Intelligent emergency triage method and system based on knowledge prompts and expert knowledge supervision
By combining expert knowledge supervision and large language models, and utilizing structured preprocessing and knowledge graph enhancement, the problems of low efficiency and misjudgment in traditional emergency triage are solved, achieving efficient and interpretable intelligent triage.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUN YAT SEN UNIV
- Filing Date
- 2026-01-28
- Publication Date
- 2026-06-12
AI Technical Summary
Traditional emergency triage relies on human experience and large language models, which is inefficient, inconsistent, and susceptible to subjective factors. Furthermore, it does not make full use of structured vital sign data, leading to triage delays or misjudgments.
By combining expert knowledge supervision and a large language model, the system generates comprehensive prompts by preprocessing vital signs and patient complaints in a structured manner. It enhances professionalism by using knowledge graphs and aggregation analysis to generate triage suggestions that conform to clinical guidelines, and improves interpretability through closed-loop optimization.
It improves the accuracy and efficiency of emergency triage, ensures that triage results conform to clinical guidelines, enhances the professionalism and interpretability of model output, and reduces the risk of misjudgment by large language models.
Smart Images
Figure CN122201705A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart healthcare technology, and more specifically, to an intelligent emergency triage method and system based on knowledge prompts and expert knowledge supervision. Background Technology
[0002] As the frontline of hospital care for critically ill patients, the emergency department faces challenges such as high patient volume, complex conditions, and limited medical resources. Traditional emergency triage relies heavily on nurses' experience and subjective judgment, using triage tools such as the Emergency Severity Index (ESI) for assessment. However, this manual triage method suffers from low efficiency, poor consistency, and susceptibility to subjective factors, especially during surges in patient numbers, which can lead to triage delays or misjudgments, impacting treatment outcomes.
[0003] In recent years, with the development of artificial intelligence technology, especially the application of Large Language Models (LLM) in the medical field, new solutions have been provided for intelligent triage. However, relying solely on LLM for triage has the following problems: Insufficient professionalism: General-purpose LLMs lack professional knowledge in the medical field and may generate inaccurate or inconsistent recommendations with clinical guidelines. Poor interpretability: The model's decision-making process lacks transparency, making it difficult to gain the trust of medical staff. Inadequate data utilization: Relying solely on the patient's chief complaint text and failing to fully utilize structured vital sign data (such as heart rate, blood pressure, blood oxygen, etc.) leads to incomplete triage basis. To solve these problems, there is an urgent need for an intelligent triage method that integrates expert knowledge supervision and LLM reasoning, which can utilize the natural language understanding capabilities of LLMs and combine them with medical expert knowledge to improve the accuracy, efficiency, and credibility of triage. Summary of the Invention
[0004] To address the aforementioned technical issues, this invention proposes an intelligent emergency triage method and system based on knowledge prompts and expert knowledge supervision. This method dynamically integrates expert knowledge with patient information, guides a large language model to generate triage suggestions that conform to clinical guidelines, and provides judgment criteria. This improves triage efficiency while ensuring the scientific nature of medical decisions.
[0005] The first aspect of this invention provides an intelligent emergency triage method based on knowledge prompts and expert knowledge supervision, comprising the following steps: The patient's vital signs data and patient complaint text data are acquired, the vital signs data and patient complaint text data are preprocessed in a structured manner, and input into a preset prompt template to generate patient symptom information prompt content. After preprocessing the structured vital signs data, aggregation analysis is used to supplement and enhance expert knowledge. Expert knowledge conclusions are obtained by comparing the range of indicators, and expert knowledge information prompts are generated based on the judgment criteria and expert knowledge conclusions. The patient symptom information prompts and the expert knowledge information prompts are integrated, and a comprehensive patient prompt is generated according to a preset prompt template. This is then input into a large language model to generate a predicted triage result. The predicted triage result is compared with the actual doctor's triage information, and closed-loop optimization is performed based on the comparison results.
