An intelligent outpatient triage method and system based on large model

Abstract

The present invention relates to the field of intelligent medical technology, and in particular to an intelligent outpatient triage method and system based on a large model. An intelligent outpatient triage system based on a large model comprises: a data acquisition module and an outpatient triage module. The present invention determines the triage priority corresponding to the patient by sending the patient's structured data into the outpatient triage model for processing, and performs triage operations on the patient based on the patient's corresponding triage priority, without the need for medical staff to make judgments based on experience; four different judgment criteria are incorporated into the outpatient triage model, and the feature information expression effect and feature information fusion effect under the perspectives of the four different judgment criteria are improved through a multi-standard feature extraction layer, a feature transmission layer, and a difference complementary enhancement layer, so that the outpatient triage model can be adapted to more clinical scenarios and has higher adaptability to complex patient conditions.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical technology, and in particular to an intelligent outpatient triage method and system based on a large model. Background Art

[0002] Outpatient triage is a critical component in optimizing medical resource allocation and ensuring patient safety. However, current outpatient triage faces significant challenges. Traditional manual triage relies heavily on the clinical experience and subjective judgment of triage personnel, leading to significant individual variability and subjective bias in triage results. Different medical staff often produce significantly different triage results for the same patient, making consistency difficult to ensure. Furthermore, during peak hours when crowds are dense, manual triage is inefficient, easily leading to patient delays and increasing potential risks.

[0003] Currently, the triage standards widely used internationally, such as ESI (Emergency Severity Index), CTAS (Canadian Emergency Triage Urgency Scale), MTS (Manchester Triage System) and ATS (Australian Triage Scale), all have their own inherent limitations: ESI focuses on resource consumption prediction, but it is insufficient in capturing the dynamic changes in clinical urgency, especially in sensitivity to elderly patients and atypical symptoms; although CTAS has a strict time response framework, it lacks flexibility in assessing complex patients with multi-system symptoms, and its accuracy in assessing acute exacerbations of chronic diseases is limited; MTS uses a symptom-oriented algorithm, but the decision-making path is unclear when faced with complex cases with multiple symptoms, and it lacks an assessment dimension for prognostic risk; because ATS was originally designed for the Australian medical environment, its adaptability varies under different medical cultural backgrounds and resource allocation models; in summary, the adaptability of each triage standard is not high. Summary of the Invention

[0004] The present invention sends the patient's structured data into the outpatient triage model for processing, determines the patient's corresponding triage priority, and performs triage operations on the patient based on the patient's corresponding triage priority, without the need for medical staff to make judgments based on experience; four different judgment criteria are integrated into the outpatient triage model, and the feature information expression effect and feature information fusion effect under the four different judgment criteria perspectives are improved through a multi-standard feature extraction layer, a feature transmission layer, and a difference complementary enhancement layer, so that the outpatient triage model can be adapted to more clinical scenarios and has higher adaptability to complex patient conditions.

[0005] The present invention provides an intelligent outpatient triage method based on a large model, comprising:

[0006] Send the patient's structured data to the outpatient triage model for processing and output the patient's corresponding triage priority;

[0007] The outpatient triage model includes a multi-standard feature extraction layer, a feature transfer layer, a difference complementary enhancement layer, a multi-standard scoring processing layer, a weighted voting layer and a triage priority output layer. The multi-standard feature extraction layer is used to extract features of different branches for patient structured data based on knowledge graph data, and construct corresponding ESI feature data, CTAS feature data, MTS feature data and ATS feature data; the feature transfer layer is used to transfer features between ESI feature data, CTAS feature data, MTS feature data and ATS feature data through a bidirectional gating unit, and construct corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data; the difference complementary enhancement layer is used to extract features of different branches for patient structured data based on knowledge graph data ... The enhancement layer is used to perform differential analysis on the ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data, and ATS enhanced feature data, and to use the obtained differential information to reinforce the ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data, and ATS enhanced feature data to construct a feature judgment analysis vector; the multi-standard scoring processing layer is used to construct corresponding ESI scores, CTAS scores, MT scores, and ATS scores based on the feature judgment analysis vector; the weighted voting layer is used to perform weighted voting operations on the ESI scores, CTAS scores, MT scores, and ATS scores according to the corresponding voting weights to obtain the triage priority; the triage priority output layer is used to output the triage priority.

