Learning data creation method and characteristic prediction method

Abstract

The invention relates to a learning data creation method and a characteristic prediction method. Learning data used in machine learning for generating a prediction model capable of repeatedly predicting characteristics of an object to be analyzed without calibration is created. When using a prediction model to predict a K value, which is a value relating to the concentration of an object to be analyzed, a control unit (10) of an analysis device (1) repeats a predetermined number of times with respect to a plurality of objects to be analyzed, and performs a process for acquiring a predetermined number of output values indicating the characteristics of the objects to be analyzed at intervals a from a specific time. Learning data is created by associating a reference K value, which is a K value obtained at the first execution time of the repeated processing, with an output value of each of the objects to be analyzed obtained each time of the repeated processing.

Description

Technical Field

[0001] The present disclosure relates to a method for preparing learning data used for generating a prediction model for predicting characteristics of an analysis object, and a method for predicting characteristics. Background Art

[0002] Patent Document 1 describes a method for measuring the concentration of glucose using an enzyme electrode.

[0003] Prior Art Literature

[0004] Patent Literature

[0005] Patent Document 1: Japanese Patent Application No. 2020-507773

[0006] As the characteristics of the analysis object are repeatedly analyzed by the analysis device, the measurement accuracy of the characteristics of the analysis object may decrease.

[0007] Therefore, in conventional analysis devices, it is recommended to perform calibration, which is a correction operation of predicted values, when, for example, the characteristics of the analysis object have been analyzed a predetermined number of times or when a predetermined time has passed since the analysis of the characteristics of the analysis object began.

[0008] However, performing calibration in the middle of an analysis is a burdensome task for the analyst. Summary of the invention

[0009] The present disclosure aims to provide a method for preparing learning data used in machine learning for generating a prediction model capable of repeatedly predicting characteristics of an analysis object even without calibration, and a method for predicting characteristics of an analysis object.

[0010] In order to achieve the above-mentioned purpose, a method for preparing learning data according to one scheme of the present disclosure is provided in an analysis device having a reaction pool, an electrode and a control unit, wherein the control unit uses a prediction model generated by supervised machine learning to predict a K value, and includes an acquisition step, a K value calculation step and a creation step, wherein the reaction pool is provided for introducing a buffer solution and an analyte, the electrode is arranged in the reaction pool, and when a voltage of a predetermined magnitude is applied, a response current of a magnitude corresponding to the concentration of the analyte in the buffer solution in liquid contact is output, the control unit controls the timing of introducing the buffer solution and the analyte into the reaction pool respectively, and controls the acquisition of the magnitude of the response current output from the electrode as an output value representing the characteristics of the analyte, and the K value is an output value D at time t1 when the control of introducing the analyte into the reaction pool after the introduction of the buffer solution is completed. t1The output value D at the time t2 which is predetermined as a time later than the time t1 is t2 The output value D at the time t1 is represented by the difference between the values ​​of the output value D and the output value D at the time t1. t1 , the output value D at time t3+ia (i=0 to n-1: n is a predetermined integer greater than 2) at every predetermined interval a within the period starting from time t3 which is the start of the period after the time t1 and before the time t2 t3+ia , the output value D at time t2 t2 After the acquired processing is performed, the output value D of each of the analysis objects at the time t3+ia is acquired. t3+ia The processing is repeatedly performed until a predetermined number of times, and in the K value calculation step, the output value D of each of the analysis objects initially obtained in the acquisition step is t1 The output value D of each of the analysis objects t2 The difference between the K value of each of the analysis objects is calculated as the reference K value. In the preparation step, the K value having the same value as that in order to obtain each output value D t3+ia The reference K value of the analyte having the same concentration as the analyte introduced into the reaction cell is relative to each output value D of each analyte acquired each time the acquiring step is executed. t3+ia Correspondence is established to create learning data used in supervised machine learning of the prediction model.

[0011] A method for predicting characteristics of an analysis object according to an aspect of the present disclosure includes: an input step of inputting into a prediction model an output value D representing the characteristics of the analysis object at a time t3+ia (i=0 to n-1: n is a predetermined integer greater than 2) obtained at predetermined intervals a within the period starting from a time t3 which is the beginning of a period later than a time t1 and earlier than a time t2. t3+ia The prediction model is a model that is generated by supervised machine learning using the learning data produced by the method for producing learning data of one embodiment of the present disclosure and predicts a K value, wherein the K value is a value related to the concentration of the analyte, the time t1 is a time for controlling the introduction of the analyte into the reaction pool for introducing the buffer solution and the analyte, and the time t2 is a time predetermined as a time later than the time t1; and a prediction step, based on the input of the output value D t3+ia The concentration of the analyte is predicted based on the predicted value of the K value output by the prediction model.

[0012] Effects of the Invention

[0013] According to the present disclosure, it is possible to create learning data used in machine learning for generating a prediction model that repeatedly predicts the characteristics of an analysis object even without calibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a diagram showing an example of the appearance of an analysis device.

[0015] Figure 2 It is a diagram showing an example of a flow path in an analysis device.

[0016] Figure 3 It is a diagram showing the principle of measuring glucose concentration using a glucose sensor.

[0017] Figure 4 is a graph showing the relationship between the output value and the glucose concentration.

[0018] Figure 5 It is a diagram showing a configuration example of main parts of an electrical system of an analysis device including a control unit.

[0019] Figure 6 This is a flowchart showing an example of the flow of the learning data creation process.

[0020] Figure 7 This is a flowchart showing an example of the flow of data creation processing in creation example A of learning data.

[0021] Figure 8 This is an example of the evaluation result of the predicted K value predicted by prediction model A.

[0022] Fig. 9 This is a flowchart showing an example of the flow of data creation processing in creation example B of learning data.

[0023] Fig.10 This is a flowchart showing an example of the flow of data creation processing in a modification of creation example B of learning data.

[0024] Fig.11 It is a diagram showing an example of the setting range at time t4.

[0025] Fig.12 This is an example of the evaluation result of the predicted K value predicted by prediction model B. DETAILED DESCRIPTION

[0026] Hereinafter, the present embodiment will be described with reference to the accompanying drawings. It should be noted that the same components and processes are marked with the same reference numerals throughout the drawings, and repeated descriptions are omitted. The dimensional ratios of the drawings are exaggerated for the convenience of description and may be different from the actual ratios.

[0027] Figure 1 1 is a diagram showing an example of the appearance of the analysis device 1 of the present embodiment. As an example, the analysis device 1 measures a value having a correlation with the glucose concentration contained in urine and predicts the glucose concentration contained in urine, but the analysis object is not limited to urine and may be other objects such as blood. Furthermore, the characteristics of the analysis object measured by the analysis device 1 are not limited to the value having a correlation with the glucose concentration and may be other characteristics. The value having a correlation with the glucose concentration will be described in detail later.

[0028] like Figure 1 As shown, the analysis device 1 is equipped with a standard liquid container 2 filled with a standard liquid, a buffer solution container 3 filled with a buffer solution, a cleaning liquid container 4 filled with a cleaning liquid, and a drain container 5 for filling unnecessary liquid used in the analysis device 1. The standard liquid container 2, the buffer solution container 3, the cleaning liquid container 4, and the drain container 5 are connected to the analysis device 1 through tubes.

[0029] Figure 2 : is a diagram showing an example of flow paths of various liquids in the analysis device 1. The analysis device 1 includes a cleaning liquid supply pump 6A that sucks up a cleaning liquid from a cleaning liquid container 4 and supplies it to a predetermined flow path, a standard liquid supply pump 6B that sucks up a standard liquid from a standard liquid container 2 and supplies it to a predetermined flow path, a buffer solution delivery pump 6D that sucks up a buffer solution from a buffer solution container 3 and supplies it to a predetermined flow path, and a sample pump 6E that sucks up a sample, which is an example of an analyte such as blood or urine collected from a subject, from a sample container (not shown) and supplies it to a predetermined flow path. In addition, the analysis device 1 includes a drainage pump 6C that sucks up various unnecessary liquids remaining in the flow path and discharges them to a drainage container 5. Hereinafter, the cleaning liquid supply pump 6A, the standard liquid supply pump 6B, the drainage pump 6C, the buffer solution delivery pump 6D, and the sample pump 6E may be collectively referred to as "pump 6".

[0030] A filter 18 for removing foreign matter is provided at a predetermined position of the flow path. Note that the sample is sucked into the nozzle 9A by inserting the nozzle 9A into the sample container and driving the sample pump 6E.

[0031] The manifold 7B controls the connection destination of the flow path by opening and closing the flow path using the three-way electromagnetic valves 17A to 17C, thereby changing the branching pattern of the flow path.

[0032] The manifold 7B is connected to the sample pump 6E, and a flow path for sucking the sample into the nozzle 9A is formed by the control of the three-way electromagnetic valves 17A to 17C. In addition, the manifold 7B is connected to the air suction port and the nozzle 9A, and a flow path for air to flow into the nozzle 9A is formed by the control of the three-way electromagnetic valves 17A to 17C. When the air flows into the nozzle 9A, the sample sucked into the nozzle 9A is discharged from the front end of the nozzle 9A. In addition, a path for discharging the unnecessary sample sucked into the nozzle 9A to the drainage container 5 via the manifold 7A is formed by the control of the three-way electromagnetic valves 17A to 17C. In addition, the manifold 7B is connected to the flow path for supplying the cleaning liquid, and a flow path for supplying the cleaning liquid to the front end of the nozzle 9A is formed by the control of the three-way electromagnetic valves 17A to 17C.

[0033] It should be noted that the analysis device 1 is provided with a nozzle 9B and a nozzle 9C in addition to the nozzle 9A, and the liquid is sucked and discharged by the same control as the control of the nozzle 9A. Figure 2 In the example shown, the nozzle 9B is installed at a position opposite to the cleaning tank 11B filled with the cleaning solution, and the nozzle 9C is installed at a position opposite to the standard solution tank 11A filled with the standard solution. Hereinafter, the nozzle 9A, the nozzle 9B, and the nozzle 9C are sometimes collectively referred to as "the nozzle 9". The number of nozzles 9 installed in the analysis device 1 is not limited. The nozzle 9 can be moved up and down and left and right by, for example, the actuator 21.

