An acute abdomen triage decision support method and system
By combining a multi-task collaborative prediction model with an evidence-based medicine knowledge base, the problems of model opacity and decision conflict in the triage of acute abdominal pain are solved, and efficient and interpretable triage decision support is achieved.
Patent Information
- Application Number
- CN202611225635.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-06
AI Technical Summary
In existing technologies, the decision-making process of machine learning-based clinical prediction models in the triage of acute abdominal pain is not transparent, resulting in low trust among clinical users. Conflicting prediction results from multiple models also interfere with the accuracy and efficiency of triage.
A multi-task collaborative prediction model is adopted, combined with an evidence-based medicine knowledge base, and feature vectors are fused by a shared encoder and expert network weights are dynamically allocated to generate an interpretable triage decision support method, including high-risk identification and specialist triage recommendation.
It improves the clinical credibility and interpretability of prediction results, resolves multi-model decision conflicts, and achieves efficient and transparent triage decision support.
Smart Images
Figure CN123053451A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical information technology, specifically relating to a triage decision support method and system for acute abdominal pain. Background Technology
[0002] Acute abdomen is one of the most common reasons for visits to the emergency department. Its causes are complex and its condition can change rapidly, placing extremely high demands on the accuracy and timeliness of emergency triage. Triage nurses need to complete the identification of high-risk patients and the determination of their appropriate specialty within 3-5 minutes.
[0003] In existing technologies, predictive models are used to assist in triage decisions to improve efficiency and accuracy. However, while existing machine learning-based clinical prediction models demonstrate good statistical performance, their decision-making process remains opaque to clinical users. When triage nurses receive high-risk warnings or departmental recommendations, they cannot understand why the model made that prediction or whether the basis for the prediction is reliable, severely limiting clinical acceptance and trust.
[0004] Furthermore, when multiple prediction models are used for prediction, if different models give contradictory prediction results, it may affect or even interfere with the judgment of triage nurses, thereby affecting the accuracy and efficiency of triage. Summary of the Invention
[0005] The purpose of this invention is to propose a triage decision support method and system for acute abdominal pain, in order to solve the problems in the prior art.
[0006] Therefore, the present invention provides a triage decision support method for acute abdominal pain, comprising: Obtain patients' clinical data; The clinical data is preprocessed to construct a standardized feature vector; The standardized feature vector is input into multiple basic prediction models for parallel prediction to generate a basic prediction probability vector. Based on the clinical data, evidence-based knowledge fragments related to the clinical data are retrieved from the evidence-based medicine knowledge base, and structured knowledge feature vectors are extracted. The standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector are input into the multi-task collaborative prediction model, fused into an enhanced feature vector by a shared encoder, and knowledge-enhanced prediction is performed by dynamically allocating expert network weights through a gating network. The prediction results are output, including the output high-risk probability, subject-specific recommendation probability distribution, and prediction confidence. The prediction results of the multi-task collaborative prediction model are verified for consistency with the evidence-based knowledge fragments, and a consistency verification result is generated. Based on the prediction results of the multi-task collaborative prediction model, the evidence-based knowledge fragments, and the consistency verification results, a personalized clinical decision interpretation is generated. Based on the high-risk probability and the sub-specialty recommendation probability distribution, a triage decision prompt signal of the corresponding level is triggered, and the personalized clinical decision interpretation and evidence-based decision basis are visualized.
[0007] In some embodiments, the step of inputting the standardized feature vector into multiple basic prediction models for parallel prediction to generate multiple basic prediction modules in the basic prediction probability vector includes: A high-risk early warning model group, comprising three high-risk early warning models, each of which outputs a corresponding high-risk probability vector. as well as The subject-specific model group includes three subject-specific models, each of which is used to output a corresponding departmental affiliation probability vector.
[0008] In some embodiments, the step of retrieving evidence-based knowledge fragments related to the clinical data from an evidence-based medicine knowledge base and extracting structured knowledge feature vectors based on the clinical data includes: Key clinical features are extracted from the clinical data, and natural language queries are automatically generated based on the structured templates corresponding to the key clinical features. The natural language query is encoded into a query vector. The semantic similarity between the query vector and each knowledge unit vector is calculated in the evidence-based medicine knowledge base. The results are sorted in descending order of similarity and the search results are output. The knowledge unit vector is a high-dimensional vector converted from the knowledge text in the evidence-based medicine knowledge base. Based on the search results, select the most relevant evidence-based knowledge fragments; Structured knowledge features are extracted from the multiple evidence-based knowledge fragments to construct the structured knowledge feature vector; The structured knowledge feature vector includes the risk level recommended by the guidelines, the departments recommended by the guidelines, and the matching degree of key risk indicators.
[0009] In some embodiments, the standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector are input into a multi-task collaborative prediction model, fused into an enhanced feature vector by a shared encoder, and then used to dynamically allocate expert network weights through a gating network for knowledge-enhanced prediction. The prediction result is output, including: The standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector are mapped into a 512-dimensional enhanced feature vector through a deep neural network; Based on the enhanced feature vector, the activation weights of multiple expert networks are calculated, and the two expert networks with the highest weights are selected for activation. The activated expert network performs knowledge-enhanced predictions and outputs the prediction results.
[0010] In some embodiments, the expert network includes high-risk early warning experts, specialist triage experts, and comprehensive decision-making experts; When performing a high-risk early warning task, activate the high-risk early warning expert and the comprehensive decision-making expert; When performing a specialist admission task, the specialist admission expert and the comprehensive decision-making expert are activated; When a complex case scenario is identified, the three expert networks mentioned above are activated to make a joint decision.
