Triage data processing method and device, computer equipment and storage medium

Abstract

The invention relates to a triage data processing method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring doctor-seeing information of a patient and department introduction information of each department of a current doctor-seeing hospital; obtaining a first triage result according to the treatment information, the introduction information of each department and a trained triage language model; obtaining a second triage result according to the treatment information, the introduction information of each department and a trained multi-level text matching model; wherein the multi-level text matching model is a model used for representing an association relationship between the treatment information of the patient and a triage result conforming to a preset multi-level department mapping relationship; and obtaining a final triage result according to the first triage result and the second triage result. By adopting the method, the flexibility and accuracy of medical triage can be improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a triage data processing method, apparatus, computer equipment, and storage medium. Background Technology

[0002] With the development of data processing technology, automated triage systems have emerged. These systems can effectively free up manpower and alleviate the pressure on hospital outpatient services to a certain extent.

[0003] Traditional automated triage systems employ a point-and-click model, where patients select the appropriate department based on a list of prompts within the system, according to their symptoms. However, traditional automated triage methods or systems are only applicable to a specific hospital and often can only identify the department corresponding to a single symptom. In the healthcare field, hospitals and clinics at different levels vary significantly in terms of the departments they offer and their medical resources, resulting in varying granularity and complexity of automated triage systems. This can confuse patients who lack basic medical knowledge, making it difficult for them to accurately determine where to seek medical attention. Therefore, the problems of poor triage flexibility and inaccuracy persist. Summary of the Invention

[0004] Therefore, it is necessary to provide a triage data processing method, device, computer equipment, and storage medium that can improve the flexibility and accuracy of medical triage in response to the above-mentioned technical problems.

[0005] Firstly, a triage data processing method is provided, the method comprising: Obtain the patient's medical information and departmental information of the hospital where they are currently receiving treatment; Based on the patient information, departmental introductions, and the trained triage language model, the first triage result is obtained; Based on the patient's medical history, departmental information, and a trained multi-level text matching model, a second triage result is obtained. The multi-level text matching model represents the association between the patient's medical history and the triage result, which conforms to a pre-defined multi-level departmental mapping. The final triage result is obtained based on the results of the first and second triages.

[0006] In some embodiments, obtaining a patient's medical records includes: Obtain consultation information input by the user and patient information; Extract symptom keywords from consultation information to generate prompts for further questions about the condition; Receive interactive information input by the user based on prompts for follow-up questions about the patient's condition, and extract the chief complaint information from the interactive information; and The chief complaint and patient information are used as medical information.

[0007] In some embodiments, a second triage result is obtained based on patient visit information, departmental information, and a trained multi-level text matching model, including: Convert medical information into medical text vectors; Based on the department introduction information and the preset multi-level department mapping relationship, the department introduction information is mapped to the multi-level department mapping relationship, and multi-level department text vectors are generated; and The second triage result is obtained based on the patient's text vector, the multi-level department text vector, and the trained multi-level text matching model.

[0008] In some embodiments, the training method for a multi-level text matching model includes: Obtain historical patient visit data from medical institutions at all levels; Extract sample medical visit information and corresponding positive and negative sample department information from historical medical visit data; Based on preset department levels and multi-level department mapping relationships, positive sample department description information is mapped to the corresponding department levels, generating positive sample multi-level department description information; and The multi-level text matching model is trained based on sample medical visit information, positive sample multi-level department introduction information, and negative sample department introduction information.

[0009] In some embodiments, a multi-level text matching model is trained based on sample patient visit information, positive sample multi-level department description information, and negative sample department description information, including: The loss function under a single embedding dimension is calculated based on the sample visit information, the positive sample multi-level department introduction information, and the negative sample department introduction information. Calculate the similarity of positive and negative samples under a single embedding dimension; The positive and negative sample similarity scores are concatenated to obtain the loss function for each sample; and The total loss function is obtained based on the loss function of each sample, and the multi-level text matching model is trained using the total loss function.

[0010] In some embodiments, the training method for the triage language model includes: Extract sample medical visit information and corresponding positive sample department information from historical medical visit data, and generate sample data pairs; Supervised fine-tuning training of the triage language model was performed based on sample data. The reward mechanism is used to provide feedback and correct the training of the triage language model after supervised fine-tuning; and Candidate output data are selected from the output results corresponding to the current training batch. The training strategy is updated based on the candidate output data. The updated training strategy is then used to optimize the triage language model after feedback correction training.

