Intelligent triage system and method for emergency treatment grading

Abstract

The invention provides an intelligent triage system for emergency treatment grading and a method thereof. The intelligent triage system comprises a patient end, a nurse end, a physician end and a cloud end, the patient end is configured to obtain multi-dimensional vital sign information of an emergency patient and generate symptom expression description according to the vital sign information; the cloud end is configured to construct a symptom dependency network based on a preset symptom association relationship according to symptom expression description, past medical history abstracts of emergency patients and treatment time information, perform hierarchical emergency degree adjustment on the symptom dependency network, update the emergency degree of each symptom item through influence conduction among nodes, generate a dynamic emergency degree feature set, and send the dynamic emergency degree feature set to the cloud end; inputting the dynamic urgency degree feature set into a pre-trained triage decision model, generating a grading evaluation result including grading grades and intervention timeliness requirements, and determining recommended disposal departments and triage schemes including treatment priorities and necessary examination items according to the grading evaluation result; the nurse end and the doctor end are configured to receive and display the triage scheme.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to an intelligent triage system and method for emergency triage. Background Technology

[0002] In traditional emergency triage scenarios, patient classification and triage primarily rely on the experience and judgment of medical staff. Medical staff obtain descriptions of symptoms by communicating with patients or accompanying persons, manually record vital sign data, and retrieve past medical history information from paper medical records or limited information systems. However, this manual information collection method is inefficient and prone to omissions or inaccuracies.

[0003] In symptom analysis, healthcare professionals often rely on personal experience to trace relationships between symptoms, lacking a systematic and objective approach, making it difficult to fully and accurately grasp the complex connections between symptoms. Assessing the urgency of symptoms is often based on simplistic rules and experience, failing to consider the mutual influence and transmission between symptoms, resulting in inaccurate urgency assessments.

[0004] In the triage decision-making process, due to the lack of scientific models and algorithms, triage results are easily influenced by the subjective factors of medical staff. Different medical staff may arrive at different triage plans, making it difficult to guarantee the consistency and accuracy of triage. Moreover, traditional triage methods cannot be dynamically adjusted in real time based on changes in patients' vital signs and the availability of emergency resources, failing to meet the needs of modern emergency medicine for efficiency, precision, and dynamism. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent triage system and method for emergency triage.

[0006] To achieve the above objectives, a first aspect of this disclosure provides an intelligent triage system for emergency department stratification, comprising: a patient terminal, a nurse terminal, a physician terminal, and a cloud terminal for communicative connection with the patient terminal, the nurse terminal, and the physician terminal; the patient terminal is configured to acquire multi-dimensional vital sign information of an emergency patient and generate a symptom description of the emergency patient based on the vital sign information; the cloud terminal is configured to construct a symptom dependency network based on a preset symptom association relationship, according to the symptom description of the emergency patient uploaded by the patient terminal, a summary of the emergency patient's past medical history, and consultation time information received from the patient terminal, wherein nodes in the symptom dependency network correspond to specific symptom items, and edges... The attributes include the probability of symptom co-occurrence and the strength of causal influence. A hierarchical urgency adjustment is performed on the symptom dependency network, updating the urgency of each symptom item through inter-node influence transmission, generating a dynamic urgency feature set. This dynamic urgency feature set is input into a pre-trained triage decision model, which generates a graded assessment result including grading levels and intervention timeliness requirements through a dual-track inference mechanism. Based on the graded assessment result, the recommended treatment department is determined, along with a triage plan including consultation priority and necessary examination items. The nurse's end and the physician's end are configured to receive and display the triage plan sent from the cloud, corresponding to the emergency patient and including the consultation priority and necessary examination items.

[0007] A second aspect of this disclosure provides an intelligent triage method for emergency department stratification, applied in a cloud environment, wherein the cloud environment is the same as described in any one of the first aspects; the intelligent triage method for emergency department stratification includes: Based on the symptom description, medical history summary, and consultation time information of the emergency patient uploaded by the patient's end, a symptom dependency network is constructed based on preset symptom associations. Nodes in the symptom dependency network correspond to specific symptom items, and the attributes of the edges include the probability of symptom co-occurrence and the strength of causal influence. The symptom description is generated by the patient's end based on the multi-dimensional vital sign information of the emergency patient. Hierarchical priority calculation is performed on the symptom dependency network, and the urgency score of each symptom item is updated through the influence transmission between nodes to generate a dynamic urgency feature set. The dynamic urgency feature set is input into a pre-trained triage decision model, and a graded assessment result including grade level and intervention timeliness requirements is generated through a dual-track reasoning mechanism. Based on the graded assessment result, the recommended treatment department and a triage plan including consultation priority and necessary examination items are determined. The triage plan is pushed by the cloud to the nurse's and physician's ends of the corresponding recommended treatment department to display the consultation priority and necessary examination items on the nurse's and physician's ends.

[0008] This invention provides an intelligent triage system and method for emergency department grading. Compared with existing technologies, it has the following advantages: The system automatically collects multi-dimensional vital sign data from the patient and generates structured symptom descriptions; the cloud integrates past medical history and consultation time stamps to construct a complete patient profile. Compared with traditional manual recording methods, it significantly reduces the risk of information omissions or errors. The cloud constructs a hierarchical dependency network based on symptom concurrency probability and causal strength, updating symptom urgency scores in real time through a node influence transmission mechanism. This model can quantify the dynamic correlation between complex symptoms, reducing the error rate of urgency assessment and achieving a qualitative leap from static rules to dynamic deduction. Simultaneously, the pre-trained model employs a dual-track inference mechanism: one track performs initial triage based on a symptom-department correlation graph, while the other track dynamically optimizes based on emergency department resource occupancy and intervention timeliness requirements. This improves the consistency of triage results and shortens the average triage time.

[0009] Furthermore, based on the triage assessment results, a triage plan including visit priority, recommended departments, and necessary examinations is generated in real time and synchronized to the nurse / physician's end. Through linkage with the emergency dispatch system, the triage strategy can be dynamically adjusted to match resource changes, improving emergency bed turnover and reducing examination redundancy. This achieves intelligent integration across the entire data-model-resource chain, realizing closed-loop optimization from information collection to resource scheduling, and providing efficient, accurate, and dynamically adaptable triage.

[0010] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0011] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a block diagram of an intelligent triage system for emergency care, as illustrated in the embodiments of the instruction manual.

[0012] Figure 2 This is a flowchart illustrating an intelligent triage method for emergency department grading, as shown in the embodiments of the instruction manual.

[0013] Figure 3 This is a block diagram of an intelligent triage device for emergency triage, as shown in the embodiments of the specification. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0016] This disclosure provides an intelligent triage system for emergency care. Figure 1 This is a block diagram illustrating an intelligent triage system for emergency care, according to one embodiment. Specifically, the intelligent triage system for emergency care includes: a patient terminal 100, a nurse terminal 200, a physician terminal 300, and a cloud terminal 400 for communicating with the patient terminal 100, the nurse terminal 200, and the physician terminal 300.

[0017] The system comprises the following components: Patient Terminal 100 (which may include at least one of a smart bracelet, smartphone, or fitness equipment) for collecting and initially processing patient information; Nurse Terminal 200 (which may be an authenticated terminal used by nurses, accessible via applications or mini-programs, for receiving triage plans and assisting with nursing care); Physician Terminal 300 (which may be an authenticated terminal used by physicians for receiving triage plans and conducting diagnostic and treatment procedures); and Cloud Terminal 400 (a server cluster located on the network, responsible for data storage, processing, and decision generation, enabling communication and collaboration between the terminals).

[0018] The patient terminal is configured to acquire multi-dimensional vital sign information of emergency patients and generate a description of the symptoms of the emergency patients based on the vital sign information. Among them, multidimensional vital sign information can cover multiple physiological indicators of the patient's body, such as body temperature, blood pressure, heart rate, respiratory rate, and blood oxygen saturation, to comprehensively assess the patient's physical condition. Symptom description is a detailed written summary and record of the various symptoms presented by the patient, such as the location, nature, and degree of pain, and whether there are symptoms such as fever, cough, and vomiting.

[0019] In this embodiment, the patient terminal acquires multi-dimensional vital sign information of emergency patients in real time through built-in sensors (such as a body temperature sensor, blood pressure monitor interface, heart rate monitoring module, etc.) or by connecting to external medical devices. Then, using preset algorithms and rules, this vital sign information is analyzed and integrated, combined with the patient's self-description (if any), to generate a description of the emergency patient's symptoms, and this information is uploaded to the cloud.

[0020] The cloud platform is configured to construct a symptom dependency network based on a preset symptom association relationship, using the symptom description, medical history summary, and consultation time information of the emergency patient uploaded by the patient terminal. Nodes in the symptom dependency network correspond to specific symptom items, and the attributes of the edges include the probability of symptom co-occurrence and the strength of causal influence. The platform performs hierarchical urgency adjustment on the symptom dependency network, updates the urgency of each symptom item through the influence transmission between nodes, generates a dynamic urgency feature set, and inputs the dynamic urgency feature set into a pre-trained triage decision model. Through a dual-track reasoning mechanism, it generates a graded assessment result including grade level and intervention timeliness requirements, and determines the recommended treatment department and a triage plan including consultation priority and necessary examination items based on the graded assessment result. Among them, symptom association is the inherent connection and pattern between different symptoms, including the co-occurrence relationship between symptoms (that is, when one symptom occurs, another symptom often occurs as well) and the causal relationship (that is, one symptom may lead to the occurrence of another symptom).

[0021] In this embodiment of the disclosure, by learning from and analyzing a large amount of historical emergency room case data, data mining and machine learning algorithms are used to identify association rules and patterns among various symptoms, and to establish symptom association models. These models can quantify the concurrent probability and causal influence strength among different symptoms.

[0022] Among them, the symptom dependency network is a network structure used to represent the relationships between symptoms. Nodes represent specific symptom items, and edges represent the associations between symptoms. The attributes of the edges include the probability of symptom concurrency and the strength of causal influence, which is used to intuitively display the complex relationships between symptoms.

[0023] In this embodiment, the cloud constructs a symptom dependency network based on the description of the emergency patient's symptoms, a summary of the patient's medical history, and the time of their visit, all uploaded by the patient. This is done in conjunction with pre-defined symptom associations. Specifically, each symptom is treated as a node, and edges are added between these nodes based on the relationships between symptoms. Each edge is then assigned a concurrency probability and a causal influence strength attribute. This creates a network structure that reflects the relationships between the patient's symptoms.

[0024] Hierarchical urgency adjustment is a process of dynamically adjusting the urgency of symptoms based on the hierarchical relationship and influence transmission mechanism between symptoms. By considering the causal influence and concurrent relationships between symptoms, the urgency of each symptom item is updated to more accurately reflect the actual urgency level of the patient.

[0025] In this embodiment of the disclosure, a hierarchical relationship and influence transmission path exist between different symptoms in the symptom dependency network. When the urgency of a symptom changes, it affects the urgency of other related symptoms through the influence transmission mechanism of edges. For example, if the urgency of a basic symptom increases, and this symptom is the cause of other symptoms, then the urgency of the affected symptoms will also increase accordingly. Through this hierarchical adjustment of urgency, the overall urgency of the patient can be assessed more comprehensively and accurately.

[0026] The dynamic urgency feature set is a feature set composed of the urgency of each symptom item after hierarchical urgency adjustment. It is used to describe the patient's current urgency status and provide a basis for subsequent triage decisions.

[0027] In this embodiment of the disclosure, after the hierarchical urgency adjustment of the symptom dependency network is completed, the urgency information of each symptom item is extracted, and this information is combined into a feature set, namely, a dynamic urgency feature set. This set can dynamically reflect the changes in the urgency of the patient's symptoms, providing more accurate and comprehensive input data for the triage decision model.

[0028] Among them, the pre-trained triage decision model is a machine learning model trained using a large amount of historical emergency case data and corresponding triage results. It can analyze and judge the input patient information and generate reasonable triage decisions.

[0029] In this embodiment, a large amount of historical emergency room case data is collected, including patients' vital signs, symptoms, past medical history, consultation time, and final triage results (such as triage level, intervention timeliness requirements, recommended treatment department, etc.). This data is divided into training and testing sets. The training set is used to train a machine learning model (such as a neural network, decision tree, support vector machine, etc.), adjusting the model's parameters to accurately learn the mapping relationship between patient information and triage results. Then, the testing set is used to evaluate and validate the trained model, ensuring high accuracy and generalization ability.

[0030] Among them, the dual-track reasoning mechanism is a decision generation mechanism that combines two different reasoning methods, typically including rule-based reasoning and data-based reasoning. By combining the advantages of the two reasoning methods, it improves the accuracy and reliability of triage decisions.

[0031] In this embodiment, the dual-track reasoning mechanism performs reasoning based on preset rules and a knowledge base. These rules are summarized from medical expertise and clinical experience; for example, certain symptom combinations correspond to specific triage levels or treatment departments. A pre-trained triage decision-making model performs reasoning based on data; by learning from a large amount of historical case data, the model can discover potential patterns and regularities. The dual-track reasoning mechanism comprehensively analyzes and weighs the results of these two reasoning methods to generate a tiered assessment result that includes both the triage level and the intervention timeliness requirements.

[0032] The triage assessment results are obtained by evaluating the patient's emergency level and the timeliness requirements of intervention. They are used to guide subsequent medical treatment and ensure that the patient receives timely and appropriate treatment.

[0033] In this embodiment, a tiered assessment result is generated through a dual-track reasoning mechanism. This result comprehensively considers multiple factors, including the patient's symptoms, urgency characteristics, and past medical history. The tiers are typically divided according to the severity of the patient's condition; for example, Level 1 is the most urgent, requiring immediate treatment; Level 2 is less urgent, requiring prompt intervention. The intervention timeliness requirement specifies the time limit from the patient's arrival at the clinic to the commencement of appropriate treatment, ensuring that the patient receives treatment within the optimal timeframe.

[0034] The recommended treatment department is the most suitable department for further diagnosis and treatment based on the patient's condition and symptoms, which helps to improve the efficiency of medical resource utilization and the patient's treatment outcome.

[0035] In this embodiment, the cloud platform determines the most suitable department for the patient's treatment based on the grading assessment results, combined with the scope of diagnosis and treatment and expertise of various departments in the hospital, using preset rules and algorithms. For example, if the patient's main symptom is chest pain and the grading assessment result is high-level, the system may recommend treatment by the cardiology department or the emergency department.

