Triage auxiliary decision-making method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a triage auxiliary decision-making method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining the treatment data of a target patient, carrying out the processing of the treatment data, and extracting the feature information representing the illness state of the target patient; matching with a preset association rule base based on the feature information, acquiring supplementary information of the target patient under the condition of successful matching, and updating the feature information according to the supplementary information; inputting the updated feature information into a pre-trained triage model, and outputting auxiliary decision information of the target patient; wherein the triage model is obtained by training based on historical triage data. Through the method provided by the embodiment of the invention, the problem of poor information caused by unclear complaint or limited sign information of the patient is effectively solved, more reliable auxiliary decision suggestions are output, and the risk of misclassification caused by insufficient information is reduced from the whole process.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a triage auxiliary decision-making method, device, electronic device and storage medium. Background Technology

[0002] As the core frontline for treating critically ill patients in hospitals, the accuracy and efficiency of triage in the emergency department directly affect the safety of patients' lives and the rational allocation of medical resources.

[0003] Currently, triage is generally based on the personal experience and subjective judgment of triage nurses. When faced with patients whose complaints are vague, whose vital signs are limited, or whose conditions are complex and concealed, misclassification is prone to occur, leading to delays in treatment for critically ill patients or the waste of valuable rescue resources on non-emergency patients. In recent years, although some studies have attempted to introduce machine learning models to assist triage decisions, these existing technical solutions are mostly limited to passive classification of existing static datasets. When dealing with the dynamically changing data characteristics and computational loads of emergency scenarios, their prediction accuracy, robustness, and interpretability all need improvement, failing to meet the actual clinical needs for highly reliable auxiliary decision-making. Therefore, a solution is urgently needed to address these problems. Summary of the Invention

[0004] This application provides a triage auxiliary decision-making method, device, electronic device, and storage medium to address the deficiencies in the prior art.

[0005] This application provides a triage auxiliary decision-making method, the method comprising: Acquire the medical data of the target patient, process the medical data, and extract feature information that characterizes the condition of the target patient; The feature information is matched with a pre-set association rule base. If the match is successful, the supplementary information of the target patient is obtained, and the feature information is updated according to the supplementary information. The updated feature information is input into the pre-trained triage model, and the auxiliary decision-making information of the target patient is output; wherein, the triage model is trained based on historical triage data.

[0006] According to an embodiment of this application, a triage-assisted decision-making method is provided, wherein the medical data includes numerical physiological indicators and textual chief complaint data; The process of processing the medical data to extract feature information characterizing the condition of the target patient includes: For the numerical physiological indicators, outlier detection and removal are performed, and missing value imputation is performed on the data after outlier removal to obtain the first feature sub-information. For the text-based chief complaint data, terminology consistency standardization processing is performed to obtain the second feature sub-information; The feature information is constructed by combining the first feature sub-information and the second feature sub-information.

[0007] According to an embodiment of this application, a triage-assisted decision-making method is provided, wherein the textual chief complaint data undergoes terminology consistency standardization processing to obtain second feature sub-information, including: For the text-based chief complaint data, terminology consistency standardization processing is performed to obtain standardized chief complaint data; The standardized chief complaint data is input into a pre-trained medical entity recognition model, which outputs the identified symptoms, signs, and disease entities as the second feature sub-information; wherein, the medical entity recognition model is constructed based on a bidirectional long short-term memory network and a conditional random field framework.

[0008] According to an embodiment of this application, a triage auxiliary decision-making method is provided, wherein the feature information is constructed by using the first feature sub-information and the second feature sub-information, including: The first feature sub-information and the second feature sub-information are combined, and feature engineering processing is performed on the combined data to generate the feature information; The feature engineering process includes: segmenting and discretizing numerical features, and performing word frequency-inverse document frequency weighting and encoding conversion on textual entity features.

[0009] According to an embodiment of this application, a triage-assisted decision-making method includes the following process for constructing the pre-set association rule base: Obtain the historical triage data and construct a feature dataset based on the historical triage data; The feature dataset is scanned to extract at least one high-frequency symptom combination that meets the preset frequency conditions, forming a frequent itemset; Based on the frequent itemset, multiple scans are performed to iteratively calculate the support and confidence of the at least one high-frequency symptom combination; Based on the support and confidence of the at least one high-frequency symptom combination, strong rules associated with the target high-risk disease are determined to construct the association rule base.

[0010] According to an embodiment of this application, a triage-assisted decision-making method is provided, wherein the step of matching the feature information with a pre-set association rule base, and obtaining supplementary information of the target patient if the match is successful, includes: The chief complaint symptoms in the feature information are compared with the rule antecedents in the association rule base to generate a comparison result; wherein, the chief complaint symptoms are the discomfort symptoms actively stated by the target patient, and the rule antecedents are the condition part of the rule; If the comparison result indicates a successful comparison, a supplementary questioning process for the target patient is triggered; wherein, the supplementary questioning process is to automatically push at least one key chief complaint follow-up question corresponding to the triggered rule; Receive the response results to the at least one key chief complaint inquiry, and use the response results as the supplementary information.

[0011] According to an embodiment of this application, a triage auxiliary decision-making method is provided, wherein the triage model includes a tiered prediction module and a severe illness prediction module; the auxiliary decision-making information includes triage level and severe illness risk prediction results; The step of inputting the updated feature information into the pre-trained triage model and outputting the auxiliary decision-making information for the target patient includes: The updated feature information is input into the triage prediction module and the severe illness prediction module respectively, and the triage level of the target patient and the severe illness risk prediction result of the target patient are output.

[0012] A triage-assisted decision-making method is provided according to an embodiment of this application. The hierarchical prediction module is configured to adopt a multi-model integration architecture, which integrates at least the following four neural network models: text convolutional neural network, bidirectional long short-term memory network, occlusion correction pre-trained language model, and meta-controller pre-trained language model.

[0013] According to an embodiment of this application, a triage auxiliary decision-making method is provided, wherein the updated feature information is input into the triage prediction module, and the triage level of the target patient is output, including: Based on the updated feature information, real-time load, and data features of the updated feature information, the weights of the integrated neural network model output results are dynamically adjusted by the hierarchical prediction module to jointly determine the triage level of the target patient.

[0014] According to an embodiment of this application, a triage-assisted decision-making method is provided, wherein the weights of the output results of the integrated neural network model are dynamically adjusted by the hierarchical prediction module based on the updated feature information, real-time load, and data features of the updated feature information, including: When the system throughput requirement exceeds a preset threshold, the weights of the text convolutional neural network model are increased, and the weights of the bidirectional long short-term memory network model are decreased.

[0015] According to an embodiment of this application, a triage auxiliary decision-making method is provided, wherein the severe illness prediction module is a gradient boosting decision tree model; the auxiliary decision-making information also includes an interpretable report corresponding to the severe illness risk prediction result; The step of inputting the updated feature information into the pre-trained triage model and outputting the auxiliary decision-making information for the target patient further includes: The updated feature information is input into the critical illness prediction module, and the contribution of each feature in the updated feature information to the critical illness risk prediction result is calculated. Based on the contribution of each feature to the prediction of the severe illness risk, an interpretability report is generated that explains the degree of influence of each feature.

