KR102567388B1 - A method of prescribing aid providing test set by using interview content - Google Patents

Abstract

The present disclosure relates to a prescription assistance method for assisting the prescription of a doctor for a patient. Specifically, a test set to be prescribed for the patient is provided using the interview log, which records the conversation between the doctor and the patient. In particular, the method is to consider the intention of the doctor. In the present disclosure, in order to reduce the time spent on tasks other than interviews during the treatment process, the provided prescription assistance method fully considers the patterns of the doctor.

Description

A prescribing assistance method that provides a test set using interview contents {A METHOD OF PRESCRIBING AID PROVIDING TEST SET BY USING INTERVIEW CONTENT}

The present disclosure relates to a prescription assistance method for assisting a doctor's prescription for a patient, and specifically, provides a set of tests to be prescribed for a patient using an interview recording a conversation between a doctor and a patient. It's about how intent is taken into account.

When a doctor treats a patient in a hospital, the time available for interviewing a patient has a significant effect on improving the quality of patient care. As the interview time lengthens, the patient can receive more meticulous treatment, and accordingly, accurate diagnosis and prescription are provided to the patient, and the patient's satisfaction also increases.

On the other hand, the doctor's medical treatment process includes not only interviews with patients, but also various tasks such as recording medical treatment contents, prescribing tests or drugs necessary for patients, and recording prescription results. At this time, as the business hours other than the interview with the patient increase during the limited treatment time, the interview time with the patient becomes shorter. A problem arises that the patient's satisfaction is also lowered according to the short interview time. For example, in the case of a large university hospital in Korea, appointments are managed at a level where doctors have to treat 3 patients every 10 minutes, and the time to treat 1 patient is limited to about 3 minutes. .

As one of the ways to solve this problem, there will be a method of automatically prescribing drugs and recording the prescription results in order to reduce the time required for tasks other than interviews. In particular, a prescription assistance method of recommending drugs to be prescribed to the patient or tests to be taken by the patient based on the contents of the interview between the doctor and the patient may be considered.

Considering the above-mentioned prescription assistance method, in the process of making a prescription drug list, a doctor can individually select and configure a list of drugs. Therefore, it is difficult to say that the doctor's "pattern" was sufficiently considered when deciding whether or not to prescribe individual drugs using the interview contents.

In the present disclosure, a prescription assistance method is provided as a way to reduce the time required for work other than interviews in the course of treatment, but the prescription assistance method is sufficiently considered so as to overcome the above-mentioned problems. .

One problem to be solved by the present disclosure is to automatically generate a medical record for a patient in consideration of a doctor's treatment pattern for the patient.

One problem to be solved by the present disclosure is to classify the content of an interview between a doctor and a patient into a specific pattern and to generate auxiliary prescription information corresponding to the pattern.

One problem to be solved by the present disclosure is to generate a test set including test items to be performed on a patient using interview contents between a doctor and a patient.

One problem to be solved by the present disclosure is to generate comprehensive information including a test set, a drug set, consultation request information, visit date information, and the like by using interview contents between a doctor and a patient.

The problems to be solved in the present disclosure are not limited to the above-described problems, and problems not mentioned will be clearly understood by those skilled in the art from the present disclosure and the accompanying drawings. .

According to one embodiment, a method for providing prescription assistance information of a server includes: acquiring medical records of a target doctor, each of the medical records including an interview record and a prescription test item list; generating learning data sets using the medical records; generating a pattern classification model using the learning data sets; obtaining a first analysis target interview record of the target doctor; obtaining first candidate examination items using the first interview record to be analyzed and the pattern classification model; and providing the first candidate test items, wherein the generating of the learning data sets includes: generating test sets using the medical records; In consideration of the first prescription test item list of the first medical record, data corresponding to the first test set among the test sets is labeled with data obtained by processing the first interview record of the first medical record. generating a learning data set; and processing data corresponding to a second test set among the test sets into a second interview record of the second medical record in consideration of a second prescription test item list of the second medical record. and generating a second training data set by labeling data, wherein the pattern classification model receives data obtained by processing the first analysis target interview record and outputs data corresponding to one of the test sets. and the first candidate test items are test items included in a test set corresponding to the outputted data, and the generating of the test sets includes a first prescription test item list of a first medical record to an n-th medical record. A method is provided, which includes extracting test items included in an n-th prescription test item list (n is a natural number equal to or greater than 2), and determining the test sets in consideration of the extracted test items.

The solutions to the problems of the present disclosure are not limited to the above-described solutions, and solutions not mentioned will be clearly understood by those skilled in the art from the present disclosure and the accompanying drawings. You will be able to.

According to an embodiment of the present disclosure, medical records for patients are automatically generated after medical treatment, so that time spent on administrative tasks of a doctor can be reduced.

According to an embodiment of the present disclosure, the time required for a task of recording medical treatment contents during a medical treatment process of a doctor may be reduced.

According to an embodiment of the present disclosure, a task of computerarily recording medical treatment contents during a medical treatment process of a doctor may be omitted.

According to an embodiment of the present disclosure, an operation of computerarily searching for and selecting test items or medicines necessary for a patient during a doctor's treatment process may be omitted.

Effects according to the present disclosure are not limited to the above-mentioned effects, and effects not mentioned will be clearly understood by those skilled in the art from the present disclosure and the accompanying drawings.

1 is a diagram showing configurations of a prescription assistance system according to an embodiment.
2 is a diagram illustrating configurations of an auxiliary server according to an embodiment.
3 is a diagram illustrating configurations of a main server according to an exemplary embodiment.
4 is a flowchart illustrating a prescription assisting method according to an exemplary embodiment.
5 is a diagram illustrating a process in which a prescription assistance system performs a prescription assistance method according to an embodiment.
6 is a diagram illustrating a process of generating an interview record according to an exemplary embodiment.
7 is a diagram illustrating a process of generating model input data according to an exemplary embodiment.
8 is a diagram illustrating an input/output process of a pattern classification model according to an embodiment.
9 is a diagram illustrating a pattern database according to an exemplary embodiment.
10 is a diagram illustrating another output form of a pattern classification model according to an exemplary embodiment.
11 is a diagram illustrating a method of providing prescription auxiliary information according to an exemplary embodiment.
12 is a diagram illustrating an input/output form of a pattern classification model and a learning data set for learning the pattern classification model according to an embodiment.
13 is a flowchart illustrating a method of generating a training data set according to an exemplary embodiment.
14 is a diagram illustrating a method of generating test sets using frequently used test item lists according to an embodiment.
15 is a diagram illustrating prescription test item lists according to an embodiment.
16 is a diagram illustrating a method of generating check sets according to an exemplary embodiment.
17 is a diagram illustrating a method of classifying interview records according to examination sets and classification criteria according to an exemplary embodiment.
18 and 19 are diagrams illustrating the structure of a pattern classification model according to an embodiment.
20 is a diagram illustrating a structure of an encoder constituting a pattern classification model according to an embodiment.
21 is a diagram illustrating a method of using a basic inspection model and an additional inspection model according to an embodiment.
22 is a diagram illustrating an output form of an additional inspection model according to an exemplary embodiment.
23 is a flowchart illustrating a method of generating a training data set of an additional examination model according to an exemplary embodiment.
24 is a diagram illustrating a case in which a plurality of additional inspection models are used according to an embodiment.
25 is a diagram illustrating a data transmission/reception relationship between components of a prescription assistance system according to an embodiment.
26 is a diagram showing an input/output form of a drug prescription model and a learning data set for learning the drug prescription model according to an embodiment.
27 is a flowchart illustrating a method of generating a training data set for training a drug prescription model according to an embodiment.
28 is a diagram illustrating a comprehensive medical care information providing method according to an embodiment.

According to one embodiment, a method for providing prescription assistance information of a server includes: acquiring medical records of a target doctor, each of the medical records including an interview record and a prescription test item list; acquiring frequently used examination item lists set by the target doctor; generating learning data sets using the frequently used examination item lists and the medical records; generating a pattern classification model using the learning data sets; obtaining a first analysis target interview record of the target doctor; obtaining first candidate examination items using the first interview record to be analyzed and the pattern classification model; and providing the first candidate test items, wherein the generating of the learning data sets includes: generating test sets using the frequently used test item lists; In consideration of the first prescription test item list of the first medical record, data corresponding to the first test set among the test sets is labeled with data obtained by processing the first interview record of the first medical record. generating a learning data set; and processing data corresponding to a second test set among the test sets into a second interview record of the second medical record in consideration of a second prescription test item list of the second medical record. and generating a second training data set by labeling data, wherein the pattern classification model receives data obtained by processing the first analysis target interview record and outputs data corresponding to one of the test sets. and the first candidate test items are test items included in a test set corresponding to the output data.

The frequently used test item lists and the test sets correspond to each other on a one-to-one basis.

The frequently used inspection item lists include a first frequently used inspection item list including at least two or more inspection items and a second frequently used inspection item list including at least two or more inspection items, wherein the first frequently used inspection item list includes the first frequently used inspection item list. The list and the second frequently used examination item list include at least one identical examination item and at least one different examination item.

The method may further include obtaining a second analysis target interview record of the target doctor; and acquiring second candidate examination items by using the second interview record to be analyzed and the pattern classification model, wherein the pattern classification model inputs data obtained by processing the first interview record to be analyzed. receiving and outputting data corresponding to the first examination set among the examination sets, receiving data obtained by processing the second analysis target interview record, and outputting data corresponding to the second examination set among the examination sets; , The second candidate check items are check items included in the second check set, and the first candidate check items and the second candidate check items are not identical to each other, but include at least one check item in common.

The first prescription test item list of the first medical record and the second prescription test item list of the second medical record are not identical to each other, but include at least one common test item.

The first interview record of the first medical record is obtained by converting first sound data obtained by recording the first treatment process of the target doctor, and the first prescription test item list of the first medical record is In the first treatment process, a test item prescribed by the target doctor for the patient is included.

Test items included in the first prescription test item list of the first medical record are the same as test items included in the first test set.

The first test set includes test items included in the first prescription test item list of the first medical record.

The generating of the learning data sets may include generating a third learning data set by labeling data indicating that there is no pattern in data obtained by processing a third interview record of a third medical record; None of them is identical to the third prescription test item list of the third medical record.

The generating of the learning data sets may include generating a third learning data set by labeling data corresponding to the first test set to data obtained by processing a third interview record of a third medical record; The number of test items commonly included in the third prescription test item list of the third medical record and the first test set is equal to the number of test items commonly included in the third prescription test item list and the second test set. and the number of test items included in the first test set is greater than the number of test items included in the second test set.

The pattern classification model receives the processed data of the first analysis target interview record and outputs pattern data corresponding to one of the test sets, and the first candidate test items are tests corresponding to the pattern data. These are the check items included in the set.

The generating of the pattern classification model may include selecting a pre-learned base model with data different from the training data sets; and training the base model using the training data sets.

The base model is a bidirectional encoder representations form transformer (BERT) or a generative pre-trained transformer 3 (GPT-3).

The method may include recording a treatment process for a patient of the target doctor; and obtaining the first analysis target interview record from the recorded medical treatment process.

According to another embodiment, a method for generating a model using a server includes: acquiring medical records of a target doctor, each of the medical records including an interview record and a prescription test item list; acquiring frequently used examination item lists set by the target doctor; generating learning data sets using the frequently used examination item lists and the medical records; and training a base model using the training data sets, wherein the generating of the training data sets comprises: generating test sets using the frequently used test item lists; In consideration of the first prescription test item list of the first medical record, data corresponding to the first test set among the test sets is labeled with data obtained by processing the first interview record of the first medical record. generating a learning data set; and processing data corresponding to a second test set among the test sets into a second interview record of the second medical record in consideration of a second prescription test item list of the second medical record. A method is provided that includes labeling data to create a second training data set.

According to another embodiment, a method for providing prescription assistance information of a server, comprising: acquiring medical records of a target doctor, each of the medical records including an interview record and a prescription test item list; generating learning data sets using the medical records; generating a pattern classification model using the learning data sets; obtaining a first analysis target interview record of the target doctor; obtaining first candidate examination items using the first interview record to be analyzed and the pattern classification model; and providing the first candidate test items, wherein the generating of the learning data sets includes: generating test sets using the medical records; In consideration of the first prescription test item list of the first medical record, data corresponding to the first test set among the test sets is labeled with data obtained by processing the first interview record of the first medical record. generating a learning data set; and processing data corresponding to a second test set among the test sets into a second interview record of the second medical record in consideration of a second prescription test item list of the second medical record. and generating a second training data set by labeling data, wherein the pattern classification model receives data obtained by processing the first analysis target interview record and outputs data corresponding to one of the test sets. and the first candidate test items are test items included in a test set corresponding to the outputted data, and the generating of the test sets includes a first prescription test item list of a first medical record to an n-th medical record. A method is provided, which includes extracting test items included in an n-th prescription test item list (n is a natural number equal to or greater than 2), and determining the test sets in consideration of the extracted test items.

The determining of the check sets may include generating candidate check sets using the extracted check items, each of the candidate check sets including at least two or more of the extracted check items, and the candidate check sets. and selecting at least some of the test sets as the test sets.

The generating of the test sets may include generating first candidate test sets using the extracted test items, wherein each of the first candidate test sets includes at least two or more of the extracted test items. , calculating co-occurrence determination values of each of the first candidate test sets for the first to nth prescription test item lists, and calculating the co-occurrence determination value of each of the first candidate test sets; The method further includes selecting at least some of the secondary candidate test sets as secondary candidate test sets.

In the step of selecting the secondary candidate test sets, a primary candidate test set having the largest co-occurrence determination value among the primary candidate test sets is selected as the secondary candidate test sets.

The co-occurrence determination value is a value obtained by dividing the number of lists including both the first primary candidate test set and the second primary candidate test set among the first to nth prescription test item lists by n, A value obtained by dividing the number of lists including the first primary candidate test set among the 1 to nth prescription test item lists by n, and the second primary candidate test set among the first to nth prescription test item lists It is calculated by dividing the number of lists to be multiplied by n by the value obtained.

The generating of the test sets includes performing first and second steps, wherein the first step comprises each of the (M-1) order candidate test sets for the first to nth prescription test item lists. Calculating a co-occurrence decision value of , and the second step may include at least some of the (M-1) th candidate test sets based on the co-occurrence decision value of each of the (M-1) th candidate test sets. and selecting as M-order candidate check sets, where M is a natural number equal to or greater than 3.

