Automated analysis device

Abstract

This automated analysis device comprises: a pump that feeds a mobile phase; a separating column through which the mobile phase fed by the pump and a specimen being analyzed are passed to separate components in the specimen; an acquiring unit that acquires pressure data when the pump feeds the mobile phase; a storage unit that stores a normal value of the pressure data; and a calculating unit that, each time a specimen is measured using the separating column, calculates first data indicating the degree of deviation of the pressure data from the normal value on the basis of a statistical method that takes into account the degree of similarity between the pressure data and the normal value, and calculates the remaining service life of the separating column on the basis of the first data. This makes it possible to provide an automated analysis device in which it is possible to predict more accurately the remaining service life of a separating column used in a liquid chromatograph.

Description

Automatic analyzer

[0001] The present invention relates to an automatic analyzer.

[0002] An automatic analyzer is a device that performs qualitative or quantitative analysis of specific components contained in biological samples such as blood, urine, cerebrospinal fluid, etc. (hereinafter sometimes referred to as specimens, samples, etc.). For example, in facilities such as hospitals and medical examination facilities that need to process a large number of patient specimens in a short time, it is an essential device. Some such automatic analyzers employ a liquid chromatograph, which is a technique for qualitatively and quantitatively analyzing components dissolved in a liquid sample by sending a liquid sample in which a plurality of components are mixed to a separation column by a liquid feed pump and separating the substance to be measured in the separation column.

[0003] As a prior art related to a liquid chromatograph, Patent Document 1 discloses a liquid chromatograph configured to record the liquid feed pressure detected by a pressure sensor of a liquid feed pump at a predetermined timing during a series of operations related to analysis for each analysis run of a sample.

[0004] International Publication No. 2020 / 170378

[0005] The separation column used in a liquid chromatograph gradually becomes clogged due to contamination or the like during use, and the pressure loss increases. Therefore, the pressure in the flow path of the liquid feed pump increases as the number of uses of the separation column increases. On the other hand, using the separation column in a high-pressure state can cause measurement errors and equipment failures, so a service life as a usage limit is set. Generally, the upper limit value of the increased flow path pressure is determined as the service life, and the separation column can be used until this value is reached. Also, since there is a risk of sample loss when the service life is reached during measurement, it is required to predict the remaining service life, which is the number of measurements from the measurement time point until the separation column reaches its service life, and to pre-stop the device for separating column replacement.

[0006] Clogging of the separation column is monitored based on the pressure in the flow path of the liquid delivery pump. However, pressure changes vary from column to column and depend on usage conditions, making it difficult to accurately predict the remaining lifespan of the separation column. Furthermore, there are various abnormal modes that affect pressure, and there is no obvious indicator that explicitly shows clogging of the separation column. In addition, it is difficult to see signs of clogging from pressure data during normal operation in the initial stages of use after the separation column has been installed, making it unsuitable for lifespan prediction. To improve the accuracy of remaining lifespan prediction of the separation column, it is necessary to select and use pressure data that is suitable for predicting clogging of the separation column.

[0007] The present invention has been made in view of the above, and aims to provide an automated analyzer that can more accurately predict the remaining lifespan of a separation column used in a liquid chromatograph.

[0008] The present invention includes multiple means for solving the above problems, but one example is a pump for delivering a mobile phase, a separation column for separating components in a sample by passing the mobile phase delivered by the pump through the sample to be analyzed, an acquisition unit for acquiring pressure data when the pump delivers the mobile phase, a storage unit for storing normal values ​​of the pressure data, and a calculation unit that, each time a sample is measured using the separation column, calculates first data indicating the degree of deviation of the pressure data from the normal value based on a statistical method that takes into account the degree of similarity between the pressure data and the normal value, and calculates the remaining life of the separation column based on the first data.

[0009] According to the present invention, it is possible to provide an automated analyzer that can more accurately predict the remaining lifespan of a separation column used in a liquid chromatograph.

[0010] A functional block diagram illustrating the overall configuration of the automated analyzer. A diagram showing an example of the remaining life calculation process. A flowchart showing the content of the first data calculation process. A diagram showing an example of the remaining life display. A flowchart showing the content of the remaining life calculation process. A diagram showing an example of the threshold input setting screen. A diagram showing an example of the threshold input setting screen. A flowchart showing the content of the remaining life calculation process. A diagram showing an example of machine learning input and output.

[0011] Embodiments of the present invention will be described with reference to the drawings. In the drawings used in these embodiments, similar or corresponding components are denoted by the same or similar reference numerals, and repeated descriptions of these components will be omitted as appropriate.