[0011] In some embodiments, the consistency verification of the prediction results of the multi-task collaborative prediction model with the evidence-based knowledge fragment, and the generation of consistency verification results, includes: A hybrid scoring method combining keyword matching and semantic similarity was used to calculate the consistency score; The consistency level is determined based on the consistency score. When the consistency level is completely consistent, proceed directly to the step of generating personalized clinical decision interpretation; When the consistency level is partially consistent, the steps of generating personalized clinical decision interpretation are performed after marking the differences. When the consistency level is inconsistent, a manual review warning is triggered, and the process proceeds to the step of generating personalized clinical decision interpretation.
[0012] In some embodiments, generating personalized clinical decision interpretations based on the prediction results of the multi-task collaborative prediction model, the evidence-based knowledge fragments, and the consistency verification results includes: The structured instruction text is constructed based on the prediction results, the evidence-based knowledge fragments, and the consistency verification results. Based on a large language model, a three-part personalized clinical decision interpretation is generated from the structured instruction text.
[0013] In some embodiments, the visualization of the personalized clinical decision interpretation and evidence-based decision-making basis includes: High-risk levels are indicated by red circular symbols combined with high-risk text labels. The recommendation probability of each specialty is displayed using a horizontal progress bar, and the length of the horizontal progress bar is proportional to the recommendation probability. The predicted confidence level is displayed using text labels; The evidence-based decision-making basis is presented in structured text, including key excerpts from clinical guidelines retrieved by RAG and sources of evidence.
[0014] In some embodiments, the preprocessing of the clinical data to construct a standardized feature vector includes: The clinical data was cleaned to remove outliers; Missing values were processed from the cleaned clinical data; The clinical data after missing values were processed were coded into variables to construct the standardized feature vector.
[0015] On the other hand, a triage decision support system for acute abdominal pain is also provided, including: The patient information collection module, configured in the data layer, is used to acquire patients' clinical data. The preprocessing module, configured in the information layer, is used to preprocess the clinical data and construct standardized feature vectors; The basic model prediction module, configured in the information layer, is used to input the standardized feature vector into multiple basic prediction models for parallel prediction and generate a basic prediction probability vector. The knowledge retrieval module, configured in the knowledge layer, is used to retrieve evidence-based knowledge fragments related to the clinical data from the evidence-based medicine knowledge base, and extract structured knowledge feature vectors based on the clinical data. A multi-task collaborative prediction module, configured in the knowledge layer, is used to input the standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector into the multi-task collaborative prediction model, fuse them into a unified high-dimensional feature representation through a shared encoder, and perform knowledge-enhanced prediction by dynamically allocating expert network weights through a gating network, and output prediction results. The prediction results include high-risk probability, subject-specific recommendation probability distribution, and prediction confidence. A consistency verification module, configured in the knowledge layer, is used to verify the consistency between the prediction results of the multi-task collaborative prediction module and the evidence-based knowledge fragments retrieved by the knowledge retrieval module, and generate a consistency verification result. The decision interpretation generation module, configured in the intelligence layer, is used to generate personalized clinical decision interpretations based on the prediction results of the multi-task collaborative prediction module, the evidence-based knowledge fragments retrieved by the knowledge retrieval module, and the consistency verification results. The early warning and visualization module, configured in the intelligent layer, is used to trigger triage decision prompt signals of corresponding levels based on the high-risk probability and the departmental recommendation probability distribution, and to visualize the personalized clinical decision interpretation and evidence-based decision basis.
[0016] Beneficial effects: 1. This invention uses the structured knowledge features retrieved by RAG as one of the input features for MoE prediction, thereby achieving a paradigm shift from pure data-driven prediction to evidence-based medical knowledge-enhanced reasoning, which significantly improves the clinical credibility and interpretability of the prediction results.
[0017] 2. This invention adopts the MoE architecture to integrate high-risk early warning and specialist triage into a shared expert network system, which solves the decision conflict problem caused by the independent operation of multiple models in the prior art, and realizes information sharing and collaborative optimization of high-risk identification and specialist triage recommendation.
[0018] 3. This invention adopts a two-layer output structure of rapid prediction and deep interpretation, which not only meets the rapid decision-making needs of emergency scenarios with high timeliness, but also provides a deep evidence-based analysis approach for complex cases, realizing progressive decision support through human-machine collaboration.
[0019] 4. This invention uses a hybrid scoring method that combines keyword matching and semantic similarity to calculate consistency scores, and performs three-level judgment and differentiation processing based on the scores. This is equivalent to introducing a two-person verification process in the clinical decision-making process, effectively covering two types of system failure modes: basic model prediction errors and knowledge retrieval biases, and significantly enhancing the safety and reliability of decision recommendations. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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 based on these drawings without creative effort.
[0021] Figure 1 A flowchart for decision support methods for triage in acute abdominal pain.
[0022] Figure 2 This is a flowchart for preprocessing clinical data.
[0023] Figure 3 A flowchart for knowledge retrieval for RAG.
[0024] Figure 4 A flowchart for knowledge-enhanced prediction for multiple expert networks.
[0025] Figure 5 The flowchart for performing consistency verification.
[0026] Figure 6 This is an architecture diagram of a triage decision support system for acute abdominal pain.
[0027] Figure 7 This is a schematic diagram of the system interface. Detailed Implementation
[0028] The invention will be more readily understood by referring to the following detailed description of preferred embodiments and included examples. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. In case of conflict, the definitions in this specification shall prevail.
[0029] like Figures 1-5 as well as Figure 7 As shown, a triage decision support method for acute abdominal pain includes: S1, Obtain patient's clinical data; Specifically, when a patient with acute abdominal pain arrives at the emergency triage station, the triage nurse creates a medical record for the patient in the hospital's emergency information system. After the nurse enters their chief complaint of abdominal pain, the system automatically triggers the acute abdominal pain intelligent decision support module, displaying a structured "Acute Abdominal Pain Patient Triage Questionnaire." The triage nurse then asks questions and enters the patient's standardized clinical data item by item.