[0036] In one implementation, the grading level and intervention time limit requirements in the grading assessment results are analyzed to determine the patient's treatment priority. The treatment priority is positively correlated with the grading level and intervention time limit requirements. If a patient is rated as Level 1 and has a short intervention time limit, their treatment priority is determined to be the highest, and priority medical resources should be allocated to them.

[0037] Based on the core symptom characteristics in the grading assessment results, a pre-defined departmental division of labor guideline is retrieved, and the most suitable clinical department for treating this type of symptom is matched to generate a recommended treatment department list. According to information such as "retrosternal pain" and "abnormal heart rate" in the core symptom characteristics, the departmental division of labor guideline is retrieved, and departments such as cardiology and emergency medicine, which are suitable for treatment, are matched to generate a recommended list based on the degree of matching.

[0038] The hierarchical structure of the departmental division guidelines was analyzed to identify the major departmental categories, subspecialties, and the range of symptoms they specialize in treating. The guidelines are divided into major departmental categories (e.g., internal medicine, surgery) and subspecialties (e.g., cardiology, gastroenterology), with each subspecialty clearly defining its area of ​​expertise in treating symptoms such as chest pain and arrhythmias.

[0039] The symptom features in the core grading features are decomposed into primary symptom components, secondary symptom components, and accompanying symptom components, each corresponding to different weights in department matching. The symptom features in the core grading features are decomposed into primary symptom components (retrosternal pain), secondary symptom components (abnormal heart rate), and accompanying symptom components (squeezing sensation), among which the primary symptom component has the highest weight in department matching.

[0040] At the department category level, the main symptom component is matched with the core treatment scope of each department category in the departmental division guidelines to screen out possible department categories. At the department category level, the main symptom component "retrosternal pain" matches the core treatment scope of internal medicine, therefore internal medicine is selected as a possible department category.

[0041] At the subspecialty level, based on the selected departmental categories, the secondary symptom components are matched with the symptom range of expertise of each subspecialty within the corresponding category to further screen potential subspecialties. Based on the internal medicine category, the secondary symptom component "abnormal heart rate" matches the symptom range of expertise of cardiology, thus further screening cardiology as a potential subspecialty.

[0042] At the specific departmental level, based on the selected subspecialties, the accompanying symptom components are finely matched with the characteristic diagnostic and treatment items of each specific department under the corresponding subspecialty, and the suitability parameters of the departmental matching are calculated. Under the cardiology subspecialty, the accompanying symptom component "squeezing sensation" is matched with the characteristic diagnostic and treatment items of the cardiovascular emergency department, and a high suitability parameter is calculated.

[0043] Extract matching items and discrepancies between core symptom features and departmental areas of expertise during the matching process, and record the matching strength parameters for each department. Record matching items (e.g., management of acute chest pain) and discrepancies (e.g., experience in managing rare arrhythmias) between core symptom features and the cardiovascular emergency department's areas of expertise, and determine the matching strength parameters accordingly.

[0044] Based on the departmental emergency response capability assessment results, the matching strength parameter is adjusted to increase the ranking priority of departments with strong emergency response capabilities. Referring to the departmental emergency response capability assessment, if the cardiovascular emergency department has superior emergency response speed and equipment, its matching strength parameter is adjusted to increase its ranking priority in the recommended list.

[0045] Based on the multi-level matching results, matching strength parameters, and emergency response capability assessment results, a list of recommended treatment departments is generated, ranked by priority. Taking all the above factors into account, the recommended treatment department list is generated, ranked from highest to lowest priority as follows: cardiovascular emergency department, emergency department general clinic, etc.

[0046] Based on the specialty characteristics of the recommended treatment department and the intervention needs in the tiered assessment results, essential examination items were determined. These essential examination items include key examinations for rapidly verifying diagnostic hypotheses. Combining the specialty characteristics of the cardiovascular emergency department and the intervention needs in the tiered assessment results, essential examination items were determined to be electrocardiogram (ECG), myocardial enzyme spectrum testing, and chest CT scans, etc. These examinations can rapidly verify diagnostic hypotheses such as acute myocardial infarction.

[0047] Based on real-time emergency resource load information, the order of recommended treatment departments is adjusted, prioritizing departments with currently sufficient resources. If, upon obtaining real-time emergency resource load information, the cardiovascular emergency department is experiencing a high volume of patients and resource shortages, while the emergency department's general outpatient clinic has relatively ample resources, the order of recommended treatment departments is appropriately adjusted, prioritizing the emergency department's general outpatient clinic.

[0048] Based on the patient priority, the revised list of recommended departments, and the required examinations, a structured triage plan is generated. This plan includes timeline requirements and suggested execution order. The patient priority, revised department list, and required examinations are integrated into a structured triage plan, clearly defining the timeline requirements and execution order for each examination and visit, such as completing an electrocardiogram (ECG) before proceeding to the recommended department.

[0049] A unique identifier code is added to the triage scheme, which is associated with the patient's visit time marker and identity identifier. A unique identifier code is generated, which contains the patient's visit time marker and identity identifier information, to facilitate the tracking, querying, and management of the triage scheme and ensure accurate association with patient information.

[0050] The triage plan, with its unique identifier, is synchronized to the emergency dispatch system, triggering the system's resource reservation and personnel notification processes. After synchronization, the system automatically triggers resource reservation processes, such as reserving electrocardiogram (ECG) equipment and examination rooms, and notifies relevant medical staff to prepare for patient reception, ensuring the effective execution of the triage plan.

[0051] In this embodiment, the cloud platform generates a triage plan that includes priority of treatment and necessary examinations based on the triage assessment results and recommended departments, combined with the hospital's treatment procedures and resources. The priority of treatment is determined based on the triage assessment results, ensuring that patients with more severe conditions receive priority treatment. Necessary examinations are selected based on the patient's symptoms and condition, focusing on tests crucial for diagnosis and treatment, such as complete blood counts, electrocardiograms, and imaging examinations.

[0052] The method also includes a training step for a triage decision model, collecting historical emergency triage case data, which includes multi-dimensional medical information, manual triage results, and final diagnosis and treatment conclusions.

[0053] Step S211: Preprocess historical emergency triage case data, removing cases with a data missing rate exceeding a preset proportion, performing word segmentation and standardization on textual information, and normalizing numerical information. During preprocessing, cases with excessive missing key information are removed, textual information such as symptom descriptions is segmented and converted into standard terminology, and numerical information such as vital signs is normalized to a uniform range to ensure consistent data format.

[0054] Step S212: Construct training and validation sample sets. Divide the preprocessed historical emergency triage case data into training and validation sample sets according to a preset ratio. The training sample set is used for model parameter learning, and the validation sample set is used for model performance evaluation. The preprocessed data is divided into training and validation sample sets in an 8:2 ratio. The training sample set is used for the model to learn triage rules and parameters, and the validation sample set is used to evaluate the model's generalization ability during training.

[0055] Step S213: Initialize the network structure of the triage decision model. The network structure includes a feature allocation layer, a core inference path, an associated inference path, a fusion inference layer, a timeliness mapping layer, a priority calibration module, and an output layer. Set the initial parameters for each layer. Initialize the network structure of the model, determine the number of neurons and activation functions for each layer, such as using a fully connected structure for the feature allocation layer, and using a parallel neural network structure for the core inference path and the associated inference path. Set the initial parameters such as weights and biases for each layer.

[0056] Step S214: Input the multi-dimensional medical information from the training sample set into the initialized triage decision model, and calculate the predicted triage result output by the model through forward propagation. The multi-dimensional medical information from the training samples is input into the model, and through the forward propagation process including feature allocation, dual-track inference, and fusion, the predicted triage result is calculated.

[0057] Step S215: Calculate the loss value based on the difference between the predicted triage results and the manual triage results. Use the backpropagation algorithm to adjust the parameters of each layer of the model to minimize the loss value. Compare the model's predicted triage results with the manual triage results, calculate the loss value using the loss function, and then use the backpropagation algorithm to adjust the parameters of each layer from the output layer to the input layer to reduce the loss value.

[0058] Step S216: After each preset round of training, the performance metrics of the model are evaluated using a validation sample set. The performance metrics include the grading accuracy, the intervention time prediction error, and the department matching accuracy.

[0059] Every 10 training rounds, the model performance is evaluated using a validation sample set, and the following calculations are made: grading accuracy (the proportion of predicted grading that matches the actual grading), intervention time prediction error (the deviation between predicted time and actual time), and department matching accuracy (the proportion of recommended departments that match the actual departments visited).

[0060] Step S217: When the performance index reaches the preset threshold and remains stable for multiple consecutive rounds, stop model training and save the model parameters at this time as a pre-trained triage decision model. When the performance index of the validation sample set reaches the preset standard and shows no significant fluctuation for 5 consecutive rounds, stop training, save the current model parameters, and form a pre-trained model that can be used for actual triage.

[0061] Step S218: Deploy the pre-trained triage decision model to the inference engine of the emergency triage system, and configure the model call interface and data interaction protocol. The pre-trained triage decision model is encapsulated as a callable service module and deployed to the inference engine of the emergency triage system. At this time, it is necessary to configure the interaction protocol between the model and the front-end data acquisition module and the back-end database, and clarify the data transmission format, encoding method and verification rules. For example, it is stipulated that multi-dimensional medical information must be input into the model in structured JSON format, and the model output triage assessment results must include field identifiers, confidence intervals and data generation time.

[0062] Step S2181: Connect to the real-time data channel of the emergency triage system and establish a dynamic acquisition mechanism for model input data. Through interface adaptation, the pre-trained triage decision model can receive data streams from emergency terminals, monitoring equipment, and the electronic health record system in real time. Set up a data buffer queue, and trigger the model input event when the completeness of fields of multi-dimensional medical information reaches a preset standard. For example, when the symptom description, vital sign records, and past medical history summary are all obtained and the timestamps are aligned, the data is packaged and input into the model.

[0063] Step S2182: Set the resource scheduling strategy for model inference, allocating computing resource priorities and concurrent processing limits. Based on the predicted number of patients during peak emergency periods, allocate dedicated computing resources to the triage decision model, setting a single batch processing limit and a timeout retry mechanism. When the number of requests during the same period exceeds the concurrency threshold, a queuing mechanism is activated, processing data according to the order of the patient's appointment time. For example, reserve a fixed number of computing cores and memory space for the model to ensure that a maximum of 10 patient information entries are processed per batch, and automatically trigger a retry process if the process is not completed within the timeout period.

[0064] Step S219: The method further includes online performance monitoring of the deployed triage decision model, recording inference time, result consistency, and error rate indicators. During model operation, the performance monitoring module is activated to collect inference time, number of cases processed per hour, consistency ratio between results and manual triage, and number of abnormal outputs in real time. These indicators are written to the monitoring log, and statistical reports are generated hourly. For example, the distribution range of single-case inference time is recorded, the degree of consistency with manual triage results is calculated, and inference failure events caused by data format errors are marked.

[0065] Step S2191: Periodically extract actual triage cases and compare them with the model output results to generate a model deviation report. Each week, a certain number of cases are extracted from the emergency department triage records, covering different grading results and departmental assignment types. The grading assessment results output by the model are compared with the final diagnostic grading by the attending physician. The deviation rate of the model under specific symptom combinations, age groups, or underlying disease types is statistically analyzed, and the characteristic distribution of the deviation is analyzed. For example, the model's grading deviation for the symptom combination of "chest pain with dyspnea" is analyzed in detail, recording the proportion of cases in which the model underestimates or overestimates the urgency.

[0066] Step S2192: When the model bias rate exceeds a preset threshold, the incremental training process is triggered, and biased cases are used to supplement the training sample set. If the model bias rate exceeds the set standard for two consecutive weeks, multi-dimensional medical information and corrected grading results are extracted from the biased cases and added to the training sample set, and incremental training is started. Incremental training uses the original training parameters and validation mechanism, but only performs a limited number of rounds of parameter fine-tuning for the newly added samples to avoid damaging the original performance of the model. For example, when the grading bias rate of the "abdominal pain with hypotension" case reaches 15%, 100 new data points of this type of case are added to the training set for 10 rounds of incremental training.

[0067] The nurse's terminal and the doctor's terminal are configured to receive and display the triage plan sent from the cloud, which includes the treatment priority and the required examination items for the emergency patient.

[0068] In this embodiment, the nurse's and doctor's terminals establish a communication connection with the cloud to receive in real-time triage plans corresponding to emergency patients, including their priority and required examinations. This information is clearly displayed on the software interfaces of both the nurse's and doctor's terminals. Nurses can arrange the order of patients' visits according to their priority and assist patients in completing necessary examinations; doctors can quickly understand the patient's condition and examination results based on the triage plan and conduct targeted diagnosis and treatment.

[0069] The aforementioned technical solution automatically collects multi-dimensional vital sign data on the patient's end and generates structured symptom descriptions. The cloud integrates past medical history and consultation time stamps to construct a complete patient profile. Compared to traditional manual recording methods, this significantly reduces the risk of information omissions or errors. The cloud constructs a hierarchical dependency network based on symptom concurrency probability and causal strength, updating symptom urgency scores in real time through a node influence transmission mechanism. This model can quantify the dynamic relationships between complex symptoms, reducing the error rate of urgency assessment and achieving a qualitative leap from static rules to dynamic deduction. Simultaneously, the pre-trained model employs a dual-track inference mechanism: one track performs initial triage based on a symptom-department association graph, while the other track dynamically optimizes based on emergency room resource occupancy and intervention timeliness requirements. This improves the consistency of triage results and shortens the average triage time.

[0070] Furthermore, based on the triage assessment results, a triage plan including visit priority, recommended departments, and necessary examinations is generated in real time and synchronized to the nurse / physician's end. Through linkage with the emergency dispatch system, the triage strategy can be dynamically adjusted to match resource changes, improving emergency bed turnover and reducing examination redundancy. This achieves intelligent integration across the entire data-model-resource chain, realizing closed-loop optimization from information collection to resource scheduling, and providing efficient, accurate, and dynamically adaptable triage.

[0071] In a preferred embodiment, the step of inputting the dynamic urgency feature set into a pre-trained triage decision model and generating a graded assessment result including grading level and intervention timeliness requirements through a dual-track inference mechanism includes: The dynamic urgency feature set is input into the feature allocation layer of the triage decision model. The feature allocation layer is used to divide the dynamic urgency features into core symptom feature streams and related symptom feature streams based on the clinical attributes of the dynamic urgency features in the dynamic urgency feature set.