[0008] Preferably, the corresponding ESI feature data, CTAS feature data, MTS feature data and ATS feature data are constructed through the multi-standard feature extraction layer, specifically including the following steps:

[0009] The multi-standard feature extraction layer has built-in ESI feature extraction unit, CTAS feature extraction unit, MTS feature extraction unit and ATS feature extraction unit, which are used to extract features of different branches respectively;

[0010] Any one of the ESI feature extraction unit, CTAS feature extraction unit, MTS feature extraction unit and ATS feature extraction unit is recorded as the target feature extraction unit, and the following steps are performed for the target feature extraction unit: in the target feature extraction unit, the patient structured data is traversed, and relevant knowledge graph triples are queried from the knowledge graph corresponding to the target feature extraction unit, and all relevant knowledge graph triplets are combined into a knowledge graph feature graph. In the process of feature extraction by the target feature extraction unit, the self-attention mechanism is executed based on the patient structured data and the knowledge graph feature graph, and the corresponding query vector is constructed with the knowledge graph feature graph; after the feature extraction operation is completed by the feature extraction unit, the corresponding ESI feature data, CTAS feature data, MTS feature data and ATS feature data are output.

[0011] Preferably, the corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data are constructed through the feature transfer layer, specifically including the following steps:

[0012] Building a bidirectional gating unit between any two feature data among the ESI feature data, the CTAS feature data, the MTS feature data and the ATS feature data;

[0013] For the ESI feature data, any one of the remaining three CTAS feature data, MTS feature data, and ATS feature data is selected as the target feature data. The following contents are executed in the bidirectional gating unit between the ESI feature data and the target feature data: the ESI feature data and the target feature data are concatenated and recorded as the overall feature data Z. The forget gate calculation is performed on the overall feature data: f t =σ(W f Z+b f ), where f t is the forget gate vector, σ is the sigmoid function, W f is the forget gate weight matrix, b f is the forget gate bias value, and the input gate is calculated for the overall feature data: i t =σ(W i Z+b i ), where i t is the input gate vector, W i is the input gate weight matrix, b i is the input gate bias value, and feature transfer is completed through the following formula: G=f t *R+i t *R, where G is the ESI feature transfer data corresponding to the ESI feature data, R is the target feature data, and * is the dot product operation; the three ESI feature transfer data corresponding to the ESI feature data are weighted summed and then residually connected with the ESI feature data to obtain the corresponding ESI enhanced feature data; the ESI feature data, CTAS feature data, MTS feature data and ATS feature data are traversed to construct the corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data.

[0014] Preferably, constructing a feature judgment analysis vector through a difference complementary enhancement layer specifically includes the following steps:

[0015] The ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data are respectively sent to the semantic mapping network for processing, and the corresponding ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector are output;

[0016] The ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector are summed and averaged to obtain the average feature vector. The ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector are then differentially calculated with the average feature vector to construct the corresponding ESI differential feature vector, CTAS differential feature vector, MTS differential feature vector and ATS differential feature vector. The self-attention mechanism is performed on the ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector, respectively. The corresponding query vector is constructed with the corresponding ESI differential feature vector, CTAS differential feature vector, MTS differential feature vector and ATS differential feature vector to obtain the ESI feature vector to be fused, CTAS feature vector to be fused, MTS feature vector to be fused and ATS feature vector to be fused. The ESI feature vector to be fused, CTAS feature vector to be fused, MTS feature vector to be fused and ATS feature vector to be fused are then concatenated according to channels and a convolution operation is performed to obtain the feature judgment analysis vector.

[0017] Preferably, the triage priority is obtained through a weighted voting layer, which specifically includes the following steps:

[0018] According to the ESI score, CTAS score, MT score and ATS score, the decision vectors corresponding to the ESI score, CTAS score, MT score and ATS score are determined from the decision vector library; the corresponding decision vectors are added to the end of the ESI score, CTAS score, MT score and ATS score to construct the corresponding ESI decision analysis vector, CTAS decision analysis vector, MT decision analysis vector and ATS decision analysis vector, and then the ESI decision analysis vector, CTAS decision analysis vector, MT decision analysis vector and ATS decision analysis vector are sent to the multi-layer perceptron for processing, and the corresponding voting weights are output. The ESI score, CTAS score, MT score and ATS score are weightedly voted according to the corresponding voting weights to obtain the triage priority.