[0016] According to an embodiment of this application, a triage-assisted decision-making method, after determining strong rules associated with a target high-risk disease based on the support and confidence of the at least one high-frequency symptom combination to construct the association rule base, the method further includes: New triage data is acquired according to a preset period, and strong rule mining is re-executed based on the new triage data to update the association rule base.

[0017] This application embodiment also provides a triage auxiliary decision-making device, the device comprising: The acquisition module is used to acquire the medical data of the target patient, process the medical data, and extract feature information that characterizes the condition of the target patient. The update module is used to match the feature information with a preset association rule base. If the match is successful, it obtains supplementary information of the target patient and updates the feature information according to the supplementary information. The decision module is used to input the updated feature information into the pre-trained triage model and output the auxiliary decision information of the target patient; wherein the triage model is trained based on historical triage data.

[0018] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the triage auxiliary decision-making method as described above.

[0019] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the triage auxiliary decision-making method as described above.

[0020] This application also provides a computer program product, including a computer program that, when executed by a processor, implements any of the triage auxiliary decision-making methods described above.

[0021] This application provides a triage auxiliary decision-making method, device, electronic device, and storage medium. It acquires and processes the medical data of a target patient to extract feature information characterizing the patient's condition. Based on this feature information, it matches it with a pre-set association rule base. If a match is successful, it acquires supplementary information about the target patient and updates the feature information accordingly. The updated feature information is then input into a pre-trained triage model, outputting auxiliary decision-making information for the target patient. The triage model is trained based on historical triage data. Therefore, this application, through real-time matching of initially extracted feature information using a pre-set association rule base, can proactively identify information gaps in the initial triage judgment and trigger targeted information supplementation processes. This dynamically and purposefully guides the collection of key clinical information, effectively compensating for information gaps caused by unclear patient complaints or limited physical examination information. This significantly improves the quality of the feature information input into the triage model, resulting in more reliable auxiliary decision-making suggestions and reducing the risk of misclassification due to insufficient information throughout the overall process. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the triage-assisted decision-making method provided in the embodiments of this application.

[0024] Figure 2 This is a schematic diagram of the chief complaint recommendation process provided in the embodiments of this application.

[0025] Figure 3 This is a schematic diagram of the hierarchical prediction process provided in the embodiments of this application.

[0026] Figure 4 This is a schematic diagram of the critical illness prediction process provided in the embodiments of this application.

[0027] Figure 5 This is a complete flowchart of the triage auxiliary decision-making method provided in the embodiments of this application.

[0028] Figure 6 This is a schematic diagram of the triage auxiliary decision-making device provided in the embodiments of this application.

[0029] Figure 7This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the embodiments of this application.

[0031] The following description, in conjunction with the accompanying drawings, describes a triage auxiliary decision-making method, apparatus, electronic device, and storage medium according to embodiments of this application.

[0032] It is important to note that the emergency department is the department in a hospital with the highest concentration of critically ill patients, the most diverse range of illnesses, and the heaviest workload in terms of resuscitation and management. Patients admitted are mostly those with urgent, complex, and rapidly changing conditions. The emergency department is a crucial and unique department within a hospital; if medical staff fail to assess patients' conditions and their progression early enough, it will not only increase the incidence of errors but also raise patient mortality rates. Emergency triage is the first step in the emergency nursing process, involving the rapid assessment of emergency patients' conditions and the quick classification to determine treatment methods or further priority. Triage errors are known as emergency misclassification. Emergency misclassification primarily involves classification by level. According to data from a tertiary hospital over the past four years, 11.2% of emergency patients experienced this level misclassification. Emergency misclassification can lead to a series of serious consequences. Missing critically ill patients will delay their treatment, missing precious golden rescue time; over-classifying critically ill patients will waste valuable emergency resources, delay the treatment of waiting patients, and affect patient care and satisfaction. Rapid and accurate emergency triage means quickly matching patients with emergency resources. On the one hand, it can shorten the time patients spend seeking medical attention, waiting for treatment, and even being transferred; on the other hand, it can reduce the mortality rate of critically ill patients, improve their quality of life, reduce the occurrence of medical disputes, and create a safe and healthy medical environment.

[0033] Emergency departments urgently need to triage patients quickly and accurately, making the establishment of an intelligent triage system imperative. This would facilitate the effective identification of high-risk patients and contribute to epidemic control. However, according to doctors' experience, mis-triage in emergency departments often stems from: (1) overcrowding in the emergency department; and (2) over-reliance on doctors' subjective experience in general triage schemes. While methods such as logistic regression and decision trees have been developed for classifying patients by triage level, methods that provide additional questioning to help doctors obtain crucial diagnostic information and compensate for patient information gaps have not yet emerged. This results in some critically ill patients being triaged to lower levels and having to wait in line with non-critical patients, leading to delays in their treatment. Therefore, this application provides a triage auxiliary decision-making method to address at least one of the above problems.

[0034] Figure 1 This is a flowchart illustrating the triage-assisted decision-making method provided in the embodiments of this application, as shown below. Figure 1 As shown, the method includes the following: Step 100: Obtain the medical data of the target patient, process the medical data, and extract the feature information that characterizes the condition of the target patient.

[0035] Specifically, medical data refers to the multi-source heterogeneous information about patients initially collected in the triage scenario, which usually includes, but is not limited to, the patient's chief complaint (such as "chest tightness, abdominal pain"), basic vital signs (such as blood pressure, heart rate, body temperature), basic personal information (such as age, gender), and mode of arrival.

[0036] Processing medical data is a crucial preprocessing stage, aiming to transform raw, messy data, potentially containing missing or anomalies, into high-quality information suitable for machine analysis. This process specifically involves data cleaning, missing value imputation, and medical terminology standardization of textual complaints (e.g., unifying "palpitation" as "heart palpitations"). Through these processes, characteristic information representing the target patient's condition is ultimately extracted. This characteristic information is a structured or vectorized data set that comprehensively and concisely reflects the patient's current clinical status.

[0037] Step 200: Match the feature information with a preset association rule base. If the match is successful, obtain the supplementary information of the target patient and update the feature information according to the supplementary information.

[0038] It should be noted that the association rule base is based on a large amount of historical triage case data and was pre-learned through data mining technology. Its content is represented by a series of "IF-THEN" rules, such as "IF chief complaint includes 'melena' AND low blood pressure THEN high-risk disease is 'gastrointestinal bleeding'".

[0039] Specifically, matching refers to comparing the current patient's characteristic information with the rule conditions (IF part) in the rule base. When a match is successful, it indicates that the current patient may belong to a high-risk condition that is easily missed or misdiagnosed. At this time, the system does not directly give a final judgment, but actively obtains supplementary information about the target patient. For example, it automatically generates and asks, "Do you have symptoms such as dizziness or palpitations?" The system then incorporates the patient's answer as supplementary information into the original characteristic information, realizing dynamic updates of the characteristics and thus making up for the deficiency of insufficient initial information.