The generating of the check sets is repeated until the first step and the second step satisfy a preset condition, and finally, after the second step is finished, the selected K-th candidate test set is obtained as the test set. It further includes the steps of

obtaining a second analysis target interview record of the target doctor; and acquiring second candidate examination items by using the second interview record to be analyzed and the pattern classification model, wherein the pattern classification model inputs data obtained by processing the first interview record to be analyzed. receiving and outputting data corresponding to the first examination set among the examination sets, receiving data obtained by processing the second analysis target interview record, and outputting data corresponding to a second examination set among the examination sets; The second candidate test items are test items included in the second test set, and the first candidate test items and the second candidate test items are not identical to each other, but include at least one test item in common.

The first prescription test item list of the first medical record and the second prescription test item list of the second medical record are not identical to each other, but include at least one common test item.

The first interview record of the first medical record is obtained by converting first sound data obtained by recording the first treatment process of the target doctor, and the first prescription test item list of the first medical record is In the first treatment process, a test item prescribed by the target doctor for the patient is included.

Test items included in the first prescription test item list of the first medical record are the same as test items included in the first test set.

The first test set includes test items included in the first prescription test item list of the first medical record.

The generating of the learning data sets may include generating a third learning data set by labeling data indicating that there is no pattern in data obtained by processing a third interview record of a third medical record; None of them is identical to the third prescription test item list of the third medical record.

The generating of the learning data sets may include generating a third learning data set by labeling data corresponding to the first test set to data obtained by processing a third interview record of a third medical record; The number of test items commonly included in the third prescription test item list of the third medical record and the first test set is equal to the number of test items commonly included in the third prescription test item list and the second test set. and the number of test items included in the first test set is greater than the number of test items included in the second test set.

The generating of the pattern classification model may include selecting a pre-learned base model with data different from the training data sets; and training the base model using the training data sets.

The base model is a bidirectional encoder representations form transformer (BERT) or a generative pre-trained transformer 3 (GPT-3).

recording a treatment process for a patient of the target doctor; and obtaining the first analysis target interview record from the recorded medical treatment process.

According to another embodiment, a method for providing prescription assistance information of a server, comprising: acquiring medical records of a target doctor, each of the medical records including an interview record and a prescription test item list; obtaining basic test sets using test items included in the prescription test item lists; generating basic learning data sets using the medical records and the basic examination sets; generating additional learning data sets using at least some of the medical records and the examination items; generating a basic inspection model using the basic training data sets; generating an additional test model using the additional training data sets; obtaining a first analysis target interview record of the target doctor; acquiring a first basic test set using the first interview record to be analyzed and the basic test model; obtaining at least one additional examination item using the first analysis target interview record and the additional examination model; and providing first candidate test items using the first basic test set and the at least one additional test item, wherein the basic training data sets include at least a first basic training data set and a second basic learning data set. The first basic learning data set includes a data set obtained by processing a first interview record of the first medical record in consideration of a first prescription test item list of the first medical record. It is generated by labeling data corresponding to a test set, and the second basic learning data set is processed by a second interview record of the second medical record in consideration of a second prescription test item list of the second medical record. generated by labeling data corresponding to a second test set among the test sets, wherein the additional training data sets include at least a first additional training data set and a second additional training data set; The data set is generated by labeling at least a first additional test item to data obtained by processing the first interview record of the first medical record in consideration of the first prescription test item list of the first medical record, and the second The additional learning data set is created by labeling at least a second additional test item to data obtained by processing the second interview record of the second medical record in consideration of the second prescription test item list of the second medical record. The first prescription test item list includes at least some of the test items included in the first test set and the first additional test items, and the second prescription test item list includes the test items included in the second test set. at least some of them and the second additional inspection item.

The first candidate inspection items include inspection items included in the first basic inspection set and the at least one additional inspection item.

The first prescription test item list includes test items included in the first test set and the first additional test item, and the second prescription test item list includes test items included in the second test set and It is composed of the second additional inspection item.

Additional test items obtained using the additional test model are not included in the test items included in the basic test sets.

The method may further include obtaining a second analysis target interview record of the target doctor; obtaining a second basic test set using the second interview record to be analyzed and the basic test model; and obtaining data indicating that there is no additional examination item by using the second analysis target interview record and the additional examination model.

The method further includes providing test items included in the second basic test set as second candidate test items.

The basic learning data sets include a third basic learning data set, and the third basic training data set is a third interview record of the third medical record in consideration of a third prescription test item list of the third medical record. The processed data is generated by labeling data corresponding to a third test set among the test sets, the additional training data sets include a third additional training data set, and the third additional training data set includes the first additional training data set. 3 It is generated by labeling data indicating that there is no additional examination item in the processed data of the third interview record of the medical record.

obtaining a third analysis target interview record of the target doctor; and obtaining a third basic examination set by using the third interview record to be analyzed and the basic examination model, wherein the obtaining of the third basic examination set comprises: the third interview record to be analyzed; and acquiring a probability value indicating the third basic examination set by inputting model input data processed from the basic examination model into the basic examination model, and if the probability value is equal to or less than a preset value, the third analysis target interview record. The step of using the additional examination model is performed for , and if the probability value is greater than or equal to a preset value, the step of using the additional examination model is not performed for the interview record to be analyzed.

The outputting of the basic test set and the at least one additional test item may include displaying the test items included in the basic test set on a first area on a display and the at least one additional test item on a second area on the display. Including displaying

According to another embodiment, a method for providing prescription assistance information of a server, comprising: obtaining medical records of a target doctor, each of the medical records including an interview record and a prescription test item list; acquiring frequently used examination item lists set by the target doctor; generating basic test sets using the frequently used test item lists; generating basic learning data sets using the medical records and the basic examination sets; generating additional learning data sets using at least some of the test items included in the prescription test item lists and the medical records; generating a basic inspection model using the basic training data sets; generating an additional test model using the additional training data sets; obtaining a first analysis target interview record of the target doctor; acquiring a first basic test set using the first interview record to be analyzed and the basic test model; obtaining at least one additional examination item using the first analysis target interview record and the additional examination model; and providing first candidate test items using the first basic test set and the at least one additional test item, wherein the basic training data sets include at least a first basic training data set and a second basic learning data set. The first basic learning data set includes a data set obtained by processing a first interview record of the first medical record in consideration of a first prescription test item list of the first medical record. It is generated by labeling data corresponding to a test set, and the second basic learning data set is processed by a second interview record of the second medical record in consideration of a second prescription test item list of the second medical record. generated by labeling data corresponding to a second test set among the test sets, wherein the additional training data sets include at least a first additional training data set and a second additional training data set; The data set is generated by labeling at least a first additional test item to data obtained by processing the first interview record of the first medical record in consideration of the first prescription test item list of the first medical record, and the second The additional learning data set is created by labeling at least a second additional test item to data obtained by processing the second interview record of the second medical record in consideration of the second prescription test item list of the second medical record. The first prescription test item list includes at least some of the test items included in the first test set and the first additional test items, and the second prescription test item list includes the test items included in the second test set. A method including at least some of these and the second additional inspection item is provided.

There is a one-to-one correspondence between the frequently used check item lists and the basic check sets.

According to another embodiment, a method for providing prescription assistance information of a server, comprising: acquiring medical records of a target doctor, each of the medical records including an interview record and a prescription test item list; generating basic learning data sets using the medical records and basic examination sets, wherein each of the basic examination sets includes at least one examination item; generating additional learning data sets using at least some of the medical records and test items included in the prescription test item lists; generating a basic inspection model using the basic training data sets; generating an additional test model using the additional training data sets; obtaining a first analysis target interview record of the target doctor; acquiring a first basic test set using the first interview record to be analyzed and the basic test model; obtaining at least one additional examination item using the first analysis target interview record and the additional examination model; and providing first candidate test items using the first basic test set and the at least one additional test item.

The foregoing objects, features and advantages of the present disclosure will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. However, the present disclosure may apply various changes and may have various embodiments. Hereinafter, specific embodiments will be illustrated in the drawings and described in detail.

Like reference numbers designate essentially like elements throughout this disclosure. In addition, components having the same function within the scope of the same idea appearing in the drawings of each embodiment will be described using the same reference numerals, and overlapping descriptions thereof will be omitted.

Numbers (eg, first, second, etc.) used in the description process of the present disclosure are merely identifiers for distinguishing one component from another.

In addition, the suffixes "module" and "unit" for components used in the following embodiments are given or used interchangeably in consideration of ease of writing the specification, and do not have meanings or roles that are distinguished from each other by themselves.

Unless specifically stated or clear from the context, the term "about" in reference to a numerical value shall be understood to mean the recited numerical value and up to +/- 10% of that numerical value, and in reference to a numerical range "about" The term "is to be understood to mean a range from 10% lower than the lower value of the numerical range to 10% higher than the upper value of the numerical range.

In the following embodiments, terms such as "include" or "have" mean that features or elements described in the specification exist, and the possibility of one or more other features or elements being added is excluded in advance. It is not.

In the drawings, the size of components may be exaggerated or reduced for convenience of explanation. For example, the size and thickness of each component shown in the drawings are arbitrarily shown for convenience of explanation, and the present invention is not necessarily limited to those shown.

When an embodiment is otherwise implementable, a specific process sequence may be performed differently from the described sequence. For example, two processes described in succession may be performed substantially simultaneously, or may be performed in an order reverse to the order described.

The present disclosure relates to a prescription assisting method for assisting a doctor's prescription record for a patient and a prescription assisting system performing the same.

In the present disclosure, assisting in prescription means providing related information to a doctor when the doctor prescribes medicines or tests after treating the patient. Specifically, when a doctor prescribes necessary drugs to a patient or tests to be taken by a patient after performing medical treatment on a patient, the prescription assistance method provides the doctor with candidate drugs or candidate test items, and the doctor receives the provided A prescription for a patient may be performed in consideration of candidate medicines or candidate test items.

[System Configuration]

Hereinafter, the configuration of the prescription assistance system will be described with reference to FIGS. 1 to 3 .

1 is a diagram showing configurations of a prescription assistance system 10 according to an embodiment.

Referring to FIG. 1, the prescription auxiliary system 10 includes a recording device RD, an auxiliary server 1000, a main server 2000, an electronic medical record (EMR) server 3000, a hospital server 4000, and a terminal. (TD).

Hereinafter, each component of the prescription assistance system 10 will be described in detail.

The recording device RD may record the doctor's treatment process. Specifically, the recording device RD may record the doctor's interview with the patient. For example, the recording device RD may be located in an examination room and record sounds generated in the examination room. The sound generated in the examination room may include a doctor's voice, a patient's voice, and a third person's voice (eg, a nurse, caregiver, guardian, etc.).

The recording device RD may obtain sound data by recording the doctor's treatment process. The sound data may refer to data of a voice signal in which contents recorded during a doctor's medical treatment are reflected.

A portable recorder, an installation type recorder, or a desktop or tablet may be used as the recording device RD. Alternatively, a smart phone may be used as the recording device (RD), and furthermore, a smart watch, a smart band, a smart ring, and a smart neckless, etc. A wearable device equipped with a recording function may be used as the recording device RD.

2 is a diagram showing configurations of an auxiliary server 1000 according to an embodiment.

The auxiliary server 1000 may convert the acoustic data into interview records.

Here, the sound data converted into the interview record may be sound data obtained from the above-described recording device RD.

Also, here, the interview record may be understood as data obtained by converting audio data into text form, but is not limited thereto.

The secondary server 1000 may transmit data to the main server 2000 or receive data from the main server 2000. To this end, the secondary server 1000 and the main server 2000 communicate through wired or wireless communication. Data can be sent and received.

Referring to FIG. 3 , an auxiliary server 1000 may include a voice recognition model 1100 and a speaker classification model 1200 .

The auxiliary server 1000 may generate the interview transcript using the speech recognition model 1100 and/or the speaker classification model 1200 .

The voice recognition model 1100 may convert sound data into text data.

The speaker classification model 1200 may classify speakers after converting audio data into text. For example, the speaker classification model 1200 may classify the voices of speakers included in the audio data and extract audio data or text data corresponding to each speaker's voice.

Alternatively, the speaker classification model 1200 may distinguish speakers from audio data before text conversion. For example, the speaker classification model 1200 may process the audio data by classifying the audio signal included in the audio data according to the speaker and extracting the audio signal of a specific speaker as necessary.

The EMR server 3000 may refer to a server in which patient records are stored and utilized. Specifically, the EMR server 3000 may refer to a server in which data for executing an EMR program is stored or an EMR program is executed.

The EMR program is a program for electronic medical records. It refers to a program for computerizing patient charts such as patient personal information, medical history, examination results, treatment results, surgery records, hospitalization and discharge records, outpatient treatment records, and health checkup records in hospitals. do.

The EMR server 3000 may include a database that stores various types of information about the patient described above.

The EMR server 3000 may provide information about patients to a doctor through the terminal TD. For example, the EMR server 3000 may output an EMR interface through the terminal TD. The doctor may read information about the patient or record information about the patient using the EMR interface in the terminal TD.

The EMR server 3000 may transmit and receive data to and from the main server 2000 .

The EMR server 3000 may transmit/receive data with the terminal TD.

The EMR server 3000 may transmit and receive data to and from the hospital server 4000 .

The EMR server 3000 may be located in a hospital where doctors treat patients.

Alternatively, the EMR server 3000 may be located outside the hospital. For example, the EMR server 3000 may be implemented using Internet-based cloud computing technology. In this case, the terminal TD may access the EMR server 3000, receive necessary information, and display the information on the screen.

The hospital server 4000 may refer to a server in which a computer system necessary for management or operation of a hospital is built.

If the above-described EMR server 3000 is a server focused on patient management, the hospital server 4000 is a server for hospital management, administrative tasks such as finance/accounting/asset/personnel/performance, and information exchange between servers. It can be understood as a server for performing the like.

For example, the hospital server 4000 includes a hospital information system (HIS), a laboratory information system (LIS), a clinical device information system (CDIS), a customer relationship management system ( Customer relationship management (CRMS), data warehouse, etc. can be built.

The hospital server 4000 may store medical records of patients as well as medical records of medical treatments performed by individual doctors. Here, the medical record may include a record of an interview between a doctor and a patient, medications prescribed by the patient in the corresponding medical treatment, or test items prescribed to the patient.

The hospital server 4000 may provide the patient's medical records or the doctor's medical records to the EMR server 3000 or the main server 2000 .