[0034] The manifold 7A is connected to the drainage pump 6C, and the opening and closing of the flow path is controlled by the pinch valves 8A to 8E, thereby selecting from which of the plurality of flow paths connected to the manifold 7A unnecessary liquid is sucked, and the unnecessary liquid sucked by the drainage pump 6C is discharged to the drainage container 5. Hereinafter, the pinch valves 8A to 8E may be collectively referred to as "pinch valves 8".

[0035] The three-way solenoid valve 17D controls the inflow direction of the buffer solution. The buffer solution supplied to the flow path by the buffer solution delivery pump 6D passes through the three-way solenoid valve 17D and is introduced into the reaction pool 14 via the heating unit 13. The heating unit 13 is a device that sets the buffer solution to a predetermined temperature. Hereinafter, the three-way solenoid valves 17A to 17D are sometimes collectively referred to as "three-way solenoid valves 17".

[0036] In addition, a degassing module 15 and a degasser 16 are provided on the flow path of the buffer solution. The degassing module 15 is a component for separating unnecessary components such as air dissolved in the buffer solution from the buffer solution and for branching the liquid containing the unnecessary components to a flow path connected to the manifold 7A and having a check valve 19. The degasser 16 is a component for preventing bubbles from being retained in the buffer solution feeding pump 6D by degassing the buffer solution before the buffer solution enters the buffer solution feeding pump 6D, thereby achieving smooth liquid feeding.

[0037] In such a structure, the sample sucked into the nozzle 9A is introduced into the reaction cell 14 filled with a buffer solution. The reaction cell 14 is provided with a glucose sensor 12, and the concentration of glucose 25 contained in the sample (hereinafter referred to as "glucose concentration") is measured by the glucose sensor 12.

[0038] Figure 3 1 is a diagram showing the principle of measuring the glucose concentration by the glucose sensor 12. The glucose sensor 12 measures the glucose concentration using an enzyme electrode in which a glucose oxidase (Glucose Oxidase: GOD)-immobilized enzyme membrane 20 and a hydrogen peroxide electrode 22 are combined.

[0039] The GOD immobilized enzyme membrane 20 is an enzyme membrane in which GOD 27 is wrapped by a polycarbonate membrane 20A and a cellulose acetate membrane 20B. The sample introduced into the reaction cell 14 contains, for example, glucose 25 and protein 26, but the polycarbonate membrane 20A has a hole of such a size that only glucose 25 can pass through without allowing protein 26 to pass through, specifically, a hole of about 300 angstroms. Therefore, only glucose 25 passes through the polycarbonate membrane 20A. The glucose 25 entering the GOD immobilized enzyme membrane 20 is decomposed into gluconic acid 28 and hydrogen peroxide 29 by the action of GOD 27 as shown in formula (1).

[0040] (Chemical formula 1)

[0041] C6H 12 O6+O2+H2O → C6H 12 O7+H2O2 …(1)

[0042] The cellulose acetate membrane 20B has holes of a size that allows only hydrogen peroxide 29 to pass through, specifically, holes of 5 angstroms or more and 6 angstroms or less. Therefore, the decomposed hydrogen peroxide 29 passes through the cellulose acetate membrane 20B and reaches the hydrogen peroxide electrode 22. The hydrogen peroxide electrode 22 is an electrode in which platinum is used for the anode 22A and silver is used for the cathode 22B. A voltage source 24 is connected between the anode 22A and the cathode 22B. When a voltage of a predetermined magnitude is applied from the voltage source 24, the hydrogen peroxide 29 that has reached the hydrogen peroxide electrode 22 undergoes an oxidation-reduction reaction by the voltage applied by the voltage source 24, and current flows between the anode 22A and the cathode 22B.

[0043] As described above, the polycarbonate membrane 20A has holes of a size that only allow hydrogen peroxide 29 to pass through, so it will not be affected by the interference reaction based on the reducing substance. Therefore, a stable current continues to flow between the anode 22A and the cathode 22B. The current flowing between the anode 22A and the cathode 22B represents a magnitude related to the glucose concentration in the buffer. In this way, the current flowing between the anode 22A and the cathode 22B corresponding to the glucose concentration is called a "response current".

[0044] It should be noted that the reaction in the anode 22A and the reaction in the cathode 22B are respectively expressed by equation (2) and equation (3).

[0045] (Chemical formula 2)

[0046] H2O2 → 2H2+O2+2e - … (2)

[0047] O2+4H+4e - → 2H2O …(3)

[0048] An ammeter 23 is connected between the anode 22A and the cathode 22B, and the ammeter 23 performs AD (Analog to Digital) conversion on the response current, thereby obtaining an output value indicating the magnitude of the response current.

[0049] The glucose concentration is calculated based on the output value D at time t1 when the control of introducing the sample into the reaction cell 14 into which the buffer solution has been introduced is performed. t1 The output value D at time t2 which is predetermined as a time later than time t1 t2 Calculate the difference.

[0050] Figure 4 This is a diagram showing the relationship between the elapsed time from the start of the measurement of the response current and the output value. Figure 4 The horizontal axis of represents the elapsed time [seconds] from the start time of the response current measurement, and the vertical axis represents the output value. Graph 32A represents the output value at each time point.

[0051] The output value D at time t1 when the sample is introduced into the reaction cell 14 t1 It is called the base value, and the output value D at time t2 t2 The time t2 is the time when the output value that changes due to the introduction of the sample into the reaction cell 14 stabilizes. The time t0 represents the time when a predetermined amount of buffer solution is introduced into the reaction cell 14. Output value stability means that the fluctuation range of adjacent output values ​​along the time series converges to a range that can be regarded as no change in the output value.

[0052] exist Figure 4In the analysis, the difference between the basal value and the reaction value is referred to as the K value. The K value is a characteristic value that is correlated with the glucose concentration. Therefore, the glucose concentration of the sample can be predicted based on the K value. In the present disclosure, predicting the glucose concentration based on the K value is referred to as "measuring the glucose concentration." Specifically, the K value of a standard sample with a known glucose concentration is obtained using the analysis device 1, and a relationship between the glucose concentration and the K value is obtained in advance. Then, the K value of a sample with an unknown glucose concentration is measured using the analysis device 1, and the glucose concentration is predicted based on the relationship obtained in advance and the measured value of the K value.

[0053] In the above example, in principle, it takes time until time t2 to measure the glucose concentration in the sample. However, the analysis device 1 uses a prediction model pre-generated by supervised machine learning to predict the K value based on the degree of change in the output value before reaching time t2, thereby shortening the time required for measuring the glucose concentration.

[0054] Specifically, the control unit 10 obtains n (n is an integer greater than or equal to 2) output values ​​D at intervals a, i.e., at each elapsed time a, within a predetermined period starting at time t3 which is the beginning of a period after time t1 and before time t2. t3+ia Here, “i” is an integer index ranging from 0 to (n-1), and the output value D t3+ia Represents the output value at time t3+ia.

[0055] The control unit 10 converts the n output values ​​D obtained after time t3 into t3+ia Input to the prediction model to predict the K value. Later, sometimes the n output values ​​D obtained at intervals a from time t3 t3+ia Indicated as "specific output value".

[0056] The number n is a value predetermined by, for example, an experiment before measuring the glucose concentration. The larger the value of the number n, the higher the accuracy of the glucose concentration measurement, but the longer the time required for the glucose concentration measurement. It should be noted that the interval a and the time t3 are also values ​​predetermined by, for example, an experiment before measuring the glucose concentration. If the maximum value of the number n is set to n max , then at time t3+a(n max -1)<time t2 to set the number n, interval a and time t3.

[0057] In this way, the analysis device 1 measures the glucose concentration of the sample. However, in the analysis device 1 that uses the enzyme electrode to measure the glucose concentration, as the glucose concentration is repeatedly measured, the K value obtained sometimes deviates from the K value corresponding to the actual glucose concentration, and the prediction accuracy of the glucose concentration decreases. This phenomenon is called a "drift phenomenon."

[0058] For this purpose, the analyzer 1 has a calibration function. Calibration is a process for correcting the deviation between the obtained K value and the K value corresponding to the actual glucose concentration, and has the effect of making the glucose concentration measured by the analyzer 1 close to the actual glucose concentration even when the glucose concentration is repeatedly measured.

[0059] The calibration in the analysis device 1 is described. The analysis device 1 sucks up a standard solution having a predetermined glucose concentration from the standard solution container 2 and introduces it into the reaction cell 14, and predicts the K value of the standard solution by the method described above. When the predicted K value of the standard solution deviates from the standard value of the K value corresponding to the glucose concentration of the standard solution, the analysis device 1 determines a correction coefficient of the predicted K value of the standard solution so that the predicted K value of the standard solution approaches the standard value of the K value corresponding to the glucose concentration of the standard solution. In the subsequent measurement of the glucose concentration, the analysis device 1 corrects the K value using the correction coefficient determined in the calibration, thereby making the measurement result of the glucose concentration close to the actual glucose concentration of the sample.

[0060] Such calibration is generally performed in response to a user's instruction when, for example, the glucose concentration of a sample is measured a predetermined number of times or when a predetermined time has passed since the start of glucose concentration measurement.

[0061] However, the calibration of the analyzer 1 of the present disclosure can be performed at a timing that the user considers particularly necessary, such as when starting to use the analyzer 1. As will be described in detail later, the analyzer 1 has a prediction model that predicts the K value reflecting the correction factor, so that even when the user repeatedly measures the glucose concentration of the sample, the user can continue to measure the glucose concentration without calibration.

[0062] The control unit 10 controls the timing of introducing the buffer solution and the sample into the reaction pool 14 by controlling the pump 6, the pinch valve 8, the heating unit 13, the three-way solenoid valve 17 and the actuator 21, controls the supply of the cleaning liquid and the standard liquid to the analysis device 1, controls the discharge of the unnecessary liquid into the drainage container 5, controls the voltage source 24, and calibrates and obtains the output value of the ammeter 23.

[0063] The control unit 10 of the present disclosure is configured using, for example, a computer 30 . Figure 5 1 is a diagram showing a configuration example of a main part of an electrical system of the analyzing device 1 including a control unit 10 configured using a computer 30 .