[0030] In this embodiment, the clinical data collection is based on the key decision-making element system for triage of acute abdominal pain constructed in previous research, which includes structured data with 6 dimensions and 32 items. Specifically, it includes: General feature collection is used to collect information such as age, gender, admission method, past medical history, reproductive history, menopausal history, abdominal surgery history, and allergy history. Vital signs collection is used to collect body temperature, heart rate, systolic blood pressure, diastolic blood pressure, and blood oxygen saturation, and automatically calculate the MEWS score; Abdominal pain feature collection is used to collect information such as the location of abdominal pain, the nature of the pain, the triggers of the pain, the duration of the pain, the NRS pain score, and accompanying symptoms. Abdominal physical examination is used to collect information on tenderness, rebound tenderness, muscle guarding, the triad of peritonitis, scleral icterus, renal percussion tenderness, Murphy's sign, etc. Comprehensive scoring is used to collect MEWS score, NRS pain score, shock index, etc. Emergency outcome data will be used for subsequent model validation and effectiveness evaluation.
[0031] This step ensures the standardization and completeness of data collection by combining standardized options with numerical input. The system automatically reads the entered vital signs and calculates the MEWS score, eliminating the need for re-entry. Simultaneously, through interface integration with the Hospital Information System (HIS), it supports tracing and supplementing missing data from the electronic medical record system using the patient ID number.
[0032] S2, preprocess the clinical data to construct a standardized feature vector; Specifically, after collecting the above-mentioned clinical data, the clinical data undergoes three-step preprocessing to ensure the authenticity and integrity of the data, which facilitates subsequent model calculations.
[0033] In this embodiment of the application, data preprocessing includes data cleaning, missing value handling, and variable encoding.
[0034] Data cleaning includes excluding duplicate and outlier records, and removing records that do not conform to clinical practice. For example, outliers such as blood oxygen saturation greater than 100% or less than 0 are excluded, and logical consistency checks are performed on the values of each variable to ensure that the samples included in the analysis are true, accurate, and free of redundancy.
[0035] Missing value handling included imputation using the MICE (Multiple Imputation, Components, and Interpolation) algorithm based on random forests. Specifically, the `mice` package in R was used to perform multiple imputation, with 10 imputation iterations generated to produce 10 complete imputed datasets for each missing value. The imputation model used the observed values of each variable as the dependent variable and all other variables as independent variables for prediction and imputation. After imputation, the modeling results of the 10 imputed datasets were merged to obtain the final estimated model parameters.
[0036] Variable encoding involves converting continuous variables into categorical variables based on the optimal cutoff value. Binary variables are assigned values of 1 / 0 for "yes / no" or "present / absent," while multi-category variables are handled using dummy variables. For example, gender is set to "male = 1, female = 0"; abdominal pain location is classified using the abdominal nine-point system, converting it into nine independent binary dummy variables. Each variable is assigned "yes = 1 / no = 0" to indicate whether the patient belongs to the corresponding region, thus avoiding the model from incorrectly introducing ordinal relationships between categories.
[0037] After the above preprocessing, a 27-dimensional standardized feature vector is finally constructed, which serves as the input for subsequent basic model prediction and RAG knowledge retrieval.
[0038] S3, the standardized feature vector is input into multiple basic prediction models for parallel prediction to generate a basic prediction probability vector, wherein the multiple basic prediction modules include: A high-risk early warning model group, comprising three high-risk early warning models, each of which outputs a corresponding high-risk probability vector. as well as The subject-specific model group includes three subject-specific models, each of which is used to output a corresponding departmental affiliation probability vector.
[0039] Specifically, the 27-dimensional standardized feature vector is simultaneously input into six pre-trained base prediction models for independent parallel prediction. The six models are a high-risk early warning Top 3 model and a subject-specific Top 3 model.
[0040] The top three high-risk early warning models include the XGBoost model, the LightGBM model, and the Random Forest model, which output the probability P that a patient is high-risk. 高危 =[p1,p2,p3], where p1 is the high-risk probability of the XGBoost model, p2 is the high-risk probability of the LightGBM model, and p3 is the high-risk probability of the Random Forest model.
[0041] The top 3 departmental classification models include the XGBoost model, the LightGBM model, and the Random Forest model, which output the probability P of a patient belonging to each department. 科室 =[p4,p5,p6], which correspond to the predicted probability distributions of the three optimal models for each department (internal medicine / surgery / gynecology).
[0042] The six models mentioned above are all optimal models validated by a multi-center prospective cohort in previous studies. Optimal hyperparameters were determined using five-fold cross-validation and grid search. The six-dimensional basic prediction probability vectors constitute new feature vectors, which are passed to the MoE module of the knowledge layer as one of the input features of the expert network. The basic model predictions are fully automated and deterministic; the same input always produces the same output, ensuring the reproducibility of the predictions.
[0043] S4, based on the clinical data, retrieve evidence-based knowledge fragments related to the clinical data from the evidence-based medicine knowledge base, and extract structured knowledge feature vectors, including: Key clinical features are extracted from the clinical data, and natural language queries are automatically generated based on the structured templates corresponding to the key clinical features. The natural language query is encoded into a query vector. The semantic similarity between the query vector and each knowledge unit vector is calculated in the evidence-based medicine knowledge base. The results are sorted in descending order of similarity and the search results are output. The knowledge unit vector is a high-dimensional vector converted from the knowledge text in the evidence-based medicine knowledge base. Based on the search results, select the most relevant evidence-based knowledge fragments; Structured knowledge features are extracted from the multiple evidence-based knowledge fragments to construct the structured knowledge feature vector; The structured knowledge feature vector includes the risk level recommended by the guidelines, the departments recommended by the guidelines, and the matching degree of key risk indicators.