[0006] In this solution, the vital signs data and patient complaint text data are preprocessed in a structured manner, input into a preset prompt template, and the resulting patient symptom information prompts are generated, specifically as follows: The patient's basic physiological indicators, laboratory test results, and pain scores are obtained as vital sign data. The vital sign data are then standardized and missing values are filled in. Collect patient complaint text data, map the patient complaint text data to a standard medical terminology database, filter irrelevant descriptions, generate word vector sequences of patient complaint text data, and construct an entity recognition model using bidirectional long short-term memory neural network, dilated convolutional neural network and conditional random field; A bidirectional long short-term memory neural network is used to extract forward and reverse features from the word vector sequence to obtain the bidirectional contextual feature representation of each word in the word vector sequence. A dilated convolutional neural network is introduced to extract the deep feature representation of local keywords in the word vector sequence. The bidirectional contextual feature representation and deep feature representation are modeled using conditional random fields. Entity recognition results are obtained using BIO annotation. The entity recognition results are then used to generate standardized expressions according to a preset prompt template. Knowledge graph enhancement is applied to the standardized expressions to generate patient symptom information prompts.
[0007] In this solution, standardized expressions are enhanced with knowledge graphs to generate prompts for patient symptom information, specifically: Extract symptom terms from standardized expressions, access relevant medical knowledge graphs, locate nodes of symptom terms in the relevant medical knowledge graphs, perform adaptive neighbor expansion based on symptom term nodes, preserve the original relationships in the relevant medical knowledge graphs, and construct symptom subgraphs. The symptom term nodes are initialized, the degree centrality of neighboring nodes in the symptom subgraph is calculated, basic weights are defined based on the degree centrality and clinical prior knowledge, a graph attention network is used to learn the representation of the symptom subgraph, attention coefficients are calculated for each neighbor of the symptom term node, and attention weights are normalized. The attention weights are combined with the basic weights to aggregate neighbor features and update the representation of symptom term nodes. Through multi-layer graph attention layer iteration, high-order semantics are obtained to generate structured descriptions, which are then converted into clinically readable prompts.
[0008] In this solution, the preprocessed structured vital sign data will be enhanced with expert knowledge through aggregation analysis, specifically as follows: An expert knowledge base is established based on predefined clinical rules and thresholds. Structured preprocessed vital sign data and their time series are obtained. An independent trajectory is constructed for each vital sign field. A dynamic trajectory segment is established based on the independent trajectory using a sliding window mechanism. In the aggregation analysis, a stacking strategy with multiple similarity fusion is introduced to dynamically correct the judgment results of expert rules. For the dynamic trajectory segment, cosine similarity, Jaccard index, dynamic time warping and expert rule matching degree are used to calculate the similarity in the expert knowledge base space, screen similar emergency typical trajectory patterns, and output preliminary decisions independently for each similarity. Using a random forest as the meta-model, the system inputs each similarity score and a preliminary decision, learns the optimal weight combination, and obtains a comprehensive similarity by weighting and summing each similarity score using the optimal weight combination. Based on the comprehensive similarity, it triggers expert knowledge association, and obtains expert knowledge conclusions by comparing the index ranges.
[0009] In this solution, the patient symptom information prompts and the expert knowledge information prompts are combined according to a preset prompt template to generate comprehensive patient prompts, specifically as follows: Obtain patient symptom information prompts and expert knowledge information prompts, perform feature encoding on the patient symptom information prompts and expert knowledge information prompts, and generate corresponding symptom text features, symptom numerical features and knowledge features; Cross-attention calculation is performed on the symptom text features, symptom numerical features and knowledge features. A three-modal correlation matrix is established based on the cross-attention value to achieve adaptive feature alignment. The feature sequence after adaptive feature alignment is then enhanced temporally using a gated recurrent unit. An urgency weight is introduced based on the degree of abnormality of vital signs, a consistency weight is introduced based on the mutual evidence strength of multimodal features, and a timeliness weight is introduced based on the data collection time. A genetic algorithm is used to optimize the urgency weight, consistency weight, and timeliness weight to obtain the best weight combination. Dynamic weighted fusion is performed using the optimal weight combination, and the fused content is filtered using a preset prompt template. Information is then integrated according to the preset prompt template to generate comprehensive prompt content for the patient.
[0010] In this solution, a large language model is used to generate predicted triage results, specifically: The comprehensive prompts for patients are input into a pre-trained large language model. Danger signals are identified in the large language model. When danger alarm keywords are present, warning information is generated immediately. Otherwise, the comprehensive prompts are compared with the model knowledge base to calculate the typicality score of the symptom combination. Knowledge instances that meet the typicality score criteria are identified. Based on the acquired knowledge instances, multi-round question answering is activated to generate deep reasoning data. The deep reasoning data is imported into an uncertainty-based classifier to generate predictive triage results with uncertainty scores.
