A medical guide semantic understanding method fusing time and intensity features

Abstract

The application relates to the technical field of medical guidance, and discloses a medical guidance semantic understanding method fusing time and intensity characteristics, which comprises the following steps: constructing a multi-dimensional symptom feature vector library based on a medical symptom knowledge graph; generating an input text based on patient voice, determining a query vector based on the input text, determining target information through the multi-dimensional symptom feature vector library, and determining a three-dimensional joint probability based on the target information; determining a time weighting factor and an intensity weighting factor based on the input text, determining a coupling weighting weight based on the time weighting factor and the intensity weighting factor, and determining a fusion weighting vector based on the time weighting factor, the intensity weighting factor, the coupling weighting weight and the three-dimensional joint probability; and determining a triage decision through a multi-task decision layer based on the fusion weighting vector. Through the improvement of the guidance method, the technical problems that the symptom timeliness weight and the intensity grade cannot be effectively extracted, the symptom-time-intensity joint risk assessment mechanism is lacked, and the multi-symptom co-occurrence relationship is insufficiently mined are solved.

Description

Technical Field

[0001] This invention relates to the field of medical triage technology, specifically to a medical triage semantic understanding method that integrates time and intensity features. Background Technology

[0002] The medical triage system is a core component of smart hospital construction. Its main function is to automatically determine the department and urgency of the patient based on the patient's symptom description, providing the patient with accurate triage and guidance services. It undertakes the important mission of optimizing the medical process, alleviating outpatient pressure, and improving the efficiency of medical services.

[0003] Semantic understanding technology based on natural language processing is a key supporting module of medical triage systems. Existing semantic understanding methods for medical triage mainly use pre-trained language models to semantically encode patient descriptions and output department recommendation results through a classifier.

[0004] However, in actual clinical triage scenarios, existing technologies suffer from a fundamental technical deficiency: they cannot effectively extract and utilize the timeliness and intensity characteristics of symptom descriptions. In clinical practice, the time attributes and intensity levels of symptoms are key factors in determining the urgency and type of a disease. "Sudden chest pain for 1 hour" and "intermittent onset for 3 days" point to completely different disease spectrums, and "severe chest pain" and "mild chest tightness" also have drastically different risk implications under different symptom contexts. However, existing methods treat time and intensity modifiers as ordinary tokens in attention calculations, failing to assign them independent decision weights. This results in a significant loss of medical semantic information regarding time and intensity characteristics during feature extraction. This lack of a technical mechanism makes it difficult for the system to distinguish the essential difference in urgency between "sudden chest pain less than 1 hour" and "chest pain lasting 3 days," and it also fails to identify the potentially high-risk condition hidden behind the low-intensity description of "mild chest tightness accompanied by cold sweats." Consequently, high-risk diseases such as acute coronary syndrome and aortic dissection, which are indicated by specific time patterns and intensity characteristics, are missed, seriously affecting triage accuracy and patient safety. Summary of the Invention

[0005] To overcome the aforementioned technical problems in the prior art, this invention provides a medical triage semantic understanding method that integrates time and intensity features. By improving the existing triage method, it effectively solves the technical problems of existing technologies, such as the inability to effectively extract the timeliness weights and intensity levels of symptoms, the lack of a symptom-time-intensity joint risk assessment mechanism, insufficient mining of co-occurrence relationships of multiple symptoms, and the lack of interpretability of the decision-making process.

[0006] To achieve the above objectives, embodiments of the present invention provide a medical triage semantic understanding method, the method comprising: constructing a multidimensional symptom feature vector library based on a medical symptom knowledge graph, the multidimensional symptom feature vector library including a joint matrix; generating input text based on patient speech, determining a query vector based on the input text, determining target information based on the query vector through the multidimensional symptom feature vector library, and determining a three-dimensional joint probability based on the target information; determining a time-weighted factor and an intensity-weighted factor based on the input text, determining a coupling weighting weight based on the time-weighted factor and the intensity-weighted factor, and determining a fusion weighted vector based on the time-weighted factor, the intensity-weighted factor, the coupling weighting weight, and the three-dimensional joint probability; and determining a triage decision through a multi-task decision layer based on the fusion weighted vector.