[0044] Specifically, while the base model makes predictions in parallel, the RAG (Retrieval Enhanced Generation) module independently initiates the knowledge retrieval process. This retrieval process is completely independent of the base model's predictions, ensuring that the retrieved clinical knowledge originates from objective descriptions of the patient's symptoms themselves, rather than statistical judgments based on the model.
[0045] In this embodiment of the application, RAG knowledge retrieval adopts a three-stage retrieval process to achieve accurate knowledge matching: The first stage involves retrieval triggering and query construction. The system extracts key clinical features from the patient's 27-dimensional characteristics, such as the location of abdominal pain, positive physical signs, and abnormal vital signs. It then automatically generates natural language queries based on structured templates of chief complaint, accompanying symptoms, positive physical signs, and vital signs. For example, for a patient complaining of right lower quadrant abdominal pain accompanied by fever and positive peritoneal irritation signs, the system automatically constructs the query: "Patients with acute abdomen presenting with right lower quadrant abdominal pain, positive peritoneal irritation signs, and fever, clinical practice guidelines recommended."
[0046] The second stage involves semantic similarity calculation and matching. A pre-trained embedding model in the medical domain is used to encode the query into a query vector. The cosine similarity between the query vector and all knowledge unit vectors is calculated in a vector database, and the vectors are sorted in descending order of similarity. Here, the knowledge unit vectors are high-dimensional vector representations of knowledge text from the evidence-based medicine knowledge base; semantically similar texts are closer in the vector space.
[0047] The third stage involves reordering and precisely filtering the search results. The system prioritizes knowledge units with the highest semantic similarity to the patient's condition. When the similarity is close, it prioritizes literature with a higher level of evidence. Finally, it selects the top-5 evidence-based knowledge fragments with the highest relevance.
[0048] After retrieving the Top-5 evidence-based knowledge fragments, the knowledge feature extraction unit extracts three types of structured knowledge features from the fragments to construct a 5-10 dimensional knowledge feature vector.
[0049] The three types of structured knowledge features specifically include: The guidelines recommend risk levels, specifically high-risk / non-high-risk, coded as 1 / 0; The departments recommended by the guidelines are specifically coded as vectors of [surgery, internal medicine, gynecology]; The key risk indicator matching degree is calculated by keyword matching to determine the overlap ratio between the patient's symptoms and the guideline description, with a value range of 0-1.
[0050] The constructed knowledge feature vectors are passed to the MoE (Expert Hybrid Model) module in the knowledge layer to inject and enhance knowledge into the prediction process.
[0051] S5, the standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector are input into the multi-task collaborative prediction model, fused into an enhanced feature vector by a shared encoder, and knowledge-enhanced prediction is performed by dynamically allocating expert network weights through a gating network. The prediction result is output, which includes the output high-risk probability, subject-specific recommendation probability distribution, and prediction confidence, including: The standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector are mapped into a 512-dimensional enhanced feature vector through a deep neural network; Based on the enhanced feature vector, the activation weights of multiple expert networks are calculated, and the two expert networks with the highest weights are selected for activation. The activated expert network performs knowledge-enhanced predictions and outputs the prediction results.
[0052] Specifically, the shared encoder of the MoE multi-task collaborative prediction module receives heterogeneous features from three sources: 27-dimensional original patient clinical features, a 6-dimensional base model prediction probability vector, and a 5-10 dimensional RAG knowledge feature vector, totaling approximately 38-43 dimensions of features. These features are mapped to a unified high-dimensional feature representation of 512 dimensions, namely the enhanced feature vector, through a two-layer fully connected deep neural network.
[0053] The shared encoder in this embodiment implements three functions: Used for feature standardization, mapping input data from different sources and with different dimensions to a unified feature space; To reduce computational redundancy, all experts share the same encoder, avoiding redundant calculations. Semantic alignment enables features from different modalities to be aligned in a high-dimensional space, forming semantic associations.
[0054] The gating network receives a 512-dimensional enhanced feature vector from the shared encoder as input. It then calculates the activation weights of three expert networks using a multilayer perceptron (MLP) structure. The output layer uses the Softmax activation function to generate a normalized weight vector G=[W1,W2,W3]. This invention employs a Top-K sparse gating mechanism (K=2), activating only the two experts with the highest weights during each inference iteration. The weights of unselected experts are reset to zero and re-normalized.
[0055] This application improves computational efficiency through sparse activation, avoiding full computation for all experts and encouraging clear specialization among experts during training. It also enhances interpretability by analyzing activation weight distribution to trace the basis of decisions. Specifically, the gating network's decisions are guided by RAG knowledge features; when these features indicate a high-risk scenario, the gating network tends to assign higher activation weights to experts providing high-risk warnings.
[0056] In some embodiments, the expert network layer consists of three complementary specialized experts, each performing domain fine-tuning based on a large language model, specifically including: High-risk early warning experts focus on high-risk assessment, deeply integrating the prediction results of the Top 3 high-risk early warning model to output the patient's high-risk probability and risk stratification; Specialty triage expert, focusing on specialty triage recommendations, deeply integrating the prediction results of the Top 3 specialty triage model, and outputting the recommendation probability distribution of internal medicine / surgery / gynecology; As a comprehensive decision-making expert and a bridge expert, we take into account the inherent relationship between high-risk factors and specialization, and provide comprehensive decision-making suggestions for overall optimization.
[0057] In this embodiment of the application, sparse activation specifically includes: When performing high-risk early warning tasks, activate high-risk early warning experts and comprehensive decision-making experts; When performing the task of specialization-based admission, activate the specialization-based admission expert and the comprehensive decision-making expert; In complex case scenarios, a collaborative decision-making process involving three experts is activated. The comprehensive decision-making expert, acting as a bridging expert, is activated in both tasks to ensure information sharing and collaboration between them.
