JP7469775B2 - Hospitalization Probability Prediction System - Google Patents

Description

The present invention relates to a hospitalization probability estimation system.

In the field of psychiatric care, visiting nursing services are provided to patients in coordination with the medical services provided by hospitals. These visiting nursing services involve nurses visiting patients' homes to provide various types of care to the patient.

When a patient's condition worsens, they end up being admitted to a hospital. This means that the business entity providing the home nursing service is no longer able to provide services, putting a strain on profits. Therefore, it is important to detect the deterioration of a patient's condition in advance, provide appropriate care before hospitalization becomes necessary, and prepare for the patient's hospitalization, both in terms of improving services for patients and stabilizing management.

In this regard, a technology has been proposed that processes the text information recorded in medical records, vectorizes the frequency of words along with the timing of recording, and calculates the probability that a particular symptom is occurring based on the frequency of occurrence of words that indicate a particular symptom (for example, Patent Document 1).

JP 2018-169795 A

However, although this technology can calculate the probability of a particular symptom occurring, it cannot estimate the possibility of a patient being hospitalized. As a result, with this technology, determining whether a patient should be hospitalized requires a doctor to consider multiple symptoms in combination, which does not solve the problem mentioned above.

Therefore, there is a need for a hospitalization probability estimation system that can estimate the possibility of hospitalization of a patient based on information obtained through home nursing care.

The present invention provides a hospitalization probability estimation system that includes a nursing record acquisition unit that acquires nursing records, which are records of psychiatric home nursing care; an objective variable addition unit that adds an objective variable indicating whether or not a patient will be hospitalized to the acquired nursing record; a teacher data generation unit that generates teacher data by vectorizing text data indicating the patient's condition or nursing content contained in the nursing record; a learning unit that trains a model using the teacher data to generate a trained model; an inference data generation unit that generates inference data by vectorizing text data indicating the patient's condition or nursing content contained in the nursing record to be processed, which is separately acquired by the nursing record acquisition unit; and a hospitalization probability estimation unit that applies the inference data to the trained model to estimate the possibility of hospitalization.

The present invention provides a hospitalization probability estimation system that estimates the possibility of hospitalization of a patient based on information obtained through home nursing care.

FIG. 1 is a block diagram showing a configuration of a hospitalization probability estimation system. FIG. 13 is a diagram showing the data structure of a user information file. FIG. 2 is a diagram showing the data structure of a nursing record file. FIG. 2 is a diagram showing the data structure of a teacher data file. FIG. 13 is a diagram showing the structure of an inference data file. 13 is a flowchart showing the operation of the control unit of the hospitalization probability estimation system for generating training data and learning a model. 13 is a flowchart showing the operation of an inference process by a control unit of the hospitalization probability estimation system. FIG. 1 is a diagram illustrating the flow of word embedding. FIG. 13 is a diagram showing an example of output of hospitalization risk as an inference result. FIG. 13 is a diagram showing an example of output of an increase or decrease in hospitalization risk. FIG. 13 illustrates an example output of hospitalization risk statistics. FIG. 13 is a diagram showing an example output of hospitalization risk statistics trends. FIG. 13 is a diagram showing an example of output of a transition in hospitalization risk by patient.

The hospitalization probability estimation system according to one embodiment of the present invention is described in detail below.

Figure 1 is a block diagram showing the configuration of the hospitalization probability estimation system 1 of this embodiment. As shown in Figure 1, the hospitalization probability estimation system 1 can be configured using a computer such as a so-called server device or a personal computer.

Specifically, the hospitalization possibility estimation system 1 includes a control unit 101, a memory unit 102, an input/output unit 103, and a communication unit 104.

The control unit 101 includes a calculation device such as a CPU (central processing unit).

The storage unit 102 includes storage devices such as memory and a hard disk drive. The storage unit 102 stores a user information file 102A, a nursing record file 102B, a transaction file 102C, a teacher data file 102D, an inference data file 102E, an inference result file 102F, a tokenizer 102G, an embedding layer 102H, and a model 102I.

