Clinical test assistance device, clinical test assistance program, clinical test assistance model, and clinical test assistance method

Abstract

A clinical test assistance device (100) comprises: a measured value acquisition unit (121) that acquires measured values regarding each of a plurality of specimens that are successively collected from a plurality of test subjects; a bias adjustment unit (122) that applies a bias value to a portion of the plurality of measured values; a training unit (123) that generates a trained model which has been trained, treating a dataset that includes a set of measured values to which a bias value was not applied and a set of measured values to which a bias value was applied as training data, such that by setting the measured value of an evaluation specimen and the measured value of a specimen collected before the evaluation specimen as input values, the trained model outputs a value indicating the probability that an error was occurring in a measurement system when the measured value of the evaluation specimen was inputted; and a suitability determination unit (124) that, by setting the respective measured values of a target specimen and a specimen collected before the target specimen as input values of the trained model, determines whether an error was occurring in the measurement system when the measured value of the target specimen was inputted.

Description

Clinical test support device, clinical test support program, clinical test support model, and clinical test support method

[0001] One aspect of the present disclosure relates to a technique for supporting clinical testing.

[0002] Clinical testing is a test used by doctors to evaluate a patient's health condition for medical treatment. Clinical testing involves analyzing various components contained in a sample taken from a patient. A sample is something taken from or excreted from the human body, such as blood or urine. Clinical testing measures the component values ​​(concentrations) of enzymes such as ALT (Alanine Transaminase) and sodium. Doctors refer to the results of clinical testing when treating patients. Clinical testing requires high accuracy. Furthermore, the results of clinical testing need to be fed back to the medical practice promptly.

[0003] To ensure the quality of clinical testing, it is necessary to regularly check the accuracy of the measuring equipment (measurement system) used in clinical testing. Generally, the component values ​​of commercially available quality control samples of known concentrations are measured several times a day. If the component values ​​measured for the quality control sample are within the acceptable range, it is assumed that there are no errors in the measurement system. Factors that can cause errors in a measurement system include dirt on the instrument (probe) that comes into contact with the sample during collection and equipment malfunction.

[0004] In clinical testing, component values ​​are measured for a large number of samples collected from a large number of patients. Specifically, component values ​​are measured sequentially, such as for the sample from patient P1, the sample from patient P2, and the sample from patient P3. Before starting the clinical testing, clinical laboratory technician C performs a first accuracy check using a quality control sample. If the measured values ​​for the quality control sample are within the acceptable range, clinical laboratory technician C begins clinical testing for a large number of samples obtained from a large number of patients. After completing the clinical testing of a predetermined number of samples, clinical laboratory technician C performs a second accuracy check using the quality control sample. If the measured component values ​​are abnormal (outside the acceptable range), the results of the clinical testing performed between the first and second accuracy checks are deemed unreliable. Abnormal measurements for the quality control sample in the second accuracy check indicate that an error occurred in the measurement system during the clinical testing performed between the first and second accuracy checks.

[0005] Clinical test results need to be fed back to the medical practice quickly. Medical institutions, especially large hospitals, perform clinical tests on a huge number of samples. Therefore, to ensure smooth clinical testing, accuracy checks must also be performed quickly. However, with accuracy check methods that use quality control samples, accuracy can only be checked when the component values ​​of the quality control samples are measured. This means that errors in the measurement system that occur between measurements cannot be detected early, making immediate evaluation difficult.

[0006] Against this background, a quality control method called PBRTQC (Patient-based Real-time Quality Control) has been proposed, which uses the measured values ​​of specimens directly rather than using quality control samples to confirm the accuracy of measurement systems. PBRTQC is a method that monitors the measured values ​​of specimens in real time, based on the premise that the distribution of a certain number of consecutive measured values ​​is always constant, and determines that an error has occurred in the measurement system when an abnormal trend is observed in multiple measured values.

[0007] Patient-Based Real-Time Quality Control: Review and Recommendations : Tony Badrick, Andreas Bietenbeck, Mark A Cervinski, Alex Katayev, Huub H van Rossum, Tze Ping Loh on behalf of the International Federation of Clinical Chemistry, and Laboratory Medicine Committee on Analytical Quality : Clinical Chemistry, Volume 65, Issue 8, 1 August 2019, Pages 962-971, https: / / doi.org / 10.1373 / clinchem.2019.305482

[0008] PBRTQC allows for earlier detection of errors in the measurement system compared to methods using quality control samples. However, PBRTQC requires the appropriate setting of various parameters, such as the boundary values ​​of the acceptable range, how to handle extreme measurement values, and the number of samples used for error judgment, which is cumbersome. Furthermore, PBRTQC is not capable of detecting errors in the measurement system early enough to meet the needs of clinical testing sites. For this reason, PBRTQC is currently not widely used in clinical testing sites.

[0009] One aspect of the present disclosure was completed in consideration of the above background, and its main purpose is to efficiently detect errors in a measurement system by inputting measurement values ​​of a patient's sample, etc. into a mathematical model.

[0010] a learning unit that uses a dataset including a set of measurement values ​​without a bias value and a set of measurement values ​​with a bias value added as training data, and that generates a trained model that is trained to output a value indicating the likelihood that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input, by setting the measurement values ​​of an evaluation sample, which is a sample to be evaluated, and the measurement values ​​of samples collected before the evaluation sample as input values; and a determination unit that determines whether an error has occurred in the measurement system when the measurement values ​​of the target sample are input, by setting the measurement values ​​of a target sample, which is a sample collected from a single subject, and a previous sample, which is a sample collected before the target sample, as input values ​​of the trained model.

[0011] According to one aspect of the present disclosure, errors in a measurement system can be detected more efficiently.

[0012] 1 is a functional block diagram of a clinical test support device. FIG. 2 is a flowchart showing the flow of an embodiment of the present disclosure. FIG. 3 is a conceptual diagram of an error detection model. FIG. 4 is a schematic diagram showing the configuration of data used in the learning stage. FIG. 5 is a schematic diagram showing an overview of bias value addition in the learning stage. FIG. 6 is a schematic diagram showing an overview of threshold setting in the setting stage. FIG. 7 is a schematic diagram showing an overview of error detection in the verification stage. FIG. 8 is a schematic diagram showing a specific example of error detection. FIG. 9 is a schematic diagram showing calculation results of period median and period mean using the MA method and the ML method. FIG. 10 is a schematic diagram showing the results of TP systematic error. FIG. 11 is a schematic diagram showing the results of Alb systematic error. FIG. 12 is a schematic diagram showing the results of Na systematic error. FIG. 13 is a schematic diagram showing the results of Cl systematic error. FIG. 14 is a schematic diagram showing the results of AST systematic error. FIG. 15 is a schematic diagram showing the results of ALT systematic error. FIG. 16 is a schematic diagram showing the results of RBC systematic error. FIG. 17 is a schematic diagram showing the results of Hb systematic error. FIG. 18 is a schematic diagram showing the results of TP random error. FIG. 19 is a schematic diagram showing the results of Alb random error. FIG. 19 is a schematic diagram showing the results of Na random error. FIG. 19 is a schematic diagram showing the results of Cl random error. 1 is a schematic diagram showing the results of random error of AST, ALT, RBC, and Hb.