[0040] Step 300: Input the updated feature information into the pre-trained triage model and output the auxiliary decision-making information of the target patient; wherein the triage model is trained based on historical triage data.

[0041] Specifically, this more comprehensive and accurate updated feature information is input into the pre-trained triage model. This model is a machine learning or deep learning model trained on massive amounts of historical triage data, which has internally learned the complex mapping relationship between complex features and triage decisions.

[0042] After the model runs, it outputs auxiliary decision-making information for the target patient. This information usually includes suggested triage level (such as Level 1 critical, Level 2 severe, etc.) and prediction of severe illness risk, providing triage nurses with objective and quantitative decision-making references, thereby improving the accuracy and efficiency of triage.

[0043] The above describes the steps of the triage auxiliary decision-making method provided in this application embodiment. As can be seen from the above description, according to the triage auxiliary decision-making method provided in this application embodiment, by acquiring the medical data of the target patient and processing the medical data, feature information characterizing the target patient's condition is extracted; the feature information is matched with a pre-set association rule base; if the match is successful, supplementary information of the target patient is acquired, and the feature information is updated according to the supplementary information; the updated feature information is input into a pre-trained triage model, and auxiliary decision-making information for the target patient is output; wherein, the triage model is trained based on historical triage data. Therefore, this application embodiment, through real-time matching of the initially extracted feature information using a pre-set association rule base, can proactively identify information gaps in the initial triage judgment and trigger a targeted information supplementation process, thereby dynamically and specifically guiding the collection of key clinical information. This effectively compensates for information gaps caused by unclear patient complaints or limited physical signs, significantly improving the quality of the feature information finally input into the triage model, thus outputting more reliable auxiliary decision-making suggestions and reducing the risk of misclassification due to insufficient information in the overall process.

[0044] Based on the above embodiments, in this embodiment, the medical data includes numerical physiological indicators and textual chief complaint data; Step 100 processes the medical data to extract feature information characterizing the target patient's condition, including: Step 110: For the numerical physiological indicators, outlier detection and removal are performed, and missing value imputation is performed on the data after outlier removal to obtain the first feature sub-information.

[0045] Step 120: Perform terminology consistency standardization processing on the text-type complaint data to obtain the second feature sub-information.

[0046] Step 130: Construct the feature information using the first feature sub-information and the second feature sub-information.

[0047] It should be noted that the medical data is divided into numerical physiological indicators (such as quantifiable measurements such as systolic blood pressure of 120 mmHg, heart rate of 85 beats / min, and body temperature of 37.2℃) and textual complaint data (such as natural language descriptions such as patients' self-reported symptoms such as "chest pain", "shortness of breath", and "nausea and vomiting").

[0048] Specifically, for numerical physiological indicators, outlier detection and removal are first performed. For example, statistical methods (such as Z-score) or clinically reasonable ranges (such as systolic blood pressure not lower than 50 mmHg or higher than 250 mmHg) are used to identify and remove obviously erroneous or physiologically impossible extreme values. Subsequently, missing values ​​are imputed for any remaining items in the cleaned data, which can be done using algorithms such as K-nearest neighbors based on similar cases for reasonable estimation. The well-organized and complete set of numerical data obtained after these two steps constitutes the first feature sub-information.

[0049] Simultaneously, for text-based complaint data, terminology standardization is performed. This is a crucial step in overcoming the diversity, ambiguity, and non-standardization of medical text expressions. For example, using medical knowledge graphs or standard terminology dictionaries, colloquial terms like "palpitation" and "chest tightness" are uniformly mapped to the standard medical terms "heart palpitation" and "chest pain," respectively. Synonyms and abbreviations (such as "CA" standardized to "cancer") are also aligned. The set of text data with consistent semantic representation after standardization is the second feature sub-information.

[0050] Finally, by using techniques such as vectorization and splicing, the first feature sub-information and the second feature sub-information are effectively combined and integrated to form a comprehensive, unified, and machine-readable feature information vector or data structure. This feature information integrates the patient's objective physiological indicators and subjective symptom descriptions, providing a high-quality, low-noise input foundation for subsequent intelligent matching and prediction.

[0051] Further, step 120 performs terminology consistency standardization processing on the textual complaint data to obtain second feature sub-information, including: Step 121: Perform terminology consistency standardization processing on the text-type chief complaint data to obtain standardized chief complaint data.

[0052] Step 122: Input the standardized chief complaint data into the pre-trained medical entity recognition model, and output the identified symptoms, signs and disease entities as the second feature sub-information; wherein, the medical entity recognition model is constructed based on a bidirectional long short-term memory network and a conditional random field framework.

[0053] Specifically, the original text-based chief complaint data undergoes terminology standardization. This is a text normalization process based on a medical knowledge base (such as the ICD-10 Disease Classification and Standard Medical Terminology Dictionary). Its goal is to map and replace diverse, ambiguous, and even misspelled colloquial expressions with unified and unambiguous standard medical terms. For example, "palpitation" and "heart palpitation" are unified as "heart palpitation"; "shortness of breath" and "difficulty breathing" are unified as "dyspnea"; and "abdominal pain" is standardized as "upper abdominal pain" or "lower abdominal pain" depending on the specific location and nature. After this step, standardized chief complaint data with standardized expressions and unified terminology is obtained.

[0054] Subsequently, this cleaned and standardized chief complaint data was input into a pre-trained medical entity recognition model. This model is a natural language processing model specifically designed for medical texts, and its core architecture is constructed using a bidirectional long short-term memory network and a conditional random field framework. The bidirectional long short-term memory network can fully capture the contextual semantic information of each word in the text sequence, thereby deeply understanding the specific clinical meaning; while the conditional random field, as a sequence labeling framework, accurately delineates entity boundaries and assigns them the correct category labels based on the word formation rules and contextual dependencies of medical entities. After being trained on a large number of labeled medical texts, the model has the ability to recognize specific types of medical entities. After the model runs, it outputs the identified symptoms, signs, and disease entities from the input sentences. For example, from "chest pain with shortness of breath after activity," it identifies "chest pain" (symptom entity) and "shortness of breath" (symptom entity), and may associate them with potential "angina pectoris" or "heart failure" (disease entities).

[0055] Furthermore, step 130 constructs the feature information using the first feature sub-information and the second feature sub-information, including: The first feature sub-information and the second feature sub-information are combined, and feature engineering processing is performed on the combined data to generate the feature information; The feature engineering process includes: segmenting and discretizing numerical features, and performing word frequency-inverse document frequency weighting and encoding conversion on textual entity features.

[0056] Specifically, the first feature sub-information and the second feature sub-information are first combined, aligned and associated according to the patient identifier, to form an initial composite data view containing all the multimodal information of the patient.