[0064] The computer 30 includes a CPU (Central Processing Unit) 30A, which is an example of a processor in charge of the processing executed by the control unit 10, a RAM (Random Access Memory) 30B used as a temporary work area of ​​the CPU 30A, a nonvolatile memory 30C, and an input / output interface (I / O) 30D. The CPU 30A, the RAM 30B, the nonvolatile memory 30C, and the I / O 30D are respectively connected via a bus 30E.

[0065] The nonvolatile memory 30C is an example of a storage device that maintains stored information even when the power supplied to the nonvolatile memory 30C is cut off, and a semiconductor memory, for example, may be used, but a hard disk may also be used. The nonvolatile memory 30C stores, for example, a control program for causing the computer 30 to function as the control unit 10 and various parameters referred to when measuring the glucose concentration.

[0066] It should be noted that the nonvolatile memory 30C does not necessarily need to be built in the computer 30 , and may be a removable storage device such as a memory card that can be attached to and detached from the computer 30 .

[0067] The I / O 30D is connected to components that receive instructions from the control unit 10 and operate, such as the pump 6, the pinch valve 8, the heating unit 13, the three-way electromagnetic valve 17, the actuator 21, the ammeter 23, and the voltage source 24, but the components connected to the I / O 30D are not limited to these. It is also possible to connect components corresponding to the functions of the analysis device 1, such as a communication unit (not shown) that is connected to a communication line and communicates with an external device.

[0068] <Example A of creating learning data>

[0069] Next, a learning data preparation process for preparing learning data used for generating a prediction model for predicting a K value will be described in detail.

[0070] Figure 6 This is a flowchart showing an example of the flow of the learning data creation process executed by the control unit 10 when receiving a creation instruction of learning data from the user. The CPU 30A of the control unit 10 reads the control program from the nonvolatile memory 30C and executes the learning data creation process.

[0071] Before starting the learning data preparation process, a sample set is prepared in advance for the analyzer 1. The sample set refers to a plurality of samples having predetermined different glucose concentrations, and each sample is placed in a different sample container. Hereinafter, the predetermined glucose concentration in each sample is referred to as a "predetermined concentration."

[0072] In step S10, the control unit 10 drives the buffer solution delivery pump 6D while closing the pinch valve 8C on the flow path connecting No. 2 of the manifold 7A and the reaction cell 14, thereby introducing a specified amount of buffer solution into the reaction cell 14 provided with the glucose sensor 12 in such a manner that the buffer solution and the glucose sensor 12 are in liquid contact.

[0073] In step S20 , the control unit 10 controls the voltage source 24 to apply a voltage of a predetermined magnitude to the hydrogen peroxide electrode 22 in the glucose sensor 12 .

[0074] In step S30, the control unit 10 starts control to obtain the output value corresponding to the magnitude of the response current measured by the glucose sensor 12 at intervals a. It should be noted that the time when the output value is obtained is set to time t0. The control unit 10 uses, for example, a timer function included in the CPU 30A to manage the time, but the time information can also be obtained from the outside.

[0075] While acquiring the output value at each interval a, the control unit 10 determines whether the time t1 has been reached in step S40. The time t1 is the time when the output value is stable before the sample is introduced into the reaction cell 14. The time t1 is the time determined before starting the learning data creation process based on the results of a previously performed experiment or the like.

[0076] It should be noted that the control unit 10 may refer to the fluctuation range of adjacent output values ​​along the time series and determine that the time t1 has been reached when the output value stabilizes. In this case, the time t1 may differ for each learning data creation process.

[0077] If the time t1 has not been reached, the determination process of step S40 is repeatedly executed, and the control unit 10 continues the process of acquiring the output value at each interval a until the time t1 is reached. On the other hand, if the time t1 has been reached, the process proceeds to step S50.

[0078] In step S50, the control unit 10 controls the actuator 21 to move the nozzle 9A to the top of the sample container into which any sample not yet selected in the sample group is placed, and then lowers the nozzle 9A to allow the nozzle 9A to enter the sample container. In this state, the control unit 10 drives the sample pump 6E to suck the sample into the nozzle 9A.

[0079] Furthermore, the control unit 10 controls the actuator 21 to move the nozzle 9A that has attracted the sample to the top of the reaction cell 14, and then lowers the nozzle 9A so that the tip of the nozzle 9A enters the reaction cell 14. In this state, the control unit 10 drives the sample pump 6E to introduce a predetermined amount of the sample in the nozzle 9A into the reaction cell 14. After the sample is introduced into the reaction cell 14, the control unit 10 controls the actuator 21 to raise the nozzle 9A to a position where the tip of the nozzle 9A comes out of the reaction cell 14.

[0080] In this way, by introducing the sample into the reaction cell 14 after the output value has stabilized, the output value after the sample has been introduced can more accurately reflect the value of the change in the response current obtained only by the redox reaction of the hydrogen peroxide 29 generated by the glucose 25 contained in the sample, compared to the case where the sample is introduced before the output value has stabilized.

[0081] In step S60 , after the control unit 10 introduces the sample into the reaction cell 14 , it stores the output value first obtained as a basic value in the nonvolatile memory 30C.

[0082] Next, while acquiring output values ​​at intervals a, the control unit 10 determines whether time t3 has been reached in step S70. Time t3 is a time determined in advance by experiments before executing the learning data creation process, and represents the elapsed time from time t0.

[0083] When it is determined that the time t3 has not been reached, the determination process of step S70 is repeatedly executed, and the control unit 10 continues the process of acquiring the output value at each interval a until the time t3 is reached.

[0084] On the other hand, when it is determined that the time has been reached t3, the process proceeds to step S80.

[0085] In step S80, the control unit 10 obtains n output values ​​at intervals a and uses each of the obtained output values ​​as a specific output value D t3+ia (i=0 to n-1) is stored in the non-volatile memory 30C.

[0086] While obtaining n specific output values ​​D t3+ia , while in step S90 the control unit 10 determines whether the time t2 has been reached.

[0087] If the time t2 has not been reached, the determination process of step S90 is repeatedly executed, and the control unit 10 continues to obtain n specific output values ​​D t3+ia On the other hand, when reaching time t2, the process proceeds to step S100.

[0088] In step S100, the control unit 10 obtains the output value D at time t2.t2 , the output value D t2 The response value is stored in the nonvolatile memory 30C.

[0089] In step S110, the control unit 10 controls the voltage source 24 to stop the application of the voltage to the hydrogen peroxide electrode 22 in the glucose sensor 12, thereby completing the acquisition of the output value.

[0090] In step S120 , the control unit 10 reads the base value obtained in the process of step S60 and the reaction value obtained in the process of step S100 from the nonvolatile memory 30C, and subtracts the base value from the reaction value to calculate the K value. The control unit 10 stores the calculated K value in the nonvolatile memory 30C.

[0091] In step S130, the control unit 10 converts the n specific output values ​​D obtained by the process of step S80 into t3+ia The K value calculated by the process of step S120 is associated with the number of measurements and the predetermined concentration of the sample sucked into the nozzle 9A by the process of step S50, and is stored in the nonvolatile memory 30C. For each sample included in the sample group, the number of measurements, the predetermined concentration of the sample, and the specific output value D are associated with each sample. t3+ia And the K value establishes the corresponding data called "first result data".

[0092] In step S140, the control unit 10 determines whether there is a sample in the sample group for which the first result data has not yet been obtained. If there is a sample for which the first result data has not yet been obtained, the process proceeds to step S150.

[0093] In order to prepare for obtaining the first result data of the next sample included in the sample set, in step S150, the control unit 10 drives the discharge pump 6C while the pinch valve 8C is open, thereby discharging the buffer mixed with the sample from the reaction cell 14, emptying the reaction cell 14, and transferring to step S10. Thus, the control unit 10 repeatedly performs the processing of steps S10 to S150 until it is determined in the determination processing of step S140 that there is no sample for which the first result data has not been obtained. Therefore, the first result data for each sample included in the sample set is obtained.

[0094] It should be noted that in each of the repeated processes of step S10 to step S150 , the order of selecting the samples selected by the process of step S50 is the same in any of the repeated processes.

[0095] On the other hand, when it is determined in the determination process of step S140 that there is no sample in the sample group for which the first result data has not been obtained, the process proceeds to step S160 .

[0096] In step S160, it is determined whether the acquisition of the first result data for the sample group has been repeated M times. Here, "M" is an index indicating the number of times the acquisition of the first result data for each sample group is repeated, and is an integer greater than 2. There is no upper limit to the value of the number of repetitions M, and the number of repetitions M is the number of times predetermined before the learning data preparation process is executed. The number of repetitions M is pre-stored in, for example, the non-volatile memory 30C.

[0097] If the acquisition of the first result data for the sample group has not been repeated M times, the process proceeds to step S150, and the control unit 10 performs the process of step S10 after emptying the reaction pool 14. Thus, the control unit 10 repeatedly performs the processes of step S10 to step S160 until it is determined in the determination process of step S160 that the acquisition of the first result data for the sample group has been repeated M times. Therefore, M pieces of first result data for each sample included in the sample group are obtained.

[0098] Hereinafter, in order to distinguish whether the first result data of each sample included in the sample group is obtained by the acquisition of the first result data in the sample group unit, the expression "measurement week" is sometimes used. The measurement week is a value that increases by 1 week each time the control unit 10 acquires the first result data of each sample included in the sample group. In other words, the number of times the processing of steps S10 to S150 is repeated is equivalent to the number of samples included in the sample group. Therefore, each measurement week is equivalent to the number of executions of acquiring the first result data in the sample group unit. In this way, the number of measurements of each sample included in the sample group is also represented by the measurement week, so the number of measurements of each sample constituting the first result data is the same value as the measurement week.

[0099] It should be noted that the first result data of the sample group unit, which is a collection of the first result data of the individual samples included in the sample group, is simply referred to as “the first result data of the sample group”.

[0100] The control unit 10 acquires the first result data of the sample group in each measurement week to thereby acquire the first result data of the M sample groups.

[0101] When it is determined in the determination process of step S160 that the acquisition of the first result data of the sample group has been repeated M times, the process proceeds to step S170.

[0102] In this case, the first result data of M sample groups are obtained. Therefore, in step S170, the control unit 10 performs data preparation processing for preparing learning data for generating a prediction model, and ends Figure 6 The learning data creation process shown.

