.

In the second embodiment, by learning the occurrence pattern of the symptom and the like, a prediction model for predicting whether the symptom worsens as the patient is hospitalized is constructed. Specifically, the learning device 100 learns a prediction model that predicts the presence or absence of hospitalization from medical data by using learning data to be learned as a case of being hospitalized and learning data as a case of not being hospitalized.

FIG. 19 is a diagram illustrating an example of learning data according to the second embodiment. As shown in FIG. 19, the learning data is composed of medical data every 6 months and a label indicating whether or not there is a hospitalization record within 3 months from the 6 months onward. FIG. 19A is medical data of an unwell person with a label (affected) in which the patient was hospitalized due to the plague. FIG. 19 (b) is attendance record data of ordinary persons labeled as not hospitalized (no illness).

As shown in FIG. 19, the learning device 100 according to the second embodiment has "6 months worth of medical care data and label (affected)" and "6 months worth of medical care data and label (not affected)". Is used as training data to train the prediction model. The learning device 100 predicts whether or not to be hospitalized within 3 months from the medical data for 6 months of a certain person after learning. The shaded area in FIG. 19 indicates fever, which is one of the elements of medical data.

Then, the learning device 100 also executes the weight change in the second embodiment as in the first embodiment. Specifically, since the characteristics of the medical data after the hospitalization period when the label "not affected" is set and the characteristics of the medical data of the poorly ill person who are set the label "affected" may be similar, learning. It leads to deterioration of accuracy. Therefore, also in Example 2, the effect on learning is obtained by changing the weight of the medical data of the recovered patient who has been hospitalized in the past but has already returned to the hospital after the hospitalization period to be smaller than usual. Reduce the degree.

Here, the hospitalization period and the setting of the weight will be described with reference to FIG. FIG. 20 is a diagram illustrating determination of hospitalization period and weight setting. First, as shown in FIG. 20, the medical care data is data having "patient No., date, day of the week, symptom, blood pressure, medication" as elements. The "patient No." is information that identifies the patient, the "date" is the date of examination, and the "day of the week" is the day of the week of examination. The "symptom" is a patient's symptom, such as fever, abdominal pain, chills, and normality (no abnormality). "Blood pressure" is a measurement of a patient's blood pressure. "Medication" is information indicating whether or not a patient has been administered a drug. The elements shown here are examples, and other elements such as the time of examination, the presence or absence of surgery, and the result of a blood test can be used.

At the time of tensorizing such learning data, the learning device 100 determines whether each learning data corresponds to before or after the "hospitalization period". In the example of FIG. 20, when the learning device 100 detects from "August 24, 2015" to "October 4, 2015" as the "hospitalization period", it starts from "August 24, 2015" when the hospitalization starts. The weight is set to "1" for the training data that corresponds to the previous one, and the weight is set to "0.5" for the training data after "October 4, 2015" after the end of hospitalization.

That is, the pre-hospital period before the hospitalization period is determined to be important for extracting as a core tensor which is a partial pattern that causes hospitalization, and "1" is set as the weight. On the other hand, it is determined that the post-hospital period after the hospitalization period is not important as a target to be extracted as a core tensor, which is a partial pattern that causes hospitalization, and a weight of "0.5" is set. In this way, a part of the weight of the tensor data is changed for the learning data corresponding to the end of the hospitalization period in which the label "not affected" is set.

Next, the tensor data will be described. FIG. 21 is a diagram illustrating the tensorization according to the second embodiment. The learning device 100 generates a tensor from the learning data with each attribute that is supposed to characterize the tendency to be hospitalized as each dimension. For example, the learning device 100 generates a four-dimensional fourth-order tensor using the four elements of month, date, symptom, and presence / absence of medication. In the case of data for 6 months, the number of monthly elements is "6", the maximum number of dates in each month is 31, the number of date elements is "31", and the type of symptom is fever. If it is normal or abdominal pain, the number of symptom elements is "3", and since there is no medication, the number of factors with or without medication is "2". Therefore, the tensor generated from the training data is a "6 x 31 x 3 x 2" tensor, and the value of the element corresponding to each month, the symptom on the date, and the presence or absence of medication in the training data is 1, and the value of the element not Becomes 0.

As shown in FIG. 21, the tensor generated by the learning device 100 has monthly data in the horizontal direction, date in the vertical direction, symptoms in the depth direction, medication on the left side, and no medication on the right side. The dates indicate the first day from the top, and the symptoms indicate normal, fever, and abdominal pain from the front. For example, FIG. 21 (a) shows the elements that were normal but had medication on the first day of the monthly examination, and FIG. 21 (b) shows the examination on the second day of the monthly month 1. Indicates an element with fever but no medication.

