A preliminary guiding method and device based on dialogue and electronic medical record

Abstract

The disclosure provides a preliminary guidance method based on a dialogue and an electronic medical record, comprising the following steps: obtaining a patient guidance dialogue and an electronic medical record; preprocessing the patient guidance dialogue and the electronic medical record, and extracting short text data; inputting the short text data into a pre-trained guidance model to obtain a predicted doctor list and a predicted value; the pre-trained guidance model comprises a joint model of a BERT embedding layer, a BERT model and a CRF; obtaining auxiliary data; based on the auxiliary data, correcting the predicted value by using a post-processing method to obtain a corrected predicted value; and outputting the predicted doctor list and the corrected predicted value. The disclosure can perform preliminary guidance prediction on dialogues and electronic medical records. Especially in the case of non-standard input dialogue, the doctor prediction with high accuracy can still be output, and the rapid and accurate matching of patients and doctors can be realized.

Description

Technical Field

[0001] This disclosure relates to the field of medical information technology, and in particular to a preliminary triage method based on dialogue and electronic medical records. Background Technology

[0002] Internet hospitals have broken down the limitations of time and space, expanding medical services through the application of internet and other information technologies, and building an integrated online and offline medical service model covering pre-diagnosis, diagnosis, and post-diagnosis. Internet hospitals have provided timely assistance, playing a vital role in online follow-up consultations, remote consultations, and online prescriptions, thereby increasing people's willingness to accept online medical care.

[0003] Currently, many internet hospital services suffer from a lack of initial triage, or inaccurate triage procedures. Most internet hospital platforms direct patients to departments, not doctors. This is partly due to patients' lack of medical knowledge and inaccurate symptom descriptions, and partly because existing triage programs don't adequately utilize basic patient information, leading to an overemphasis on highly qualified specialists who may not be suitable for the patient's condition. These problems make it difficult for patients to find the most appropriate doctor, thus impacting their overall healthcare experience. Summary of the Invention

[0004] This disclosure provides a preliminary triage method based on dialogue and electronic medical records. It relies solely on patients inputting dialogue and electronic medical records into the triage program. Even when the input dialogue is not standardized, it can output a relatively accurate prediction of the attending physician, thereby achieving rapid and accurate matching between patients and doctors. This enhances the intelligence of the preliminary triage program of the Internet hospital, improves the patient experience, promotes the rational use of medical resources, and ultimately improves the service level of the Internet hospital.

[0005] To achieve the aforementioned objectives, the technical solution provided in this disclosure is as follows:

[0006] Firstly, a preliminary triage method based on dialogue and electronic medical records includes the following steps:

[0007] S1: Obtain patient triage dialogue and electronic medical records;

[0008] S2: Preprocess the patient triage dialogue and electronic medical records to extract short text data;

[0009] S3: Input short text data into the pre-trained triage model to obtain a list of predicted attending doctors and predicted values. The pre-trained triage model includes a joint model of BERT embedding layer, BERT model and CRF.

[0010] S4: Obtain auxiliary data;

[0011] S5: Based on auxiliary data, post-processing methods are used to correct the predicted values ​​and obtain the corrected predicted values;

[0012] S6: Output the predicted list of doctors to be treated and the corrected predicted values.

[0013] Preferably, the patient triage dialogue and electronic medical record include: the patient's symptom description information in the triage dialogue and the electronic medical record.

[0014] Preferably, the preprocessing of patient triage dialogues and electronic medical records in S2, and the extraction of short text data, includes the following steps:

[0015] S21: Use word segmentation tools in conjunction with a medical dictionary to segment symptom description information in the triage dialogue;

[0016] S22: Remove stop words, noise words, and punctuation marks from the word segmentation results;

[0017] S23: Use custom templates to extract gender, age group, and past medical history information from electronic medical records;

[0018] S24: Combine the medical dictionary to merge synonyms in the word segmentation results to obtain simplified word segmentation results;

[0019] S25: Combine the simplified word segmentation results, gender, age group, and past medical history information into short text data.

[0020] Preferably, the word segmentation tool supports custom dictionaries, including jieba, hanlp, ICTCLAS, or LTP.