The terminal TD may provide prescription assistance information generated by the prescription assistance system 10 . For example, prescription assistance information (ex. candidate test items, candidate medicines, etc.) generated according to a prescription assistance method described below may be provided to the terminal TD. Specifically, the terminal TD may receive prescription assistance information from the main server 2000 or the EMR server 3000 .

Hereinafter, the supplementary prescription information is described as candidate test items to be received by the patient, but the technical spirit of the present disclosure is not limited thereto.

The terminal TD can be understood as a device assigned to a pseudo-person. For example, a doctor's desktop, laptop, tablet, or smart phone may be used as the terminal TD.

The terminal TD can transmit/receive data with the main server 2000, the EMR server 3000, and/or the hospital server 4000. For example, the terminal TD may access the EMR server 3000 and receive prescription assistance information. For another example, the terminal TD may receive prescription assistance information from the main server 2000 .

The terminal TD may output the provided supplemental prescription information.

The doctor may record medical treatment contents using the supplementary prescription information provided through the terminal TD. For example, a doctor may determine test items to be prescribed to a patient using a test set or candidate test items provided through the terminal TD. At this time, the doctor may perform verification, such as checking and correcting test sets or candidate test items provided through the terminal TD, prior to determining test items to be prescribed for the patient.

The terminal TD may transmit and receive data with the EMR server 3000 while transmitting and receiving data with the main server 2000 . For example, the doctor may check information on the next patient using the terminal TD after treating the patient. At this time, the terminal TD is a prescription generated from the interview record for the patient from the main server 2000. Receiving auxiliary information and receiving information on the next patient from the EMR server 3000 may be performed in parallel.

3 is a diagram showing configurations of a main server 2000 according to an embodiment.

Referring to FIG. 3 , the main server 2000 may include a memory unit 2100, an input unit 2200, an output unit 2300, a communication unit 2400, and a control unit 2500.

The memory unit 2100 may store programs for performing the prescription assisting method and information necessary for the prescription assisting method. The memory unit 2100 may include a tokenization model 2110, an embedding model 2120, a pattern classification model 2130, and a pattern database 2140.

The tokenization model 2110 may tokenize text data in the form of paragraphs or sentences. Tokenization means dividing text data such as paragraphs or sentences into a plurality of tokens according to a specific criterion. Each token may correspond to a word, a syllable, a word, or a sentence according to a classification criterion.

In order to analyze the interview records using the pattern classification model 2130 described later, it is necessary to change the data format of the interview records, and as part of the process, the tokenization model 2110 divides the interview records into a plurality of tokens. can

The embedding model 2120 may embed text data. Embedding refers to an operation of converting text data into a specific representation that is easy for a computer to analyze. Data converted through the embedding process may be in the form of a vector, and such data is referred to as an embedding vector.

As part of the process of changing the data format of the interview transcript described above, the embedding model 2120 may convert the tokens into a vector form.

The pattern classification model 2130 may receive data and output pattern data. Here, the pattern data indicates a pattern of data input to the pattern classification model 2130, and prescription auxiliary information obtained from the main server 2000 may be determined according to the determined pattern. For example, a specific test set may be provided as prescription auxiliary information according to pattern data output from the pattern classification model 2130, and the specific test set may include candidate test items. For another example, specific candidate examination items may be provided as prescription auxiliary information according to pattern data output from the pattern classification model 2130 .

Meanwhile, the pattern classification model 2130 may receive data and output supplemental prescription information. Here, the prescription auxiliary information may be a test set, data corresponding to the test set, or candidate test items. In other words, in the above-described embodiment, the process of outputting pattern data and determining supplemental prescription information according to the pattern data may be compressed and output of supplemental prescription information directly from the pattern classification model 2130 .

Information on patterns may be stored in the pattern database 2140 . For example, test sets corresponding to each of the patterns may be stored in the pattern database 2140 . The main server 2000 may obtain pattern data indicating a specific pattern by using the pattern classification model 2130, and acquire it from the pattern database 2140 as supplemental prescription information to provide a test set corresponding to a specific pattern. there is.

The memory unit 2100 may store various types of data in addition to the pattern database 2140 . For example, the memory unit 2100 may store medical records of doctors or information of patients obtained from the EMR server 3000 or the hospital server 4000 . For another example, the memory unit 2100 may store learning data sets necessary for learning a pattern classification model 2130 to be described later.

The memory unit 2100 may temporarily or semi-permanently store the above-described models and databases as well as various other data. Examples of the memory unit 2100 include a hard disk drive (HDD), a solid state drive (SSD), a flash memory, a read-only memory (ROM), and a random access memory (RAM). etc. may be present. The memory unit 2100 may be provided in a form embedded in the main server 2000 or in a detachable form.

Meanwhile, some of the above-described models stored in the memory unit 2100 may be omitted. For example, if the tokenization or embedding process is performed in an external server other than the main server 2000, the tokenization model 2110 or the embedding model 2120 may be omitted. For another example, when data output from the pattern classification model 2130 is not pattern data but data corresponding to a test set or candidate test items, the pattern database 2140 may be omitted.

The input unit 2200 may receive a user input from a user. Here, the user may be understood as a developer who manages the main server 2000 and a person who builds and manages the prescription assistance system 10 as a whole. User input may be made in various forms including key input, touch input, and voice input. Examples of the input module 2600 include a traditional type of keypad, keyboard, mouse, touch sensor for detecting a user's touch, and various types of input means for detecting or receiving various types of user inputs. is a concept

The output unit 2300 may output the state of the main server 2000, data obtained or generated by the main server 2000, or data stored in the main server 2000. The output unit 2300 is a comprehensive concept that includes a display that outputs an image, a speaker that outputs sound, a haptic device that generates vibration, and other various types of output means.

The communication unit 2400 may communicate with an external device. The main server 2000 may transmit/receive data with at least some of the components of the prescription assistance system 10 through the communication unit 2400 . For example, the main server 2000 transmits and receives data with at least some of the recording device RD, the auxiliary server 1000, the EMR server 3000, the hospital server 4000, or the terminal TD as needed. can

The controller 2500 may control overall operations of the main server 2000 . For example, the control unit 2500 may load and execute the tokenization model 2110, the embedding model 2120, or the pattern classification model 2130 of the memory unit 2100, and the pattern database 2140 You can search for the data you need from . The control unit 2500 may be implemented as a central processing unit (CPU) or a device similar thereto according to hardware or software or a combination thereof. It may be provided in the form of an electronic circuit that performs a control function by processing electrical signals in hardware, and may be provided in the form of a program or code that drives a hardware circuit in software.

[Prescription assistance method]

Hereinafter, a prescription assistance method performed by the prescription assistance system 10 will be described with reference to FIGS. 4 to 11 .

4 is a flowchart illustrating a prescription assisting method according to an exemplary embodiment.

5 is a diagram illustrating a process in which the prescription assistance system 10 performs a prescription assistance method according to an embodiment.

Referring to FIG. 4 , the prescription assistance method includes recording the interview process between the doctor and the patient (S1100), acquiring the interview record between the doctor and the patient (S1300), and analyzing the pattern of the interview record (S1500). ), and outputting supplementary prescription information (S1700).

Referring to FIGS. 4 and 5 , recording the interview process between the doctor and the patient ( S1100 ) may be performed by the recording device RD.

The recording device RD may initiate recording in a variety of ways. For example, the recording device RD may start recording by receiving a doctor's input. Here, the doctor's input may include button input, voice input, or gesture input. For another example, the recording device RD may detect when a patient enters the treatment room through a separately provided sensor and start recording. As another example, the recording device RD may start recording by receiving data instructing to start recording from at least one of the components of the prescription assistance system 10 .

The recording device RD may transmit data to the main server 2000 or receive data from the main server 2000 . For example, the recording device RD may transmit acquired sound data to the main server 2000 . For another example, the recording device RD may receive data instructing to start recording from the main server 2000 .

Referring to FIGS. 4 and 5 , obtaining a record of an interview between a doctor and a patient ( S1300 ) may be performed by the auxiliary server 1000 .

The auxiliary server 1000 may obtain audio data of a recorded interview between a doctor and a patient. For example, sound data generated by the recording device RD may be transmitted to the auxiliary server 1000 . For another example, sound data generated by the recording device RD may be transmitted to the main server 2000 and transmitted from the main server 2000 to the auxiliary server 1000 .

The auxiliary server 1000 may acquire the interview record by converting the acquired acoustic data. In this case, as described above, the interview record may include only data corresponding to the doctor's voice or both data corresponding to the doctor's voice and data corresponding to the patient's voice.

The secondary server 1000 may transmit interview records to the main server 2000 .

Referring to FIGS. 4 and 5 , the step of analyzing the interview record pattern ( S1500 ) may be performed by the main server 2000 .

The main server 2000 may determine a pattern for the interview record received from the auxiliary server 1000 . For example, the main server 2000 may output pattern data indicating a specific pattern by using the interview record and the pattern classification model 2130 .

The main server 2000 may generate prescription assistance information using the obtained pattern data. For example, the main server 2000 may obtain test sets or candidate test items corresponding to the obtained pattern data.

The main server 2000 may provide prescription assistance information to the EMR server 3000 or the terminal TD.

Referring to FIGS. 4 and 5 , the step of outputting supplemental prescription information ( S1700 ) may be performed by the terminal TD.

The terminal (TD) may acquire prescription assistance information. For example, the terminal TD may access the EMR server 3000 and acquire prescription assistance information stored in the EMR server 3000 . For another example, the terminal TD may acquire prescription assistance information from the main server 2000 .

The terminal TD may output the acquired supplementary prescription information through a display or the like.

In the above-described prescription assisting method, the step of analyzing the pattern of the interview record (S1500) can be understood as a process of identifying the doctor's practice pattern. As a doctor performs medical treatment for numerous patients, common parts exist among the results of the medical treatment (eg, prescription drugs or prescription test items), and the results of the medical treatment can be classified or classified based on the common parts. and the treatment processes corresponding to the differentiated outcomes can be defined as one treatment pattern.

In other words, since each of the treatments performed by the doctor on the patients can be classified into a specific treatment pattern, if only the pattern of treatment performed by the doctor on the new patient is determined, the result corresponding to the determined pattern is the outcome of the treatment. It can be. In this case, the doctor only needs to check the result and modify the result if necessary, thereby omitting the exhausting work of selecting or determining the result of treatment from the beginning.

Hereinafter, the process of the prescription assistance method will be described in more detail with reference to FIGS. 6 to 10 .

6 is a diagram illustrating a process of generating an interview record according to an exemplary embodiment.

The interview record is corpus (corpus) data, and may refer to text data obtained by converting sound data or data generated by processing the same. For example, an interview record may be understood as data including text corresponding to a doctor's voice, text corresponding to a patient's voice, and/or text corresponding to a third party's voice.

As described above, the auxiliary server 1000 may obtain acoustic data and generate a record of the interview using the acquired acoustic data.

Specifically, the auxiliary server 1000 transforms audio data into text data using the speech recognition model 1100, and converts the converted text data into text data using the speaker classification model 1200 to a speaker (ex. doctor, patient, or third person), etc., and text data corresponding to a specific speaker can be obtained as an interview record.

The voice recognition model 1100 may transform sound data into text data through processes such as feature extraction, pattern classification, and language processing of an input voice signal.

The voice recognition model 1100 may be implemented using a dynamic time warping (DTW) algorithm, a hidden Markov model (HMM) algorithm, a knowledge-based voice recognition technology, a neural network, and the like.

The speaker classification model 1200 may classify a speaker from text data obtained by converting sound data, and may select speaker text data to be included in the interview record. For example, the speaker classification model 1200 selects only a text corresponding to a doctor's voice from among text converted from sound data and obtains the interview record. For another example, the speaker classification model 1200 may obtain an interview record including texts corresponding to a doctor's voice and texts corresponding to a patient's voice among texts obtained by converting sound data.

On the other hand, the treatment pattern to be analyzed in the prescription assistance method is the doctor's treatment pattern, and in this case, the patient's voice in the interview is not essential and may act as noise. desirable. Hereinafter, it is described that the interview record includes only text corresponding to the doctor's voice, but the technical spirit of the present specification is not limited thereto, and the interview record includes text corresponding to the doctor's voice as well as other speakers (ex. patient or Of course, text corresponding to the voice of a third party) may be included.

The speaker classification model 1200 may identify speakers even before converting sound data into text data.

Meanwhile, when the speaker classification model 1200 is not needed, such as when texts corresponding to voices of all speakers during an interview are used as interview records, the speaker classification model 1200 may be omitted.

The above-described interview records are data to be analyzed by a prescription assisting method, and may be understood as interview records to be analyzed.

7 is a diagram illustrating a process of generating model input data according to an exemplary embodiment.

Referring to FIG. 7 , interview records may be processed into model input data through tokenization and embedding processes.

Model input data may refer to data input to the pattern classification model 2130 . To be input to the pattern classification model 2130, interview records need to be pre-processed in a specific format, and representative pre-processing processes are tokenization and embedding, which will be described later.

The tokenization model 2110 may segment the interview record obtained by converting the acoustic data into a plurality of tokens.

The tokenization model 2110 may use a token dictionary to tokenize the interview transcript.

Keywords that are standards for tokenizing text data may be stored in the token dictionary, and tokenization may be performed based on keywords stored in the token dictionary.

Here, the keywords may basically include a specific morpheme, a specific word, or a specific phrase. Keywords may be set differently depending on which of the tokenization techniques described later is used.

A variety of tokenization techniques may be used in implementing the tokenization model 2110. For example, the tokenization model 2110 includes word-based tokenization, character-based tokenization, subword tokenization, byte pair encoding (BPE). encoding), wordpiece, SentencePiece, or Unigram.

In this case, a tokenization scheme for implementing the tokenization model 2110 may be determined according to a base model of a pattern classification model 2130 to be described later. For example, if the pattern classification model 2130 is generated using a bidirectional encoder representations form transformer (BERT) model, the tokenization model 2110 may be implemented as a word piece. For another example, when the pattern classification model 2130 is generated using a generative pre-trained transformer 3 (GPT-3) model, the tokenization model 2110 may be implemented with byte pair encoding.

In the tokenization process, the tokenized data of the interview record may include a start token indicating the start point at the front (ex. [CLS]) and an end token indicating the end point at the end (ex. [SEP]). can

The embedding model 2120 may change the data obtained by tokenizing the interview record into a number or code form.