[0072] The feature allocation layer is a specific hierarchical structure in the pre-trained triage decision model, used to classify the input dynamic urgency feature set, dividing it into different feature streams based on the clinical attributes of the features. Dynamic urgency features are individual feature elements in the feature information set that reflects the urgency of each patient's symptoms after hierarchical urgency adjustment by the symptom dependency network. They include the symptom's urgency value and other related attribute information. Clinical attributes are feature attributes closely related to medical clinical practice, used to describe the characteristics of symptoms in clinical diagnosis and treatment, such as symptom severity, whether it is a major manifestation of the disease, and its correlation with other symptoms. The core symptom feature stream is a feature stream composed of symptom features with key clinical significance that directly reflect the severity and urgency of the patient's condition, divided from the dynamic urgency feature set by the feature allocation layer. These symptoms are usually the main reasons why patients seek emergency care. The associated symptom feature stream is a feature stream composed of symptom features that are associated with the core symptoms, although they may not be the key factors directly determining the severity of the condition, but they will have a certain impact on the patient's overall condition, treatment process, and prognosis.

[0073] In this embodiment, the feature allocation layer pre-defines a series of rules and algorithms based on clinical medical knowledge and experience. When the dynamic urgency feature set is input into this layer, the system analyzes and judges the clinical attributes of each dynamic urgency feature. For example, by comparing the urgency values ​​of symptoms, ranking the importance of reference symptoms among common emergency diseases, and analyzing the causal relationship strength between symptoms and other symptoms, it determines whether each feature belongs to a core symptom or a related symptom. Then, core symptom features are aggregated into a core symptom feature stream, and related symptom features are aggregated into a related symptom feature stream.

[0074] The core reasoning path of the triage decision model sequentially performs lethality assessment, intervention window calculation, and resource demand prediction on the core symptom feature stream to generate core hierarchical features.

[0075] The core reasoning path is a reasoning process within the pre-trained triage decision model specifically designed to process the core symptom feature flow. Through in-depth analysis of core symptoms, it extracts key information to support tiered assessment. Fatality assessment evaluates the potential life-threatening severity of the core symptoms, determining the risk of serious complications or even death for the patient under current symptom conditions. Intervention window calculation calculates the optimal timeframe for effective intervention based on the development trend and severity of the core symptoms—the time interval from patient visit to when intervention must begin. Resource demand prediction forecasts the medical resources required for treatment and management of core symptoms, including human resources (e.g., the number of physicians and nurses) and material resources (e.g., medications and equipment). Core tiered features are a set of feature information comprehensively reflecting the crucial role of core symptoms in tiered assessment, obtained after fatality assessment, intervention window calculation, and resource demand prediction.

[0076] In this embodiment of the disclosure, the core reasoning path first utilizes a pre-set medical model and algorithm, combined with historical case data and clinical research results, to perform a lethality score on each symptom in the core symptom feature stream. For example, for chest pain, the risk probability of the symptom leading to the patient's death is comprehensively assessed by considering factors such as the nature of the pain (e.g., squeezing, stabbing), location (e.g., precordial region, retrosternal region), and duration, and by referring to the incidence and mortality data of related diseases (e.g., myocardial infarction, aortic dissection).

[0077] Based on the fatality assessment results and the development pattern of core symptoms, time series analysis and predictive models are used to calculate the upper and lower limits of the time from patient presentation to the necessary intervention. For example, for patients with acute stroke, the intervention window is determined based on factors such as the time of symptom onset, the degree of neurological deficit, and the effective time window for thrombolytic therapy (generally 3-4.5 hours after onset).

[0078] Based on the type and severity of the core symptoms, and referring to hospital treatment guidelines and resource utilization standards, the various medical resources required to treat the symptoms are predicted. For example, for severely trauma patients, the required number of operating rooms, anesthesiologists, blood transfusion volumes, and special equipment (such as internal fixation devices) may be predicted. The results of lethality assessment, intervention window calculation, and resource demand prediction are integrated and coded to generate core hierarchical features.

[0079] The associated reasoning path of the triage decision model is used to sequentially perform complication risk analysis, underlying disease impact assessment and treatment conflict detection on the associated symptom feature flow, generating associated adjustment features.

[0080] The associative reasoning path is the reasoning process used in the pre-trained triage decision model to process the feature flow of associated symptoms, aiming to analyze the impact of associated symptoms on core symptoms and the overall condition. Complication risk analysis assesses the risk that associated symptoms may trigger or worsen other diseases or symptoms in the patient, analyzing the impact of these complications on the patient's condition and prognosis. Underlying disease impact assessment considers the role of the patient's pre-existing underlying diseases on current associated symptoms and the overall condition, determining whether underlying diseases will alter the manifestation, progression, and response to treatment of associated symptoms. Treatment conflict detection checks for contradictions or mutual influences between proposed treatment measures for associated and core symptoms, avoiding treatment conflicts that could lead to worsening of the patient's condition or poor treatment outcomes. Associative adjustment features are a set of feature information reflecting the adjusting effect of associated symptoms on the triage assessment, obtained after complication risk analysis, underlying disease impact assessment, and treatment conflict detection.

[0081] In this embodiment of the disclosure, the associative reasoning path utilizes a medical knowledge graph and a risk assessment model to analyze each symptom in the associated symptom feature stream. For example, for the associated symptom of foot ulcers in diabetic patients, the risk of complications such as infection and gangrene that may be caused is analyzed, taking into account factors such as the patient's blood sugar control and the degree of foot neuropathy, to calculate the probability of complications and their impact on the condition.

[0082] By reviewing the patient's past medical history and considering the characteristics of associated symptoms, the impact of underlying diseases on associated symptoms can be assessed. For example, for patients with chronic obstructive pulmonary disease (COPD) who experience cough symptoms, the analysis can determine whether COPD exacerbates the cough, whether the nature of the cough changes (e.g., from a dry cough to a cough with sputum), and the impact of COPD on the choice and efficacy of cough medications.

[0083] Based on treatment plans for core and associated symptoms, potential treatment conflicts are detected using drug interaction databases and clinical treatment guidelines. For example, in patients with both hypertension and asthma, certain antihypertensive medications (such as beta-blockers) may trigger asthma attacks, conflicting with asthma treatment. The results of complication risk analysis, underlying disease impact assessment, and treatment conflict detection are comprehensively analyzed and quantified to generate correlation-adjusted characteristics.

[0084] Based on the core grading features and the associated adjustment features, the output layer of the triage decision model generates a grading assessment result that includes the grading level and the intervention time limit requirements.

[0085] The grading system categorizes patients into different levels based on the severity and urgency of their condition, typically into Level 1 (most urgent), Level 2 (less urgent), and Level 3 (generally urgent), to guide the priority allocation of medical resources and the order in which patients seek medical attention. The intervention timeliness requirement specifies the latest time for intervention and treatment, ensuring that patients receive timely and effective treatment within an appropriate timeframe to prevent their condition from worsening.

[0086] In this embodiment, the output layer employs a specific algorithm and model to perform weighted fusion and comprehensive analysis of core grading features and related adjustment features. For example, core grading features are assigned higher weights based on their importance in the grading assessment, while related adjustment features are assigned corresponding weights based on their impact on core symptoms. By calculating the weighted feature values ​​and combining them with preset grading standards and intervention time limits, the patient's grading level and intervention time limit requirements are determined.

[0087] For example, if the core grading features indicate a very severe patient condition, and the correlation adjustment features suggest a high risk of complications and treatment conflicts, the output layer will classify the patient as Level 1 Emergency and set a short intervention timeframe, such as immediate treatment. Conversely, if the core grading features and correlation adjustment features indicate a relatively mild patient condition, the output layer will classify the patient as a lower grading level and set a relatively longer intervention timeframe. The final result generates a grading assessment that includes the grading level and intervention timeframe requirements.

[0088] In a preferred embodiment, the step of generating a grading assessment result, including the grading level and the intervention timeliness requirement, through the output layer of the triage decision model based on the core grading features and the correlation adjustment features includes: A dynamic weighting strategy based on clinical risk is adopted to weight and fuse the core grading features and the associated adjustment features to generate fused features.

[0089] Among these strategies, the dynamic weighting strategy based on clinical risk is used to dynamically adjust the weight of different features during the feature fusion process according to the patient's clinical risk status. Clinical risk encompasses multiple aspects, including the severity of the patient's condition, the likelihood of potential complications, and the impact of underlying diseases on the current condition. By assessing these risk factors in real time and assigning corresponding weights to each feature, the feature fusion process can more accurately reflect the true urgency of the patient's condition.

[0090] In this embodiment, a clinical risk assessment model is constructed that comprehensively considers multiple clinical indicators, such as vital signs (heart rate, blood pressure, respiratory rate, body temperature, etc.), symptom severity, past medical history, and laboratory test results. By quantifying and scoring these indicators, the patient's overall clinical risk value is calculated.

[0091] Then, the dynamic weights of the core grading features and the associated adjustment features are determined based on the clinical risk value. Generally, the higher the clinical risk, the greater the weight of the core grading features, because the core symptoms have a more critical impact on the condition at this time; conversely, when the clinical risk is relatively low, the weight of the associated adjustment features will be appropriately increased to fully consider the potential impact of associated symptoms on the condition.

[0092] Finally, the core hierarchical features and the associated adjustment features are weighted and summed according to the determined weights to obtain the fused features. For example, let the core hierarchical feature vector be X=[x1,x2,...,xn], the associated adjustment feature vector be Y=[y1,y2,...,ym], the weight of the core hierarchical features be wx, and the weight of the associated adjustment features be wy (wx+wy=1), then the formula for calculating the fused feature vector Z is: Z=wx×X+wy×Y.

[0093] The fusion features are mapped to intervention time window parameters through the timeliness mapping layer of the triage decision model. The intervention time window parameters are used to quantify the treatment response time limits corresponding to different levels.

[0094] The timeliness mapping layer is a specific hierarchical structure in the pre-trained triage decision model. Its function is to transform fused features into intervention time window parameters that quantify the response time limits for different treatment levels. This layer establishes a mapping relationship between fused features and intervention time, enabling the model to directly provide reasonable intervention time recommendations based on the patient's condition characteristics. The intervention time window parameter describes the time range within which patients of different levels should receive intervention treatment. It typically includes the earliest and latest intervention times, clarifying that patients receiving treatment within this time range will achieve the best treatment results and avoid disease deterioration.

[0095] In this embodiment, the timeliness mapping layer employs machine learning algorithms (such as neural networks, decision trees, etc.) to construct a mapping model. During the model training phase, a large amount of historical emergency case data is collected, including the patients' fusion characteristics and the actual intervention time. Using the fusion characteristics as input and the intervention time as output, the mapping model is trained to learn the intrinsic relationship between the fusion characteristics and the intervention time.

[0096] In practical applications, when the fused features are input into the time-mapping layer, the mapping model processes and calculates the fused features based on the trained parameters, outputting the corresponding intervention time window parameters. For example, for patients with very severe conditions (whose fused features indicate extremely high clinical risk), the mapping model will output a shorter intervention time window, meaning that both the earliest and latest intervention times are close to the current time, to ensure that the patient can receive treatment as soon as possible; while for patients with relatively mild conditions, the intervention time window will be relatively longer.

[0097] The priority calibration module of the triage decision model is invoked, and the triage threshold is adjusted in conjunction with the current resource load of the emergency department to adapt the intervention time window parameters to the clinical scenario, thereby generating the target intervention time window parameters.

[0098] The priority calibration module is a functional module within the triage decision model. It adjusts the triage thresholds based on the current resource load of the emergency department, thereby adapting the intervention time window parameters to the clinical scenario. Its purpose is to ensure that the triage assessment results and intervention time recommendations meet actual clinical needs under different resource conditions, and to rationally allocate medical resources. The current resource load of the emergency department refers to the utilization of various medical resources available to the emergency department at a given moment, including the number of medical staff, beds, equipment availability, and drug inventory. The level of resource load directly affects the patient reception capacity and treatment efficiency. The triage threshold is the critical value used to classify patients into different triage levels. In the initial triage assessment, the patient's triage level is determined by comparing the fusion characteristics with the preset triage thresholds. However, these triage thresholds are set under ideal resource conditions, and changes in resource load in actual clinical practice may affect the rationality of the triage. The target intervention time window parameter is the intervention time window parameter adjusted by the priority calibration module in conjunction with the current resource load of the emergency department.

[0099] In this embodiment of the disclosure, the priority calibration module first monitors the current resource load of the emergency department in real time, and determines the current resource load status (such as high load, medium load, low load) by comparing it with the preset resource load standard.

[0100] Then, the stratification threshold is adjusted based on the resource load status. For example, when the emergency department is under high load, in order to prioritize the treatment of the most seriously ill patients, the stratification threshold will be appropriately increased, so that more patients are classified into lower stratification levels, thereby extending their intervention time window and alleviating resource pressure; conversely, when the resource load is low, the stratification threshold will be appropriately decreased, so that more patients can receive timely treatment.

[0101] Finally, the patient's grading level is reassessed based on the revised grading threshold, and the intervention time window parameters are adjusted accordingly to generate the target intervention time window parameters. For example, if the patient was originally determined to be a Level II emergency based on fusion characteristics and the initial grading threshold, the intervention time window would be 30-60 minutes; however, under high resource load conditions, the grading threshold is increased, and the patient is reclassified as a Level III emergency, and the intervention time window is adjusted to 60-120 minutes.

[0102] Based on the target intervention time window parameters, the output layer of the triage decision model generates a graded assessment result including the grade level and the intervention time limit requirements.

[0103] In this embodiment, the output layer receives the target intervention time window parameters and the graded level information adjusted by the priority calibration module. This information is then integrated and formatted to generate a graded assessment result according to a preset output format. For example, the graded level is represented by explicit text or numerical identifiers (such as "Level 1 Emergency" or "1"), and the intervention time limit requirement is represented by a specific time range (such as "treatment must begin within 30 minutes"). The generated graded assessment result is output to provide a scientific basis for the triage and treatment of emergency patients, guiding medical staff to rationally arrange the order of patient visits and treatment time, improving the utilization efficiency of emergency medical resources and the treatment effect on patients.

[0104] In a preferred embodiment, the core symptom feature stream is sequentially subjected to lethality assessment, intervention window calculation, and resource requirement prediction via the core reasoning path of the triage decision model to generate core hierarchical features, including: The core symptom feature stream is input into the lethality assessment submodule of the core inference path, and the lethal symptoms in the core symptom features are detected by the pre-trained risk recognition unit.