[0103] Figure 7 This is a flowchart showing an example of the flow of data creation processing in creation example A of learning data.

[0104] First, in step S200 , the control unit 10 initializes the measurement cycle M to “1”.

[0105] In step S210, the control unit 10 obtains the specific output value D corresponding to each predetermined concentration from the first result data of the sample group in the measurement week of the Mth week. t3+ia .

[0106] In step S220, the control unit 10 has the following functions to obtain each output value D t3+ia The reference K value of the sample of the same predetermined concentration as the sample introduced into the reaction cell 14 is relative to the specific output value D obtained by the process of step S210. t3+ia The benchmark K value is obtained by measuring the K value in the first week. Figure 6 The K value calculated by the process of step S120 shown above exists for each predetermined concentration of the sample.

[0107] The specific output value D is set according to each specified concentration of the sample group. t3+ia The corresponding data groups established with the benchmark K value are respectively the learning data for the prediction model and the specific output value D t3+ia is the input data of the learning data, and the reference K value is the output data of the learning data. The control unit 10 stores each of the created learning data, for example, in the nonvolatile memory 30C.

[0108] In step S230, the control unit 10 determines whether the measurement cycle M is M. max . M max Indicates the upper limit value of the measurement cycle M. In the measurement cycle M, the upper limit value M is not max In the case of a situation where a specific output value D is not obtained t3+ia The first result data of the measurement week M. Therefore, the process moves to step S240.

[0109] In step S240, the control unit 10 adds "1" to the measurement cycle M and updates the measurement cycle M, and then proceeds to step S210. Thus, the control unit 10 repeatedly executes the processing of steps S210 to S240 until the measurement cycle M is determined to be M in the determination processing of step S230. max Therefore, based on the first result data of the sample group in all the measurement weeks M, a specific output value D is generated for each measurement week and for each predetermined concentration of the sample included in the sample group. t3+ia Establish learning data corresponding to the benchmark K value.

[0110] On the other hand, in the determination process of step S230, it is determined that the measurement cycle M is M. max In the case of , the learning data is created based on the first result data of the sample group in all measurement weeks M, so the end Figure 7 The data creation process shown in FIG. 1 is completed, and the data creation process is completed. Figure 6 The learning data creation process shown.

[0111] Table 1 shows the Figure 6 An example of learning data created by the learning data creation process shown in Table 1. Table 1 shows the measurement cycle M, specific output value, and reference K value of the sample in the order of the measurement samples from the top of Table 1.

[0112]

Table 1

[0113] In Table 1, "M" included in the top item row indicates a measurement week. The combination of a specific output value and a reference K value along the row direction in each measurement week M represents learning data created from the first result data of the sample group of the corresponding measurement week M.

[0114] For the convenience of explanation, an index C is used to distinguish samples with different prescribed concentrations included in a sample group. The index C is an integer greater than 1. For example, index C=1 represents a sample with a prescribed concentration of A [mg / dL] in the sample group, and index C=2 represents a sample with a prescribed concentration of B [mg / dL] in the sample group. The C types of samples included in the sample group are distinguished by establishing a value of index C corresponding to each prescribed concentration. In addition, the upper limit value of index C, that is, the number of samples included in the sample group, is represented as C. max .

[0115] It should be noted that in order to clarify the specific output value D t3+ia The specific output value D is the specific output value of the sample with the specified concentration in which measurement week. The specific output value D is distinguished using the index C and the measurement week M. t3+ia Specific output value D[M:C] t3+ia It represents a specific output value at time t3+ia of the first result data of the sample having a predetermined concentration associated with the index C in the first result data of the sample group whose measurement week is the Mth week.

[0116] In addition, the reference K value exists for each specified concentration of the sample, so it is expressed as KC (C = 1 ~ C max) to distinguish the benchmark K value. Specifically, the benchmark K value K1 is a benchmark K value calculated based on a sample with a specific prescribed concentration that first obtained the first result data in the measurement week of the first week, and the benchmark K value K2 represents a benchmark K value calculated based on a sample with another specific prescribed concentration that secondly obtained the first result data in the measurement week of the first week.

[0117] That is, Table 1 shows the reference K value KC: (C = 1 ~ C max ) relative to a specific output value D[M:C] t3+ia :(M=1~M max 、C=1~C max 、i=0~(n max -1) Establish corresponding learning data.

[0118] By associating the reference K value corresponding to the predetermined concentration of each sample as output data with the learning data produced based on the first result data of the sample group in any measurement cycle M, learning data from which the influence of the drift phenomenon is eliminated can be obtained.

[0119] It should be noted that according to Figure 7 As can be seen from the data preparation process shown in FIG. 1 , when preparing learning data by the preparation example A of learning data, the K value in the measurement week M after the second week is unnecessary. Therefore, in the measurement week M after the second week, the control unit 10 performs the K value in the measurement week M after the second week. Figure 6 The processing of step S80 in the learning data creation process shown in FIG. t3+ia After that, the process may be transferred to step S130 without executing the process of step S90 to step S120. In this case, the control unit 10 does not calculate the K value for each sample in the measurement cycle M after the second week. Therefore, in step S130 in the measurement cycle M after the second week, the control unit 10 only needs to calculate the n specific output values ​​D obtained by the process of step S80. t3+ia The number of measurements of the sample sucked into the nozzle 9A by the process of step S50 and the predetermined concentration may be associated with each other and stored in the nonvolatile memory 30C.

[0120] <Generation of prediction model A>

[0121] Using the learning data prepared according to the learning data preparation example A, supervised machine learning based on a convolutional neural network (CNN) is performed to generate a prediction model A. The prediction model A is based on the specific output value D of the sample obtained by the analysis device 1. t3+ia An example of a prediction model to predict the K value of a sample.

[0122] Here, as an example, CNN is used to generate the prediction model A, but known supervised machine learning can be used to generate the prediction model A. Specifically, logistic regression analysis, decision tree, random forest, k-neighbor method, support vector machine (SVM), deep learning, etc. can be used instead of CNN.

[0123] <Evaluation of prediction model A>

[0124] The prediction accuracy of the K value when the prediction model A is used to predict the K value of a sample having a known glucose concentration is evaluated.

[0125] Specifically, for a sample group having the same combination of predetermined concentrations as the sample group when the learning data is prepared, Figure 6 The learning data creation process shown in Table 1 creates prior data having the same data structure as the learning data shown in Table 1. The upper limit value M of the measurement cycle M when creating prior data is max The upper limit value M when creating learning data can be max However, the upper limit M is set to be higher than that in the case of creating learning data. max As an example, the upper limit value M of the measurement cycle M when creating prior data is max Set to "140".

[0126] From each prior data thus created, prior data of a sample having the same predetermined concentration is extracted for each evaluation cycle M and used as evaluation data for the prediction model A. When the specific output value D[M:C] included in the evaluation data is t3+ia When input into prediction model A, the K value output by prediction model A, i.e. the predicted K value, is obtained, and the prediction accuracy of the K value in prediction model A is evaluated based on the degree of deviation between the benchmark K value and the predicted K value contained in the same evaluation data.

[0127] Table 2 shows an example of evaluation data for prediction model A.

[0128]

Table 2

[0129] In order to obtain the specific output value D[M:C] of the learning data t3+ia The specific output value of the evaluation data is represented by D L [M:C] t3+ia In addition, in order to distinguish it from the reference K value KC of the learning data, the reference K value of the evaluation data is expressed as K LC. As can be seen from Table 2, evaluation data was created based on the sample of the predetermined concentration corresponding to index C=7.

[0130] The deviation rate is used as an index indicating the prediction accuracy of the K value in the prediction model A. The deviation rate of the predicted K value is expressed by equation (4).

[0131] (Mathematical formula 1)

[0132] Deviation rate of predicted K value [%] = (predicted K value - benchmark K value) / benchmark K value × 100 … (4)

[0133] In order to compare with the deviation rate of the predicted K value, a specific output value D is calculated as the obtained evaluation data. L [M:C] t3+ia The deviation rate of the K value of the sample of the evaluation data preparation source calculated together with the actual measured K value is expressed by formula (5).

[0134] (Mathematical formula 2)

[0135] Deviation rate of measured K value [%] = (measured K value - reference K value) / reference K value × 100 … (5)

[0136] Figure 8 This is an example of the evaluation results showing the deviation rate of the predicted K value in the evaluation data shown in Table 2 predicted by the prediction model A and the deviation rate of the actually measured K value in the evaluation data shown in Table 2. Figure 8 The horizontal axis of the evaluation results shown represents the measurement cycle M of the evaluation data, and the vertical axis represents the deviation rate. Figure 8 The graph 33A shown represents the predicted K value, and the graph 33B represents the measured K value.

[0137] like Figure 8 As shown, in the previous method of repeatedly measuring the actual K value without using the prediction model A and without calibration in the middle, it was found that the deviation rate tended to increase with the increase of the measurement cycle M, resulting in a maximum deviation of 3.5% from the benchmark K value.

[0138] On the other hand, the predicted K value when the prediction model A is used to predict the K value of the sample does not show a tendency that the deviation rate increases with the increase of the measurement cycle M, and it is found that the deviation rate tends to converge within a certain range regardless of the measurement cycle M. Specifically, the predicted K value deviates only by about 2.4% from the reference K value. Therefore, even if the glucose concentration is repeatedly measured by the analyzer 1 that predicts the K value using the prediction model A, it is not necessary to calibrate the analyzer 1 in the middle of the repeated glucose concentration measurements.

[0139] That is, the analysis device 1 that measures the glucose concentration of a sample using the prediction model A can suppress the error of the measured glucose concentration within a certain range even if the glucose concentration is repeatedly measured.

[0140] <Example B of creating learning data>

[0141] Next, a learning data preparation process for preparing learning data used for generating a prediction model for predicting a K value using a method different from the learning data preparation example A will be described in detail.

[0142] The learning data creation process in the learning data creation example B is Figure 6 The data creation process of step S170 in the learning data creation process in the learning data creation example A shown in FIG. 1 is different. The other processes are the same as Figure 6 The learning data preparation process is the same as shown in the figure, so the following Figure 6 The data creation processing performed in step S170 is described below.