[0013] First, an overview of the clinical test support device according to the present disclosure will be described. Next, an overview of the error detection model for detecting errors in the measurement system according to the present disclosure will be described. Next, details of the processing performed according to the present disclosure will be described, divided into embodiments of the learning stage and the setting stage.

[0014] FIG. 1 is a functional block diagram of a clinical test support device 100. The components of the clinical test support device 100 are implemented by hardware including computing units such as a CPU (Central Processing Unit) and various coprocessors, such as a GPU (Graphics Processing Unit), that provide auxiliary functions to the CPU, storage devices such as memory and storage, and wired or wireless communication lines connecting these, as well as software stored in the storage devices and supplying processing instructions to the computing units. The computer program may be composed of a device driver, an operating system, various application programs located at higher levels than these, and libraries that provide common functions to these programs. The blocks described below represent functional blocks rather than hardware configurations.

[0015] The clinical test support device 100 includes a user interface processing unit 110, a data processing unit 120, and a data storage unit 130. The user interface processing unit 110 accepts user operations via input interfaces such as a mouse, keyboard, and touch panel, and is also responsible for user interface-related processing such as image display and audio output. The users of the clinical test support device 100 are primarily assumed to be clinical laboratory technicians. The data storage unit 130 stores component values ​​of various components contained in samples such as blood collected from patients. Hereinafter, the patient from whom the sample is collected will be referred to as a "subject." Furthermore, the measured value of a specific component in the subject's sample will be referred to as a "measurement value." The data processing unit 120 performs various processes based on the data acquired by the user interface processing unit 110 and the data stored in the data storage unit 130.

[0016] The user interface processing unit 110 includes an input unit 111 that receives input from a user (clinical laboratory technician), and an output unit 112 that outputs various types of information such as images and sounds to the user.

[0017] The data processing unit 120 includes a measurement value acquisition unit 121, a bias adjustment unit 122, a learning unit 123, an appropriateness determination unit 124, and a threshold setting unit 125. The measurement value acquisition unit 121 registers measurement values ​​and the like input to the input unit 111 in the data storage unit 130. Note that the measurement values ​​and the like may be transmitted to the clinical test support device 100 from the measuring device used by the user (clinical laboratory technician) for the measurement via a communication network (not shown). The measurement value acquisition unit 121 also acquires measurement values ​​and the like stored in the data storage unit 130. The bias adjustment unit 122 adds a bias value to the measurement values ​​acquired by the measurement value acquisition unit 121 from the data storage unit 130. The addition of the bias value will be described in detail later. The learning unit 123 executes a learning process for an error detection model (described later) so that errors in the measurement system can be detected. The appropriateness determination unit 124 determines whether an error in the measurement system has occurred when the measurement values ​​and the like are input to the error detection model, and calculates a determination period (described later), etc. The propriety determining unit 124 also generates a screen or sound indicating the determination result regarding the occurrence of an error in the measurement system, and outputs it to the output unit 112. The threshold setting unit 125 sets a threshold (described later) that serves as a detection standard for an error in the measurement system.

[0018] 2 is a flowchart showing the flow of an embodiment of the present disclosure. The embodiment of the present disclosure is divided into a "learning stage" and a "setting stage." In the learning stage, a learning process is performed by continuously inputting multiple measured values, etc. into the error detection model (S10). The purpose of the learning stage is to generate and train the error detection model.

[0019] In the learning phase, measurements obtained for 90,000 samples are prepared. As will be described in detail later, bias values ​​are added to some of the measurements. The bias value is a numerical value that simulates the error added to a measurement when an error occurs in the measurement system. Hereinafter, measurements without a bias value added are referred to as "raw data," and measurements with a bias value added are referred to as "bias data (processed data)." The error detection model is trained to output a value y close to 0 (no error in the measurement system) when raw data is input, and a value close to 1 (error in the measurement system) when bias data is input. In this embodiment, 90,000 samples held by Osaka Municipal University were used to train (generate) the error detection model. In the following description, error detection models that have completed the training process are referred to as "trained models." Details of the learning phase will be described later with reference to Figures 3 to 5. The number of samples used to train the error detection model may be greater or less than 90,000.

[0020] The purpose of the setting stage is to determine a threshold value for the output value in the error detection model (trained model) that determines whether an error has occurred in the measurement system. For example, assume that the threshold value is 0.7. If the output value when the measured value of the sample of subject S is set in the trained model is less than 0.7, it is determined that no error has occurred in the measurement system, and if it is 0.7 or more, it is determined that an error has occurred in the measurement system.

[0021] In the setting stage, measurements of 90,000 samples are used that are either prepared separately from the 90,000 samples used in the learning stage or are the same (but without bias values). No bias values ​​are added to the measurements of the 90,000 samples used in the setting stage. The raw data of the 90,000 samples is set as the trained model, and an appropriate threshold is determined by statistically examining the output values ​​obtained from this raw data (S12). Among error detection models, those whose thresholds have been determined after going through the learning and setting stages are called "generative models." Details of the setting stage will be described later in relation to FIG. 6.

[0022] In the clinical test of this embodiment, the measured values ​​obtained from the sample are input into a trained error detection model (generative model), and in parallel with the clinical test, it is determined whether an error has occurred in the measurement system. In this embodiment, verification was also performed to confirm the effectiveness of the generative model (verification stage).

[0023] The purpose of the validation phase is to confirm how quickly and accurately the generative model can detect errors in the measurement system. In the validation phase, the performance of the generative model was confirmed using 90,000 samples, separate from those used in the training and configuration phases. Details of the validation phase will be described later with reference to Figures 7 and 8.

[0024] Next, an outline of an error detection model for detecting errors in a measurement system will be described. FIG. 3 is a conceptual diagram of the error detection model. The error detection model in this embodiment is expressed by a multiple regression analysis formula of formula (1). The error detection model is １ ~x ｎ and the objective variable y. Also, the explanatory variable x １ ~x ｎ The coefficient of a １ ~a ｎ Let's say.

[0025] Here, the explanatory variable x in equation (1) １ ~x ｎ The details of the explanatory variable x １ represents the measured value of a certain component A in a sample (target sample) collected from a subject S. ２ ~x １１ represents the measurement values ​​of the most recent 10 samples (past samples) input into the error detection model. That is, the explanatory variable x ２ ~x １１ The measured values ​​of the 10 samples input into the error detection model immediately before subject S are input to [variable]. Hereinafter, the measured values ​​of the most recent samples will be referred to as "most recent values." Note that the subject from which the sample from which the most recent value was measured is not limited to subject S. Furthermore, the number of measured values ​​included in the most recent values ​​may be 9 or less, or 11 or more.

