Automatic analysis device, recommended action notification system, and recommended action notification method

Abstract

The present application provides a system that recommends an appropriate response to an operator from an abnormality precursor stage. The recommended action notification system includes a plurality of automatic analysis devices (51) including a first automatic analysis device, and a learning device (52) connected through a network (53) that recommends an action to be performed on the first automatic analysis device, including: a processing section (63) that receives specimen analysis result data or maintenance result data from the first automatic analysis device, inputs associated device data including the specimen analysis result data or the maintenance result data to a learning model (72), and recommends a prescribed action to be performed on the first automatic analysis device to an operator in a case where a probability value output by the learning model that the prescribed action is to be performed is equal to or higher than a prescribed threshold value; and an updating section (71) that updates the learning model based on a learning data set from the plurality of automatic analysis devices.

Description

Technical Field

[0001] This invention relates to an automatic analysis device, a recommended action notification system for informing an operator of recommended actions for the automatic analysis device, and a recommended action notification method. Background Technology

[0002] In recent years, automatic analysis devices have increasingly used alarms to notify operators of detected anomalies, and these devices often include pre-set response suggestions for these alarms. However, almost all of them suggest multiple response methods, requiring the operator to collect device data to determine the cause and then select the appropriate response from the suggested options.

[0003] Patent document 1 describes how, in order to easily trace the cause of poor accuracy management, information related to the measurement data acquired by the specimen analysis device and information related to the operation of the specimen analysis device are displayed together through time series charts, etc., to generate data that allows the operator to clearly understand the status of the specimen analysis device. Existing technical documents Patent documents

[0004] Patent Document 1: Japanese Patent Application Publication No. 2019-174424 Summary of the Invention The technical problem that the invention aims to solve

[0005] In order to determine the appropriate response based on alarms from the automatic analysis device, operators need knowledge of what device data is required to determine the cause, as well as sufficient experience to infer the cause from the collected device data, consider the operating conditions, and determine the appropriate response method for the device. Inexperienced operators may sometimes find it difficult to choose the right response method.

[0006] According to the method described in Patent Document 1, by displaying information such as the accuracy management of standard samples, which was previously only displayed on a screen or paper, in conjunction with information related to the operation of the sample analysis device, an experienced operator can easily and intuitively determine whether the poor accuracy management is due to the device itself. However, for less experienced operators, it may still be difficult to deduce the cause and take appropriate measures based on the displayed information.

[0007] For example, in automated analysis devices, the measurement and maintenance results of samples are monitored. If they exceed a preset threshold, an anomaly is detected, and the operator is notified. However, the reasons for measurement results exceeding the threshold could be varied, including abnormalities in the sample, reagents, or various mechanisms of the device. Therefore, simply confirming the detected abnormal data makes it difficult to determine the cause and decide on the appropriate response. In the past, the response methods notified to the operator during anomaly detections included all possible solutions for the various possible causes.

[0008] In view of the above problems, the present invention provides a system and method for recommending appropriate coping methods to operators starting from the stage of abnormal signs. Technical solutions to solve technical problems

[0009] One embodiment of the present invention provides a recommended action notification system comprising a plurality of automatic analysis devices, each containing a first automatic analysis device; and a learning device connected via a network to the plurality of automatic analysis devices, which recommends actions to be performed by the first automatic analysis device to an operator, including: The processing unit receives specimen analysis result data or maintenance result data from the first automatic analysis device, inputs associated device data containing the specimen analysis result data or maintenance result data into the learning model, and recommends to the operator to perform the prescribed action on the first automatic analysis device if the probability value of the recommended action output by the learning model is above a predetermined threshold; and The update unit updates the learning model based on the learning dataset from multiple automated analysis devices. Invention Effects

[0010] A system and method are provided to recommend appropriate coping methods to operators starting from the stage of abnormal signs.

[0011] Other issues and new features become clear from the description and accompanying drawings in this specification. Attached Figure Description

[0012] Figure 1A This is a diagram showing the overall situation of the recommended action notification system. Figure 1B This is a functional block diagram of the recommended action notification system. Figure 2 This is a schematic diagram of an automatic analysis device. Figure 3 This is a flowchart of the learning model update process. Figure 4 This is an example of a motion evaluation input screen. Figure 5 This is an example of associated device data. Figure 6This is an example of a learning model based on neural networks. Figure 7 This is an example of using tables. Figure 8 This is a flowchart of the process of using a learning model to output recommended actions. Figure 9 This is an example of a recommended action display screen. Figure 10 This is an example of a learning model for the action "changing the light source". Figure 11 This is an example of photometer inspection data, i.e., time-series absorbance data. Figure 12 This is an example of reaction monitoring data. Figure 13 This is an example of a recommended action display screen. Figure 14 This is an example of a learning model for the action "cleaning the outer wall of the sample probe". Figure 15 This is an example of measurement results when the outer wall of the sample probe was not cleaned. Figure 16A This is an example of reaction monitoring data (ALB reaction process) when the outer wall of the sample probe is not cleaned. Figure 16B This is an example of reaction monitoring data (CRE reaction process) when the outer wall of the sample probe is not cleaned. Figure 17 This is an example of a recommended action display screen. Detailed Implementation

