JP7656499B2 - Learning model generation method, program, information processing method, information processing device, and learning model data generation method - Google Patents

Description

The present invention relates to a method for generating a learning model, a program, an information processing method, an information processing device, and a method for generating learning model data.

In hospital laboratories, testing centers, and other testing facilities, a large number of clinical tests are constantly being carried out, and clinicians use the test results to make diagnostic and treatment decisions regarding the patient's condition. Before deciding on a treatment plan based on the test results, the validity of the test results is verified, and if any values are determined to be abnormal, retests are conducted to increase the reliability of the test.

A delta verification method has been disclosed that increases the reliability of tests by setting a standard value for the difference between the previous value and the current value for each test item, and extracting abnormal values through data processing that compares the values of samples from individual patients with the previous values (see, for example, Patent Documents 1 and 2).

Japanese Patent Application Publication No. 5-151282 Japanese Patent Application Publication No. 11-296605

However, in the technologies described in Patent Documents 1 and 2, whether or not there is an abnormality is determined based on the difference in test values in the test results of a single item, and since the fluctuations in other related test values, the test intervals, and the chronological order of the tests are not taken into consideration, there are concerns that the detection accuracy may decrease.

The objective of this disclosure is to provide a method for generating a learning model that can improve the accuracy of detecting abnormalities in test values.

A method for generating a learning model according to one aspect of the present disclosure causes a computer to execute a process of acquiring a plurality of specimen data in a clinical test, generating a first specimen data group including abnormal specimen data based on the acquired plurality of specimen data, and generating a learning model that outputs classification information when the specimen data group is input based on training data including the generated first specimen data group, the second specimen data group not including the abnormal specimen data, and classification information regarding normality or abnormality of the specimen data group.

A program according to one aspect of the present disclosure acquires a specimen data group consisting of a plurality of specimen data arranged in chronological order, and causes a computer to execute a process of inputting the acquired specimen data group into a learning model that has been trained to output classification information regarding normality or abnormality of the last specimen data in chronological order included in the specimen data group when the specimen data group consisting of a plurality of specimen data arranged in chronological order is input, and outputting the classification information.

A method for generating learning model data according to one aspect of the present disclosure includes acquiring multiple pieces of specimen data in a clinical test, generating a first specimen data group including abnormal specimen data based on the acquired multiple pieces of specimen data, and causing a computer to execute a process of associating the generated first specimen data group and a label indicating abnormality with a second specimen data group not including the abnormal specimen data and a label indicating normality as training data for a learning model, and storing the association in a storage unit.

This disclosure can improve the accuracy of detecting abnormalities in test values.

1 is a block diagram illustrating an example of the configuration of an information processing system. FIG. 13 is a diagram showing an example of the contents of a specimen DB. FIG. 2 is a diagram showing an example of the contents of a training DB. FIG. 2 is an explanatory diagram showing an overview of a first learning model. 13 is a flowchart illustrating an example of a procedure for generating training data. 13 is a flowchart illustrating an example of a process for generating a first learning model. 13 is a flowchart illustrating an example of a procedure for detecting an abnormality. FIG. 2 is a schematic diagram showing an example of a screen displayed on a display device. FIG. 2 is a schematic diagram showing an example of a screen displayed on a display device. FIG. 2 is a schematic diagram showing an example of a screen displayed on a display device. A flowchart showing an example of a generation processing procedure of a first learning model in the second embodiment. An explanatory diagram showing an overview of a second learning model in the third embodiment. A flowchart showing an example of a learning process procedure for a first learning model in the third embodiment. An explanatory diagram showing an overview of the third learning model in the fourth embodiment. 23 is a flowchart showing an example of a disease information acquisition process procedure according to the fourth embodiment.

This disclosure will be specifically described with reference to drawings showing embodiments thereof.

First Embodiment
1 is a block diagram showing a configuration example of an information processing system. The information processing system includes an information processing device 1 and a testing facility device 2. The information processing device 1 and the testing facility device 2 are communicatively connected via a network N such as a LAN (Local Area Network) or the Internet.

The testing facility device 2 is installed in a testing facility such as a hospital testing room or testing center. The testing facility device 2 is, for example, a server computer, a personal computer, etc., and is connected to a display device 3, an input device 4, and a testing device 5, etc. The display device 3 is, for example, a liquid crystal display or an organic EL (Electro Luminescence) display, etc. The input device 4 is an operation unit that accepts user operations, and is, for example, a pointing device such as a keyboard or a touch panel. The testing device 5 performs various testing (analysis) processes on the specimen and generates clinical test results (specimen data). The specimen data obtained by the testing device 5 is transmitted to the information processing device 1 via the testing facility device 2.

The information processing device 1 is, for example, a server computer, a personal computer, etc., and is capable of various information processing and sending and receiving information. The information processing device 1 acquires specimen data obtained by the testing device 5 via the testing facility device 2. The information processing device 1 acquires classification information regarding normality or abnormality of the acquired specimen data. The information processing device 1 functions as an abnormality detection device for the specimen data. The information processing device 1 displays the generated classification information on the display device 3. Note that the information processing device 1 may be a local server installed in the same facility as the testing facility device 2, or may be a cloud server communicatively connected to the testing facility device 2 via the Internet, etc.

Classification information is information that indicates whether the specimen data obtained by a clinical test is normal or abnormal (the presence or absence of abnormality). The specimen data may include multiple test values corresponding to multiple types of test items. Classification information includes normal and abnormal, and abnormalities are further classified into multiple types depending on the cause, such as sample mix-up and dilution error. Here, normal means that the test value of the specimen data is not an abnormal value caused during the test, i.e., the clinical test result is appropriate. Abnormal means that the test value of the specimen data is an abnormal value caused by the testing process (test error), i.e., the clinical test result is not appropriate.

The person in charge at the testing facility can determine whether there is an abnormality in the sample data and the cause of the abnormality based on the classification information displayed on the display device 3. If the sample data is normal, it is reported to the clinic as correct sample data, and if the sample data is abnormal, retesting is performed in a manner appropriate to the cause of the abnormality, allowing for efficient and appropriate testing.

Test results in clinical testing vary due to the influence of various factors, including not only the physiological state of the patient, but also the test conditions mentioned above, human error during the test, and abnormalities in the test reagents and testing equipment. Data handled daily in clinical testing generally yields test values that are closely related over time if they are from the same patient. Furthermore, the range of possible test values related to the human body is limited, although to varying degrees, and each test value is considered to be highly related to each other. Therefore, it is believed that there is a correlation between differences in test values from the same test and between multiple test items. Conventionally, the validity of test results has been determined using statistical methods such as delta validation.