[0021] Preferably, before inputting the short text data into the pre-trained triage model in step S3 to obtain the predicted list of attending physicians and the predicted values, the method further includes:

[0022] S0. Train the triage model to be trained;

[0023] The training of the triage model to be trained in S0 includes:

[0024] S01: Obtain a sample dataset of historical patient triage dialogues, relevant patient medical records, and electronic medical records;

[0025] S02: Preprocess the sample dataset to extract the sample short text dataset;

[0026] S03: Based on relevant patient medical records, obtain patient visit doctor information, use the patient visit doctor information as a label, label each short text data in the sample short text dataset, and construct a training dataset;

[0027] S04: Train the triage model to be trained based on the training dataset;

[0028] S05: Obtain the pre-trained triage model.

[0029] Preferably, the relevant patient medical records include: the patient's actual medical records within a certain period of time after completing the triage dialogue.

[0030] Preferably, step S3, which involves inputting short text data into a pre-trained triage model to obtain a predicted list of attending physicians and predicted values, includes:

[0031] Short text data is input into the BERT embedding layer to obtain a vector representation of the text data. This vector representation is then input into the BERT model. The feature sequence obtained from the BERT model is used as input to the CRF model to annotate and decode the short text, obtaining a label sequence. The output is the label sequence probability value, and the prediction equation for the label sequence probability value is... as follows:

[0032] ,………..(1)

[0033] Where X is the feature sequence and y is the label sequence. It is the transition score between feature sequences of short text data. It is the ratio of the total score of correctly labeled feature sequences in short text data to the total score of all possible labeled feature sequences, used to obtain the predicted value of the label sequence. The formula is as follows:

[0034] ,………..(2)

[0035] in, This represents all possible labels.

[0036] Preferably, the acquisition of auxiliary data includes: acquiring the consultation ratio of each department and a list of all attending physicians within a selected time period.

[0037] Preferably, the step of correcting the predicted value based on auxiliary data using a post-processing method to obtain the corrected predicted value includes:

[0038] The corrected prediction value is obtained based on formula (3):

[0039] ,………..(3)

[0040] in, It is to correct the predicted value. It is the percentage of patients seen in each department within a selected time period, divided by the total number of doctors in that department. It is the predicted value of the label sequence. This refers to the percentage of labels appearing, and the total number of doctors in that department is the total number of doctors belonging to that department on the list of all attending physicians.

[0041] Secondly, a preliminary triage device based on dialogue and electronic medical records is provided. This device includes:

[0042] Acquisition Unit: Used to acquire patient triage dialogues and electronic medical records;

[0043] Preprocessing unit: used to preprocess patient triage dialogues and electronic medical records, and extract short text data;

[0044] Prediction unit: used to input short text data into a pre-trained triage model to obtain a list of predicted attending doctors and predicted values. The pre-trained triage model includes a joint model of BERT embedding layer, BERT model and CRF.

[0045] Auxiliary data unit: used to acquire auxiliary data;

[0046] Correction unit: Used to correct the predicted value based on auxiliary data and using post-processing methods to obtain the corrected predicted value;

[0047] Output unit: Used to output the predicted list of attending physicians and corrected prediction values.

[0048] The above technical solution has at least the following advantages compared with the existing technology:

[0049] The above-described solution utilizes various data sources, including patient triage dialogues, electronic medical records, and data on hospital doctor visits. By using this data, the impact of inaccurate patient symptom descriptions can be minimized, and the seasonality of diseases can be taken into account to improve clinic prediction, ultimately achieving rapid and accurate matching of patients and doctors. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of a preliminary triage method based on dialogue and electronic medical records disclosed herein;

[0052] Figure 2 This is a flowchart illustrating the training method of the public triage model;

[0053] Figure 3This is a structural diagram of the triage model disclosed in this publication;

[0054] Figure 4 This disclosure presents a block diagram of a preliminary triage device based on dialogue and electronic medical records. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0056] Unless otherwise defined, the technical or scientific terms used in this disclosure shall have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms “first,” “second,” and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an,” “a,” or “the,” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising,” “including,” or “including,” and similar terms mean that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. The terms “connected,” “linked,” or “connected,” and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0057] It should be noted that the terms "up", "down", "left", "right", "front", and "back" used in this disclosure are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0058] This disclosure addresses the common pain points in current internet hospital services, such as the lack of a preliminary triage process or inaccurate preliminary triage procedures. It provides a preliminary triage method based on dialogue and electronic medical records. This method can address issues such as patients' lack of medical knowledge, inaccurate symptom descriptions, and insufficient utilization of basic patient information by existing triage procedures.