[0143] Fig. 9 This is a flowchart showing an example of the flow of the data creation process in the learning data creation example B. The CPU 30A of the control unit 10 reads the control program from the nonvolatile memory 30C and executes the data creation process.

[0144] First, in step S300 , the control unit 10 initializes the measurement cycle M to “1”.

[0145] In step S310, the control unit 10 obtains the specific output value D corresponding to each predetermined concentration from the first result data of the sample group in the measurement week of the Mth week. t3+ia .

[0146] In step S320, the control unit 10 obtains the K value from the first result data of the measurement week of the Mth week of the sample having a predetermined concentration in the sample included in the sample group. Hereinafter, the sample having the predetermined concentration is referred to as a "specific concentration sample". The specific concentration sample is an example of a specific analyte.

[0147] The number of specific concentration samples does not necessarily need to be one, and may be multiple. In the analysis device 1, the measurement accuracy is sometimes different depending on the measured glucose concentration, so the measurement accuracy for a specific glucose concentration is sometimes higher than the measurement accuracy of other glucose concentrations. Therefore, when the number of specific concentration samples is one, it is preferred to select a sample included in the sample group whose glucose concentration estimated by, for example, the reference K value is closest to the actual glucose concentration of the sample as the specific concentration sample. In this case, the measurement accuracy of the glucose concentration in the analysis device 1 is sometimes improved compared to the case where other samples are specific concentration samples.

[0148] Such a specific concentration sample is determined in advance by, for example, experiments before executing the learning data creation process.

[0149] It should be noted that when the number of specific concentration samples is set to two, for example, the sample included in the sample group whose glucose concentration estimated by the reference K value is closest to the actual glucose concentration of the sample and the second closest sample can also be selected as specific concentration samples.

[0150] In step S330 , the control unit 10 calculates, for each specific concentration sample, a drift value indicating the degree of deviation between the reference K value of the specific concentration sample and the K value of the specific concentration sample acquired by the process in step S320 .

[0151] There is no limitation on the method for calculating the drift value as long as it is a drift value that indicates the degree of deviation between the reference K value of the specific concentration sample and the K value of the specific concentration sample in the measurement week of the Mth week. For example, as shown in formula (6), the drift value is calculated by the difference between the reference K value of the specific concentration sample and the K value of the specific concentration sample in the measurement week of the Mth week.

[0152] (Mathematical formula 3)

[0153] Drift value of the Mth week = K value of the sample with a specific concentration in the measurement week of the Mth week - reference K value of the sample with a specific concentration… (6)

[0154] It should be noted that the terms on the right side of equation (6) may be exchanged to calculate the drift value as (reference K value−K value).

[0155] Furthermore, as shown in Formula (7), for example, the drift value may be calculated from the ratio of the K value of the specific concentration sample in the Mth measurement cycle to the reference K value of the specific concentration sample.

[0156] (Mathematical formula 4)

[0157] Drift value in week M = K value of sample with specific concentration in measurement week M / K value of sample with specific concentration… (7)

[0158] It should be noted that the denominator and the numerator on the right side of equation (7) may be exchanged to calculate the drift value as reference K value / K value.

[0159] In step S340, the control unit 10 compares the drift value calculated by the process of step S330 with the specific output value D of each sample included in the sample group in the Mth measurement week. t3+ia Furthermore, the control unit 10 will correspond to each other in order to obtain each output value D t3+ia The samples introduced into the reaction cell 14 have the same reference K value of the sample with the same predetermined concentration and the specific output value D of each sample included in the sample group in the measurement week of the Mth week. t3+ia Establish correspondence.

[0160] The specific output value D is set according to each specified concentration of the sample group. t3+ia , drift value and benchmark K value to establish corresponding data groups, which are the learning data of the prediction model, the specific output value D t3+ia The drift value is input data of the learning data, and the reference K value is output data of the learning data. The control unit 10 stores each of the created learning data, for example, in the nonvolatile memory 30C.

[0161] As can be seen from the process of step S340 , all learning data created based on the first result data of the sample group in the same measurement cycle M are associated with the same drift value.

[0162] It should be noted that, when there are a plurality of samples having a specific concentration, a plurality of drift values ​​are obtained, and therefore, a plurality of drift values ​​are associated with one piece of learning data.

[0163] In step S350, the control unit 10 determines whether the measurement cycle M is M. max . In the measurement cycle M is not the upper limit value M max In the case of a situation where a specific output value D is not obtained t3+ia The first result data of the measurement week M. Therefore, the process moves to step S360.

[0164] In step S360, the control unit 10 adds "1" to the measurement cycle M to update the measurement cycle M, and then proceeds to step S310. Thus, the control unit 10 repeatedly executes the processing of steps S310 to S360 until the measurement cycle M is determined to be the upper limit value M in the determination processing of step S350. max Therefore, based on the first result data of the sample group in all the measurement cycles M, a specific output value D is generated for each measurement cycle M and for each predetermined concentration of the sample included in the sample group. t3+ia , drift value and benchmark K value to establish corresponding learning data.

[0165] On the other hand, in the determination process of step S350, it is determined that the measurement cycle M is the upper limit value M. max In the case of , the learning data is created based on the first result data of the sample group in all measurement weeks M, so the end Fig. 9 The data creation process shown in FIG. 1 is completed, and the data creation process is completed. Figure 6 The learning data creation process shown.

[0166] Table 3 shows that Figure 6 The process of step S170 in the learning data preparation process shown in FIG. Fig. 9 The learning data shown in Table 3 is the learning data when there is only one sample of a specific concentration. For the convenience of explanation, the drift value generated based on the first result data of the sample group in the measurement week of the Mth week is represented as ΔK (M).

[0167]

Table 3

[0168] That is, Table 3 shows the drift value ΔK(M): (M=1~M max ), and the reference K value KC: (C = 1 ~ C max ) relative to a specific output value D[M:C] t3+ia :(M=1~M max 、C=1~C max 、i=0~(n max -1) Establish corresponding learning data.

[0169] Depend on Fig. 9 As can be seen from the data preparation process shown in FIG. 1 , when preparing learning data by the preparation example B of learning data, in the measurement weeks M after the second week, it is sufficient to calculate the K value of only the specific concentration sample among the samples included in the sample group. Therefore, in the measurement weeks M after the second week, the control unit 10 calculates the K value of only the specific concentration sample by Figure 6 The processing of step S80 in the learning data creation process shown in FIG. t3+ia Thereafter, the process may be transferred to step S130 without executing the processes of step S90 to step S120. In this case, the control unit 10 does not compare the K value with the specific output value D in step S130 for the sample for which the K value has not been calculated. t3+ia Establish correspondence.

[0170] That is, in the first result data of the sample group in the measurement week M after the second week, the K value is set only in the first result data obtained from the sample with the specific concentration. In the measurement week M after the second week, the K value of only the sample with the specific concentration among the samples included in the sample group is calculated, so compared with the case where the K value of each sample included in the sample group is calculated for each measurement week M, the time for preparing the learning data can be shortened.

[0171] <Modification of Learning Data Creation Example B>

[0172] In the learning data preparation example B, the same drift value is set in each of the learning data prepared based on the first result data of the sample group in the measurement week of the Mth week. That is, in the learning data preparation example B, the same drift value is set in the learning data in units of the measurement week M, so it is necessary to calculate the K value of the specific concentration sample for each measurement week M.

[0173] Hereinafter, a method for preparing learning data without calculating the K value of a specific concentration sample for each measurement cycle M will be described.

[0174] To this end, the control unit 10 aggregates the measurement cycles M into a plurality of measurement cycles M that are continuous for the same number of cycles in order from the first measurement cycle. max When the measurement weeks M of "6" are collected every two weeks, the control unit 10 divides the measurement weeks M into three sets of measurement weeks M: a set of measurement weeks M of the first and second weeks, a set of measurement weeks M of the third and fourth weeks, and a set of measurement weeks M of the fifth and sixth weeks. On this basis, the control unit 10 processes the set of measurement weeks M as a new unit of the number of repetitions.

[0175] In order to distinguish between the measurement week M and the set of measurement weeks M, the set of measurement weeks M will be referred to as "data set week M". p Here, “M p " is an integer greater than 2. p The upper limit value M pmax The upper limit value M of the measurement cycle M max / Q indicates. "Q" is a data group. p The number of measurement cycles M is an integer greater than 1. The user sets the upper limit value M for the measurement cycle M. max , with the data group M p The upper limit value M pmax The number of measurement cycles Q is preset to be an integer. When the number of measurement cycles Q is set to "2", two cycles of measurement cycles M correspond to the number of data set cycles M. p It should be noted that when the measurement week number Q is "1", the data group week M p= Measurement cycle M, but this is equivalent to the method of preparing learning data in the preparation example B of learning data described above. Therefore, in order to set the data cycle M p Including a plurality of measurement cycles M, the number of measurement cycles Q is set to 2 or more.

[0176] In this modification, the learning data preparation process Figure 6 The data creation process performed in step S170 in the learning data creation process in the learning data creation example A shown in FIG. 1 is replaced with the data creation process performed in step S170 in the learning data creation example A shown in FIG. Fig. 9 The data production process shown is different from the data production process. Other processes are similar to Figure 6 The learning data preparation process is the same as shown in the figure, so the following Figure 6 The data creation processing performed in step S170 is described below.

[0177] Fig.10 2 is a flowchart showing an example of the flow of the data creation process in the present modification. The CPU 30A of the control unit 10 reads the control program from the nonvolatile memory 30C and executes the data creation process.

[0178] First, in step S400, the control unit 10 groups the data into p Initialized to "1".

[0179] In step S410, the control unit 10 starts from the Mth p The specific output value D corresponding to each predetermined concentration is obtained from the first result data of the sample group of each measurement week M included in the data group of the week. t3+ia .

[0180] Data Group Week M p The one-week volume includes a plurality of measurement weeks M, so the K value of the specific concentration sample is calculated multiple times in the same data set week. Therefore, in step S420, the control unit 10 calculates the K value of the specific concentration sample from the Mth week. p The first result data of the sample with a specific concentration in the initial measurement week M included in the weekly data set week obtains the K value.