[0057] Subsequently, key feature engineering processes were performed on this combined data to deeply uncover its inherent patterns and improve model performance. Among these processes, segmenting and discretizing numerical features is a common technique for continuous variables such as age and blood pressure. For example, instead of directly inputting "65 years old" as a continuous numerical value, it is mapped to the discrete interval of "elderly (≥60 years old)" based on its clinical significance; or "systolic blood pressure 145 mmHg" is categorized into the "grade 1 hypertension (140-159 mmHg)" interval. This processing transforms continuous numerical values ​​into categorical features with clear clinical classification significance, helping the model capture nonlinear relationships and reduce the interference of outliers. Simultaneously, word frequency-inverse document frequency weighting and encoding conversion of textual entity features is a core step in transforming non-numerical medical entities into numerical vectors. Term frequency-inverse document frequency (TF-IDF) weighting is used to assess the importance of an entity in a specific patient's description. On one hand, it calculates the frequency (term frequency, TF) of the entity in the current patient's complaint; on the other hand, it reduces the weight (inverse document frequency, IDF) of entities common in all patient complaints (such as "discomfort"), thereby highlighting specific symptom or disease entities with discriminative significance. After weighting, these weighted entities are converted into fixed-length, purely numerical feature vectors through encoding transformation, enabling the textual information to be understood and computed by mathematical models.

[0058] The triage-assisted decision-making method provided in this embodiment constitutes a complete and robust feature construction pipeline from raw data to high-quality model input, laying a crucial data foundation for subsequent realization of high-precision and high-reliability intelligent triage decisions.

[0059] Based on the above embodiments, in this embodiment, the construction process of the preset association rule base includes: Obtain the historical triage data and construct a feature dataset based on the historical triage data; The feature dataset is scanned to extract at least one high-frequency symptom combination that meets the preset frequency conditions, forming a frequent itemset; Based on the frequent itemset, multiple scans are performed to iteratively calculate the support and confidence of the at least one high-frequency symptom combination; Based on the support and confidence of the at least one high-frequency symptom combination, strong rules associated with the target high-risk disease are determined to construct the association rule base.

[0060] Specifically, historical triage data includes complete medical records, triage decisions, and final diagnoses of past emergency room patients. Based on this data, a feature dataset is constructed. This involves creating a standardized and structured set of patient features through preliminary data processing steps, where each record represents a patient's multi-dimensional clinical information. Subsequently, data mining techniques (such as the FP-Growth algorithm) are used to scan the feature dataset, efficiently counting the occurrence of all symptoms, signs, and other features, and extracting at least one high-frequency symptom combination (i.e., a frequent itemset) that meets preset frequency conditions. For example, "chest pain" and "dyspnea" often occur simultaneously.

[0061] Building upon this foundation, multiple scans based on frequent itemsets are performed using algorithms such as Apriori to iteratively calculate the support and confidence of these high-frequency combinations. Finally, based on preset support and confidence thresholds, association patterns that are both frequent and highly predictive are selected to identify strong rules associated with the target high-risk disease. For example, the rule "chest pain radiating to the left shoulder → acute myocardial infarction (confidence 85%)" is generated.

[0062] Furthermore, after determining strong rules associated with the target high-risk disease based on the support and confidence of the at least one high-frequency symptom combination to construct the association rule base, the method further includes: New triage data is acquired according to a preset period, and strong rule mining is re-executed based on the new triage data to update the association rule base.

[0063] Specifically, after completing the initial construction and identifying a series of strong rules associated with the target high-risk disease based on historical triage data, this embodiment introduces a dynamic update mechanism. The system automatically acquires new triage data according to a preset cycle (e.g., monthly, quarterly). This data originates from the latest real cases processed by the system, reflecting the latest treatment situations and potential new symptom-disease association patterns. Subsequently, the system re-executes strong rule mining based on the new triage data. This means re-running the entire construction process (including data preprocessing, frequent itemset extraction, support and confidence calculation, etc.) on a new, larger dataset. Through this process, the system can discover new association rules arising from seasonal disease prevalence, new treatment findings, or changes in population characteristics, while also validating or eliminating old rules that are no longer applicable due to updated treatment guidelines. Finally, by incorporating newly mined strong rules that conform to current data statistical characteristics into the knowledge system, the original association rule base is iteratively updated. This mechanism ensures that the system maintains the cutting-edge nature, accuracy, and clinical relevance of its triage recommendations over the long term.

[0064] The triage process of the association rule base is described in detail in one embodiment below.

[0065] Figure 2 This is a schematic diagram of the chief complaint recommendation process provided in the embodiments of this application, such as... Figure 2 As shown, the construction of the association rule base comprehensively utilizes two classic algorithms: Apriori and FP-growth. The Apriori section on the left presents a simplified example of an emergency triage dataset, where Uid represents the patient's unique identifier, and the Terms column records the patient's set of chief complaints, such as "low blood pressure, palpitations, hematemesis." The algorithm generates and tests candidate itemsets by repeatedly scanning such transaction datasets to find all frequent itemsets that meet the minimum support (i.e., frequency of occurrence) threshold. For example, the table clearly shows that the combination "black stool, dizziness" appears 3 times. The FP-growth section on the right illustrates another, more efficient method: First, the algorithm scans the original data (i.e., the list of transactions, where letters A, B, C, etc., are used to abstractly represent different symptoms in the diagram), counts the frequency of each item, sorts them, and generates a linked list of nodes (e.g., B:8 means symptom B appears 8 times in all transactions). Next, it constructs an FP-tree to compress and store the entire dataset, where each node represents a symptom and records its occurrence count along that path. Finally, by recursively traversing the FP-tree, all frequent itemsets and their support (Sup) can be quickly mined; for example, the table shows that the support of itemset {A,B} is a specific value. In actual construction of the association rule base, the FP-growth algorithm is first used to efficiently generate frequent itemsets, and then the idea of ​​the Apriori algorithm is used to iteratively calculate the support and confidence based on these itemsets, thereby filtering out strong association rules. The two complement each other, jointly completing the mining process from raw symptom data to usable triage knowledge rules.

[0066] The triage-assisted decision-making method provided in this embodiment transforms massive, discrete historical triage experiences into clear, quantifiable, and directly accessible structured knowledge that can be called by computers through a systematic data mining process.

[0067] Based on the above embodiments, in this embodiment, step 200 involves matching the feature information with a preset association rule base. If the match is successful, supplementary information about the target patient is obtained, including: Step 210: Compare the chief complaint symptoms in the feature information with the rule antecedents in the association rule base to generate a comparison result; wherein, the chief complaint symptoms are the discomfort symptoms actively stated by the target patient, and the rule antecedents are the condition part of the rule.

[0068] Step 220: If the comparison result indicates that the comparison is successful, trigger a supplementary questioning process for the target patient; wherein, the supplementary questioning process is to automatically push at least one key chief complaint follow-up question corresponding to the triggered rule.

[0069] Step 230: Receive the response results to the follow-up questions on the at least one key chief complaint, and use the response results as the supplementary information.