[0058] The fusion layer is responsible for intelligently aggregating the prediction results of multiple expert networks according to the weights assigned by the gating network. Weighted fusion is performed separately for two tasks: high-risk assessment fuses the outputs of high-risk early warning experts and comprehensive decision-making experts; specialist triage decision-making fuses the outputs of specialist triage experts and comprehensive decision-making experts.
[0059] The output layer generates clinically usable triage conclusions and confidence scores, including high-risk probability and confidence, probability distribution of departmental recommendations, priority ranking of recommendations for each specialty, and knowledge contribution, i.e., the degree of influence of quantitative knowledge features on prediction, with a value range of 0-1.
[0060] Because RAG knowledge features are embedded in the input features, the prediction process is no longer purely data-driven statistical inference, but rather knowledge-enhanced reasoning that incorporates evidence-based medicine. Validated in a real emergency environment with backend data from 486 patients, the system's average response time was 10.85 seconds (median 7.67 seconds), with a response rate of 99.59% (≤90 seconds), an average fault-free uptime rate of 99.99%, and zero system crashes, fully meeting the real-time and high reliability requirements of emergency triage.
[0061] S6, verify the consistency between the prediction results of the multi-task collaborative prediction model and the evidence-based knowledge fragment, and generate consistency verification results including: A hybrid scoring method combining keyword matching and semantic similarity was used to calculate the consistency score; The consistency level is determined based on the consistency score. When the consistency level is completely consistent, proceed directly to the step of generating personalized clinical decision interpretation; When the consistency level is partially consistent, the steps of generating personalized clinical decision interpretation are performed after marking the differences. When the consistency level is inconsistent, a manual review warning is triggered, and the process proceeds to the step of generating personalized clinical decision interpretation.
[0062] Specifically, the consistency verification module, as the quality control layer of the system, cross-validates the guidelines and recommendations of RAG retrieval with the MoE prediction results, which is equivalent to the double-checking method in clinical practice.
[0063] In this embodiment, a hybrid scoring method combining keyword matching and semantic similarity is used to calculate the consistency score, as shown in the formula: Consistency score = 0.7 * keyword score + 0.3 * semantic similarity score.
[0064] The keyword score is calculated by detecting whether the risk level and recommended department keywords in the prediction results appear in the search text. The semantic similarity score is obtained by calculating the cosine similarity between the MoE predicted description and the RAG knowledge features.
[0065] The results are categorized into three levels based on the consistency score: Completely consistent, specifically with a consistency score greater than 0.9, directly proceed to S7 to generate an explanation; Partially consistent, specifically, the consistency score is between 0.7 and 0.9. After marking the differences, proceed to S7; Inconsistency occurs when the consistency score is less than 0.7, triggering a manual review warning and proceeding to S7.
[0066] Consistency verification effectively reduces the risk of misleading recommendations due to system prediction errors or knowledge retrieval biases, significantly enhancing the security and reliability of decision-making recommendations.
[0067] S7, Based on the prediction results of the multi-task collaborative prediction model, the evidence-based knowledge fragments, and the consistency verification results, generate personalized clinical decision interpretations, including: The structured instruction text is constructed based on the prediction results, the evidence-based knowledge fragments, and the consistency verification results. Based on a large language model, a three-part personalized clinical decision interpretation is generated from the structured instruction text. Specifically, the decision interpretation generation unit, based on a large language model such as Doubao, integrates patient clinical information, original text of knowledge fragments retrieved from RAG, MoE knowledge-enhanced prediction results, and consistency verification results into a structured Prompt, generating a three-part personalized clinical decision interpretation: The first paragraph describes the clinical significance of the predicted results, specifically the degree of high risk and the possible corresponding pathological state. The second paragraph contains the key clinical features and guidelines that support the prediction, specifically citing authoritative guidelines and sources of evidence retrieved from RAG, and clearly indicating the sources of evidence. The third section provides triage suggestions and precautions, including recommended first-line departments, alternative departments, and key points for differential diagnosis.
[0068] The generated explanations clearly indicate the evidence-based sources of the predictions, the contribution of knowledge features to the predictions, and the consistency verification results, ensuring that the decisions are transparent and traceable.
[0069] S8. Based on the high-risk probability and the departmental recommendation probability distribution, trigger the corresponding level of triage decision prompt signal, and visualize the personalized clinical decision interpretation and evidence-based decision basis.
[0070] Specifically, the system presents the output results of S5 to S7 in a visual manner on the decision support interface. For example... Figure 7 As shown, the system interface adopts a left-right split layout, with the left side being the patient information collection area and the right side being the prediction results and decision-making basis display area, which conforms to the clinical thinking habit of nurses to assess first and then judge.
[0071] The visualization includes: High-risk information is visualized, with high-risk patients specifically identified by red circular icons and high-risk text labels, providing a strong visual warning. The departmental recommendations are visualized using a horizontal progress bar to show the recommendation probability of each specialty. The length of the progress bar is proportional to the probability value, which intuitively presents the preferred specialty and the differences in the recommendation level of each selected specialty. The prediction confidence score is displayed, providing the confidence score of the MoE model output in the form of text labels; The decision-making basis is presented in a structured manner, showcasing key excerpts from clinical guidelines retrieved by RAG and sources of evidence, as well as a three-part personalized clinical decision interpretation text.
[0072] After reviewing the decision support report generated by the system, the triage nurse makes the final triage decision based on the patient's clinical presentation, system recommendations, and their own professional judgment. The system does not force nurses to accept its recommendations; nurses retain full clinical decision-making authority, achieving human-machine collaborative decision-making.
[0073] In some embodiments, a four-dimensional assessment of clinical outcomes is also included.