The input/output unit 103 includes an input device such as a keyboard, a pointing device such as a mouse, a display device such as a display, a printing device such as a printer, etc.

The communication unit 104 includes a communication device such as a network device, and communicates with the cloud 3 and the client terminal 4 via a public communication line network 2 such as the Internet.

Figure 2 shows the data structure of user information file 102A. User information file 102A stores information about users.

As shown in FIG. 2, the user information file 102A stores a user ID, which is an identifier uniquely assigned to a user, and user information, which is information about the user for each user ID.

User information includes, for example, name, date of birth, place of business, usage status, medical care expense category, etc.

Figure 3 shows the data structure of nursing record file 102B. Nursing record file 102B stores the contents of nursing records created during home visits to psychiatric patients for each user ID.

As shown in FIG. 3, the nursing record file 102B stores basic visit information, basic ADL (Activities of Daily Living) information, vital signs, and nursing care details for each user ID.

Basic visit information includes, for example, the name, age at the time of the visit, medical care expense category, place of business, start date and time of the visit, end date and time of the visit, an absence flag indicating that the patient was absent, duration of stay, a special instruction visit flag indicating that it was a special support visit, a 5 visits per week after discharge flag indicating that the visit occurred within 5 weeks of discharge, a multiple visit flag indicating that multiple visits occurred in one day, and a multiple visit flag indicating that the visit was made by multiple people.

Examples of basic ADL information include eating, sleeping, taking medicine, cleaning, environmental improvement, financial management, interpersonal relations, eating, moving, excretion, changing clothes, bathing, grooming, eating (notes), sleeping (notes), taking medicine (notes), cleaning (notes), environmental improvement (notes), financial management (notes), interpersonal relations (notes), eating (notes), moving (notes), excretion (notes), changing clothes (notes), bathing (notes), and grooming (notes). Here, the above (notes) may include text data.

Vital signs include, for example, body temperature, pulse rate, systolic blood pressure, and diastolic blood pressure.

Examples of nursing care content include attentive listening, coordination with the patient's doctor and related institutions, administration of medication and injections, treatment of skin and bedsores, assistance with medical examinations, guidance and consultation for family members, rehabilitation, advice, notes on nursing care content (text), changes in symptoms, content of symptoms (text), and special notes (text). Here, the items marked as (text) above may include text data.

Figure 4 shows the data structure of the teacher data file 102D. The teacher data file 102D stores teacher data for training a model for each user ID.

As shown in FIG. 4, the teacher data file 102D stores, for each user ID, visit information, ADL information, vital signs, nursing care content, month of visit, vector text which is vectorized text, and objective variables.

Examples of home nursing information include age at time of visit, length of stay, number of days since last visit, number of absences (past 14 days), change in symptoms, special instruction visit flag, 5 visits per week after discharge flag, multiple visit flag, and multiple visits flag.

Examples of ADL information include eating (regular), eating (irregular), eating (refusal), eating (overeating), eating (other), sleep (good sleep), sleep (insomnia), sleep (difficulty falling asleep), sleep (early awakening), sleep (awakening during the night), sleep (other), medication (regular), medication (irregular), medication (refusal), medication (failure to take medication), medication (overdosing), medication (other), money management (self-management), money management (management of others), money management (wasting), cleanliness, environmental improvement, interpersonal relationships, meals, movement, excretion, changing clothes, bathing, and grooming.

Vital signs include, for example, body temperature, pulse rate, systolic blood pressure, and diastolic blood pressure.

Nursing duties include, for example, listening, coordinating with the patient's doctor and related institutions, administering medication and injections, treating skin and pressure ulcers, assisting with medical examinations, instructing and consulting with family members, rehabilitation, and giving advice.

Examples of visit months include visit month flag (January) to visit month flag (December) for the months January to December.

Vectorized text includes data included in nursing records that indicates the patient's condition or nursing content, such as ADL notes, details of symptoms, and nursing content notes and special notes.

An example of an objective variable is the hospitalization flag, which indicates whether or not a patient has been hospitalized.