[0011] In some embodiments, obtaining a final triage result based on a first triage result and a second triage result includes: The combined score of the first and second triage results is calculated based on the inverse sorting fusion algorithm, and the final triage result is determined based on the combined score.

[0012] Secondly, a triage data processing device is provided, the device comprising: The information acquisition module is used to acquire the patient's medical information and the department introduction information of each department of the current hospital. The first prediction module is used to obtain the first triage result based on the patient information, the introduction information of each department, and the trained triage language model; The second prediction module is used to obtain a second triage result based on the patient's medical information, departmental information, and a trained multi-level text matching model. The multi-level text matching model is used to represent the association between the patient's medical information and the triage result that conforms to a preset multi-level departmental mapping relationship. The results fusion module is used to obtain the final triage result based on the first triage result and the second triage result.

[0013] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect.

[0014] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the steps of the method described in the first aspect.

[0015] The aforementioned triage data processing method, device, computer equipment, and storage medium integrate the first triage result based on the triage language model and the second triage result based on the multi-level text matching model and conforming to the multi-level department mapping relationship. The resulting final triage result effectively takes into account both the breadth of medical semantic understanding and the structured characteristics of hospital department systems at different levels. Therefore, it can flexibly and dynamically adapt to the granularity of department settings and differences in medical resources of different hospitals while improving the accuracy and reliability of triage results. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the triage data processing method in some embodiments; Figure 2 A schematic diagram illustrating the technical concept of triage data processing methods in some application examples; Figure 3 This is a structural block diagram of the triage data processing device in some embodiments; Figure 4 This is a diagram showing the internal structure of a computer device in some embodiments. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] The triage data processing method provided in this application can be applied to computer equipment, such as terminal devices. These terminal devices can be, but are not limited to, various personal computers, laptops, smartphones, tablets, hospital consultation terminals, and portable wearable devices.

[0019] In some embodiments, such as Figure 1 As shown, a triage data processing method is provided, which can be applied to... Figure 1 Taking the terminal in the example, the explanation includes the following steps: Step S102: Obtain the patient's medical information and the department introduction information of each department of the current hospital.

[0020] Here, "patient" refers to an individual awaiting treatment at an outpatient or emergency room who requires services such as triage. "Currently seeking treatment" refers to the hospital specified by the user (which can be the patient or someone else).

[0021] For example, the specific ways to obtain a patient's medical information may include, but are not limited to: retrieving information such as the patient's historical medical records, chief complaint, and present medical history from the hospital information system as medical information; obtaining key information such as the patient's online description of their symptoms and medical history as medical information; or extracting medical information from the collected interaction information after real-time human-computer interaction with the patient through the triage language model involved in this application.

[0022] For example, obtaining the department introduction information of each department of the current hospital may include, but is not limited to: accessing the database or knowledge base system of the current hospital and extracting structured data such as the scope of diagnosis and treatment, expert specialties, equipment configuration, technical advantages and routinely treated diseases of each department as the department introduction information of each department.

[0023] Step S104: Based on the patient information, departmental introductions, and the trained triage language model, obtain the first triage result.

[0024] The trained triage language model here is a deep learning model used to represent the complex relationships between patient complaints, symptoms, medical history, and other consultation information and departmental information such as the scope of treatment and specialties. The triage language model can include, but is not limited to: LLM (Large Language Model), reasoning models based on medical knowledge graphs, BERT (Bidirectional Encoder Representations from Transformers) variants specifically trained for clinical text, or other specialized classification and matching models built using natural language processing (NLP) and machine learning techniques.

[0025] Step S106: Based on the patient information, departmental introductions, and the trained multi-level text matching model, obtain the second triage result.

[0026] The trained multi-level text matching model is used to represent the association between a patient's medical information and triage results that conform to a preset multi-level departmental mapping relationship. This multi-level departmental mapping relationship can be established based on national or industry standards and can follow a progressive departmental classification system from broad to specific levels.

[0027] For example, if the current hospital's department information is gastroenterology, it can be mapped to the following multi-level department mapping relationship: Multi-level department mapping relationship = {Level 1 department → Level 2 department → Level 3 department}; After mapping: Multi-level department mapping relationship = {Internal Medicine → Gastroenterology → Empty}; In other words, the information of the department visited (gastroenterology) is mapped to the first-level department as internal medicine, the second-level department as gastroenterology, and the third-level department as empty (i.e. there is no third level), thus forming a structured hierarchical data relationship.