The embedding model 2120 may be implemented using various embedding techniques. For example, in implementing the embedding model 2120, one-hot encoding, term frequency-inverse document frequency (TF-IDF), latent semantic analysis (LSA), Word2Vec, and glove , or FastText, etc. may be used.

The embedding model 2120 may be classified according to a method of specifying a target to be embedded. For example, there may be token embedding performed for each token or word, segment embedding performed for sentences connecting tokens, and position embedding for embedding the order of tokens. And, when the embedding model 2120 performs embedding on tokenized interview records, it can perform embedding on at least one of the above-mentioned subjects.

Meanwhile, an implementation algorithm and an embedding target of the embedding model 2120 may be determined according to a basic model of the pattern classification model 2130 to be described later. For example, when the basic model of the pattern classification model 2130 is the BERT model, the embedding model 2120 uses the Word2Vec algorithm and can perform token embedding, segment embedding, and position embedding.

The prescription assistance system 10 may obtain prescription assistance information using model input data.

The main server 2000 may obtain candidate examination items using the model input data and the pattern classification model 2130 .

Hereinafter, a process of obtaining candidate examination items in the main server 2000 will be described with reference to FIGS. 8 to 10 .

8 is a diagram illustrating an input/output process of a pattern classification model according to an embodiment.

9 is a diagram illustrating a pattern database according to an exemplary embodiment.

The pattern classification model 2130 may receive model input data and output pattern data indicating a specific pattern. After the pattern is classified, that is, after the pattern data is output, prescription auxiliary information corresponding to the pattern data can be obtained.

Referring to FIG. 8 , the pattern classification model 2130 may receive model input data and output probability values for each of the first pattern, the second pattern, and the third pattern. At this time, it can be understood that probability values for each of the patterns constitute pattern data.

Meanwhile, the pattern classification model 2130 may output pattern data indicating a specific pattern. For example, the pattern classification model 2130 may output pattern data indicating the first pattern.

The main server 2000 may determine a pattern for model input data using pattern data output from the pattern classification model 2130 . For example, as shown in FIG. 8 , the main server 2000 may determine the first pattern having the highest probability value as model input data or an interview record pattern before being processed into model input data.

The main server 2000 may obtain a test set based on pattern data or the determined pattern. For example, when a pattern determined by considering pattern data is a first pattern, a first test set corresponding to the first pattern may be obtained. For another example, when pattern data output from the pattern classification model 2130 indicates a first pattern, a first test set corresponding to the first pattern may be obtained.

The main server 2000 may obtain candidate test items based on the obtained test set. For example, when the test set obtained from the main server 2000 is the first test set, the first test item, the second test item, and the third test item included in the first test set are candidate test items. can be obtained with

The pattern database 2140 may be used in a process in which the main server 2000 acquires test sets and candidate test items. Specifically, referring to FIG. 9 , a first pattern in the pattern database 2140 corresponds to a first test set, and the first test set may include test A, test B, and test C. If the first pattern is determined as the pattern of the model input data or the interview record, the main server 2000 consequently obtains test A, test B, and test C included in the first test set as candidate test items. can

As described above, when specifying a pattern of an interview record using the pattern classification model 2130, information on a test set corresponding thereto is essential. In other words, the pattern database 2140 can be understood as a component that must be prepared in order to output pattern data from the pattern classification model 2130 and obtain practical information on the test set therefrom.

Meanwhile, data output from the pattern classification model 2130 does not necessarily have to be pattern data indicating a certain pattern. For example, the pattern classification model 2130 may output probability values for each test set, a specific test set, or data indicating the probability values.

The pattern classification model 2130 may use a deep learning algorithm. In this case, the training data sets may be used to create the pattern classification model 2130. A method of generating the pattern classification model 2130 and a method of generating training data sets will be described later.

In the above, when the pattern classification model 2130 outputs pattern data corresponding to the model input data, it can be understood that the model input data is classified into a specific pattern.

Here, the meaning of pattern classification is to find out what pattern the model input data, the interview record before the model input data is processed, the acoustic data before being converted into the interview record, and furthermore, the doctor's treatment for the patient has.

When a doctor treats a patient, the doctor prescribes necessary tests based on a disease or symptom the patient suffers from and patient information. At this time, when looking at medical records formed by treating many patients, tests prescribed for patients may have the same or high similarity. More specifically, when the doctor prescribes the first to nth test sets for the first to nth patients (n is a natural number greater than or equal to 2), at least some of the first to nth test sets are the same or commonly included. The number of inspection items to be tested may be relatively large. In this case, a plurality of patterns may be formed by classifying the same prescribed test set or test sets having a high degree of similarity among the prescribed test sets into one pattern. In this case, when a doctor newly treats a patient, the treatment process may be classified into one of the previously formed patterns. In this way, in order for the pattern classification model 2130 to be trained to output a specific pattern or to output a value in consideration of the pattern, it is necessary to clearly classify and generate patterns from the learning data set as will be described later. In other words, if the deep learning model is trained with a training data set in which patterns are not clearly distinguished, the quality (or accuracy) cannot be guaranteed. It is a very important process to output data for a specific pattern.

10 is a diagram showing another output form of the pattern classification model 2130 according to an embodiment.

Referring to FIG. 10 , the pattern classification model 2130 may receive model input data and output one or more inspection items.

Here, inspection items output from the pattern classification model 2130 may constitute one inspection set.

Unlike the case described above, the pattern classification model 2130 does not output pattern data or inspection sets, but outputs inspection items constituting one inspection set, which is understood as eventually classifying model input data into a specific inspection set. It can be. In other words, although the output form is different from the pattern classification model 2130 shown in FIG. 8 , it is the same in that the doctor's practice pattern is considered in that the model input data is classified into a specific pattern or a specific test set.

As described in FIG. 8 , the pattern classification model 2130 may use a deep learning algorithm, and may be trained to receive model input data and output test items constituting one test set.

The main server 2000 may use an additional examination model in addition to the pattern classification model 2130 in generating prescription assistance information.

More specifically, the main server 2000 may acquire prescription assistance information by inputting the acquired interview records or data obtained by processing the interview records to the pattern classification model 2130 and the additional examination model.

Here, the pattern classification model 2130 can be used to classify the pattern of the interview record as described above and obtain a test set according to the pattern.

Here, the additional examination model may be used to obtain examination items to be additionally performed based on the interview record.

The main server 2000 generates prescription auxiliary information using a test set obtained by using the pattern classification model 2130 for the interview record and at least one additional test item obtained by using the additional test model for the interview record. can do.

The purpose of using an additional examination model in addition to the pattern classification model 2130 is to generate supplementary information in consideration of the doctor's practice pattern, but when the doctor prescribes an examination item that was not usually prescribed, the patient additionally performs another examination. It is intended to generate supplementary information for prescription taking into account even exceptional circumstances, such as when desired. In this way, as the purpose of the additional examination model is to add exceptional examination items to prescription assistance information, exceptional examination items may not be used in implementing the existing pattern classification model 2130 . When exceptional test items are used in the implementation of the pattern classification model 2130, it can be a factor that can hinder the accuracy. As the exceptional test items are not used, as a result, the accuracy of the pattern classification model 2130 can be improved.

Additional inspection models will be described in detail later.

In the above, the case of using the interview record in obtaining prescription assistance information has been described, but the technical idea of the present disclosure is not limited thereto.

For example, the main server 2000 may generate supplemental prescription information using acoustic data related to interview content. Specifically, the main server 2000 receives acoustic data from the recording device RD, processes the received acoustic data to generate model input data, inputs the generated model input data to the pattern classification model 2130, Pattern data can be created.

Here, the model input data generated by processing the acoustic data is data obtained by processing the acoustic signal, and may be generated through a preprocessing process such as noise filtering and feature extraction. In this case, it may be understood that a doctor's treatment pattern is considered, but a sound signal rather than a text is used. In this case, the auxiliary server 1000 in the prescription auxiliary system 10 may be omitted. Also, in a process of generating a learning data set described later, the training data set may be composed of sound data and pattern data or sound data and a test set, rather than interview records.

Hereinafter, with reference to FIG. 11, a process of providing and using prescription assistance information to a doctor will be described.

11 is a diagram illustrating a method of providing prescription auxiliary information according to an exemplary embodiment.

Prescription assistance information obtained from the main server 2000 may be output through the terminal TD. Specifically, prescription assistance information is transmitted from the main server 2000 to the EMR server 3000, and the EMR server 3000 may output the prescription assistance information through the doctor's personal terminal TD.

Referring to FIG. 11, in the EMR program output from the terminal TD, patient information (name, resident registration number, gender, age, etc.) and medical treatment information (treatment date, treatment contents, suspected disease, medical image, etc.) are output, and Inspection A, inspection B, and inspection C can be output together as prescription auxiliary information. In this case, the prescription auxiliary information may be displayed as candidate test items or as a test set. Prescription auxiliary information may be displayed as candidate test items or required test items.

Meanwhile, as described above, when prescription assistance information is generated by using the additional examination model together with the pattern classification model 2130, examination items included in the examination set and additional examination items may be displayed separately. For example, as shown in FIG. 11, inspection A, inspection B, and inspection C included in the inspection set obtained through the pattern classification model 2130 are necessary inspection items, and inspection D obtained through the additional inspection model is It can be divided into additional inspection items and displayed.

In addition to outputting the prescription assistance information through the terminal TD, additional information may be further provided.

Specifically, the EMR server 3000 may obtain recommended examination items in consideration of candidate examination items that are supplementary prescription information. For example, the EMR server 3000 searches or retrieves past medical records of doctors to find a list of past prescription examination items including candidate examination items, and among the examination items included in the found past prescription examination items list, candidate examination items. Inspection items other than these may be obtained as recommended inspection items. Specifically, if the prescription auxiliary information is test A, test B, and test C, and there is a list of past prescription test items including test A, test B, test C, and test E among past medical records, test E is a recommended test item. can be obtained with

Recommended test items may be output through the terminal TD. The terminal TD may acquire recommended examination items from the EMR server 3000 and output the recommended examination items through the EMR program.

Meanwhile, there may be a plurality of past prescription examination item lists including candidate examination items. In this case, priorities may be assigned among past prescription examination item lists including candidate examination items. The priority may be determined in consideration of the number of occurrences of each past prescription examination item list among past medical records, that is, the number of times a doctor prescribes a corresponding past prescription examination item list.

Recommended test items may be output based on priority through the terminal TD. For example, the terminal TD may output five test items having a higher priority among recommended test items.

Candidate examination items may be verified by a physician. For example, as described above, candidate examination items may be output through the terminal TD, and final examination items may be determined based on a received doctor's input. To this end, as shown in FIG. 11 , the terminal TD may output a “modify” button and a “confirm” button together with candidate examination items, and the candidate examination items are selected based on the doctor's input to the corresponding buttons. It can be modified or confirmed as final inspection items.

The determined final inspection items may be provided to and stored in the EMR server 3000 . Also, final examination items may be provided and stored in the hospital server 4000 . In addition, final inspection items may be provided and stored in the main server 2000 .

Final inspection items can be provided to necessary departments within the hospital. For example, each of the final examination items may be transmitted to a terminal of an examination room where the corresponding examination item is to be performed, and preparations for the examination may be made in the examination room. For another example, final inspection items may be transmitted to a terminal of the hospital administration department, and documents such as reception, acceptance, verification, hospitalization/discharge may be performed in consideration of the final inspection items in the hospital administration department.

In the above, the process of performing the prescription assistance method in the prescription assistance system 10 has been described. A key technology essential in the prescription assistance method is to acquire prescription assistance information, and its accuracy may vary depending on how the pattern classification model 2130 of the main server 2000 is implemented (generated).

Hereinafter, a method of implementing the pattern classification model 2130 will be described in detail with reference to FIGS. 12 to 20 .

[How to implement a pattern classification model]

12 is a diagram illustrating an input/output form of a pattern classification model 2130 and a learning data set for learning the pattern classification model according to an embodiment.

The pattern classification model 2130 may be implemented using a deep learning algorithm.

Here, CNN (Convolution Neural Network), RNN (Recurrent Neural Network), LSTM (Long-Short Term Memory), Seq2seq (Sequence to sequence), Transformer, BERT (Bidirectional Encoder Representations from Transformers), or GPT- 3 (Generative Pre-trained Transformer-3) or the like may be used.

Referring to FIG. 12 , the pattern classification model 2130 may be trained to receive model input data and output pattern data. At this time, a learning data set for learning the pattern classification model 2130 may be prepared. The training data set may include pattern data for learning in the same form as pattern data in input data for learning in the same form as the above-described model input data.

For example, the doctor's first interview with the first patient is recorded to obtain the first interview record, the first interview record is processed through the above-described preprocessing process to obtain first learning input data, and the first study A first training data set may be generated by labeling input data with first pattern data indicating a first pattern. When the deep learning model learned using the generated first training data set receives the second interview record obtained by recording the doctor's second interview with the second patient, the first pattern for the second interview record It is possible to output first pattern data indicating or output pattern data indicating another pattern. When the first pattern data is output, it can be understood that the second interview record or the second interview content is similar to the first interview record or the first interview content, and the test items to be prescribed are the same.

As described above, in order for the pattern classification model 2130 to operate, generation of a training data set for training the pattern classification model 2130 and selection of a deep learning algorithm used as a basic model of the pattern classification model 2130 are very important. It is important.

Hereinafter, a method of generating a training data set will be described with reference to FIG. 13 .

13 is a flowchart illustrating a method of generating a training data set according to an exemplary embodiment.

Referring to FIG. 13 , the method for generating a learning data set includes data collection (S2100), test set determination (S2300), interview records classified (S2500), and interview records are labeled. A performing step (S2700) may be included.

Each step is described in detail below.

The main server 2000 may collect data necessary for generating a learning data set (S2100).

Prior to the description of the data collection process, the pattern classification model 2130 is for analyzing the treatment pattern of a doctor (hereinafter referred to as a 'target doctor') who receives prescription assistance information, and is a pattern classification model according to the target doctor. 2130 may be implemented differently. To this end, a training data set for training the pattern classification model 2130 may also vary depending on the target doctor.

For example, the learning data set may be obtained by collecting medical records of patients of a target doctor, classifying the collected medical records according to a predetermined method, and labeling a specific pattern or test set.

The learning data set may be generated based on medical records of a target doctor receiving prescription assistance information.

As described above, each of the medical records of the target doctor may include a record of an interview conducted with the patient by the target doctor and a list of prescription test items prescribed for the patient. For example, a first medical record of a target doctor for a first patient may include at least a first interview record and a first prescription test item list.