However, because it is difficult to distinguish using statistical inference alone whether an abnormality in a test value is due to the patient's pathology, such as a panic value, or due to the testing process, such as a test error, a retest is conducted when a test value is determined to be abnormal to determine whether there is an abnormality in the testing process. Conducting retests frequently can prevent abnormalities in the testing process from being overlooked, but on the other hand, there is the problem that the retesting requires additional costs and time. This information processing system achieves efficient testing by providing classification information indicating the cause of the abnormality using a first learning model 113 described below.

The information processing device 1 includes a control unit 10, a storage unit 11, and a communication unit 12. The information processing device 1 may be a multi-computer consisting of multiple computers, or may be a virtual machine virtually constructed by software.

The control unit 10 is a processor that uses one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), etc. The control unit 10 uses built-in memory such as ROM (Read Only Memory) or RAM (Random Access Memory) to control each component and execute processing.

The storage unit 11 is a non-volatile storage device such as a hard disk or SSD (Solid State Drive). The storage unit 11 stores programs and data referenced by the control unit 10, including the program (program product) 1P. The control unit 10 reads and executes the program 1P, causing a general-purpose server computer to function as an information processing device specific to the present disclosure. The storage unit 11 may be composed of multiple storage devices, or may be an external storage device connected to the information processing device 1.

The program 1P stored in the memory unit 11 may be recorded in a computer-readable manner on a recording medium. The memory unit 11 stores the program 1P read from the recording medium 1A by a reading device (not shown). The program 1P may also be downloaded from an external computer (not shown) connected to a communication network (not shown) and stored in the memory unit 11. The program 1P may be deployed so as to be executed on a single computer, or on one site, or distributed across multiple sites and on multiple computers interconnected by a communication network.

The memory unit 11 also stores a specimen DB (Data Base) 111, a training DB 112, and a plurality of first learning models 113. The specimen DB 111 is a database that stores specimen data related to specimens in clinical tests. The training DB 112 is a database that stores training data (learning data) used to learn the first learning model 113. The first learning model 113 is a machine learning model that has learned the training data. It is assumed that the first learning model 113 will be used as a program module that constitutes artificial intelligence software. The memory unit 11 may further store a second learning model 114 and a third learning model 115. The second learning model 114 and the third learning model 115 will be described in detail in other embodiments.

The communication unit 12 is a communication module for performing communication-related processing. The control unit 10 can send and receive information to and from the testing facility device 2 via the communication unit 12.

In this embodiment, the information processing device 1 is not limited to the above configuration, and may include, for example, a display unit that displays images, an input unit that accepts operational input, etc.

The testing facility equipment 2 includes a control unit 20, a memory unit 21, a communication unit 22, and an input/output unit 23.
The control unit 20 is a processor using one or more CPUs, GPUs, etc. The control unit 20 uses built-in memories such as ROM and RAM to control each component and execute processes.

The memory unit 21 is a non-volatile storage device such as a hard disk or SSD. The memory unit 21 stores programs and data referenced by the control unit 20. The communication unit 22 is a communication module for performing communication-related processing. The control unit 20 can send and receive information to and from the information processing device 1 via the communication unit 22.

The input/output unit 23 is an input/output I/F (interface) for connecting an external device. The above-mentioned display device 3, input device 4, and testing device 5 are connected to the input/output unit 23. The control unit 20 causes the display device 3 to display classification information regarding whether the specimen data is normal or abnormal via the input/output unit 23. The control unit 20 accepts information input to the input device 4 via the input/output unit 23. The control unit 20 accepts specimen data output from the testing device 5 via the input/output unit 23.

2 is a diagram showing an example of the contents of the specimen DB 111. The specimen data DB 111 stores, for example, a specimen ID, a testing facility name, patient attribute information, a test date and time, and a test value for each of a plurality of specimen data in association with each other. The specimen ID is identification information for identifying a specimen. The testing facility name is information for identifying the testing facility that performed the test on the specimen. The patient attribute information is attribute information of a patient (subject) related to the specimen, and includes, for example, a patient ID, a patient name, an age, and a sex. Each patient can be uniquely identified, for example, by a patient ID and a patient name. The test date and time is the date and time when the test on the specimen was performed, or the date and time when blood was drawn for a blood test. The test value is a test value related to the specimen. The test value column stores each test value in association with each of a plurality of test items. The specimen data stored in the specimen DB 111 may further include information such as the type of department in charge of the testing facility (type of medical department in a hospital), a test equipment ID, a lot number of a test reagent, a date of receipt of the specimen data, a test method, and a name of a person in charge of the test.

Each time the information processing device 1 receives sample data from the testing facility device 2, it stores the received sample data in the sample DB 111. The sample DB 111 may store sample data in chronological order for each patient. The contents stored in the sample DB 111 are not limited to the example shown in FIG. 2.

Figure 3 is a diagram showing an example of the contents of the training DB 112. The training DB 112 stores training data used for learning the first learning model 113 described below. The training DB 112 stores, for example, specimen data group information and classification information related to the specimen data group in association with a training data ID that identifies the training data.

The specimen data group information string contains items that are the same as or a part of the specimen data stored in specimen DB 111, such as the test date and time of the specimen data and test values for multiple test items. The specimen data group information contains information on multiple specimen data that make up one specimen data group. A specimen data group is composed of specimen data in chronological order that originates from the same patient. Specifically, a specimen data group is composed of a predetermined number of specimen data that are adjacent in chronological order among the specimen data obtained for the same patient. Figure 3 shows an example in which one specimen data group is composed of four specimen data, but the number of specimen data that make up a specimen data group is not limited.

As will be described in more detail later, the specimen data groups are divided into specimen data groups that do not contain abnormal specimen data resulting from testing errors (abnormal specimen data) and specimen data groups that contain such abnormal specimen data. In the example shown in FIG. 3, training data ID001-1 and training data ID001-2 are specimen data groups that do not contain abnormal specimen data, and training data ID001-3 is a specimen data group that includes abnormal specimen data. Training data ID001-3 is obtained by replacing all test values in the specimen data in the fourth row of training data ID001-2 with abnormal values.

The classification information column stores classification information regarding whether the sample data group is normal or abnormal. Classification information includes, for example, normal and abnormal. Abnormalities resulting from testing errors are classified into multiple types depending on the cause, such as mix-up, dilution error, misidentification of units, reagent error, and sample clogging in the testing equipment.