[0059] like Figure 1 As shown, the initial triage method 100 based on dialogue and electronic medical records includes six steps.

[0060] S1: Obtain patient triage dialogue and electronic medical records;

[0061] In some embodiments, patient dialogue is exported from the online consultation triage program. The triage program first guides first-time patients through a patient registration process. During registration, patients need to fill out an electronic medical record, providing basic personal information such as gender, age, and past medical history. The electronic medical record is created through input methods such as selection, avoiding user-input errors. The triage program can be configured to allow patients to input their own dialogue or guide them to select certain tags. During patient input, prompts can be provided to ensure the input is as complete and standardized as possible. For example, after a patient inputs certain symptoms, the dialogue program prompts the patient to provide the duration and severity of the symptoms. The character limit for the personal input section in the triage program can be set.

[0062] S2: Preprocess the patient triage dialogue and electronic medical records to extract short text data;

[0063] In some embodiments, the patient referral dialogue is first extracted. A word segmentation tool combined with a medical dictionary is used to segment the dialogue. The symptom description information in the dialogue is then segmented using the same tool. Stop words, noise words, and punctuation are removed from the segmented results. Gender, age group, and past medical history information are extracted from the electronic medical record using a custom template. Synonyms are merged in the segmented results using a medical dictionary to obtain simplified segmentation results. The simplified segmentation results, gender, age group, and past medical history information are then merged into short text data.

[0064] In some embodiments, a medical-specific dictionary is first established. This can be expanded using a basic Chinese medical-specific dictionary. The following are some commonly used medical-specific dictionaries: CMeSH (Chinese Medical Science Subject Headings), corresponding to the English version of MeSH, is the standard subject heading in the Chinese medical field. CDT (Chinese Disease and Diagnosis Codes), is a disease and diagnosis classification standard published by the Ministry of Health of China, containing a large number of Chinese medical terms and vocabulary. CUMT (Chinese Medical Terminology), is a medical terminology published by the National Medical Products Administration of China, containing commonly used words and terms in the Chinese medical field.

[0065] In some embodiments, Chinese word segmentation is an important step in Chinese natural language processing. The Chinese word segmenter can be jieba, hanlp, ICTCLAS, or LTP. The selected segmenter needs to support custom dictionaries. Segmentation performance benchmarks include: precision, recall, and the F-1 score.

[0066] All of the above word segmenters are open-source tools, available for free use and modification, and have high word segmentation accuracy and speed.

[0067] In some embodiments, key information from electronic medical records can be extracted by defining matching templates, such as medical history, physical examination reports, and diagnostic records. Template matching can be implemented using rule-based methods. Extracting key information typically involves directly converting it into phrases or words. For example, age can be extracted and converted into age ranges. Past medical history can be extracted and directly converted into disease names or treatments such as hospitalization and surgery.

[0068] In some embodiments, the method of merging triage dialogues and information on gender, age group, and past medical history into short text data can be either direct overlay merging or deduplication merging. Patients may mention their past medical history during the dialogue, which may result in records that do not match the electronic medical record. The deduplication merging process can increase the weight of past medical history because the recorders of past medical history are professionals.

[0069] S3: Input the short text data into the pre-trained triage model to obtain the predicted list of attending doctors and the predicted values; the pre-trained triage model includes a joint model of BERT embedding layer, BERT model and CRF;