[0013] Figure 1A This is an overall diagram of the recommended action notification system 50 of this embodiment. The recommended action notification system 50 includes multiple automatic analysis devices 51a-z and a learning device 52 connected via a network 53. The automatic analysis devices 51 are installed in clinical examination rooms, clinical examination centers, etc. There is no particular limitation on the number of automatic analysis devices 51 connected to the learning device 52. The hardware of the learning device 52 is an information processing device including storage devices such as HDDs (hard disk drives) and SSDs (solid state drives). (Further details will follow later.) Figure 1BIn this embodiment, the functions performed by the automatic analysis device 51 or the learning device 52 are represented in the form of a functional block diagram. These functions are implemented by reading the software program code into main memory and executing the read program code by the processor. In this embodiment, the functions implemented by the information processing device through a program are sometimes referred to as "sections". Furthermore, when the data processed by the learning device 52 is large-scale data, the data itself does not need to be stored in the storage device of the learning device 52, but can be stored, for example, in object storage in the cloud, along with the data path used to access the object data. The learning device 52 can be implemented by a single server or by a distributed processing server, and is not limited to the physical hardware structure.

[0014] Figure 1B This is a functional block diagram showing an outline of the recommended action notification system 50. Details will be described later. In the automatic analysis device 51, the dataset generation unit 61 generates a learning dataset that includes the action's effect information and associated device data ("associated device data") based on the action's effect information input by the device operator from the input unit 11 and the device data stored in the database 62. Furthermore, as described later, actions triggered by the anomaly detection unit 60 detecting an anomaly based on specimen measurement results or maintenance results and notifying the operator (e.g., displaying it on the display unit 10) are used as the objects for generating the learning dataset.

[0015] The learning dataset generated by the dataset generation unit 61 is sent to the learning device 52 via the network 53. In the learning device 52, the update unit 71 generates a teacher signal based on the learning dataset and updates the learning model 72. The updated learning model 72 is then published to the automatic analysis device 51 via the network 53. On the other hand, whenever a specimen measurement result or maintenance result is output, the automatic analysis device 51 inputs the associated device data stored in the database 62 into the learning model 72 via the processing unit 63 and outputs a recommended action. The recommended action is displayed on the display unit 10 of the automatic analysis device 51 to notify the operator.

[0016] Figure 2The diagram shows an outline of the automatic analysis device 51. The automatic analysis device 51 mainly includes a sample tray 1, a reagent tray 2, a reaction tray 3, a reaction tank 4, a sampling mechanism 5, a pipetting mechanism 6, a stirring mechanism 7, a photometric mechanism 8, a cleaning mechanism 9, a display unit 10, an input unit 11, a storage unit 12, and a control unit 13. In the sample tray 1, multiple sample containers 16 containing collected samples are fixed and arranged on the circumference of a circular disk 17. The circular disk 17 is circumferentially rotated in a positionable manner by a drive mechanism consisting of a motor, a rotating shaft, etc. (not shown). In the reagent tray 2, multiple reagent bottles 18 containing reagents for mixing and reacting with the samples are fixed and arranged on the circumference of a circular disk 19, surrounded by a temperature-controlled cold storage chamber 20. Furthermore, the circular disk 19 is circumferentially rotated in a positionable manner by a drive mechanism consisting of a motor, a rotating shaft, etc. (not shown). Multiple reaction vessel holders 22 are mounted on the reaction tray 3. These holders hold reaction vessels 21 for containing samples and reagents and are intermittently transferred via a drive mechanism 23 that rotates and stops in a constant cycle. The reaction tank 4 is a thermostatic bath positioned along the movement path of the reaction vessels 21. To promote the chemical reaction between the samples and reagents, the reaction solution containing the sample and reagent mixture within the reaction vessels 21 is maintained at a constant temperature, for example, using temperature-controlled water. The reaction vessels 21 move within the reaction tank 4. The sampling mechanism 5 comprises a probe 24, an arm 26 mounted on a support shaft 25, and a drive mechanism capable of reciprocating between the sample tray 1 and the reaction tray 3 about the support shaft 25. Following a predetermined sequence, it provides samples transferred to a fixed position in a sample container 16 by the rotation of the sample tray 1 to the reaction vessels 21. Similarly, the pipetting mechanism 6 includes a probe 27, an arm 29 mounted on a support shaft 28, and a drive mechanism capable of reciprocating between the reagent tray 2 and the reaction tray 3 with the support shaft 28 as the rotation center. Following a predetermined sequence, it supplies reagents transferred to a fixed position in the reagent bottle 18 via the rotation of the reagent tray 2 to the reaction vessel 21. Additionally, different types of samples and reagents are placed into their respective sample containers 16 and reagent bottles 18, and the required amounts are supplied to the reaction vessel 21. The stirring mechanism 7 is used to stir and mix the samples and reagents transferred to the reaction vessel 21 at this position (stirring position). The stirring mechanism 7 is a non-contact stirring mechanism, including: a fixing part 31 that fixes a piezoelectric element (not shown) to a position where sound waves can be irradiated from the side of the reaction vessel 21 at the stirring position; a piezoelectric element driver 14 that drives the piezoelectric element; and a stirring mechanism controller 15.