A mix-up is an abnormality in a test value caused by a mix-up of samples (mix-up of all test values related to sample data) between different patients who underwent testing on the same reception date. Mix-ups may include mix-ups of test values for a specific test item between different patients who underwent testing on the same reception date. A dilution error is an abnormality caused by an error in the dilution ratio or dilution amount when diluting a sample with saline, buffer solution, reagent, etc. A unit misidentification is an abnormality caused by a misidentification of a unit related to a test value. A reagent error is an abnormality caused by an error in the use of a reagent in a test using a reagent, or a shortage or excess of a reagent. A specimen clogging is an abnormality caused by a specimen clogging inside a tube or the like in a testing device used for the test. The classification information column may store normal or abnormal (presence or absence of abnormality) for each type of abnormality. The contents stored in the training DB 112 are not limited to the example shown in FIG. 3.

Figure 4 is an explanatory diagram showing an overview of the first learning model 113. The first learning model 113 is a machine learning model that takes as input a sample data group consisting of multiple sample data in chronological order, and outputs classification information regarding the normality or abnormality of the last (latest, i.e., most recent) sample data in the chronological order contained in the sample data group. Specifically, the first learning model 113 takes as input a sample data group, and outputs a classification result of normality or abnormality (presence or absence of abnormality).

Multiple first learning models 113 are stored for each type of abnormality. That is, a first first learning model 113 that receives a group of sample data as input and outputs whether or not there is a mix-up, a second first learning model 113 that outputs whether or not there is a dilution error, a third first learning model 113 that outputs whether or not there is a unit misidentification, a fourth first learning model 113 that outputs whether or not there is a reagent mistake, and a fifth first learning model 113 that outputs whether or not there is a sample clogging are respectively stored. Since each first learning model 113 has the same configuration, the configuration of the first learning model 113 that outputs whether or not there is an abnormality caused by a mix-up will be described below.

The information processing device 1 performs machine learning to learn predetermined training data to generate a first learning model 113 in advance. The information processing device 1 then inputs a sample data group acquired from the testing device 5 to the first learning model 113 and outputs classification information for the sample data group.

The first learning model 113 is constructed, for example, by XGBoost (eXtreme Gradient Boosting). XGBoost is an ensemble learning method that combines gradient boosting and random forest. The first learning model 113 is configured to perform boosting by constructing multiple decision trees (weak learners) and constructing a new decision tree using information from the previous decision tree. Specifically, a new decision tree is constructed with an error (gradient of the loss function) that could not be predicted by the previous decision tree as the objective variable. In each decision tree, the input data is classified according to conditions on the way from the root to the branches, and when it reaches the terminal leaf node, the value given to the terminal leaf node is output as the predicted value.

The input of the first learning model 113 is a sample data group. The sample data group includes, for example, the current, previous, previous-before, and previous-before sample data related to the same patient in the same testing facility. The sample data includes the test date and time (the interval from the previous test date and time) and test values corresponding to multiple test items. The multiple sample data are input to the first learning model 113 as sample data group data arranged in chronological order. Note that the number of sample data included in the sample data group is not limited to four. The sample data group may include sample data to be determined for the presence or absence of abnormality (the last sample data in the chronological order) and at least one sample data acquired earlier than the sample data to be determined. Note that in general, in clinical settings, it is not past data but the most recent data that needs to be checked first as a test error or pathological abnormality, so in this method, it is assumed that the past data is correct, and whether the most recent (latest in the chronological order) data is correct is verified.

The output of the first learning model 113 is the classification result of the last sample data in the time series contained in the sample data group as normal or abnormal. When a sample data group is input, the first learning model 113 outputs the accuracy for each of the normal and abnormal classes. The first learning model 113 can output the class whose accuracy is equal to or greater than a threshold value.

The first learning model 113 may be a multi-class classification model, and may classify normality or multiple types of abnormality using a single first learning model 113. In this case, the first learning model 113 receives as input a sample data group consisting of multiple sample data in chronological order, and outputs classification information indicating whether the last sample data in the chronological order contained in the sample data group is normal or multiple types of abnormality.

Although an example in which the first learning model 113 is XGBoost has been described above, the configuration of the first learning model 113 is not limited as long as it is capable of identifying classification information regarding normality or abnormality of the last sample data in the time series contained in the sample data group. The first learning model 113 may be, for example, a neural network such as a Transformer, a Convolution Neural Network (CNN), a Recurrent Neural Network (RNN), or a Long Short Term Memory (LSTM), or may use other learning algorithms such as a support vector machine, a logistics regression, or a random forest.

In the learning phase, which is a stage preceding the operational phase in which anomaly detection is performed, the information processing device 1 generates a first learning model 113 using training data and stores the generated first learning model 113 in the storage unit 11. Then, in the operational phase, the stored first learning model 113 is used to detect anomalies in the sample data.

Figure 5 is a flowchart showing an example of a training data generation process. The following process is executed by the control unit 10 in accordance with the program 1P stored in the storage unit 11 of the information processing device 1.

The control unit 10 of the information processing device 1 acquires multiple pieces of chronologically ordered specimen data relating to the same patient based on the information stored in the specimen DB 111 (step S11). The control unit 10 extracts specimen data relating to the same testing facility and the same patient from the large amount of specimen data stored in the specimen DB 111, for example, and arranges them in chronological order.

The control unit 10 extracts specimen data including test values of test items with high co-occurrence frequencies from the acquired specimen data (step S12). Test items that are frequently performed at the same time, i.e., test items with high co-occurrence frequencies, can be extracted by using, for example, association analysis, market basket analysis, and co-occurrence cluster mining (see Kenichi Fukui et al., Artificial Intelligence, Vol. 30, No. 2, 2015). On the other hand, there are several methods for easily extracting test items with high co-occurrence frequencies, and they can be extracted, for example, as follows. First, a frequency table of test items that co-occur with a certain test item is created, and a brute-force frequency table between the two items is created for all test items. Next, in the frequency table, the co-occurrence frequency including all items in the set is confirmed for a test item set limited to the top test item, and test items with a co-occurrence frequency above a certain threshold are narrowed down to extract a test item group (multiple specimen data) with high co-occurrence frequencies. This makes it possible to generate a specimen data group with few missing values.

The control unit 10 generates a sample data group that does not include abnormal sample data based on the multiple chronologically ordered sample data with high co-occurrence frequency obtained (step S13). In detail, the control unit 10 generates a first sample data group consisting of four sample data including the most recent sample data from the multiple chronologically ordered sample data obtained. Similarly, the control unit 10 slides the reference sample data in order to generate multiple sample data groups, such as a second sample data group consisting of four sample data including the second most recent sample data in chronological order, .... In this way, a sample data group that does not include abnormal sample data is generated using the test values obtained in the clinical test as they are. The control unit 10 may generate sample data by extracting only predetermined management items necessary for input elements to the first learning model 113.