Figure 5 shows the structure of the inference data file 102E. The inference data file 102E stores data that the trained model uses for inference.

As shown in FIG. 5, the inference data file 102E stores, for each user ID, visit information, ADL information, vital signs, nursing care content, month of visit, and vector text, which is vectorized text.

Examples of home nursing information include age at time of visit, length of stay, number of days since last visit, number of absences (past 14 days), change in symptoms, special instruction visit flag, 5 visits per week after discharge flag, multiple visit flag, and multiple visits flag.

Examples of ADL information include eating (regular), eating (irregular), eating (refusal), eating (overeating), eating (other), sleep (good sleep), sleep (insomnia), sleep (difficulty falling asleep), sleep (early awakening), sleep (awakening during the night), sleep (other), medication (regular), medication (irregular), medication (refusal), medication (failure to take medication), medication (overdosing), medication (other), money management (self-management), money management (management of others), money management (wasting), cleanliness, environmental improvement, interpersonal relationships, meals, movement, excretion, changing clothes, bathing, and grooming.

Vital signs include, for example, body temperature, pulse rate, systolic blood pressure, and diastolic blood pressure.

Nursing duties include, for example, listening, coordinating with the patient's doctor and related institutions, administering medication and injections, treating skin and pressure ulcers, assisting with medical examinations, instructing and consulting with family members, rehabilitation, and giving advice.

Examples of visit months include visit month flag (January) to visit month flag (December) for the months January to December.

Vectorized text includes data included in nursing records that indicates the patient's condition or nursing content, such as ADL notes, details of symptoms, and nursing content notes and special notes.

Transaction file 102C stores data in progress.

The inference result file 102F stores the probability of hospitalization output by the trained model for each date and user ID.

Figure 6 is a flowchart showing the operation of the control unit 101 of the hospitalization probability estimation system 1 in the process of generating training data and learning a model.

As shown in FIG. 6, in step 601, the control unit 101 acquires and converts data. If the raw data is a spreadsheet, it is converted to CSV (Comma Separated Value) format. If the raw data is an electronic medical record, the data structure of the nursing record file 102B is extracted in CSV format from the database. Each piece of data in the generated CSV format is stored in the nursing record file 102B.

In step 602, the control unit 101 assigns a target variable to each nursing record. For example, if the patient was hospitalized within the month of the visit, the hospitalization flag is set to "1," and if the patient was not hospitalized, the hospitalization flag is set to "0."

In step 603 , the control unit 101 creates features. Specifically, the control unit 101 causes a feature adding unit to calculate and add, to the nursing record data in the nursing record file 102B, one or more variables selected from the duration of stay, the number of days since the previous visit, and the number of absences (in the past 14 days).

Next, the control unit 101 deletes nursing records that are missing basic ADL information and changes in symptoms from the nursing record file 102B.

Next, the control unit 101 encodes the categorical variables from the ADL basic information using an encoder.

Next, if the medical care expense category is missing, the control unit 101 copies it from another nursing record with the same user ID to complete it.

Next, the control unit 101 converts the visit month into a visit month flag and adds it.

Next, if the vital signs are abnormal, the control unit 101 deletes the nursing record from the nursing record file 102B.

Next, the control unit 101 complements "-1" if a vital sign is missing.

Next, the control unit 101 supplements and corrects the text data as appropriate.

In step 604 , the control unit 101 selects data suitable for the teacher data from the nursing record file 102B, and stores it in the transaction file 102C. For example, the control unit 101 selects only nursing records whose medical expense category is "mental".

In step 605 , the control unit 101 converts text data indicating the patient's condition or nursing content contained in the nursing record into a vector by word embedding for the data stored in the transaction file 102C, and replaces the original text data with the obtained vector and stores it.

The text data included in the nursing record that indicates the patient's condition or nursing content is selected from one or more of the following: notes on ADL, details of symptoms, notes on nursing content, and special notes on nursing content.

The control unit 101 performs word embedding using the following procedure.