[0019] Preferably, the training of the outpatient triage model specifically includes the following steps:

[0020] Obtain several outpatient triage training samples, which include patient structured data. Label the outpatient triage training samples according to triage priority, and form an outpatient triage training set with all labeled outpatient triage training samples. Train the outpatient triage model with the outpatient triage training set. During training, use the labeled triage priority as the training target to determine whether the training conditions are met. If the training conditions are met, output the trained outpatient triage model; otherwise, continue to train the outpatient triage model with the outpatient triage training set.

[0021] The present invention also provides an intelligent outpatient triage system based on a large model, comprising:

[0022] The data acquisition module is used to obtain the patient's basic information, chief complaint data and basic vital signs data, pre-process the patient's basic information, chief complaint data and basic vital signs data, and construct the corresponding patient structured data;

[0023] The outpatient triage module is used to feed the patient's structured data into the outpatient triage model for processing and output the patient's corresponding triage priority.

[0024] The present invention has the following advantages:

[0025] The present invention sends the patient's structured data into the outpatient triage model for processing, determines the patient's corresponding triage priority, and performs triage operations on the patient based on the patient's corresponding triage priority, without the need for medical staff to make judgments based on experience; four different judgment criteria are integrated into the outpatient triage model, and the feature information expression effect and feature information fusion effect under the four different judgment criteria perspectives are improved through a multi-standard feature extraction layer, a feature transmission layer, and a difference complementary enhancement layer, so that the outpatient triage model can be adapted to more clinical scenarios and has higher adaptability to complex patient conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a structural diagram of the outpatient triage model adopted in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of the large model-based intelligent outpatient triage system adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0029] Example 1, a large model-based intelligent outpatient triage method, comprising:

[0030] Obtain the patient's basic information, chief complaint data and basic vital signs data. It should be noted that for outpatient triage in the emergency department, the patient needs to arrive at the emergency triage desk first. The nurse will enter the patient's basic information, chief complaint data and basic vital signs data at the emergency triage desk. The basic information generally includes the patient's age, weight, gender and allergy history. The chief complaint data generally refers to the patient's self-reported symptom description, such as chest pain for two hours. Vital signs data refers to data such as blood pressure, heart rate and body temperature. By analyzing these entered data, the urgency of the patient's condition can be judged, and then triage operations can be performed; the patient's basic information, chief complaint data and basic vital signs data are preprocessed to construct the corresponding patient structured data. It should be noted that the preprocessing here includes directly filling the original numerical content, performing named entity recognition on non-numerical text, and mapping the colloquial content to medical professional terms, and then performing word embedding operations to construct numerical content. The patient's structured data is a vector composed of a series of numerical values;

[0031] The patient's structured data is fed into the outpatient triage model for processing, and the patient's corresponding triage priority is output. The triage priority is generally divided into levels 1-5, with level 1 being the most urgent. The triage priority reflects the urgency of the patient's condition. The triage operation is performed on the patient based on the patient's corresponding triage priority. The triage operation for the patient specifically involves assigning the patient to various waiting areas and setting waiting times. The outpatient triage model can output a reference triage priority based on the entered patient data, eliminating the need for nurses to perform triage operations on patients based on experience, which can improve the efficiency of outpatient triage.