[0017] The control unit 13 sends control signals to the sample tray 1, reagent tray 2, reaction tray 3, reaction tank 4, sampling mechanism 5, pipetting mechanism 6, stirring mechanism 7, photometric mechanism 8, and cleaning mechanism 9 to control their respective actions and execute the function of the recommended action notification system of this embodiment. Specifically, it executes... Figure 1B The anomaly detection unit 60, dataset generation unit 61, and processing unit 63 are shown to perform their functions. The display unit 10 displays various screens such as analysis items and analysis results, while the input unit 11 accepts input of various information such as analysis items. The storage unit 12 stores various information such as predetermined sequences (programs) and analysis items used to control each mechanism. Figure 1B The updated learning model 72 published by the database 62 or learning device 52 shown is also stored in the storage unit 12.

[0018] Figure 3 This is the update flowchart for learning model 72. Furthermore, since the input and output of learning model 72 differ depending on the learned action, it is recommended that the action notification system 50 generate the learning model separately for each action. The action being learned is selected within the system, but it is assumed that the selected action is an operator action performed when the device malfunctions, such as maintenance, reagent replacement, or sample dilution and remeasurement.

[0019] The learning model is set to use the device anomaly data pre-collected by the recommended action notification system 50, the recommended action corresponding to the device anomaly data, and the anomaly cause as teacher signals for initial learning.

[0020] If the anomaly detection unit 60 of the automatic analysis device 51 detects certain anomalies from the results of the specimen inspection or maintenance performed by the device, it notifies the operator of an alarm (S10). The operator performs an action to deal with the anomaly (S01). The control unit 13 of the automatic analysis device records the operator's actions as the device's action log (S11). If it is determined from the action log that the operator has performed an action that is a learning target, the dataset generation unit 61 of the automatic analysis device displays an input screen on the display unit 10 after the action is performed, which evaluates the effectiveness of the operator's action (S12). Thus, this flowchart applies to situations where the anomaly detection unit 60 detects the occurrence of a certain anomaly in the device, and the operator uses this as a trigger to perform an action. Actions performed during routine maintenance are not included in the learning dataset. This is because by collecting only cases where the effect of an action is clear as the learning dataset, learning can be optimized.

[0021] Figure 4An example of the action evaluation input screen 80 is shown. The action performed by the operator is displayed on the execution content display unit 81 (the specific name of the action performed by the operator is displayed in the [Action Name] field in the figure). The operator inputs the effectiveness evaluation of the action by selecting any of the effect buttons 82 (S02). In addition, the cause of the abnormality determined by the operator's action is input to the abnormality cause selection unit 83 (S02). In order to reduce the operator's input effort and suppress performance deviation, the input is performed by selection. The action evaluation and abnormality cause input by the operator are recorded in the storage unit 12 of the automatic analysis device 51 (S13).

[0022] The dataset generation unit 61 generates a learning dataset for each action, which includes operator feedback, reasons for errors, and related device data (S14). The generated learning dataset is sent to the learning device 52 (S15).

[0023] The associated device data is collected within a pre-defined period based on the date and time of the operator's action, in order to learn the model's updates. Figure 5 The example shown is related to device data. Although termed device data, it is not limited to the device's structure and hardware, but broadly includes information on the structure, hardware, reagents, samples, and analytical results data related to the device's operation. Different device data indicate different precursors depending on the effective execution of the operator's action (the cause of the abnormality); therefore, associated device data is assigned for each action. Figure 5 The document summarizes and illustrates the types of associated device data for each action, categorized into dataset groups. A dataset group is obtained by grouping actions according to their content. The three dataset groups, "Device," "Reagent," and "Specimen," respectively contain actions related to devices, reagents, and specimens, and the types of associated device data collected show a high degree of commonality.

[0024] Furthermore, the dataset used for learning preferably contains time information related to the action. This time information includes the date and time of the action's execution and the date and time of the occurrence of any abnormalities (errors) detected by the anomaly detection unit 60 that indicate the action's execution is valid.