The interview record may refer to data recorded by a target doctor during treatment of patients. For example, the interview record may refer to a text obtained by converting audio data of a recorded interview, or at least a part of the converted text.

The prescription test item list may include test items prescribed for the patient by a target doctor. The prescription test item list may be understood as corresponding to the interview record. For example, when a target doctor treats a first patient, a first interview record and a list of first prescription test items may be collected.

The medical records are not only the medical records collected before the time point at which the pattern classification model 2130 starts to be trained, but also the interview records collected after the target doctor starts using the prescription assistance system 10. may include

Meanwhile, among the medical records, the interview records may include records generated by using data augmentation as well as records obtained by the target doctor actually treating the patient. A record created using the data augmentation technique may have a form similar to or include similar content to an actual interview record.

As described above, the learning data set may be generated by labeling a specific test set or a specific pattern in the interview records of the medical records.

Since test sets or patterns to be labeled must first be determined for labeling, a method for determining test sets or patterns prior to labeling will be described first. Meanwhile, if the test sets are specified, each test set may be determined as one pattern. Hereinafter, a method of determining the test sets will be described with reference to FIGS. 14 to 17 .

The main server 2000 may determine a test set for generating a learning data set (S2300).

14 is a diagram illustrating a method of generating test sets using frequently used test item lists according to an embodiment.

Examination sets may be created using frequently used examination item lists. Frequently used inspection item lists may be obtained in the above-described data collection step ( S2100 ).

In treating patients, the target doctor may group frequently prescribed examination items into a single list. This is because there are cases in which test items prescribed for patients are similar. For example, in the case of first-time patients, since there are basic test items that must be received for diagnosis, the prescribed test items may be the same for first-time patients complaining of the same symptoms. For another example, even in the case of a re-examination patient, there may be test items that need to be performed to monitor the progress of the disease or the effect of treatment, so if the diagnosed disease or symptom is the same or the prescribed drug is the same, the re-examination patient In the case of the above, the prescribed test items may be the same.

In general, a target doctor can prescribe for 70-90% of patients with only a few test item lists, and these test item lists can be entered as frequently used test item lists in an EMR program or the like.

Accordingly, when frequently used test item lists are used as test sets for classifying patterns of interview records, accuracy of prescription auxiliary information can be secured to a certain level or higher.

A list of frequently used examination items may be set by a target doctor. For example, as described above, the target doctor may arbitrarily group specific examination items for convenience in providing medical care to patients and create frequently used examination item lists.

A list of frequently used inspection items may be plural. The list of frequently used items may include first through m-th lists (m is a natural number greater than or equal to 2). Here, each of the first to m-th lists may include one or more inspection items. Also, at least some of the first to m-th lists may include the same inspection item. For example, referring to FIG. 14 , the frequently used test item list includes at least first to third lists, and the first to third lists may commonly include Free T3 and Free T4 test items. .

Meanwhile, frequently used examination item lists do not necessarily have to be set by a target doctor. A list of frequently used examination items may be set based on a medical record of a target doctor. For example, among the prescription test item lists included in the medical records of the target doctor, x number of prescription test item lists having a high number of appearances (x is a natural number of 2 or more) may be set as frequently used test item lists. .

The number of frequently used inspection item lists may be arbitrarily set. For example, when a list of frequently used examination items is set by a target doctor, as many lists as the number set by the target doctor may be set. For another example, when a list of frequently used examination items is set based on a medical record of a target doctor, the number of lists may be 5 or more, preferably 5 or more and 20 or less.

A plurality of test sets may be determined using frequently used test item lists. For example, each of frequently used test item lists may be determined as one test set. Alternatively, each of frequently used test item lists may define a pattern and become a test set of the pattern. Specifically, as shown in FIG. 14 , a first list among frequently used test item lists may be determined as a first test set of a first pattern.

Examination sets may be determined using prescription examination item lists of medical records.

15 is a diagram illustrating prescription test item lists according to an embodiment.

16 is a diagram illustrating a method of generating check sets according to an exemplary embodiment.

Referring to FIG. 15 , it can be seen that some prescription test item lists of medical records include specific test items.

Referring to FIG. 16 , test sets may be created by extracting test items from the prescription test item lists shown in FIG. 15 and considering the extracted test items and test set determination criteria.

The criteria for determining the test set are described below.

As a criterion for determining the test set, the possibility of co-occurrence may be considered.

For example, candidate test sets may be generated using test items extracted from prescription test item lists, and test sets may be determined by appropriately combining the generated candidate test sets. It is described in more detail below.

Prior to determining the check sets, the total number of required check sets may first be specified. For example, the total number of required check sets may be a (a is a natural number of 2 or more). Here, a may be determined from 5 or more to 30 or less, but is not limited thereto.

Next, each of the test items extracted from the prescription test item lists may be determined as candidate test sets. In this case, a test item having a low frequency of appearance may not be determined as a candidate test set. For example, if a test item appears less than b times in the prescription test item lists, the test item may not be determined as a candidate test set.

When the number of candidate check sets is more than a, the number of candidate check sets may be reduced by combining two or more of the candidate check sets. At this time, co-occurrence possibility can be used as a method of combining candidate test sets.

The co-occurrence possibility may refer to a probability that candidate test sets appear together in comparison to a probability that candidate test sets separately appear in the prescription test item lists. The co-occurrence possibility may be variously referred to as a co-occurrence judgment value, a judgment value, a probability value, or a reference value, and may be calculated through Equation (1) below.

For example, the co-occurrence probabilities may be calculated for all cases of selecting any two candidate test sets among the candidate test sets, and the two candidate test sets having the largest value among the calculated co-appearance probabilities can be combined into one.

The process of combining the candidate check sets described above may be repeated until the number of candidate check sets becomes a or less. For example, two test items with the highest co-occurrence probability among test items extracted from prescription test item lists are combined, and each of the combined set and the remaining non-combined test items are used as primary candidate test sets. can be configured. Thereafter, the two primary candidate test sets with the highest co-occurrence probability among the primary candidate test sets are combined again, and the combined set and the remaining uncombined primary candidate test sets constitute secondary candidate test sets. It can be. When n-th candidate test sets are determined by repeating the combination of the above-described candidate test sets, when the number of n-th candidate test sets is a or less, the n-th candidate test sets may be determined as final test sets.

In the above, the case where two candidate test sets with the highest co-occurrence probability are combined has been described in combining the candidate test sets, but the technical idea of the present specification is not limited thereto, and two pairs of Of course, each of the candidate test set groups (eg, a pair of candidate test set groups includes two candidate test sets) may be combined into one candidate test set.

Co-occurrence probability can be understood as a criterion for judging which combination of candidate test sets is suitable for patterning a target doctor's treatment or test prescription when the total number of required test sets is limited and the number of candidate test sets is greater than that. can

Other equations other than Equation (1) may be used as a method of calculating the co-occurrence probability. For example, values such as the number or ratio of lists including both candidate test sets among prescription test item lists may be calculated as co-occurrence possibilities.

Meanwhile, as a criterion for determining a test set, a likelihood may be considered.

For example, candidate test sets may be generated using test items extracted from prescription test item lists, and test sets that are mainly prescribed in a real situation may be determined by considering likelihood from the candidate test sets.

First, the total number of required test sets may be specified. For example, the total number of necessary test sets may be the above-mentioned a.

Next, the sum of likelihoods for all cases of generating a number of candidate test sets from the prescription test item lists is calculated, and a number of candidate test sets having the largest sum of calculated likelihoods are determined as the final test sets. can

Here, the likelihood of each candidate test set may be calculated using a probability value at which the corresponding candidate test set is prescribed when a distribution of actual prescription frequencies of the candidate test sets is given. In this case, a distribution of actual prescription frequencies of candidate test sets may be obtained using prescription test item lists.

As a criterion for determining the test set, the number of occurrences may be considered. Specifically, among the prescription test item lists, prescription test item lists having a relatively large number of occurrences may be set as the test set.

For example, among prescription test item lists, prescription test item lists whose number of occurrences is equal to or greater than the third threshold value may be determined as the test set. The third threshold value may be determined at 1/5 or more of the total number of prescription test item lists. Alternatively, the third threshold value may be determined at 1/10 or more of the total number of prescription test item lists. Alternatively, the third threshold value may be determined at 1/20 or more of the total number of prescription test item lists. Alternatively, the third threshold value may be determined between 1/20 and 1/5 of the total number of prescription test item lists.

As another example, the upper r number of prescription inspection item lists having a high number of appearances among the prescription inspection item lists may be determined as a test set. Here, r may be determined at 5 or more. Alternatively, r may be determined at 10 or more. Alternatively, r may be determined to be greater than or equal to 20. Alternatively, r may be determined from 5 or more and 20 or less.

In the above, the first method of generating a test set using frequently used test item lists and the second method of generating a test set using prescription test item lists have been described.

The test set may be generated using the first method or the second method described above, and may also be generated using the first method and the second method. For example, the test set may be created to include both test sets created using frequently used test item lists and test sets created using prescription test item lists. At this time, duplicate check sets may be deleted.

Meanwhile, in determining the test set, the number of test items included in each test set and the determined total number of test sets may affect the efficiency of patterning the treatment method of the target doctor. For example, patterns may be diversified as the total number of test sets increases, but it may be difficult to secure data required for learning for each pattern. In addition, as the number of test items included in the test set increases, patterns for classifying interview records may become clearer, but similarly, it may be difficult to secure data required for learning for each pattern. Therefore, the test set needs to be determined by considering the total number of test sets and the total number of test items included in each test set.

Meanwhile, in generating the test set, unsupervised learning may be used. For example, a test set may be obtained by clustering test items included in prescription test item lists of medical records. In this case, spectral clustering may be used as a clustering technique.

A process of generating a check set using spectral clustering may proceed as follows.

First, test items included in prescription test item lists of medical records may be extracted, and an affinity matrix may be generated using the extracted test items.

Here, the similarity matrix may refer to a matrix in which test items are arranged on the x-axis and y-axis, and similarities between the test items are calculated and recorded. The degree of similarity between the test items is calculated as the frequency or number of times that the first test item and the second test item are prescribed together based on the list of prescription test items in the medical records when there is an arbitrary first test item and the second test item. It can be.

Next, a graph can be constructed based on the generated similarity matrix. In the constructed graph, each of the inspection items may be displayed as a node, and the similarity between the inspection items may be displayed in the form of an edge connecting the nodes. The graph can be constructed in the form of a fully connected graph in which all nodes are connected, an ε-neighborhood graph that displays only edges closer than ε, or a k-nearest neighbor graph in which only k neighbors around each node are connected.

Nodes can be divided into two or more clusters in the constructed graph, and nodes of each cluster, that is, check items of each cluster can be determined as one check set. Here, a technique such as minimum out or normalized out may be used as a method of classifying nodes into clusters.

Meanwhile, in the above, the case of using the spectral clustering technique has been mainly described, but the technical spirit of the present specification is not limited thereto, and K-means clustering, DBSCAN clustering, Affinity Propagation clustering, or Hierarchical clustering or the like may be used.

The main server 2000 may classify the interview records (S2500).

17 is a diagram illustrating a method of classifying interview records according to test sets and classification criteria, according to an exemplary embodiment.

The interview records may be classified using the test sets determined through the above method and the classification criterion.

Referring to FIG. 17 , the interview records of each of the medical records may be classified into one of several patterns. Here, each pattern may be understood as corresponding to each of the determined test sets. For example, the first pattern may correspond to the first test set, and the second pattern may correspond to the second test set.

Basically, the interview records of the medical record may be classified into patterns corresponding to the same examination set as the prescription examination item list of the medical record. For example, referring to FIGS. 15 to 17 , a first prescription test item list of a first medical record includes test A, test B, test C, test D, and test E, and the first test set is likewise When examination A, examination B, examination C, examination D, and examination E are included, the first interview record of the first medical record may be classified into a first pattern corresponding to the first examination set.

Meanwhile, if there is no test set identical to the list of prescription test items in the medical record, the interview record of the medical record may be classified as 'no pattern'. In this way, when classified as no pattern, a learning data set in which a specific interview record is labeled as 'no pattern' can be generated, and the pattern classification model 2130 learned through this can also output 'no pattern'. . When 'no pattern' is output, the target doctor can determine a test item to be newly prescribed, and therefore, it is possible to prevent a test prescription from being erroneously output due to incorrect candidate test items being output.

Alternatively, if there is no test set identical to the list of prescription test items in the medical record, the interview record of the medical record may not be used for learning.

Alternatively, when there is no test set identical to the prescription test item list of the medical record, similarity between the prescription test item list and each test set may be considered.

Specifically, the similarity between test items included in the list of prescription test items of the medical record and test items included in each test set may be compared, and the interview records of the medical record may be classified as the most similar test set.

When comparing the prescription test item list and the test set, it can be understood that the similarity increases as more test items included in the prescription test item list are included in the test set. In addition, it may be understood that the similarity increases as the number of test items not included in the prescription test item list among the test items included in the test set decreases.

In other words, if the interview record of the medical record is labeled with a specific pattern or a test set corresponding thereto, the test set corresponding to the specific pattern among the test sets has the largest number of test items included in the prescription test item list of the medical record. It can be understood as including inspection items.

For example, referring to FIGS. 15 to 17 , a fourth prescription test item list of a fourth medical record includes test A, test B, test C, test I, and test J, and the first test set is test A. , test B, test C, test D, and test E, and the third test set includes test A, test B, test C, test I, test J, and test P, the fourth prescription test item list and The number of test items commonly included in the third test set is greater than the number of test items commonly included in the fourth prescription test item list and the first test set. Therefore, the fourth prescription test item list has a higher similarity with the third test set than the first test set, and has the highest similarity with the third test set even when other test sets are considered, and thus the fourth interview record of the fourth medical record. may be classified as a third pattern corresponding to the third test set.

Meanwhile, there may be a plurality of test sets including the most test items included in the prescription test item list of the medical record. In this case, the number of test items included in each test set and/or the number of test items included in the prescription test item list may be considered.

For example, referring to FIGS. 15 and 16 , the second prescription test item list of the second medical record includes test A, test B, test C, test D, test E, and test F, and the first test set is A fifth test set may include test A, test B, test C, test D, and test E, and the fifth test set may include test A, test B, test C, test D, test E, test K, and test L. In this case, the number of test items common to the second prescription test item list is five in both the first test set and the fifth test set.