In Example 2, as in Example 1, the above-mentioned tensor will be simplified and described as shown in FIG. 21 (c). That is, the monthly, date, symptom, and the presence or absence of medication are expressed in a cube shape, the presence or absence of medication for each month and date is expressed separately, and the symptom for each month and date is expressed separately. To do. By doing so, the learning device 100 can learn a prediction model with high learning accuracy by using the same method as in the first embodiment.

By the way, although the examples of the present invention have been described so far, the present invention may be implemented in various different forms other than the above-mentioned examples.

[Learning]
The learning process described above can be executed any number of times. For example, it can be executed using all the training data, or can be executed a predetermined number of times. Further, as a method for calculating the classification error, a known calculation method such as the least squares method can be adopted, and a general calculation method used in NN can also be adopted. The learning data and attendance record data can also be acquired from an external device.

[Example of weight setting]
In the above embodiment, the weight "0.5" is set for the learning data corresponding to the past person and after the medical treatment period, and the weight "1.0" is set for other than that, but the weight is not limited to this. .. Specifically, it is also possible to adopt a method that can change the behavior of the vacation of a former person after the medical treatment period to a behavior different from that of a new person who is in poor physical condition.

For example, the learning device 100 uniformly changes the weight of the attendance classification to a value smaller than 1 without distinguishing the vacation type. For example, the learning device 100 sets the average number of vacation days per month for the learning period (6 months) before the weight change to Da, and one month for a period sufficiently longer than the learning period (for example, one year). Let the average number of vacation days be Db. Then, the learning device 100 gives Db / Da as a correction value of the weight for the learning data corresponding to the past person and after the medical treatment period. By doing so, it is usually expected that Da> Db, so that the vacation days pattern of the learning data corresponding to the past person and after the medical treatment period can be brought closer to the original vacation days pattern of the past person.

Further, the learning device 100 calculates the average number of vacation days Dc per month for a normal person, and gives Dc / Da as a weight correction value for the learning data corresponding to the person who has a history and after the medical treatment period. By doing so, the behavior of the past person can be made closer to the behavior of the normal person. Further, the learning device 100 is less affected by the vacation situation, and in order to surely distinguish it from the new leave, the weight of the learning data corresponding to the past person and after the medical treatment period is as extreme as 0.1. You can also set a small value for. Further, the learning device 100 can change the weight of the quasi-absence to a value smaller than 0.5 among the vacation types such as annual leave and quasi-absence.

In addition, the learning device 100 can also replace the attendance record data of a past person and the corresponding attendance record data after the medical treatment period with the attendance record data of a normal person and generate it as learning data. That is, the learning device 100 can also replace the learning data corresponding to the past person and after the medical treatment period with the learning data of the normal person.

[Weight change target]
In the first embodiment, an example of changing the weight of the learning data corresponding to the past person and after the medical treatment period has been described, but the present invention is not limited to this, and for example, the learning data corresponding to the past person is related to before and after the medical treatment period. Instead, "0.5" can be set as the weight. Further, in the first embodiment, among the elements of the attendance record data, the weight of the attendance classification is changed to 0.5, and the weight of the other elements (monthly, etc.) is set to 1.0. The element that changes the weight is not limited to this, and can be changed arbitrarily.

Further, in the first embodiment, an example of changing the weight of the tensor data of a person who has a medical treatment experience in the past has been described, but the present invention is not limited to this. For example, data on employees who have childcare or long-term care and who have work disruptions that are not caused by their physical condition can also be noise in new medical treatment predictions, so these employees are targeted. The conditions for changing the weight of the tensor data can be changed arbitrarily.

Further, in the above embodiment, an example of learning by automatically extracting a period of 6 months or 3 months from the attendance record data at the time of learning and assigning labels of "with medical treatment" and "without medical treatment" has been described. It is not limited to this. For example, when the period extraction or the label is given by the user in advance, the given data or the given label can be used.

In addition, an example of setting a label is also an example, and various labels that can distinguish the presence or absence of a person with poor physical condition are used, such as a person with poor physical condition and a normal person, a person with leave and no leave, etc. You can also do it. Further, the values described in each embodiment are also examples, and can be arbitrarily changed.