[0025] The updating unit 71 of the learning device 52 receives such a learning dataset (S21) and updates the learning model 72 using teacher signals based on the received learning dataset (S22). The updated learning model is published to the automatic analysis device 51 at a predetermined time.

[0026] Any learner, such as a neural network, regression tree, or Bayesian recognizer, can be used to learn the model 72. Figure 6The diagram illustrates an example of a neural network-based learning model. In learning model 90, information input to input layer 91 is transmitted to intermediate layer 92, and then sequentially to output layer 93, outputting an inference result based on information input from output layer 93 to input layer 91. Input layer 91, intermediate layer 92, and output layer 93 each have multiple input units, intermediate units, and output units, marked with circles. Furthermore, neural networks typically have multiple intermediate layers, represented in the diagram as intermediate layer 92. Information input to each input unit of input layer 91 is weighted by coupling coefficients between the input units and intermediate units and then input to each intermediate unit. The value of the intermediate unit in intermediate layer 92 is calculated by summing the values ​​from the input units. Similarly, the outputs from each intermediate unit in intermediate layer 92 are weighted by coupling coefficients between the intermediate units and output units and then input to each output unit. The value of the output unit in output layer 93 is calculated by summing the values ​​from the intermediate units. Thus, the processing in intermediate layer 92 is equivalent to performing a non-linear transformation on the input data input to input layer 91 and outputting it as the output data of output layer 93.

[0027] Figure 6 The learning model 90 shown is a learning model for the action "reaction unit cleaning," and is illustrated along with the input and output data examples. Associated device data (or its processing data, hereinafter referred to as associated device data without distinction) is input to the input layer 91, and the corresponding calculation results are output from the output layer 93. The associated device data input to the input layer 91 includes time-series measurement results 94 such as analysis results and maintenance data defined for each learning model, and analytical information 95 such as reagent information and specimen information that affect the action. The output data from the output layer 93 is the reasoning result related to the action "reaction unit cleaning," and includes action-related information 97 and abnormality causes 98. The action-related information 97 includes whether the action (here, "reaction unit cleaning") is recommended, the average remaining time until an error occurs, and the average remaining time until the action is performed. The average remaining time until an error occurs refers to the average remaining time required until the abnormality detection unit 60 detects the abnormality (error) corresponding to the action, and the average remaining time until the action is performed refers to the average remaining time required until the operator performs the action. The mean time remaining until an error occurs is a standard for knowing when the necessity of an action will arise, even if the action is not recommended. The mean time remaining until the action is performed is a standard for knowing when the action will generally be performed, taking into account the experience of users and operators in each facility.

[0028] The output values ​​of "Recommendation for Reaction Unit Cleaning (Action)" and "Cause of Abnormality (Unit Contamination)" are probability values. Whether an output probability value is above or below a predefined threshold determines whether it is a recommended action or a cause of an abnormality. For example, the closer the output value of "Recommendation for Reaction Unit Cleaning" is to 1, the more recommended it is; the closer it is to 0, the less recommended it is. Similarly, the closer the output value of "Cause of Abnormality (Unit Contamination)" is to 1, the more reliable it is; the closer it is to 0, the more suspicious it is. On the other hand, the output values ​​of "Average Remaining Time Until Error Occurs" and "Average Remaining Time Until Action is Executed" are the values ​​of the remaining time itself.

[0029] In the learning model update process (S22), the update unit 71 generates a teacher signal corresponding to the input and output data of the learning model 90 based on the learning dataset. The input value of the teacher signal is the association device data. On the other hand, in this example, the output value of the teacher signal is as follows: If the effectiveness evaluation of the action is effective, the output value of "Recommend Cleaning Response Unit" is 1; if it is ineffective, the output value of "Recommend Cleaning Response Unit" is 0. In addition, if the cause of the anomaly is "Unit Contamination", the output value of "Cause of Anomaly (Unit Contamination)" is 1; otherwise, the output value of "Cause of Anomaly (Unit Contamination)" is 0. Furthermore, regarding time information, the elapsed time from a predetermined reference time, which is the date and time of the error occurrence and the date and time of the action execution, is calculated and used as the output value of the teacher signal. The predetermined reference time may be, for example, the date and time of the last action execution.

[0030] The update unit 71 inputs the generated teacher signal input value (associated device data) to the input layer 91 and adjusts the value of the coupling coefficient between the units so that the value output from the output layer 93 becomes the output value of the teacher signal.

[0031] To improve the accuracy of the learning model 72, a large training dataset is needed. However, to enhance the learning effect, it is desirable to use a training dataset from an approximate automated analysis device 51 to construct the learning model. Each automated analysis device 51 differs in its application and inspection content, and like its approximate automated analysis devices, it shares common anomaly occurrence trends. Therefore, its methods for handling anomalies are more relevant. For example, automated analysis devices in facilities with long operating times require more frequent lamp replacements, or automated analysis devices in facilities that analyze both serum and plasma require more frequent probe cleaning compared to facilities that only analyze serum. The anomaly occurrence trends differ depending on the device's application.