[0028] Among them, multi-level text matching models may include, but are not limited to, vectorized representation models such as BGE (BAAI General Embedding) model and Qwen-GTE (Qwen-General Text Embeddings) model.

[0029] Step S108: Obtain the final triage result based on the first and second triage results. In this step, the first triage result obtained from the triage language model is fused with the second triage result obtained from the multi-level text matching model, which conforms to the preset multi-level department mapping relationship, to obtain a more accurate final triage result that is flexibly adapted to the current hospital.

[0030] The above-mentioned triage data processing method integrates the first triage result based on the triage language model and the second triage result based on the multi-level text matching model and conforming to the multi-level department mapping relationship. The resulting final triage result effectively takes into account both the breadth of medical semantic understanding and the structured characteristics of hospital department systems at different levels. Therefore, it can flexibly and dynamically adapt to the granularity of department settings and differences in medical resources of different hospitals while improving the accuracy and reliability of triage results.

[0031] In some embodiments, obtaining a patient's medical information includes: obtaining consultation information input by a user and the patient's patient information; extracting symptom keywords from the consultation information to generate follow-up question prompts; receiving interactive information input by the user based on the follow-up question prompts, and extracting chief complaint information from the interactive information; and using the chief complaint information and the patient information as medical information.

[0032] In real-world applications, patients may omit key symptoms when describing their chief complaint (e.g., only describing "headache" but omitting "blurred vision"), use vague non-medical terms (e.g., "stomach discomfort"), or have other users register for the patient without accurately knowing the patient's key symptoms. In this embodiment, based on the consultation information input by the user (which may be the patient or someone else) and combined with the patient's information, a dynamic follow-up questioning mechanism based on symptom keywords is introduced. This mechanism dynamically generates follow-up question prompts and guides the user to interact, thereby extracting the patient's chief complaint information from targeted interactive questions and answers. This results in more complete and reliable clinically usable medical information, improving the accuracy and reliability of the output of subsequent triage language models and multi-level text matching models.

[0033] In some embodiments, obtaining a second triage result based on medical visit information, departmental introduction information, and a trained multi-level text matching model includes: converting medical visit information into medical visit text vectors; mapping departmental introduction information to a multi-level departmental mapping relationship based on departmental introduction information and a preset multi-level departmental mapping relationship, and generating multi-level departmental text vectors; and obtaining a second triage result based on the medical visit text vectors, multi-level departmental text vectors, and the trained multi-level text matching model.

[0034] In this embodiment, text extraction can be performed based on a trained multi-level text matching model to generate a medical visit text vector corresponding to the medical visit information and a department introduction text vector corresponding to the department introduction information. Furthermore, by combining a preset multi-level department mapping relationship, the information in the department introduction text vector is mapped to the multi-level department mapping relationship to generate a multi-level department text vector that conforms to a general preset multi-level department mapping relationship. The medical visit text vector and the multi-level department text vector are then matched and calculated to obtain a second triage result that conforms to a general preset multi-level department mapping relationship.

[0035] In this embodiment, by converting unstructured patient information and department introduction information into text vector representations, and further structuring the department introduction text vectors into multi-level department text vectors, it is possible to quantitatively compare the patient's patient information with information from different levels of departments. This ensures that the output triage results are adapted to a structured and universal departmental architecture, and also captures the semantic-level correlation between text vectors, thereby improving the accuracy of the second triage results.

[0036] In some embodiments, the training method for a multi-level text matching model includes: S1: Obtain historical outpatient data from medical institutions at all levels; Specifically, the process begins by collecting historical medical records from multiple levels of medical institutions. These records may include, but are not limited to, fields such as patient information, chief complaint information, and department visited. The data may also undergo preprocessing such as data cleaning and formatting.

[0037] S2: Extract sample medical visit information and corresponding positive and negative sample department introduction information from historical medical visit data; Specifically, text extraction techniques based on LLM (e.g., Qwen, GPT) can be used to extract standardized patient information, patient chief complaints (as sample consultation information), information about the department the patient visited (as positive sample department information), and information about other departments (as negative sample department information).