The control unit 10 generates a specimen data group including abnormal specimen data by replacing a portion of the acquired specimen data with an abnormal value (step S14). In detail, the control unit 10 generates abnormal specimen data by replacing a portion of the last specimen data in the chronological order among the specimen data included in the specimen data group with an abnormal value. The control unit 10 generates a specimen data group consisting of the generated abnormal specimen data and the remaining three specimen data that use the test values obtained in the clinical test as is.

A specific example of a method for generating abnormal sample data is shown below. The control unit 10 generates abnormal sample data resulting from a mix-up by replacing all test values included in the sample data for a first patient at a specified test date and time with test values for sample data for a second patient, different from the first patient, who was tested at the specified test date and time. The control unit 10 may replace only test values for a specific test item included in the sample data for the first patient at a specified test date and time with test values for the second patient.

The control unit 10 generates abnormal specimen data resulting from a dilution error (calculation error) by replacing the test values for one or more test items included in the specimen data with test values calculated on the assumption that dilution was performed at a dilution ratio different from the specified one. The control unit 10 generates abnormal specimen data resulting from a unit misidentification (calculation error) by replacing the test values for one or more test items included in the specimen data with test values calculated on the assumption that a unit was misidentified.

The control unit 10 generates abnormal specimen data caused by an incorrect reagent by replacing the test values for one or more test items included in the specimen data with test values calculated under the assumption that a reagent other than the specified one was used. The control unit 10 generates abnormal specimen data caused by specimen clogging by replacing the test values for one or more test items included in the specimen data with test values calculated under the assumption that a specimen clogging has occurred (e.g., zero, a fraction of the actual value, or a specific maximum value, etc.). In this way, the control unit 10 generates abnormal specimen data using actually obtained specimen data. Note that the method and type of abnormal specimen data generation are not limited to the above examples.

The control unit 10 generates training data, which is a data set in which classification information is labeled as a correct answer value, for each of the generated specimen data groups that do not include abnormal specimen data and the generated specimen data groups that include abnormal specimen data (step S15). Specifically, the control unit 10 generates training data in which a label indicating normality or abnormality is assigned to each specimen data group for each type of abnormality. The training data may be data in which a label indicating normality or multiple types of abnormalities is assigned to each specimen data group.

The control unit 10 stores the generated training data in the training DB 112 (step S16), and ends the series of processes. The control unit 10 generates multiple information groups based on the large amount of specimen data collected in the specimen DB 111, and accumulates the generated information groups in the training DB 112 as training data. Usually, when specimen data is determined to be abnormal in a clinical test, only normal specimen data after retesting is often ultimately recorded in the database, making it difficult to collect abnormal specimen data. According to the above-mentioned process, abnormal specimen data is virtually generated, and training data can be accumulated efficiently. Note that, in order to objectively evaluate the accuracy of anomaly detection, it is preferable that the abnormal specimen data is generated taking into account the frequency of occurrence of abnormalities in actual clinical tests.

In the above-mentioned processing, the specimen data may include, for example, the name of the testing facility, patient attribute information, a testing device ID, etc. In other words, the input elements to the first learning model 113 may include the name of the testing facility, patient attribute information, a testing device ID, etc. In this case, when generating a specimen data group, it is possible to eliminate the need for processing to classify specimen data for the same testing facility or the same patient.

Figure 6 is a flowchart showing an example of a process for generating the first learning model 113. The following process is executed by the control unit 10 according to the program 1P stored in the storage unit 11 of the information processing device 1, for example, after the process of Figure 5 is completed.

The control unit 10 of the information processing device 1 acquires a set of training data extracted from the information group based on the information stored in the training DB 112 (step S21). The control unit 10 uses the acquired training data to generate a first learning model 113 that outputs information indicating whether the last specimen data in the time series included in the specimen data group is normal or abnormal when a specimen data group is input (step S22).

Specifically, the control unit 10 uses a group of sample data included in the training data to construct a first decision tree. The control unit 10 fits the input data into the first decision tree and obtains a first residual between the input data and the predicted value included in the leaf node. The control unit 10 then uses the obtained first residual to construct a second decision tree by a similar procedure and obtains a second residual between the first residual and the predicted value included in the leaf node. The second decision tree is constructed by combining the first and second decision trees to achieve higher accuracy. The control unit 10 subsequently constructs decision trees sequentially by a similar procedure.

The control unit 10 adjusts parameters using, for example, gradient descent to optimize (minimize) the loss function in the first learning model 113, and sequentially learns and integrates decision trees. Note that, in order to prevent so-called overfitting, when optimizing the loss function, the control unit 10 does not use all explanatory variables, but instead selects and uses a randomly determined number of explanatory variables. The control unit 10 completes learning when the loss function satisfies a predetermined criterion.

The control unit 10 stores definition information related to the trained first learning model 113 in the storage unit 11 as the trained first learning model 113 (step S23), and ends the series of processes. Through the above-mentioned processes, the control unit 10 can construct a trained first learning model 113 capable of appropriately outputting classification information related to normality or abnormality for a group of sample data. The control unit 10 executes the above-mentioned processes for each first learning model 113.

Using the first learning model 113 generated as described above, the information processing device 1 detects anomalies in the sample data. Below, the processing procedure executed by the information processing device 1 in the operation phase is described.

Figure 7 is a flowchart showing an example of a processing procedure for detecting an anomaly. The following processing is executed by the control unit 10 in accordance with the program 1P stored in the memory unit 11 of the information processing device 1. The control unit 10 may perform the following processing in real time, for example, each time sample data is transmitted from the testing facility device 2, or may perform the processing at any timing based on the collected sample data.

The control unit 10 of the information processing device 1 acquires the specimen data from the testing facility device 2 (step S31) and stores the acquired specimen data in the specimen DB 111.

Based on the information stored in the specimen DB 111, the control unit 10 acquires a specimen data group including the acquired specimen data (step S32). The specimen data group is, for example, specimen data related to the same patient as the received specimen data, and is four specimen data items adjacent in time series to the received specimen data arranged in chronological order. The control unit 10 may extract specimen data including test values of test items that frequently co-occur with the current (latest) specimen data from multiple specimen data acquired in the past, and acquire a specimen data group including specimen data including the extracted test values of the frequently co-occurring test items. The method of extracting frequently co-occurring specimen data may be the same as that used when generating the training data.

The control unit 10 inputs the acquired sample data group as input data to the first learning model 113 (step S33). The control unit 10 acquires the classification information output from the first learning model 113 (step S34).