[0032] Therefore, by setting the learning model 72 to be further subdivided into learning models for each group of facilities with similar usage conditions, the accuracy of the learning model can be further improved. Figure 7This is an example of using Table 100. The learning device 52 collects information on the usage status of the automatic analysis device for each facility and manages it using Table 100. Typical usage examples are defined as groups (group A is shown here). Based on the similarity between the usage status defined for the group and the usage status of each facility, it is determined whether the facility is an object of the group. The usage status information for the group definition and the usage status information of each facility are registered in the usage status information column 101. The correlation value between the usage status defined for the group and the usage status of each facility is registered in the similarity column 102, and whether the facility is an object of group (group A) is registered in the object determination column (group A) 103. In this example, facility A is determined to be an object facility of group A, and other facilities are excluded. In this way, by collecting data from automatic analysis devices with high similarity in usage patterns, the learning model learns, thereby constructing a more reliable learning model. Furthermore, instead of approximating all items of the usage status as a group condition, one or more items of the usage status can be approximated as a group condition based on the cause of the anomaly.

[0033] Figure 8 The flowchart shows the automatic analysis device 51 outputting recommended actions using the learning model 72. Preferably, the learning model is used to perform diagnosis in real time whenever the automatic analysis device 51 outputs specimen analysis results, maintenance results, etc. (S31). This allows for early detection of device malfunctions and notification to the operator. The processing unit 63 selects a learning model for performing the diagnosis (S32). A learning model that sets the output results of the automatic analysis device 51 as its associated device data can be extracted; sometimes multiple learning models are extracted. For the extracted learning model, associated device data containing the output results corresponding to that learning model is input (S33). The inference results, such as recommended user actions, output from the learning model are obtained (S34) and displayed on the screen (S35).

[0034] Figure 9This example shows a display screen of recommended actions displayed on the display unit 10. The recommended action display screen 110 includes a recommended action display unit 111, a recommended implementation period display unit 112, and a reference information display unit 113. As described above, whether an action recommendation is possible depends on whether the probability value output from the learning model is above or below a certain threshold. The recommended action is not limited to one; multiple actions may be recommended. Thus, when multiple actions are recommended, the recommendation level can be determined based on the probability value, and actions with a large difference between the threshold and the probability value are displayed as high-priority recommended actions. In this example, the recommendation level of each recommended action is displayed via bar 114, and the recommended actions are displayed in descending order of priority. Here, the higher-priority action "reaction unit cleaning" is displayed on the previous recommended action display screen 110a, and the action "reaction unit replacement" is displayed on the subsequent recommended action display screen 110b.

[0035] The recommended action display unit 111 displays the name of the recommended action and its recommendation level. The larger the shaded area of ​​bar 114, the greater the recommendation level.

[0036] The recommended implementation period display unit 112 displays not only the output value from the learning model but also the implementation history of the action in the automatic analysis device. If an implementation is scheduled, it also displays the date and time. Shaded dates in the calendar represent today, and a black circle indicates the recommended action "reaction unit cleaning." Furthermore, as reference information, the error occurrence date and time of the automatic analysis device and the action implementation date and time are estimated based on the average remaining time output by the learning model and displayed as reference information. This allows for determination of whether the operator actually performs the recommended action based on the timing under normal, average conditions.

[0037] The reference information display unit 113 can display device data that is important for determining whether a recommended action needs to be predetermined, such as the implementation history of blank measurement data of the reaction unit during the cleaning of the reaction unit. Therefore, even if the operator does not collect device data again, it can be confirmed whether the recommended action is appropriate based on the operator's experience.

[0038] use Figures 10-12This section explains the learning example for the action "light source replacement". Light source replacement is generally performed at the earlier of two times: periodically or after a certain period of use, or when the photometer reading exceeds a threshold. Even if a replacement threshold is set for the photometer reading, the light does not necessarily turn off after exceeding the threshold; therefore, periodic replacement or replacement after a certain period of use is recommended. However, while periodic replacement is recommended, the appropriate interval may vary depending on the operating time of the devices in each facility. For example, a facility used for 5 hours a day may require more frequent replacements than a facility used for 10 hours a day. Therefore, in the following explanation, the learning dataset used for learning is data collected from automated analysis devices in facilities with similar device operating times.