[0007] Preferably, the construction of a multidimensional symptom feature vector library based on a medical symptom knowledge graph includes: determining symptom entities, disease types, and association information between the symptom entities and the disease types based on the medical symptom knowledge graph, wherein the association information includes time features and intensity features; constructing a probability matrix based on the symptom entities and the disease types, wherein the probability matrix includes the symptom entities, the disease types, and the probability of the occurrence of the symptom entities; constructing a time matrix based on the disease types and the time features, wherein the time matrix includes the disease types, the time features, and a time correction factor; constructing an intensity matrix based on the disease types and the intensity features, wherein the intensity matrix includes the disease types, the intensity features, and an intensity correction factor; and constructing a joint matrix based on the probability matrix, the time matrix, and the intensity matrix, wherein the joint matrix includes the correspondence between the probability matrix, the time matrix, and the intensity matrix.

[0008] Preferably, determining the query vector based on the input text includes: obtaining the text information after the input text has been converted; using a bidirectional long short-term memory network conditional random field model to annotate the text information with medical entities to obtain medical entities, the medical entities including symptom entities, time entities, intensity entities, and modifier entities; determining medical features based on the medical entities through a medical pre-trained model, the medical features including symptom features, time features, intensity features, and modifier entities; and constructing a query vector based on the medical features.

[0009] Preferably, determining the time weighting factor and intensity weighting factor based on the input text includes: constructing a dual-channel weighting layer, the dual-channel weighting layer including a time channel and an intensity channel; determining time modifiers based on the input text through the time channel, calculating the syntactic dependency distance between the time modifiers and symptom entities, and determining the time weighting factor when the syntactic dependency distance meets a first preset condition; determining the symptom intensity level based on the input text through the intensity channel matching, and determining the intensity weighting factor when the symptom intensity level meets a second preset condition.

[0010] Preferably, the dual-channel weighting layer further includes a time-intensity coupled attention layer; through the coupled attention layer, the coupled weighting weight is calculated based on the time weighting factor and the intensity weighting factor using a coupled amplification effect function.

[0011] Preferably, the dual-channel weighted layer further includes a symptom co-occurrence map; based on the symptom co-occurrence map, a co-occurrence enhancement factor is determined by a graph attention network; and the fusion weighting vector is determined based on the coupling weighting weight, the co-occurrence enhancement factor, and the three-dimensional joint probability.

[0012] Preferably, the multi-task decision layer includes a disease prediction task layer, a department recommendation task layer, an urgency grading task layer, and a risk scoring task layer; the disease prediction task layer determines the final probability of the disease based on the fusion weighted vector; the department recommendation task layer determines the recommended department based on the final probability of the disease; and the urgency grading task layer determines the urgency based on the final probability of the disease, the coupling weighting weight, the co-occurrence enhancement factor, and the time weighting factor.

[0013] Preferably, the step of determining triage decisions through a multi-task decision layer based on the fusion weighted vector further includes: calculating the output distribution entropy value, and triggering follow-up questioning script generation or transferring to a human agent channel when the entropy value meets a third preset condition.

[0014] To achieve the above objectives, embodiments of the present invention provide a medical triage semantic understanding system, including a control module for implementing a medical triage semantic understanding method. The control module includes: a vector module for determining and constructing a multi-dimensional symptom feature vector library based on a medical symptom knowledge graph, the multi-dimensional symptom feature vector library including a probability matrix, a time matrix, an intensity matrix, and a joint matrix; an input module for generating input text based on patient speech, determining a query vector based on the input text, and determining a three-dimensional joint probability based on the query vector using the multi-dimensional symptom feature vector library; a weighting module for determining a time-weighted factor and an intensity-weighted factor based on the input text, determining a coupling weighting weight based on the time-weighted factor and the intensity-weighted factor, determining a coupling weighting weight based on the coupling weighting weight and the three-dimensional joint probability, and determining a fusion weighted vector based on the coupling weighting weight, the time-weighted factor, the intensity-weighted factor, and the three-dimensional joint probability; and a decision module for determining a triage decision through a multi-task decision layer based on the fusion weighted vector.