[0181] It should be noted that the number of specific concentration samples does not necessarily need to be one, but may be multiple. In this case, the control unit 10 selects each specific concentration sample from the Mth p The first result data of the sample with a specific concentration in the initial measurement week M included in the weekly data set week obtains the K value.

[0182] In step S430 , the control unit 10 calculates a drift value for each specific concentration sample.

[0183] In step S440, the control unit 10 compares the drift value calculated by the process of step S430 with the Mth pThe specific output value D of each sample in the sample group of each measurement week M included in the data group of week t3+ia Furthermore, the control unit 10 will correspond to each other in order to obtain each output value D t3+ia The sample introduced into the reaction cell 14 has the same predetermined concentration as the reference K value of the sample and the Mth p The specific output value D of each sample in the sample group of each measurement week M included in the data group of week t3+ia Establish correspondence.

[0184] According to the M p The specific output value D is set for each specified concentration of the sample group of each measurement week M included in the data group of the week. t3+ia , drift value and benchmark K value to establish corresponding data groups, which are the learning data of the prediction model, the specific output value D t3+ia The drift value is input data of the learning data, and the reference K value is output data of the learning data. The control unit 10 stores each of the created learning data, for example, in the nonvolatile memory 30C.

[0185] From the processing of step S440, it can be seen that according to the same data group, p The same drift value is set for all the learning data created using the first result data of the sample group of each measurement week M included in .

[0186] It should be noted that, when there are a plurality of samples having a specific concentration, a plurality of drift values ​​are obtained, and therefore a plurality of drift values ​​are set for one set of learning data.

[0187] In step S450, the control unit 10 determines the data group period. p Is it the upper limit value M? pmax In the data set week M p Not the upper limit M pmax In the case of a situation where a specific output value D is not obtained t3+ia The first result data of the measurement week M. Therefore, the process moves to step S460.

[0188] In step S460, the control unit 10 groups the data into p Add "1" and update the data group week M p Thus, the control unit 10 repeatedly executes the processing of step S410 to step S460 until it is determined in the determination processing of step S450 that the data group is peripheral. p The upper limit value M pmax Therefore, according to all the data sets, p The first result data of the sample group in each measurement week M included in the pAnd according to the data group week M p The specific output value D is generated for each predetermined concentration of the sample included in the sample group of each measurement week M included in t3+ia , drift value and benchmark K value to establish corresponding learning data.

[0189] On the other hand, in the determination process of step S450, it is determined that the data group is peripheral. p The upper limit value M pmax In the case of all data sets, p The first result data of the sample group in each measurement week M included in the above is used to create learning data, so the above process is completed. Fig.10 The data creation process shown in FIG. 1 is completed, and the data creation process is completed. Figure 6 The learning data creation process shown.

[0190] Table 4 shows that Figure 6 The process of step S170 in the learning data preparation process shown in FIG. Fig.10 The learning data shown in Table 4 is an example of learning data created in the case of the data creation process shown in Table 4. The learning data shown in Table 4 is the learning data when there is one specific concentration sample.

[0191]

Table 4

[0192] That is, Table 4 shows the drift value ΔK(1+Q(M p -1)): (M p =1~M pmax ), and the reference K value KC: (C = 1 ~ C max ) relative to a specific output value D[M:C] t3+ia :(M=1~M max 、C=1~C max 、i=0~(n max -1) Establish corresponding learning data.

[0193] Depend on Fig.10 As can be seen from the data creation process shown in the figure, when creating learning data using this variation, it is only necessary to calculate the initial data group period. p The K value of each sample included in the sample group in the first measurement week M, the K value of each sample included in the sample group in the second week and after the data group week M p Each data set included in the p The K value of only the sample with a specific concentration among the samples included in the sample group in the first measurement week M may be used.

[0194] Therefore, the measurement week M after the second week does not correspond to the 1st + Q (M p-1) Week 1 measurement In week 2, for each sample included in the sample group, Figure 6 The processing of step S80 in the learning data creation process shown in FIG. t3+ia After that, the process may not be performed in steps S90 to S120 but may be transferred to step S130.

[0195] Furthermore, the measurement week M after the second week is equivalent to the 1st + Q (M p -1) In the measurement cycle M of the week, the control unit 10 controls the samples other than the specific concentration sample by Figure 6 The processing of step S80 in the learning data creation process shown in FIG. t3+ia After that, the process may not be performed in steps S90 to S120 but may be transferred to step S130.

[0196] In this case, the control unit 10 does not compare the K value with the specific output value D in step S130 for the sample for which the K value has not been calculated. t3+ia Establish correspondence.

[0197] Thus, according to this modification, even if the K value of the specific concentration sample is not calculated for each measurement cycle M, the K value of the specific concentration sample can be calculated for each data set cycle M. p It only needs to be calculated for each initial measurement cycle M included in , so compared with the creation example B of the learning data, the creation time of the learning data can be shortened.

[0198] In the above-mentioned learning data preparation example B and the modified example of learning data preparation example B, the difference between the reference K value and the K value of the specific concentration sample is used as the drift value, but a value other than the difference between the reference K value and the K value may be used as the drift value.

[0199] For example, the control unit 10 Figure 6 The specific output value D is obtained in the process of step S80 of the learning data preparation process shown in FIG. t3+ia At the same time, we also obtain Fig.11 The output value D at time t4 after time t3 and before time t2 is shown. t4 .

[0200] On this basis, the control unit 10 Figure 6 In the process of S120, instead of calculating the K value, the base value is calculated as the output value D. t1 With the output value D t4 The difference (D t4 -D t1 ) as the output difference.

[0201] In the description of each process included in the subsequent learning data preparation process, the reference K value is replaced by the value obtained by Figure 6 The output difference value (hereinafter referred to as "reference output difference value") calculated by the processing of step S120 shown in FIG. 1 is replaced by the value K obtained by performing the measurement in each measurement cycle M. Figure 6 The output difference value is calculated by the process of step S120 shown in FIG. Therefore, the drift value is replaced by the value shown in equations (8) and (9).

[0202] (Mathematical formula 5)

[0203] Data Group Week M p Drift value = data set cycle M p The output difference of the specific concentration sample in the initial measurement cycle M included in the above is the reference output difference of the specific concentration sample… (8)

[0204] Data Group Week M p Drift value = data set cycle M p The output difference of the specific concentration sample in the initial measurement cycle M included in the formula is the reference output difference of the specific concentration sample… (9)

[0205] As described above, the terms on the right side of equation (8) may be exchanged to calculate the drift value as (reference output difference - output difference), and the denominator and numerator on the right side of equation (9) may be exchanged to calculate the drift value as reference output difference / output difference.

[0206] The time t4 is set in advance by the user and stored in, for example, the nonvolatile memory 30C. When the time t4 is set to the time t2, the output difference value is the K value, and the reference output difference value is the reference K value.

[0207] <Generation of prediction model B>

[0208] Using the learning data of Table 4 prepared according to the modification of the learning data preparation example B, supervised machine learning based on CNN is performed to generate a prediction model B. The prediction model B is based on the specific output value D of the sample obtained by the analysis device 1. t3+ia An example of a prediction model that predicts the K value of a sample based on its drift value.

[0209] CNN is used in the generation of prediction model B, but known supervised machine learning can be used in the generation of prediction model B as in prediction model A. Specifically, logistic regression analysis, decision tree, random forest, k-neighbor method, SVM, deep learning, etc. can be used instead of CNN.

[0210] It should be noted that it is needless to say that prediction model B can also be prepared using the learning data in Table 3 prepared according to preparation example B of learning data.

[0211] <Evaluation of prediction model B>

[0212] The prediction accuracy of the K value when predicting the K value of a sample having a known glucose concentration using the prediction model B generated by performing supervised machine learning using the learning data shown in Table 4 was evaluated.

[0213] Specifically, for a sample group having the same combination of predetermined concentrations as the sample group when the learning data is prepared, the process of step S170 is executed as Fig.10 The data shown is produced and processed Figure 6 The learning data creation process shown in Table 4 creates prior data having the same data structure as the learning data shown in Table 4. The upper limit value M of the measurement cycle M when creating prior data max It can also be the upper limit value M when creating learning data max However, it is set to be less than the upper limit value M when creating learning data. max As an example, the upper limit value M of the measurement cycle M when creating prior data is max Set to "140". That is, the data group period p The upper limit value M pmax is "70".

[0214] From each prior data thus created, prior data of a sample having the same predetermined concentration is extracted for each evaluation cycle M and used as evaluation data for the prediction model B. When the specific output value D included in the evaluation data is L [M:C] t3+ia When the drift value ΔK(M) is input into prediction model B, the predicted K value output by prediction model B is obtained, and the prediction accuracy of the K value in prediction model B is evaluated based on the degree of deviation between the reference K value and the predicted K value contained in the same evaluation data.

[0215] Table 5 shows an example of evaluation data for prediction model B. As can be seen from Table 5, the evaluation data is created based on the sample of the predetermined concentration corresponding to index C = 7. It should be noted that the number of measurement cycles Q = 2 is set.

[0216]

Table 5

[0217] In the prediction model B, the prediction accuracy of the K value is also evaluated, and therefore the deviation rates shown in equations (4) and (5) are calculated.

[0218] Fig.12 This is an example of the evaluation results showing the deviation rate of the predicted K value in the evaluation data shown in Table 5 predicted by the prediction model B and the deviation rate of the actually measured K value in the evaluation data shown in Table 5. Fig.12 The horizontal axis of the evaluation results shown represents the measurement cycle M of the evaluation data, and the vertical axis represents the deviation rate. Fig.12 The graph 33C shown represents the predicted K value, and the graph 33B represents the measured K value.

[0219] like Fig.12 As shown, in the previous method of repeatedly measuring the actual K value without using the prediction model B and without calibration in the middle, a maximum deviation of 3.5% was produced relative to the reference K value. On the other hand, the predicted K value when the prediction model B is used to predict the K value of the sample did not show a tendency that the deviation rate increases with the increase of the measurement cycle M, and it was found that the deviation rate tends to converge within a certain range regardless of the measurement cycle M. Specifically, the predicted K value only deviates by about 1.05% relative to the reference K value. Therefore, even if the glucose concentration is repeatedly measured by the analyzer 1 that predicts the K value using the prediction model B, it is not necessary to calibrate the analyzer 1 in the middle of the repeated glucose concentration measurements.