[0011] In this solution, the predicted triage results are compared with the actual doctor triage information, and closed-loop optimization is performed based on the comparison results, specifically as follows: The predicted triage results output by the large model are compared with the actual triage information of doctors in the expert knowledge base to verify the results and check the logic. If the predicted triage results output by the large model conflict with the clinical rules, the correction mechanism is automatically triggered and replaced with the preset conclusion of the expert knowledge base. Obtain the difference data between the predicted triage results and the actual doctor triage information, evaluate the difference level based on the difference data, preset the model update frequency and processing method for different difference levels, continuously monitor the difference data corresponding to the large model, and perform targeted optimization of the large model.
[0012] The second aspect of the present invention provides an intelligent emergency triage system based on knowledge prompts and expert knowledge supervision, realizing an intelligent emergency triage method based on knowledge prompts and expert knowledge supervision. The system includes a patient information input module, a patient symptom information generation module, a triage expert knowledge information generation module, a triage prediction information fusion module, a triage prediction module, a closed-loop feedback module, and a prompt template management module. The patient information input module acquires the patient's vital signs data and the patient's chief complaint text, and performs structured preprocessing on the vital signs data and the patient's chief complaint text. The patient symptom information generation module inputs the structured preprocessed vital sign data and the patient's chief complaint text data into a preset prompt template to generate patient symptom information prompt content. The triage expert knowledge information generation module uses aggregation analysis to supplement and enhance expert knowledge after the structured preprocessed vital sign data, obtains expert knowledge conclusions through indicator range comparison, and generates expert knowledge information prompts based on the judgment criteria and expert knowledge conclusions. The triage prediction information fusion module will combine the generated patient symptom information prompts and expert knowledge information prompts according to a preset prompt template to generate comprehensive patient prompts. The triage prediction module inputs the comprehensive patient prompts into the large language model to generate predicted triage results. The closed-loop feedback module compares the predicted triage results with the actual doctor triage information and performs closed-loop optimization based on the comparison results. The prompt template management module is responsible for the management and maintenance of prompt template information in the patient symptom information generation module, the triage expert knowledge information generation module, and the triage prediction information fusion module.
[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention combines vital sign data with patient complaint text to generate comprehensive and standardized "patient symptom information prompts"; it also generates interpretable "expert knowledge information prompts" through predefined medical rules and indicator ranges, enhancing the professionalism of the model output. Aggregate analysis ensures that the model output conforms to clinical guidelines, reducing the "illusion" risk of LLM (Laboratory Management Model). Dynamically integrating expert knowledge with patient information guides the large language model to generate triage suggestions that conform to clinical guidelines, while providing judgment criteria, improving the reliability and interpretability of the results. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.
[0015] Figure 1 A flowchart of an intelligent emergency triage method based on knowledge prompts and expert knowledge supervision is shown; Figure 2 The flowchart illustrates the process of using knowledge graphs to enhance the generation of patient symptom information prompts. Figure 3 A flowchart illustrating the use of aggregation analysis to enhance expert knowledge is shown. Figure 4 A block diagram of an intelligent emergency triage system based on knowledge prompts and expert knowledge supervision is shown. Detailed Implementation
[0016] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0018] Figure 1 A flowchart of an intelligent emergency triage method based on knowledge prompts and expert knowledge supervision is shown.
[0019] like Figure 1 As shown, this embodiment provides an intelligent emergency triage method based on knowledge prompts and expert knowledge supervision, including: S102, acquire the patient's vital signs data and patient's chief complaint text data, perform structured preprocessing on the vital signs data and patient's chief complaint text data, input them into a preset prompt template, and generate patient symptom information prompt content. S104: After the structured preprocessing of vital signs data, the data is enhanced with expert knowledge through aggregation analysis. Expert knowledge conclusions are obtained by comparing the range of indicators. Based on the judgment criteria and expert knowledge conclusions, expert knowledge information prompts are generated. S106, the patient symptom information prompts and the expert knowledge information prompts are integrated, and a comprehensive patient prompt is generated according to a preset prompt template. This is then input into a large language model to generate a predicted triage result. The predicted triage result is compared with the actual doctor's triage information, and closed-loop optimization is performed based on the comparison results.
[0020] It should be noted that the patient's basic physiological indicators, laboratory test results, and pain scores are collected as vital sign data. Basic physiological indicators include heart rate, blood pressure, blood oxygen saturation, body temperature, and respiratory rate; laboratory test results include blood glucose and blood gas analysis; and pain scores include NRS / VAS scores. The vital sign data are standardized and missing values are imputed. Patient complaint text data is collected, which may include symptom descriptions, medical history fragments, and triggering factors. This patient complaint text data is mapped to a standard medical terminology database, and irrelevant descriptions are filtered out to generate a word vector sequence for the patient complaint text data.