The control unit 10 calculates the contribution of the test value corresponding to each test item in the sample data input to the first learning model 113, thereby identifying the test value that contributed to the output classification information (step S35). The method for calculating the contribution is not limited, but SHAP (SHapley Additive exPlanation), LIME (Local Interpretable Model-Agnostic Explanations), etc., can be used. The greater the contribution of the test value, the greater the influence on the classification of the test value.

Based on the calculated contribution degree, the control unit 10 identifies, for example, test values whose contribution degree is equal to or greater than a predetermined value as test values that contributed to the classification information. The control unit 10 may identify test values that contributed to the classification information by, for example, selecting a predetermined number of test values in order of decreasing contribution degree. Note that, if an Attention mechanism is used for the first learning model 113, the Attention value may be used as the contribution degree. The control unit 10 executes the processes of steps S33 to S35 for each of the multiple first learning models 113. The control unit 10 may identify the test item related to the test value that contributed to the classification information.

The control unit 10 obtains the results of the verification of the appropriateness of the sample data using other methods (step S36). The control unit 10 may use, for example, the well-known delta verification method and occurrence history zone verification method as a method for verifying the appropriateness of the sample data.

The delta verification method compares the test values of a specified test item contained in the currently acquired sample data with the previous values for the same patient, and determines whether the current test value is within the upper and lower limits (e.g., ±2SD). If the current test value is outside the upper and lower limits, the current test value is determined to be abnormal. For details on the verification method using the delta verification method, see, for example, JP-A-5-151282 and JP-A-11-296605.

The occurrence history zone verification method finds the frequency distribution of test values based on their occurrence history, sets the allowable range for previous value checks, i.e., the normal zone, non-linearly, and performs single-item checks, inter-item checks, and previous value checks. If the current test value is outside the normal zone, the current test value is determined to be abnormal. For details about the verification method using the occurrence history zone verification method, please refer to Patent Publication No. 2828609, for example.

The control unit 10 generates screen information including the acquired classification information and verification results (step S37). The control unit 10 displays the generated screen information on the display device 3 (step S38), and ends the series of processes.

8 to 10 are schematic diagrams showing an example of a screen displayed on the display device 3. FIG. 8 shows an example of a result list screen 31. The result list screen 31 displays a list of detection results including classification information by the first learning model 113 and verification results by the delta verification method and the occurrence history zone verification method for the specimen data to be displayed. The result list is, for example, in a table format, and the name of the abnormality detection method (e.g., the delta method meaning the delta verification method, the zone method meaning the occurrence history zone verification method, the machine learning method meaning the first learning model 113, etc.), the initial test value, and the latest test value in the case of retesting are displayed on the horizontal axis. In addition, each item of the specimen data (e.g., specimen ID, patient ID, test value, test item, etc.) is displayed on the vertical axis. Furthermore, multiple items indicating the type of abnormality are displayed below the machine learning method. The column of one or more test values or test items in the method (item indicating the type of abnormality) in which the abnormality was detected shows the abnormality of the specimen data in a recognizable manner by a predetermined display mode. Specifically, a symbol indicating the presence or absence of an abnormality, a numerical value of the accuracy of the presence or absence of an abnormality predicted by the machine learning method, etc. are displayed.

The result list may further include detection results (degree of anomaly) using other anomaly detection methods such as Hotelling's theory, and integrated results of anomaly determination (automatic estimation results based on the degree of anomaly in Hotelling's theory or the accuracy of predictions using machine learning methods). Note that estimation methods using Hotelling's theory, etc. will be described in detail in other embodiments.

The control unit 10 of the information processing device 1 performs processing such as coloring, framing, lighting/flashing, and symbol display indicating the presence or absence of an abnormality on the test value or test item column of the abnormality type column corresponding to the classification information based on the classification information by the first learning model 113 and the identification result of the test value that contributed to the classification information. In this case, the control unit 10 may change the display mode of the abnormality depending on the classification information or the contribution degree, for example, by changing the density of the display color depending on the contribution degree of the test value or test item, or by changing the display color depending on the classification information (type of abnormality). Similarly, the control unit 10 performs processing such as coloring, framing, lighting/flashing, and symbol display indicating the presence or absence of an abnormality on the test value or test item column of the method name column in which an abnormality was detected based on the verification results by the delta verification method and the occurrence history zone verification method. For the delta verification method and the occurrence history zone verification method, various processes may be performed on the test value or test item column used in the verification.

The result list screen 31 may display time series data of test values for the patient ID to be displayed. Furthermore, the result list screen 31 is not limited to displaying each sample data (sample ID), but may display a list of detection results for multiple sample data acquired on the same test date. This makes it easier to estimate the subject of the mix-up, for example, if a mix-up is detected.

When the control unit 10 receives a tap operation on any test value while the list of results is displayed as shown in FIG. 8, it causes the display device 3 to display a first details screen 32 shown in FIG. 9. The first details screen 32 includes a graph showing the time series change of the selected test value. The vertical axis of the graph is the test value, and the horizontal axis is the test date. The control unit 10, for example, refers to the specimen DB 111 and acquires test values for the current specimen data and a predetermined number of previous specimen data (the past four specimens in the example of FIG. 9) for the same patient. The test value corresponds to the test item related to the test value selected by the tap operation. The control unit 10 plots each acquired test value on the graph.

The graph displays the normal variation range and the candidate abnormality range for the current test value. The control unit 10 estimates the normal variation range of the current test value based on, for example, the past four test values. The method of estimating the normal variation range is not limited, but the control unit 10 may estimate the normal variation range using a model that has been trained using a learning algorithm such as Gaussian process regression. The control unit 10 performs processing such as coloring differently on the area on the graph that corresponds to the normal variation range and the area that corresponds to the candidate abnormality range that is outside the normal range. By visualizing the normal variation range and the candidate abnormality range, the state of the current test value can be visually and easily grasped. Although FIG. 9 shows an example of the first detailed screen 32 that includes information about one test item, the first detailed screen 32 may display information about multiple test items simultaneously (in parallel).

In addition, when the control unit 10 receives a tap operation for the "delta method" or "zone method" while the result list is displayed as shown in FIG. 8, the control unit 10 causes the display device 3 to display the second detailed screen 32 shown in FIG. 10. The second detailed screen 32 includes a frequency distribution map of test values by the delta verification method, which is displayed in response to the selection of the "delta method," as shown in FIG. 10A, and a frequency distribution map of test values by the occurrence history zone verification method, which is displayed in response to the selection of the "zone method," as shown in FIG. 10B. Both frequency distribution maps display the normal range and abnormal range of the current test value identified by the delta verification method and the occurrence history zone verification method in a distinguishable manner, such as by different colors or boundary display. The control unit 10 generates a frequency distribution map based on the verification results. Furthermore, the control unit 10 plots the current test value at the corresponding position on the generated frequency distribution map.