[0026] In clinical testing, it is assumed that subject S may have undergone the same test in the past. Therefore, in order to improve the accuracy of the error detection model, the previous measurement value (previous value) of component A of subject S is also included in the explanatory variable x １２ Set as.

[0027] Explanatory variable x １３ ~x ｎ represents the attribute information of the subject S. The attribute information refers to the age, sex, inpatient / outpatient status, and medical department of the subject S. More specifically, the explanatory variable x １３ represents the age of the subject S. That is, the explanatory variable x １３ For example, if subject S is 6 years old, then x １３ 6 is input.

[0028] Explanatory variable x １４ represents gender. If subject S is male, 1 is x, and if female, 0 is x. １４ The explanatory variable x １５ indicates whether the patient is an inpatient or outpatient. For example, if the subject S is an inpatient, 1 is given, and if the subject S is an outpatient, 0 is given. １５ The explanatory variable x １６ ~x ｎ represents the department where the subject S received medical treatment. １６ , if it is respiratory surgery, x １７ In this manner, the explanatory variable x １６ ~x ｎ A value of 1 is entered for each explanatory variable. A value of 0 is entered for explanatory variables for which a value of 1 has not been entered.

[0029] The attribute information is not limited to the age, sex, inpatient / outpatient status, and medical department of the subject S, and may include other information. In other words, the number of explanatory variables may be greater than n.

[0030] The objective variable y is a numerical value (output value) that indicates the degree of probability of an error in the measurement system. The closer the value of the objective variable y is to 0, the lower the probability that an error has occurred in the measurement system, and the closer it is to 1, the higher the probability.

[0031] Although the error detection model in this embodiment is expressed by a multiple regression analysis formula such as formula (1), it may be expressed by other mathematical models. For example, the error detection model may be expressed by a neural network, eXtreme Gradient Boosting, etc.

[0032] 4 is a schematic diagram showing the configuration of data used in the learning stage. The learning process involves adjusting the coefficient a of the error detection model so that the objective variable y is 1 when an error occurs in the measurement system and 0 when no error occurs. １ ~a ｎ The purpose is to adjust the

[0033] More specifically, the configuration of the data used in the learning process in this embodiment will be described. First, the measured values ​​etc. obtained from 90,000 specimens are divided into 200 data sets. One data set contains data (measured values ​​etc.) for 450 specimens. In this case, the data for one specimen is called a record. A record is a set of explanatory variables x １ ~x ｎ n input values ​​d corresponding to １ ~d ｎ It consists of:

[0034] Furthermore, the 450 records in one dataset are divided into 50 records and 400 records. １ (variable x １ ) is the raw data, i.e., the unprocessed measurements. The input values ​​d １ ´ (variable x １ ) is bias data, i.e., processed measurement values ​​in which bias values ​​are added to raw data. The bias values ​​vary, but details will be explained later in Figure 5. Records containing raw data are called raw records (first set), and records containing bias data are called bias records (second set).

[0035] The learning process is performed by continuously inputting records into the error detection model. If the input record is a raw record, the output value D corresponding to the objective variable y is set to 0 (no error), and the coefficient a １ ~aｎ In addition, if the input record is a bias record, if the added bias value is greater than the reference bias value, the output value D is set to 1 (error exists), and if it is equal to or less than the reference bias value, the output value D is set to 0. １ ~a ｎ The reference bias value is a threshold value that serves as a reference for error judgment, and more specifically, it is the maximum value among the bias values ​​that are not detected as an error in the measurement system. If the model is trained to judge errors even in bias data to which a very small bias value that has no effect on clinical testing is added, the model will not meet the needs of clinical testing sites. With this in mind, the maximum bias value that is permissible in clinical testing is set as the reference bias value. As described above, the coefficient a is calculated using 90,000 records. １ ~a ｎ The adjustment of the error detection model is the learning process.

[0036] 4, the bias data of record 51 is not the raw data of record 1 with a bias value added, but the raw data obtained in the 51st instance with a bias value added. In other words, the subject from which record 1 was collected is different from the subject from which record 51 was collected.

[0037] 5 is a schematic diagram showing an outline of adding a bias value in the learning stage. As described above, records 51 to 450 of each data set contain bias data in which a bias value is added to raw data. For example, the measured value d of component A of a certain sample is １ (raw data) is 100. At this time, the measured value d １ If a bias value of +2% is added to １ '=100×1.02=102 bias data are generated.

[0038] In each data set, the same bias value is added to the bias data of records 51 to 450. The added bias value is selected from a set of bias values ​​that have a mean value of 0 and a variation in the form of a normal distribution with a predetermined variance. This is based on the inventors' finding that the probability of an error occurring in a measurement system in clinical testing follows a normal distribution with a mean of 0.

[0039] For example, suppose the bias value selected from the set of bias values ​​is "+5%." In this case, a bias value of "+5%" is uniformly added to the raw data contained in records 51 to 450 of dataset 1. Next, suppose the bias value selected from the set of bias values ​​is "-3%." In this case, a bias value of "-3%" is uniformly added to the raw data contained in records 51 to 450 of dataset 2. In this manner, bias values ​​are added to datasets 1 to 200.

[0040] As described above, the bias values ​​added to the bias data in records 51 to 450 of each data set are selected from a set of bias values ​​that vary in a normal distribution with a mean value of 0 and a predetermined variance. Therefore, the bias data may be close to the raw data or may deviate from the raw data. In this embodiment, the number of biases added to the raw data is distributed in a normal distribution. Therefore, there is a relatively large amount of bias data to which small bias values ​​are added. Specifically, in this embodiment, bias data to which a bias value equal to or less than the reference bias value is added is assumed to have no measurement system error, and an error detection model is trained. Furthermore, bias data to which a bias value greater than the reference bias value is added is assumed to have a measurement system error, and an error detection model is trained. Through this training, the error detection model is able to appropriately detect errors that should be detected without detecting bias values ​​that are so small that they are unnecessary to detect in clinical practice.

[0041] The number of bias values ​​added to raw data only needs to be distributed discretely to some extent. Therefore, the number of bias values ​​may be distributed in the form of a uniform distribution rather than a normal distribution. Bias values ​​may also be randomly selected with a predetermined probability from a predetermined range or an infinite range. By performing a learning process using bias values ​​with a variance in the number distribution, the performance of the error detection model can be improved.

[0042] Here, a detailed flow of the learning process of the error detection model will be described. First, a user (clinical laboratory technician) inputs the measured value of a subject for a certain component A and attribute information into the input unit 111 of the clinical test support device 101. Based on the information input into the input unit 111, the measured value acquisition unit 121 obtains the input value d １ ~d ｎ The measurement value acquisition unit 121 generates a record consisting of the most recent value d ２ ~d １１ and the previous value d １２ is acquired from the data storage unit 130. The measurement value acquisition unit 121 registers a record in the data storage unit 130.