[0038] S3: Based on the preset department level and multi-level department mapping relationship, map the positive sample department introduction information to the corresponding department level to generate positive sample multi-level department introduction information.

[0039] Specifically, departments can be pre-classified according to national standards or custom standards to obtain multiple department levels and the relationships between these levels, and a multi-level department mapping relationship can be constructed. For example, the multi-level department mapping relationship = {Level 1 Department → Level 2 Department → Level 3 Department}; the number of levels can be more or less, which is not limited here.

[0040] For example, the formatted sample data is as follows: {Patient Information, Patient Chief Complaint, Department Visited (Level 1), Department Visited (Level 2), Department Visited (Level 3)} After mapping, we can obtain: {Patient Information xxx, Chief Complaint xxx, Department Visited (Internal Medicine), Department Visited (Gastroenterology), Department Visited (blank)} S4: Train the multi-level text matching model based on sample medical information, positive sample multi-level department introduction information, and negative sample department introduction information.

[0041] Specifically, the multi-level text matching model can be fine-tuned based on a pre-trained vectorized model. Query: Patient information + Patient chief complaint (sample medical visit information) Pos: Introduction to the scope of medical services corresponding to the departments after multi-level mapping (positive sample multi-level department introduction information) Neg: Description of the scope of services offered by other departments (information on departments with negative sample information) The generated training data is in the following format: Training sample data = (Query, Pos, Neg) Training method: The training objective combines the triplet intra-batch softmax loss function of Matryoshka Representation Learning (MRL). For fine-tuning the multi-level text matching model, the AdamW optimizer can be used with a small learning rate, such as 5e-6, and it can be combined with the Cosine LR Scheduler.

[0042] In some embodiments, training a multi-level text matching model based on sample consultation information, positive sample multi-level department introduction information, and negative sample department introduction information includes: calculating a loss function under a single embedding dimension based on sample consultation information, positive sample multi-level department introduction information, and negative sample department introduction information under a single embedding dimension; calculating positive sample similarity and negative sample similarity under a single embedding dimension respectively; concatenating positive sample similarity and negative sample similarity to obtain the loss function for each sample; and obtaining a total loss function based on the loss function for each sample, and using the total loss function to train the multi-level text matching model.

[0043] Specifically, This represents the softmax loss for a single embedding dimension. For a given query (sample medical information) with embedding Query, Embedding of Pos for multi-level departmental information in positive samples Negative sample department information embedding (Neg) Where B is the batch size, D is the embedding dimension, and N is the number of negative samples per query, then the loss for a single dimension d is:

[0044] in, These represent the embedding of queries (sample medical information), positive samples (positive sample multi-level department introduction information), and negative samples (negative sample department introduction information) truncated to dimension d, respectively.

[0045] Furthermore, L2 normalization can be performed on the query, positive sample, and negative sample embeddings to obtain:

[0046] The formula for calculating positive sample similarity is:

[0047] The formula for calculating negative sample similarity is:

[0048] After concatenating the similarity scores of positive and negative samples:

[0049] Where τ is the temperature parameter.

[0050] Then, calculate the loss function for each sample: Next, the total loss function can be calculated using Matryoshka Representation Learning:

[0051] Finally, the total loss function can be used to train the multi-level text matching model. In this embodiment, the fine-tuning training method improves the prediction accuracy of the multi-level text matching model, thereby making the multi-level text matching model compatible with the triage data processing method involved in this application.

[0052] In some embodiments, the training method for the triage language model may include: S21: Extract sample visit information and corresponding positive sample department information from historical visit data, and generate sample data pairs.

[0053] S22: Supervised fine-tuning training of the triage language model based on sample data.

[0054] In the Supervised Fine-Tuning (SFT) phase, the sample data is first sampled, and patient information and chief complaint information (medical visit information) are input into, for example, an LLM (Large Language Model) to obtain samples with chain-of-thought.

[0055] Input data: A set of labeled sample data pairs ,in For the question (sample medical information). This is the standard answer (positive sample department introduction information corresponding to the sample medical information).

[0056] Training method: During the model fine-tuning stage, the training objective can use cross-entropy loss and the AdamW optimizer, with a small learning rate such as 5e-6 and a Cosine LR schedule to minimize the negative log-likelihood loss function.

[0057] in, For model parameters, The t-th word in the answer.