[0070] Specifically, the system extracts the patient's chief complaint (i.e., the discomfort actively stated by the patient, such as "abdominal pain" or "dizziness") from the current patient's characteristic information and compares it in real time with the antecedents (i.e., the combination of symptom conditions contained in the "if" part) of each rule stored in a pre-built association rule base. For example, if there is a rule in the rule base that reads "Rule antecedent: abdominal pain + hematemesis → conclusion: high risk of upper gastrointestinal bleeding," the system will check whether the patient's current chief complaint includes "abdominal pain" and "hematemesis." The comparison result indicates whether there is a complete or partial match.

[0071] When the comparison result indicates a successful match, it means that the patient's current presentation has triggered a high-risk association rule, but the information may not be complete. For example, the patient may only have said "abdominal pain" without mentioning "vomiting blood." In this case, the system will immediately trigger a follow-up questioning process for that target patient. This process is not random; instead, it automatically pushes at least one key chief complaint follow-up question corresponding to the triggered rule, aiming to verify or eliminate key conditions in the rule. Continuing with the example above, the system will automatically push the follow-up question to the triage nurse or patient: "Is there any vomiting blood or melena?" This key chief complaint follow-up question is directly derived from the rule logic and has strong clinical relevance.

[0072] Finally, the system receives the patient's or medical staff's responses to follow-up questions. For example, if the patient answers "I have black stools," this confirmed or newly added key information is treated as supplementary information. This process simulates the clinical questioning thinking of experienced doctors, transforming a one-off, potentially incomplete initial triage into a dynamic, progressively more comprehensive, interactive diagnostic process. This effectively bridges the information gap in initial doctor-patient communication and provides crucial incremental information for subsequent, more accurate model-based decisions.

[0073] The triage-assisted decision-making method provided in this embodiment enables real-time optimization and enrichment of initial features during the golden time of triage, transforming the traditional passive and static data processing mode into an active and interactive information gain process. This significantly reduces the information gap between doctors and patients, providing a solid information guarantee for subsequent models to make more accurate and reliable decisions, and fundamentally reducing the risk of misclassification due to information omissions.

[0074] Based on the above embodiments, in this embodiment, the triage model includes a triage prediction module and a severe illness prediction module; the auxiliary decision-making information includes triage level and severe illness risk prediction results; Step 300 inputs the updated feature information into the pre-trained triage model and outputs the auxiliary decision-making information for the target patient, including: The updated feature information is input into the triage prediction module and the severe illness prediction module respectively, and the triage level of the target patient and the severe illness risk prediction result of the target patient are output.

[0075] Specifically, the triage model consists of two well-defined, collaborative sub-modules: a triage prediction module and a critical illness prediction module. The core task of the triage prediction module is to categorize patients' urgency and resource needs based on emergency triage criteria. Its direct output is the triage level, such as classifying patients as "Level 1 (critical)," "Level 2 (severe)," "Level 3 (emergency)," or "Level 4 (non-emergency)," providing a direct basis for prioritizing the allocation of nursing resources. Meanwhile, the critical illness prediction module focuses on assessing the risk of a patient's condition deteriorating to a critical state or experiencing adverse outcomes in the short term. Its output is a quantitative prediction of critical illness risk, typically expressed as a probability value (e.g., 0.85) or a risk level (e.g., "high risk"), aiming to provide early warnings for patients requiring close monitoring or early intervention.

[0076] When making decisions, the system inputs the updated feature information, supplemented by previous information, into the triage prediction module and the critical illness prediction module in parallel, and outputs two core auxiliary decision-making information simultaneously: one is the triage level of the target patient from the triage prediction module, which is related to the priority of medical treatment; the other is the critical illness risk prediction result of the target patient from the critical illness prediction module, which is related to the prognostic risk warning.

[0077] The triage-assisted decision-making method provided in this embodiment effectively compensates for the lack of in-depth risk information that may be overlooked at a single triage level, and significantly improves the comprehensiveness, foresight, and patient safety protection capabilities of triage decisions.

[0078] Based on the above embodiments, in this embodiment, the hierarchical prediction module is configured to adopt a multi-model integration architecture, which integrates at least the following four neural network models: text convolutional neural network, bidirectional long short-term memory network, occlusion correction pre-trained language model, and meta-controller pre-trained language model.

[0079] The step of inputting the updated feature information into the triage prediction module and outputting the triage level of the target patient includes: Based on the updated feature information, real-time load, and data features of the updated feature information, the weights of the integrated neural network model output results are dynamically adjusted by the hierarchical prediction module to jointly determine the triage level of the target patient.

[0080] Specifically, when the system throughput requirement exceeds a preset threshold, the weights of the text convolutional neural network model are increased, while the weights of the bidirectional long short-term memory network model are decreased.

[0081] It should be noted that the hierarchical prediction module is a multi-model ensemble architecture that integrates multiple advanced neural networks and has intelligent weight adjustment capabilities. It incorporates at least four neural networks with complementary characteristics: Text Convolutional Neural Network (TextCNN) is good at extracting key local features from text sequences; Bidirectional Long Short-Term Memory Network (Bi-LSTM) is good at understanding the context and temporal dependencies of symptom descriptions; and the Masking Correction Pre-trained Language Model (MacBert) and Meta-Controller Pre-trained Language Model (MCBert), as large-scale pre-trained models, have powerful general semantic understanding and deep contextual representation capabilities, and can better handle the complex semantics and rare expressions of medical terms.

[0082] In specific decision-making, the module's intelligence is reflected in its dynamic weight adjustment mechanism. It doesn't simply fix the voting weights of each model, but rather performs real-time calculations based on three factors: first, the content and complexity of the updated feature information itself; second, the system's current real-time load (such as concurrent requests and computing resource utilization); and finally, the data characteristics presented by this feature information (such as text length, number of entities, and distribution of numerical features). Based on a comprehensive analysis of these factors, the module dynamically adjusts the weights of the output results of the four integrated models.

[0083] A typical strategy is as follows: during peak periods when system throughput demand exceeds a preset threshold or when resources are strained, in order to improve processing speed, the system increases the weights of text convolutional neural network models that are computationally efficient and adept at quickly capturing local patterns, while decreasing the weights of bidirectional long short-term memory network models that are computationally more time-consuming and focus on long sequence modeling. Through this adaptive adjustment, the advantages of each model are optimally balanced between accuracy and efficiency, ultimately outputting an optimal triage level for the target patient that integrates the wisdom of multiple parties and adapts to the system state at that time, in a jointly determined manner.

[0084] The following embodiment describes the specific implementation process of the hierarchical prediction module.