[0043] When 90,000 records are registered in the data storage unit 130, the bias adjustment unit 122 generates 200 sets of data sets divided into 450 records each. In addition, the bias adjustment unit 122 adjusts the measured values ​​d included in records 51 to 450 in each data set. １ The bias value added by the bias adjustment unit 122 is selected from a set of bias values ​​that have a mean value of 0 and a variation in the form of a normal distribution with a predetermined variance, as described above.

[0044] When the bias adjustment unit 122 completes adding the bias value, the learning unit 123 continuously inputs records to the error detection model. The error detection model receives the input and sets an output value D. Based on the output value D, the learning unit 123 calculates the input value d １ ~d ｎ Coefficient a １ ~a ｎThe input is performed for each dataset, and records 1 to 450 are input in order for each dataset. For example, consider a situation where record 1 of dataset 2 is input to the error detection model. In this case, the most recent value of record 1 of dataset 2 (explanatory variable x ２ ~x １１ ) contains the measurements x of records 441 to 450 of dataset 1. １ When all records have been input, from record 1 of dataset 1 to record 450 of dataset 200, the learning process for the error detection model is complete.

[0045] 6 is a schematic diagram showing an overview of threshold setting in the setting stage. In the setting stage, a set of output values ​​consisting of real values ​​between 0 and 1 is obtained by inputting multiple records into the trained model. Based on this set of output values, a threshold T for detecting errors in the measurement system is set.

[0046] More specifically, in the setting stage, 90,000 records generated from the measurements of 90,000 samples, which are different from or the same as those in the learning stage (but no bias values ​​are added), are divided into 200 datasets. Each dataset contains raw records for 450 samples. Note that in the setting stage, no bias values ​​are added to records 51 to 450 of each dataset.

[0047] Next, 200 sets of data, each containing 450 sets of raw records, i.e., 90,000 records, are input to the trained model. All input records are raw records. Therefore, it is desirable that all output values ​​y be 0. In reality, most of the output values ​​y are close to 0, but values ​​close to 1 are also output. The threshold T is set from the set of output values ​​consisting of values ​​between 0 and 1 that are output in this way.

[0048] When the output value y is obtained in the manner shown in FIG. 6 , the threshold setting unit 125 sets a threshold T that serves as a reference for error detection. More specifically, when the total number of output values ​​y is 100%, the threshold setting unit 125 sets the output value y that corresponds to the smallest 99% as the threshold T. That is, the threshold setting unit 125 sets the bottom 99th percentile value when the output value y is obtained as shown in FIG. 6 as the threshold T. In FIG. 6 , the bottom 99th percentile value of the output value y is 0.85. Therefore, the threshold setting unit 125 sets 0.85 as the threshold T. In this way, the threshold T is set for the trained model, and the generative model 200 is generated. Note that the threshold T can be changed arbitrarily depending on the tolerable false positive rate, in which an error is detected even when no error actually occurred.

[0049] 7 is a schematic diagram showing an overview of error detection in the verification stage. In the verification stage, multiple records are input to the generative model 200. At this time, the propriety determining unit 124 calculates the speed from the occurrence of an error in the measurement system to its detection.

[0050] More specifically, the appropriateness determining unit 124 obtains 90,000 records stored in the data storage unit 130, i.e., 200 sets of data sets each consisting of 450 records. The records obtained by the appropriateness determining unit 124 at this time are different from the records used in the setting stage. Unlike the setting stage, bias values ​​are added to the records used in the verification stage. Specifically, in the data sets obtained by the appropriateness determining unit 124, bias values ​​are added to the raw data of records 51 to 450 by the bias adjustment unit 122. That is, records 51 to 450 include bias data. However, unlike the learning stage, the bias values ​​added in the verification stage are fixed values.

[0051] The propriety determining unit 124 continuously inputs 90,000 sets of records acquired from the data storage unit 130 to the generative model 200. Specifically, records 1 to 450 of the 200 sets of data are input to the generative model 200 in order. The generative model 200 outputs an output value y each time a record is input. Records 1 to 50 contain only raw data. Therefore, it is unlikely that the output value y will be 0.85 or greater for records 1 to 50. In other words, the propriety determining unit 124 is likely to determine that the output value y is normal, in other words, that no error has occurred in the measurement system.

[0052] However, records after record 51 contain bias data. In other words, an error has been occurring in the measurement system since record 51 was input. Accordingly, if records after record 51 are repeatedly input, the amount of bias data contained in the immediately preceding value increases. Therefore, from record 51 onwards, the output value y shows an overall increasing trend. When the propriety determination unit 124 inputs a record after record 51, the output value y is assumed to be equal to or greater than the threshold value T = 0.85. In this case, the propriety determination unit 124 determines that the output value y is abnormal, in other words, that an error has occurred in the measurement system.

[0053] As described above, in the verification stage, the propriety determining unit 124 calculates the time from the occurrence of an error in the measurement system to its detection. That is, the propriety determining unit 124 calculates the number of times the record 51 is input to the generative model 200 until the generative model 200 outputs an output value y equal to or greater than the threshold T=0.85. Hereinafter, this number of inputs will be referred to as the "determination period."

[0054] The judgment period is calculated by the propriety determination unit 124 each time the output value y exceeds the threshold value T=0.85 when records are continuously input in each data set. That is, a total of 200 judgment periods are calculated. When 200 judgment periods have been calculated, the propriety determination unit 124 also calculates the median (median period) and average (average period) of the calculated 200 judgment periods. The speed from the occurrence of an error in the measurement system to its detection is evaluated by the average period and the median period.

[0055] FIG. 8 is a schematic diagram showing a specific example of error detection. FIG. 8 shows the processing flow from when the propriety determining unit 124 inputs dataset 1 into the generative model 200 to when an error in the measurement system is detected. In the verification stage, the same bias value is uniformly added to the bias data of records 51 to 450. That is, this simulates the situation in which an error occurs in the measurement system when record 51 is input, resulting in a certain error in the measurement value. In FIG. 8, a bias value of +2% is added to the measurement value for records from record 51 onwards. At this time, the propriety determining unit 124 calculates the determination period when records from record 51 onwards are input.

[0056] When the appropriateness judgment unit 124 inputs record 1 into the generative model 200, the output value y was 0. Because the output value y is below the threshold value T=0.85, the appropriateness judgment unit 124 judges that the model is normal, that is, that no error has occurred in the measurement system. When the appropriateness judgment unit 124 inputs record 2 into the generative model 200, the output value y was 0.04. Although this is an increase compared to when record 1 was input, the output value y is below the threshold value T=0.85. Therefore, the appropriateness judgment unit 124 judges that no error has occurred in the measurement system.

[0057] Thereafter, input is repeated up to record 50. As described above, no bias is applied to the measurement values ​​up to record 50, so the output value y is below the threshold value T=0.85. In other words, during this time, the suitability determination unit 124 continues to determine that no error has occurred in the measurement system.