[0074] Specifically, after the system's clinical application, standardized effectiveness evaluation will be performed from the following four dimensions: Triage accuracy assessment: The study included four sub-indicators: under-triage rate, over-triage rate, departmental triage accuracy rate, and tiered triage accuracy rate. The chi-square test was used to compare the differences between the system-assisted triage and traditional triage in each accuracy index (P<0.05 was considered statistically significant).
[0075] Triage time assessment: Record the time from patient arrival to triage completion, compare the time consumption of traditional triage and system-assisted triage, and use non-parametric tests to analyze the differences.
[0076] System performance evaluation: Based on the system's background monitoring logs, the following indicators are statistically analyzed: average response time, median response time, response time compliance rate (with ≤90 seconds as the clinically acceptable threshold), number of system crashes, system error rate, and mean time between failures (MTBF) rate.
[0077] User acceptance assessment: Based on the Technology Acceptance Model (TAM), a 24-item Likert 5-point questionnaire with 6 dimensions (information quality, system quality, service quality, perceived usability, perceived usefulness, and willingness to use) was designed to comprehensively evaluate the clinical acceptability of the system from the perspective of nursing staff.
[0078] The above assessment provides evidence-based support for system iteration, optimization, and widespread application. A prospective, non-concurrent controlled study (780 patients, including 390 in the intervention group and 390 in the traditional triage control group) validated that even though the control group had used standardized triage tools including the MEWS score and electronic pre-screening triage system (baseline triage accuracy of 81.54%), the intelligent decision support system still achieved a significant improvement in triage accuracy (P<0.05), reducing the under-triage rate from 12.05% to 8.21%, without increasing triage time (P=0.499). The overall acceptance score of the 25 triage nurses reached a mean of 4.94 (out of 5), with all dimensions having a mean score of at least 4.88.
[0079] In some embodiments, the evidence-based medicine knowledge base contains 43 high-quality medical articles selected through systematic literature analysis, covering clinical practice guidelines (17 articles), clinical decision support literature (21 articles), expert consensus (3 articles), and systematic reviews (2 articles).
[0080] The knowledge base construction process includes knowledge source selection and text cleaning, structured knowledge extraction, knowledge vectorization encoding, and a dynamic maintenance mechanism that integrates regular review, triggered updates, and version management.
[0081] Among these, knowledge source selection and text cleaning remove duplicate content, irrelevant chapters and formatting marks from the included literature, retain core clinical knowledge, and standardize medical terminology in accordance with the International Classification of Diseases (ICD-11).
[0082] Structured knowledge extraction extracts core information such as typical symptoms, high-risk characteristics, triage departments, and triage recommendations of diseases from literature, and organizes and labels the sources of evidence in a structured manner.
[0083] Knowledge vectorization encoding uses a pre-trained embedding model in the medical field to transform knowledge text into a high-dimensional vector representation, which is stored in a vector database along with the original text and structured information. Texts with similar semantics are closer in the vector space, supporting intelligent retrieval based on semantic similarity.
[0084] The following is a complete demonstration of the system's data flow from data input to decision output through a typical acute abdominal case.
[0085] Case: A 65-year-old male patient presented with right lower abdominal pain for 6 hours, body temperature of 38.5℃, heart rate of 95 beats / min, and positive peritoneal irritation signs.
[0086] Time T0: Input data: Standardized data for 32 items including age = 65, gender = male, abdominal pain location = right lower quadrant, body temperature = 38.5℃, peritoneal irritation sign = positive. After data cleaning, missing value imputation, and variable encoding, the output is a 27-dimensional standardized feature vector X. patient .
[0087] Time T1: Perform parallel processing.
[0088] Basic model prediction: X patient Input six pre-trained models for parallel prediction, and output a 6-dimensional base prediction probability vector X. pred =[0.87,0.85,0.89,0.88,0.86,0.90], where the first three dimensions represent the high-risk probability and the last three dimensions represent the surgical probability.
[0089] Simultaneously, RAG knowledge retrieval was performed: key features were extracted: right lower quadrant pain, fever, and positive peritoneal irritation signs. A query was constructed: "Patients with acute abdomen presenting with right lower quadrant pain, positive peritoneal irritation signs, and fever, and recommendations from clinical practice guidelines." Top-5 guideline fragments were retrieved via vector retrieval, with the retrieved guideline text stating: "The presence of peritoneal irritation signs is a strong predictor of high risk for acute abdomen, requiring immediate surgical evaluation..." Knowledge feature vector X was then extracted and output. knowledge =[1,1,0,0,0.67], where risk = high risk, department = surgery, and match degree = 0.67, while retaining the original guideline text for subsequent consistency verification.
[0090] Time T2: MoE knowledge-enhanced prediction, input X patient +Xpred +X knowledge The shared encoder is mapped to 512-dimensional augmented features. The gating network calculates expert weights [0.6, 0, 0.4], activating high-risk early warning experts and comprehensive decision-making experts.
[0091] The weighted fusion output is as follows: high risk probability 0.89 (2% higher than the pure data-driven 0.87), surgical probability 0.88, knowledge contribution 0.15, and the prediction result text is "high risk, surgical triage recommended".
[0092] Time T3: Consistency verification involves comparing the MoE prediction (high risk, surgical) with the RAG guideline text (peritoneal irritation signs...immediate surgical evaluation required). Based on keyword matching and semantic similarity calculation, a comprehensive consistency score greater than 0.85 is considered completely consistent, and the process proceeds to interpretation generation.
[0093] Time T4: The large language model receives patient information, RAG knowledge fragments, MoE predictions, and consistency verification results to construct a structured Prompt and generate a three-part personalized decision explanation.
[0094] Final output: High risk; The recommended surgery category is shown in the horizontal progress bar, which indicates 88% for surgery, 8% for internal medicine, and 4% for gynecology. High prediction confidence; Decision basis: Excerpt from the structured presentation guidelines and the above three-part explanation.