[0015] To achieve the above objectives, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a medical triage semantic understanding method.

[0016] The present invention has at least the following technical effects through the technical solution provided by the present invention:

[0017] By enhancing the medical semantics of temporal features through a time-disease correction factor matrix, the system can distinguish between sudden and intermittent chest pain, significantly improving the detection rate of high-risk diseases such as acute coronary syndrome and aortic dissection. Similarly, by enhancing the medical semantics of intensity features through an intensity-disease correction factor matrix, the system can differentiate between different risk levels and avoid isolated misjudgments based on intensity grades. Furthermore, by using nonlinear coupling calculations of three-dimensional joint probabilities, the system achieves joint modeling of time, intensity, and entity dimensions. When multiple dimensions point to high risk, a coupling amplification effect is triggered, improving diagnostic accuracy. A symptom co-occurrence graph attention mechanism captures joint diagnostic patterns, enabling in-depth mining of symptom relationships. An interpretable output link provides clear decision-making support for healthcare professionals and patients, enhancing system credibility. Finally, a continuous learning mechanism supports dynamic updates to the knowledge base, ensuring the system's long-term effectiveness.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 The flowchart illustrates the specific implementation of the medical triage semantic understanding method that integrates time and intensity features, as provided in this embodiment of the invention. Detailed Implementation

[0021] 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.

[0022] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0023] Please see Figure 1 This invention provides a medical triage semantic understanding method that integrates time and intensity features. The method can be executed by a control module, which collects, analyzes, and processes data. The control module can include various components, such as at least one processor, microprocessor, control circuit, or cloud server. The control module can include a vector module, an input module, a weighting module, and a decision module. The method includes:

[0024] Step S10: Construct a multi-dimensional symptom feature vector library based on the medical symptom knowledge graph. This step can be performed by the vector module.

[0025] A medical symptom knowledge graph is a graph structure that reflects the association between symptom entities and disease types. It can be constructed based on a large amount of historical medical records. A medical symptom knowledge graph can include multiple nodes and multiple edges.

[0026] The nodes of a medical symptom knowledge graph include entities representing different disease types and different symptoms. For example, disease types can include... Migraine entity dizziness, headache

[0027] Edges can be directed, such as unidirectional or bidirectional edges. Edges describe the association between symptom entities and disease types, and their content can include temporal and intensity characteristics of the symptom. Temporal characteristics refer to information related to the time of the symptom, such as the duration of the symptom, the time of onset, and the time interval between two symptom episodes. Intensity characteristics reflect the severity of the symptom, such as mild pain, severe pain, extreme pain, mild sweating, and profuse sweating.

[0028] When the temporal and / or intensity characteristics of a symptom meet certain requirements, the symptom can be a symptom corresponding to a specific disease, or a disease can present with multiple different symptoms, each with distinct temporal and intensity characteristics. For example, sudden chest pain lasting one hour and accompanied by severe pain may indicate an acute myocardial infarction or other diseases. The edge connecting chest pain and acute myocardial infarction could include sudden onset within one hour and severe pain, while the edges connecting chest pain and other diseases could include the corresponding temporal and intensity characteristics of the symptoms. Acute myocardial infarction can also be connected to other symptoms such as radiating pain, profuse sweating, shortness of breath, and chest tightness by edges, with different edges including corresponding temporal and intensity characteristics. In one possible implementation, the edges can also include probability information of a disease corresponding to the occurrence of a certain symptom. For example, there is a 45% probability of chest pain occurring during an acute myocardial infarction.