[0058] The propriety determining unit 124 inputs record 51 into the generative model 200. A bias is applied to the measured values ​​from record 51 onwards. In other words, an error has occurred in the measurement system. As a result of inputting record 51 into the generative model 200, an output value y = 0.05 is output. Although an error has occurred in the measurement system, the output value y is below the threshold T = 0.85, which is the error detection standard. Therefore, the propriety determining unit 124 determines that the system is normal, that is, that no error has occurred in the measurement system.

[0059] Next, the propriety determining unit 124 inputs record 52 into the generative model 200. At this time, the output value y = 0.16. As described above, from record 51 onwards, the output value y increases due to an increase in the number of bias data included in the immediately preceding value. Therefore, the output value y increases further compared to when record 51 was input. However, the output value y is still below the threshold value T = 0.85. Therefore, the propriety determining unit 124 determines that no error has occurred in the measurement system.

[0060] The suitability determination unit 124 continues to input records from record 53 onwards into the generative model 200. As the input is repeated, the number of bias data in the most recent value also increases. Furthermore, as the number of bias data in the most recent value increases, the output value y also gradually increases.

[0061] The propriety determination unit 124 inputs record 56 into the generative model 200. At this time, the output value y is 0.90. The output value y exceeds the threshold value T of 0.85. Therefore, the propriety determination unit 124 determines that an abnormality has occurred, that is, that an error has occurred in the measurement system. In this way, the propriety determination unit 124 detects an error in the measurement system.

[0062] Next, the propriety determination unit 124 calculates the determination period from the time an error occurs in the measurement value, i.e., from the time record 51 is input to the generative model 200, until the error is detected. The propriety determination unit 124 detects the error based on the output value y when record 56 is input to the generative model 200. Therefore, the propriety determination unit 124 calculates that the determination period for dataset 1 is 6.

[0063] The above-described process is similarly performed for the remaining 199 sets of data. When calculation of the assessment periods for all data sets, from data set 1 to data set 200, has been completed, the appropriateness determination unit 124 calculates the median period and the average period. The appropriateness determination unit 124 generates a result screen (not shown) indicating the median period and the average period, and displays it on the output unit 112. Alternatively, the appropriateness determination unit 124 may generate a sound indicating the median period and the average period, and output it from the output unit 112.

[0064] The appropriateness determination unit 124 may calculate indicators such as accuracy rate, sensitivity, specificity, positive predictive value, and negative predictive value to evaluate the accuracy of the generative model 200 in determining whether an error has occurred in the measurement system. The accuracy rate is the rate at which the appropriateness determination unit 124 correctly determines that a record input to the generative model 200 is abnormal if it is a bias record, and normal if it is a raw record. The sensitivity is the rate at which the appropriateness determination unit 124 correctly determines that a bias record is abnormal when input to the generative model 200. The specificity is the rate at which the appropriateness determination unit 124 correctly determines that a raw record is normal when input to the generative model 200. The positive predictive value is the rate at which the bias record is included among the records determined to be abnormal by the appropriateness determination unit 124. The negative predictive value is the rate at which the raw record is included among the records determined to be normal by the appropriateness determination unit 124.

[0065] Thereafter, the error detection process is performed while changing the bias value to be uniformly added, and the median and average values ​​for the period are calculated. In the verification stage, it is verified how the median and average values ​​for the period change when the magnitude of the bias value to be added is changed.

[0066] FIG. 9 is a schematic diagram showing the calculation results of the period median and period average using the MA method and the ML method. FIG. 9 shows the calculation results of the period median and period average when measuring the component values ​​of sodium contained in the samples of a large number of subjects. FIG. 9 shows the results of calculating the period median and period average using a method using a generative model 200 (ML method: machine-learning method). For comparison, FIG. 9 also shows the results of calculating the period median and period average using the MA method (moving-average method), a type of PBRTQC.

[0067] As shown in Figure 9, in both the MA and ML methods, the larger the bias value added to the raw data, the shorter the judgment period. In other words, the generative model can detect errors earlier. However, the ML method can detect errors in the measurement system earlier than the MA method.

[0068] In the case of sodium, the allowable error (allowable bias) when an error occurs in the measurement system is generally considered to be ±2%. For example, 200 sets of data including biased data to which a +2% bias is applied are input to the generative model 200. In this case, the calculation results of the period median and period mean values ​​using the MA method and the ML method are compared. As shown in Figure 9, the period median value using the MA method was 12.0, while the value using the ML method was 7.0. Similarly, the period mean value using the MA method was 19.0, while the value using the ML method was 9.2. In this way, the ML method allows for early detection of errors in the measurement system.

[0069] [Summary] The clinical test support device 100 has been described above based on the embodiments of the learning stage, setting stage, and verification stage. Measurement values ​​and attribute information of multiple subjects for a certain component are provided as input to the clinical test support device 100. At this time, calculations using the generative model 200 enable rapid detection of errors in the measurement system.

[0070] Traditionally, to ensure the quality of clinical testing, measurement system accuracy checks have been performed periodically. More specifically, in measurement system accuracy checks, the component values ​​of known concentrations in a quality control sample are measured several times a day using the measurement device to be checked. However, because accuracy checks can only be performed when the component values ​​of the quality control sample are measured, measurement system errors that occur between measurements cannot be detected early, potentially resulting in delays in feedback to the clinical setting. In response to these issues, a quality control method called PBRTQC has been proposed. PBRTQC utilizes the component values ​​of the subject's sample directly for quality control. However, PBRTQC requires appropriate setting of various parameters during execution, making it cumbersome. Furthermore, the ability of PBRTQC to detect measurement system errors early does not meet the needs of clinical testing. Against this background, the inventors of the present disclosure hypothesized that measurement system errors could be efficiently detected by using a mathematical model expressed in the form of a multiple regression analysis formula. In the present disclosure, an efficient error detection method in a measurement system is realized by inputting measurement values ​​and attribute information into a mathematical model expressed in the form of a multiple regression analysis formula, i.e., an error detection model.

[0071] The error detection model is trained using a dataset containing two types of records. One is a raw record (first set) containing raw data (measured values) to which no bias value, which simulates errors due to errors in the measurement system, has been added. The other is a bias record (second set) containing bias data in which a bias value has been added to the raw data. The inventors have discovered that the probability of an error occurring in a measurement system when adding a clinical test follows a normal distribution. Therefore, in the present disclosure, a bias value is randomly selected from a set of bias values ​​whose number is distributed in a normal distribution manner and added to the raw data. This makes it possible to generate a highly versatile error detection model that reflects the occurrence of errors in the measurement system in actual clinical tests.

[0072] In this embodiment, records are continuously input to the generative model 200. Each record includes the measurement value of a sample (target sample) collected from a subject. The record also includes the measurement values ​​of the most recent 10 samples (past samples) input to the generative model 200 prior to the measurement of the target sample. When a record is input, the generative model 200 determines whether the measurement value of the target sample included in the record is raw data or bias data. That is, when a record is input, the generative model 200 detects whether an error has occurred in the measurement system. As described above, the error detection model in this embodiment performs a learning process that reflects the probability of an error occurring in the measurement system in an actual clinical test. Furthermore, the error detection model in this embodiment performs a learning process so as not to detect slight measurement system errors that do not affect the clinical test as errors. Therefore, the generative model 200 can quickly detect serious measurement system errors that affect the clinical test, regardless of their severity. Furthermore, the detection frequency of minor measurement system errors that do not affect the clinical test can be reduced.