[0085] Figure 3 This is a schematic diagram of the hierarchical prediction process provided in the embodiments of this application, such as... Figure 3 As shown, this paper illustrates the technical principles and division of labor among the four core neural network models in the multi-model ensemble architecture adopted by the graded prediction module of this application. Specifically, Text CNN (Text Convolutional Neural Network) focuses on extracting local key features from the patient's chief complaint text through its convolutional layer structure, such as... Figure 3As shown, it focuses on specific words (such as symptom descriptors) and their local combinations; Bi-LSTM (Bidirectional Long Short-Term Memory Network) models the context and temporal dependencies of symptom descriptions through its forward and backward LSTM layers, such as understanding sequence information that negates or describes changes in the condition, like "no cardiac enlargement"; MC-BERT (Meta-Controller BERT) and BERT are both pre-trained language models based on the Transformer architecture. BERT learns deep, general semantic representations through pre-training tasks such as Masked Language Modeling (Masked LM) and Next Sentence Prediction (NSP), while MC-BERT further introduces a meta-controller mechanism to dynamically predict and sample words, enhancing the model's adaptability and generative ability to specific domains (such as medical text). The Fine-Tuning section at the bottom of the figure indicates that these pre-trained models will be fine-tuned using high-quality medical feature data in emergency triage scenarios to ensure that their semantic understanding capabilities accurately match the needs of triage decisions. Each of these four models has its own advantages: TextCNN is good at capturing local symptom keywords, Bi-LSTM is good at analyzing the temporal logic of symptoms, while MacBert and MCBert, as powerful pre-trained models, provide deep semantic understanding and generalization capabilities. Their integration together constitutes a robust hierarchical prediction engine that can analyze patient condition information from multiple dimensions and levels.

[0086] The triage-assisted decision-making method provided in this embodiment systematically and significantly improves the accuracy, reliability, and clinical applicability of emergency intelligent triage through multi-level technological innovation.

[0087] Based on the above embodiments, in this embodiment, the critical illness prediction module is a gradient boosting decision tree model; the auxiliary decision information also includes an interpretable report corresponding to the critical illness risk prediction result; The step of inputting the updated feature information into the pre-trained triage model and outputting the auxiliary decision-making information for the target patient further includes: The updated feature information is input into the critical illness prediction module, and the contribution of each feature in the updated feature information to the critical illness risk prediction result is calculated. Based on the contribution of each feature to the prediction of the severe illness risk, an interpretability report is generated that explains the degree of influence of each feature.

[0088] Specifically, the critical illness prediction module uses a gradient boosting decision tree model (such as XGBoost), which is a powerful ensemble learning algorithm. By sequentially building multiple decision trees, each new tree is dedicated to correcting the prediction error of the previous tree, thus enabling efficient modeling of complex nonlinear relationships and high-order interactive features. It is very suitable for handling mixed features including numerical physiological indicators and structured medical entities, and outputs a critical illness risk prediction probability value between 0 and 1.

[0089] To ensure traceability and verifiability of decision-making, this embodiment requires the generation of a corresponding interpretability report along with the output risk value. To this end, after calculating the risk prediction result, the system further calculates the contribution of each feature in the updated feature information to the critical illness risk prediction result. This is typically achieved through advanced model interpretation techniques such as Shapley Gain Interpretation (SHAP), which, based on cooperative game theory, fairly allocates the merit or fault of each feature (such as "age 65 years old," "systolic blood pressure 180 mmHg," and "chief complaint of chest pain") to the final predicted value relative to the baseline average, thereby quantifying its contribution and identifying which key factors increased or decreased the risk probability. For example, the calculation might show that "age" contributes a +0.15 risk increase, while "blood oxygen saturation" contributes a -0.08 risk decrease. Finally, based on the contribution of each feature, the system automatically generates a structured and interpretable report. This report clearly explains the degree of influence of each feature, for example, by presenting a list of feature importance, contribution values, or visual charts, intuitively showing that "advanced age," "low blood pressure," and "complaint of shortness of breath" are the top three driving factors predicted to be high-risk.

[0090] The following embodiment describes the specific implementation process of the critical illness prediction module.

[0091] Figure 4 This is a schematic diagram of the critical illness prediction process provided in the embodiments of this application, such as... Figure 4 As shown, the severe illness prediction module and its interpretability engine are explained: This module is based on high-performance gradient boosting decision tree models such as CatBoost, XGBoost, and LightGBM. Through learning from a large number of structured features, it outputs accurate severe illness risk prediction results. To achieve decision transparency, the system introduces the SHAP interpretation framework, such as... Figure 4 As shown, it attributes the difference between the predicted value of a single sample (e.g., 0.4) and the base rate (e.g., 0.1) of the entire sample to the contribution of each feature (e.g., Age=65, BP=180) (e.g., +0.3, +0.15), thereby quantitatively revealing how each feature affects the final risk assessment and generating an intuitive and interpretable report.

[0092] The triage-assisted decision-making method provided in this embodiment uses a gradient boosting decision tree model as the core of the severe illness prediction module, which enables it to efficiently and accurately capture the complex nonlinear relationships between high-dimensional features and output a quantified probability of severe illness risk, providing key risk warning information for triage.

[0093] Figure 5 This is a complete flowchart of the triage auxiliary decision-making method provided in the embodiments of this application. The following is a combination of... Figure 5 This application provides a complete description of the triage-assisted decision-making method provided in the embodiments.

[0094] like Figure 5 As shown, the overall input is divided into prospective patient information (i.e., real-time triage data and basic information of the current patients) and retrospective patient information (i.e., rich historical medical record data used for model training and knowledge mining). This accurately corresponds to the reality in emergency scenarios where the acquisition of new patient information is limited, but the analysis of historical data is in-depth.

[0095] The core development system executes four key modules sequentially: First, the data processing module performs three-level cleaning, terminology standardization, medical entity recognition based on Bi-LSTM-CRF, and data augmentation through adversarial training on multi-source heterogeneous data to construct high-quality feature representations; next, the chief complaint recommendation module uses FP-Growth and Apriori algorithms to mine association rules for features. When real-time data triggers high-risk rules, the system proactively generates and pushes key follow-up questions, enabling intelligent interaction with an additional round of chief complaint questioning; subsequently, the severe illness prediction model (using XGBoost and other gradient boosting decision trees) and the grading prediction model (using TextCNN, Bi-LSTM, MacBert, and MCBert multi-model ensemble) process the updated feature information in parallel, outputting quantified severe illness probabilities and predicted grades respectively; finally, the severe illness risk results can also be fed back to the interaction stage to further optimize the questioning strategy.

[0096] The triage-assisted decision-making method provided in this application constructs a high-quality feature engineering system from the source through an original three-level cleaning and fine-grained standardization process combined with precise medical entity recognition based on Bi-LSTM-CRF. By integrating FP-Growth and Apriori algorithms for dynamic association rule mining, it automatically extracts high-risk warning knowledge from massive historical data and can proactively trigger targeted supplementary questions, effectively bridging the information gap between doctors and patients. By integrating four neural networks—TextCNN, Bi-LSTM, MacBert, and MCBert—and introducing a dynamic weight adjustment mechanism, the triage prediction combines local feature capture, temporal logic analysis, and deep semantic understanding capabilities, significantly improving the accuracy and environmental adaptability of triage. Simultaneously, by employing the XGBoost model for severe case risk prediction and coupling it with the SHAP interpretability framework, it provides a clear and transparent feature contribution report while outputting high-precision risk quantification, making AI decision-making a reliable decision-making partner for clinicians.