In this case, the interview records of the corresponding medical records may be classified into test sets containing a large number of test items or patterns corresponding thereto. For example, referring to FIG. 17 , the second interview record of the aforementioned second medical record is classified into a fifth pattern corresponding to a fifth test set having a greater number of test items among the first test set and the fifth test set. It can be.

In this way, it may be understood that the reason why the interview records are classified into test sets having a relatively large number of test items is to prevent omission of test items to be received by the patient. In addition, it may be advantageous in terms of convenience to delete unnecessary examination items rather than adding examination items omitted by the target doctor in the above-described verification process.

However, contrary to the foregoing, it is needless to say that the interview records may be classified into test sets with a relatively small number of test items.

As another example, referring to FIGS. 15 and 16 , the sixth prescription test item list of the sixth medical record includes test A, test B, test C, and test D, and the first test set includes test A, test B, A fifth test set may include test A, test B, test C, test D, test E, test K, and test L. In this case, the number of test items common to the sixth prescription test item list in both the first test set and the fifth test set is four.

In this case, the number of test items included in each test set is greater than the number of test items included in the prescription test item list, and the interview record may be classified into a test set having a smaller number of test items or a pattern corresponding thereto. there is. For example, referring to FIG. 17 , the sixth interview record of the aforementioned sixth medical record is classified into a first pattern corresponding to the first test set having fewer test items among the first test set and the fifth test set. It can be.

As such, it may be understood that the classification of the interview records into test sets having a relatively small number of test items is to prevent unnecessary increase of test items to be received by the patient.

However, contrary to the foregoing, it goes without saying that the interview records may be classified into test sets having a relatively large number of test items.

Meanwhile, one interview record may be classified into two or more test sets or patterns. For example, if the prescription test item list of the medical record is the union of test items included in two test sets, the interview record of the corresponding medical record may be classified into each of the two test sets or two patterns corresponding to each of the two test sets. there is.

The main server 2000 may label the classified interview records (S2700).

A training data set may be created using the interview transcript and examination set. For example, the learning data set may be generated by labeling an examination set (or a pattern corresponding thereto) in which the interview record is classified in the interview record.

The training data set may be generated in consideration of the input data type and the output data type of the pattern classification model 2130 .

For example, as described with reference to FIG. 8 , model input data obtained by processing interview records may be input to the pattern classification model 2130 , and pattern data may be output from the pattern classification model 2130 .

In this case, the training data set may be generated by labeling model input data obtained by processing the interview record with pattern data indicating a pattern in which the corresponding interview record is classified. Alternatively, the training data set may be generated by labeling probability values of each pattern (eg, a pattern in which the corresponding interview record is classified is 1, and the others are 0) in model input data obtained by processing the interview record.

As another example, when model input data is input to the pattern classification model 2130 and a test set is output from the pattern classification model 2130, the learning data set is the test interview record in which the interview record is classified in the model input data processed by the interview record. Sets can be labeled and created.

As described above, a pattern may be labeled or a test set may be labeled in the interview record, but as a result, it may be understood that what is provided to a target doctor is a test set or test items included in the test set.

Specifically, when a learning data set is created by labeling an interview record with a pattern, the output form of the pattern classification model 2130 learned through this becomes a specific pattern, and the test set corresponding to the specific pattern output to the target doctor or The test items that the test set contains will be provided. In addition, when the learning data set is generated by labeling the test set in the interview record, the output form of the pattern classification model 2130 learned through this becomes a specific test set, and the test set or the test set included in the test set is similar to the target doctor. Inspection items will be provided.

As another example, as described in FIG. 10 , model input data may be input to the pattern classification model 2130 and inspection items may be output from the pattern classification model 2130 . In this case, the learning data set may be generated by labeling test items included in a test set in which the corresponding interview record is classified in model input data obtained by processing the interview record. Alternatively, the learning data set may be generated by labeling each of the test items included in the test set in which the corresponding interview record is classified in model input data obtained by processing the interview record, and a plurality of training data sets are generated for one interview record. It can be.

Hereinafter, an artificial neural network model used to generate the pattern classification model 2130 will be described with reference to FIGS. 18 to 20 .

18 and 19 are diagrams illustrating the structure of a pattern classification model 2130 according to an exemplary embodiment.

20 is a diagram illustrating a structure of an encoder constituting a pattern classification model 2130 according to an embodiment.

Various artificial neural network models are used in natural language processing (NLP) technology. For example, as described above, CNN, RNN, LSTM, Seq2seq, Transformer, BERT, or GPT-3 may be used.

The pattern classification model 2130 may be generated by learning any one of the above-described artificial neural network models with training data sets.

On the other hand, when the amount of data for learning is insufficient, it may be more efficient to train an artificial neural network already trained with other data using collected data rather than learning the artificial neural network from the beginning. For example, among the above-described artificial neural network models, BERT is implemented using a transformer and is a language model pre-trained with a large amount of unlabeled text data. A natural language processing model having a desired input/output shape can be created through fine-tuning in which a separate neural network is added to the pre-trained BERT and further trained with the collected data to readjust the parameters of the BERT. For another example, GPT-3 has also been learned using hundreds of billions of data sets, and can be used to generate a model having a desired input/output shape through fine tuning.

The pattern classification model 2130 may be generated using a pre-trained model, and in this case, the pre-trained model is referred to as a base model.

Meanwhile, the pattern classification model 2130 may be implemented by learning a base model using transfer learning.

For example, in generating the pattern classification model 2130, BERT may be used as a base model. Referring to FIG. 18 , the pattern classification model 2130 may have a structure in which a classifier is combined with a pretrained BERT model having a plurality of input terminals. Here, the classifier may include a feed forward layer and a softmax function.

BERT includes a plurality of encoders, and each encoder has the same structure as that of the transformer. For example, referring to FIG. 20, BERT includes N encoders, and each encoder includes a multi-head attention layer and a position-wise feed forward layer. can do.

In the multi-head attention layer, a plurality of self-attention is connected in parallel, and in each self-attention layer, a weight value that an input value has with respect to values input to other self-attention layers can be calculated.

The position-wise feed forward layer may serve as a fully-connected layer, and values to be input to the next encoder may be calculated.

Depending on the input/output type of the pattern classification model 2130, the input/output type of BERT used may also vary.

Referring to FIG. 18 , the pattern classification model 2130 may have a structure that receives model input data obtained by processing interview records and outputs a probability value for each pattern. The pattern classification model 2130 may learn the aforementioned training data sets. Here, the training data sets are, for example, a first training data set in which data obtained by processing a first interview record is labeled with a first pattern, and a second training data set in which data obtained by processing a second interview record is labeled with a second pattern. etc. may be included.

Referring to FIG. 19 , sentences or paragraphs of the interview record may be classified and input to BERT, and values output from BERT may be values corresponding to specific examination items through a feedforward layer.

In generating the pattern classification model 2130, it is also possible to use a model other than the BERT model as a base model. For example, when generating the pattern classification model 2130, GPT-3 may be used as a base model.

In the above, the implementation method of the pattern classification model 2130 has been described. The purpose of the pattern classification model 2130 is to output data related to a test set or pattern by receiving an interview record in consideration of a medical treatment pattern of each doctor.

The pattern classification model 2130 may be implemented differently for each doctor. Specifically, the learning data sets for training the pattern classification model 2130 are generated by considering the medical records of individual doctors, a first pattern classification model for a first doctor and a second pattern classification model for a second doctor. Even if the same data is input, the output may be different.

The pattern classification model 2130 may be implemented in the main server 2000 as described above. Specifically, the main server 2000 acquires the medical records of the target doctor from the EMR server 3000 and/or the hospital server 4000, and generates a pattern classification model 2130 through training data set generation and model learning. can

Meanwhile, the creation of the learning data set may be performed by a server other than the main server 2000 or an external learning server not included in the prescription assistance system 10 . At this time, the main server 2000 may learn the pattern classification model 2130 by obtaining learning data from another server or an external learning server. Alternatively, the pattern classification model 2130 may be learned in another server or an external learning server and provided to the main server 2000 .

In the above, the method of using the interview records has been mainly described in using the pattern classification model 2130, but the technical spirit of the present disclosure is not limited thereto, and instead of the interview records, the audio data recorded with the interview contents may be used. there is. Specifically, the pattern classification model 2130 may be trained to output a test set by receiving data obtained by processing audio data recorded from interviews, and in the process, the above-described method of using interview records may be similarly applied. .

[Using similarity-based classification model]

In the above, a method of learning a deep learning model using training data sets as a method of classifying patterns of interview records has been described.

Meanwhile, pattern classification may be implemented using a text categorization or document classification (clustering) technique and a similarity measurement method as will be described later. Hereinafter, a case in which the main server 2000 includes a similarity-based classification model and the similarity-based classification model classifies patterns of interview records will be described.

The similarity-based classification model may classify patterns of interview records instead of the pattern classification model 2130 . Accordingly, the process of acquiring the interview record, the process of outputting candidate examination items, the process of verifying candidate examination items, the process of transmitting and storing data, and the like described in the above prescription assistance method can be applied as they are. However, at least a part of the tokenization process or the embedding process, which is a process of preprocessing the interview record, may be omitted or modified.

Meanwhile, the similarity-based classification model may be used together with the pattern classification model 2130.

Like the pattern classification model 2130, the similarity-based classification model may receive an interview record and output test set or pattern data.

The similarity-based classification model can vectorize the input interview records. For example, the similarity-based classification model may obtain an interview vector by vectorizing an input interview record.

As a method of vectorizing the interview transcript, the Doc2Vec package in which the entire interview transcript is vectorized may be used. Alternatively, the similarity-based classification model may vectorize words/sentences/paragraphs in the interview record using Word2Vec/Sent2Vec/Paragraph2Vec, etc., and obtain the average vector as the interview vector.

The similarity-based classification model may calculate the similarity of the interview vector with pre-stored pattern vectors.

Here, the pattern vectors may be understood as representing a treatment pattern of a target doctor as a vector. Each of the pattern vectors may correspond to a specific pattern. For example, the first pattern vector may correspond to the first pattern.

Pattern vectors may be generated based on medical records of a target doctor. A method for generating pattern vectors will be described later.

Similarity can be understood as a measure of how similar two vectors are. The degree of similarity may be calculated as cosine similarity, Euclidean distance, Jaccard similarity, or Levenshtein distance.

The similarity-based classification model may determine the pattern of the interview transcript based on the similarity between the interview vector and the pattern vectors. Specifically, the similarity-based classification model may determine a pattern corresponding to a pattern vector having the highest similarity with the interview vector among pattern vectors as the pattern of the interview transcript.

The pattern vector may be generated using medical records of a target doctor.

First, interview records of medical records of a target doctor may be clustered based on a preset criterion.

For example, interview records having the same corresponding prescription test item list may be clustered into the same group. Specifically, when the test items included in the first prescription test item list of the first medical record and the test items included in the second prescription test item list of the second medical record are the same, the first interview record of the first medical record and the second interview records of the second medical records may be clustered into the same group.

As another example, if corresponding prescription test item lists satisfy the same or similar conditions, the interview records may be clustered into the same group. Specifically, when comparing the test items included in the first prescription test item list of the first medical record with the test items included in the second prescription test item list of the second medical record, the number of test items not included in each other is previously determined. If the number is less than or equal to the set number, the first interview record of the first medical record and the second interview record of the second medical record may be clustered into the same group.

Meanwhile, in the process of clustering the interview records, the above-described method for determining the test set and the method for classifying the interview records may be used. For example, each of the frequently used examination item lists described in FIG. 14 may be defined as one group, and the interview records of the medical records may be classified into a specific group based on the prescription examination item list of the medical records. For another example, as described in FIG. 16, test sets are created using test items extracted from prescription test item lists, each of the test sets is defined as one group, and the prescription test item list of the medical record is created. Considering, the interview records of the medical records can be classified into a specific group. Since the method of classifying the interview records has been described in FIGS. 15 to 17, it will be omitted.

A pattern vector of a corresponding group may be calculated from the clustered interview records. For example, interview vectors of the same group may be vectorized to obtain interview vectors, and an average value of the obtained interview vectors may be obtained as a pattern vector.

[Using additional inspection models]

In the above, a method for obtaining a test set from interview records has been described.

As described above, when the test set considering the doctor's practice pattern or the test items included in the test set are provided as prescription-assisted information, the data used for learning is relatively small compared to predicting individual test items. However, the accuracy of the supplementary prescription information provided may be high.

However, providing the test set as supplementary information in consideration of the doctor's practice pattern only outputs the test set with a relatively high frequency of prescription, so individual test items with a relatively low frequency of prescription are omitted from the auxiliary information for prescription. As a result, the inconvenience of having to add the corresponding test items directly by the doctor arises.

Hereinafter, with reference to FIGS. 21 to 23 , a method of predicting an additional test item in addition to a test set to solve the above-described inconvenience and providing prescription assistance information with higher accuracy will be described.

21 is a diagram illustrating a method of using a basic inspection model and an additional inspection model according to an embodiment.

22 is a diagram illustrating an output form of an additional inspection model according to an exemplary embodiment.

The main server 2000 may obtain prescription assistance information using the basic examination model and the additional examination model. For example, referring to FIG. 21 , the main server 2000 acquires a basic inspection set by inputting model input data into a basic inspection model, and obtains additional inspection items by inputting model input data into an additional inspection model, Test items included in the test set and additional test items may be obtained as candidate check items.

Here, the basic inspection model may be understood as the same model as the pattern classification model 2130 described above. In other words, the basic inspection model may receive model input data and output an inspection set or pattern data corresponding to the inspection set.

Hereinafter, for convenience of description, a test set acquired through a basic test model is referred to as a 'basic test set', and a test item included in the basic test set is referred to as a 'basic test item'. In addition, since the description of the pattern classification model 2130 can be applied as it is to the basic inspection model, it will be omitted, and the additional inspection model will be described in more detail.

The additional inspection model may output additional inspection items by receiving an interview record or data obtained by processing the interview record. For example, referring to FIG. 21 , the additional inspection model may receive model input data and output additional inspection items. For another example, referring to FIG. 22 , the additional inspection model may receive model input data and output probability values for a plurality of inspection items, and the inspection item with the highest probability value is obtained as an additional inspection item. It can be. For another example, the additional inspection model outputs an additional inspection set, and inspection items included in the additional inspection set may be obtained as additional inspection items.

Data input to the additional inspection model may be the same as data input to the basic inspection model. For example, as shown in FIG. 21 , data input to the basic examination model and the additional examination model may be model input data obtained by processing interview records.