Further, although the attendance record data of 6 months was used as the data used for the prediction, it is not limited to this and can be changed arbitrarily such as 4 months. In addition, an example of assigning a label to the 6-month attendance record data depending on whether or not the patient has been treated within the next 3 months has been described, but the present invention is not limited to this, and any option such as within 2 months Can be changed to. Further, in the above embodiment, the example in which the learning device 100 converts the attendance record data into a tensor has been described, but it is also possible for another device to acquire the tensorized data and execute each of the above processes. In the embodiment, the tensor is actually the 4th floor tensor, but the tensor on the 3rd floor is used for simplification of the explanation, but this is just an example and can be changed arbitrarily. Moreover, the number of dimensions can be changed arbitrarily.

[Assumed system]
In the above embodiment, the example of predicting the employees who may learn the attendance record data and receive medical treatment has been described, but the present invention is not limited to this. For example, it can be applied to failure prediction using electronic component operation data, attack prediction using communication data, traffic congestion prediction using road traffic data, and the like.

[neural network]
In this embodiment, various neural networks such as RNN (Recurrent Neural Networks) and CNN (Convolutional Neural Network) can be used. Further, as a learning method, various known methods other than error back propagation can be adopted. Further, the neural network has a multi-stage structure composed of, for example, an input layer, an intermediate layer (hidden layer), and an output layer, and each layer has a structure in which a plurality of nodes are connected by edges. Each layer has a function called "activation function", the edge has "weight", and the value of each node is the value of the node of the previous layer, the value of the weight of the connecting edge (weight coefficient), and the layer has. Calculated from the activation function. As the calculation method, various known methods can be adopted.

In addition, learning in a neural network is to modify parameters, that is, weights and biases, so that the output layer has correct values. In the back-propagation method, the neural network is defined with a "loss function" that indicates how far the value of the output layer is from the correct state (desired state), and the steepest descent method, etc. Is used to update the weights and biases so that the loss function is minimized.

[system]
Information including processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified. In addition, the specific examples, distributions, numerical values, etc. described in the examples are merely examples and can be arbitrarily changed.

Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific forms of distribution and integration of each device are not limited to those shown in the figure. That is, all or a part thereof can be functionally or physically distributed / integrated in any unit according to various loads, usage conditions, and the like. Further, each processing function performed by each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.

[hardware]
FIG. 22 is a diagram illustrating a hardware configuration example. As shown in FIG. 22, the learning device 100 includes a communication device 100a, an HDD (Hard Disk Drive) 100b, a memory 100c, and a processor 100d. Further, the parts shown in FIG. 22 are connected to each other by a bus or the like.

The communication device 100a is a network interface card or the like, and communicates with another server. The HDD 100b stores a program or DB that operates the function shown in FIG.

The processor 100d reads a program that executes the same processing as each processing unit shown in FIG. 6 from the HDD 10b or the like and expands the program into the memory 10c to operate a process that executes each function described in FIG. 6 or the like. That is, this process executes the same function as each processing unit of the learning device 100. Specifically, the processor 100d outputs a program having the same functions as the learning data generation unit 111, the history determination unit 112, the weight setting unit 113, the tensor generation unit 114, the learning unit 115, the prediction unit 116, etc. from the HDD 100b or the like. read out. Then, the processor 100d executes a process of executing the same processing as the learning data generation unit 111, the history determination unit 112, the weight setting unit 113, the tensor generation unit 114, the learning unit 115, the prediction unit 116, and the like.

In this way, the learning device 100 operates as an information processing device that executes the learning method by reading and executing the program. Further, the learning device 100 can realize the same function as that of the above-described embodiment by reading the program from the recording medium by the medium reading device and executing the read program. The program referred to in the other embodiment is not limited to being executed by the learning device 100. For example, the present invention can be similarly applied when another computer or server executes a program, or when they execute a program in cooperation with each other.

This program can be distributed over networks such as the Internet. In addition, this program is recorded on a computer-readable recording medium such as a hard disk, flexible disk (FD), CD-ROM, MO (Magneto-Optical disk), or DVD (Digital Versatile Disc), and is recorded from the recording medium by the computer. It can be executed by being read.

100 Learning device 101 Communication unit 102 Storage unit 103 Weight information DB
104 Attendance record data DB
105 Learning data DB
106 Tensor DB
107 Learning result DB
108 Prediction target DB
110 Control unit 111 Learning data generation unit 112 Past person determination unit 113 Weight setting unit 114 Tensor generation unit 115 Learning unit 116 Prediction unit

Claims (6)