[0105] The lethality assessment submodule is a functional module within the core inference path specifically designed to assess the lethality of a patient's core symptoms. It analyzes the input symptom features using a series of algorithms and models to determine the presence of fatal symptoms. The pre-trained risk identification unit is a model component pre-trained on a large amount of labeled clinical data. It can identify symptoms with a fatal risk among the core symptom features. This data includes the symptom presentations of various common emergency conditions and their corresponding lethality annotations.

[0106] In this embodiment, the pre-trained risk identification unit typically employs a deep learning model, such as a convolutional neural network (CNN) or a recurrent neural network (RNN) and its variants (such as LSTM and GRU). During the training phase, a large amount of clinical case data, including core symptom features and whether the symptoms are fatal, is input into the model for training. The model learns the complex relationship between symptom features and fatality rates, adjusting its parameters to optimize identification accuracy.

[0107] In practical applications, the core symptom feature stream is input into a pre-trained risk identification unit. This unit analyzes and calculates each symptom feature, extracts its feature representation, and matches it with the fatal symptom feature patterns learned during training. If a symptom feature is highly similar to a known fatal symptom pattern, the symptom is determined to be fatal, and the corresponding identification result is output.

[0108] The fatality assessment is performed based on the fatal symptoms to determine the onset speed, severity, and systemic involvement of the fatal symptoms, and to generate a fatality score vector.

[0109] Fatality assessment is a comprehensive and quantitative evaluation of detected fatal symptoms. It aims to determine the onset speed, severity, and extent of impact on various systems in the patient's body, providing crucial information for subsequent grading. Onset speed refers to the rate at which fatal symptoms develop and reach their full severity, reflecting the speed of disease progression and essential for timely intervention. Severity refers to the degree of damage caused by fatal symptoms, typically quantified using a series of clinical indicators and scoring criteria. The number and extent of affected systems indicate the number and scope of body systems impacted by the fatal symptoms, such as the cardiovascular, respiratory, and nervous systems; the more systems involved, the more complex and severe the condition. Each dimension of the fatality scoring vector corresponds to a quantitative score for the onset speed, severity, and affected systems of the fatal symptoms, comprehensively representing the fatality level of the symptoms.

[0110] In this embodiment of the disclosure, the rate of onset can be assessed based on the time changes after symptom onset and the dynamic changes in relevant clinical indicators. For example, for patients with acute myocardial infarction, the rate of onset can be assessed by the duration of ST segment elevation on the electrocardiogram and the rate of change in myocardial enzyme indicators.

[0111] Severity assessment typically employs existing clinical scoring systems, such as the trauma severity score for trauma patients and the oxygenation index for patients with acute respiratory failure.

[0112] Assessment of systemic involvement requires comprehensive consideration of the patient's symptoms, signs, and relevant ancillary examination results. For example, neurological examination, head CT or MRI scans may be used to assess whether the nervous system is involved; chest X-rays, echocardiography, and other examinations may be used to assess whether the cardiopulmonary system is involved.

[0113] Each assessment indicator is assigned a corresponding quantitative score, and these scores are combined into a lethality score vector. For example, the onset speed can be divided into three levels: rapid, moderate, and slow, each corresponding to a different score; the severity can be divided into mild, moderate, and severe according to the scoring system, and assigned corresponding scores; the range of systems involved can be scored according to the number of systems involved.

[0114] The lethality score vector is input into the intervention window calculation submodule, and the intervention window calculation is performed in conjunction with the preset clinical intervention criteria to determine the optimal implementation time period for different intervention measures and generate time window features.

[0115] The intervention window calculation submodule, located within the core reasoning path, is responsible for calculating the optimal implementation timeframe for different intervention measures based on the lethality score vector. It incorporates pre-defined clinical intervention criteria to provide patients with personalized intervention timing recommendations. These pre-defined clinical intervention criteria are developed based on extensive clinical research and practical experience, specifying optimal intervention time ranges and intervention guidelines for different lethal symptoms and disease severity. These criteria provide the basis for calculating the intervention window. The time window features are a set of characteristics describing the optimal implementation timeframe for different intervention measures, including the earliest intervention time, the latest intervention time, and the optimal intervention time interval, used to guide medical staff in rationally scheduling patient treatment.

[0116] In this embodiment, the intervention window calculation submodule stores a preset clinical intervention standard database, which contains various fatal symptoms and their corresponding intervention measures and optimal intervention time ranges. When a fatality score vector is input into this submodule, the module matches and queries the clinical intervention standard database based on information such as the onset speed, severity, and scope of affected systems in the score vector to find the most suitable intervention measure and corresponding optimal intervention time range for the current patient.

[0117] Meanwhile, considering individual patient differences and actual clinical conditions, the module may fine-tune the optimal intervention time range by incorporating dynamic factors (such as the patient's age, underlying diseases, and current vital signs). Ultimately, a feature vector containing information such as the earliest intervention time, the latest intervention time, and the optimal intervention time interval is generated, i.e., the time window feature.

[0118] Based on the time window characteristics and the lethality score vector, an intervention delay risk parameter is calculated, which describes the severity of the adverse consequences of missing the optimal intervention time.

[0119] The intervention delay risk parameter measures the severity of adverse consequences such as worsening of the patient's condition, development of complications, or even death if the optimal intervention time is missed. It comprehensively considers multiple factors in the time window characteristics and the lethality score vector. The severity of adverse consequences refers to the extent to which the patient's condition worsens, physical function is impaired, quality of life declines, and the risk of death increases due to intervention delay; it is an important basis for assessing the risk of intervention delay.

[0120] In this embodiment, firstly, an intervention delay risk assessment model is established. This model can be a regression or classification model based on machine learning algorithms (such as decision trees, support vector machines, neural networks, etc.). During the model training phase, a large amount of clinical case data is collected, including patients' time window characteristics, lethality score vectors, and actual intervention delays and adverse consequences. Using the time window characteristics and lethality score vectors as inputs and the severity of adverse consequences as outputs, the model is trained to learn the complex relationship between the input features and the severity of adverse consequences.

[0121] The time window features and lethality score vector are input into a trained intervention delay risk assessment model. The model calculates and predicts based on the learned relationships, outputting a quantified value as the intervention delay risk parameter. A higher value indicates a more severe adverse consequence from missing the optimal intervention time. For example, for patients with acute myocardial infarction, if the optimal intervention time is reperfusion therapy within 90 minutes of onset, the intervention delay risk parameter will increase with the increase in delay time when the actual intervention time exceeds this time window, reflecting an increased risk of serious complications such as myocardial necrosis and heart failure.

[0122] The intervention delay risk parameters are input into the resource demand prediction submodule to analyze the types and quantities of medical resources required to address the fatal symptom and generate resource demand characteristics.

[0123] The resource demand forecasting submodule is used to predict the types and quantities of medical resources needed to address life-threatening symptoms based on intervention delay risk parameters. It helps in the rational allocation of medical resources, ensuring patients receive timely and effective treatment. Medical resource types include human resources (such as doctors, nurses, and emergency personnel), material resources (such as medicines, medical equipment, and hospital beds), and financial resources. Different types of resources play different roles in addressing life-threatening symptoms. Resource demand characteristics describe the types and quantities of medical resources required to address life-threatening symptoms, guiding the allocation and management of medical resources.

[0124] In this embodiment, a resource demand prediction submodule establishes a resource demand prediction model, which is constructed based on clinical experience, expert knowledge, and historical data statistical analysis. It considers the relationship between factors such as different fatal symptoms, disease severity, and the risk of intervention delays, and the demand for medical resources.

[0125] When intervention delay risk parameters are input into this submodule, the model analyzes and calculates according to preset rules and algorithms. For example, acutely ill patients with high intervention delay risk parameters may require more advanced medical resources, such as intensive care unit (ICU) beds, specialized interventional cardiologists, and expensive emergency medications; while patients with low intervention delay risk parameters require relatively fewer medical resources. The required types and quantities of medical resources are determined, and a resource demand feature vector is generated. Each dimension in this vector corresponds to a type of medical resource, and its value represents the quantity of that resource required.

[0126] The core grading features are generated by integrating the lethality score vector, the time window features, the intervention delay risk parameters, and the resource requirement features.

[0127] In this embodiment, a feature fusion method is used to integrate the lethality score vector, time window features, intervention delay risk parameters, and resource requirement features. Common feature fusion methods include concatenation fusion and weighted fusion.

[0128] The individual feature vectors are directly concatenated into a longer feature vector in a certain order. For example, let the lethality score vector be V1=[v11,v12,...,v1n], the time window feature vector be V2=[v21,v22,...,v2m], the intervention delay risk parameter be a scalar r, and the resource demand feature vector be V3=[v31,v32,...,v3k], then the core hierarchical feature vector V after concatenation and fusion is: V=[v11,v12,...,v1n,v21,v22,...,v2m,r,v31,v32,...,v3k].

[0129] In a preferred embodiment, the step of sequentially performing complication risk analysis, underlying disease impact assessment, and treatment conflict detection on the associated symptom feature stream through the associative reasoning path of the triage decision model to generate associative adjustment features includes: The associated symptom feature stream is input into the complication risk analysis submodule of the associated inference path, and the complication risk analysis is performed by the complication prediction unit to generate a complication risk feature vector.

[0130] The complication risk analysis submodule is a module within the associative reasoning path used to analyze the risk of complications arising from associated symptoms in patients. It evaluates the input symptom features using internal algorithms and models. The complication prediction unit is the core component of this submodule, built based on machine learning or deep learning algorithms. It can predict the types of complications a patient may experience and their probabilities based on associated symptom features. The complication risk feature vector has each dimension corresponding to a possible complication, and the vector value represents the degree of risk of that complication, used to quantify the likelihood of a patient developing various complications.

[0131] In this embodiment of the disclosure, the complication prediction unit first preprocesses the input associated symptom feature stream, including data cleaning (removing noise and erroneous data), feature standardization (mapping feature values ​​of different ranges to a unified interval), and feature encoding (converting non-numerical features into numerical features, such as symptom category features).

[0132] Predictive models can be decision trees, random forests, neural networks, etc. Taking neural networks as an example, during the training phase, a large amount of clinical data containing associated symptom features and actual complication occurrences is collected as a training set. The associated symptom features are used as input, and the occurrence of complications (represented by binary encoding as whether a certain complication has occurred) is used as output to train the neural network. The weights and biases in the network are adjusted through the backpropagation algorithm to minimize the error between the model's predictions and the actual results.

[0133] In practical applications, the preprocessed associated symptom feature stream is input into the trained complication prediction unit. Based on the learned relationship between the features and complications, the model calculates the probability of each complication and combines these probabilities into a complication risk feature vector. For example, for an emergency patient, the probabilities of developing complications such as lung infection, arrhythmia, and renal failure might be calculated to be 0.3, 0.2, and 0.1, respectively, resulting in a complication risk feature vector of [0.3, 0.2, 0.1, …].

[0134] The complication risk feature vector is input into the underlying disease impact assessment submodule to retrieve the patient's past medical history summary and perform the underlying disease impact assessment to obtain the interaction relationship between the underlying disease and the patient's symptoms, and generate the underlying disease impact coefficient.

[0135] The underlying disease impact assessment submodule is a module within the associative reasoning path used to assess the impact of a patient's underlying diseases on the risk of complications. It determines the role of underlying diseases in disease progression by analyzing the relationship between past medical history and current symptoms. The past medical history summary uses key information extracted from the patient's electronic medical record system, including the types, duration, and treatment of past underlying diseases, providing a basis for assessing the impact of underlying diseases. The interaction between underlying diseases and symptoms describes how underlying diseases affect the occurrence, development, and development of current emergency symptoms and complications. For example, diabetic patients may be more prone to complications such as blood glucose fluctuations and ketoacidosis under stressful conditions like infection. The underlying disease impact coefficient represents the degree to which the underlying disease amplifies or reduces the risk of complications; a coefficient greater than 1 indicates that the underlying disease increases the risk of complications, while a coefficient less than 1 indicates that the underlying disease reduces the risk of complications.

[0136] In this embodiment, a knowledge base is established within the submodule to analyze the interaction between underlying diseases and symptoms. This knowledge base is built upon extensive clinical research, medical literature, and expert experience. The knowledge base includes rules and influence weights governing the relationships between various underlying diseases and common emergency symptoms and complications.

[0137] When the complication risk feature vector is input into the submodule, the patient's past medical history summary is retrieved simultaneously. The basic disease information in the past medical history is matched against rules in the knowledge base to identify rules that match the current patient's condition. For example, if the patient has a history of hypertension and is currently experiencing acute stroke symptoms, the knowledge base may contain rules indicating that hypertension increases the risk of complications such as cerebral hemorrhage after stroke, and assign corresponding influence weights.

[0138] Based on the matched rules and influence weights, and combined with the data in the complication risk feature vector, the underlying disease influence coefficient is calculated. Weighted summation or other comprehensive calculation methods can be used to integrate the influence of different underlying diseases on the risk of different complications, resulting in a comprehensive underlying disease influence coefficient. For example, for patients with multiple coexisting underlying diseases, the influence coefficient of each underlying disease on the risk of various complications is calculated separately, and then a weighted sum is applied based on the severity and interaction of the underlying diseases to obtain the final underlying disease influence coefficient.

[0139] The complication risk feature vector is adjusted based on the underlying disease impact coefficient to strengthen the complication risk component of the complication risk vector that aggravates the underlying disease, thereby generating a target complication feature vector.

[0140] In this embodiment of the disclosure, a corresponding adjustment strategy is formulated based on the mechanism of action of the underlying disease influence coefficient. Generally, if the underlying disease influence coefficient is greater than 1, it indicates that the underlying disease increases the risk of complications, and the corresponding risk component in the complication risk feature vector needs to be amplified; if the underlying disease influence coefficient is less than 1, it indicates that the underlying disease reduces the risk of complications, and the corresponding risk component is reduced.

[0141] For each dimension (corresponding to a complication) in the complication risk feature vector, a multiplication operation is performed based on the underlying disease influence coefficient. For example, if the complication risk feature vector is V=[v1,v2,...,vn] and the underlying disease influence coefficient is k, then the i-th dimension vi' of the adjusted target complication feature vector V' is k×vi.

[0142] If the underlying disease has different degrees of influence on different complications, i.e., there are multiple underlying disease influence coefficients (each corresponding to a different complication), then the risk value of each dimension is adjusted according to the corresponding coefficient. For example, for complication i, its underlying disease influence coefficient is ki, then vi' = ki × vi, thus obtaining a target complication feature vector that more accurately reflects the actual situation.