[0029] The multidimensional symptom feature vector library includes a joint matrix. The joint matrix is ​​used to describe the correspondence between symptom entities, disease types, temporal features, intensity features, and probability information.

[0030] Step S20 involves generating input text based on the patient's speech, determining a query vector based on the input text, determining target information based on the query vector using a multi-dimensional symptom feature vector library, and determining a three-dimensional joint probability based on the target information. This step can be performed by the input module.

[0031] Input text refers to information related to the patient's symptoms as described by the patient. For example, a patient might input "sudden onset of chest pain for 1 hour, severe and unbearable, accompanied by cold sweats" through an input device. Input devices can include various methods, such as at least one of a microphone, touchscreen, keyboard, mini-program, or self-service terminal.

[0032] In one possible implementation, the input terminal may include a touchscreen and multiple optional options on the touchscreen, including symptom options, time options, and intensity options. Symptom options can correspond to multiple symptoms, such as headache or chest tightness; time options can correspond to multiple time-related options, such as duration (1 hour), sudden onset, or two days ago; and intensity options can correspond to multiple options related to symptom intensity, such as mild, moderate, or severe. Patients can select and press the relevant options based on their actual situation, improving the efficiency and accuracy of the input text acquisition, avoiding vague user descriptions that could affect subsequent judgment accuracy, or preventing the input of excessive irrelevant information.

[0033] A query vector is a vector generated by the input module after extracting symptom-related information from the input text. It is used to query the multidimensional symptom feature vector to obtain target information related to the input text. Target information may include symptom entities, disease types, posterior probabilities of the disease, temporal features, temporal correction factors, intensity features, and intensity correction factors, as well as their corresponding relationships, in the joint matrix.

[0034] The three-dimensional joint probability is used to describe the types of diseases a patient may have and the probability of having those diseases. For example, if a patient inputs "sudden chest pain for 1 hour, severe and unbearable accompanied by cold sweats" through the input terminal, the input module can output the probability of acute myocardial infarction as 0.39, the probability of aortic dissection as 0.33, and the probability of angina pectoris as 0.28. In one possible implementation, the three-dimensional joint probability can be calculated based on the time features, intensity features, and probability information in the joint matrix.

[0035] Step S30 involves determining the time-weighted factor and intensity-weighted factor based on the input text, determining the coupling weighting based on the time-weighted factor and intensity-weighted factor, and determining the fusion weighting vector based on the coupling weighting and the three-dimensional joint probability. This step can be performed by the weighting module.

[0036] The time-weighted factor and intensity-weighted factor are weights used to unilaterally adjust the three-dimensional joint probability of a patient having a certain disease. When the input text contains specific time-related words, such as "sudden onset 1 hour ago" or "three consecutive days," the weighting module can determine the corresponding time-weighted factor. When the input text contains specific intensity-related words, such as "severe pain" or "unbearable pain," the weighting module can determine the corresponding intensity-weighted factor.

[0037] In one possible implementation, the coupling weighting can be positively correlated with the time weighting factor and the intensity weighting factor, respectively. The weighting module can calculate the coupling weighting in various ways. For example, the weighting module can calculate the sum or product of the time weighting factor and the intensity weighting factor, and use the product of this sum or product with another coefficient as the coupling weighting. The coefficient can be a preset value and can be set according to actual needs.

[0038] In one possible implementation, the weighting module determines the fusion weighting vector based on a time weighting factor, an intensity weighting factor, a coupling weighting weight, and a three-dimensional joint probability. The fusion weighting vector may include... .in, , , , These can represent the three-dimensional joint probabilities of different diseases. It can represent a time-weighted factor. It can represent an intensity weighting factor. It can represent coupled weighted weights.

[0039] Step S40: Based on the fusion weighted vector, a triage decision is determined through a multi-task decision layer. This step can be executed by a decision module.