[0070] In some embodiments, the triage model is built upon the BERT model. BERT (Bidirectional Encoder Representations from Transformers) is a bidirectional pre-trained language model based on Transformers, and is one of the most advanced pre-trained models in the field of natural language processing. The purpose of the BERT embedding layer is to convert the input text sequence into a vector representation for subsequent model training and inference. Figure 2As shown, the triage model comprises an input layer, a BERT model, and a CRF model. The BERT model and CRF are a joint model. The BERT model includes an embedding layer and multiple Transformer Encoder layers. The BERT embedding layer consists of three parts: Token embedding, Segment embedding, and Position embedding. Token embedding segments the input text into words or subwords and maps each word or subword to a vector. Segment embedding divides the input text into multiple segments and assigns a vector to each segment. Position embedding assigns a vector to the position of each input element. These three embedding vectors are concatenated to form a one-dimensional vector sequence as the model's input. The BERT model is structured as a stack of two Transformer Encoder layers, each consisting of a multi-head self-attention mechanism and a feedforward neural network. The outputs of these layers pass through a linear transformation layer and are then processed using a Softmax function for classification or regression. The final output of the Transformer Encoder layer is a vector. The multi-head self-attention mechanism in BERT can improve the input of the convolutional layer. Each word in the description of patient symptoms is represented as a vector, and a sentence is represented as an embedding matrix. Each row of the embedding matrix is ​​a word vector. Using BERT is an important step that can better map natural language into a word vector matrix and preserve contextual information.

[0071] like Figure 2 As shown, the joint approach of the BERT and CRF model is to input these output vectors into the CRF layer for label prediction. The CRF layer relies on the label of the previous word to predict the label of the current word, outputting label sequence probability values. During training, the model performs loss calculations and gradient backpropagation based on the differences between entity labels and annotations. The prediction equation for label sequence probability values ​​is... as follows:

[0072] ,………..(1)

[0073] Where X is the feature sequence and y is the label sequence. It is the transition score between feature sequences of short text data. It is the ratio of the total score of correctly labeled feature sequences in short text data to the total score of all possible labeled feature sequences, used to obtain the predicted value of the label sequence. The formula is as follows:

[0074] ,………..(2)

[0075] in, This represents all possible labels during the training of the triage model. A value of 1 indicates that the predicted entity label and the annotation are consistent.

[0076] In some embodiments, the output of the triage model is a list of patients and predicted values. In general, the output of the triage model is a list of doctors.

[0077] In some embodiments, the purpose of a triage model is to provide pre-screening and triage. Pre-screening and triage involves conducting an initial medical assessment of patients before their outpatient visit to a hospital. By evaluating and analyzing the patient's condition, more accurate medical advice and treatment plans are provided. In offline hospitals, pre-screening and triage are typically handled by nurses or doctors. With the development of artificial intelligence, patients can complete pre-screening and triage before their outpatient visit via telephone or online consultation. Pre-screening and triage in internet hospitals are usually completed by a triage model. The triage model can output doctors, depending primarily on the granularity of the data annotation during model training.

[0078] S4: Obtain auxiliary data;

[0079] In some embodiments, triage dialogues, especially online ones, may lack complete information. Because diseases have seasonal and regional characteristics, supplementary data, such as department popularity, is often an important basis for triage. Department popularity refers to the number and frequency of departments patients choose to visit within a certain period. Analyzing department popularity can provide a better understanding of patients' medical needs. Collecting and analyzing department popularity data is crucial. Hospitals can collect various department popularity data, such as the number of patients, frequency of visits, and visit time, through outpatient registration systems, medical records, etc. This data can be obtained through electronic medical record systems and analyzed and processed using data analysis software. It is also necessary to collect a complete list of attending physicians.

[0080] In some embodiments, obtaining the number of consultations for each department within a selected time period can reflect the department's popularity. Department popularity statistics are generally performed using different time granularities such as annual, monthly, weekly, and daily.

[0081] S5: Based on auxiliary data, post-processing methods are used to correct the predicted values ​​and obtain the corrected predicted values;

[0082] In some embodiments, department popularity can be used to supplement a patient's symptom description. Patients may lack awareness of some symptoms or easily omit them in their descriptions. Meanwhile, a particular department may experience a surge in visits during the same period. Combining incomplete symptom descriptions with department visit volume can help infer the appropriate department.

[0083] In some embodiments, the Bayesian equation is a mathematical formula used to infer the probability of an event occurring. Its basic principle is to combine prior knowledge with new evidence to derive the posterior probability.

[0084] Specifically, the Bayesian equation describes how, given certain prior knowledge, we can update the estimate of the probability of an event based on new evidence. The equation is shown below:

[0085] P(A|B) = P(B|A) P(A) / P(B)

[0086] Where P(A|B) represents the probability of event A occurring given that B has occurred; P(B|A) represents the probability of event B occurring given that event A has occurred; P(A) is the prior probability, representing the probability of event A occurring without new evidence; and P(B) is a normalization constant used to ensure that the sum of the posterior probabilities is 1.