[0039] Figure 10 An example of a learning model is a learning dataset that collects time-series measurement data related to the light source as associated device data, timing of measurement data anomalies caused by the light source as action-related time information, and the operator's light source replacement timing, and recommends effective light source replacement timing. The input layer 121 of the learning model 120 is fed time-series photometer inspection data of the light source as associated device data, reaction monitoring data for each major wavelength and each sub-wavelength of the time series, and the output layer 122 outputs whether light source replacement can be recommended, the average remaining time until the reaction monitoring data of the automatic analysis device using a similar facility begins to show anomalies, and the average remaining time until light source replacement is performed.

[0040] Figure 11 The image shows an example of photometer readings used in learning or reasoning, specifically time-series absorbance data. As time passes after the light source is replaced with a new one, there is a trend of decreasing light intensity and gradually increasing absorbance. Figure 12 An example of response monitoring data used for learning or inference is shown. When the lifespan of the light source is nearing its end, the light intensity becomes unstable, resulting in absorbance spike noise 131 during the response process, indicating a linear anomaly. The spike noise 131 tends to increase over time. The learning model 120 infers whether light source replacement can be recommended by understanding such features from the time-series data.

[0041] The learning process of learning model 120 and the reasoning and action "reaction unit cleaning" using learning model 120 are the same, therefore repeated explanations are omitted. Figure 13 The recommended action display screen 140 is shown. The recommended action display screen 140 includes a recommended action display unit 141, a recommended implementation period display unit 142, and a reference information display unit 143. Because each display unit and... Figure 9The display shown is the same, so repeated descriptions are omitted. The operator can recommend light source replacement based on time series data based on associated device data, as well as the average error occurrence timing in the automatic analysis device and the operator's replacement timing, and perform light source replacement at an effective timing.

[0042] use Figures 14-17 This section describes a learning example of the action "sample probe outer wall cleaning". The outer wall of the sample probe needs to be cleaned daily after measurement. Neglecting cleaning may reduce sample dispensing accuracy. Reduced dispensing accuracy will affect the specimen measurement data; based solely on the measurement data, it's unclear whether the abnormal value was originally from a specimen or if the cause is reduced dispensing accuracy. Here, if abnormal measurement data is found and sample probe outer wall cleaning is performed, the user is asked to input whether it has an effect. That is, if it has an effect, the cause can be presumed to be reduced sample dispensing accuracy. Based on this collected learning dataset, a learning model recommending sample probe outer wall cleaning is generated based on the measurement data. In this case, the degree of sample probe contamination varies depending on the type of specimen used (serum, plasma, whole blood) and the usage time (24-hour use, etc.). Therefore, in the following explanation, the learning dataset used in the learning process is data collected from automated analysis devices in facilities with similar usage conditions that affect the degree of sample probe contamination.

[0043] Figure 14 In the example of the learning model, a learning dataset is collected, including time-series reaction monitoring data of the primary and secondary wavelengths as data of the associated device, timing of anomalies in the reaction monitoring data due to contamination of the outer wall of the sample probe as time information related to the action, and timing of the operator's cleaning of the outer wall of the sample probe, and an effective timing for cleaning the outer wall of the sample probe is recommended. The input layer 151 of the learning model 150 is input with reaction monitoring data of each primary and secondary wavelength as time-series data of the associated device, and the output layer 152 outputs whether cleaning of the outer wall of the sample probe can be recommended, the average remaining time until anomalies begin to appear in the reaction monitoring data of an automated analysis device using a similar facility as reference information, and the average remaining time until cleaning of the outer wall of the sample probe is performed.

[0044] Figure 15 The image shows measurement results when the outer wall of the sample probe was not cleaned. Figure 16A , 16B The data shown is reaction monitoring data when the outer wall of the sample probe was not cleaned. Figure 16A These are monitoring data from the ALB reaction process. Figure 16BThis is monitoring data from the CRE reaction process. Due to reduced sample aliquot accuracy, initial and remeasurement data for the same specimen may yield different results in multiple projects. Learning Model 150, by understanding the aforementioned characteristics from the time-series data of the reaction process, infers whether sample probe outer wall cleaning can be recommended.

[0045] The learning process of Learning Model 150 and the inference and action "reaction unit cleaning" using Learning Model 150 are the same, therefore repeated explanations are omitted. Figure 17 The recommended action display screen 160 is shown. The recommended action display screen 160 includes a recommended action display unit 161 and a recommended implementation period display unit 162. Because each display unit and... Figure 9 The display unit shown is the same, so repeated descriptions are omitted. The operator can recommend sample probe outer wall cleaning based on time-series data from the associated device, the average error timing of the automated analyzer used in the same operation, and the operator's cleaning timing, and perform sample probe outer wall cleaning at an effective timing. The reference information display unit can be set on the recommended action display screen 160, displaying information such as... Figure 16A The time series data shown.