According to this embodiment, the information processing device 1 can accurately detect abnormalities in the sample data using the first learning model 113. The first learning model 113 accurately classifies the sample data as normal or abnormal based on multiple test values in a time series related to test items that frequently co-occur. Since the detection results obtained by various methods are integrated and displayed, the detection results can be easily compared. Since the information processing device 1 can classify multiple types of abnormalities, subsequent measures such as retesting can be efficiently taken by isolating the cause of the abnormality.

According to this embodiment, abnormalities in the testing process, i.e. tests with a high probability of testing errors, can be estimated, so that the frequency of retesting can be minimized while efficiently increasing the reliability of test results (reducing oversight of testing errors). Furthermore, while the need for retesting is generally determined by experienced clinical laboratory technicians, this embodiment can automatically extract test subjects for retesting and automatically present the reason for retesting, allowing retesting to be performed efficiently without relying on the experience of the tester. By identifying the cause of the abnormality, it is possible to determine whether or not to retest depending on the cause of the abnormality, making it possible to automate retesting.

In each of the above flowcharts, an example in which a series of processes is executed by the control unit 10 of the information processing device 1 has been described, but the processing entity of each process is not limited. The above processes may be executed in part or in whole by the control unit 20 of the testing facility device 2. They may also be executed by the testing device 5. The first learning model 113 may be generated by the information processing device 1 and learned by the testing facility device 2. The first learning model 113 generated by the information processing device 1 may be installed in the testing facility device 2, and anomaly detection may be performed by the testing facility device 2. In addition, the classification information is not limited to being output to the display device 3 via the testing facility device 2. The control unit 10 may output the classification information to a device other than the display device 3 and display it. The information processing device 1 is not limited to being configured separately from the testing facility device 2, and may be included in the testing facility device 2.

Second Embodiment
In the second embodiment, actual abnormal specimen data obtained by a clinical test is used to train the first learning model 113. The following mainly describes the differences from the first embodiment, and the same reference numerals are used to designate the same components as in the first embodiment, and detailed descriptions thereof will be omitted.

The information processing device 1 of the second embodiment generates training data using actual abnormal sample data obtained by clinical testing. The information processing device 1 performs pre-learning of the first learning model 113 in the first embodiment, and uses newly generated training data to train the first learning model 113 in two stages (fine tuning).

Figure 11 is a flowchart showing an example of a process for generating the first learning model 113 in the second embodiment. The following process is executed by the control unit 10 according to the program 1P stored in the storage unit 11 of the information processing device 1.

The control unit 10 of the information processing device 1 acquires abnormal specimen data including abnormal values from the testing facility device 2 in association with the type of abnormality (cause of the abnormality) (step S41). The acquired abnormal specimen data includes abnormal values actually obtained in a clinical test.

The control unit 10 refers to the specimen DB 111 to acquire one or more past specimen data relating to the same patient as the acquired abnormal specimen data, for example, the past three specimen data, and generates a specimen data group including the actual abnormal specimen data and the past specimen data (step S42).

The control unit 10 generates training data, which is a data set in which classification information indicating the type of abnormality is labeled as a correct answer value for the generated sample data group (step S43). The control unit 10 stores the generated training data in the training DB 112 (step S44).

The control unit 10 uses the newly acquired training data to generate a first learning model 113 that outputs information indicating whether the last sample data in the time series contained in the sample data group is normal or abnormal when the sample data group is input (step S45). In this case, the control unit 10 trains the first learning model 113 in advance using training data that does not include actual abnormal sample data from clinical testing, by the process described in the first embodiment. The control unit 10 then performs training of the first learning model 113, using parameters from the advance training as initial values and optimizing the parameters using the new training data.

The control unit 10 stores definition information regarding the trained first learning model 113 in the memory unit 11 (step S46) and ends the series of processes.

According to this embodiment, by learning the first learning model 113 in two stages, learning can be performed efficiently and the accuracy of the first learning model 113 can be improved. For example, by pre-learning the first learning model 113 using a group of specimen data acquired from multiple testing facilities and fine-tuning using abnormal specimen data actually obtained at each testing facility, it is possible to efficiently and effectively generate a first learning model 113 suitable for each testing facility even if a sufficient number of abnormal data cannot be prepared for each testing facility.

Third Embodiment
In the third embodiment, the first learning model 113 is learned by combining with an unsupervised learning model. The following mainly describes the differences from the first embodiment, and the same reference numerals are used for the configurations common to the first embodiment, and detailed descriptions thereof are omitted.

The information processing device 1 of the third embodiment stores a second learning model 114, which is an unsupervised learning model, in the memory unit 11. The information processing device 1 uses the classification results from the second learning model 114 to learn the first learning model 113.

FIG. 12 is an explanatory diagram showing an overview of the second learning model 114 in the third embodiment. The second learning model 114 is a machine learning model configured to input a sample data group and output information indicating the presence or absence of an abnormality (abnormal or normal) in the sample data group. The second learning model 114 outputs a classification result of the presence or absence of an abnormality in the sample data group, for example, by a method of LOF (Local Outlier Factor). The second learning model 114 calculates the degree of anomaly of the sample data group and classifies the presence or absence of an abnormality by defining a case where the degree of anomaly is greater than a certain threshold as an abnormality.

The information processing device 1 performs unsupervised learning using a large amount of previously collected specimen data groups that do not contain abnormal specimen data, i.e., normal specimen data groups. The second learning model 114 is trained to identify whether the degree of abnormality of a new specimen data group is within the normal range, with specimen data groups that do not contain abnormal specimen data as normal data. The second learning model 114 identifies specimen data groups other than those that do not contain abnormal specimen data, i.e., specimen data groups that contain abnormal specimen data, as abnormal values. The second learning model 114 classifies the presence or absence of an abnormality using an outlier score (LOF) that evaluates the degree of deviation of the specimen data group, which is the input data, from the normal data. This allows the second learning model 114 to be constructed so that, when a specimen data group is input, it can appropriately output information indicating the presence or absence of an abnormality.

The second learning model 114 may include a feature extractor that extracts features of the specimen data group and the classification model described above, and may classify the presence or absence of anomalies in the specimen data group based on the features extracted by the feature extractor. The feature extractor is, for example, a convolutional neural network model that extracts features of the specimen data group and outputs the extracted features. The CNN may use a trained model that has been trained in advance, for example by transfer learning.