[0073] The present disclosure is not limited to the above-described embodiments and modifications, and the components can be modified without departing from the spirit of the present disclosure. Various configurations can be formed by appropriately combining multiple components disclosed in the above-described embodiments and modifications. Furthermore, some components can be deleted from all the components shown in the above-described embodiments and modifications.

[0074] [Modification] Errors in the measurement system can be caused by random errors and constant systematic errors. Random errors are errors that cause variations in the measured values ​​due to defects in the measuring equipment, etc. Constant systematic errors are errors that cause a constant bias in the measured values ​​regardless of the concentration of the quality control sample. As a modification, an error detection model that reflects these errors may be generated. Specifically, when the error detection model is expressed by equation (1), a new term a that represents the random errors and constant systematic errors is ｎ＋１ x ｎ＋１ You can also add explanatory variables x ｎ＋１The difference from the previous measurement value or the standard deviation calculated from the measurement values ​​of the most recent multiple samples may be input as the parameter.

[0075] The clinical test support device 100 may be connected to a terminal (not shown) via a communication network. The terminal may also include a user interface processing unit 110. That is, the terminal may include an input unit 111 that accepts input from a user (clinical laboratory technician) and an output unit 112 that outputs various information, such as images and audio, to the user. In this case, the clinical test support device 100 also includes a communication unit (not shown). The communication unit receives measurement values ​​and the like from the terminal via the communication network. The communication unit also transmits a result screen to the terminal via the communication network. The output unit 112 of the terminal may display the result screen received from the clinical test support device 100. Alternatively, the communication unit of the clinical test support device 100 transmits audio indicating the period median and period mean values ​​to the terminal via the communication network. The output unit 112 of the terminal may output the audio received from the clinical test support device 100.

[0076] The terminal is primarily intended to be a general-purpose computer such as a workstation in a server room in a hospital or a laptop PC in a hospital examination room, but may also be a portable communication terminal such as a tablet or smartphone. The clinical test support device 100 may also be configured to include a web server. The terminal may also be configured to include a portable communication terminal with a web browser installed thereon. The terminal may be connected to the communication network wirelessly or via a wired connection.

[0077] A system (clinical test support system) consisting of the clinical test support device 100, a communication network, and terminals may be installed on the same premises of the same medical institution. Alternatively, the clinical test support system may be provided as a so-called cloud service by installing the clinical test support device 100 at an external service provider, etc.

[0078] Generally, the more data used for learning, the more robust the model that can be constructed. Therefore, additional embodiments in which the amount of data is increased will be shown with reference to FIGS. 10 to 25 described below. Specifically, the measured values ​​etc. obtained from 250,000 specimens are divided into 1,000 datasets. One dataset contains data (measured values ​​etc.) for 250 specimens. Furthermore, the 250 records in one dataset are divided into 50 records and 200 records. The input value d of the 50 records １ (variable x １ ) is the raw data, i.e., the unprocessed measurements. The input values ​​d １ ´ (variable x １ ) is bias data, i.e., processed measurement values ​​in which bias values ​​are added to raw data. The learning method is otherwise the same. The number of data items was also increased for verification. Specifically, the suitability determination unit 124 acquires 100,000 records stored in the data storage unit 130, i.e., 400 sets of data consisting of 250 records each. In the data items acquired by the suitability determination unit 124, bias values ​​are added to the raw data items of records 51 to 250 by the bias adjustment unit 122. In other words, records 51 to 250 contain bias data. The verification method is also the same as before, except for the number of data items.

[0079] Figures 10 to 17 show examples of results for systematic error (bias shift), which is an error that biases measurement results, for each biochemical measurement item and blood test item. Figure 10 is a schematic diagram showing the results for systematic error of TP. Figure 11 is a schematic diagram showing the results for systematic error of Alb. Figure 12 is a schematic diagram showing the results for systematic error of Na. Figure 13 is a schematic diagram showing the results for systematic error of Cl. Figure 14 is a schematic diagram showing the results for systematic error of AST. Figure 15 is a schematic diagram showing the results for systematic error of ALT. Figure 16 is a schematic diagram showing the results for systematic error of RBC. Figure 17 is a schematic diagram showing the results for systematic error of Hb. In the schematic diagrams shown in Figures 10 to 17, the horizontal axis represents additive bias (Bias) and the vertical axis represents the number of samples until error detection (NPed). In these schematic diagrams, the graphs show the median value of NPed, and the dotted line on the vertical axis represents the allowable error limit. In these schematic diagrams, the EWMA (exponentially weighted moving average) method, the MA method, and the ML method are compared. As shown in Figures 10 to 17, in each evaluation item, the ML method according to the present disclosure is sharper than the other methods (EWMA and MA methods) and can detect abnormalities earlier. In addition, in each evaluation item, the ML method according to the present disclosure has a higher ability to detect errors early than the other methods.

[0080] Examples of random error results for each biochemical measurement item and blood test item are shown in Figures 18 to 25. Figure 18 is a schematic diagram showing the results of random error for TP. Figure 19 is a schematic diagram showing the results of random error for Alb. Figure 20 is a schematic diagram showing the results of random error for Na. Figure 21 is a schematic diagram showing the results of random error for Cl. Figure 22 is a schematic diagram showing the results of random error for AST. Figure 23 is a schematic diagram showing the results of random error for ALT. Figure 24 is a schematic diagram showing the results of random error for RBC. Figure 25 is a schematic diagram showing the results of random error for Hb. In the schematic diagrams shown in Figures 18 to 25, the horizontal axis represents the coefficient of variation (CV), which is the added variability, and the vertical axis represents NPed. In these schematic diagrams, the graphs show the median value of NPed, and the dotted line on the vertical axis represents the allowable error limit. In these schematic diagrams, the ML method and the MSD (moving standard deviation) method are compared. As shown in Figures 18 to 25, in all evaluation items, the ML method according to the present disclosure is below the other method (MSD method) and can always detect small NPed at any CV value. In addition, in all items, the ML method according to the present disclosure has a higher ability to detect errors early than other methods.