[0040] Triage decision refers to the initial recommendations provided to patients, which may include at least one of the following: disease prediction, departmental recommendation, and urgency. Disease prediction refers to the prediction of at least one disease that the patient may have.

[0041] The decision-making module can calculate the probability of different diseases based on the fusion weighted vector, select the disease with the highest probability as the disease prediction result, and recommend the patient to visit the department corresponding to the disease.

[0042] Urgency level describes the degree to which a patient needs medical attention. Urgency levels are categorized into different grades, such as Level 1, Level 2, and Level 3, ranked from lowest to highest severity. Level 1 urgency indicates the patient's condition is relatively mild and they can wait in line for treatment. Level 2 urgency indicates the patient's condition is relatively severe and they can be prioritized for treatment. Level 3 urgency indicates the patient's condition is extremely severe and requires immediate attention.

[0043] In one possible implementation, the decision-making module may include an output terminal for outputting triage decisions for patient confirmation. The output terminal may include various devices, such as a speaker, display screen, or printer. The printer can print paper copies of the triage decisions.

[0044] The medical triage semantic understanding method provided in some embodiments of this invention can pre-determine a patient's disease based on the patient's input text before the patient's visit, thereby facilitating the recommendation of appropriate departments and reducing the burden of department selection for patients, thus improving the efficiency of their medical visits. By determining the coupling weighting, the correlation between the patient's symptoms, time, and intensity can be fully considered, and correction coefficients related to the probability of disease occurrence can be determined. This corrects the probability of different diseases in the three-dimensional joint probability, improving the accuracy of determining the probability of disease occurrence and the accuracy of department recommendations, avoiding delays in patient visits, and enhancing the patient's user experience. Through a multi-task decision layer, triage decisions can be output, making the triage decisions visible so that patients can refer to the triage decisions to visit the corresponding departments. By determining the urgency, the method can also determine the degree of urgency required for a patient's visit, facilitating the immediate or priority treatment of patients requiring urgent care, thus helping to ensure the patient's safety.

[0045] In one possible implementation, constructing a multidimensional symptom feature vector library based on a medical symptom knowledge graph includes the following:

[0046] The vector module identifies symptom entities, disease types, and the relationships between them based on a medical symptom knowledge graph. These relationships include temporal and intensity features. The vector module can process the medical symptom knowledge graph using data processing models, such as the TransE knowledge graph embedding model. The TransE knowledge graph embedding model maps the medical symptom knowledge graph into low-dimensional knowledge embedding vectors, extracting symptom entities, disease types, disease probabilities, temporal features, and intensity features from the medical symptom knowledge graph.

[0047] The vector module constructs a probability matrix based on symptom entities and disease types. The probability matrix includes symptom entities, disease types, and the probability of symptom entities occurring.

[0048] In one possible implementation, the vector module can statistically analyze the prior probabilities of diseases in the medical symptom knowledge graph. and conditional likelihood probability .

[0049] The prior probability of a disease describes the proportion of disease D in the total number of diseases. The conditional likelihood probability describes the proportion of the frequency of symptom S among diagnosed diseases D.

[0050] Vector modules can provide prior probabilities of diseases. and conditional likelihood probability Determine the posterior probability The posterior probability is used to describe the probability of having a given symptom S and the prevalence of various diseases D.

[0051] The vector module can calculate the posterior probability based on the first algorithm, which may include formula (1):

[0052]

[0053] in, The normalized denominator refers to calculating the product of the prior probability and the likelihood probability of each disease after iterating through all diseases, and then summing all the products to ensure that the sum of the probabilities of all diseases under the same symptom S is equal to 1.

[0054] For example, the probability matrix may include the following table (1):

[0055]

[0056] In one possible implementation, the vector module can construct a time matrix based on disease type and time features, the time matrix including disease type, time features and time correction factor.