[0097] The distributed digital identity resolution device provided in the embodiments of this application is described below. The distributed digital identity resolution device described below can be referred to in correspondence with the distributed digital identity resolution method described above.

[0098] Figure 6 This is a schematic diagram of the triage auxiliary decision-making device provided in the embodiments of this application, as shown below. Figure 6 As shown in the embodiment of this application, the triage auxiliary decision-making device includes: The acquisition module 601 is used to acquire the medical data of the target patient, process the medical data, and extract feature information that characterizes the condition of the target patient. The update module 602 is used to match the feature information with a preset association rule base. If the match is successful, it obtains supplementary information of the target patient and updates the feature information according to the supplementary information. The decision module 603 is used to input the updated feature information into the pre-trained triage model and output the auxiliary decision information of the target patient; wherein the triage model is trained based on historical triage data.

[0099] The triage auxiliary decision-making device provided in this application embodiment acquires the medical data of a target patient, processes the medical data, and extracts feature information characterizing the patient's condition. Based on this feature information, it matches it with a pre-set association rule base. If a match is successful, it acquires supplementary information about the target patient and updates the feature information accordingly. The updated feature information is then input into a pre-trained triage model, outputting auxiliary decision-making information for the target patient. The triage model is trained based on historical triage data. Therefore, this application embodiment, through real-time matching of the initially extracted feature information using a pre-set association rule base, can proactively identify information gaps in the initial triage judgment and trigger a targeted information supplementation process. This dynamically and purposefully guides the collection of key clinical information, effectively compensating for information gaps caused by unclear patient complaints or limited physical signs. This significantly improves the quality of the feature information input into the triage model, resulting in more reliable auxiliary decision-making suggestions and reducing the risk of misclassification due to insufficient information throughout the overall process.

[0100] Based on the above embodiments, in this embodiment, the medical data includes numerical physiological indicators and textual chief complaint data; The acquisition module 601 is specifically used for: For the numerical physiological indicators, outlier detection and removal are performed, and missing value imputation is performed on the data after outlier removal to obtain the first feature sub-information. For the text-based chief complaint data, terminology consistency standardization processing is performed to obtain the second feature sub-information; The feature information is constructed by combining the first feature sub-information and the second feature sub-information.

[0101] Based on the above embodiments, in this embodiment, the device further includes a processing module, specifically used for: For the text-based chief complaint data, terminology consistency standardization processing is performed to obtain standardized chief complaint data; The standardized chief complaint data is input into a pre-trained medical entity recognition model, which outputs the identified symptoms, signs, and disease entities as the second feature sub-information; wherein, the medical entity recognition model is constructed based on a bidirectional long short-term memory network and a conditional random field framework.

[0102] Based on the above embodiments, in this embodiment, the processing module is further specifically used for: The first feature sub-information and the second feature sub-information are combined, and feature engineering processing is performed on the combined data to generate the feature information; The feature engineering process includes: segmenting and discretizing numerical features, and performing word frequency-inverse document frequency weighting and encoding conversion on textual entity features.

[0103] Based on the above embodiments, in this embodiment, the device further includes a construction module, specifically used for: Obtain the historical triage data and construct a feature dataset based on the historical triage data; The feature dataset is scanned to extract at least one high-frequency symptom combination that meets the preset frequency conditions, forming a frequent itemset; Based on the frequent itemset, multiple scans are performed to iteratively calculate the support and confidence of the at least one high-frequency symptom combination; Based on the support and confidence of the at least one high-frequency symptom combination, strong rules associated with the target high-risk disease are determined to construct the association rule base.

[0104] Based on the above embodiments, in this embodiment, the update module 602 is specifically used for: The chief complaint symptoms in the feature information are compared with the rule antecedents in the association rule base to generate a comparison result; wherein, the chief complaint symptoms are the discomfort symptoms actively stated by the target patient, and the rule antecedents are the condition part of the rule; If the comparison result indicates a successful comparison, a supplementary questioning process for the target patient is triggered; wherein, the supplementary questioning process is to automatically push at least one key chief complaint follow-up question corresponding to the triggered rule; Receive the response results to the at least one key chief complaint inquiry, and use the response results as the supplementary information.

[0105] Based on the above embodiments, in this embodiment, the triage model includes a triage prediction module and a severe illness prediction module; the auxiliary decision-making information includes triage level and severe illness risk prediction results; The decision module 603 is specifically used for: The updated feature information is input into the triage prediction module and the severe illness prediction module respectively, and the triage level of the target patient and the severe illness risk prediction result of the target patient are output.

[0106] Based on the above embodiments, in this embodiment... The hierarchical prediction module is configured to adopt a multi-model integration architecture, which integrates at least the following four neural network models: text convolutional neural network, bidirectional long short-term memory network, occlusion correction pre-trained language model, and meta-controller pre-trained language model.

[0107] Based on the above embodiments, in this embodiment, the decision module 603 is further configured to: Based on the updated feature information, real-time load, and data features of the updated feature information, the weights of the integrated neural network model output results are dynamically adjusted by the hierarchical prediction module to jointly determine the triage level of the target patient.

[0108] Based on the above embodiments, in this embodiment, the device further includes an adjustment module, specifically used for: When the system throughput requirement exceeds a preset threshold, the weights of the text convolutional neural network model are increased, and the weights of the bidirectional long short-term memory network model are decreased.

[0109] Based on the above embodiments, in this embodiment, the critical illness prediction module is a gradient boosting decision tree model; the auxiliary decision information also includes an interpretable report corresponding to the critical illness risk prediction result; The decision module 603 is further used for: The updated feature information is input into the critical illness prediction module, and the contribution of each feature in the updated feature information to the critical illness risk prediction result is calculated. Based on the contribution of each feature to the prediction of the severe illness risk, an interpretability report is generated that explains the degree of influence of each feature.

[0110] Based on the above embodiments, in this embodiment, the device further includes an update module, specifically used for: After determining strong rules associated with the target high-risk disease based on the support and confidence of the at least one high-frequency symptom combination, and constructing the association rule base, New triage data is acquired according to a preset period, and strong rule mining is re-executed based on the new triage data to update the association rule base.

[0111] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device can be a robot or other electronic device. This electronic device may include: a processor 710, a communications interface 720, a memory 730, and a communication bus 740. The processor 710, communications interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions from the memory 730 to execute a triage-aid decision-making method, including: Acquire the medical data of the target patient, process the medical data, and extract feature information that characterizes the condition of the target patient; The feature information is matched with a pre-set association rule base. If the match is successful, the supplementary information of the target patient is obtained, and the feature information is updated according to the supplementary information. The updated feature information is input into the pre-trained triage model, and the auxiliary decision-making information of the target patient is output; wherein, the triage model is trained based on historical triage data.

[0112] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application embodiment, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in at least one embodiment of this application embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] On the other hand, embodiments of this application also provide a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the triage auxiliary decision-making methods provided by the above methods, including: Acquire the medical data of the target patient, process the medical data, and extract feature information that characterizes the condition of the target patient; The feature information is matched with a pre-set association rule base. If the match is successful, the supplementary information of the target patient is obtained, and the feature information is updated according to the supplementary information. The updated feature information is input into the pre-trained triage model, and the auxiliary decision-making information of the target patient is output; wherein, the triage model is trained based on historical triage data.