[0081] A clinical test support device, a clinical test support program, and a clinical test support method according to one aspect of the present disclosure may have the following configuration: [1] A clinical test support device comprising: a measurement acquisition unit that acquires measurement values ​​for each of a plurality of samples consecutively collected from a plurality of subjects; a bias adjustment unit that performs bias adjustment on some of the plurality of measurement values; a learning unit that uses a dataset including a first set that is a set of raw measurement values ​​that have not been bias adjusted and a second set that is a set of processed measurement values ​​that have been bias adjusted as learning data, and sets as input values ​​the measurement values ​​of a target sample that is a sample to be evaluated and the measurement values ​​of previous samples that are samples collected before the target sample, thereby generating a mathematical model for determining whether the measurement values ​​of the target sample belong to the first set or the second set; and an appropriateness determination unit that determines the appropriateness of the target sample of the first subject by setting as input values ​​of the mathematical model after learning processing the measurement values ​​of the target sample collected from a first subject and the measurement values ​​of previous samples collected before the target sample of the first subject, wherein the bias adjustment unit adjusts the measurement values ​​using a bias value randomly selected from a set including a plurality of types of bias values ​​during bias adjustment. [2] The clinical test support device of [1], wherein the bias adjustment unit adjusts the measurement value using a bias value selected from a set of normally distributed bias values ​​during bias adjustment. [3] The clinical test support device of [1], wherein the mathematical model is a multiple regression analysis formula including a plurality of input values, and the learning unit adjusts coefficients of the plurality of input values ​​in the learning process. [4] The clinical test support device of [3], wherein the learning unit sets attribute information of the first subject as input values ​​to the mathematical model in addition to the measurement value of the target sample collected from the first subject and the measurement value of the previous sample, and the attribute information includes one or more of the first subject's age, sex, inpatient / outpatient status, and medical department. [5] The clinical test support device of [1], wherein the learning unit further sets a difference value between the measurement value of the target sample for the first subject and the measurement value of the previous sample for the first subject as an input value to the mathematical model.[6] The clinical test support device according to [1], further comprising a threshold setting unit that sets, as a threshold, an output value that defines a subset corresponding to a predetermined proportion of the entire set of output values ​​in a set of output values ​​obtained by inputting a plurality of the raw measurement values ​​into the mathematical model after the learning process, and the suitability determination unit determines the suitability of the target sample of the first subject by comparing the measurement value of the target sample of the first subject with the threshold. [7] A clinical test support program that causes a computer to perform the following functions: acquire measurement values ​​for each of a plurality of samples collected consecutively from a plurality of subjects; perform bias adjustment on some of the plurality of measurement values; use a dataset including a first set, which is a set of raw measurement values ​​not subjected to bias adjustment, and a second set, which is a set of processed measurement values ​​that have been bias adjusted, as learning data, and set as input values ​​the measurement values ​​of a target sample that is a sample to be evaluated and measurement values ​​of previous samples that are samples collected before the target sample, thereby generating a mathematical model for determining whether the measurement values ​​of the target sample belong to the first set or the second set; and determine whether the target sample of the first subject is appropriate by setting as input values ​​of the mathematical model after learning processing the measurement values ​​of the target sample collected from the first subject and previous samples collected before the target sample of the first subject; and during the bias adjustment, adjusts the measurement values ​​by a bias value randomly selected from a plurality of types of bias values.[8] A clinical test support method comprising the steps of: a computer acquiring measurement values ​​for each of a plurality of samples successively collected from a plurality of subjects; a computer bias-adjusting some of the measurement values; a computer using a dataset including a first set, which is a set of raw measurement values ​​not subjected to bias adjustment, and a second set, which is a set of processed measurement values ​​that have been bias-adjusted, as learning data, and setting, as input values, measurement values ​​of a target sample that is a sample to be evaluated and measurement values ​​of previous samples that are samples collected before the target sample, to generate a mathematical model for determining whether the measurement values ​​of the target sample belong to the first set or the second set; and a computer determining whether the target sample of the first subject is appropriate by setting, as input values ​​of the mathematical model after learning processing, measurement values ​​of the target sample collected from a first subject and previous samples collected before the target sample of the first subject; wherein, during the bias adjustment, the measurement values ​​are adjusted using a bias value randomly selected from a plurality of types of bias values.

[0082] A clinical test support device, a clinical test support program, a clinical test support model, and a clinical test support method according to one aspect of the present disclosure may have the following configuration: [1] A clinical test support device that determines whether an error has occurred in a measurement system that performs measurements on samples, comprising: an acquisition unit that acquires measurement values ​​for each of a plurality of samples continuously collected from a plurality of subjects; an adjustment unit that adds a bias value representing an error due to an error in the measurement system to some of the measurement values; a learning unit that uses a dataset including a set of measurement values ​​without the bias value added and a set of measurement values ​​with the bias value added as training data, and sets measurement values ​​of an evaluation sample that is a sample to be evaluated and measurement values ​​of samples collected before the evaluation sample as input values ​​to generate a trained model that is trained to output a value indicating the likelihood that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input; and a determination unit that determines whether an error has occurred in the measurement system when the measurement values ​​of the target sample are input by setting measurement values ​​of a target sample that is a sample collected from a single subject and measurement values ​​of previous samples that are samples collected before the target sample as input values ​​of the trained model. [2] The clinical test support device according to [1], wherein the adjustment unit adds a bias value selected at random from a set including a plurality of types of bias values. [3] The clinical test support device according to [1] or [2], wherein the adjustment unit adds a bias value selected from a set of normally distributed bias values. [4] The clinical test support device according to any one of [1] to [3], wherein the trained model is a mathematical model expressed by a multiple regression analysis formula including a plurality of input values, and the learning unit adjusts coefficients of the plurality of input values ​​during learning. [5] The clinical test support device according to any one of [1] to [4], wherein the learning unit further sets attribute information of the one subject as an input value. [6] The clinical test support device according to [5], wherein the attribute information includes one or more of the age, sex, inpatient / outpatient status, and department of the one subject.[7] The clinical test support device according to any one of [1] to [6], wherein the learning unit further sets, as an input value, a difference value between the measurement value of the evaluation specimen and the measurement value of a specimen collected before the evaluation specimen. [8] The clinical test support device according to any one of [1] to [7], further comprising: a setting unit that sets, as a threshold, an output value that defines a subset corresponding to a predetermined proportion of the entire set of output values ​​obtained by inputting a plurality of measurement values ​​to which no bias value has been added to the trained model, the subset corresponding to a predetermined proportion of the entire set of output values, wherein the determination unit determines whether an error has occurred in the measurement system when the measurement value of the target specimen is input by comparing the output value obtained by inputting the measurement value of the target specimen to the trained model with the threshold. [9] A clinical test support program for determining whether an error has occurred in a measurement system that performs measurements on samples, comprising: an acquisition unit that acquires measurement values ​​for each of multiple samples collected consecutively from multiple subjects; an adjustment unit that adds a bias value representing an error due to an error in the measurement system to some of the multiple measurement values; a learning unit that uses a dataset including a set of measurement values ​​without added bias values ​​and a set of measurement values ​​with added bias values ​​as training data, and sets the measurement values ​​of an evaluation sample, which is a sample to be evaluated, and the measurement values ​​of samples collected before the evaluation sample as input values ​​to generate a trained model that has been trained to output a value indicating the probability that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input; and a determination unit that determines whether an error has occurred in the measurement system when the measurement values ​​of the target sample are input by setting the measurement values ​​of a target sample, which is a sample collected from a single subject, and a previous sample, which is a sample collected before the target sample, as input values ​​of the trained model.