[0021] A bidirectional long short-term memory (LSTM) neural network, a dilated convolutional neural network (DCNN), and a conditional random field (CRF) are used to construct an entity recognition model to identify and label key entities (e.g., diseases, examinations, locations, symptoms) in medical texts. Adversarial training is introduced into the model's training to improve its robustness to input noise. Small perturbations are injected into word vectors during training to force the model to learn more stable feature representations. The LSTM is used to extract forward and reverse features from the word vector sequence. For example, given a patient's complaint of "chest pain for 2 hours, accompanied by shortness of breath," the forward-order LSTM encodes the text from left to right, capturing contextual dependencies, such as chest pain potentially being associated with myocardial infarction. The reverse-order LSTM encodes the text from right to left, supplementing subsequent information, such as shortness of breath potentially indicating pulmonary embolism. This yields bidirectional contextual feature representations for each word in the word vector sequence. A dilated convolutional neural network is introduced to extract deep feature representations of local keywords in the word vector sequence, such as chest pain radiating to the left shoulder. Dilated convolution expands the receptive field, capturing long-distance dependencies and avoiding the locality limitations of ordinary CNNs. Conditional random fields (CRF) are used to model the bidirectional contextual feature representation and deep feature representation into sequence labels to ensure label rationality. BIO annotation is used to obtain entity recognition results, such as "chest / B-symptom pain / I-symptom 2 / O hours / O, / O accompanied / O breathing / B-symptom difficulty / I-symptom". The entity recognition results are then used to generate standardized expressions according to a preset prompt template. Knowledge graph enhancement is applied to these standardized expressions to generate patient symptom information prompts.
[0022] Figure 2 The flowchart illustrates the process of using knowledge graphs to enhance the generation of patient symptom information prompts.
[0023] According to an embodiment of the present invention, knowledge graph enhancement is applied to standardized expressions to generate patient symptom information prompts, specifically as follows: S202, extract symptom terms from the standardized expressions, access the relevant medical knowledge graph, locate the nodes of the symptom terms in the relevant medical knowledge graph, perform adaptive neighbor expansion based on the symptom term nodes, retain the original relationships in the relevant medical knowledge graph, and construct a symptom subgraph. S204, initialize the representation of the symptom term node, calculate the degree centrality of the neighbor nodes in the symptom subgraph, define basic weights based on the degree centrality and clinical prior knowledge, use a graph attention network to learn the representation of the symptom subgraph, calculate the attention coefficient for each neighbor of the symptom term node, and normalize the attention weights. S206, using the attention weights combined with the basic weights to aggregate neighbor features, update the representation of symptom term nodes, and obtain high-order semantics to generate structured descriptions through multi-layer graph attention layer iterations, and convert the structured descriptions into clinically readable prompts.
[0024] It should be noted that the relevant medical knowledge graph includes nodes such as symptoms, diseases, examinations, and anatomical sites, as well as their relationships. The graph retrieves nodes corresponding to symptom terms, and if a term is ambiguous, context disambiguation is used to select the most relevant node. A symptom subgraph is constructed by adaptively extracting its k-hop neighbors, centered on the symptom term node. The breadth of entity neighbors retrieved from the relevant medical knowledge graph is dynamically selected based on the complexity of the patient's symptoms. The degree centrality of neighbor nodes in the symptom subgraph is calculated, where degree centrality is the number of direct connections between a node and other nodes, representing the importance of the node. Based on the degree centrality and clinical prior knowledge, basic weights are defined; for example, the weight of a possible cause is higher than that of accompanying symptoms. Preferably, the basic weights are dynamically adjusted using the co-occurrence frequency of relationships in the graph; for example, if the statistical probability of diagnosing myocardial infarction after chest pain is high, the weight is increased. By combining prior knowledge with a data-driven attention mechanism, the limitations of static rules are avoided. The attention weights of neighboring nodes are calculated in the graph attention network to reflect the contribution of neighbors to the central node. For example, the weight of myocardial infarction to chest pain is higher than that of gastritis. The neighbor features are weighted and aggregated to update the representation of the central symptom node. Through multi-layer graph attention network iteration, higher-order semantics are captured, such as the implicit association between chest pain → myocardial infarction → elevated troponin.