[0143] The target complication feature vector is input into the treatment conflict detection submodule to perform the treatment conflict detection, so as to analyze the compatibility between the recommended treatment plan for the emergency patient's symptoms and the treatment plan for the underlying disease, and generate treatment conflict risk parameters. The treatment conflict risk parameters are used to quantify the degree of impact of the conflict on the treatment effect.

[0144] Among them, the recommended treatment plan for symptoms is a treatment measure recommended by the triage decision model or other clinical guidelines based on the patient's core emergency symptoms and related symptoms, including drug treatment, surgical treatment, physical therapy, etc.

[0145] In this embodiment, a treatment plan knowledge base is established within the submodule, containing treatment plan information for various common diseases, including drug names, dosages, usage methods, treatment cycles, and interaction rules between different treatment plans. Simultaneously, the knowledge base also records contraindications and precautions between underlying diseases and medications used to treat common emergency symptoms.

[0146] The system extracts recommended treatment plans for the patient's symptoms and underlying disease treatment plans from triage decision models or other relevant systems. It then matches the drug information and treatment measures in these two plans with rules in a treatment plan knowledge base to identify potential conflicts. For example, if the patient's underlying disease is hypertension and they are taking a beta-blocker long-term, while the current emergency symptoms recommend the use of a drug containing a beta-agonist, there is a conflicting interaction between these two drugs.

[0147] Treatment conflict risk parameters are calculated based on the matched conflict points and their influence weights in the knowledge base. The Analytic Hierarchy Process (AHP) or other multi-attribute decision-making methods can be used to comprehensively consider factors such as the type, severity, and probability of occurrence of the conflict, assigning appropriate weights to each conflict point. The final treatment conflict risk parameters are then obtained through weighted summation or other comprehensive calculation methods. For example, different weights can be assigned to different types of conflicts (such as drug interactions, treatment contraindications, etc.), and the weight values ​​can be adjusted according to the specific circumstances of the conflict (such as drug dosage, patient's physical condition, etc.) to finally calculate a comprehensive treatment conflict risk parameter.

[0148] The adjusted complication risk feature vector, the underlying disease impact coefficient, and the treatment conflict risk parameter are integrated to generate the associated adjusted feature.

[0149] Various feature fusion methods can be used, such as concatenation fusion and weighted fusion. Concatenation fusion directly concatenates the individual feature vectors in a certain order into a longer feature vector; weighted fusion assigns appropriate weights to each feature based on its importance, and then adds or concatenates the weighted features.

[0150] Implementation of splicing and fusion: Let the adjusted complication risk feature vector be V1=[v11,v12,...,v1n], the underlying disease influence coefficient be a scalar k, and the treatment conflict risk parameter be a scalar r. Then the spliced ​​and fused correlation adjusted feature vector V is: V=[v11,v12,...,v1n,k,r], V=w1×V1+w2×k+w3×r, where V1 can be appropriately transformed into a scalar or expanded in dimension to fuse with the scalar feature according to specific circumstances. Through feature fusion, a correlation adjusted feature that can comprehensively reflect the relationship between the patient's associated symptoms, underlying diseases, and treatment plans is generated, providing richer and more accurate information for subsequent triage decisions.

[0151] In a preferred embodiment, the step of inputting the associated symptom feature stream into the complication risk analysis submodule of the associated inference path, and performing the complication risk analysis through the complication prediction unit to generate a complication risk feature vector includes: The associated symptom feature stream is input into the complication risk analysis submodule of the associated inference path, and the complication risk analysis is performed by the complication prediction unit to identify potential complication triggers in the associated symptom features.

[0152] Among these, potential complication triggers are factors among the associated symptom characteristics that can induce or increase the likelihood of a certain complication. These factors may be specific symptom manifestations or specific combinations of symptoms. For example, in diabetic patients, a persistent state of hyperglycemia combined with infection symptoms may be a potential trigger for the complication of diabetic ketoacidosis.

[0153] In this embodiment of the disclosure, the complication prediction unit first preprocesses the input associated symptom feature stream. This includes data cleaning to remove noisy and erroneous data, ensuring the accuracy and reliability of the input data; feature standardization to map feature values ​​of different ranges and scales to a unified interval, avoiding the impact of excessive differences in feature values ​​on the training effect of the model; and feature encoding to convert non-numerical features (such as symptom category descriptions) into numerical features so that the model can process them.

[0154] The prediction model can be a convolutional neural network (CNN), a recurrent neural network (RNN), or its variants (such as LSTM and GRU). Taking CNN as an example, during the training phase, a large amount of clinical data containing associated symptom features and actual complication triggers is collected as a training set. The associated symptom features are used as input, and the complication triggers are used as output labels to train the CNN model. The backpropagation algorithm is used to continuously adjust the convolutional kernel parameters and fully connected layer weights in the network, minimizing the error between the model's predictions and the actual results. After multiple iterations of training, the model gradually learns the complex mapping relationship between symptom features and complication triggers.

[0155] In practical applications, the preprocessed associated symptom feature stream is input into a trained complication prediction unit. The model first performs convolutional operations on the input features through convolutional layers to extract local feature information; then, pooling layers reduce the dimensionality of the features, reducing computation while preserving important features; finally, fully connected layers integrate and classify the extracted features, outputting the identified potential complication triggers. For example, given the input associated symptom feature stream, the model might identify potential complication triggers such as infection, inflammatory response, and metabolic disorders.

[0156] Based on the potential complication triggers, the types of potential complications with a high risk of occurrence are identified.

[0157] High-risk complications refer to situations where, in the presence of specific potential complication triggers, the likelihood of a patient developing a complication is significantly higher than normal. Once these complications occur, they can pose a serious threat to the patient's health, requiring timely intervention and treatment. Complication types categorize and summarize potential complications to facilitate subsequent risk assessment and treatment decisions. Common complication types include infectious complications (such as lung infections and urinary tract infections), cardiovascular complications (such as arrhythmias and heart failure), and metabolic complications (such as diabetic ketoacidosis and hyperosmolar coma).

[0158] In this embodiment of the disclosure, a knowledge base is established that contains the correspondence between potential complication triggers and their corresponding types. This knowledge base is constructed based on extensive clinical research, medical literature, and expert experience, and records in detail the types of complications that various potential complication triggers may cause and their mechanisms of occurrence. For example, the knowledge base may record that the potential complication trigger of "infection" may cause multiple types of complications such as "lung infection," "urinary tract infection," and "sepsis," and describes in detail how infection leads to these complications through different pathways and mechanisms.

[0159] Once potential complication triggers are identified, they are matched against rules in a knowledge base. Using techniques such as string matching and semantic analysis, knowledge base entries matching the triggers are identified, and based on the concurrency type information recorded in these entries, high-risk complication types are marked. For example, if the identified potential complication trigger is "lung inflammation," matching the knowledge base reveals that potentially high-risk complication types include "lung infection" and "respiratory failure," which will then be marked.

[0160] Based on the type of complication, the duration of symptoms of the emergency patient, and the patient's basic physical condition, the probability of occurrence of the potential complications is calculated, and a complication risk feature vector is generated.

[0161] The duration of symptoms refers to the length of time from the onset of symptoms to the present moment for emergency room patients. The length of symptom duration can affect the risk of complications; generally, the longer the symptoms last, the higher the likelihood of complications. The patient's baseline health status refers to their overall health condition before the onset of illness, including the presence of underlying diseases (such as hypertension, diabetes, heart disease, etc.), immune system function, age, and gender. Baseline health status has a significant impact on the probability of complications; for example, patients with underlying diseases may have a higher risk of developing complications when presenting with emergency room symptoms.

[0162] In this embodiment of the disclosure, probabilistic calculation models such as logistic regression and Bayesian networks are used to calculate the probability of potential complications. Taking the logistic regression model as an example, this model predicts the probability of complications by establishing a logical relationship between whether or not a complication occurs (a binary variable, 1 for occurrence and 0 for non-occurrence) and influencing factors such as complication type, symptom duration, and basic physical condition.

[0163] Information such as complication types, symptom duration, and underlying physical condition of emergency patients is collected and quantified. For complication types, one-hot encoding can be used to convert them into numerical features; symptom duration can be directly quantified in days or hours; underlying physical condition can be comprehensively assessed by evaluating factors such as the patient's age, underlying diseases, and immune system to provide a comprehensive score or classification index.

[0164] The probabilistic computation model was trained using collected clinical data. Quantified features were used as input, and the occurrence of complications was used as the output label. Regression coefficients were estimated using methods such as maximum likelihood estimation. After training, the model learned the relationship between different influencing factors and the probability of complication occurrence.

[0165] For each identified potential complication type, the patient's complication type, symptom duration, and baseline physical condition are substituted into a trained probability calculation model to calculate the probability of that complication occurring. All potential complication probabilities are then arranged in a specific order to generate a complication risk feature vector. For example, if a patient may have three complications—lung infection, arrhythmia, and renal failure—with calculated probabilities of 0.3, 0.2, and 0.1 respectively, the complication risk feature vector would be [0.3, 0.2, 0.1].

[0166] In a preferred embodiment, the step of constructing a symptom dependency network based on a preset symptom association relationship, according to the received description of the emergency patient's symptoms, a summary of the emergency patient's past medical history, and the time of visit information uploaded by the patient, includes: Extract symptom items from the symptom description, and determine independent symptom items and compound symptom clusters based on the symptom items, the summary of past medical history, and the consultation time information. The compound symptom cluster consists of multiple related symptom items that occur simultaneously.

[0167] Among them, symptom items are single, independent symptom descriptions extracted from the symptom presentation, such as "headache," "fever of 38.5℃," and "cough with sputum." Each symptom item is a specific depiction of the patient's symptoms. Independent symptom items are symptoms that appear alone in the patient's current symptom presentation and have no obvious direct connection with other symptoms. For example, a patient with only a simple headache symptom, without any other accompanying symptoms that have an obvious concurrent or secondary relationship with it, can be considered an independent symptom item. Complex symptom clusters are a collection of multiple symptom items that appear simultaneously and are related to each other. These symptoms are often intrinsically linked in terms of pathogenesis and pathophysiological processes, collectively reflecting some abnormal state of the patient's body. For example, a patient may simultaneously experience fever, cough, sputum, and chest pain; these symptoms may all be related to a lung infection, constituting a complex symptom cluster.

[0168] In this embodiment, the Named Entity Recognition (NER) method from Natural Language Processing (NLP) technology is employed. First, a medical dictionary containing common symptom terms is constructed as the basis for symptom recognition. Then, rule-based matching algorithms or deep learning-based sequence labeling models (such as the BiLSTM-CRF model) are used to process the symptom description text. The model accurately identifies and extracts symptom items by learning the contextual features and language patterns of symptom terms in the text.

[0169] Analysis was conducted using a combination of the patient's medical history summary and consultation time information. For each extracted symptom, it was examined to see if it had a known association with other symptom items in the patient's medical history or during the current course of the illness. If a symptom item did not have a clear pathophysiological connection with other symptom items, and there was no record in the patient's medical history indicating that they would occur simultaneously, it was identified as an independent symptom item. For related symptom items, their occurrence time and sequence were further analyzed. If multiple symptom items occurred within the same time period, and the patient's medical history or medical knowledge indicated that they might be caused by the same etiology or have an interrelationship, these symptom items were grouped into a complex symptom cluster. For example, if a patient has a history of chronic obstructive pulmonary disease (COPD) in their medical history, and at this consultation they simultaneously presented with symptoms such as cough, sputum production, and worsening dyspnea, and the consultation time showed that these symptoms occurred simultaneously within a short period of time, these symptoms could be identified as a complex symptom cluster, possibly related to an acute exacerbation of COPD.

[0170] Based on the associated symptom items in the complex symptom cluster, a preset symptom association rule base is retrieved, and a target association type between any two associated symptom items is determined from the preset association types. The preset association types include concurrent association types, secondary association types, and exclusion association types.

[0171] Concurrent association refers to the simultaneous occurrence of two or more symptoms at the same time or within a short period of time, without a clear causal order between them, usually caused by the same etiology or pathophysiological process. For example, when a patient has a cold, they may experience symptoms such as fever, cough, and runny nose at the same time; these symptoms are in a concurrent association relationship.

[0172] Secondary association indicates that the appearance of one symptom is caused by the development or evolution of another symptom, with a clear sequence and causal relationship. For example, if a patient first develops a lung infection, and then the infection leads to pleurisy, resulting in chest pain, then the lung infection and chest pain are examples of a secondary association.

[0173] Exclusionary associations mean that two symptoms would not normally occur simultaneously in the same patient; their occurrence is often mutually exclusive. For example, during an acute asthma attack, a patient typically does not experience both respiratory alkalosis and respiratory acidosis simultaneously because their pathogenesis is contradictory.

[0174] In this embodiment of the disclosure, each associated symptom item in the complex symptom cluster is used as a query condition for retrieval in the symptom association rule base. During the retrieval process, an efficient index structure and query algorithm are employed to quickly locate the rule records containing these symptom items.

[0175] For each retrieved rule, the degree of match between the recorded symptom association types and the preset association types is analyzed. By comparing information such as the temporal sequence and pathophysiological mechanisms between symptoms, it is determined whether the association type between two associated symptom items conforms to one of concurrent, secondary, or exclusionary association types. For example, if a rule records that two symptoms occur within the same time period and there is no clear causal relationship description, the association type is determined to be concurrent; if a rule clearly indicates that one symptom is caused by the development of another symptom, it is determined to be secondary; if a rule indicates that these two symptoms cannot occur simultaneously in medicine, it is determined to be exclusionary. Finally, the target association type between any two associated symptom items is determined from the preset association types.

[0176] Based on the target association type and the frequency ratio of the co-occurrence of the two symptoms in historical cases, the concurrence probability parameter between the associated symptom items is calculated.

[0177] The co-occurrence frequency ratio of two symptoms is the ratio of the number of cases in historical records where two specific symptoms co-occur to the total number of all cases containing at least one of these two symptoms. This ratio reflects the frequency with which these two symptoms occur simultaneously in practice. The concurrency probability parameter quantifies the likelihood of two related symptom items occurring simultaneously. It comprehensively considers the target association type and the co-occurrence frequency ratio of symptoms in historical records, providing important edge attribute information for subsequent construction of the symptom dependency network.