[0220] That is, the analysis device 1 that measures the glucose concentration of a sample using the prediction model B can suppress the error of the measured glucose concentration within a certain range even if the glucose concentration is repeatedly measured.

[0221] In the above, a method of preparing learning data for a prediction model used in the analysis device 1 is described using the embodiment, but the disclosed method of preparing learning data is only an example, and the method of preparing learning data is not limited to the scope described in the embodiment. Various changes or improvements can be made to the embodiment without departing from the scope of the present disclosure, and the method to which such changes or improvements are made is also included in the disclosed technical scope.

[0222] For example, the following may be changed without departing from the scope of the present disclosure: Figure 6 The internal processing sequence and Figure 7 , Fig. 9 as well as Fig.10 The internal processing sequence in the data creation process shown.

[0223] In the above-mentioned embodiment, as an example, the method of implementing the learning data preparation process and the data preparation process by software is described. However, it is also possible to implement the same by hardware. Figure 6 The flowchart of the learning data preparation process shown in FIG. Figure 7 , Fig. 9 as well as Fig.10 In this case, compared with the case where each process is implemented by software, the processing speed is increased.

[0224] In the above-mentioned embodiment, the processor refers to a processor in a broad sense, including a general-purpose processor (such as CPU30A) and a dedicated processor (such as GPU: Graphics Processing Unit, ASIC: Application Specific Integrated Circuit, FPGA: Field Programmable Gate Array, programmable logic device, etc.).

[0225] The actions of the processor in the above-mentioned embodiments can be completed not only by one processor, but also by a plurality of processors in physically separate locations. Furthermore, the order of the actions of the processor is not limited to the order described in the above-mentioned embodiments, but can also be changed appropriately.

[0226] In the above embodiment, an example is described in which the control program is pre-stored in the nonvolatile memory 30C. However, the storage location of the control program is not limited to the nonvolatile memory 30C. The control program of the present disclosure can also be provided in the form of being recorded on a storage medium readable by the computer 30.

[0227] For example, the control program can be provided in the form of being recorded on an optical disc such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), and a Blu-ray Disc. Furthermore, the control program can also be provided in the form of being recorded on a removable semiconductor memory such as a USB (Universal Serial Bus) memory and a memory card. The nonvolatile memory 30C, CD-ROM, DVD-ROM, Blu-ray Disc, USB, and memory card are examples of non-transitory storage media.

[0228] Furthermore, the analyzing device 1 may download the control program from an external device connected to the communication line via the communication unit, and store the downloaded control program in the nonvolatile memory 30C of the analyzing device 1. In this case, the CPU 30A of the analyzing device 1 reads the control program downloaded from the external device from the nonvolatile memory 30C and executes the learning data creation process and the data creation process.

[0229] The following are additional notes of the present disclosure.

[0230] (Note 1)

[0231] A method for preparing learning data, wherein:

[0232] In an analysis device having a reaction pool, an electrode and a control unit, when the control unit uses a prediction model generated by supervised machine learning to predict a K value, the method for preparing learning data includes an acquisition step, a K value calculation step and a preparation step.

[0233] The reaction pool is used for introducing a buffer solution and an analyte.

[0234] The electrode is disposed in the reaction cell and, when a voltage of a predetermined magnitude is applied, outputs a response current of a magnitude corresponding to the concentration of the analyte in the buffer solution in liquid contact.

[0235] The control unit controls the timing of introducing the buffer solution and the analyte into the reaction cell, and controls the magnitude of the response current output from the electrode as an output value representing a characteristic of the analyte.

[0236] The K value is an output value D at time t1 when control is performed to introduce the analyte into the reaction cell after the introduction of the buffer solution. t1 The output value D at the time t2 which is predetermined as a time later than the time t1 is t2 The difference between the two represents a value related to the concentration of the analyte.

[0237] In the acquisition step, for each of the plurality of analytes having different predetermined concentrations, the output value D at the time t1 is first acquired at least once. t1 , the output value D at time t3+ia (i=0 to n-1: n is a predetermined integer greater than 2) at every predetermined interval a within the period starting from time t3 which is the start of the period after the time t1 and before the time t2 t3+ia , the output value D at time t2 t2After the acquired processing is performed, the output value D of each of the analysis objects at the time t3+ia is acquired. t3+ia The processing is repeated until a predetermined number of times.

[0238] In the K value calculation step, the output value D of each of the analysis objects initially acquired in the acquisition step is calculated. t1 The output value D of each of the analysis objects t2 The difference between the K value of each of the analysis objects is calculated as the reference K value.

[0239] In the production step, the output value D t3+ia The reference K value of the analyte having the same concentration as the analyte introduced into the reaction cell is relative to each output value D of each analyte acquired each time the acquiring step is executed. t3+ia Correspondence is established to create learning data used in supervised machine learning of the prediction model.

[0240] (Note 2)

[0241] The method for preparing learning data according to Supplement 1, wherein:

[0242] In each execution of the acquisition step, the output value D at time t4 is acquired for each of the analysis objects. t4 , and in each subsequent execution of the acquisition step, the output value D at the time t1 is acquired for each of the analysis objects. t1 , the time t4 is a time after the time t3 and before the time t2,

[0243] The method for preparing learning data includes a calculation step and a drift value calculation step.

[0244] In the calculation step, for a specific analyte having a predetermined specific concentration among the analytes, the output value D of the specific analyte obtained in each execution of the acquisition step is calculated. t1 With the output value D t4 The difference between the values ​​of the output of the specific analysis object in each execution of the acquisition step is calculated as the output difference of the specific analysis object in each execution of the acquisition step.

[0245] In the drift value calculation step, a drift value is calculated for each execution of the acquisition step, the drift value indicating a degree of deviation between the output difference value of the specific analysis object in the first execution of the acquisition step and the output difference value of the specific analysis object in each execution of the acquisition step.

[0246] In the creating step, the output values ​​D of each of the analysis objects in each execution of the acquiring step, including the output values ​​D of each of the analysis objects in each execution of the acquiring step, are obtained. t3+ia And get each output value D t3+ia The input data of the drift value in the same execution times will have the same value as that in order to obtain each output value D t3+ia The reference K value of the analyte having the same concentration as the analyte introduced into the reaction cell is associated as output data, and learning data used in the supervised machine learning of the prediction model is prepared.

[0247] (Note 3)

[0248] The method for preparing learning data according to Supplement 2, wherein:

[0249] The time t4 is set to the time t2.

[0250] (Note 4)

[0251] The method for preparing learning data according to Appendix 2 or 3, wherein:

[0252] The specific analyte is the analyte whose concentration estimated from the reference K value of each analyte is closest to the actual concentration of the analyte.

[0253] (Note 5)

[0254] The method for preparing learning data according to Appendix 2 or 3, wherein:

[0255] The specific analyte is a plurality of analyte having different concentrations that are pre-specified among the analyte.

[0256] In the drift value calculation step, the drift value of each specific analyte is calculated each time the acquisition step is performed.

[0257] In the creating step, the output values ​​D of each of the analysis objects in each execution of the acquiring step, including the output values ​​D of each of the analysis objects in each execution of the acquiring step, are obtained. t3+ia And get each output value D t3+ia The input data of the plurality of said drift values ​​in the same execution times will have the same value as that in order to obtain each output value D t3+ia The reference K value of the analyte having the same concentration as the analyte introduced into the reaction cell is associated as output data, and learning data used in the supervised machine learning of the prediction model is prepared.

[0258] (Note 6)

[0259] The method for preparing learning data according to any one of Supplements 2 to 5, wherein:

[0260] In the drift value calculation step, when a plurality of consecutive execution times of the acquisition step are collected and processed as a unit of new execution times of the acquisition step so that each of the new execution times includes a plurality of output difference values ​​of the specific analyte having the same concentration, the output value D obtained by the acquisition step initially performed in each of the new execution times is used. t1 With the output value D t4 the output difference calculated by the difference between the output difference and the output difference, calculating the drift value of at least one of the specific analysis objects,

[0261] In the production step, the output value D is obtained by t3+ia The at least one drift value calculated by the new execution times that is the same as the new execution times and the output value D constituting the learning data obtained by each execution times constituting the new execution times for which the drift value is calculated t3+ia Establish a correspondence for each one and produce the learning data.

[0262] (Note 7)

[0263] The method for preparing learning data according to any one of Supplements 2 to 6, wherein:

[0264] The degree of deviation is represented by the difference between the output difference value of the specific analysis object in the initial execution of the acquisition step and the output difference value of the specific analysis object in each execution of the acquisition step, or is represented by the ratio of the output difference value of the specific analysis object in each execution of the acquisition step to the output difference value of the specific analysis object in the initial execution of the acquisition step.

[0265] (Note 8)

[0266] A method for predicting a characteristic is a method for predicting a characteristic of an analysis object, comprising:

[0267] An input step of inputting, into the prediction model, an output value D representing the characteristic of the analysis object at time t3+ia (i=0 to n-1: n is a predetermined integer greater than 2) obtained at predetermined intervals a within the period starting from time t3 which is the beginning of a period later than time t1 and earlier than time t2. t3+ia, the prediction model is a model that is generated by supervised machine learning using the learning data produced by the method for producing learning data described in Supplement 1 and predicts a K value, the K value being a value related to the concentration of the analyte, the time t1 being a time for controlling the introduction of the analyte into the reaction pool for introducing a buffer solution and the analyte, and the time t2 being a time predetermined as a time later than the time t1; and

[0268] Prediction step, according to the input output value D t3+ia The concentration of the analyte is predicted based on the predicted value of the K value output by the prediction model.

[0269] (Note 9)

[0270] A method for predicting a characteristic is a method for predicting a characteristic of an analysis object, wherein:

[0271] In the case of using a prediction model to predict the concentration of an analyte when the analyte is repeatedly introduced into a reaction cell for introducing a buffer solution and an analyte, the method for predicting the characteristics of the analyte comprises an input step and a prediction step.

[0272] The prediction model is a model that is generated by supervised machine learning using the learning data produced by the method for producing learning data described in Supplement 2 and predicts a K value that is a value related to the concentration of the analyte.