Meanwhile, data input to the additional inspection model may be different from data input to the basic inspection model.

Additional inspection items may be understood as inspection items other than the basic inspection items. Here, the basic inspection items are inspection items included in the inspection set output from the basic inspection model as described above. In other words, the additional test items are included in the rest of the test items except for the basic test items among all the test items prescribed by the doctor for the patients in the past.

Here, the test set including the basic test items may refer to the basic test model or the test sets described in the pattern classification model 2130 . For example, a test set including basic test items may refer to a frequently used test item list described in FIG. 14 or a test set generated from prescription test item lists described in FIG. 16 .

Meanwhile, additional test items may include test items that are not used in the above-described method for determining a test set. Specifically, as described in FIG. 16 , in the process of determining candidate test sets by extracting test items from prescription test item lists, some test items with a low frequency of occurrence may not be determined as candidate test sets. Test items not determined as candidate test sets may be included in additional test items.

Additional inspection models can be implemented using deep learning algorithms. CNN, RNN, LSTM, Seq2seq, Transformer, BERT, or GPT-3 may be used as a deep learning algorithm. The additional inspection model may be implemented similarly to the method of implementing the pattern classification model 2130 described above. For example, an additional test model may be created by training a training data set using BERT as a base model. In this case, the training data set for learning the additional inspection model may be different from the training data set for learning the pattern classification model 2130 .

A training data set for training the additional examination model may be created using the interview records and additional examination items.

23 is a flowchart illustrating a method of generating a training data set of an additional examination model according to an exemplary embodiment.

Referring to FIG. 23 , the method for generating training data for the additional examination model includes collecting data (S3100), determining a test set (S3300), classifying interview records (S3500), and determining additional examination items. (S3700), and labeling additional test items in the interview record (S3900).

The step of collecting data (S3100), the step of determining a test set (S3300), and the step of classifying interview records (S3500) are the same as steps S2100, S2300, and S2500 described in FIG. The description is omitted.

The main server 2000 may determine additional inspection items (S3700).

The additional test items may be determined based on the interview records, patterns or test sets in which the interview records are classified, and/or a prescription test item list corresponding to the interview records. For example, when the first interview record of the first medical record is classified as the first test set, the test included in the first prescription test item list of the first medical record among the basic test items included in the first test set Inspection item(s) other than the items may be determined as additional inspection item(s).

Additional inspection items may be determined individually for the interview transcript. Further, additional examination items are obtained when the examination set in which the interview records are classified and the list of prescription examination items corresponding to the interview records are not identical.

The main server 2000 may label additional examination items in the interview record (S3900).

The training data set may be created by labeling the interview record or data obtained by processing the interview record with an additional test item determined for the interview record. If there are a plurality of additional test items corresponding to the interview record, an additional test set including the additional test items may be labeled in the corresponding interview record to generate a learning data set. Alternatively, when there are a plurality of additional test items, a plurality of training data sets may be created by labeling each additional test item with respect to the interview record.

The training data set may be configured differently according to the output form of the aforementioned additional inspection model.

Meanwhile, additional inspection items may be determined when a preset condition is satisfied. For example, when candidate additional examination items are acquired by excluding basic examination items from among examination items included in the list of examination items prescribed by the target doctor according to the above-described method, additional candidate examinations whose number of appearances is equal to or greater than a preset number. Only items can be determined as additional inspection items.

Here, the preset number of times may be determined from 1 to 5 times or less. Alternatively, the preset number of times may be determined to be 5% or less of the total number of prescription test item lists used to generate the learning data set.

As described above, when an additional test item is determined only when a preset condition is satisfied, a problem of deterioration in prediction accuracy of the additional test model, which may occur due to an extremely low frequency prescription test item, can be prevented in advance.

Meanwhile, the number of additional inspection models may be plural.

24 is a diagram illustrating a case in which a plurality of additional inspection models are used according to an embodiment.

Referring to FIG. 24 , model input data may be input to a plurality of additional inspection models, and each of the additional inspection models may output a value indicating whether a specific additional inspection item is necessary.

For example, a first additional inspection model may receive model input data and output a value related to whether a first additional inspection item is needed, and a second additional inspection model may receive model input data and output a value related to a second additional inspection item. It is possible to output a value on whether or not is necessary, and the first additional inspection item and the second additional inspection item are different inspection items.

In this case, each additional examination model may be learned as a training data set including an additional examination item and an interview record corresponding to the additional examination item. For example, the first additional examination model may be learned with training data sets including the first additional examination item.

As described above, when a plurality of additional inspection models are used, accuracy may be improved in that the need for one additional inspection item can be predicted from one model.

In the main server 2000, the additional inspection model may be used only when a preset condition is satisfied. For example, whether to use an additional inspection model may be determined based on a value output from the pattern classification model 2130 . Specifically, as described in FIG. 8, when a probability value indicating a specific pattern is output from the pattern classification model 2130, an additional inspection item is determined using the additional inspection model only when the output probability value is less than or equal to a preset value. can be obtained

When the pattern classification model 2130 outputs a probability value indicating a specific pattern, the probability value is understood as the probability that the interview record input to the pattern classification model 2130 is the corresponding pattern. It can be, and if the probability is high enough, it was considered unnecessary to use an additional test model. However, the additional inspection model does not necessarily have to be used conditionally.

[Performer of Prescribing Auxiliary Methods and Their Variations]

Hereinafter, with reference to FIG. 25, a data transmission/reception relationship between components of the above-described prescription assistance system 10 will be described.

25 is a diagram illustrating a data transmission/reception relationship between components of the prescription assistance system 10 according to an embodiment.

First, the EMR server 3000 may transmit a medical treatment record of a doctor to the main server 2000 . At this time, the EMR server 3000 may transmit medical records stored therein to the main server 2000. Alternatively, the EMR server 3000 may acquire medical records from the hospital server 4000 and provide the obtained medical records to the main server 2000 .

The main server 2000 may learn a model using medical records. Specifically, the main server 2000 may generate learning data sets using medical records, and acquire the pattern classification model 2130 by learning an artificial neural network model with the generated learning data sets.

The terminal TD may transmit a control signal to the recording device RD. The terminal TD may receive the doctor's input and transmit a control signal instructing the recording device RD to start recording.

The recording device RD may perform recording by receiving the control signal. The recording device RD may record for a preset time. Alternatively, the recording device RD may end recording by receiving an additional control signal from the terminal TD.

The recording device RD may transmit sound data acquired through recording to the main server 2000 .

The main server 2000 may transmit sound data to the auxiliary server 1000 .

The auxiliary server 1000 may obtain an interview record by converting the acoustic data, and transmit the obtained interview record to the main server 2000 .

The main server 2000 may obtain prescription assistance information by performing pattern analysis or a prescription assistance method on the interview record.

The main server 2000 may transmit the obtained supplemental prescription information to the EMR server 3000 .

The EMR server 3000 may transmit prescription assistance information to the terminal TD.

The terminal TD may output supplemental prescription information through a display, receive a doctor's input, and perform verification of the supplementary prescription information. The terminal TD may obtain final prescription information through a verification operation.

The terminal TD may transmit final prescription information to the EMR server 3000, and the EMR server 3000 may store the final prescription information in the server. Alternatively, the EMR server 3000 may transmit final prescription information to the hospital server 4000.

Meanwhile, at least some of the components of the prescription assistance system 10 may be integrally formed. Here, integrally formed means that functions performed in two or more components are performed in one component.

For example, the secondary server 1000 and the main server 2000 may be integrally formed. For example, a voice recognition program of the auxiliary server 1000 may be installed in the main server 2000 to convert sound data into interview records in the main server 2000 . In this case, the auxiliary server 1000 in the prescription auxiliary system 10 may be omitted.

When the auxiliary server 1000 and the main server 2000 are integrally formed, or when the function of the auxiliary server 1000 is mounted on the main server 2000, the data transmission/reception relationship described in FIG. 25 is changed as follows. can

The main server 2000 may obtain an interview record by converting the acoustic data acquired from the recording device RD, and obtain prescription assistance information by performing pattern analysis on the obtained interview record.

Other than the above, the data transmission/reception relationship described in FIG. 25 may be applied as it is.

As another example, the secondary server 1000 and the EMR server 3000 may be integrally formed. For example, a voice recognition program of the auxiliary server 1000 may be installed in the EMR server 3000 to convert acoustic data into interview records in the EMR server 3000. In this case, the auxiliary server 1000 in the prescription auxiliary system 10 may be omitted.

When the auxiliary server 1000 and the EMR server 3000 are integrally formed, or when the functions of the auxiliary server 1000 are installed in the EMR server 3000, the data transmission/reception relationship described in FIG. 25 is changed as follows. can

The main server 2000 may transmit sound data acquired from the recording device RD to the EMR server 3000. The EMR server 3000 may obtain an interview record by converting the received acoustic data, and may provide the obtained interview record to the main server 2000 .

Other than the above, the data transmission/reception relationship described in FIG. 25 may be applied as it is.

As another example, the main server 2000 and the EMR server 3000 may be integrally formed. For example, the memory unit 2100 of the main server 2000 described above may be mounted in the EMR server 3000 so that the EMR server 3000 may perform a prescription assistance method. As another example, data and programs for executing an EMR program are embedded in the main server 2000, and the main server 2000 may provide an EMR interface to a doctor through the terminal TD. Hereinafter, for convenience of description, a case in which the function of the main server 2000 is installed in the EMR server 3000 and a prescription assistance method is performed in the EMR server 3000 will be described. In this case, the main server 2000 in the prescription assistance system 10 may be omitted.

When the EMR server 3000 and the main server 2000 are integrally formed, the data transmission/reception relationship described in FIG. 25 may be changed as follows.

The EMR server 3000 may generate the pattern classification model 2130 using medical records stored in the server or medical records acquired from the hospital server 4000 . For example, the EMR server 3000 may generate training data sets using medical records, and may generate a pattern classification model 2130 by training a base model using the generated training data sets.

The EMR server 3000 may acquire sound data from the recording device RD and transmit the acquired sound data to the auxiliary server 1000 .

The auxiliary server 1000 may transmit to the EMR server 3000 the interview record obtained by converting the acoustic data.

The EMR server 3000 may perform pattern analysis using the acquired interview record and the pattern classification model 2130 and obtain prescription assistance information.

The EMR server 3000 may transmit the acquired prescription assistance information to the terminal TD.

Other than the above, the data transmission/reception relationship described in FIG. 25 may be applied as it is.

As another example, the main server 2000, the EMR server 3000, and the hospital server 4000 may be integrally formed. For example, data and programs for executing an EMR program are loaded in the hospital server 4000 so that the hospital server 4000 can provide an EMR interface to a doctor through the terminal TD. Also, since the memory unit 2100 of the main server 2000 is mounted in the hospital server 4000, the prescription assistance method can be performed in the hospital server 4000. In this case, the main server 2000 and the EMR server 3000 in the prescription assistance system 10 may be omitted.

When the functions of the EMR server 3000 and the main server 2000 are installed in the hospital server 4000, the data transmission/reception relationship described in FIG. 25 may be changed as follows.

The hospital server 4000 may generate the pattern classification model 2130 using the doctor's medical records stored therein. Specifically, the hospital server 4000 may create a learning data set using medical records of target doctors stored therein, and may generate a pattern classification model 2130 by learning a base model.

The hospital server 4000 may obtain sound data from the recording device RD and transmit the acquired sound data to the auxiliary server 1000 .

The hospital server 4000 may acquire an interview record from the auxiliary server 1000 and obtain prescription auxiliary information by performing pattern analysis on the interview record.

The hospital server 4000 may transmit the acquired prescription assistance information to the terminal TD.

The terminal TD may transmit final prescription information to the hospital server 4000 .

Examples of integration described above may be combined with each other. For example, the auxiliary server 1000 and the main server 2000 may be integrally configured, and the EMR server 3000 and the hospital server 4000 may be integrally configured. For another example, the auxiliary server 1000, the main server 2000, the EMR server 3000, and the hospital server 4000 may be integrally configured.

In addition, in the case of the recording device RD, it may be mounted on at least a part of the components of the prescription assistance system 10 .

In the above, cases in which each component of the prescription assistance system 10 is integrated with each other have been described. On the other hand, information collected in the hospital requires considerable care in handling, including personal information about patients, and for this reason, it may be recommended to separate the server inside the hospital and the external server. In other words, the servers of the prescription subsystem 10, for example, the auxiliary server 1000, the main server 2000, the EMR server 3000, and the hospital server 4000 may all be located inside the hospital, Data communication with external servers may be blocked. However, the technical concept of the present disclosure is not limited thereto, and at least some of the prescription assistance systems 10 may be located inside the hospital and other parts may be located outside the hospital, and each server of the prescription assistance system 10 may be required. Depending on this, data communication with an external server may be possible.

[How to manage the main server]

Hereinafter, a method of modifying or supplementing the above-described prescription assistance system 10 will be described.

As described above, the main server 2000 of the prescription assistance system 10 may receive the interview record and output a test set. To this end, data for implementing the pattern classification model 2130 of the main server 2000 is required. It is essential, and the amount needs to be secured to some extent. In particular, the pattern classification model 2130 of the main server 2000 needs to be prepared for each doctor, and more than a certain number of medical records of each doctor are required.

At this time, for a minimum period of time (ex. 1 week, 2 weeks, 30 days, 1 month, 2 months, 3 months, or 6 months, etc.) so that the doctor can use the prescription assistance system 10 as quickly as possible, The pattern classification model 2130 is formed by collecting medical records or collecting as few medical records as possible (ex. 20, 30, 50, 100, 200, 300, 500, or 1000, etc.) A method of implementing and supplementing later may be used.

For example, based on the provision time of providing the pattern classification model 2130 to the doctor, the pattern classification model 2130 is implemented using the doctor's past medical records collected before the provision time, and the pattern classification model 2130 is collected after the provision time. The pattern classification model 2130 may be additionally trained using additional medical records of doctors.

As another example, the main server 2000 provides the pattern classification model 2130 to the doctor, receives the pattern data output from the pattern classification model 2130 or the doctor's feedback input on the examination set, and responds to the received feedback input. Based on this, at least some of the parameters of the pattern classification model 2130 may be readjusted.

On the other hand, in the case of examinations received by patients in hospitals, they may change over time. For example, there may be cases in which a test item of a new method is introduced in a hospital or a name of a test method or test item is changed due to a change in an existing test method.