[0012] <First Embodiment> The first embodiment will be described with reference to Figures 1 to 6.

[0013] Figure 1 is a functional block diagram that schematically shows the overall configuration of the automated analyzer.

[0014] In Figure 1, the automated analyzer 100 is generally composed of a liquid chromatograph 1 and an acquisition unit 2.

[0015] The liquid chromatograph 1 mainly comprises a liquid delivery pump 11, a switching valve 12, a separation column 13, and a detector 14.

[0016] In the liquid chromatograph 1, the mobile phase 15 is delivered by the liquid delivery pump 11 to the downstream switching valve 12. At the switching valve 12, the mobile phase 15 and the sample 16 (specimen) to be analyzed merge and are introduced into the separation column 13. A portion of the mobile phase 15 containing the sample 16 passes through the separation column 13, separating the components within the sample, and the eluate is delivered to the detector 14. The detector 14 detects and analyzes the target substance from the eluate in the separation column 13.

[0017] The acquisition unit 2 includes a pressure sensor or equivalent function (not shown) provided on the liquid delivery pump 11, and acquires the discharge pressure of the liquid delivery pump 11 as digital data (pressure data).

[0018] The management device 3 is a computer composed of memory, a processor, a storage device, etc., and includes a memory unit 31, a calculation unit 32, a display unit 33, etc. The management device is also connected to input devices 4 such as a microphone, mouse, and keyboard for the user to input settings and various information, and output devices 5 such as a display, projector, or speaker that displays analysis results and various information to the user. A touch panel display or the like in which the input device 4 and output device 5 are integrated may also be used. The storage device functions as the memory unit 31, and stores various data including various programs, settings used by the programs, and pressure data. The memory is loaded with the functions of the calculation unit 32 and the display unit 33 as programs and executed by the processor. The memory unit, calculation unit, and output devices may be located within the automatic analyzer 100.

[0019] The memory unit 31 stores normal data, which is pressure data during normal operation. Here, normal data refers to pressure data obtained during initial measurements in a state where there are no malfunctions in the device components that affect the performance of the automatic analyzer 100, and where it is assumed that the separation column 13 is not clogged or only slightly clogged immediately after installation.

[0020] The calculation unit 32 performs a remaining life calculation process (described later) to calculate (predict) the remaining life of the separation column 13 based on the normal data stored in the storage unit 31 and the pressure data acquired by the acquisition unit 2. Here, the remaining life of the separation column 13 is the number of measurements remaining until the separation column 13 reaches the end of its life. The separation column 13 has a predetermined upper limit for the pressure in the flow path of the liquid transfer pump 11, and it is determined whether or not the end of its life has been reached by comparing this upper limit with the pressure data obtained in one measurement. Specifically, it is determined that the end of its life has been reached if the maximum value of the pressure data obtained in one measurement reaches that upper limit.

[0021] The display unit 33 outputs the result of the remaining lifespan calculation process calculated by the calculation unit 32 to the output device 5 and presents it to the user.

[0022] Figure 2 shows an example of the remaining lifespan calculation process.

[0023] In Figure 2, when measurement is performed by the automatic analyzer 100, the acquisition unit 2 first acquires pressure data 21, which is a time series of pressure values ​​discharged from the liquid delivery pump 11 at the time of measurement, using a pressure sensor or the like (processing 201).

[0024] Furthermore, the calculation unit 32 receives pressure data 21 from the acquisition unit 2 and uses the normal data 22 stored in the storage unit 31 to calculate first data 23 indicating the degree of deviation from the pressure data during normal operation (processing 202). Based on the pressure data 21, the first data 23, and a threshold 24 set in advance by the user from the input device, the remaining lifespan 25 of the separation column is calculated (processing 203).

[0025] Here, we will explain how to calculate the first data 23.

[0026] The first data 23 is calculated based on a statistical method that takes into account the degree of similarity between the input pressure data 21 and the pressure data during normal operation (normal data 22).

[0027] First, a partial time series is extracted from the pressure data 21 with a predetermined window width to obtain partial pressure data. Here, the length (window width) of the partial pressure data is an arbitrary natural number that is within the range where the pressure value is stable (flat) and does not exceed the length of the pressure data 21. By optimizing the length of the partial pressure data, the first data can more sensitively detect anomalies.