[0114] In another aspect, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the triage assistance decision-making methods provided by the methods described above, including: Acquire the medical data of the target patient, process the medical data, and extract feature information that characterizes the condition of the target patient; The feature information is matched with a pre-set association rule base. If the match is successful, the supplementary information of the target patient is obtained, and the feature information is updated according to the supplementary information. The updated feature information is input into the pre-trained triage model, and the auxiliary decision-making information of the target patient is output; wherein, the triage model is trained based on historical triage data.

[0115] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of this application, and are not intended to limit them; although the embodiments of this application have been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A triage-assisted decision-making method, characterized in that, include: Acquire the medical data of the target patient, process the medical data, and extract feature information that characterizes the condition of the target patient; The feature information is matched with a pre-set association rule base. If the match is successful, the supplementary information of the target patient is obtained, and the feature information is updated according to the supplementary information. The updated feature information is input into the pre-trained triage model, and the auxiliary decision-making information of the target patient is output; wherein, the triage model is trained based on historical triage data.

2. The triage-assisted decision-making method according to claim 1, characterized in that, The medical data includes numerical physiological indicators and text-based chief complaint data; The process of processing the medical data to extract feature information characterizing the condition of the target patient includes: For the numerical physiological indicators, outlier detection and removal are performed, and missing value imputation is performed on the data after outlier removal to obtain the first feature sub-information. For the text-based chief complaint data, terminology consistency standardization processing is performed to obtain the second feature sub-information; The feature information is constructed by combining the first feature sub-information and the second feature sub-information.

3. The triage-assisted decision-making method according to claim 2, characterized in that, The terminology consistency standardization process is performed on the text-based chief complaint data to obtain the second feature sub-information, including: For the text-based chief complaint data, terminology consistency standardization processing is performed to obtain standardized chief complaint data; The standardized chief complaint data is input into a pre-trained medical entity recognition model, which outputs the identified symptoms, signs, and disease entities as the second feature sub-information; wherein, the medical entity recognition model is constructed based on a bidirectional long short-term memory network and a conditional random field framework.

4. The triage-assisted decision-making method according to claim 2, characterized in that, The process of constructing the feature information using the first feature sub-information and the second feature sub-information includes: The first feature sub-information and the second feature sub-information are combined, and feature engineering processing is performed on the combined data to generate the feature information; The feature engineering process includes: segmenting and discretizing numerical features, and performing word frequency-inverse document frequency weighting and encoding conversion on textual entity features.

5. The triage-assisted decision-making method according to claim 1, characterized in that, The construction process of the pre-set association rule base includes: Obtain the historical triage data and construct a feature dataset based on the historical triage data; The feature dataset is scanned to extract at least one high-frequency symptom combination that meets the preset frequency conditions, forming a frequent itemset; Based on the frequent itemset, multiple scans are performed to iteratively calculate the support and confidence of the at least one high-frequency symptom combination; Based on the support and confidence of the at least one high-frequency symptom combination, strong rules associated with the target high-risk disease are determined to construct the association rule base.

6. The triage-assisted decision-making method according to claim 1, characterized in that, The process involves matching the feature information with a pre-set association rule base. If a match is successful, supplementary information about the target patient is obtained, including: The chief complaint symptoms in the feature information are compared with the rule antecedents in the association rule base to generate a comparison result; wherein, the chief complaint symptoms are the discomfort symptoms actively stated by the target patient, and the rule antecedents are the condition part of the rule; If the comparison result indicates a successful comparison, a supplementary questioning process for the target patient is triggered; wherein, the supplementary questioning process is to automatically push at least one key chief complaint follow-up question corresponding to the triggered rule; Receive the response results to the at least one key chief complaint inquiry, and use the response results as the supplementary information.

7. The triage-assisted decision-making method according to claim 1, characterized in that, The triage model includes a triage prediction module and a severe illness prediction module; the auxiliary decision-making information includes triage level and severe illness risk prediction results. The step of inputting the updated feature information into the pre-trained triage model and outputting the auxiliary decision-making information for the target patient includes: The updated feature information is input into the triage prediction module and the severe illness prediction module respectively, and the triage level of the target patient and the severe illness risk prediction result of the target patient are output.

8. The triage-assisted decision-making method according to claim 7, characterized in that, The hierarchical prediction module is configured to adopt a multi-model integration architecture, which integrates at least the following four neural network models: text convolutional neural network, bidirectional long short-term memory network, occlusion correction pre-trained language model, and meta-controller pre-trained language model.

9. The triage-assisted decision-making method according to claim 8, characterized in that, The step of inputting the updated feature information into the triage prediction module and outputting the triage level of the target patient includes: Based on the updated feature information, real-time load, and data features of the updated feature information, the weights of the integrated neural network model output results are dynamically adjusted by the hierarchical prediction module to jointly determine the triage level of the target patient.

10. The triage-assisted decision-making method according to claim 9, characterized in that, The weights of the integrated neural network model output results are dynamically adjusted by the hierarchical prediction module based on the updated feature information, real-time load, and data features of the updated feature information, including: When the system throughput requirement exceeds a preset threshold, the weights of the text convolutional neural network model are increased, and the weights of the bidirectional long short-term memory network model are decreased.

11. The triage-assisted decision-making method according to claim 7, characterized in that, The critical illness prediction module is a gradient boosting decision tree model; the auxiliary decision information also includes an interpretable report corresponding to the critical illness risk prediction results; The step of inputting the updated feature information into the pre-trained triage model and outputting the auxiliary decision-making information for the target patient further includes: The updated feature information is input into the critical illness prediction module, and the contribution of each feature in the updated feature information to the critical illness risk prediction result is calculated. Based on the contribution of each feature to the prediction of the severe illness risk, an interpretability report is generated that explains the degree of influence of each feature.

12. The triage-assisted decision-making method according to claim 5, characterized in that, After determining strong rules associated with the target high-risk disease based on the support and confidence of the at least one high-frequency symptom combination to construct the association rule base, the method further includes: New triage data is acquired according to a preset period, and strong rule mining is re-executed based on the new triage data to update the association rule base.

13. A triage auxiliary decision-making device, characterized in that, include: The acquisition module is used to acquire the medical data of the target patient, process the medical data, and extract feature information that characterizes the condition of the target patient. The update module is used to match the feature information with a preset association rule base. If the match is successful, it obtains supplementary information of the target patient and updates the feature information according to the supplementary information. The decision module is used to input the updated feature information into the pre-trained triage model and output the auxiliary decision information of the target patient; wherein the triage model is trained based on historical triage data.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the triage auxiliary decision-making method as described in any one of claims 1 to 12.

15. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the triage auxiliary decision-making method as described in any one of claims 1 to 12.

16. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the triage auxiliary decision-making method according to any one of claims 1 to 12.

Images