[0032] See also Figure 1The outpatient triage model includes a multi-standard feature extraction layer, a feature transfer layer, a difference complementary enhancement layer, a multi-standard scoring processing layer, a weighted voting layer and a triage priority output layer. The multi-standard feature extraction layer is used to extract features of different branches for patient structured data based on knowledge graph data, and construct corresponding ESI feature data, CTAS feature data, MTS feature data and ATS feature data. It should be noted that the ESI feature data, CTAS feature data, MTS feature data and ATS feature data here correspond to the four major international emergency triage standards: ESI (Emergency Severity Index), CTAS (Additional The system extracts features from patient structured data through the knowledge graph corresponding to the four international emergency triage standards, incorporates multiple judgment criteria into the outpatient triage process, and enhances the clinical interpretability and transparency of the outpatient triage model; the feature transfer layer is used to transfer features between ESI feature data, CTAS feature data, MTS feature data and ATS feature data through a bidirectional gating unit to construct corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data. Data, through the feature transfer process, can fuse the valuable features under different judgment criteria, and filter out irrelevant or possibly interfering features, so that the feature information under each judgment criterion has the perspective under other judgment criteria; the difference complementary enhancement layer is used to perform difference analysis on the ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data, and use the obtained difference information to reinforce the ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data to construct a feature judgment analysis vector. Since different judgment criteria have corresponding specialties, for example For example, ESI mainly considers resource requirements and life-threatening situations, while CTAS focuses on symptom severity and time sensitivity. The differences between features under different judgment criteria are used to reflect the specialties corresponding to different judgment criteria, and then the features related to these specialties are retained and strengthened, and the features under different judgment criteria are further integrated. The multi-standard scoring processing layer is used to perform full-connection operations on different branches of the feature judgment analysis vector to construct the corresponding ESI score, CTAS score, MT score, and ATS score. The ESI score, CTAS score, MT score, and ATS score reflect the urgency of the patient's emergency condition from the perspective of different judgment criteria.The weighted voting layer is used to determine the voting weights corresponding to the ESI score, CTAS score, MT score and ATS score according to the ESI score, CTAS score, MT score and ATS score, and perform weighted voting operations on the ESI score, CTAS score, MT score and ATS score according to the corresponding voting weights to obtain the triage priority. By weighted summing the scoring results under different judgment criteria, it is possible to consider multiple judgment criteria in the process of emergency outpatient triage, making it more adaptable; the triage priority output layer is used to output the triage priority;

[0033] This application sends the patient's structured data into the outpatient triage model for processing, determines the patient's corresponding triage priority, and performs triage operations on the patient based on the patient's corresponding triage priority, without the need for medical staff to make judgments based on experience; incorporates four different judgment criteria into the outpatient triage model, and uses a multi-standard feature extraction layer, a feature transmission layer, and a difference complementary enhancement layer to improve the feature information expression effect and feature information fusion effect under the four different judgment criteria perspectives, so that the outpatient triage model can be adapted to more clinical scenarios and has higher adaptability to complex patient conditions.

[0034] Through the multi-standard feature extraction layer, based on the knowledge graph data, feature extraction of different branches is performed on the patient structured data to construct the corresponding ESI feature data, CTAS feature data, MTS feature data, and ATS feature data. The specific steps include the following:

[0035] The multi-standard feature extraction layer has built-in ESI feature extraction units, CTAS feature extraction units, MTS feature extraction units, and ATS feature extraction units, which are used to extract features from different branches respectively. It should be noted that the ESI feature extraction units, CTAS feature extraction units, MTS feature extraction units, and ATS feature extraction units here are all set based on the encoder in the Transformer model.

[0036] In the ESI feature extraction unit, the patient structured data is traversed, and relevant ESI knowledge graph triples are queried from the ESI knowledge graph. All relevant ESI knowledge graph triplets are combined into an ESI knowledge graph feature graph. During the feature extraction process by the ESI feature extraction unit, a self-attention mechanism is executed based on the patient structured data and the ESI knowledge graph feature graph, and the corresponding query vector is constructed using the ESI knowledge graph feature graph. After feature extraction is completed, the ESI feature data is output;

[0037] In the CTAS feature extraction unit, the patient structured data is traversed, and relevant CTAS knowledge graph triples are queried from the CTAS knowledge graph. All relevant CTAS knowledge graph triplets are combined into a CTAS knowledge graph feature graph. During the feature extraction process by the CTAS feature extraction unit, a self-attention mechanism is executed based on the patient structured data and the CTAS knowledge graph feature graph, and the corresponding query vector is constructed using the CTAS knowledge graph feature graph. After feature extraction is completed, the CTAS feature data is output;

[0038] In the MTS feature extraction unit, the patient structured data is traversed, and relevant MTS knowledge graph triples are queried from the MTS knowledge graph. All relevant MTS knowledge graph triplets are combined into an MTS knowledge graph feature graph. During the feature extraction process by the MTS feature extraction unit, a self-attention mechanism is executed based on the patient structured data and the MTS knowledge graph feature graph, and the corresponding query vector is constructed using the MTS knowledge graph feature graph. After feature extraction is completed, the MTS feature data is output;