Patient medical data consisting of multiple records with each element being a symptom indicating at least the date, month, and the content of the patient's symptom.
Each predetermined period of the medical care data is set as one learning data which is an explanatory variable of supervised learning, and hospitalization is set for the symptom within a specified period from the predetermined period set for each learning data. When there is a hospitalization period, it indicates a person who is in poor physical condition who has been hospitalized, and the patient is hospitalized when there is no hospitalization period for which hospitalization is set for the symptom within the specified period from the predetermined period. The morbidity information used as the correct answer information which is the objective variable of the supervised learning indicating a normal person who does not have the above, and the morbidity information used for the processing of the supervised learning based on each learning data by the learning device and the medical treatment. Weight information that indicates the weight that changes the shape of the tensor data generated from the data ,
A medical data structure that includes
A storage unit that stores a target core tensor, a weighting rule for a recovery patient who has been hospitalized but has been discharged in the past, a data structure for medical use, and the target core tensor stored in the storage unit. A learning device having a generation unit that generates tensor data from each training data based on the weighting rule and the medical data structure.
When the first patient corresponding to the recovery patient, wherein according to weight information and said weighting rules for training data after past hospitalization among the learning data corresponding to the first patient, the past The weight of any element of the training data after the hospitalization period is changed to generate the first tensor data, and the first core tensor is generated from the first tensor data so as to be similar to the target core tensor. and, by entering the first core tensor to the learning device for performing learning using the tensor data, a first that is set for the learning data after output and the previous hospitalization period by the learning device A medical data structure used in a process for causing the learner to perform supervised learning using a difference from morbidity information. The learning device, when the first patient corresponding to the recovery patient, according to the weight information and the weighting rules, each of the first dimension of the element of the date of medical data, monthly and symptoms The medical data structure according to claim 1, wherein the weights of the elements of the second dimension different from the first dimension are left as they are. It is medical care data of a patient composed of graph data composed of a plurality of nodes and an edge connecting the plurality of nodes.
The plurality of nodes consist of date, month and symptom nodes.
There are as many nodes for the date, month, and symptom as there are dates, months, and symptoms, respectively.
Each node stores values according to the date, month, and symptom status.
The edge connects the date node, the monthly node, and the relevant node of the symptom nodes.
When the learning device changes the weight of each of the elements of the first dimension of the date, month, and symptom of the medical data in the case where the first patient corresponds to the recovery patient, the said. Tensor data showing graph data in which the values stored in the node corresponding to the first dimension are changed based on the weight is generated, and the first tensor data is converted from the first tensor data so as to be similar to the target core tensor. The core tensor of the above is generated, and the first core tensor is input to the learner.
When the first patient does not correspond to the recovery patient, tensor data showing graph data in a state where the value stored in the node is not changed is generated, and the tensor data is used to resemble the target core tensor. Generate 2 core tensors, input the 2nd core tensor into the learner,
The learner
When the first core tensor is input, the relationship between the node and the edge showing the characteristics of the recovery patient is learned, and the relationship is learned.
When the second core tensor is input, the relationship between the node and the edge, which is characteristic of a normal person who has never been hospitalized, is learned.
The medical data structure according to claim 2. According to claim 1, the weighting rule is information for changing the value of a node or the length of an edge of the graph structure so that the shape of the graph structure of each medical data changes before and after the weight change. Described medical data structure. The learning device is included in the first tensor data generated from the learning data based on the medical data of the recovery patient, which is similar to the third tensor data generated from the learning data based on the medical data of the poor physical condition. By changing the weight of any of the elements, the direction of the principal component of the core tensor generated from the third tensor data by the tensor decomposition and the first tensor data generated from the first tensor data. The medical use according to claim 1, wherein the learning device is made to learn each of the characteristic amount of the poorly ill person and the characteristic amount of the recovery patient by differentiating from the main component direction of the core tensor. Data structure. The learning device
By the admission to the affected information is set, when the first patient corresponding to the recovery patient, in accordance with the weighting rule and the weight information, generated from the clinical data of the first patient By changing the weight of any element of the training data after the past hospitalization period with respect to the training data after the past hospitalization period among the trained data, the training data before the weight is changed. The shape of the original tensor data generated from is changed to the first tensor data, and the first core tensor is generated from the first tensor data so as to be similar to the target core tensor. The first core tensor is input to the learner, and supervised learning is executed using the difference between the output result of the learner and the morbidity information set for the learning data after the past hospitalization period. And
By the said admission to the affected information is not set, generated from the case of the first patient does not correspond to the recovery patient, according to the weight information and the weighting rules, the first patient of the medical data Each of the trained data is changed to each tensor data without changing the weight of any element of the training data after the past hospitalization period, and each core from the tensor data so as to be similar to the target core tensor. A process of generating a tensor, inputting each core tensor into the learning device, and executing supervised learning using the difference between the output result of the learning device and the morbidity information set for each learning data. The medical data structure according to claim 1, which is used in the above.