[0058] S23: Use a reward mechanism to provide feedback and correct the output of the triage language model after supervised fine-tuning training.

[0059] In the Reinforcement Learning (RL) phase, a desired reward mechanism can be constructed based on human feedback. Specifically, Format rewards ( ): Check whether the answer is given in accordance with the specified format and strategy.

[0060] Answer Rewards ( ): Verify the accuracy of the triage results.

[0061] Comprehensive reward mechanism: ; The method for determining whether the output is correct can be based on rule-based strategies or LLM-based reasoning.

[0062] For example, a rule-based approach may include: requiring a large language model to generate an answer in a specified format, obtaining the answer through format parsing, and then performing text matching.

[0063] LLM-based reasoning methods can include: constructing a Prompt (cue word) using a triage question-triage answer-LLM response triple, using LLM as the evaluator, and generating a yes / no evaluation on whether the answer and response are consistent.

[0064] S24: Select candidate output data from the output results corresponding to the current training batch, update the training strategy based on the candidate output data, and use the updated training strategy to optimize the triage language model after feedback correction training.

[0065] Specifically, reinforcement learning-related optimization strategies can be employed, such as Generalized Reward-based Policy Optimization (GRPO), which calculates relative advantages in response batches to optimize the model's output.

[0066] The optimization objective function of GRPO can be expressed as:

[0067] Specifically, this may include: Expected Value Calculation: In this section, This indicates that the expected value is calculated for all input queries q (sample medical information), and the distribution of the query comes from the training data $$P(Q)$$. This indicates that for each query q, from the old strategy Sample G candidate output data.

[0068] Policy Ratio: This section measures the new policy. generate The probability compared to the old strategy The changes.

[0069] Clipped Objective section: This section is accessed through... Function limits the scope of variation of the new strategy This is to prevent drastic changes from causing the model to crash.

[0070] Advantage Estimate: This section measures the advantage. The relative quality within the same set of candidate output data.

[0071] KL Divergence Regularization: This part is used to compute the new policy. Reference Strategy The KL divergence between the two strategies is used to prevent the new strategy from deviating too far from the reference strategy.

[0072] More specifically, the optimizer used to train the triage language model can be the AdamW optimizer, which can employ a small learning rate, such as 5e-6, and combined with the Cosine LR Scheduler to stabilize the training process.

[0073] In this embodiment, training the triage language model using the above method can improve the prediction accuracy of the triage language model, thereby making the triage language model compatible with the triage data processing method involved in this application.

[0074] In some embodiments, obtaining a final triage result based on a first triage result and a second triage result includes: calculating a comprehensive score of the first triage result and the second triage result based on a reciprocal ranking fusion algorithm, and determining the final triage result based on the comprehensive score.

[0075] In this embodiment, the RRF (Reciprocal Rank Fusion) algorithm can be used to fuse multiple triage results from multiple ranking lists. For example, given k ranking lists (e.g., triage results from different model outputs), the first triage result (represented as...) r 1( d ) and the results of the second triage (represented as r 2( d The RRF score of ) is calculated as follows:

[0076] in, For triage result d in the first... i The ranking is determined by the number of sorted lists. c is a smoothing constant. Finally, the final triage result is obtained. This embodiment can filter more accurate triage results from multiple triage results from different models. Since different models (such as LLM based on semantic understanding and multi-level text matching models based on structured matching) can perform complementary analyses from different dimensions such as semantic breadth and structural precision, this solution effectively overcomes the limitations of a single model by fusing the outputs of these heterogeneous models, significantly improving the comprehensiveness of triage decisions and the reliability of the final result.

[0077] More specifically, you can refer to Figure 2 As shown, Figure 2 The diagram illustrates the technical concepts of triage data processing methods in some application examples. For Figure 2 For a better understanding, please refer to the detailed explanation of the relevant parts in any one or more of the above embodiments, which will not be repeated here.

[0078] It should be understood that, although Figures 1 to 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figures 1 to 2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0079] In some embodiments, such as Figure 3 As shown, a triage data processing device is provided, including: an information acquisition module 310, a first prediction module 320, a second prediction module 330, and a result fusion module 340, wherein: The information acquisition module 310 is used to acquire the patient's medical information and the department introduction information of each department of the current hospital. The first prediction module 320 is used to obtain the first triage result based on the patient information, the introduction information of each department and the trained triage language model; The second prediction module 330 is used to obtain a second triage result based on the patient's medical information, departmental information, and a trained multi-level text matching model; wherein, the multi-level text matching model is a model used to represent the association between the patient's medical information and the triage result that conforms to the preset multi-level departmental mapping relationship; and The result fusion module 340 is used to obtain the final triage result based on the first triage result and the second triage result.