[0025] Figure 3 A flowchart is shown for enhancing expert knowledge using aggregation analysis.
[0026] According to an embodiment of the present invention, the structured preprocessed vital sign data is enhanced with expert knowledge through aggregation analysis, specifically as follows: S302, establish an expert knowledge base based on predefined clinical rules and thresholds, acquire structured preprocessed vital sign data and their time series, construct an independent trajectory for each vital sign field, and establish a dynamic trajectory segment based on the independent trajectory using a sliding window mechanism; S304, In the aggregation analysis, a stacking strategy with multiple similarity fusion is introduced to dynamically correct the judgment results of expert rules. For the dynamic trajectory segment, cosine similarity, Jaccard index, dynamic time warping and expert rule matching degree are used to calculate the similarity in the expert knowledge base space, screen similar emergency typical trajectory patterns, and output preliminary decisions independently for each similarity. S306 uses a random forest as a meta-model, inputs each similarity score and preliminary decision, learns the optimal weight combination, obtains the comprehensive similarity by weighted summation of each similarity score through the optimal weight combination, triggers expert knowledge association based on the comprehensive similarity, and obtains expert knowledge conclusions by comparing the index range.
[0027] It should be noted that in trajectory modeling, baseline patient data (such as a history of hypertension) is added as a reference trajectory. Expert knowledge conclusions are obtained through processes such as trajectory modeling, pattern matching, and knowledge association. The trajectory specifically refers to the pattern of vital signs changing over time. Dynamic aggregation analysis based on a multi-similarity fusion Stacking strategy expands the capabilities of traditional aggregation analysis. Through the multi-similarity fusion Stacking strategy, the judgment results of expert rules are dynamically corrected, overcoming the limitations of single rules in complex scenarios. Cosine similarity is used to calculate numerical similarity, suitable for vectorized features, such as the combined state of multiple indicators, calculating the vector angle between the current vital signs and typical pathological states; Jaccard index is used to calculate set similarity, suitable for discretized indicators, such as whether there is fever or confusion, comparing the intersection ratio between the patient's vital signs and rule conditions; dynamic time warping is used to calculate temporal similarity, suitable for time series data, such as continuous blood pressure fluctuation trends, aligning the patient's time axis with typical pathological patterns and calculating the minimum path distance; expert rule matching degree is used to calculate clinical similarity, determining whether it conforms to a specific disease evolution pattern. Each similarity score independently outputs a preliminary judgment. A random forest is used as the meta-model, inputting multiple similarity scores and original rule conclusions to learn the optimal weight combination. The meta-model is trained using similarity features from historical emergency cases and final diagnosis labels. For trajectory segments deviating from expectations, adjustments are made based on reasonable fluctuations in similar cases, and clinical explanatory labels are added, such as interpreting a persistent drop in blood pressure accompanied by an increased heart rate as a precursor to shock. When an unknown pattern appears, an expert review process is automatically triggered, the new trajectory is stored in the learning library, and the similarity weights are updated through online learning.
[0028] It should be noted that the process involves acquiring patient symptom information prompts and expert knowledge information prompts, encoding these prompts into corresponding textual, numerical, and knowledge features. Cross-attention calculations are then performed on these features: text-numerical attention is calculated to determine the correlation between the symptom description and the vital signs data; text-knowledge attention is obtained by assessing the matching degree between the symptoms and expert knowledge; and numerical-knowledge attention is obtained by verifying the support of the vital signs data for the knowledge conclusions. A three-modal correlation matrix is established based on the cross-attention values to achieve adaptive feature alignment. Mismatched features are corrected; for example, if the text describes chest pain but the electrocardiogram is normal, the weight of the text feature is reduced; conversely, if blood pressure drops suddenly but shock symptoms are not mentioned, the weight of the knowledge feature is increased. The adaptively aligned feature sequences are augmented temporally using gated recurrent units to capture dynamic patterns such as progressive blood pressure decline. An urgency weight is introduced based on the severity of vital sign abnormalities: urgency weight = number of abnormal indicators / total number of indicators × severity coefficient. A consistency weight is introduced based on the cross-verification strength of multimodal features: consistency weight = sum of cross-modal attention scores / 3. A timeliness weight is introduced based on the data acquisition time: timeliness weight = ... , The data age and timeliness weights emphasize the importance of the latest data. A genetic algorithm is used to optimize the urgency weight, consistency weight, and timeliness weight. The optimal weight combination is obtained through tournament selection, arithmetic crossover, and Gaussian perturbation mutation. The optimal weight combination is used for dynamic weighted fusion. The fused content is used to filter the preset prompt template. The information is integrated according to the preset prompt template to generate comprehensive prompt content for patients.