The configuration of the second learning model 114 is not limited to the above example, and it is sufficient if the presence or absence of an abnormality in a group of sample data can be identified by an unsupervised learning algorithm. The second learning model 114 may be configured by an anomaly detection method that defines the degree of anomaly, such as the k-nearest neighbor method, one-class SVM, or Hotelling's theory.

The input data to the second learning model 114 is not limited to a specimen data group including multiple specimen data, but may include at least the last specimen data in the specimen data group in time series, i.e., the most recent (most recent) specimen data.

Figure 13 is a flowchart showing an example of a learning process procedure for the first learning model 113 in the third embodiment. The following process is executed by the control unit 10 according to a program 1P stored in the storage unit 11 of the information processing device 1. The storage unit 11 stores a first learning model 113 that has been previously learned.

The control unit 10 of the information processing device 1 acquires a sample data group (step S51). The acquired sample data group may be the same as the sample data group acquired in step S32 of FIG. 7, for example.

The control unit 10 inputs the acquired sample data group as input data to the second learning model 114 (step S52). The control unit 10 acquires the estimation result of the presence or absence of an abnormality output from the second learning model 114 (step S53).

The control unit 10 compares the classification result obtained by the second learning model 114 with the classification information obtained by the first learning model 113 (step S54). The control unit 10 outputs the comparison result to the testing facility device 2 (step S55).

When the classification information from the first learning model 113 differs from the classification result from the second learning model 114, the control unit 10 acquires the classification result from the second learning model 114 as correction information (step S56). The control unit 10 re-learns the first learning model 113 using the correction information for the classification information, and updates the first learning model 113 (step S57). Specifically, the control unit 10 re-learns using the sample data group input to the first learning model 113 and the correction information for the classification information as training data, and updates the first learning model 113. The control unit 10 optimizes parameters so that the classification information output from the first learning model 113 approximates the corrected estimated information, and regenerates the first learning model 113.

The comparison result between the classification result by the second learning model 114 and the classification information by the first learning model 113 may be displayed, for example, using the result list screen 31 shown in FIG. 8. In addition to the classification information by the second learning model 114, the control unit 10 acquires the degree of abnormality for each test value (test item) in the second learning model 114. The control unit 10 displays each acquired degree of abnormality in the test value column of the anomaly detection method of the second learning model 114 (Hotelling's theory in the example of FIG. 8), and applies a display mode such as coloring to the test value column determined to be abnormal. In addition, the control unit 10 compares the classification result by the second learning model 114 with the classification information by the first learning model 113, and derives an integrated result of anomaly judgment (normal or abnormal) based on, for example, the degree of abnormality in the second learning model 114 and the accuracy of prediction by the first learning model 113. The control unit 10 displays the derived overall judgment result in the anomaly judgment column.

The information processing device 1 may be configured to construct a single integrated model by integrating the first learning model 113 and the second learning model 114, and to detect abnormalities in the sample data using the integrated model.

According to this embodiment, the first learning model 113 can be further optimized by combining the first learning model 113 and the second learning model 114. The first learning model 113 can effectively detect a variety of anomalies that are different from past anomalies.

Fourth Embodiment
In the fourth embodiment, abnormalities caused by pathological conditions, diseases, and physiological conditions related to the condition of a subject (hereinafter referred to as "abnormal test values caused by diseases") are detected. The following mainly describes the differences from the first embodiment, and the same reference numerals are used to designate the same components as the first embodiment, and detailed descriptions thereof will be omitted.

The information processing device 1 of the fourth embodiment stores a third learning model 115 in the storage unit 11, which detects abnormalities in test values caused by a disease. The information processing device 1 uses the third learning model 115 to detect abnormalities in the specimen data caused by the patient's disease, which are different from abnormalities during testing caused by testing errors.

Figure 14 is an explanatory diagram showing an overview of the third learning model 115 in the fourth embodiment. The third learning model 115 is a machine learning model that receives sample data composed of test values corresponding to a plurality of test items as input and outputs disease information regarding whether the sample data is normal or abnormal. The third learning model 115 is generated and learned in advance by deep learning using a neural network in the information processing device 1 or an external device. The learning algorithm may be a recurrent neural network (RNN) when time series data is acquired.

The third learning model 115 includes an input layer that inputs specimen data, an output layer that outputs disease information, and an intermediate layer (hidden layer) that extracts features. The intermediate layer has multiple nodes that extract features of the input data, and passes the features extracted using various parameters to the output layer. When specimen data is input to the input layer, calculations are performed in the intermediate layer, and disease information is output from the output layer.

The input of the third learning model 115 is sample data including test values corresponding to multiple test items. The sample data may further include attribute information such as the patient's age and gender.

The output of the third learning model 115 is disease information indicating whether the specimen data is normal or abnormal due to the patient's disease. The abnormalities may be classified into multiple types according to the disease name (e.g., diabetes, hypertension, etc.). The output layer includes nodes corresponding to the set disease information (normal or each disease name), and outputs the accuracy for each disease information as a score. The information processing device 1 can set the disease information with the highest score, or the disease information with a score equal to or greater than a threshold, as the output value of the output layer. Note that the output layer may have one output node that outputs the most accurate disease risk, instead of having multiple output nodes that output the accuracy of each disease information.

The control unit 10 learns the third learning model 115 by collecting in advance as training data a group of information in which input information obtained from a large number of clinical tests performed in the past is given known test results (normal or disease name). The control unit 10 learns various parameters that constitute the third learning model 115, for example using the backpropagation method, so as to output disease information corresponding to the sample data. In this way, when sample data is input, the third learning model 115 is constructed, which has been trained to be able to appropriately output disease information.

Figure 15 is a flowchart showing an example of a disease information acquisition processing procedure in the fourth embodiment. The following processing is executed by the control unit 10 according to the program 1P stored in the storage unit 11 of the information processing device 1.

The control unit 10 of the information processing device 1 identifies a sample data group that is classified as normal based on the classification information output from the first learning model 113, and extracts the latest (last in chronological order) sample data from the identified sample data group (step S61).

The control unit 10 inputs the acquired specimen data as input data to the third learning model 115 (step S62). The control unit 10 acquires disease information indicating whether the specimen data is normal or abnormal, which is output from the third learning model 115 (step S63).

The control unit 10 outputs the acquired disease information to the testing facility device 2 (step S64), and ends the series of processes. The control unit 10 may generate screen information that displays the acquired disease information together with classification information, etc.