[0080] In some embodiments, the information acquisition module 310 is further configured to acquire consultation information input by the user and patient information of the patient; extract symptom keywords from the consultation information to generate follow-up question prompts; receive interactive information input by the user based on the follow-up question prompts, extract chief complaint information from the interactive information; and use the chief complaint information and patient information as medical information.

[0081] In some embodiments, the second prediction module 330 is further configured to convert the patient visit information into a patient visit text vector; map the department introduction information to the multi-level department mapping relationship according to the department introduction information and the preset multi-level department mapping relationship, and generate a multi-level department text vector; and obtain a second triage result according to the patient visit text vector, the multi-level department text vector and the trained multi-level text matching model.

[0082] In some embodiments, the second prediction module 330 is further configured to acquire historical outpatient data of medical institutions at various levels; extract sample outpatient information and corresponding positive and negative sample department introduction information from the historical outpatient data; map the positive sample department introduction information to the corresponding department level according to the preset department level and multi-level department mapping relationship to generate positive sample multi-level department introduction information; and train the multi-level text matching model based on the sample outpatient information, positive sample multi-level department introduction information and negative sample department introduction information.

[0083] In some embodiments, the second prediction module 330 calculates the loss function under a single embedding dimension based on the sample consultation information, positive sample multi-level department introduction information, and negative sample department introduction information; calculates the positive sample similarity and negative sample similarity under a single embedding dimension respectively; concatenates the positive sample similarity and negative sample similarity to obtain the loss function for each sample; and obtains the total loss function based on the loss function for each sample, and uses the total loss function to train the multi-level text matching model.

[0084] In some embodiments, the first prediction module 320 is further configured to extract sample visit information and positive sample department introduction information corresponding to the sample visit information from historical visit data, and generate sample data pairs; perform supervised fine-tuning training on the triage language model based on the sample data pairs; use a reward mechanism to provide feedback correction training on the output results of the triage language model after supervised fine-tuning training; and use candidate output data from the output results corresponding to the current training batch, update the training strategy based on the candidate output data, and optimize the triage language model after feedback correction training with the updated training strategy.

[0085] In some embodiments, the result fusion module 340 is further configured to calculate a comprehensive score of the first triage result and the second triage result based on a reciprocal sorting fusion algorithm, and determine the final triage result based on the comprehensive score.

[0086] Specific limitations regarding the triage data processing device can be found in the limitations of the triage data processing method described above, and will not be repeated here. Each module in the aforementioned triage data processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0087] In some embodiments, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a diagnostic data processing method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0088] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0089] In some embodiments, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps: acquiring the patient's medical information and department descriptions of various departments in the currently visited hospital; obtaining a first triage result based on the medical information, department descriptions, and a trained triage language model; obtaining a second triage result based on the medical information, department descriptions, and a trained multi-level text matching model; wherein the multi-level text matching model is a model used to characterize the association between the patient's medical information and the triage results that conform to a preset multi-level department mapping relationship; and obtaining a final triage result based on the first and second triage results.

[0090] In some embodiments, when the processor executes the computer program, it further performs the following steps: acquiring consultation information input by the user and patient information of the patient; extracting symptom keywords from the consultation information to generate follow-up question prompts; receiving interactive information input by the user based on the follow-up question prompts, extracting chief complaint information from the interactive information; and using the chief complaint information and patient information as medical consultation information.

[0091] In some embodiments, when the processor executes the computer program, it further performs the following steps: converting the patient information into a patient text vector; mapping the department introduction information to the multi-level department mapping relationship according to the department introduction information and the preset multi-level department mapping relationship, and generating a multi-level department text vector; and obtaining a second triage result based on the patient text vector, the multi-level department text vector, and the trained multi-level text matching model.