[0178] In this embodiment of the disclosure, all cases containing the target associated symptom item are selected from the historical case database. The number of times n12 of both symptoms occur simultaneously in these cases is counted, along with the total number N of cases containing symptom 1, symptom 2, or both. The frequency ratio of the two co-occurrences is calculated as P12 = Nn12.

[0179] The frequency ratio is adjusted according to the type of target association. For concurrent association types, since the two symptoms occur simultaneously and there is no clear causal relationship, the concurrent probability parameter P can be directly taken as a certain multiple of the frequency ratio P12 (determined according to the actual situation and experience, such as 1 time), that is, P=k1×P12, where k1 is the concurrent association adjustment coefficient.

[0180] For secondary association types, considering the causal sequence, the concurrent probability parameter needs to be appropriately reduced, for example, P=k2×P12, where k2<1 is the secondary association adjustment coefficient, and its value can be determined according to factors such as the time interval of the secondary relationship and the complexity of the pathophysiological process.

[0181] For exclusionary association types, theoretically, the two symptoms should not occur simultaneously, with a concurrence probability parameter P=0. In this way, by comprehensively considering the target association type and historical case data, the concurrence probability parameter between associated symptom items can be calculated.

[0182] Based on the descriptions of symptom causal relationships and degree of influence in clinical guidelines, assess the strength of the causal influence among the associated symptom items.

[0183] The causal effect strength is an indicator used to measure the extent to which one symptom influences another. It reflects the tightness and magnitude of the causal relationship between symptoms and is of great significance for understanding the pathogenesis of diseases and developing treatment plans.

[0184] In this embodiment of the disclosure, natural language processing technology is used to parse the clinical guideline text. Descriptive statements regarding the causal relationship and degree of impact of symptoms are identified, and key information, such as causal symptom pairs and qualitative descriptions of the degree of impact (e.g., "mild impact," "moderate impact," "severe impact"), is extracted.

[0185] A causal influence strength quantification table is established to convert the qualitative descriptions in clinical guidelines into specific numerical values. For example, the values ​​are defined as follows: "mild influence" corresponds to 0.2-0.4, "moderate influence" to 0.5-0.7, and "severe influence" to 0.8-1.0. Based on the analyzed qualitative descriptions of influence, the corresponding numerical ranges are found in the quantification table, and appropriate adjustments are made based on specific symptom pairs and clinical realities to ultimately determine the causal influence strength values ​​between related symptom items.

[0186] The symptom dependency network is constructed by using each of the independent symptom items and each of the associated symptom items as nodes of the symptom dependency network, and using the concurrency probability parameter and the causal influence strength as attributes of the edges between nodes.

[0187] In this embodiment, each extracted independent symptom item and each symptom item in the associated symptom items are treated as a node in the symptom dependency network. A unique identifier is assigned to each node, such as a numerical code or a hash value of the symptom name, to ensure accurate identification and referencing of each node during network construction and management.

[0188] Edge Construction: For any two nodes (i.e., two symptom items), if a relationship exists between them (determined according to the previous steps), an edge is drawn between them. The direction of the edge is determined based on the target association type between the two symptom items (for secondary association types, the edge points from the causal symptom to the resulting symptom; for concurrent and exclusionary association types, the edge can be undirected or bidirectional). Simultaneously, the calculated concurrency probability parameter and the assessed causal influence strength are stored as attributes of the edge in the edge's relevant information.

[0189] Network visualization: The constructed symptom dependency network is visualized using graphical visualization tools (such as Graphviz, Gephi, etc.). By adjusting the size, color, and edge thickness and color of nodes, the relationships and attribute information between symptoms are presented intuitively, enabling medical staff to more clearly understand the patient's symptom characteristics and disease progression.

[0190] In a preferred embodiment, the hierarchical urgency adjustment of the symptom dependency network, updating the urgency of each symptom item through the influence propagation between nodes, and generating a dynamic urgency feature set, includes: Each node in the symptom-dependent network is assigned an initial urgency baseline value, which is determined based on the basic weight of the symptom in the emergency triage criteria.

[0191] The initial urgency baseline value is an initial quantified value for each node (i.e., each symptom item) in the symptom dependency network, representing the basic urgency level of the symptom in an emergency situation. Emergency triage standards are a set of standardized systems used in the medical field to classify and assess the severity of emergency patients' conditions. For example, the common four-level emergency triage system (Level 1 for critically ill patients, Level 2 for severely ill patients, Level 3 for acute patients, and Level 4 for non-acute patients) assigns corresponding basic weights to various symptoms based on factors such as the degree of threat to the patient's life and health.

[0192] In this embodiment, existing emergency triage standard documents are collected and organized, and basic weight information for different symptoms is extracted. A mapping table between symptoms and basic weights is established, and each symptom item in the symptom dependency network is matched with this mapping table.

[0193] For each matched symptom item, its corresponding baseline weight is directly used as the initial urgency benchmark value. If a symptom item does not have a clearly defined baseline weight in the emergency triage criteria, its initial urgency benchmark value is determined based on its similarity to existing symptoms (judged through the pathophysiological characteristics of the symptoms, common associated diseases, etc.), using methods such as weighted averaging or expert evaluation. For example, if a new symptom has a high similarity to a known severe symptom in terms of pathogenesis and clinical manifestations, the baseline weight of that severe symptom can be referenced, and appropriately adjusted based on the degree of similarity to determine the initial urgency benchmark value of the new symptom.

[0194] Based on the causal influence strength of the edges in the symptom dependency network, the direction of influence transmission between nodes is determined.

[0195] Influence transmission direction refers to the direction of transmission of influence between symptoms in a symptom dependency network, used to clarify how a change in one symptom affects the urgency of other symptoms.

[0196] In this embodiment of the disclosure, each edge in the symptom dependency network is traversed, and the direction of the edge is determined according to the causal influence strength and directionality of the edge (for secondary association types, the edge has a clear direction, pointing from the cause symptom to the result symptom; for concurrent and exclusionary association types, bidirectional transmission can be set by default or the direction can be determined according to the actual situation).

[0197] If the causal influence strength of an edge is greater than a preset threshold (which can be determined based on clinical experience and data analysis, for example, 0.5), the direction of influence transmission is determined according to the original direction of the edge. If the causal influence strength is less than or equal to the preset threshold, the influence between symptoms is considered weak, and its directionality can be ignored, or a relatively reasonable direction of influence transmission can be determined after comprehensive judgment based on the overall network structure and clinical logic. For example, in a complex network composed of multiple symptoms, if the causal influence strength of the edge between symptom A and symptom B is low, but according to clinical knowledge, a change in symptom A is more likely to affect symptom B first, then the direction of influence transmission can be determined as from symptom A to symptom B.

[0198] According to the direction of influence propagation, the source node transmits an influence packet, including its own urgency baseline value and the corresponding causal influence intensity, to the target node in a forward direction. The target node adjusts its initial urgency baseline value according to the parameters in the influence packet and calculates the difference between the initial urgency baseline value and the adjusted urgency value to obtain the urgency deviation.

[0199] In this context, the source node is the node that generates influence and transmits information outward during the influence propagation process; it represents the initial symptom. The target node is the node that receives the influence of the source node and adjusts its own state according to the transmitted information; it represents the symptom affected by other symptoms. The influence packet is the set of information transmitted from the source node to the target node, containing the source node's urgency baseline value and the causal influence strength of the edges between the source and target nodes. The urgency deviation is the difference between the adjusted urgency of the target node and its initial urgency baseline value, used to measure the degree of change in the target node's urgency level.

[0200] In this embodiment of the disclosure, the source node encapsulates its initial urgency baseline value and the causal influence strength of the edge between it and the target node into an influence packet based on the determined influence propagation direction. After receiving the influence packet, the target node adjusts its own urgency according to the following formula: Adjusted urgency = Initial urgency baseline value + Source node urgency baseline value × Causal influence strength.

[0201] Calculate the difference between the initial urgency baseline value and the adjusted urgency value, i.e., urgency deviation = adjusted urgency - initial urgency baseline value. For example, if the initial urgency baseline value of source node A is 0.8, and the causal influence strength of the edge between it and target node B is 0.6, then the adjusted urgency of target node B = initial urgency baseline value of B + 0.8 × 0.6. The urgency deviation is the difference between the adjusted value and the initial urgency baseline value of B.

[0202] Based on the direction of influence propagation, the target node transmits the urgency deviation in reverse to the source node, and the source node adjusts its influence weight on other nodes based on the urgency deviation.

[0203] In this embodiment of the disclosure, the target node transmits the calculated urgency deviation to the source node in the reverse direction of influence propagation.

[0204] In this embodiment, after receiving an urgency deviation, the source node adjusts its influence weight on other nodes according to the following formula: Adjusted influence weight = Original influence weight + Urgency deviation × Correction coefficient (the correction coefficient can be set according to actual conditions and clinical experience, for example, 0.2). For example, if the original influence weight of source node C on target node D is 0.5, and the urgency deviation fed back to source node C by target node D is 0.2 with a correction coefficient of 0.2, then the adjusted influence weight of source node C = 0.5 + 0.2 × 0.2 = 0.54. In this way, the source node can dynamically adjust its influence on other nodes based on the feedback information from the target node.

[0205] Based on the node's hierarchical position in the symptom dependency network, multiple rounds of urgency adjustment are performed. Symptoms located upstream in the symptom dependency network complete the urgency adjustment first and transmit the impact of the urgency adjustment to symptoms located downstream in the symptom dependency network.

[0206] In this context, hierarchical position refers to the division of nodes into different levels within the symptom dependency network based on the causal relationships and the direction of influence transmission between symptoms. Upstream symptom items are those that occupy the cause position in the causal relationship and influence other symptoms; downstream symptom items are those that are influenced by upstream symptoms.

[0207] Multi-round urgency adjustment involves repeatedly executing the urgency adjustment process to gradually update and optimize the urgency of all nodes in the symptom dependency network, making the urgency assessment more accurate and in line with the actual situation.

[0208] In this embodiment, a depth-first search (DFS) or breadth-first search (BFS) algorithm is used to traverse the symptom dependency network, and the hierarchical position of each node is determined according to the direction of influence propagation. Starting from the starting node of the network (usually those symptoms without predecessor nodes), the network is traversed downwards sequentially, and symptoms directly affected by the starting node are assigned to the next level, and so on, until the entire network has been traversed.

[0209] Following a hierarchical order from highest to lowest, the urgency of nodes at each level is adjusted sequentially. In each round of adjustment, the following steps are performed on all nodes at the current level: forward propagation of the influence packet from the source node, adjustment of the urgency of the target node and calculation of the deviation, reverse feedback of the deviation from the target node, and correction of the influence weight by the source node. After the adjustment of the current level is completed, the adjusted influence is propagated to the next level, and the next round of adjustment continues. For example, the urgency of the symptom items at the first level (the most upstream) is adjusted first, and then its influence is propagated to the symptom items at the second level, and so on, until the symptom items at all levels have been adjusted.

[0210] After each round of adjustments, the causal influence strength between nodes is recalculated based on the real-time changes in the vital signs of the emergency patients, and the weight allocation of influence transmission is dynamically adjusted.

[0211] Real-time vital sign changes refer to the vital sign data of emergency patients acquired in real time during their visit through various monitoring devices (such as electrocardiogram monitors, blood pressure monitors, pulse oximeters, etc.), including changes in indicators such as heart rate, blood pressure, respiratory rate, and blood oxygen saturation. Weight allocation is the information transmission weight determined based on the strength of causal influence between nodes during the influence transmission process; it determines the degree of influence of the source node on the target node.

[0212] In this embodiment, by connecting to monitoring equipment for emergency patients, real-time vital sign data of the patients is acquired and stored in a database. A correlation model is established between changes in vital signs and the strength of the causal influence between symptoms. This model can be trained using machine learning algorithms (such as regression analysis, neural networks, etc.), with real-time vital sign data as input and adjusted values ​​for the strength of the causal influence between symptoms as output. For example, if a patient's heart rate suddenly increases, the model analysis may increase the numerical value of the causal influence between the patient and heart-related symptoms.

[0213] Based on the recalculated causal influence strength, the weight allocation of influence transmission is dynamically adjusted according to the aforementioned weight correction method. This ensures that the influence transmission between nodes can promptly reflect real-time changes in the patient's condition.

[0214] When the urgency deviation between two consecutive measurements is lower than a preset threshold, the hierarchical urgency adjustment is stopped, and the final urgency of all the symptom items is collected to form a dynamic urgency feature set.

[0215] In this embodiment, after each round of urgency adjustment, the average or maximum value of the urgency deviations of all symptom items between the current round and the previous round is calculated (an appropriate statistic can be selected based on the actual situation). The calculated deviation statistic is compared with a preset threshold. If the deviation statistic is less than the preset threshold, the urgency adjustment is considered to have converged, and the hierarchical urgency adjustment stops; otherwise, the next round of adjustment continues. After the adjustment stops, the final urgency of all symptom items in the symptom dependency network is collected, and these urgency values ​​are organized in a certain format (such as an array, list, etc.) to form a dynamic urgency feature set.

[0216] In a preferred embodiment, the nurse's terminal and the physician's terminal are further configured to: in response to emergency visit feedback information and supplementary emergency remarks information input on the nurse's terminal and / or the physician's terminal for the emergency patient, return the emergency visit feedback information and the supplementary emergency remarks information to the cloud, wherein the emergency visit feedback information is generated based on first medical orders provided based on the examination results of the required examination items, and the supplementary emergency remarks information is generated based on second medical orders provided based on the examination results of additional examination items besides the required examination items.

[0217] The emergency room visit feedback information is provided by nurses or physicians based on the results of necessary examinations for emergency patients, after professional judgment. It relates to the patient's current diagnosis and preliminary treatment plan, and is generated based on the first medical order, reflecting the treatment opinions regarding the results of necessary examinations. Supplementary emergency room remarks information is provided by nurses or physicians based on the results of additional examinations beyond the necessary ones. It is generated based on the second medical order and is used to supplement and improve the patient's condition information and treatment recommendations.

[0218] In this embodiment, nurses or physicians input emergency visit feedback information and supplementary emergency remarks based on the examination results of emergency patients on the nurse's or physician's interface. During the input process, the system can provide auxiliary functions, such as automatically linking examination results and prompting common medical orders, to improve input efficiency and accuracy.

[0219] The input emergency visit feedback information and supplementary emergency remarks information are encapsulated according to a certain data format, such as JSON format, and the information is organized into key-value pairs for easy subsequent transmission and processing.