[0029] It should be noted that the comprehensive patient prompts are input into a pre-trained large language model. The model identifies danger signals, and when danger alert keywords are present, such as cardiac arrest or suffocation, an immediate warning is generated. Otherwise, the comprehensive prompts are compared with the model's knowledge base to calculate the typicality score of the symptom combination. Knowledge instances meeting the typicality score criteria are identified. The symptom typicality score is calculated as follows: (TF-IDF value of each symptom × sum of frequencies in the knowledge graph) / (square root of symptom combination complexity). The TF-IDF value measures the relative importance of the symptom in the current chief complaint, the knowledge graph frequency reflects the typical association strength of the symptom in the medical knowledge base, and complexity is defined as the number of atypical symptoms. Atypical symptoms must meet the following criteria: the number of associated diseases in the knowledge graph is less than 3, or the simultaneous occurrence of symptoms does not conform to common clinical syndrome patterns or contains contradictory signs. Based on the acquired knowledge instances, multi-turn question answering is activated to generate deep reasoning data. This deep reasoning data is then imported into an uncertainty-based classifier to generate predicted triage results with uncertainty scores. The preferred uncertainty scores are obtained through n sampling iterations using a Monte Carlo algorithm.
[0030] It should be noted that the predicted triage results output by the large model are structured and aligned with the actual doctor triage information in the expert knowledge base. Comparisons are performed on dimensions such as triage level difference, examination matching degree, and diagnostic overlap. Result verification and logical checks are conducted. If the predicted triage results output by the large model conflict with clinical rules, a correction mechanism is automatically triggered, replacing them with the preset conclusions from the expert knowledge base. The difference data between the predicted triage results and the actual doctor triage information is obtained, and the difference level is evaluated based on this difference data. The model update frequency and processing method are preset for different difference levels. For example, when a level one difference (level deviation ≥ 2) occurs, a correction mechanism is immediately triggered. The model is frozen and reviewed by experts, and updated in real time. When a secondary discrepancy (missed key checks) occurs, it is added to the reinforcement learning dataset, and the model is updated daily through batch processing. When a tertiary discrepancy (differences in terminology) occurs, it is corrected through knowledge graph terminology mapping, and the model is updated weekly. The discrepancy data corresponding to the large model is continuously monitored, and the performance of the large model is visualized by combining visualization metrics. The preferred visualization metrics include: grade consistency rate, check recommendation recall rate, number of missed diagnoses of high-risk cases, etc. The continuous and safe evolution of the large language model in the emergency scenario is achieved by using dynamic adjustment of template weights and knowledge base enhancement.
[0031] Figure 4 A block diagram of an intelligent emergency triage system based on knowledge prompts and expert knowledge supervision is shown.
[0032] The second embodiment of the present invention provides an intelligent emergency triage system based on knowledge prompts and expert knowledge supervision, including: a patient information input module, a patient symptom information generation module, a triage expert knowledge information generation module, a triage prediction information fusion module, a triage prediction module, a closed-loop feedback module, and a prompt template management module; The patient information input module acquires the patient's vital signs data and the patient's chief complaint text, and performs structured preprocessing on the vital signs data and the patient's chief complaint text. The patient symptom information generation module inputs the structured preprocessed vital sign data and the patient's chief complaint text data into a preset prompt template to generate patient symptom information prompt content. The triage expert knowledge information generation module uses aggregation analysis to supplement and enhance expert knowledge after the structured preprocessed vital sign data, obtains expert knowledge conclusions through indicator range comparison, and generates expert knowledge information prompts based on the judgment criteria and expert knowledge conclusions. The triage prediction information fusion module will combine the generated patient symptom information prompts and expert knowledge information prompts according to a preset prompt template to generate comprehensive patient prompts. The triage prediction module inputs the comprehensive patient prompts into the large language model to generate predicted triage results. The closed-loop feedback module compares the predicted triage results with the actual doctor triage information and performs closed-loop optimization based on the comparison results. The prompt template management module is responsible for the management and maintenance of prompt template information in the patient symptom information generation module, the triage expert knowledge information generation module, and the triage prediction information fusion module.