[0046] The present invention has been described above based on embodiments and modifications, but is not limited to the embodiments and modifications described above; various modifications can be made within the scope of the invention without changing its spirit. For example, an example of a recommended action notification system consisting of multiple automatic analysis devices and learning devices has been described, but it can also be consisting of a single automatic analysis device. However, by using a learning dataset from multiple automatic analysis devices during learning, information from multiple automatic analysis devices can be collected earlier, improving the learning effect. Furthermore, by collecting learning data from automatic analysis devices other than itself, an average reasoning result from the automatic analysis devices, as described above, can be obtained, which can be used to determine whether the recommended action should actually be performed. In addition, although an example of reasoning using a learning model in an automatic analysis device has been shown, it can be implemented by providing a processing unit 63 in a learning device or other information processing device.

[0047] Furthermore, the recommended action notification system can include a structure that notifies the operator, for example, via a mobile terminal such as a smartphone. This allows the operator to receive the notification regardless of their location when a recommended action occurs. In addition to visual input, other input methods that reduce the operator's input load, such as voice input, can also be used for effective input of the operator's actions. Label Explanation

[0048] 1: Sample tray, 2: Reagent tray, 3: Reaction tray, 4: Reaction vessel, 5: Sampling mechanism, 6: Pipetting mechanism, 7: Stirring mechanism, 8: Photometric mechanism, 9: Cleaning mechanism, 10: Display unit, 11: Input unit, 12: Storage unit, 13: Control unit, 14: Piezoelectric element driver, 15: Stirring mechanism controller, 16: Sample container, 17, 19: Circular tray, 18: Reagent bottle, 20: Cold storage, 21: Reaction vessel, 22: Reaction vessel holder, 23: Drive mechanism, 24, 27: Probe, 25, 28: Support shaft, 26, 29: Arm, 31: Fixing unit, 51: Automatic analysis device, 52: Learning device, 53: Network, 60: Anomaly detection unit, 61: Dataset generation unit, 62: Database, 63: Processing unit, 71: Update unit, 72: Learning model 80: Action evaluation input screen; 81: Execution content display section; 82: Effect button; 83: Abnormal cause selection section; 90, 120, 150: Learning model; 91, 121, 151: Input layer; 92: Intermediate layer; 93, 122, 152: Output layer; 94: Time series measurement results; 95: Analysis information; 97: Action correlation information; 98: Abnormal cause; 100: Application table; 101: Application status information bar; 102: Approximation bar; 103: Object judgment bar; 110, 140, 160: Recommended action display screen; 111, 141, 161: Recommended action display section; 112, 142, 162: Recommended implementation period display section; 113, 143: Reference information display section; 114, 144, 163: Bar; 131: Spike noise.

Claims

1. A recommended action notification system, comprising a plurality of automatic analysis devices including a first automatic analysis device; And a learning device connected to multiple said automatic analysis devices via a network, recommending actions to be performed on the first automatic analysis device to the operator, characterized in that, include: The processing unit receives specimen analysis result data or maintenance result data from the first automatic analysis device, inputs associated device data containing the specimen analysis result data or maintenance result data into a learning model, and recommends to the operator to perform the prescribed action on the first automatic analysis device if the probability value of the recommended action output by the learning model is above a predetermined threshold; and The updating unit updates the learning model based on the learning dataset from multiple of the automated analysis devices. Upon receiving an anomaly detection from any of the multiple automated analysis devices, when the operator performs the prescribed action, the dataset generation unit of the automated analysis device displays an action evaluation input screen on the display unit of the automated analysis device. During a pre-set period based on the execution date and time of the prescribed action, it collects associated device data set according to the prescribed action, and generates a learning dataset containing the collected associated device data, as well as the effectiveness evaluation of the prescribed action and the cause of the anomaly input from the action evaluation input screen. When multiple actions are recommended, the degree of recommendation is determined based on the probability value output from the learning model. Recommended actions with a large difference between the specified threshold and the probability value are displayed on the screen as high-priority recommended actions.

2. The recommended action notification system as described in claim 1, characterized in that, The first automatic analysis device includes the processing unit, and the learning device includes the updating unit. The learning model updated by the updating unit is published to the first automatic analysis device.

3. The recommended action notification system as described in claim 1, characterized in that, The learning model includes: an input layer, which receives associated device data set according to the prescribed action; and an output layer, which receives the associated device data from the input layer and outputs a reasoning result related to the prescribed action. The inference results include the probability value of recommending the execution of the prescribed action and the probability value of the occurrence of a prescribed exception that the execution of the prescribed action is effective.

4. The recommended action notification system as described in claim 3, characterized in that, The inference results include the average remaining time until the specified anomaly occurs in the plurality of automated analysis devices, and the average remaining time until the specified action is performed in the plurality of automated analysis devices. The dataset generation unit of the automatic analysis device further includes in the learning dataset the elapsed time from a predetermined reference time to the occurrence of the predetermined anomaly in the automatic analysis device, and the elapsed time from the predetermined reference time to the execution of the predetermined action in the automatic analysis device.