[0028] Next, a machine learning model that outputs the last pressure value of the obtained partial pressure data is used to calculate the prediction error for the extracted partial pressure data. Here, the machine learning model is trained using normal data 22 stored in the memory unit 31. The normal data 22 to be trained can also be divided according to the solvent mixing ratio and column type at the time of measurement. Since the degradation trends of columns may differ, the prediction accuracy can be improved by classifying the normal data 22 in advance according to the solvent mixing ratio and column type at the time of measurement. On the other hand, data management becomes more complicated.

[0029] Then, the prediction error is repeatedly calculated while shifting the position from which the partial pressure data is extracted, and the pressure data 21 is converted into a prediction error series. The first data 23 is calculated by averaging the values ​​of the obtained prediction error series.

[0030] As a machine learning model, a well-known or publicly known method that can calculate the prediction error from the input partial pressure data and the model's output can be applied. For example, the flow of the first data calculation part when a Gaussian process regression model is used as the machine learning model will be explained with reference to steps S300 to S330 in Figure 3. When a pressure waveform of length I {v_1, ..., v_i, ... v_I} is given to the Gaussian process regression model, the model input and output are: input: x_i = [v_(i - (k-1)), ..., v_i]^T ∈ R^K, output: y_i = v_i. Here, K is the length of past observation data to be considered, where (I - K + 1) ≤ i ≤ L. Based on this model input and output, the Gaussian process regression model is trained to estimate the currently observed pressure value from the pressure values ​​observed from (K-1) time before to the present. The mean and variance of the output are estimated from the input partial pressure data in the trained model, and the mean of the standard score calculated from the estimated mean and variance can be used as the first data 23. Since the Gaussian process regression model can accurately estimate the output if the input is included in the training data, the first data 23 is expected to have small values ​​for partial pressure data similar to the normal data 22 and large values ​​for partial pressure data that is not similar to the normal data 22. The Gaussian process is a model used in supervised machine learning such as regression analysis and classification, and it has flexibility to the complexity of the model and robustness to overfitting, so it can handle cases where column degradation is an event involving multiple factors. In addition, when using a large amount of training data, the computational complexity can be reduced by using stochastic variational Gaussian process regression (GPR).

[0031] It should be noted that the output of a machine learning model is not limited to the above and can be appropriately modified depending on the machine learning model used. For example, when using "Auto Encoder" (GE Hinton, et al., Science, 2006) as a machine learning model, the model parameters are learned to produce an output that accurately reproduces the input partial pressure data. Therefore, if the average distance between the input partial pressure data and the output is taken as the first data 23, the first data 23 is expected to be a small value for partial pressure data similar to the normal data 22 and a large value for partial pressure data that does not resemble the normal data 22.

[0032] The display unit 33 receives the first data 23 or remaining lifespan 25 calculated by the calculation unit 32 and sends it to the output device 5 for presentation to the user (processing 204). Here, the output device 5 can display the first data 23 or remaining lifespan 25 on the display, or play the remaining lifespan 25 and related warnings as audio using a speaker. Figure 4 shows an example of the display of the remaining lifespan 25. The display may, for example, show the remaining lifespan 25 of the columns installed in the management device 3, or the remaining lifespan 25 and the first data 23 in a list, as shown in Figure 4. Also, if there is a column with a short remaining lifespan 25, a warning message may be displayed as shown in Figure 4. By reliably communicating the remaining lifespan of the column to the user through display and warnings, it is possible to avoid reaching the end of the column's lifespan during sample measurement.

[0033] Figure 5 is a flowchart showing the contents of the remaining life calculation process in the calculation unit.

[0034] First, the calculation unit 32 receives pressure data 21 from the acquisition unit 2 and normal data 22 from the storage unit 31 (step S100).

[0035] Next, using the normal data 22, the first data 23 is calculated from the pressure data 21 (step S110).

[0036] Next, it is determined whether the calculated first data 23 is equal to or greater than a preset threshold 24 (step S120). If the determination result is NO, the measurement and acquisition of pressure data in step S100 and the calculation of the first data 23 in step S110 are repeated until the first data 23 is equal to or greater than the threshold 24.

[0037] Furthermore, if the result of the determination in step S120 is YES, that is, if the calculated first data 23 is greater than or equal to the threshold 24, it is determined that the measured pressure data 21 is not normal, and as a prediction of the remaining life of the separation column 13 (calculation of remaining life), the pressure data 21 of the current measurement is first saved as stored data (step S130).

[0038] Then, the remaining lifespan is predicted from the accumulated data (step S140), and the process is terminated. Note that the processes from steps S100 to S140 are executed repeatedly, and after the completion of the process in step S140, the pressure data for the next measurement is received in step S100 and the process continues.

[0039] Here, we will explain the method for predicting remaining lifespan.