[0039] In the ATS feature extraction unit, the patient structured data is traversed, and relevant ATS knowledge graph triples are queried from the ATS knowledge graph. All relevant ATS knowledge graph triples are combined into an ATS knowledge graph feature graph. During the feature extraction process by the ATS feature extraction unit, a self-attention mechanism is executed based on the patient structured data and the ATS knowledge graph feature graph, and the corresponding query vector is constructed using the ATS knowledge graph feature graph. After feature extraction is completed, the ATS feature data is output;

[0040] It should be noted that the ESI knowledge graph, CTAS knowledge graph, MTS knowledge graph, and ATS knowledge graph are all constructed based on the corresponding professional guidance texts. They generally include named entity extraction, entity relationship extraction, and entity-entity relationship-entity triple construction, which are used to reflect the decision content under different judgment criteria and guide feature extraction methods with different focuses.

[0041] The process of feature extraction through the ESI feature extraction unit, CTAS feature extraction unit, MTS feature extraction unit and ATS feature extraction unit is actually the encoder processing process of the Transformer model. When the self-attention mechanism is executed, the traditional encoder generally constructs the corresponding value vector V, key vector K and query vector Q based on the input features, and then uses the formula H=softmax(QK T / D 0.5) V implements the self-attention mechanism, where H is the output feature, T is the matrix transpose operation, and D is the dimension size of the key vector. In this application, the ESI feature extraction unit, the CTAS feature extraction unit, the MTS feature extraction unit, and the ATS feature extraction unit use the ESI knowledge graph feature graph, the CTAS knowledge graph feature graph, the MTS knowledge graph feature graph, and the ATS knowledge graph feature graph to construct the corresponding query vector. The purpose is to introduce the corresponding knowledge graph information;

[0042] The feature transfer layer transfers features between ESI feature data, CTAS feature data, MTS feature data, and ATS feature data through a bidirectional gating unit to construct corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data, and ATS enhanced feature data. Specifically, the steps include:

[0043] A bidirectional gating unit is constructed between any two feature data among the ESI feature data, CTAS feature data, MTS feature data, and ATS feature data. The bidirectional gating unit here is set with reference to the LSTM model to control the transmission of feature information;

[0044] For the ESI feature data, any one of the remaining three CTAS feature data, MTS feature data, and ATS feature data is selected as the target feature data. The following contents are executed in the bidirectional gating unit between the ESI feature data and the target feature data: the ESI feature data and the target feature data are concatenated and recorded as the overall feature data Z. The forget gate calculation is performed on the overall feature data: f t =σ(W f Z+b f ), where f t is the forget gate vector, σ is the sigmoid function, W f is the forget gate weight matrix, b f is the forget gate bias value, and the input gate is calculated for the overall feature data: i t =σ(W i Z+b i ), where i t is the input gate vector, W i is the input gate weight matrix, b i is the input gate bias value, and feature transfer is completed through the following formula: G=f t *R+i t*R, where G is the ESI feature transfer data corresponding to the ESI feature data, R is the target feature data, and * is a dot product operation; the three ESI feature transfer data corresponding to the ESI feature data are weighted summed and then residually connected with the ESI feature data to obtain the corresponding ESI enhanced feature data. It should be noted that each feature data corresponds to three bidirectional gating units, so the ESI feature data will have corresponding three ESI feature transfer data, and the forgetting gate represents the feature information that is useless to the ESI feature data in the judgment criterion corresponding to the target feature data, and does not need to be retained, while the input gate represents the feature information that is useful to the ESI feature data in the judgment criterion corresponding to the target feature data, and needs to be retained; traverse the ESI feature data, CTAS feature data, MTS feature data and ATS feature data to construct the corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data. It should be noted that the calculation process in the bidirectional gating unit is consistent, but the data facing it is inconsistent. Therefore, in this embodiment, ESI feature data is taken as an example for introduction, and the rest will not be repeated;

[0045] The ESI enhanced feature data, the CTAS enhanced feature data, the MTS enhanced feature data, and the ATS enhanced feature data are analyzed by using a difference complementary enhancement layer, and the ESI enhanced feature data, the CTAS enhanced feature data, the MTS enhanced feature data, and the ATS enhanced feature data are reinforced with the obtained difference information to construct a feature judgment analysis vector, specifically including the following steps:

[0046] The ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data, and ATS enhanced feature data are respectively sent to the semantic mapping network for processing, and the corresponding ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector, and ATS enhanced feature vector are output. It should be noted that the semantic mapping network here is the same, which is used to map the ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data, and ATS enhanced feature data to the same semantic space. The corresponding ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector, and ATS enhanced feature vector generally include multiple dimensions in an interpretable case, such as time sensitivity assessment value and intervention urgency assessment value, and can be pre-trained in advance;

[0047] The ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector are summed and averaged to obtain the average feature vector. Then, the ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector are differentially calculated with the average feature vector respectively to highlight the strengths corresponding to different judgment criteria. The corresponding ESI differential feature vector, CTAS differential feature vector, MTS differential feature vector and ATS differential feature vector are constructed. The self-attention mechanism is executed for the ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector respectively. The corresponding ESI differential feature vector, CTAS differential feature vector, MTS differential feature vector and ATS differential feature vector are used to construct the corresponding query vector, and the ESI feature vector to be fused, the CTAS feature vector to be fused, the MTS feature vector to be fused and the ATS feature vector to be fused are obtained. Then, the ESI feature vector to be fused, the CTAS feature vector to be fused, the MTS feature vector to be fused and the ATS feature vector to be fused are concatenated according to the channels and a convolution operation is performed to obtain a feature judgment analysis vector. It should be noted that the self-attention mechanism is performed by highlighting the difference differential feature vector, which can further highlight the content corresponding to the focus content under each judgment standard.

[0048] The weighted voting layer determines the voting weights corresponding to the ESI, CTAS, MT, and ATS scores based on the ESI, CTAS, MT, and ATS scores. A weighted voting operation is then performed on the ESI, CTAS, MT, and ATS scores based on the corresponding voting weights to obtain the triage priority. The specific steps are as follows:

[0049] According to the ESI score, CTAS score, MT score and ATS score, the decision vectors corresponding to the ESI score, CTAS score, MT score and ATS score are determined from the decision vector library. The decision vector is composed of the vectors of the conditional nodes on the decision path after word embedding. It should be noted that for each judgment criterion, a judgment model will be trained based on the decision tree model. Each judgment model can complete the corresponding magnitude division, that is, each score will correspond to a decision path on the corresponding decision tree, and the corresponding decision vector will be constructed based on this;

[0050] This article uses the Canadian Triage and Urgency Scale (CTAS) as an example to illustrate the construction of a judgment model. First, several sets of triage score data implemented using CTAS are obtained as training data. The triage score here includes five levels: level 1 (immediate severe), level 2 (urgent), level 3 (urgent), level 4 (less urgent), and level 5 (non-urgent). Condition nodes are determined based on the corresponding CTAS guidance documents, such as whether there is dyspnea and whether the pulse rate is normal. The decision tree is then trained using the training data.

[0051] Add the corresponding decision vectors to the end of the ESI score, CTAS score, MT score, and ATS score to construct the corresponding ESI decision analysis vector, CTAS decision analysis vector, MT decision analysis vector, and ATS decision analysis vector. Then, send the ESI decision analysis vector, CTAS decision analysis vector, MT decision analysis vector, and ATS decision analysis vector to the multilayer perceptron for processing, output the corresponding voting weights, and perform weighted voting on the ESI score, CTAS score, MT score, and ATS score according to the corresponding voting weights to obtain the triage priority. Here, the multilayer perceptron follows the end-to-end training of the outpatient triage model for parameter adjustment;

[0052] Training the outpatient triage model includes the following steps:

[0053] Obtain several outpatient triage training samples, which include patient structured data. The patient structured data here is obtained in accordance with the patient's wishes and laws and regulations. The outpatient triage training samples are labeled by triage priority. The triage priority here is set by experts. The fuzzy comprehensive evaluation method can be used to integrate the opinions of most experts, and all labeled outpatient triage training samples are composed of an outpatient triage training set. The outpatient triage model is trained with the outpatient triage training set. During training, the labeled triage priority is used as the training target to determine whether the training conditions are met. The training conditions here can be that the accuracy of the outpatient triage model meets expectations. If the training conditions are met, the trained outpatient triage model is output; otherwise, the outpatient triage model continues to be trained with the outpatient triage training set.