[0057] In one possible implementation, the vector module can use an L1 regularized multi-class logistic regression model. Taking the disease type and temporal features corresponding to the medical symptom knowledge graph as input, the model outputs the time advantage ratio (OR) values ​​of different temporal features on the probability of disease incidence. The vector module then determines a time correction factor based on these OR values. The OR value describes the degree of influence of a temporal feature on the risk of disease. A OR value greater than 1 indicates that the temporal feature increases the risk of disease. A OR value equal to 1 indicates that the temporal feature has no effect on the risk of disease. A OR value less than 1 indicates that the temporal feature reduces the risk of disease. The L1 regularized multi-class logistic regression model is existing technology, and its specific content and training methods are not considered improvements to this solution and will not be elaborated upon here.

[0058] The time correction factor is positively correlated with the time advantage ratio (OR).

[0059] The vector module can count the total number of patients with a certain symptom in the medical symptom knowledge graph, such as chest pain. Based on the total number of patients, the number of patients diagnosed with disease type D with time feature T, diagnosed with disease type D without time feature T, diagnosed without disease type D with time feature T, and diagnosed without disease type D without time feature T can be counted separately. For example, the disease type is myocardial infarction, which is one of the multiple disease types corresponding to chest pain, and the time feature is nocturnal onset. The disease type D-time feature T association table is constructed (2):

[0060]

[0061] The disease type D-time feature T association table (2) can be used as the input of the L1 regularized multi-class logistic regression model.

[0062] L1 regularized multi-class logistic regression models can calculate P(illness | presence of this time feature) and P(illness | absence of this time feature). P(illness | presence of this time feature) describes the probability that a patient has a certain disease given a given symptom and a given time feature. For example, given chest pain, the probability that a patient experiences nighttime chest pain is a myocardial infarction. P(illness | absence of this time feature) describes the probability that a patient has a certain disease given a given symptom and a given time feature is not present. For example, given chest pain, the probability that a patient does not experience nighttime chest pain is a myocardial infarction.

[0063] The L1 regularized multi-class logistic regression model can calculate P(illness | presence of this time feature) and P(illness | absence of this time feature) using a second algorithm. The second algorithm can include formulas (2) and (3):

[0064]

[0065]

[0066] The L1 regularized multi-class logistic regression model can calculate the time advantage ratio (OR) based on P(illness | presence of this time feature) and P(illness | absence of this time feature) using a third algorithm. The third algorithm can include formula (4):

[0067]

[0068] Here, OR refers to the time advantage ratio OR value.

[0069] The vector module can calculate the time correction factor based on the time advantage ratio (OR) using the fourth algorithm. The fourth algorithm may include formula (5):

[0070]

[0071] Where e is the natural constant. k is the scaling hyperparameter, and the value of k ranges from 0.4 to 0.6, for example, 0.5.

[0072] For the same symptom or the same disease type, there may be multiple time characteristics, such as nocturnal onset, onset three days ago, onset after meals, etc. Time correction factors can be calculated separately for each of the different time characteristics.

[0073] The vector module then constructs a time matrix based on disease type, temporal characteristics, and time correction factors. For example, the time matrix may include the following table (3):

[0074]

[0075] In one possible implementation, the vector module constructs an intensity matrix based on disease type and intensity features, the intensity matrix including disease type, intensity features and intensity correction factor.

[0076] In one possible implementation, the vector module can use an ordered multi-class logistic regression model, taking the disease type and intensity features corresponding to the medical symptom knowledge graph as input, and outputting the intensity advantage ratio (OR) of different intensity features on the probability of disease occurrence through the ordered multi-class logistic regression model. The vector module then determines the intensity correction factor based on the intensity advantage ratio (OR).

[0077] The odds ratio (OR) is used to describe the degree of influence of an intensity feature on disease risk. An OR greater than 1 indicates that the intensity feature increases the risk of disease. An OR of 1 indicates that the intensity feature has no effect on disease risk. An OR less than 1 indicates that the intensity feature decreases the risk of disease. The ordered multi-class logistic regression model is existing technology; its specific content and training methods are not considered improvements in this approach and will not be elaborated upon here.