[0087] In some embodiments, the predicted values ​​are corrected based on the Bayesian equation. Obtaining the corrected predicted values ​​includes:

[0088] ,………..(3)

[0089] in, It is to correct the predicted value. It is the percentage of patients seen in each department within a selected time period, divided by the total number of doctors in that department. It is the predicted value of the label sequence. This refers to the percentage of labels appearing, and the total number of doctors in that department is the total number of doctors belonging to that department on the list of all attending physicians.

[0090] S6: Output the predicted list of attending physicians and the corrected prediction values;

[0091] In some embodiments, a predicted matching doctor can be recommended to the patient based on the revised predicted value. Only one doctor with the highest predicted value may be provided. If the revised predicted values ​​are relatively close, multiple doctors may be provided.

[0092] It should be noted that the application scenarios for initial triage based on dialogue and electronic medical records include internet hospitals, online medical self-service platforms, offline hospital intelligent triage systems, or family doctor systems. Dialogue and electronic medical records can improve patient consultation efficiency and avoid wasting medical resources. For patients, the usage threshold is low. At the same time, for doctors, initial triage allows them to gain a preliminary understanding of the patient's condition.

[0093] Figure 3 The flowchart 300 shows the training method of the triage model.

[0094] like Figure 3 As shown, the training method for the triage model includes the following steps:

[0095] S01: Obtain a sample dataset of historical patient triage dialogues, relevant patient medical records, and electronic medical records;

[0096] In some embodiments, a large amount of patient guidance dialogues and electronic medical records are required as training data. Data sources generally fall into two categories: self-collected data and publicly available data sources. Self-collected data includes capturing offline guidance dialogues using recording devices and converting them into data. Alternatively, dialogue information can be extracted from medical records.

[0097] Currently known publicly available Chinese data sources include: MedDialog, a medical question-and-answer dataset released by the Natural Language Processing and Social Computing Laboratory at Tsinghua University; and CMedQA, a medical question-and-answer dataset released by the National University of Defense Technology. A drawback of publicly available data sources is the lack of electronic medical records, which can be supplemented by randomly generated electronic medical records.

[0098] S02: Preprocess the sample dataset to extract the sample short text dataset;

[0099] The specific processing involves first using a word segmentation tool in conjunction with a medical dictionary to segment the triage dialogue; then removing stop words, noise words, and punctuation from the segmented results; if electronic medical records are available, a custom template is used to extract gender, age group, and past medical history information from the electronic medical records. Next, synonyms are merged from the segmented results using a medical dictionary; finally, the segmented results, gender, age group, and past medical history information are merged into short text data.

[0100] S03: Based on relevant patient medical records, obtain patient visit doctor information, use the patient visit doctor information as a label, and annotate each short text data in the sample short text dataset to construct a training dataset.

[0101] Regarding obtaining patient-doctor information, one approach is to use actual doctors as tags. Alternatively, individuals with medical backgrounds can annotate the conversations based on their medical knowledge. Furthermore, offline triage records can be used, leveraging the suggestions of triage staff as tags.

[0102] It's important to note that using doctors with actual visits as tags is the most accurate. However, linking actual doctors with conversation information often presents challenges. For example, patients may provide incomplete information during the triage process, and the hospital may not have adequately processed the data. Alternatively, patients may not be specifically seeking information about themselves. Therefore, if using doctors with actual visits as tags, manual data review is necessary to ensure a high degree of relevance.

[0103] In some embodiments, model training involves a large corpus. The BERT model is a pre-trained language model that can be pre-trained on a large-scale corpus to learn general representations of language, thereby improving the model's understanding of language. Compared to traditional sequence labeling models, the BERT model can utilize more contextual information, thus better capturing the semantic and syntactic information of language and lowering the barrier to entry for users. The CRF model is a conditional random field model that can be used to label or classify sequence data. The CRF model can model the state transition process of the entire sequence, thus utilizing more contextual information. By combining the advantages of BERT and CRF, the joint model can achieve state-of-the-art performance in various sequence labeling tasks.