[0040] One method for predicting remaining lifespan is to perform linear regression on a series of maximum pressure values ​​for each measurement. A series of maximum pressure values ​​for each measurement is extracted from the stored pressure data 21, and linear regression parameters are estimated for the extracted series. Based on the estimated parameters, the number of measurements required to reach the end of life is predicted. The remaining lifespan is calculated by subtracting the number of measurements taken to date from the obtained number of measurements.

[0041] Furthermore, the method for predicting remaining lifespan is not limited to the above; well-known or publicly recognized statistical methods such as nonlinear regression may be applied.

[0042] Also, the threshold value 24 used in the remaining life prediction (see step S120 in FIG. 5) is determined by the user referring to the prior experimental results. First, a plurality of durability tests are performed to measure from when the separation column 13 is attached until the end of its life, and pressure data is collected. Next, the first data 23 is calculated for the pressure data of each durability test, and a graph showing its transition is presented to the user on the output device 5 (such as a display). The user determines the threshold value by looking at the presented graph and inputs and sets the threshold value from the input device 4 (such as a keyboard).

[0043] FIGS. 6 and 7 are diagrams showing an example of an input setting screen for the threshold value.

[0044] As illustrated in FIGS. 6 and 7, in setting the threshold value related to the remaining life prediction, the presentation of the first data 23 to the user and the input / setting of the threshold value by the user can be performed through the GUI of the display unit 33.

[0045] For example, as shown in FIG. 6, the input setting screen can be provided with a transition 401 of the first data and an input setting unit 402 for the score threshold value. The user of the automatic analyzer 100 can know the transition 401 of the first data for each measurement from the information presented on the output device 5 (such as a display), and can set it by inputting an arbitrary score threshold value from the input device 4 (such as a keyboard).

[0046] Also, for example, as shown in FIG. 7, the input setting screen can be provided with a relationship 501 between the maximum pressure value and the first data and an input setting unit 502 for the threshold value. The user of the automatic analyzer can know the relationship 501 between the maximum pressure value and the first data from the information presented on the output device 5 (such as a display), and can also set a threshold value that is intuitively easy for humans to understand by using the maximum pressure value when the first data exceeds a predetermined value as the threshold value.

[0047] In the present embodiment configured as described above, it is possible to more accurately predict the remaining life of the separation column used in the liquid chromatograph. That is, since it is configured to be able to set an appropriate threshold value, it is possible to remove the pressure data during normal operation, which is considered to have little contribution to the prediction of the remaining life of the separation column, during the prediction of the remaining life, and high remaining life prediction accuracy can be achieved. Further, by predicting the remaining life of the separation column from the accumulated data based on a statistical method, it is possible to perform a prediction of the remaining life considering the individual differences and usage conditions of the separation column.

[0048] <Second Embodiment> The second embodiment will be described while referring to FIG. 8.

[0049] This embodiment shows a case where the user sets two threshold values related to the prediction of the remaining life. In this embodiment, the same members as those in the first embodiment are denoted by the same reference numerals, and the description will be omitted as appropriate.

[0050] FIG. 8 is a flowchart showing the content of the calculation process of the remaining life in the calculation unit of this embodiment.

[0051] First, the calculation unit 32 receives the pressure data 21 from the acquisition unit 2 and receives the normal data 22 from the storage unit 31 (step S200).

[0052] Next, the first data 23 is calculated from the pressure data 21 using the normal data 22 (step S210).

[0053] Subsequently, it is determined whether or not there has been one or more times until the previous measurement when the calculated first data 23 becomes equal to or greater than a preset first threshold value (step S220). The first threshold value is set in the same manner as the threshold value 24 in the first embodiment.

[0054] If the determination result in step S220 is NO, it is determined whether or not the first data is equal to or greater than the first threshold value (step S221).

[0055] If the result of the determination in step S221 is YES, the current measurement pressure data is saved as stored data (step S231), and the process returns to step S200 to perform the next measurement and acquire the next pressure data. If the result of the determination in step S221 is NO, the process returns to step S200 to perform the next measurement and acquire the next pressure data.

[0056] Furthermore, if the result of the determination in step S220 is YES, it is determined whether the first data is greater than or equal to a pre-set second threshold (step 606). The second threshold is set by the user in the same way as the first threshold (i.e., the threshold 24 in the first embodiment), and is set as a value greater than the first threshold.

[0057] If the result of the determination in step S230 is NO, the process proceeds to step S231 to save the current measurement pressure data as stored data, and then the process returns to step S200 to perform the next measurement and acquire the pressure data.