[0054] Example 2, an intelligent outpatient triage system based on a large model, see Figure 2 ,include:

[0055] The data acquisition module is used to obtain the patient's basic information, chief complaint data and basic vital signs data. It should be noted that for outpatient triage in the emergency department, the patient needs to arrive at the emergency triage desk first. The nurse will enter the patient's basic information, chief complaint data and basic vital signs data at the emergency triage desk. The basic information generally includes the patient's age, weight, gender and allergy history. The chief complaint data generally refers to the patient's self-reported symptom description, such as chest pain for two hours. The vital signs data refers to data such as blood pressure, heart rate and body temperature. By analyzing these entered data, the urgency of the patient's condition can be judged, and then triage operations can be performed; the patient's basic information, chief complaint data and basic vital signs data are preprocessed to construct the corresponding patient structured data. It should be noted that the preprocessing here includes directly filling the original numerical content, performing named entity recognition on non-numerical text, and mapping the colloquial content to medical professional terms, and then performing word embedding operations to construct numerical content. The patient's structured data is a vector composed of a series of numerical values;

[0056] The outpatient triage module is used to input the patient's structured data into the outpatient triage model for processing, and output the patient's corresponding triage priority. The triage priority is generally divided into 1-5 levels, with level 1 being the most urgent. The triage priority reflects the urgency of the patient's condition. The triage operation is performed on the patient based on the patient's corresponding triage priority. The triage operation for the patient is specifically to assign the patient to various waiting areas and set the waiting time. The outpatient triage model can output a reference triage priority based on the entered patient data, eliminating the need for nurses to perform triage operations on patients based on experience, which can improve the efficiency of outpatient triage.

[0057] It should be understood that those skilled in the art may make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims. Any portion of this specification not described in detail is prior art known to those skilled in the art.