[10] A clinical testing support model that is a trained model for causing a computer to function to determine whether an error has occurred in a measurement system that performs measurements on samples, wherein the clinical testing support model uses a dataset as training data, the dataset including: a partial set of measurement values ​​for each of a plurality of samples collected consecutively from a plurality of subjects, the partial set of measurement values ​​to which a bias value representing an error due to an error in the measurement system has not been added; and a partial set of measurement values ​​for each of a plurality of samples collected consecutively from a plurality of subjects, the partial set of measurement values ​​to which the bias value has been added; and sets as input values ​​the measurement values ​​of an evaluation sample, which is the sample to be evaluated, and the measurement values ​​of samples collected before the evaluation sample, and trains parameters to output a value indicating the probability that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input.

[11] A clinical test support method executed by a computer for determining whether an error has occurred in a measurement system that performs measurements on samples, the method comprising: an acquisition step of acquiring measurement values ​​for each of multiple samples collected consecutively from multiple subjects; an adjustment step of adding a bias value representing an error due to an error in the measurement system to some of the multiple measurement values; a learning step of using a dataset including a set of measurement values ​​without added bias values ​​and a set of measurement values ​​with added bias values ​​as training data, and setting the measurement values ​​of an evaluation sample, which is a sample to be evaluated, and the measurement values ​​of samples collected before the evaluation sample as input values ​​to generate a trained model that has been trained to output a value indicating the likelihood that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input; and a determination step of determining whether an error has occurred in the measurement system when the measurement values ​​of the target sample are input by setting the measurement values ​​of a target sample, which is a sample collected from a single subject, and a previous sample, which is a sample collected before the target sample, as input values ​​to the trained model.

[0083] 100... clinical test support device, 110... user interface processing unit, 111... input unit, 112... output unit, 120... data processing unit, 121... measurement value acquisition unit, 122... bias adjustment unit, 123... learning unit, 124... suitability determination unit, 125... threshold setting unit, 130... data storage unit, 200... generative model

Claims

1. A clinical test support device that determines whether an error has occurred in a measurement system that performs measurements on samples, comprising: an acquisition unit that acquires measurement values ​​for each of multiple samples collected continuously from multiple subjects; an adjustment unit that adds a bias value representing an error due to an error in the measurement system to a portion of the multiple measurement values; a learning unit that uses a dataset including a set of measurement values ​​to which no bias value has been added and a set of measurement values ​​to which a bias value has been added as learning data, and sets the measurement values ​​of an evaluation sample, which is a sample to be evaluated, and the measurement values ​​of samples collected prior to the evaluation sample as input values ​​to generate a trained model that has been trained to output a value indicating the likelihood that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input; and a determination unit that determines whether an error has occurred in the measurement system when the measurement values ​​of the target sample are input by setting the measurement values ​​of a target sample, which is a sample collected from a single subject, and a previous sample, which is a sample collected prior to the target sample, as input values ​​of the trained model.

2. The clinical test support device according to claim 1, wherein the adjustment section adds a bias value randomly selected from a set including a plurality of types of bias values.

3. The clinical test support device according to claim 1, wherein the adjustment unit adds a bias value selected from a set of normally distributed bias values.

4. The clinical test support device of claim 1, wherein the trained model is a mathematical model expressed by a multiple regression analysis formula including a plurality of input values, and the learning unit adjusts coefficients of the plurality of input values ​​during learning.

5. The clinical test support device according to claim 1, wherein the learning unit further sets attribute information of the one subject as an input value.

6. The clinical test support device according to claim 5, wherein the attribute information includes one or more of the subject's age, sex, inpatient / outpatient status, and medical department.

7. The clinical test support device according to claim 1, wherein the learning unit further sets, as an input value, a difference value between the measurement value of the evaluation sample and the measurement value of a sample collected prior to the evaluation sample.

8. A clinical test support device as described in claim 1, further comprising a setting unit that sets an output value that defines a subset corresponding to a predetermined proportion of the entire set of output values ​​obtained by inputting a plurality of measurement values ​​to which no bias value has been added into the trained model as a threshold, wherein the judgment unit judges whether or not an error has occurred in the measurement system when the measurement value of the target specimen is input by comparing the output value obtained by inputting the measurement value of the target specimen into the trained model with the threshold.

9. A clinical test support program for determining whether an error has occurred in a measurement system that performs measurements on samples, comprising: an acquisition unit that acquires measurement values ​​for each of multiple samples collected continuously from multiple subjects; an adjustment unit that adds a bias value representing an error due to an error in the measurement system to a portion of the multiple measurement values; a learning unit that uses a dataset including a set of measurement values ​​to which a bias value has not been added and a set of measurement values ​​to which a bias value has been added as learning data, and sets the measurement values ​​of an evaluation sample, which is a sample to be evaluated, and the measurement values ​​of samples collected prior to the evaluation sample as input values ​​to generate a trained model that has been trained to output a value indicating the likelihood that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input; and a determination unit that determines whether an error has occurred in the measurement system when the measurement values ​​of the target sample are input by setting the measurement values ​​of a target sample, which is a sample collected from a single subject, and a past sample, which is a sample collected prior to the target sample, as input values ​​of the trained model.

10. A clinical testing support model which is a trained model for causing a computer to function to determine whether or not an error has occurred in a measurement system which performs measurements on samples, the clinical testing support model using a dataset as training data which includes a partial set of measurement values ​​for each of a plurality of samples taken continuously from a plurality of subjects, the partial set of measurement values ​​to which a bias value representing an error due to an error in the measurement system has not been added, and a partial set of measurement values ​​for each of a plurality of samples taken continuously from a plurality of subjects, the partial set of measurement values ​​to which the bias value has been added, and setting as input values ​​the measurement values ​​of an evaluation sample which is the sample to be evaluated, and the measurement values ​​of a sample taken prior to the evaluation sample, so that parameters are trained to output a value indicating the probability that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input.

11. A clinical test support method executed by a computer for determining whether or not an error has occurred in a measurement system that performs measurements on samples, comprising: an acquisition step of acquiring measurement values ​​for each of a plurality of samples collected continuously from a plurality of subjects; an adjustment step of adding a bias value representing an error due to an error in the measurement system to a portion of the plurality of measurement values; a learning step of using a dataset including a set of measurement values ​​to which no bias value has been added and a set of measurement values ​​to which a bias value has been added as learning data, and setting the measurement values ​​of an evaluation sample, which is a sample to be evaluated, and the measurement values ​​of samples collected prior to the evaluation sample as input values ​​to generate a trained model that has been trained to output a value indicating the likelihood that an error has occurred in the measurement system when the measurement values ​​of the evaluation sample are input; and a determination step of determining whether or not an error has occurred in the measurement system when the measurement values ​​of the target sample are input by setting the measurement values ​​of a target sample, which is a sample collected from a single subject, and a previous sample, which is a sample collected prior to the target sample, as input values ​​of the trained model.

Images