Therefore, in preparation for the above situation, the pattern classification model 2130 may be periodically updated through additional learning and/or parameter readjustment. In this case, whether to proceed with the update may be determined by considering the error rate of the pattern classification model 2130 . Here, the error rate may be determined based on the aforementioned doctor's feedback input.

[Expanded prescription assistance system]

In the above, the prescription assistance method is performed in the prescription assistance system 10 and the prescription assistance information is obtained, and the case where the prescription assistance information is tests to be taken by the patient has been described.

Information to be derived and recorded during the doctor's treatment of the patient may include drugs prescribed to the patient, cooperation with other departments as needed, or the date of the next visit to the hospital, if necessary, in addition to examinations the patient should undergo.

Hereinafter, with reference to FIGS. 26 and 27 , a prescription assisting method performed by the prescription assisting system 10 when the prescription assisting information is about medicines will be described.

26 is a diagram showing an input/output form of a drug prescription model and a learning data set for learning the drug prescription model according to an embodiment.

27 is a flowchart illustrating a method of generating a training data set for training a drug prescription model according to an embodiment.

In the case of the drug prescription model, its input/output form is similar to the pattern classification model 2130 described above. For example, referring to FIG. 26 , the drug prescription model may receive model input data obtained by processing interview records and output a drug set or pattern data corresponding to the drug set. The drug set or pattern data can be used to obtain candidate drugs.

Like the pattern classification model 2130, the drug prescription model may be generated by learning an artificial neural network, and requires a learning data set for learning.

Here, the learning data set for learning the drug prescription model may be similar to the learning data set used for learning the pattern classification model 2130 and its generation method, and redundant content will be omitted.

Referring to FIG. 27 , the learning data set for training the drug prescription model includes collecting data (S4100), determining a drug set (S4300), classifying interview records (S4500), and It can be generated through the step of performing labeling (S4700).

The main server 2000 may obtain a medical record of a target doctor (S4100). Since this step is the same as step S2100 described in FIG. 13, details are omitted.

The main server 2000 may determine a medicine set (S4300).

For example, the main server 2000 may determine a medicine set using frequently used medicine lists.

Like the list of frequently used examination items described in FIG. 14, the frequently used drug list may be determined by a target doctor or set based on the target doctor's medical records.

A plurality of drug sets may be determined using frequently used drug lists. For example, each of frequently used drug lists may be determined as one drug set. Alternatively, each of the frequently used drug lists may define a pattern and become a drug set of the pattern.

As another example, the main server 2000 may determine a drug set using a prescription drug list in a medical record.

Here, the prescription drug list may be understood as data on drugs prescribed by a target doctor for a patient. Accordingly, the prescription drug list may include at least one drug prescribed for the patient.

Similar to that described in FIG. 16, drug sets may be created using drugs extracted from prescription drug lists. For example, drug sets may be determined using a drug set determination criterion such as likelihood or number of occurrences of the extracted drugs.

The drug sets may be determined using drug sets obtained using frequently used drug lists and drug sets obtained using prescription drug lists.

The main server 2000 may classify the interview records (S4500). Specifically, the main server 2000 may classify the interview record using the determined drug sets. For example, an interview record of a medical record may be classified into any one of the drug sets or a pattern corresponding thereto by using a prescription drug list of the medical record.

In classifying the interview records, the classification method described in FIG. 17 may be applied.

The main server 2000 may label the classified interview records (S4700).

A training data set may be created by labeling data indicating a drug set or pattern into which the interview record is classified on the processed data of the interview record. Similar to that described in step S2700, the drug set or pattern data labeled for the interview record may vary in consideration of the output form of the drug prescription model.

A drug prescription model may be generated using the learning data sets generated according to the above method. For example, a drug prescription model may be generated by training an artificial neural network model using the generated training data sets.

As medical treatment for a patient is performed, the medical information to be recorded may include not only the above-described prescription test item list and prescription drug list, but also information on a consultation request, information on the next visit date, and the like.

Such medical treatment information may be obtained using interview records similarly to the above-described prescription assisting method. In other words, similar to the pattern classification model 2130 for examination prescription or the drug prescription model for drug prescription, a model outputting consultation request information and a model outputting visit date information may be implemented.

Hereinafter, referring to FIG. 28, a comprehensive medical information providing system that calculates various medical information using interview records will be described.

The comprehensive medical information providing system has the same configuration as the above-described prescription assistance system 10, but the configuration of the main server 2000 may be different.

Referring to FIG. 28 , the main server 2000 of the comprehensive medical information providing system may include a test prescription model, a drug prescription model, a consultation decision model, and a visit date decision model.

The test prescription model is the same as the pattern classification model 2130 described above. The test prescription model may output a test set or pattern data corresponding to the test set using the interview record. The main server 2000 may output candidate test items using test set or pattern data.

As described above, the drug prescription model may output a drug set or pattern data corresponding to the drug set using the interview record. The main server 2000 may output candidate drug items using drug set or pattern data.

The collaboration decision model may receive the interview record and generate collaboration information.

For example, the collaborative consultation decision model extracts a specific keyword from the interview record, and based on the extracted keyword, obtains data indicative of whether or not the collaborative consultation is performed and which department requires the collaborative consultation.

Here, the keyword may be a word or sentence related to collaboration. For example, the keyword may mean a keyword for a first medical department to which the target doctor belongs, a keyword for a medical department such as a second medical department other than the first medical department, and a keyword indicating information about a doctor other than the target doctor. .

As another example, the consultation decision model may generate consultation information in consideration of the treatment pattern of the target doctor, like the pattern classification model 2130 . For example, the collaboration decision model may be generated using collaboration-related learning data sets. Here, the collaboration-related learning data sets include a step of acquiring medical records of a target doctor, a step of dividing the medical records of a target doctor into a first medical record for which collaboration has been determined and a second medical record for which collaboration has not been determined, and the first treatment Labeling data indicating consultation in the first interview record of the record or data indicating consultation target department to create a first learning data set and labeling data indicating no consultation in the second interview record of the second medical record and generating a second training data set. The collaboration decision model may be generated by learning the above-described artificial neural network or the base model with collaboration-related learning data sets.

The date of visit decision model may receive an interview record and output information about the date of visit.

For example, the date of visit determination model may extract a specific keyword from the interview record and obtain a candidate visit date or data indicating a candidate visit date based on the extracted keyword.

Here, the keyword may be a word or sentence related to the visit date. For example, the keyword may mean a date, a time, an adverb used with a date or time, a doctor's last uttered sentence in an interview record, or a patient's last uttered sentence.

Meanwhile, the above-described test prescription model, drug prescription model, consultation decision model, and hospital visit date decision model may have different input data.

For example, referring to FIG. 28 , an interview record including only the doctor's speech may be entered into the test prescription model, the drug prescription model, and the consultation decision model, and the doctor's speech and the patient's speech may be entered into the visit date decision model. An interview transcript containing all of the content may be entered. While the content of examination prescriptions, drug prescriptions, and consultation requests for patients depend on the doctor's speech content and not on the patient's speech content, the visit date decision model may depend on both the doctor's speech content and the patient's speech content. because there is

In addition, model input data obtained by processing the interview records may be input to the test prescription model and the drug prescription model, but the interview records may be input in an unprocessed state to the consultation decision model and the hospital visit date decision model. The test prescription model and drug prescription model are implemented in a way that considers the doctor’s practice pattern and may require preprocessing such as tokenization and embedding, whereas the consultation decision model and visit date decision model require keyword extraction before preprocessing. This is because unrecorded interview records need to be entered.

To this end, the comprehensive medical information providing system may perform the following functions.

The auxiliary server 1000 may convert the acoustic data to obtain a partial interview record on the doctor's speech and an entire interview record on the doctor's speech and the patient's speech. The secondary server 1000 may provide partial interview records and full interview records to the main server 2000 .

The main server 2000 processes the partial interview records and inputs the acquired partial model input data into the test prescription model and the drug prescription model, inputs the partial interview records into the consultation decision model, and inputs the entire interview records into the visit date decision model. can be entered.

On the other hand, depending on the implementation form of the test prescription model, drug prescription model, consultation decision model, and visit date decision model, the input data of each model is partial interview records, partial model input data obtained by processing partial interview records, full interview records, Alternatively, it can be the entire model input data obtained by processing the entire interview record.

The above comprehensive medical information providing system is significant in that it automates all necessary medical records for patients. Doctors can be freed from administrative tasks to some extent as the main contents of patient care, such as prescription test items and records of prescription drugs, are automated to request for consultation and record the date of visit. Accordingly, treatment time for the patient may be increased. At this time, the data required by each model of the comprehensive medical information provision system or the data to be entered, especially the contents of the interview record as described above, may vary, and accordingly, the data provided by the comprehensive medical information provision system may have a certain level of accuracy. can have

Embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded in a non-temporary computer readable recording medium. The non-transitory computer readable recording medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the non-transitory computer readable recording medium may be specially designed and configured for the present invention or may be known and usable to those skilled in the art of computer software. Examples of non-transitory computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROM and DVD, and magneto-optical media such as floptical disks ( magneto-optical media), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine language codes such as those produced by a compiler. The hardware device may be configured to act as one or more software modules to perform processing according to the present invention and vice versa.

The features, structures, effects, etc. described in the embodiments above are included in at least one embodiment of the present invention, and are not necessarily limited to only one embodiment. Furthermore, the features, structures, effects, etc. illustrated in each embodiment can be combined or modified with respect to other embodiments by those skilled in the art in the field to which the embodiments belong. Therefore, contents related to these combinations and variations should be construed as being included in the scope of the present disclosure.

In addition, although the embodiment has been described above, this is only an example and does not limit the technical idea of the present disclosure, and those skilled in the art to which the present disclosure belongs will not deviate from the essential characteristics of the present embodiment. It will be seen that various modifications and applications not exemplified above are possible. That is, each component specifically shown in the embodiment can be implemented by modifying it. And differences related to these variations and applications should be construed as being included in the scope of the present disclosure as defined in the appended claims.

10: Prescription assistance system
1000: secondary server
2000: main server
3000: EMR Server
4000: hospital server

Claims (17)

A method for providing prescription auxiliary information of a server,
acquiring medical records of a target doctor, each of the medical records including an interview record and a list of prescription test items;
generating learning data sets using the medical records;
generating a pattern classification model using the learning data sets;
obtaining a first analysis target interview record of the target doctor;
obtaining first candidate examination items using the first interview record to be analyzed and the pattern classification model; and
Including; providing the first candidate check items;
The step of generating the learning data sets,
generating test sets using the medical records;
In consideration of the first prescription test item list of the first medical record, data corresponding to the first test set among the test sets is labeled with data obtained by processing the first interview record of the first medical record. generating a training data set; and
A second training data set by labeling the data corresponding to the second test set among the test sets in consideration of the second prescription test item list of the second medical record to data processed from the second interview record of the second medical record. Including the step of generating,
The pattern classification model,
receiving data obtained by processing the first analysis target interview record and outputting data corresponding to any one of the test sets;
The first candidate test items are test items included in a test set corresponding to the output data;
The step of generating the check sets is,
Extracting test items included in the first prescription test item list of the first medical record to the nth prescription test item list of the nth medical record (n is a natural number equal to or greater than 2); and
Determining the test sets in consideration of the extracted test items,
The step of determining the test sets in consideration of the extracted test items,
generating first candidate test sets using the extracted test items, wherein each of the first candidate test sets includes at least two or more of the extracted test items;
Calculating a co-occurrence determination value of each of the first candidate test sets for the first to nth prescription test item lists; and
Selecting at least some of the primary candidate test sets as secondary candidate test sets based on a co-occurrence determination value of each of the primary candidate test sets.
method. According to claim 1,
At least some of the secondary candidate checksets are selected as the checksets.
method. delete According to claim 1,
In the step of selecting the secondary candidate check sets,
Among the primary candidate test sets, a primary candidate test set having the largest co-occurrence determination value is selected as the secondary candidate test sets.
method. According to claim 1,
The co-occurrence determination value is,
A value obtained by dividing the number of lists including both the first primary candidate test set and the second primary candidate test set among the first to nth prescription test item lists by n,
The value obtained by dividing the number of lists including the first primary candidate test set among the first to nth prescription test item lists by n, and the second primary candidate test set among the first to nth prescription test item lists Calculated by dividing the number of lists containing by the value obtained by multiplying the value divided by n,
method. According to claim 1,
The step of generating the check sets is,
Including performing the first and second steps,
The first step may include calculating co-occurrence determination values of each of the (M-1) th candidate test sets for the first to nth prescription test item lists, and
The second step is a step of selecting at least some of the (M-1) candidate test sets as M-order candidate test sets based on the co-occurrence determination value of each of the (M-1)-order candidate test sets. Including more,
M is a natural number greater than or equal to 3;
method. According to claim 6,
The step of generating the check sets is,
The first step and the second step are repeated until a preset condition is satisfied, and finally, after the second step is finished, obtaining a selected K-th candidate test set as the test set.
method. delete delete According to claim 1,
The first interview record of the first medical record is obtained by converting first sound data obtained by recording a first medical treatment process of the target doctor;
The first prescription test item list of the first medical record includes test items prescribed by the target doctor for the patient in the first treatment process.
method. According to claim 1,
The test items included in the first prescription test item list of the first medical record are the same as the test items included in the first test set.
method. According to claim 1,
The first test set includes test items included in the first prescription test item list of the first medical record.
method. According to claim 1,
The step of generating the learning data sets,
Labeling data indicating that there is no pattern in the data processed from the third interview record of the third medical record to generate a third learning data set;
None of the test sets is the same as the third prescription test item list of the third medical record,
method. According to claim 1,
The step of generating the learning data sets,
Generating a third learning data set by labeling data corresponding to the first test set to data obtained by processing a third interview record of a third medical record;
The number of test items commonly included in the third prescription test item list of the third medical record and the first test set is equal to the number of test items commonly included in the third prescription test item list and the second test set. do,
The number of test items included in the first test set is greater than the number of test items included in the second test set.
method. According to claim 1,
The step of generating the pattern classification model,
selecting a pre-learned base model with data different from the training data sets; and
Including; learning the base model using the training data sets;
method. According to claim 15,
The base model is a bidirectional encoder representations form transformer (BERT) or a generative pre-trained transformer 3 (GPT-3),
method. According to claim 1,
recording a treatment process for a patient of the target doctor; and
Acquiring the first analysis target interview record from the recorded medical treatment process; further comprising,
method.

Images