[0220] The packaged data is sent to the cloud server via a wireless network (such as Wi-Fi, 4G / 5G, etc.). During transmission, encryption technology (such as SSL / TLS protocol) is used to encrypt the data to ensure the security and integrity of the information. After receiving the data, the cloud server decrypts and parses it, and stores the information in the corresponding database.

[0221] The cloud platform is also configured to: associate the supplementary emergency remarks information as an additional field with the emergency visit feedback information to generate comprehensive emergency information, and check for conflicts between the first medical order information and the second medical order information based on a rule engine, mark the conflicts, and send the marked conflicts to the nurse's end and / or the doctor's end.

[0222] The comprehensive emergency information is a collection of information that integrates emergency room visit feedback and supplementary emergency room notes. It comprehensively reflects the patient's diagnosis and treatment recommendations based on different examination items. The rule engine processes and judges the input data according to predefined rules and outputs corresponding results. In this scenario, the rule engine is used to check for conflicts between the first and second medical orders. Conflicts are parts of the first and second medical orders that are contradictory, inconsistent, or mutually influential, such as conflicting medication dosages, conflicting examination times, or conflicting treatment plans.

[0223] In this embodiment, the cloud server reads emergency visit feedback information and supplementary emergency remarks from the database, adds the supplementary emergency remarks as an additional field to the emergency visit feedback information, and generates comprehensive emergency information. During the association process, the accuracy of the information correspondence is ensured; for example, the remarks for additional examination items are associated with the corresponding patient information and examination results.

[0224] A set of rules for checking for conflicts in medical orders is predefined in the rule engine. These rules can be formulated based on clinical medical knowledge, hospital treatment standards and procedures, such as rules on drug interactions and the order of examinations. Rules can be represented in the form of logical expressions, decision trees, etc.

[0225] The rules engine takes the first and second medical orders as input and checks them according to predefined rules. By traversing all rules, it determines if there are any conflicts in the medical order information. If a conflict is found, its specific content and location are recorded. Conflicts are marked, for example, with special colors, symbols, or identifiers, so that they can be clearly identified by nurses and physicians. The marked conflicts are then sent to the nurses and / or physicians via a wireless network. During transmission, encryption technology is used to ensure information security, and the message clearly indicates relevant information about the conflict, such as the patient involved, the content of the medical order, and the type of conflict.

[0226] The nurse's end and the doctor's end are also configured to: upon receiving the conflict item, highlight the conflict item and receive input review information for the conflict item.

[0227] Highlighting involves displaying conflicting information in a prominent manner (such as changing color or adding flashing effects) on the nurse's or doctor's interface to distinguish it from other information and attract the nurse's or doctor's attention. Review information consists of processing opinions or explanatory information entered by the nurse or doctor after verifying and analyzing the conflicting information, used to resolve the conflict or further clarify the treatment plan.

[0228] In this embodiment, after receiving information marked with conflicting items from the cloud via a wireless network, the nurse's or doctor's end parses the information and extracts the specific content of the conflicting items. On the user interface, the conflicting item information is displayed using a highlighting method. For example, conflicting medical orders are displayed in red font with a flashing animation effect, enabling nurses or doctors to quickly locate the conflicting items.

[0229] Nurses or physicians carefully review conflicting information, considering the patient's actual condition, clinical experience, and professional knowledge to determine whether the conflict truly exists and how to resolve it. The user interface provides interactive elements such as input boxes or selection boxes to facilitate the input of review information by nurses or physicians. Review information may include explanations of the conflict, revised medical orders, and items requiring further investigation.

[0230] The cloud platform is also configured to: update and adjust the comprehensive emergency information based on the review information, and generate a comprehensive emergency report based on the updated comprehensive emergency information using a push template corresponding to the risk level. The comprehensive emergency report is then pushed to the patient for display.

[0231] The risk level is determined based on factors such as the severity of the patient's condition and the complexity of the treatment plan, categorizing patients into different risk levels, such as low risk, medium risk, and high risk, according to the comprehensive emergency information. The risk level classification can be determined based on hospital-established standards and clinical experience. The push template is a pre-configured format and content framework for generating comprehensive emergency reports; different risk levels correspond to different push templates. The push template includes common content such as basic patient information, diagnosis, treatment plan, and precautions, with the level of detail and emphasis varying depending on the risk level. The comprehensive emergency report is a detailed report generated based on the updated and adjusted comprehensive emergency information and the selected push template. It comprehensively summarizes the patient's emergency visit, providing clear information about their condition and treatment guidance.

[0232] In this embodiment, after receiving review information from the nurse's or physician's end, the cloud server updates and adjusts the comprehensive emergency information. Based on the type and content of the review information, it modifies or supplements the corresponding medical orders, diagnoses, treatment plans, etc. For example, if the review information modifies the medication dosage, the cloud server will update the medication records in the comprehensive emergency information.

[0233] Based on the updated and adjusted comprehensive emergency information, a predefined risk assessment model is used to evaluate the patient's risk level. The risk assessment model can consider multiple factors, such as the patient's vital signs, examination results, disease type, age, etc., and derive the patient's risk level through weighted calculation or other algorithms.

[0234] Based on the risk level determined by the assessment, an appropriate push notification template is selected from a pre-designed template library. For example, for high-risk patients, a detailed and focused template is selected, emphasizing the severity of the condition and the urgency of treatment; for low-risk patients, a concise and clear template is selected, providing basic information about the condition and treatment suggestions.

[0235] The updated and adjusted comprehensive emergency information is filled in and formatted according to the selected push template to generate a comprehensive emergency report. During the generation process, ensure the report content is accurate and the format is standardized. The generated comprehensive emergency report is pushed to the patient's device via wireless network. During the push process, notifications can be sent to the patient via SMS, app notifications, etc., to remind them to view the report. After receiving the report, the patient's device parses and displays it, allowing the patient to easily understand their condition and treatment progress.

[0236] This disclosure also provides an intelligent triage method for emergency department stratification, applied in a cloud environment, wherein the cloud environment is the same as any one of the foregoing embodiments; see also Figure 2 As shown, the intelligent triage method for emergency department grading includes: In step S21, based on the symptom description of the emergency patient, the summary of the patient's past medical history, and the consultation time information uploaded by the patient terminal, a symptom dependency network is constructed based on a preset symptom association relationship. The nodes in the symptom dependency network correspond to specific symptom items, and the attributes of the edges include the symptom concurrency probability and the causal influence strength. The symptom description is generated by the patient terminal based on the multi-dimensional vital sign information of the emergency patient obtained. In step S22, a hierarchical priority calculation is performed on the symptom dependency network, and the urgency score of each symptom item is updated through the influence transmission between nodes to generate a dynamic urgency feature set. In step S23, the dynamic urgency feature set is input into the pre-trained triage decision model, and a graded assessment result including the grade level and intervention timeliness requirements is generated through a dual-track inference mechanism; In step S24, a recommended treatment department is determined based on the graded assessment results, and a triage plan including the priority of treatment and the necessary examination items is determined. The triage plan is pushed to the nurse's and doctor's terminals of the corresponding recommended treatment department by the cloud, so as to display the priority of treatment and the necessary examination items on the nurse's and doctor's terminals.

[0237] In a preferred embodiment, in step S23, inputting the dynamic urgency feature set into a pre-trained triage decision model and generating a graded assessment result including grading level and intervention timeliness requirements through a dual-track inference mechanism includes: In step S231, the dynamic urgency feature set is input into the feature allocation layer of the triage decision model. The feature allocation layer is used to divide the dynamic urgency features into core symptom feature streams and associated symptom feature streams based on the clinical attributes of the dynamic urgency features in the dynamic urgency feature set. In step S232, the core symptom feature streams are subjected to lethality assessment, intervention window calculation, and resource demand prediction sequentially through the core reasoning path of the triage decision model to generate core grading features. In step S233, the associated symptom feature streams are subjected to complication risk analysis, underlying disease impact assessment, and treatment conflict detection sequentially through the associated reasoning path of the triage decision model to generate associated adjustment features. In step S234, based on the core grading features and the associated adjustment features, the output layer of the triage decision model generates a grading assessment result including the grading level and the intervention time limit requirements.

[0238] In a preferred embodiment, in step S234, generating a grading assessment result including the grading level and the intervention timeliness requirement through the output layer of the triage decision model based on the core grading features and the associated adjustment features includes: in step S2341, using a dynamic weighting strategy based on clinical risk to weight and fuse the core grading features and the associated adjustment features to generate a fused feature; in step S2342, mapping the fused feature to an intervention time window parameter through the timeliness mapping layer of the triage decision model, the intervention time window parameter being used to quantify the treatment response timeliness corresponding to different grading levels; in step S2343, calling the priority calibration module of the triage decision model, adjusting the grading threshold based on the current emergency resource load, and performing clinical scenario adaptation adjustment on the intervention time window parameter to generate a target intervention time window parameter; in step S2344, generating a grading assessment result including the grading level and the intervention timeliness requirement through the output layer of the triage decision model based on the target intervention time window parameter.

[0239] In a preferred embodiment, in step S232, the step of sequentially performing lethality assessment, intervention window calculation, and resource demand prediction on the core symptom feature stream through the core reasoning path of the triage decision model to generate core grading features includes: in step S2321, inputting the core symptom feature stream into the lethality assessment submodule of the core reasoning path, and detecting lethal symptoms in the core symptom features through a pre-trained risk identification unit; in step S2322, performing the lethality assessment based on the lethal symptoms, determining the onset speed, severity, and systemic impact of the lethal symptoms, and generating a lethality score vector; and in step S2323, inputting the lethality score vector into the intervention window calculation submodule. In step S2324, the intervention window calculation is performed based on the preset clinical intervention criteria to determine the optimal implementation time period for different intervention measures and generate time window features. In step S2325, the intervention delay risk parameter is calculated based on the time window features and the lethality score vector. The intervention delay risk parameter is used to describe the severity of adverse consequences of missing the optimal intervention time. In step S2326, the intervention delay risk parameter is input into the resource demand prediction submodule to analyze the type and quantity of medical resources required to deal with the lethal symptom and generate resource demand features. In step S2327, the core grading features are generated by fusing the lethality score vector, the time window features, the intervention delay risk parameter, and the resource demand features.

[0240] In a preferred embodiment, in step S233, the step of sequentially performing complication risk analysis, underlying disease impact assessment, and treatment conflict detection on the associated symptom feature stream through the association reasoning path of the triage decision model to generate association adjustment features includes: in step S2331, inputting the associated symptom feature stream into the complication risk analysis submodule of the association reasoning path, and performing the complication risk analysis through the complication prediction unit to generate a complication risk feature vector; in step S2332, inputting the complication risk feature vector into the underlying disease impact assessment submodule to retrieve the patient's past medical history summary to perform the underlying disease impact assessment, thereby obtaining the interaction relationship between the underlying disease and the patient's symptoms. In step S2333, the underlying disease impact coefficient is generated; in step S2333, the complication risk feature vector is adjusted based on the underlying disease impact coefficient to strengthen the complication risk component of the complication risk vector that aggravates the underlying disease, thereby generating a target complication feature vector; in step S2334, the target complication feature vector is input into the treatment conflict detection submodule to perform the treatment conflict detection, thereby analyzing the compatibility between the recommended treatment plan for the emergency patient's symptoms and the treatment plan for the underlying disease, and generating a treatment conflict risk parameter, which is used to quantify the degree of impact of the conflict on the treatment effect; in step S2335, the adjusted complication risk feature vector, the underlying disease impact coefficient, and the treatment conflict risk parameter are fused to generate the correlation adjustment feature.

[0241] In a preferred embodiment, in step S2331, the step of inputting the associated symptom feature stream into the complication risk analysis submodule of the associated inference path, and performing the complication risk analysis through the complication prediction unit to generate a complication risk feature vector includes: inputting the associated symptom feature stream into the complication risk analysis submodule of the associated inference path, performing the complication risk analysis through the complication prediction unit to identify potential complication triggers in the associated symptom features; marking the complication type of potential complications with high risk of occurrence based on the potential complication triggers; and calculating the probability of occurrence of the potential complications based on the complication type, the duration of symptoms of the emergency patient, and the patient's basic physical condition to generate the complication risk feature vector.

[0242] In a preferred embodiment, in step S21, the step of constructing a symptom dependency network based on a preset symptom association relationship, according to the received description of the emergency patient's symptoms, the summary of the emergency patient's medical history, and the consultation time information uploaded by the patient, includes: extracting symptom items from the description of the symptoms, and determining independent symptom items and compound symptom clusters based on the symptom items, the summary of the medical history, and the consultation time information, wherein the compound symptom cluster consists of multiple simultaneously occurring associated symptom items; and searching a preset symptom association rule base based on the associated symptom items in the compound symptom cluster, and determining any two from the preset association types. The target association type between the associated symptom items, the preset association type includes concurrent association type, secondary association type and exclusion association type; based on the target association type and the frequency ratio of the co-occurrence of two symptoms in historical cases, the concurrent probability parameter between the associated symptom items is calculated; based on the description of symptom causal relationship and degree of influence recorded in clinical guidelines, the causal influence strength between the associated symptom items is evaluated; the symptom dependency network is constructed by taking each of the independent symptom items and each of the associated symptom items as nodes of the symptom dependency network, and using the concurrent probability parameter and the causal influence strength as attributes of the edges between nodes.

[0243] In a preferred embodiment, in step S22, the hierarchical urgency adjustment of the symptom dependency network, updating the urgency of each symptom item through inter-node influence transmission to generate a dynamic urgency feature set, includes: assigning an initial urgency baseline value to each node in the symptom dependency network, the initial urgency baseline value being determined based on the basic weight of the symptom in the emergency triage criteria; determining the influence transmission direction between nodes based on the causal influence strength of the edges in the symptom dependency network; according to the influence transmission direction, the source node forward transmits an influence packet including its own urgency baseline value and the corresponding causal influence strength to the target node, the target node adjusting its own initial urgency baseline value according to the parameters in the influence packet, and calculating the difference between the initial urgency baseline value and the adjusted urgency. An urgency deviation is obtained; based on the direction of influence transmission, the target node transmits the urgency deviation back to the source node, and the source node adjusts its influence weight on other nodes based on the urgency deviation; according to the node's hierarchical position in the symptom dependency network, multiple rounds of urgency adjustment are performed, with symptom items located upstream in the symptom dependency network prioritizing the completion of the urgency adjustment and transmitting the impact of the urgency adjustment to symptom items downstream in the symptom dependency network; after each round of adjustment, the causal influence strength between nodes is recalculated in conjunction with the real-time vital sign changes of the emergency patient, and the weight allocation of influence transmission is dynamically adjusted; when the urgency deviation of two adjacent times is lower than a preset threshold, the hierarchical urgency adjustment is stopped, and the final urgency of all symptom items is collected to form a dynamic urgency feature set.