[0095] On the other hand, such as Figure 6 As shown, the present invention also provides a triage decision support system for acute abdominal pain, which adopts a four-layer progressive architecture design based on the DIKW (Data-Information-Knowledge-Wisdom) hierarchical model, including: The data layer includes a patient information collection module. This module embeds the standardized "Triage Questionnaire for Patients with Acute Abdominal Pain" into the hospital's emergency information system (HIS). Through an interface with the HIS, it automatically collects standardized data across 6 dimensions and 32 items when a patient arrives at the triage window. This includes general characteristics, vital signs, abdominal pain characteristics, abdominal signs, comprehensive score, and emergency outcome data. Collection methods include directly reading pre-entered vital sign information from the HIS and recording the patient's chief complaint through item-by-item questioning by the triage nurse. When the triage nurse enters "abdominal pain" in the chief complaint field, the system automatically recognizes and triggers the acute abdominal pain intelligent decision support module, displaying a structured consultation interface. Simultaneously, this module supports tracing and supplementing missing data from the electronic medical record system using the patient's ID number, ensuring the completeness of the collected data.
[0096] The information layer comprises a preprocessing module and a basic model prediction module. The preprocessing module performs standardization processing on the raw clinical data acquired by the data acquisition module, including cleaning, missing value imputation, and variable encoding. It then uses the MICE algorithm based on random forest for multiple imputation to construct a 27-dimensional standardized feature vector. The basic model prediction module includes six pre-trained machine learning models (Top 3 for high-risk warning: XGBoost, LightGBM, Random Forest; Top 3 for specialty-specific warning: XGBoost, LightGBM, Random Forest). These models receive the standardized 27-dimensional feature vector and perform independent parallel predictions, generating a 6-dimensional basic prediction probability vector, which is then output to the MoE module in the knowledge layer.
[0097] The knowledge layer comprises a RAG knowledge retrieval module, a multi-task collaborative prediction module, and a consistency verification module. The RAG knowledge retrieval module includes a knowledge base construction unit, a knowledge retrieval unit, and a knowledge feature extraction unit. It constructs a vector database based on 43 authoritative medical articles and outputs Top-5 evidence-based knowledge fragments and 5-10 dimensional structured knowledge feature vectors through a three-stage retrieval process. The multi-task collaborative prediction module, based on the MoE architecture, fuses patient original features, basic prediction probabilities, and RAG knowledge features into a 512-dimensional enhanced feature vector through a shared encoder. The activation weights of three expert networks (high-risk warning expert, specialist triage expert, and comprehensive decision-making expert) are dynamically allocated via a gating network. A Top-K (K=2) sparse gating strategy is used to perform knowledge-enhanced prediction, outputting high-risk probability, specialty recommendation probability distribution, priority ranking, and knowledge contribution. The consistency verification module uses a hybrid scoring method (keyword matching weight 0.7 + semantic similarity weight 0.3) to cross-validate the MoE prediction results with RAG guideline recommendations. It determines complete consistency, partial consistency, or inconsistency based on two threshold levels of 0.9 and 0.7, triggering corresponding processing steps.
[0098] The intelligent layer comprises a decision interpretation generation module and an early warning and visualization module. The decision interpretation generation module, based on a large language model, integrates patient information, RAG knowledge fragments, MoE prediction results, and consistency verification results into a structured prompt, generating a three-stage personalized clinical decision interpretation. The early warning and visualization module presents risk level, subspecialty recommendation, confidence level, and evidence-based basis visually on the system interface: high-risk is indicated by a red circle, the probability of each specialty recommendation is displayed by a horizontal progress bar, the prediction confidence level is displayed by text icons, and the evidence-based decision basis is displayed by structured text. This module compares the high-risk probability with a preset threshold. When the threshold is reached or exceeded, a visual or acoustic early warning is automatically triggered, suggesting that the triage nurse increase the patient's pre-screening priority, while using the prediction results as a quantitative reference for triage decisions. The system does not force nurses to accept the suggestions, preserving their full clinical decision-making authority.
[0099] The system connects all layers via standardized data interfaces, supporting independent upgrades and flexible replacements. It can be embedded into a hospital's electronic triage system, allowing triage nurses to obtain risk warnings and departmental recommendations within seconds simply by asking questions and entering patient information at the triage window, without requiring any additional steps.
[0100] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A triage decision support method for acute abdominal pain, characterized in that, include: Obtain patients' clinical data; The clinical data is preprocessed to construct a standardized feature vector; The standardized feature vector is input into multiple basic prediction models for parallel prediction to generate a basic prediction probability vector. Based on the clinical data, evidence-based knowledge fragments related to the clinical data are retrieved from the evidence-based medicine knowledge base, and structured knowledge feature vectors are extracted. The standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector are input into the multi-task collaborative prediction model, fused into an enhanced feature vector by a shared encoder, and knowledge-enhanced prediction is performed by dynamically allocating expert network weights through a gating network. The prediction results are output, including the output high-risk probability, subject-specific recommendation probability distribution, and prediction confidence. The prediction results of the multi-task collaborative prediction model are verified for consistency with the evidence-based knowledge fragments, and a consistency verification result is generated. Based on the prediction results of the multi-task collaborative prediction model, the evidence-based knowledge fragments, and the consistency verification results, a personalized clinical decision interpretation is generated. Based on the high-risk probability and the departmental recommendation probability distribution, a triage decision prompt signal of the corresponding level is triggered, and the personalized clinical decision interpretation and evidence-based decision basis are visualized.