The control unit 101 converts the text into vectors (position embedding) using a tokenizer. The tokenizer can use the best publicly known open source at the time.

The control unit 101 then passes the above vector through an embedding layer of a neural network to convert it into a tensor (word embedding). The embedding layer can use the best publicly available open source at the time.

The control unit 101 generates a tensor for, for example, the first level: a sentence sequence, the second level: a numeric sequence of the position embedding results, and the third level: a numeric sequence of the word embedding results.

The control unit 101 then calculates the third-order average of this tensor, pools it (average pooling), and converts it into a vector to create an embedding vector.

The control unit 101 stores the obtained data in the transaction file 102C.

In step 606, the control unit 101 performs data sampling and divides the data in the above transaction file 102C into teacher data:test data = 8:2 ratio.

In step 607, the control unit 101 extracts features to be used for model training from the generated training data. The control unit 101 extracts at least the visit information, the vectorized text, and the objective variable.

In step 608, the control unit 101 standardizes and scales the extracted features and outputs the resulting data to the teacher data file 102D.

In step 609, the control unit 101 sequentially reads out the teacher data from the teacher data file 102D and trains the model.

Specifically, the model is trained, the loss is calculated, and the model is saved for a predetermined number of epochs.

In step 610, the control unit 101 saves the model of the epoch with the smallest loss as the trained model.

Here, the model can use a deep learning model, but if there is a better model available at that point, that model can also be used.

Figure 7 is a flowchart showing the operation of the inference process by the control unit 101 of the hospitalization probability estimation system 1.

As shown in FIG. 7, in step 701, the control unit 101 acquires and converts the data to be processed. The method of acquiring and converting the data is the same as the method of step 601.

In step 702, the control unit 101 creates features. The processing method is the same as the processing method in step 603.

In step 703, the control unit 101 performs word embedding. The processing method is the same as the processing method in step 605.

In step 704, the control unit 101 selects the feature amount. The processing method is the same as the processing method in step 607.

In step 705, the control unit 101 performs scaling. The processing method is the same as the processing method in step 608.

In step 706, the control unit 101 applies the obtained data to be processed to the trained model to perform inference and output the possibility of hospitalization (hospitalization risk). The inference result, the possibility of hospitalization, can be output as a numerical value such as probability (%), or as a level such as high, medium, or low.

In step 707, the control unit 101 outputs the inference result of the possibility of hospitalization to the input/output unit.

Figure 8 shows the flow of word embedding. As shown in Figure 8, the control unit 101 tokenizes the input text (converts it into the smallest unit of a word string) using a tokenizer. In Figure 8, each of the separated words indicated by "tokenization" is a token.

Next, the control unit 101 performs position embedding on the obtained tokens and vectorizes the token sequence. In FIG. 8, the numerical values indicated by "position embedding" are IDs, which are identifiers uniquely assigned to each token.

Next, the control unit 101 performs word embedding for each vectorized token and converts the vector into a tensor. In FIG. 8, the string of numbers indicated by "word embedding" is a vector uniquely assigned to each token.

Next, the control unit 101 performs average pooling on the third level of the obtained tensor (the numerical sequence resulting from word embedding) to reshape it and obtain an embedding vector. In FIG. 8, the numerical sequence indicated by "average pooling" is the embedding vector after reshaping by average pooling. Here, it is shaped into 128 features.

Figure 9 is a diagram showing an example of the output of the hospitalization risk, which is the inference result. As shown in Figure 9, the control unit 101 outputs the hospitalization risk (%) and the change from the previous month (%) for each patient to the hospitalization risk screen 901.

Figure 10 is a diagram showing an example of output of an increase or decrease in hospitalization risk. As shown in Figure 10, the control unit 101 outputs the increase or decrease in hospitalization risk (%) for each patient to a hospitalization risk increase or decrease screen 1001 along with the hospitalization risk (%).

Figure 11 is a diagram showing an example of output of hospitalization risk statistics. As shown in Figure 11, the control unit 101 tally up the degrees of hospitalization risk and output them to a pie chart 1101.