5. The recommended action notification system as described in claim 3, characterized in that, The processing unit extracts the specimen analysis result data or the maintenance result data, which are then set as multiple learning models for inputting into the associated device data of the input layer. The processing unit inputs associated device data, including the specimen analysis result data or the maintenance result data, into the extracted multiple learning models. When there are multiple learning models whose probability values ​​for recommended execution of prescribed actions are above a specified threshold, the processing unit recommends multiple prescribed actions corresponding to the multiple learning models whose probability values ​​are above the specified threshold to the operator.

6. The recommended action notification system as described in claim 4, characterized in that, The processing unit displays a recommended action display screen on the display unit of the first automatic analysis device. The recommended action display screen shows: the probability value of recommending the execution of the prescribed action based on the learning model, the average remaining time until the prescribed anomaly occurs in the multiple automatic analysis devices, and the average remaining time until the prescribed action is executed in the multiple automatic analysis devices; the name of the prescribed action recommended to the operator; the date and time presumed to be the date and time of the prescribed anomaly occurring in the multiple automatic analysis devices; and the date and time presumed to be the date and time of the prescribed action being executed in the multiple automatic analysis devices.

7. The recommended action notification system as described in claim 1, characterized in that, The updating unit groups the multiple automatic analysis devices based on the similarity of their usage status, which includes usage and inspection content, and updates the learning model based on the learning dataset from the automatic analysis devices grouped into those with similar usage status.

8. An automatic analysis device for analyzing or maintaining samples, characterized in that, include: The storage unit stores a learning model, which has an input layer that receives associated device data set according to a predetermined action performed by an operator on an automatic analysis device, and an output layer that receives the associated device data input to the input layer and outputs inference results related to the predetermined action. The processing unit retrieves the learning model from the storage unit, inputs associated device data containing the results data of the specimen analysis or the results data of the maintenance into the learning model, and recommends the execution of the prescribed action to the operator when the probability value of the recommended execution of the prescribed action output by the learning model is above a predetermined threshold. A display unit that displays the name of the prescribed action recommended by the processing unit; The anomaly detection department detects anomalies. as well as The dataset generation unit, upon receiving an anomaly detection from the anomaly detection unit and upon the operator performing the prescribed action, displays an action evaluation input screen on the display unit. During a pre-set period based on the execution date and time of the prescribed action, it collects associated device data set according to the prescribed action and generates a learning dataset containing the collected associated device data, an effectiveness evaluation of the prescribed action input from the action evaluation input screen, and the cause of the anomaly. When recommending multiple actions, the display unit determines the degree of recommendation based on the probability value output from the learning model, and displays the recommended actions with a large difference between the specified threshold and the probability value as high-priority recommended actions.

9. The automatic analysis device as described in claim 8, characterized in that, The learning dataset is sent to the learning device and used to update the learning model.

10. A method for recommending actions, wherein a recommended action notification system recommends actions to be performed by a first automatic analysis device to an operator, the recommended action notification system comprising a plurality of automatic analysis devices including the first automatic analysis device, and a learning device connected to the plurality of automatic analysis devices via a network, characterized in that, When any one of the plurality of automatic analysis devices detects an anomaly, the automatic analysis device that detected the anomaly notifies the operator that the anomaly has been detected. When the operator receives a notification of the detected anomaly and performs a prescribed action, the automatic analysis device displays an action evaluation input screen on its display unit. The automatic analysis device that detects the anomaly collects associated device data set according to the prescribed action within a pre-defined period based on the execution date and time of the prescribed action, and generates a learning dataset containing the collected associated device data, the effectiveness evaluation of the prescribed action input from the action evaluation input screen, and the cause of the anomaly. Multiple automated analysis devices send the learning dataset to the learning device. The learning device updates the learning model based on the learning dataset from multiple of the automated analysis devices. The learning device publishes the updated learning model to the first automatic analysis device. The first automatic analysis device performs sample analysis or maintenance. The first automatic analysis device inputs associated device data, including the results of the specimen analysis or the results of the maintenance, into the learning model. If the probability value of recommending the execution of the prescribed action output by the learning model after receiving the associated device data input is above a predetermined threshold, the device recommends the execution of the prescribed action to the operator. When multiple actions are recommended, the degree of recommendation is determined based on the probability value output from the learning model. Recommended actions with a large difference between the specified threshold and the probability value are displayed on the screen as high-priority recommended actions.

11. The recommended action notification method as described in claim 10, characterized in that, The learning device groups the multiple automatic analysis devices based on the similarity of their usage status, which includes usage and inspection content, and updates the learning model based on the learning dataset from the automatic analysis devices grouped into groups with similar usage status.

Images