[0092] In some embodiments, when the processor executes the computer program, it further performs the following steps: acquiring historical outpatient data of medical institutions at various levels; extracting sample outpatient information and corresponding positive and negative sample department introduction information from the historical outpatient data; mapping the positive sample department introduction information to the corresponding department level according to the preset department level and multi-level department mapping relationship to generate positive sample multi-level department introduction information; and training a multi-level text matching model based on the sample outpatient information, positive sample multi-level department introduction information, and negative sample department introduction information.

[0093] In some embodiments, when the processor executes the computer program, it further performs the following steps: calculating a loss function under a single embedding dimension based on sample consultation information, positive sample multi-level department introduction information, and negative sample department introduction information; calculating positive sample similarity and negative sample similarity under a single embedding dimension respectively; concatenating the positive sample similarity and negative sample similarity to obtain the loss function for each sample; and obtaining the total loss function based on the loss function for each sample, and using the total loss function to train the multi-level text matching model.

[0094] In some embodiments, when the processor executes the computer program, it further implements the following steps: extracting sample medical information and positive sample department introduction information corresponding to the sample medical information from historical medical data, and generating sample data pairs; performing supervised fine-tuning training on the triage language model based on the sample data pairs; using a reward mechanism to provide feedback correction training on the output results of the triage language model after supervised fine-tuning training; and using candidate output data from the output results corresponding to the current training batch, updating the training strategy based on the candidate output data, and using the updated training strategy to optimize the triage language model after feedback correction training.

[0095] In some embodiments, when the processor executes the computer program, it further implements the following steps: calculating a comprehensive score of the first triage result and the second triage result based on the inverse sorting fusion algorithm, and determining the final triage result based on the comprehensive score.

[0096] In some embodiments, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program performs the following steps: acquiring the patient's medical information and the department descriptions of the various departments of the currently visited hospital; obtaining a first triage result based on the medical information, the department descriptions, and a trained triage language model; obtaining a second triage result based on the medical information, the department descriptions, and a trained multi-level text matching model; wherein the multi-level text matching model is a model used to characterize the association between the patient's medical information and the triage results that conform to a preset multi-level department mapping relationship; and obtaining a final triage result based on the first and second triage results.

[0097] In some embodiments, when the computer program is executed by the processor, it further performs the following steps: acquiring consultation information input by the user and patient information of the patient; extracting symptom keywords from the consultation information to generate follow-up question prompts; receiving interactive information input by the user based on the follow-up question prompts, extracting chief complaint information from the interactive information; and using the chief complaint information and patient information as medical consultation information.

[0098] In some embodiments, when the computer program is executed by the processor, it further performs the following steps: converting the patient information into a patient text vector; mapping the department introduction information to the multi-level department mapping relationship according to the department introduction information and the preset multi-level department mapping relationship, and generating a multi-level department text vector; and obtaining a second triage result based on the patient text vector, the multi-level department text vector, and the trained multi-level text matching model.

[0099] In some embodiments, when the computer program is executed by the processor, it further performs the following steps: acquiring historical medical visit data of medical institutions at various levels; extracting sample medical visit information and corresponding positive sample department introduction information and negative sample department introduction information from the historical medical visit data; mapping the positive sample department introduction information to the corresponding department level according to the preset department level and multi-level department mapping relationship to generate positive sample multi-level department introduction information; and training a multi-level text matching model based on the sample medical visit information, positive sample multi-level department introduction information and negative sample department introduction information.

[0100] In some embodiments, when the computer program is executed by the processor, it further performs the following steps: calculating a loss function under a single embedding dimension based on sample consultation information, positive sample multi-level department introduction information, and negative sample department introduction information; calculating positive sample similarity and negative sample similarity under a single embedding dimension respectively; concatenating the positive sample similarity and negative sample similarity to obtain the loss function for each sample; and obtaining the total loss function based on the loss function for each sample, and using the total loss function to train the multi-level text matching model.

[0101] In some embodiments, when the computer program is executed by the processor, it further performs the following steps: extracting sample medical information and positive sample department introduction information corresponding to the sample medical information from historical medical data, and generating sample data pairs; performing supervised fine-tuning training on the triage language model based on the sample data pairs; using a reward mechanism to provide feedback correction training on the output results of the triage language model after supervised fine-tuning training; and using candidate output data from the output results corresponding to the current training batch, updating the training strategy based on the candidate output data, and using the updated training strategy to optimize the triage language model after feedback correction training.