[0058] Furthermore, if the determination result in step S230 is YES, the remaining lifespan of the separation column 13 is predicted from the accumulated data using the trained machine learning model (step S240), and the process is terminated. Note that the processes in steps S200 to S231 are executed repeatedly, and after the completion of the process in step S240, the pressure data for the next measurement is received in step S200 and the process continues.

[0059] Here, we will explain how to obtain a pre-trained machine learning model.

[0060] The trained machine learning model is trained using pre-training data obtained from preliminary experiments. First, multiple durability experiments are conducted in which measurements are taken from the time the separation column 13 is installed until it reaches the end of its lifespan, and pressure data is collected. A remaining lifespan prediction is performed once for each durability experiment, and a machine learning model is trained that accepts pressure data from the first threshold to the second threshold as input for the first data, and outputs the remaining lifespan when the first data exceeds the second threshold.

[0061] As a machine learning model, for example, a linear regression model can be used to calculate an approximate straight line of the series of maximum pressure values ​​for each measurement from the input pressure data, and the remaining life can be regressioned from its slope. However, the machine learning model is not limited to the above, and any well-known or publicly available machine learning model such as a recurrent neural network can be applied. For example, as shown in Figure 9, the input layer can use values ​​such as the maximum pressure value of the measurement data, the type and mixing ratio of the mobile phase at the time of measurement, the type of column, the pressure during column conditioning (preparation run) before the start of measurement, the alarm occurrence history, and the pressure during column cleaning for each measurement, while the output layer can output values ​​such as the remaining life and the type of column degradation.

[0062] The other configurations are the same as in the first embodiment.

[0063] In this embodiment configured as described above, the same effects as in the first embodiment can be obtained.

[0064] Furthermore, the remaining lifetime of the separation column can be predicted by considering the pressure increase trend obtained from preliminary experiments, and high lifetime prediction accuracy can be achieved when sufficient prior training data is available. In addition, even if the first data shows a complex increasing trend after exceeding the first threshold, the lifetime prediction accuracy can be improved by setting an appropriate second threshold.

[0065] <Note> The present invention is not limited to the embodiments described above, and includes various modifications and combinations that do not depart from the spirit of the invention. Furthermore, the present invention is not limited to having all the configurations described in the embodiments described above, and includes those in which some of the configurations have been omitted.

[0066] 1...Liquid chromatograph, 2...Acquisition unit, 3...Control device, 4...Input device, 5...Output device, 11...Liquid pump, 12...Switching valve, 13...Separation column, 14...Detector, 15...Mobile phase, 16...Sample, 21...Pressure data, 22...Normal data, 23...First data, 24...Threshold, 25...Remaining life, 31...Storage unit, 32...Calculation unit, 33...Display unit, 100...Automatic analyzer, 401...Changes in first data, 402...Input setting unit, 501...Relationship between maximum pressure value and first data, 502...Input setting unit

Claims

1. An automated analyzer comprising: a pump for delivering a mobile phase; a separation column for separating components within a sample by passing the mobile phase delivered by the pump through the sample to be analyzed; an acquisition unit for acquiring pressure data when the pump delivers the mobile phase; a storage unit for storing normal values ​​of the pressure data; and a calculation unit that, each time the sample is measured using the separation column, calculates first data indicating the degree of deviation of the pressure data from the normal value based on a statistical method that takes into account the degree of similarity between the pressure data and the normal value, and calculates the remaining life of the separation column based on the first data.

2. An automated analyzer according to claim 1, wherein the calculation unit starts calculating the remaining lifespan of the separation column when the first data exceeds a predetermined value.

3. An automatic analyzer according to claim 1, wherein the calculation unit has a normal value model learned from the normal value stored in the storage unit, and the pressure data calculates a value corresponding to the error or distance generated from the normal value model and uses it as the first data.

4. An automatic analyzer according to claim 2, characterized in that the calculation unit calculates the remaining life of the separation column based on pressure data from the time when the first data exceeds a predetermined value.

5. An automated analyzer according to claim 2, wherein the calculation unit has a remaining life prediction timing that predicts the remaining life of the separation column after the first data exceeds the predetermined value, and the remaining life of the separation column is calculated using a trained machine learning model from the pressure data from the time the predetermined value is exceeded to the remaining life prediction timing.

6. An automatic analyzer according to claim 1, characterized in that it is equipped with an output device for notifying the user of the first data.

7. An automatic analyzer according to claim 2, characterized in that it is equipped with an input device for a user to input the predetermined value.

Images