According to this embodiment, in addition to abnormalities resulting from testing errors during clinical testing, more information can be provided by detecting abnormalities related to diseases associated with the physiological state of the subject. By adding disease information to abnormal values in the sample data, the validity of the clinical test results can be more appropriately determined.

The embodiments disclosed herein are illustrative in all respects and should not be considered limiting. The technical features described in each embodiment can be combined with each other, and the scope of the present invention is intended to include all modifications within the scope of the claims and equivalents to the scope of the claims.

REFERENCE SIGNS LIST 1 Information processing device 2 Testing facility device 3 Display device 4 Input device 5 Testing device 10, 20 Control unit 11, 21 Memory unit 12, 22 Communication unit 23 Input/output unit 111 Specimen DB
112 Training DB
113 First learning model 114 Second learning model 115 Third learning model 1P Program 1A Recording medium

Claims (20)

Obtaining a plurality of pieces of sample data including test values corresponding to a plurality of test items in a clinical test;
generating a first sample data group including abnormal sample data by replacing at least one test value included in the acquired plurality of sample data with an abnormal value ;
A method for generating a learning model that causes a computer to execute a process of generating a learning model that outputs the classification information when the sample data group is input, based on training data that includes the generated first sample data group, the second sample data group that does not include the abnormal sample data, and classification information regarding whether the sample data group is normal or abnormal. The method for generating a learning model according to claim 1 , further comprising the steps of: replacing a portion of the sample data among a plurality of sample data items at different test dates and times related to a first patient with the sample data item of a second patient, thereby generating the first group of sample data. The method for generating a learning model according to claim 1 or 2, further comprising the step of generating the first group of sample data by modifying the at least one test value in accordance with a predetermined rule. The method for generating a learning model according to claim 3 , wherein the predetermined rule includes changing a dilution ratio. The method for generating a learning model according to claim 1 , wherein the classification information includes normality and a plurality of types of abnormality including an abnormality caused by a mix-up of a sample. Obtaining abnormal specimen data obtained in the clinical test,
6. The method for generating a learning model according to claim 1, further comprising: learning the learning model based on the training data including a third sample data group including abnormal sample data obtained in the acquired clinical test, and a label indicating abnormality in the sample data group. 7. The method for generating a learning model according to claim 1, further comprising: training the learning model using a result of classification of the sample data group as normal or abnormal by an unsupervised model that has been trained by unsupervised learning to classify the sample data group as normal or abnormal. Extracting the specimen data having a high co-occurrence frequency from the plurality of specimen data;
The method for generating a learning model according to claim 1 , further comprising generating the sample data group from the extracted sample data with high co-occurrence frequency. Obtaining a specimen data group consisting of a plurality of specimen data arranged in chronological order;
inputting the acquired specimen data group into a learning model that has been trained to output classification information regarding normality or abnormality of the last specimen data in the time series included in the specimen data group when a specimen data group consisting of a plurality of specimen data arranged in chronological order is input, and outputting the classification information;
The learning model has been trained using training data including a first specimen data group including abnormal specimen data generated by replacing at least one test value included in a plurality of specimen data including test values corresponding to a plurality of test items in a clinical test with an abnormal value, a second specimen data group not including the abnormal specimen data, and classification information regarding normality or abnormality of the specimen data group.
A program that causes a computer to execute a process. The program according to claim 9 , further comprising: acquiring the specimen data group consisting of a plurality of specimen data with high co-occurrence frequencies extracted from the plurality of specimen data. The program according to claim 9 or 10, wherein the classification information includes an abnormality caused by a mix-up of a sample. The program according to claim 9 , wherein the classification information includes an abnormality caused by a calculation error when calculating any one of test values corresponding to a plurality of test items constituting the sample data. 13. The program according to claim 9, wherein the learning model outputs the classification information including a plurality of types of abnormalities, including an abnormality caused by a mix-up of samples and an abnormality caused by a calculation error when calculating one of the test values corresponding to a plurality of test items constituting the sample data. Obtaining a verification result regarding normality or abnormality of the sample data by a delta verification method or an occurrence record zone verification method;
The program according to claim 9 , further comprising: outputting the acquired verification result and the classification information in association with each other. The program according to claim 9 , further comprising: outputting, in an identifiable manner, from among test values corresponding to a plurality of test items constituting the specimen data, the test value or the test item having a high contribution to the classification information. 16. The program according to claim 9, further comprising: a graph showing time series changes in the latest sample data and past sample data in a clinical test; and an abnormality range of the latest sample data estimated based on the past sample data, the graph being output in association with the abnormality range of the latest sample data. Extracting the latest sample data determined to be normal by the learning model;
17. The program according to claim 9, further comprising: inputting the extracted latest specimen data into a second learning model that outputs disease information relating to abnormalities in the test values caused by a disease when specimen data consisting of test values corresponding to a plurality of test items is input; and outputting the disease information. Obtaining a specimen data group consisting of a plurality of specimen data arranged in chronological order;
inputting the acquired specimen data group into a learning model that has been trained to output classification information regarding normality or abnormality of the last specimen data in chronological order when a specimen data group consisting of a plurality of specimen data arranged in chronological order is input, and outputting the classification information;
The learning model has been trained using training data including a first specimen data group including abnormal specimen data generated by replacing at least one test value included in a plurality of specimen data including test values corresponding to a plurality of test items in a clinical test with an abnormal value, a second specimen data group not including the abnormal specimen data, and classification information regarding normality or abnormality of the specimen data group.
An information processing method for causing a computer to execute processing. an acquisition unit that acquires a sample data group consisting of a plurality of sample data arranged in chronological order;
and an output unit that inputs the acquired sample data group to a learning model that has been trained to output classification information regarding normality or abnormality of the last sample data in the time series when a sample data group consisting of a plurality of sample data arranged in chronological order is input, and outputs the classification information ;
The learning model has been trained using training data including a first specimen data group including abnormal specimen data generated by replacing at least one test value included in a plurality of specimen data including test values corresponding to a plurality of test items in a clinical test with an abnormal value, a second specimen data group not including the abnormal specimen data, and classification information regarding normality or abnormality of the specimen data group.
Information processing device. Obtaining a plurality of pieces of sample data including test values corresponding to a plurality of test items in a clinical test;
generating a first sample data group including abnormal sample data by replacing at least one test value included in the acquired plurality of sample data with an abnormal value ;
The method for generating learning model data causes a computer to execute a process of correspondingly storing the generated first sample data group and a label indicating abnormality, and the generated second sample data group not including the abnormal sample data and a label indicating normality in a storage unit as training data for a learning model.

Images