[0102] In some embodiments, when the computer program is executed by the processor, it further implements the following steps: calculating a comprehensive score of the first triage result and the second triage result based on the inverse sorting fusion algorithm, and determining the final triage result based on the comprehensive score.

[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the characters in this article generally indicate that the preceding and following related objects have an "or" relationship.

[0106] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

[0107] It should be noted that in the embodiments of this application, data related to user information or user data (e.g., patient's medical information, patient's patient information, chief complaint information, historical medical data, etc.) need to be obtained and processed only after the user's authorization and consent. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

Claims

1. A triage data processing method, the method comprising: Obtain the patient's medical information and departmental information of the hospital where they are currently receiving treatment; Based on the patient information, the department descriptions, and the trained triage language model, the first triage result is obtained. Based on the patient's medical information, the descriptions of each department, and the trained multi-level text matching model, a second triage result is obtained; wherein, the multi-level text matching model is a model used to characterize the association between the patient's medical information and the triage result that conforms to a preset multi-level department mapping relationship; as well as The final triage result is obtained based on the first triage result and the second triage result.

2. The method according to claim 1, characterized in that, The acquisition of the patient's medical information includes: Obtain the consultation information input by the user and the patient's patient information; Extract symptom keywords from the consultation information to generate prompts for further questions about the condition; Receive interactive information input by the user based on the prompts for further inquiry about the patient's condition, and extract the chief complaint information from the interactive information; and The chief complaint information and the patient information are used as the medical visit information.

3. The method according to claim 1, characterized in that, The second triage result is obtained based on the patient information, the department descriptions, and the trained multi-level text matching model, including: The medical information is converted into a medical text vector; Based on the department introduction information and the preset multi-level department mapping relationship, the department introduction information is mapped to the multi-level department mapping relationship, and multi-level department text vectors are generated; and The second triage result is obtained based on the patient visit text vector, the multi-level department text vector, and the trained multi-level text matching model.

4. The method according to claim 1, characterized in that, The training method for the multi-level text matching model includes: Obtain historical patient visit data from medical institutions at all levels; Extract sample medical visit information and corresponding positive and negative sample department information from the historical medical visit data; Based on preset department levels and multi-level department mapping relationships, the positive sample department introduction information is mapped to the corresponding department level to generate positive sample multi-level department introduction information; and The multi-level text matching model is trained based on the sample medical information, the positive sample multi-level department introduction information, and the negative sample department introduction information.

5. The method according to claim 4, characterized in that, The step of training the multi-level text matching model based on the sample medical visit information, the positive sample multi-level department introduction information, and the negative sample department introduction information includes: The loss function under a single embedding dimension is calculated based on the sample medical information, the positive sample multi-level department introduction information, and the negative sample department introduction information. Calculate the similarity of positive and negative samples under a single embedding dimension; The positive sample similarity and the negative sample similarity are concatenated to obtain the loss function for each sample; and The total loss function is obtained based on the loss function of each sample, and the multi-level text matching model is trained using the total loss function.

6. The method according to claim 1, characterized in that, The training method for the triage language model includes: Extract sample medical visit information and corresponding positive sample department information from historical medical visit data, and generate sample data pairs; The triage language model is trained under supervised fine-tuning based on the sample data. The output of the triage language model, after supervised fine-tuning training, is used to provide feedback and correction training through a reward mechanism; and Candidate output data is selected from the output results corresponding to the current training batch. The training strategy is updated based on the candidate output data, and the triage language model after feedback correction training is optimized and trained using the updated training strategy.

7. The method according to any one of claims 1 to 6, characterized in that, The step of obtaining the final triage result based on the first triage result and the second triage result includes: The comprehensive score of the first triage result and the second triage result is calculated based on the inverse sorting fusion algorithm, and the final triage result is determined based on the comprehensive score.

8. A triage data processing device, characterized in that, The device includes: The information acquisition module is used to acquire the patient's medical information and the department introduction information of each department of the current hospital. The first prediction module is used to obtain the first triage result based on the patient information, the introduction information of each department, and the trained triage language model; The second prediction module is used to obtain a second triage result based on the patient's medical information, the descriptions of each department, and a trained multi-level text matching model; wherein the multi-level text matching model is a model used to characterize the association between the patient's medical information and the triage result that conforms to a preset multi-level department mapping relationship; and The result fusion module is used to obtain the final triage result based on the first triage result and the second triage result.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images