2. The support method according to claim 1, characterized in that, The step of inputting the standardized feature vector into multiple basic prediction models for parallel prediction, generating multiple basic prediction modules in the basic prediction probability vector, includes: A high-risk early warning model group, comprising three high-risk early warning models, each of which outputs a corresponding high-risk probability vector. as well as The subject-specific model group includes three subject-specific models, each of which is used to output a corresponding departmental affiliation probability vector.
3. The support method according to claim 1, characterized in that, The step of retrieving evidence-based knowledge fragments related to the clinical data from the evidence-based medicine knowledge base based on the clinical data, and extracting structured knowledge feature vectors, includes: Key clinical features are extracted from the clinical data, and natural language queries are automatically generated based on the structured templates corresponding to the key clinical features. The natural language query is encoded into a query vector. The semantic similarity between the query vector and each knowledge unit vector is calculated in the evidence-based medicine knowledge base. The results are sorted in descending order of similarity and the search results are output. The knowledge unit vector is a high-dimensional vector converted from the knowledge text in the evidence-based medicine knowledge base. Based on the search results, select the most relevant evidence-based knowledge fragments; Structured knowledge features are extracted from the multiple evidence-based knowledge fragments to construct the structured knowledge feature vector; The structured knowledge feature vector includes the risk level recommended by the guidelines, the departments recommended by the guidelines, and the matching degree of key risk indicators.
4. The support method according to claim 1, characterized in that, The standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector are input into a multi-task collaborative prediction model, fused into an enhanced feature vector by a shared encoder, and then used to dynamically allocate expert network weights through a gating network for knowledge-enhanced prediction. The prediction results are output, including: The standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector are mapped into a 512-dimensional enhanced feature vector through a deep neural network; Based on the enhanced feature vector, the activation weights of multiple expert networks are calculated, and the two expert networks with the highest weights are selected for activation. The activated expert network performs knowledge-enhanced predictions and outputs the prediction results.
5. The support method according to claim 1, characterized in that, The expert network includes high-risk early warning experts, specialist triage experts, and comprehensive decision-making experts. When performing a high-risk early warning task, activate the high-risk early warning expert and the comprehensive decision-making expert; When performing a specialist admissions task, the specialist admissions expert and the comprehensive decision-making expert are activated; When a complex case scenario is identified, the three expert networks mentioned above are activated to make a joint decision.
6. The support method according to claim 1, characterized in that, The process of verifying the consistency between the prediction results of the multi-task collaborative prediction model and the evidence-based knowledge fragments, and generating consistency verification results, includes: A hybrid scoring method combining keyword matching and semantic similarity was used to calculate the consistency score; The consistency level is determined based on the consistency score. When the consistency level is completely consistent, proceed directly to the step of generating personalized clinical decision interpretation; When the consistency level is partially consistent, the steps of generating personalized clinical decision interpretation are performed after marking the differences. When the consistency level is inconsistent, a manual review warning is triggered, and the process proceeds to the step of generating personalized clinical decision interpretation.
7. The support method according to claim 1, characterized in that, The personalized clinical decision interpretation is generated based on the prediction results of the multi-task collaborative prediction model, the evidence-based knowledge fragments, and the consistency verification results, including: The structured instruction text is constructed based on the prediction results, the evidence-based knowledge fragments, and the consistency verification results. Based on a large language model, a three-part personalized clinical decision interpretation is generated from the structured instruction text.
8. The support method according to claim 1, characterized in that, The visualization presents the personalized clinical decision interpretation and evidence-based decision-making basis, including: High-risk levels are indicated by red circular markers combined with high-risk text labels. The recommendation probability of each specialty is displayed using a horizontal progress bar, and the length of the horizontal progress bar is proportional to the recommendation probability. The predicted confidence level is displayed using text labels; The evidence-based decision-making basis is presented in structured text, including key excerpts from clinical guidelines retrieved by RAG and sources of evidence.
9. The support method according to claim 1, characterized in that, The preprocessing of the clinical data to construct a standardized feature vector includes: The clinical data was cleaned to remove outliers; Missing values were handled for the cleaned clinical data; The clinical data after missing values were processed were coded into variables to construct the standardized feature vector.
10. A triage decision support system for acute abdominal pain, characterized in that, include: The patient information collection module, configured in the data layer, is used to acquire patients' clinical data. The preprocessing module, configured in the information layer, is used to preprocess the clinical data and construct standardized feature vectors; The basic model prediction module, configured in the information layer, is used to input the standardized feature vector into multiple basic prediction models for parallel prediction and generate a basic prediction probability vector. The knowledge retrieval module, configured in the knowledge layer, is used to retrieve evidence-based knowledge fragments related to the clinical data from the evidence-based medicine knowledge base, and extract structured knowledge feature vectors based on the clinical data. A multi-task collaborative prediction module, configured in the knowledge layer, is used to input the standardized feature vector, the basic prediction probability vector, and the structured knowledge feature vector into the multi-task collaborative prediction model, fuse them into a unified high-dimensional feature representation through a shared encoder, and perform knowledge-enhanced prediction by dynamically allocating expert network weights through a gating network, and output prediction results. The prediction results include high-risk probability, subject-specific recommendation probability distribution, and prediction confidence. A consistency verification module, configured in the knowledge layer, is used to verify the consistency between the prediction results of the multi-task collaborative prediction module and the evidence-based knowledge fragments retrieved by the knowledge retrieval module, and generate a consistency verification result. The decision interpretation generation module, configured in the intelligence layer, is used to generate personalized clinical decision interpretations based on the prediction results of the multi-task collaborative prediction module, the evidence-based knowledge fragments retrieved by the knowledge retrieval module, and the consistency verification results. The early warning and visualization module, configured in the intelligent layer, is used to trigger triage decision prompt signals of corresponding levels based on the high-risk probability and the departmental recommendation probability distribution, and to visualize the personalized clinical decision interpretation and evidence-based decision basis.