[0244] In a preferred embodiment, the method further includes: receiving emergency visit feedback information and supplementary emergency remarks information returned by the nurse's end and / or the physician's end, wherein the emergency visit feedback information is generated based on first medical orders provided for the results of examinations for the required examinations, and the supplementary emergency remarks information is generated based on second medical orders provided for the results of additional examinations besides the required examinations; associating the supplementary emergency remarks information as an additional field with the emergency visit feedback information to generate comprehensive emergency information, and checking the relationship between the first medical orders and the second medical orders based on a rule engine. The system identifies and marks conflicting items, and sends the marked conflicting items to the nurse's end and / or the physician's end; it receives review information returned by the nurse's end and / or the physician's end, wherein, upon receiving a conflicting item, the nurse's end and / or the physician's end highlights the conflicting item and receives input review information for the conflicting item; it updates and adjusts the comprehensive emergency information based on the review information, and generates a comprehensive emergency report based on the updated and adjusted comprehensive emergency information using a push template corresponding to the risk level, the comprehensive emergency report being pushed to the patient's end for display.

[0245] The specific implementation methods described above have been explained in detail in the foregoing embodiments, and will not be repeated here.

[0246] Figure 3 The intelligent triage device 1000 for emergency department stratification shown can be configured as a cloud platform as described in any of the foregoing embodiments. The intelligent triage device 1000 includes a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, for example, via a bus 1002. Optionally, the intelligent triage device 1000 may further include a communication component 1004, which can be used for data interaction between the device 100 and other devices, such as sending and / or receiving data. It should be noted that in actual scheduling, the communication component 1004 is not limited to one, and the structure of this intelligent triage device 1000 does not constitute a limitation on the embodiments of this application.

[0247] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0248] Bus 1002 may include a pathway for transmitting information between the aforementioned components. Bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 1002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0249] The memory 1003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium capable of carrying or storing program code and capable of being read by a computer, without limitation herein.

[0250] The memory 1003 is used to store program code for executing embodiments of the present disclosure, and its execution is controlled by the processor 1001. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the aforementioned embodiments of the intelligent triage method for emergency grading.

[0251] This disclosure also provides a computer-readable storage medium storing program code. When the program code is executed by a processor, it can implement the steps and corresponding content of the aforementioned intelligent triage method embodiment for emergency triage.

[0252] The preferred embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings. However, the present disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present disclosure, various changes, modifications, substitutions and variations can be made to these embodiments, and all such changes, modifications, substitutions and variations fall within the protection scope of the present disclosure.

[0253] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction, and such combinations should also be considered as part of this disclosure. To avoid unnecessary repetition, this disclosure will not further describe the various possible combinations. The technical scope of this application is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. An intelligent triage system for emergency department grading, characterized in that, include: The patient terminal, the nurse terminal, the physician terminal, and the cloud for communicating with the patient terminal, the nurse terminal, and the physician terminal; The patient terminal is configured to acquire multi-dimensional vital sign information of emergency patients and generate a description of the symptoms of the emergency patients based on the vital sign information. The cloud platform is configured to construct a symptom dependency network based on a preset symptom association relationship, using the symptom description, medical history summary, and consultation time information of the emergency patient uploaded by the patient terminal. Nodes in the symptom dependency network correspond to specific symptom items, and the attributes of the edges include the probability of symptom co-occurrence and the strength of causal influence. The platform performs hierarchical urgency adjustment on the symptom dependency network, updates the urgency of each symptom item through the influence transmission between nodes, generates a dynamic urgency feature set, and inputs the dynamic urgency feature set into a pre-trained triage decision model. Through a dual-track reasoning mechanism, it generates a graded assessment result including grade level and intervention timeliness requirements, and determines the recommended treatment department and a triage plan including consultation priority and necessary examination items based on the graded assessment result. The nurse's terminal and the doctor's terminal are configured to receive and display the triage plan sent from the cloud, which includes the treatment priority and the required examination items for the emergency patient.

2. The intelligent triage system for emergency department grading according to claim 1, characterized in that, The process involves inputting the dynamic urgency feature set into a pre-trained triage decision model, and generating a tiered assessment result including classification level and intervention timeliness requirements through a dual-track inference mechanism, including: The dynamic urgency feature set is input into the feature allocation layer of the triage decision model. The feature allocation layer is used to divide the dynamic urgency features into core symptom feature stream and associated symptom feature stream according to the clinical attributes of the dynamic urgency features in the dynamic urgency feature set. The core reasoning path of the triage decision model sequentially performs lethality assessment, intervention window calculation, and resource demand prediction on the core symptom feature stream to generate core hierarchical features. The associated reasoning path of the triage decision model is used to sequentially perform complication risk analysis, underlying disease impact assessment and treatment conflict detection on the associated symptom feature flow to generate associated adjustment features; Based on the core grading features and the associated adjustment features, the output layer of the triage decision model generates a grading assessment result that includes the grading level and the intervention timeliness requirements.

3. The intelligent triage system for emergency department grading according to claim 2, characterized in that, The step of generating a tiered assessment result, including the tier level and the intervention timeliness requirement, through the output layer of the triage decision model based on the core tiered features and the associated adjustment features includes: A dynamic weighting strategy based on clinical risk is adopted to weight and fuse the core grading features and the associated adjustment features to generate fused features; The fusion features are mapped to intervention time window parameters through the timeliness mapping layer of the triage decision model. The intervention time window parameters are used to quantify the treatment response time limits corresponding to different levels. The priority calibration module of the triage decision model is invoked, and the triage threshold is adjusted in conjunction with the current resource load of the emergency department to adapt the intervention time window parameters to the clinical scenario, thereby generating the target intervention time window parameters. Based on the target intervention time window parameters, the output layer of the triage decision model generates a graded assessment result including the grade level and the intervention timeliness requirements.

4. The intelligent triage system for emergency department grading according to claim 2, characterized in that, The core reasoning path of the triage decision model sequentially performs lethality assessment, intervention window calculation, and resource requirement prediction on the core symptom feature stream to generate core hierarchical features, including: The core symptom feature stream is input into the lethality assessment submodule of the core reasoning path, and the lethal symptoms in the core symptom features are detected by the pre-trained risk recognition unit. The fatality assessment is performed based on the fatal symptoms to determine the onset speed, severity, and systemic involvement of the fatal symptoms, and to generate a fatality score vector. The lethality score vector is input into the intervention window calculation submodule, and the intervention window calculation is performed in combination with the preset clinical intervention criteria to determine the optimal implementation time period for different intervention measures and generate time window features. Based on the time window characteristics and the lethality score vector, an intervention delay risk parameter is calculated, which is used to describe the severity of the adverse consequences of missing the optimal intervention time. The intervention delay risk parameters are input into the resource demand prediction submodule to analyze the types and quantities of medical resources required to address the fatal symptom and generate resource demand characteristics. The core grading features are generated by integrating the lethality score vector, the time window features, the intervention delay risk parameters, and the resource requirement features.

5. The intelligent triage system for emergency department grading according to claim 2, characterized in that, The process involves sequentially performing complication risk analysis, underlying disease impact assessment, and treatment conflict detection on the associated symptom feature stream through the associative reasoning path of the triage decision model, generating associative adjustment features, including: The associated symptom feature stream is input into the complication risk analysis submodule of the associated inference path, and the complication risk analysis is performed by the complication prediction unit to generate a complication risk feature vector; The complication risk feature vector is input into the underlying disease impact assessment submodule to retrieve the patient's past medical history summary and perform the underlying disease impact assessment to obtain the interaction relationship between the underlying disease and the patient's symptoms, and generate the underlying disease impact coefficient. The complication risk feature vector is adjusted based on the underlying disease impact coefficient to strengthen the complication risk component of the complication risk vector that aggravates the underlying disease, thereby generating a target complication feature vector. The target complication feature vector is input into the treatment conflict detection submodule to perform the treatment conflict detection, so as to analyze the compatibility between the recommended treatment plan for the emergency patient's symptoms and the treatment plan for the underlying disease, and generate treatment conflict risk parameters. The treatment conflict risk parameters are used to quantify the degree of impact of the conflict on the treatment effect. The adjusted complication risk feature vector, the underlying disease impact coefficient, and the treatment conflict risk parameter are integrated to generate the associated adjusted feature.

6. The intelligent triage system for emergency department grading according to claim 5, characterized in that, The complication risk analysis submodule, which inputs the associated symptom feature stream into the associated inference path, performs the complication risk analysis through the complication prediction unit to generate a complication risk feature vector, including: The associated symptom feature stream is input into the complication risk analysis submodule of the associated inference path, and the complication risk analysis is performed by the complication prediction unit to identify potential complication triggers in the associated symptom features; Based on the potential complication triggers, the types of potential complications with high risk of occurrence are identified; Based on the type of complication, the duration of symptoms of the emergency patient, and the patient's basic physical condition, the probability of occurrence of the potential complications is calculated, and a complication risk feature vector is generated.

7. The intelligent triage system for emergency department grading according to any one of claims 1-6, characterized in that, The step involves constructing a symptom dependency network based on a preset symptom association relationship, using the received description of the emergency patient's symptoms, a summary of the emergency patient's past medical history, and the time of their visit, uploaded from the patient's end. This includes: Extract symptom items from the symptom description, and determine independent symptom items and compound symptom clusters based on the symptom items, the summary of past medical history, and the consultation time information, wherein the compound symptom cluster consists of multiple related symptom items that occur simultaneously; Based on the associated symptom items in the complex symptom cluster, a preset symptom association rule base is retrieved, and a target association type between any two associated symptom items is determined from the preset association types. The preset association types include concurrent association types, secondary association types, and exclusion association types. Calculate the concurrency probability parameter between the associated symptom items based on the target association type and the frequency ratio of the co-occurrence of the two symptoms in historical cases; Based on the descriptions of symptom causal relationships and degree of influence in clinical guidelines, assess the strength of causal influence among the associated symptom items; The symptom dependency network is constructed by using each of the independent symptom items and each of the associated symptom items as nodes of the symptom dependency network, and using the concurrency probability parameter and the causal influence strength as attributes of the edges between nodes.

8. The intelligent triage system for emergency department grading according to any one of claims 1-6, characterized in that, The hierarchical urgency adjustment of the symptom dependency network is performed, updating the urgency of each symptom item through the influence propagation between nodes, generating a dynamic urgency feature set, including: Each node in the symptom-dependent network is assigned an initial urgency baseline value, which is determined based on the basic weight of symptoms in the emergency triage criteria. Based on the causal influence strength of the edges in the symptom dependency network, the direction of influence transmission between nodes is determined; According to the direction of influence propagation, the source node transmits an influence packet, including its own urgency baseline value and the corresponding causal influence intensity, to the target node in a forward direction. The target node adjusts its initial urgency baseline value according to the parameters in the influence packet and calculates the difference between the initial urgency baseline value and the adjusted urgency value to obtain the urgency deviation. According to the direction of influence propagation, the target node transmits the urgency deviation back to the source node, and the source node adjusts its own influence weight on other nodes based on the urgency deviation; Based on the node's hierarchical position in the symptom dependency network, multiple rounds of urgency adjustment are performed. Symptom items located upstream in the symptom dependency network complete the urgency adjustment first and transmit the impact of the urgency adjustment to symptom items located downstream in the symptom dependency network. After each round of adjustment, the causal influence strength between nodes is recalculated based on the real-time changes in the vital signs of the emergency patients, and the weight allocation of influence transmission is dynamically adjusted. When the urgency deviation between two consecutive measurements is lower than a preset threshold, the hierarchical urgency adjustment is stopped, and the final urgency of all the symptom items is collected to form a dynamic urgency feature set.

9. The intelligent triage system for emergency department grading according to any one of claims 1-6, characterized in that, The nurse's terminal and the physician's terminal are further configured to: in response to emergency visit feedback information and supplementary emergency remarks information input on the nurse's terminal and / or on the physician's terminal for the emergency patient, return the emergency visit feedback information and the supplementary emergency remarks information to the cloud, wherein the emergency visit feedback information is generated based on first medical order information provided based on the examination results of the required examination items, and the supplementary emergency remarks information is generated based on second medical order information provided based on the examination results of additional examination items besides the required examination items; The cloud is also configured to: associate the supplementary emergency remarks information as an additional field with the emergency visit feedback information to generate comprehensive emergency information, and check for conflicts between the first medical order information and the second medical order information based on a rule engine, mark the conflicts, and send the marked conflicts to the nurse's end and / or the doctor's end. The nurse's end and the doctor's end are also configured to: upon receiving the conflict item, highlight the conflict item and receive input review information for the conflict item; The cloud platform is also configured to: update and adjust the comprehensive emergency information based on the review information, and generate a comprehensive emergency report based on the updated comprehensive emergency information using a push template corresponding to the risk level. The comprehensive emergency report is then pushed to the patient for display.

10. An intelligent triage method for emergency department grading, characterized in that, Applied to the cloud, wherein the cloud is the cloud described in any one of claims 1-9; The intelligent triage method for emergency department grading includes: Based on the symptom description of the emergency patient, the summary of the patient's past medical history, and the time of visit uploaded by the patient, a symptom dependency network is constructed based on a preset symptom association relationship. The nodes in the symptom dependency network correspond to specific symptom items, and the attributes of the edges include the probability of symptom co-occurrence and the strength of causal influence. The symptom description is generated by the patient based on the multi-dimensional vital sign information of the emergency patient obtained by the patient. Hierarchical priority calculation is performed on the symptom dependency network, and the urgency score of each symptom item is updated through the influence transmission between nodes to generate a dynamic urgency feature set. The dynamic urgency feature set is input into the pre-trained triage decision model, and a graded assessment result including the grade level and intervention timeliness requirements is generated through a dual-track reasoning mechanism. Based on the tiered assessment results, a recommended treatment department is determined, along with a triage plan that includes the priority of treatment and the necessary examination items. The triage plan is pushed to the corresponding nurse and physician terminals in the cloud, so that the priority of treatment and the necessary examination items are displayed on the nurse and physician terminals.

Images