[0104] S04: Train the triage model to be trained based on the training dataset;

[0105] The triage model includes a BERT embedding layer, a BERT model, and a CRF model; such as Figure 3 As shown, the purpose of the triage model is to assign labels to the input dialogue. Training a joint model of BERT and CRF models, word segmentation, and labeling are necessary steps. The training process should be optimized using an appropriate optimizer, such as Adam or SGD. The model's hyperparameters, such as the learning rate and batch size, should be carefully tuned to obtain optimal performance. The model's performance should be evaluated on a reserved validation set to monitor its performance and prevent overfitting.

[0106] S05: Obtain the pre-trained triage model; the pre-trained triage model can be used for triage prediction. When new training data is obtained, or when doctor information is updated, the model needs to be retrained.

[0107] The above is an introduction to the method embodiments. The following describes the solution described in this disclosure further through device embodiments.

[0108] Figure 4 A block diagram 400 of a preliminary triage device based on dialogue and electronic medical records is shown. Figure 4As shown, the device 400 includes: an acquisition unit 410 for acquiring patient triage dialogues and electronic medical records; a preprocessing unit 420 for preprocessing the patient triage dialogues and electronic medical records to extract short text data; a prediction unit 430 for inputting the short text data into a pre-trained triage model to obtain a predicted list of attending physicians and predicted values, wherein the pre-trained triage model includes a joint model of BERT embedding layer, BERT model and CRF; an auxiliary data unit 440 for acquiring auxiliary data; a correction unit 450 for correcting the predicted values ​​based on the auxiliary data using post-processing methods to obtain corrected predicted values; and an output unit 460 for outputting the predicted list of attending physicians and the corrected predicted values.

[0109] In summary, this disclosure utilizes a variety of data, including patient referral dialogues and electronic medical records, as well as data on hospital doctor visits. By using this data, the impact of inaccurate patient symptom descriptions can be minimized, seasonal characteristics of diseases can be taken into account, clinic predictions can be improved, and ultimately, rapid and accurate matching of patients and doctors can be achieved.

[0110] The following points need to be explained:

[0111] (1) The accompanying drawings of the embodiments of this disclosure only involve the structures involved in the embodiments of this disclosure. Other structures can be referred to the general design.

[0112] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the present disclosure, i.e., these drawings are not drawn to scale. It will be understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.

[0113] (3) Where there is no conflict, the embodiments of this disclosure and the features in the embodiments can be combined with each other to obtain new embodiments.

[0114] The above are merely specific embodiments of this disclosure, but the scope of protection of this disclosure is not limited thereto. The scope of protection of this disclosure shall be determined by the scope of the claims.

Claims

1. A preliminary triage method based on dialogue and electronic medical records, characterized in that, Includes the following steps: S1: Obtain patient triage dialogue and electronic medical records; S2: Preprocess the patient triage dialogue and electronic medical records to extract short text data; S3: Input the short text data into the pre-trained triage model to obtain the predicted list of attending physicians and the predicted values; the pre-trained triage model includes a joint model of BERT embedding layer, BERT model and CRF, and the step of inputting the short text data into the pre-trained triage model to obtain the predicted list of attending physicians and the predicted values ​​includes: Short text data is input into the BERT embedding layer to obtain a vector representation of the text data. This vector representation is then input into the BERT model. The feature sequence obtained from the BERT model is used as input to the CRF model to annotate and decode the short text, obtaining a label sequence. The output is the label sequence probability value, and the prediction equation for the label sequence probability value is... as follows: ，………..(1) Where X is the feature sequence and y is the label sequence. It is the transition score between feature sequences of short text data. It is the ratio of the total score of correctly labeled feature sequences in short text data to the total score of all possible labeled feature sequences, used to obtain the predicted value of the label sequence. The formula is as follows: ，………..(2) in, Represents all possible labels; S4: Obtain auxiliary data; S5: Based on auxiliary data, the predicted value is corrected using post-processing methods to obtain a corrected predicted value. The process of correcting the predicted value based on auxiliary data using post-processing methods to obtain the corrected predicted value includes: The corrected prediction value is obtained based on formula (3): ，………..(3) Where P(C) is the corrected predicted value, and P(A) is the proportion of patients seeking medical advice in each department divided by the total number of doctors in that department within the selected time period. It is the predicted value of the label sequence. This refers to the proportion of labels appearing, and the total number of doctors in that department is the total number of doctors belonging to that department on the list of all attending physicians. S6: Output the predicted list of doctors to be treated and the corrected predicted values.