[0078] The intensity correction factor is positively correlated with the intensity advantage ratio (OR).

[0079] The vector module can count the total number of patients with a certain symptom in the medical symptom knowledge graph, such as chest pain. Based on the total number of patients, the number of patients diagnosed with disease type D with intensity feature Q, diagnosed with disease type D without intensity feature Q, diagnosed without disease type D with intensity feature Q, and diagnosed without disease type D without intensity feature Q can be counted separately. For example, the disease type is myocardial infarction, which is one of the multiple disease types corresponding to chest pain, and the intensity feature is severe pain. The disease type D-intensity feature Q association table is constructed (4):

[0080]

[0081] The disease type D-intensity feature Q association table (2) can be used as input for an ordered multi-class logistic regression model.

[0082] An ordered multi-class logistic regression model can calculate P(disease | presence of this intensity feature) and P(disease | absence of this intensity feature). P(disease | presence of this intensity feature) describes the probability that a patient has a certain disease given a given symptom and the patient exhibits a certain intensity feature. For example, given chest pain, the probability that a patient experiences severe chest pain indicates a heart attack. P(disease | absence of this intensity feature) describes the probability that a patient has a certain disease given a given symptom and the patient does not exhibit a certain intensity feature. For example, given chest pain, the probability that a patient experiences chest pain but not severe chest pain indicates a heart attack.

[0083] In one possible implementation, the method for calculating the strength advantage ratio (OR) and strength correction factor in an ordered multi-class logistic regression model is the same as the method for calculating the time advantage ratio (OR) and time correction factor in an L1 regularized multi-class logistic regression model. For more information on calculating the strength advantage ratio (OR) and strength correction factor, please refer to the previous description of calculating the time advantage ratio (OR) and time correction factor.

[0084] For the same symptom or the same disease type, there may be multiple intensity characteristics, such as mild, moderate, severe, or others. Intensity correction factors can be calculated separately for each of the different intensity characteristics.

[0085] The vector module then constructs an intensity matrix based on disease type, intensity features, and intensity correction factors. For example, the intensity matrix may include the following table (5):

[0086]

[0087] In one possible implementation, the vector module constructs a joint matrix based on the probability matrix, the time matrix, and the intensity matrix, the joint matrix including the correspondence between the probability matrix, the time matrix, and the intensity matrix.

[0088] In one possible implementation, the vector module can fuse the probability matrix, time matrix, and intensity matrix using an indexed triple loop traversal algorithm and the element-wise Hadamard product operator.

[0089] First, the vector module iterates through all ternary combinations of symptom entities, temporal features, and intensity features in the probability matrix, time matrix, and intensity matrix using an index traversal algorithm, and then constructs a three-dimensional joint vector group based on the symptom entities, temporal features, and intensity features. The three-dimensional joint vector group can include... .in, Represents a single three-dimensional joint vector. It can represent different symptom entities. - This represents different disease types and their corresponding posterior probabilities. - represent different temporal features and temporal correction factors, - represent different intensity features and intensity correction factors. The three-dimensional joint vector may include symptom entities, disease types, posterior probabilities of diseases, temporal features, temporal correction factors, intensity features, intensity correction factors, and their corresponding relationships.

[0090] In a possible implementation, determining a query vector based on an input text includes:

[0091] the input module acquires text information converted from the input text; the input module uses a bidirectional long short-term memory network conditional random field model to perform medical entity annotation on the text information to obtain medical entities, where the medical entities include symptom entities, temporal entities, intensity entities, and modifier entities; the input module determines medical features based on the medical entities through a medical pre-trained model, where the medical features include symptom features, temporal features, and intensity features; the input module constructs a query vector based on the medical features.