Claims

1. An intelligent outpatient triage method based on a large model, characterized in that: include: Send the patient's structured data to the outpatient triage model for processing and output the patient's corresponding triage priority; The outpatient triage model includes a multi-standard feature extraction layer, a feature transfer layer, a difference complementary enhancement layer, a multi-standard scoring processing layer, a weighted voting layer, and a triage priority output layer. The multi-standard feature extraction layer is used to extract features from different branches of patient structured data based on knowledge graph data, and construct corresponding ESI feature data, CTAS feature data, MTS feature data, and ATS feature data. The feature transfer layer is used to transfer features between ESI feature data, CTAS feature data, MTS feature data and ATS feature data through a bidirectional gating unit to construct corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data; the difference complementary enhancement layer is used to perform difference analysis on the ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data, and use the obtained difference information to reinforce the ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data to construct a feature judgment analysis vector; the multi-standard scoring processing layer is used to construct corresponding ESI scores, CTAS scores, MT scores and ATS scores based on the feature judgment analysis vector; the weighted voting layer is used to perform weighted voting operations on the ESI scores, CTAS scores, MT scores and ATS scores according to the corresponding voting weights to obtain the triage priority; the triage priority output layer is used to output the triage priority; The multi-standard feature extraction layer has built-in ESI feature extraction unit, CTAS feature extraction unit, MTS feature extraction unit and ATS feature extraction unit, which are used to extract features of different branches respectively; Any one of the ESI feature extraction unit, CTAS feature extraction unit, MTS feature extraction unit, and ATS feature extraction unit is recorded as a target feature extraction unit, and the following steps are performed for the target feature extraction unit: in the target feature extraction unit, the patient structured data is traversed, and relevant knowledge graph triples are queried from the knowledge graph corresponding to the target feature extraction unit, and all relevant knowledge graph triples are combined into a knowledge graph feature graph. In the process of feature extraction by the target feature extraction unit, a self-attention mechanism is executed based on the patient structured data and the knowledge graph feature graph, and a corresponding query vector is constructed using the knowledge graph feature graph. After the feature extraction operation is completed by the feature extraction unit, the corresponding ESI feature data, CTAS feature data, MTS feature data, and ATS feature data are output; The corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data, and ATS enhanced feature data are constructed through the feature transfer layer, specifically including the following steps: Building a bidirectional gating unit between any two feature data among the ESI feature data, the CTAS feature data, the MTS feature data and the ATS feature data; For the ESI feature data, any one of the remaining three CTAS feature data, MTS feature data, and ATS feature data is selected as the target feature data. The following contents are executed in the bidirectional gating unit between the ESI feature data and the target feature data: the ESI feature data and the target feature data are concatenated and recorded as the overall feature data Z. The forget gate calculation is performed on the overall feature data: f t =σ(W f Z+b f ), where f t is the forget gate vector, σ is the sigmoid function, W f is the forget gate weight matrix, b f is the forget gate bias value, and the input gate is calculated for the overall feature data: i t =σ(W i Z+b i ), where i t is the input gate vector, W i is the input gate weight matrix, b i is the input gate bias value, and feature transfer is completed through the following formula: G=f t *R+i t *R, where G is the ESI feature transfer data corresponding to the ESI feature data, R is the target feature data, and * is a dot product operation; the three ESI feature transfer data corresponding to the ESI feature data are weighted summed and then residually connected with the ESI feature data to obtain the corresponding ESI enhanced feature data; the ESI feature data, CTAS feature data, MTS feature data, and ATS feature data are traversed to construct the corresponding ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data, and ATS enhanced feature data; The feature judgment analysis vector is constructed through the difference complementary enhancement layer, which specifically includes the following steps: The ESI enhanced feature data, CTAS enhanced feature data, MTS enhanced feature data and ATS enhanced feature data are respectively sent to the semantic mapping network for processing, and the corresponding ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector are output; The ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector are summed and averaged to obtain the average feature vector. The ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector are then differentially calculated with the average feature vector to construct the corresponding ESI differential feature vector, CTAS differential feature vector, MTS differential feature vector and ATS differential feature vector. The self-attention mechanism is performed on the ESI enhanced feature vector, CTAS enhanced feature vector, MTS enhanced feature vector and ATS enhanced feature vector, respectively. The corresponding query vector is constructed with the corresponding ESI differential feature vector, CTAS differential feature vector, MTS differential feature vector and ATS differential feature vector to obtain the ESI feature vector to be fused, CTAS feature vector to be fused, MTS feature vector to be fused and ATS feature vector to be fused. The ESI feature vector to be fused, CTAS feature vector to be fused, MTS feature vector to be fused and ATS feature vector to be fused are then concatenated according to channels and a convolution operation is performed to obtain the feature judgment analysis vector.

2. The intelligent outpatient triage method based on a large model according to claim 1, characterized in that: The triage priority is obtained through the weighted voting layer, which includes the following steps: According to the ESI score, CTAS score, MT score and ATS score, the decision vectors corresponding to the ESI score, CTAS score, MT score and ATS score are determined from the decision vector library; the corresponding decision vectors are added to the end of the ESI score, CTAS score, MT score and ATS score to construct the corresponding ESI decision analysis vector, CTAS decision analysis vector, MT decision analysis vector and ATS decision analysis vector, and then the ESI decision analysis vector, CTAS decision analysis vector, MT decision analysis vector and ATS decision analysis vector are sent to the multi-layer perceptron for processing, and the corresponding voting weights are output. The ESI score, CTAS score, MT score and ATS score are weightedly voted according to the corresponding voting weights to obtain the triage priority.

3. The intelligent outpatient triage method based on a large model according to claim 2, characterized in that: Training the outpatient triage model includes the following steps: Obtain several outpatient triage training samples, which include patient structured data. Label the outpatient triage training samples according to triage priority, and form an outpatient triage training set with all labeled outpatient triage training samples. Train the outpatient triage model with the outpatient triage training set. During training, use the labeled triage priority as the training target to determine whether the training conditions are met. If the training conditions are met, output the trained outpatient triage model; otherwise, continue to train the outpatient triage model with the outpatient triage training set.

4. An intelligent outpatient triage system based on a large model, characterized by: The system applies the large model-based intelligent outpatient triage method according to any one of claims 1 to 3, comprising: The data acquisition module is used to obtain the patient's basic information, chief complaint data and basic vital signs data, pre-process the patient's basic information, chief complaint data and basic vital signs data, and construct the corresponding patient structured data; The outpatient triage module is used to feed the patient's structured data into the outpatient triage model for processing and output the patient's corresponding triage priority.

Images