2. The preliminary triage method based on dialogue and electronic medical records according to claim 1, characterized in that, The patient triage dialogue and electronic medical record include: the patient's symptom description information in the triage dialogue and the electronic medical record.

3. The preliminary triage method based on dialogue and electronic medical records according to claim 1 or 2, characterized in that, The preprocessing of patient triage dialogues and electronic medical records in S2, and the extraction of short text data, includes the following steps: S21: Use word segmentation tools in conjunction with a medical dictionary to segment symptom description information in the triage dialogue; S22: Remove stop words, noise words, and punctuation marks from the word segmentation results; S23: Use custom templates to extract gender, age group, and past medical history information from electronic medical records; S24: Combine the medical dictionary to merge synonyms in the word segmentation results to obtain simplified word segmentation results; S25: Combine the simplified word segmentation results, gender, age group, and past medical history information into short text data.

4. The preliminary triage method based on dialogue and electronic medical records according to claim 3, characterized in that, The word segmentation tool supports custom dictionaries, including jieba, hanlp, ICTCLAS, or LTP.

5. The preliminary triage method based on dialogue and electronic medical records according to claim 1, characterized in that, Before inputting the short text data into the pre-trained triage model in step S3 to obtain the predicted list of attending physicians and the predicted values, the method further includes: S0. Train the triage model to be trained; The training of the triage model to be trained in S0 includes: S01: Obtain a sample dataset of historical patient triage dialogues, relevant patient medical records, and electronic medical records; S02: Preprocess the sample dataset to extract the sample short text dataset; S03: Based on relevant patient medical records, obtain patient visit doctor information, use the patient visit doctor information as a label, label each short text data in the sample short text dataset, and construct a training dataset; S04: Train the triage model to be trained based on the training dataset; S05: Obtain the pre-trained triage model.

6. The preliminary triage method based on dialogue and electronic medical records according to claim 5, characterized in that, The relevant patient medical records include: the actual medical records of the patient within a certain period of time after completing the triage dialogue.

7. The preliminary triage method based on dialogue and electronic medical records according to claim 1, characterized in that, The auxiliary data to be acquired includes: the consultation rate of each department and a list of all attending physicians within a selected time period.

8. A preliminary triage device based on dialogue and electronic medical records, characterized in that, include: Acquisition Unit: Used to acquire patient triage dialogues and electronic medical records; Preprocessing unit: used to preprocess patient triage dialogues and electronic medical records, and extract short text data; Prediction Unit: Used to input short text data into a pre-trained triage model to obtain a list of predicted attending doctors and predicted values. The pre-trained triage model includes a joint model of BERT embedding layers, BERT model, and CRF, specifically including: Short text data is input into the BERT embedding layer to obtain a vector representation of the text data. This vector representation is then input into the BERT model. The feature sequence obtained from the BERT model is used as input to the CRF model to annotate and decode the short text, obtaining a label sequence. The output is the label sequence probability value, and the prediction equation for the label sequence probability value is... as follows: ，………..(1) Where X is the feature sequence and y is the label sequence. It is the transition score between feature sequences of short text data. It is the ratio of the total score of correctly labeled feature sequences in short text data to the total score of all possible labeled feature sequences, used to obtain the predicted value of the label sequence. The formula is as follows: ，………..(2) in, Represents all possible labels; Auxiliary data unit: used to acquire auxiliary data; Correction Unit: Used to correct the predicted values ​​based on auxiliary data using post-processing methods, obtaining corrected predicted values, specifically including: The corrected prediction value is obtained based on formula (3): ，………..(3) Where P(C) is the corrected predicted value, and P(A) is the proportion of patients seeking medical advice in each department divided by the total number of doctors in that department within the selected time period. It is the predicted value of the label sequence. This refers to the proportion of labels appearing, and the total number of doctors in that department is the total number of doctors belonging to that department on the list of all attending physicians. Output unit: Used to output the predicted list of attending physicians and corrected prediction values.

Images