Figure 12 is a diagram showing an example of output of the trend of hospitalization risk statistics. As shown in Figure 12, the control unit 101 tallies the degree of hospitalization risk and outputs the trend to a bar graph 1201.

Figure 13 is a diagram showing an example of output of the transition of hospitalization risk for each patient. As shown in Figure 13, the control unit 101 outputs the transition of hospitalization risk for each patient to a line graph 1301.

As described above, the hospitalization possibility estimation system 1 of this embodiment includes a nursing record acquisition unit (step 601) that acquires nursing records, which are records of psychiatric home nursing care; an objective variable addition unit (step 602) that adds an objective variable indicating whether or not the patient is hospitalized to the acquired nursing record; a teacher data generation unit (step 605) that generates teacher data by vectorizing text data indicating the patient's condition or nursing content contained in the nursing record; a learning unit (steps 609 and 610) that trains a model using the teacher data to generate a trained model; an inference data generation unit (steps 701 and 703) that vectorizes text data indicating the patient's condition or nursing content contained in the nursing record to be processed, which is separately acquired by the nursing record acquisition unit, to generate inference data; and a hospitalization possibility estimation unit (step 706) that applies the inference data to the trained model to estimate the possibility of hospitalization.

Therefore, this embodiment has the effect of providing a hospitalization probability estimation system that estimates the possibility of hospitalization of a patient from information obtained through home nursing care.

1 Hospitalization probability estimation system 2 Public communication line network 3 Cloud 4 Client terminal 101 Control unit 102 Memory unit 102A User information file 102B Nursing record file 102C Transaction file 102D Teacher data file 102E Inference data file 102F Inference result file 102G Tokenizer 102H Embedding layer 102I Model 103 Input/output unit 104 Communication unit 901 Hospitalization risk screen 1001 Hospitalization risk increase/decrease screen 1101 Pie chart 1201 Column chart 1301 Line graph

Claims (6)

A nursing record acquisition unit that acquires nursing records for each patient, the nursing records being records of psychiatric home visits and created during home visits;
an objective variable adding unit that adds an objective variable indicating whether or not the patient was hospitalized in the visiting month to the acquired nursing record;
A teacher data generation unit that generates teacher data for each patient by vectorizing text data indicating the patient's condition or nursing content included in the nursing record to which the objective variable has been added;
A learning unit that learns a model using the teacher data to generate a learned model;
an inference data generation unit that generates inference data by vectorizing text data indicating the patient's condition or nursing content included in the nursing record to be processed that is separately acquired by the nursing record acquisition unit;
a hospitalization probability estimation unit that estimates the hospitalization probability of the patient to be processed by applying the inference data to the trained model;
A hospitalization probability estimation system that includes: The text data indicating the patient's condition or nursing content included in the nursing record is
The hospitalization probability estimation system according to claim 1 , wherein one or more of the following is selected from notes on ADL, details of symptoms, notes on nursing care, and special notes on nursing care. The hospitalization probability estimation system according to claim 1, further comprising a feature adding unit that adds to the nursing record one or more features selected from the duration of stay, the number of days since the last visit, and the number of absences. The hospitalization probability estimation system according to claim 1, further comprising an encoder for encoding categorical variables included in the nursing records. The categorical variables are:
5. The hospitalization possibility estimation system according to claim 4, wherein one or more items are selected from the group consisting of basic ADL information, nursing care content, and month of visit. A trained model for causing a computer to function to output the possibility of hospitalization of a patient to be processed based on text data indicating nursing care content,
The record is a record of psychiatric home nursing care, and is trained using teacher data for each patient, which is generated by acquiring nursing records created during home nursing care for each patient, adding a target variable indicating whether or not the patient was hospitalized in the month of the visit to the acquired nursing records, and vectorizing text data indicating the patient's condition or nursing content included in the nursing records,
A trained model for causing a computer to function to accept input of inference data generated by vectorizing text data indicating the patient's condition or nursing content contained in the nursing record to be processed, which is separately obtained, and to execute a process of outputting the possibility of hospitalization of the patient to be processed.