[0033] A third aspect of the present invention provides a computer-readable storage medium comprising an intelligent emergency triage method program based on knowledge prompts and expert knowledge supervision, wherein when the intelligent emergency triage method program based on knowledge prompts and expert knowledge supervision is executed by a processor, it implements the steps of the intelligent emergency triage method based on knowledge prompts and expert knowledge supervision.
[0034] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be indirect coupling or communication connection through some interfaces, devices, or modules, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional modules can be integrated into one processing module, or each module can be a separate module, or two or more modules can be integrated into one module; the integrated modules can be implemented in hardware or in the form of hardware plus software functional modules.
[0035] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. An intelligent emergency triage method based on knowledge prompts and expert knowledge supervision, characterized in that, Includes the following steps: The patient's vital signs data and patient complaint text data are acquired, the vital signs data and patient complaint text data are preprocessed in a structured manner, and input into a preset prompt template to generate patient symptom information prompt content. After preprocessing the structured vital signs data, aggregation analysis is used to supplement and enhance expert knowledge. Expert knowledge conclusions are obtained by comparing the range of indicators, and expert knowledge information prompts are generated based on the judgment criteria and expert knowledge conclusions. The patient symptom information prompts and the expert knowledge information prompts are integrated, and a comprehensive patient prompt is generated according to a preset prompt template. This is then input into a large language model to generate a predicted triage result. The predicted triage result is compared with the actual doctor's triage information, and closed-loop optimization is performed based on the comparison results.
2. The intelligent emergency triage method based on knowledge prompts and expert knowledge supervision according to claim 1, characterized in that, The vital signs data and patient complaint text data are preprocessed in a structured manner, input into a preset prompt template, and the resulting patient symptom information prompts are generated, specifically as follows: The patient's basic physiological indicators, laboratory test results, and pain scores are obtained as vital sign data. The vital sign data are then standardized and missing values are filled in. Collect patient complaint text data, map the patient complaint text data to a standard medical terminology database, filter irrelevant descriptions, generate word vector sequences of patient complaint text data, and construct an entity recognition model using bidirectional long short-term memory neural network, dilated convolutional neural network and conditional random field; A bidirectional long short-term memory neural network is used to extract forward and reverse features from the word vector sequence to obtain the bidirectional contextual feature representation of each word in the word vector sequence. A dilated convolutional neural network is introduced to extract the deep feature representation of local keywords in the word vector sequence. The bidirectional contextual feature representation and deep feature representation are modeled using conditional random fields. Entity recognition results are obtained using BIO annotation. The entity recognition results are then used to generate standardized expressions according to a preset prompt template. Knowledge graph enhancement is applied to the standardized expressions to generate patient symptom information prompts.
3. The intelligent emergency triage method based on knowledge prompts and expert knowledge supervision according to claim 2, characterized in that, The standardized expressions are augmented with a knowledge graph to generate prompts for patient symptom information, specifically: Extract symptom terms from standardized expressions, access relevant medical knowledge graphs, locate nodes of symptom terms in the relevant medical knowledge graphs, perform adaptive neighbor expansion based on symptom term nodes, preserve the original relationships in the relevant medical knowledge graphs, and construct symptom subgraphs. The symptom term nodes are initialized, the degree centrality of neighboring nodes in the symptom subgraph is calculated, basic weights are defined based on the degree centrality and clinical prior knowledge, a graph attention network is used to learn the representation of the symptom subgraph, attention coefficients are calculated for each neighbor of the symptom term node, and attention weights are normalized. The attention weights are combined with the basic weights to aggregate neighbor features and update the representation of symptom term nodes. Through multi-layer graph attention layer iteration, high-order semantics are obtained to generate structured descriptions, which are then converted into clinically readable prompts.
4. The intelligent emergency triage method based on knowledge prompts and expert knowledge supervision according to claim 1, characterized in that, The preprocessed structured vital sign data were then augmented with expert knowledge using aggregation analysis, specifically: An expert knowledge base is established based on predefined clinical rules and thresholds. Structured preprocessed vital sign data and their time series are obtained. An independent trajectory is constructed for each vital sign field. A dynamic trajectory segment is established based on the independent trajectory using a sliding window mechanism. In the aggregation analysis, a stacking strategy with multiple similarity fusion is introduced to dynamically correct the judgment results of expert rules. For the dynamic trajectory segment, cosine similarity, Jaccard index, dynamic time warping and expert rule matching degree are used to calculate the similarity in the expert knowledge base space, screen similar emergency typical trajectory patterns, and output preliminary decisions independently for each similarity. Using a random forest as the meta-model, the system inputs each similarity score and a preliminary decision, learns the optimal weight combination, and obtains a comprehensive similarity by weighting and summing each similarity score using the optimal weight combination. Based on the comprehensive similarity, it triggers expert knowledge association, and obtains expert knowledge conclusions by comparing the index ranges.