[0273] In the input step, an output value D representing the characteristic of the analysis object at time t3+ia (i=0 to n-1: n is a predetermined integer greater than 2) obtained at predetermined intervals a within the period starting from time t3 which is the beginning of the period later than time t1 and earlier than time t2 is input into the prediction model. t3+ia and a drift value, wherein the time t1 is a time at which control of introducing the analyte into the reaction cell is performed, the time t2 is a time predetermined as a time later than the time t1, and the drift value represents an output value D at the time t1 obtained by the specific analyte in the same number of times as the number of times the analyte is introduced into the reaction cell when the specific analyte having a predetermined specific concentration is repeatedly introduced into the reaction cell. t1 and the output value D at time t4 which is after time t3 and before time t2. t4 the output difference calculated by the difference between the output difference and the output difference obtained when the specific analyte is first introduced into the reaction cell,

[0274] In the prediction step, the output value D is input t3+ia and the predicted value of the K value output by the prediction model of the drift value to predict the concentration of the analyte.

[0275] According to Supplementary Note 1, there is an effect of being able to create learning data used in machine learning for generating a prediction model that repeatedly predicts the characteristics of an analysis object even without calibration.

[0276] According to Supplementary Note 2, there is an effect that the prediction accuracy of the K value is improved compared to the case where the K value is predicted using a prediction model that is machine-learned using learning data that does not include a drift value.

[0277] According to Supplementary Note 3, there is an effect that a drift value can be calculated from the K value of a specific analyte in each execution.

[0278] According to Supplementary Note 4, there is an effect that the measurement accuracy of the concentration of the analyte is improved compared to the case where a specific analyte is selected without considering the measurement accuracy characteristics of each concentration of the analyte in the analysis device.

[0279] According to Supplementary Note 5, there is an effect that the measurement accuracy of the concentration of the analyte is improved compared with the case where the concentration of the analyte is measured using a prediction model generated from learning data created by limiting the specific analyte to one analyte having a specific concentration.

[0280] According to Supplementary Note 6, there is an effect that the time required to prepare learning data can be shortened compared to the case where the output difference of the specific analysis object is calculated every time the acquisition step is executed.

[0281] According to Supplementary Note 7, there is an effect that a method for calculating the degree of deviation between the output difference value of the specific analysis object in the initial execution of the acquisition step and the output difference value of the specific analysis object in each execution of the acquisition step can be selected from multiple calculation methods according to the characteristics of the specific analysis object.

[0282] According to Supplementary Notes 8 and 9, there is an effect that the characteristics of the analysis object can be repeatedly measured even without calibration.

Claims

1. A method for preparing learning data, wherein: In an analysis device having a reaction pool, an electrode and a control unit, when the control unit uses a prediction model generated by supervised machine learning to predict a K value, the method for preparing learning data includes an acquisition step, a K value calculation step and a preparation step. The reaction pool is used for introducing a buffer solution and an analyte. The electrode is disposed in the reaction cell and, when a voltage of a predetermined magnitude is applied, outputs a response current of a magnitude corresponding to the concentration of the analyte in the buffer solution in liquid contact. The control unit controls the timing of introducing the buffer solution and the analyte into the reaction cell, and controls the magnitude of the response current output from the electrode as an output value representing a characteristic of the analyte. The K value is an output value D at time t1 when control is performed to introduce the analyte into the reaction cell after the introduction of the buffer solution. t1 The output value D at the time t2 which is predetermined as a time later than the time t1 is t2 The difference between the two represents a value related to the concentration of the analyte. In the acquisition step, for each of the plurality of analytes having different predetermined concentrations, the output value D at the time t1 is first acquired at least once. t1 , the output value D at time t3+ia at every predetermined interval a within the period from time t3 which is the beginning of the period after the time t1 and before the time t2 t3+ia , the output value D at time t2 t2 After the acquired processing is performed, the output value D of each of the analysis objects at the time t3+ia is obtained. t3+ia The process is repeated until a predetermined number of times, wherein i=0~n-1, n is a predetermined integer greater than 2, In the K value calculation step, the output value D of each of the analysis objects initially acquired in the acquisition step is calculated. t1 The output value D of each of the analysis objects t2 The difference between the K value of each of the analysis objects is calculated as the reference K value. In the production step, the output value D t3+ia The reference K value of the analyte having the same concentration as the analyte introduced into the reaction cell is relative to each output value D of each analyte acquired each time the acquiring step is executed. t3+ia Correspondence is established to create learning data used in supervised machine learning of the prediction model.

2. The method for preparing learning data according to claim 1, wherein: In each execution of the acquisition step, the output value D at time t4 is acquired for each of the analysis objects. t4 , and in each subsequent execution of the acquisition step, the output value D at the time t1 is acquired for each of the analysis objects. t1 , the time t4 is a time after the time t3 and before the time t2, The method for preparing learning data includes a calculation step and a drift value calculation step. In the calculation step, for a specific analyte having a predetermined specific concentration among the analytes, the output value D of the specific analyte obtained in each execution of the acquisition step is calculated. t1 With the output value D t4 The difference between the values ​​of the output of the specific analysis object in each execution of the acquisition step is calculated as the output difference of the specific analysis object in each execution of the acquisition step. In the drift value calculation step, a drift value is calculated for each execution of the acquisition step, the drift value indicating a degree of deviation between the output difference value of the specific analysis object in the first execution of the acquisition step and the output difference value of the specific analysis object in each execution of the acquisition step. In the creating step, the output values ​​D of each of the analysis objects in each execution of the acquiring step, including the output values ​​D of each of the analysis objects in each execution of the acquiring step, are obtained. t3+ia And get each output value D t3+ia The input data of the drift value in the same execution times will have the same value as that in order to obtain each output value D t3+ia The reference K value of the analyte having the same concentration as the analyte introduced into the reaction cell is associated as output data, and learning data used in the supervised machine learning of the prediction model is prepared.

3. The method for preparing learning data according to claim 2, wherein: The time t4 is set to the time t2.

4. The method for preparing learning data according to claim 3, wherein: The specific analyte is the analyte whose concentration estimated from the reference K value of each analyte is closest to the actual concentration of the analyte.

5. The method for preparing learning data according to claim 3, wherein: The specific analyte is a plurality of analyte having different concentrations that are pre-specified among the analyte. In the drift value calculation step, the drift value of each specific analyte is calculated each time the acquisition step is performed. In the creating step, the output values ​​D of each of the analysis objects in each execution of the acquiring step, including the output values ​​D of each of the analysis objects in each execution of the acquiring step, are obtained. t3+ia And get each output value D t3+ia The input data of the plurality of said drift values ​​in the same execution times will have the same value as that in order to obtain each output value D t3+ia The reference K value of the analyte having the same concentration as the analyte introduced into the reaction cell is associated as output data, and learning data used in the supervised machine learning of the prediction model is prepared.

6. The method for preparing learning data according to any one of claims 2 to 5, wherein: In the drift value calculation step, when a plurality of consecutive execution times of the acquisition step are collected and processed as a unit of new execution times of the acquisition step so that each of the new execution times includes a plurality of output difference values ​​of the specific analyte having the same concentration, the output value D obtained by the acquisition step initially performed in each of the new execution times is used. t1 With the output value D t4 the output difference calculated by the difference between the output difference and the output difference, calculating the drift value of at least one of the specific analysis objects, In the production step, the output value D is obtained by t3+ia The at least one drift value calculated by the new execution times that is the same as the new execution times and the output value D constituting the learning data obtained by each execution times constituting the new execution times for which the drift value is calculated t3+ia Establish a correspondence for each one and produce the learning data.

7. The method for preparing learning data according to claim 2, wherein: The degree of deviation is represented by the difference between the output difference value of the specific analysis object in the initial execution of the acquisition step and the output difference value of the specific analysis object in each execution of the acquisition step, or is represented by the ratio of the output difference value of the specific analysis object in each execution of the acquisition step to the output difference value of the specific analysis object in the initial execution of the acquisition step.

8. A method for predicting a characteristic, comprising: An input step of inputting, into the prediction model, an output value D representing the characteristic of the analysis object at time t3+ia obtained at predetermined intervals a within the period starting from time t3 which is the beginning of the period after time t1 and before time t2. t3+ia , the prediction model is a model that is generated by supervised machine learning using the learning data produced by the method for producing learning data according to claim 1 and predicts a K value, the K value being a value related to the concentration of the analyte, the time t1 being a time for controlling the introduction of the analyte into the reaction pool for introducing the buffer solution and the analyte, the time t2 being a time predetermined as a time later than the time t1, wherein i=0~n-1, and n is a predetermined integer greater than 2; and Prediction step, according to the input output value D t3+ia The concentration of the analyte is predicted based on the predicted value of the K value output by the prediction model.

9. A method for predicting a characteristic, which is a method for predicting a characteristic of an analysis object, wherein: In the case of using a prediction model to predict the concentration of an analyte when the analyte is repeatedly introduced into a reaction cell for introducing a buffer solution and an analyte, the method for predicting the characteristics of the analyte comprises an input step and a prediction step. The prediction model is a model that is generated by supervised machine learning using the learning data generated by the method for generating learning data according to claim 2 and predicts a K value, which is a value related to the concentration of the analyte. In the input step, the output value D representing the characteristic of the analysis object at time t3+ia obtained at predetermined intervals a within the period starting from time t3 which is the beginning of the period after time t1 and before time t2 is input into the prediction model. t3+ia and a drift value, wherein the time t1 is a time at which control of introducing the analyte into the reaction cell is performed, the time t2 is a time predetermined as a time later than the time t1, and the drift value represents an output value D at the time t1 obtained by the specific analyte in the same number of times as the number of times the analyte is introduced into the reaction cell when the specific analyte having a predetermined specific concentration is repeatedly introduced into the reaction cell. t1 and the output value D at time t4 which is after time t3 and before time t2. t4 The output difference calculated by the difference between the output difference and the output difference obtained when the specific analyte is initially introduced into the reaction cell, wherein i=0 to n-1, and n is a predetermined integer greater than 2, In the prediction step, the output value D is input t3+ia and the predicted value of the K value output by the prediction model of the drift value to predict the concentration of the analyte.

Images