5. The intelligent emergency triage method based on knowledge prompts and expert knowledge supervision according to claim 1, characterized in that, The patient symptom information and expert knowledge information are combined according to a preset prompt template to generate a comprehensive patient prompt, specifically: Obtain patient symptom information prompts and expert knowledge information prompts, perform feature encoding on the patient symptom information prompts and expert knowledge information prompts, and generate corresponding symptom text features, symptom numerical features and knowledge features; Cross-attention calculation is performed on the symptom text features, symptom numerical features and knowledge features. A three-modal correlation matrix is established based on the cross-attention value to achieve adaptive feature alignment. The feature sequence after adaptive feature alignment is then enhanced temporally using a gated recurrent unit. An urgency weight is introduced based on the degree of abnormality of vital signs, a consistency weight is introduced based on the mutual evidence strength of multimodal features, and a timeliness weight is introduced based on the data collection time. A genetic algorithm is used to optimize the urgency weight, consistency weight, and timeliness weight to obtain the best weight combination. Dynamic weighted fusion is performed using the optimal weight combination, and the fused content is filtered using a preset prompt template. Information is then integrated according to the preset prompt template to generate comprehensive prompt content for the patient.
6. The intelligent emergency triage method based on knowledge prompts and expert knowledge supervision according to claim 1, characterized in that, The predicted triage results are generated using a large language model, specifically as follows: The comprehensive prompts for patients are input into a pre-trained large language model. Danger signals are identified in the large language model. When danger alarm keywords are present, warning information is generated immediately. Otherwise, the comprehensive prompts are compared with the model knowledge base to calculate the typicality score of the symptom combination. Knowledge instances that meet the typicality score criteria are identified. Based on the acquired knowledge instances, multi-round question answering is activated to generate deep reasoning data. The deep reasoning data is imported into an uncertainty-based classifier to generate predictive triage results with uncertainty scores.
7. The intelligent emergency triage method based on knowledge prompts and expert knowledge supervision according to claim 1, characterized in that, The predicted triage results are compared with the actual doctor triage information, and closed-loop optimization is performed based on the comparison results, specifically as follows: The predicted triage results output by the large model are compared with the actual triage information of doctors in the expert knowledge base to verify the results and check the logic. If the predicted triage results output by the large model conflict with the clinical rules, the correction mechanism is automatically triggered and replaced with the preset conclusion of the expert knowledge base. Obtain the difference data between the predicted triage results and the actual doctor triage information, evaluate the difference level based on the difference data, preset the model update frequency and processing method for different difference levels, continuously monitor the difference data corresponding to the large model, and perform targeted optimization of the large model.
8. An intelligent emergency triage system based on knowledge prompts and expert knowledge supervision, characterized in that, The system implements the intelligent emergency triage method based on knowledge prompts and expert knowledge supervision as described in any one of claims 1-7, comprising: a patient information input module, a patient symptom information generation module, a triage expert knowledge information generation module, a triage prediction information fusion module, a triage prediction module, a closed-loop feedback module, and a prompt template management module; The patient information input module acquires the patient's vital signs data and the patient's chief complaint text, and performs structured preprocessing on the vital signs data and the patient's chief complaint text. The patient symptom information generation module inputs the structured preprocessed vital sign data and the patient's chief complaint text data into a preset prompt template to generate patient symptom information prompt content. The triage expert knowledge information generation module uses aggregation analysis to supplement and enhance expert knowledge after the structured preprocessed vital sign data, obtains expert knowledge conclusions through indicator range comparison, and generates expert knowledge information prompts based on the judgment criteria and expert knowledge conclusions. The triage prediction information fusion module will combine the generated patient symptom information prompts and expert knowledge information prompts according to a preset prompt template to generate comprehensive patient prompts. The triage prediction module inputs the comprehensive patient prompts into the large language model to generate predicted triage results. The closed-loop feedback module compares the predicted triage results with the actual doctor triage information and performs closed-loop optimization based on the comparison results. The prompt template management module is responsible for the management and maintenance of prompt template information in the patient symptom information generation module, the triage expert knowledge information generation module, and the triage prediction information fusion module.