[0092] When a patient inputs the symptom description "Sudden chest pain for 1 hour, severe and unbearable with cold sweat" through a hospital self-service terminal or a mobile APP, the text is first transmitted to the medical entity annotation and feature extraction module for preprocessing. The bidirectional long short-term memory network conditional random field model performs medical entity annotation on the text, and the output annotation result is as follows: Sudden B-TIM Onset I-TIM Chest B-SYM Pain I-SYM 1 B-TIM Hour I-TIM Intense B-INT Severe I-INT Unbearable I-TIM With B-MOD Cold I-MOD Sweat I-MOD. Wherein B represents Begin, which is the starting character of an entity, the first Chinese character of an entity segment is marked as B-XX, I represents Inside, which is the middle or end character of an entity, and subsequent Chinese characters of the same entity are marked as I-XX. TIM is a temporal entity, SYM is a symptom entity, INT is an intensity entity, and MOD is a modifier entity.

[0093] After the medical pre-trained model extracts the context representation, four-dimensional features are extracted through an entity-aware label aggregation strategy: the symptom feature v_sym is identified as "chest pain", the temporal feature v_tim is identified as <1 hour, sudden onset>, the intensity feature v_int is identified as level L3 (severe and unbearable), and the modifier feature v_mod is identified as "cold sweat". The input module may construct a query vector Q={s=chest pain, t=<1 hour, sudden onset>, i_init=L3} based on the symptom feature, temporal feature, and intensity feature, and send this query vector to the multi-dimensional symptom feature vector library module for querying.

[0094] The bidirectional long short-term memory network conditional random field model and the medical pre-trained model are existing technologies. Their specific structures and working principles are not considered as improvements to this solution and will not be elaborated here.

[0095] The input module can break down patient-input text information into multiple independent entities, facilitating rapid and accurate identification of medical features. It also effectively extracts disease-related content from the text, preventing information omissions and ensuring the accuracy of subsequent triage decisions. The constructed query vector is compatible with the joint matrix in the multidimensional symptom feature vector library, enabling rapid matching of relevant information within the joint matrix. This reduces computational consumption, improves the output efficiency of triage decisions, and, when applied to hospital self-service terminals, reduces the time patients spend on the device, thus increasing its utilization efficiency.

[0096] The target information is determined based on the query vector through a multidimensional symptom feature vector database. This target information may include at least one three-dimensional joint vector determined from the joint matrix that matches the query vector, for example... , , The input module can calculate the product of the posterior probability, time correction factor, and intensity correction factor in each three-dimensional joint vector, and calculate the ratio of each product to the sum of the products, and confirm this ratio as the three-dimensional joint probability.

[0097] In one possible implementation, determining the time-weighted factor and intensity-weighted factor based on the input text includes:

[0098] The weighting module can construct a dual-channel weighting layer, which includes a time channel and an intensity channel;

[0099] The weighting module determines time modifiers based on the input text through the time channel, calculates the syntactic dependency distance between the time modifiers and the symptom entities, and determines the time weighting factor when the syntactic dependency distance meets the first preset condition.

[0100] The weighting module determines the symptom intensity level based on the input text through intensity channel matching, and determines the intensity weighting factor when the symptom intensity level meets the second preset condition.

[0101] The time channel is used to identify time entities (TIMs) in text information, such as sudden, 1 hour, lasting three days, early morning, intermittent, etc. The time channel can perform dependency parsing on text information, generate a syntactic tree, and calculate the syntactic dependency distance between time entities and symptom entities. The syntactic dependency distance is used to describe the closeness of the syntactic relationship between time entities and symptom entities.

[0102] The first preset condition can include when the syntactic dependency distance is less than a distance threshold. When the syntactic dependency distance is less than the distance threshold, the temporal channel can determine that the syntactic dependency distance meets the requirement and determine the corresponding temporal weighting factor. The distance threshold is a preset value and can be set according to actual needs.

[0103] The weighting module can determine the time weighting factor in various ways. For example, it can construct a first data table based on syntactic dependency distances, time weighting factors, and their corresponding relationships in historical data. The weighting module then determines the time weighting factor by querying this first data table based on the syntactic dependency distance. Alternatively, the weighting module can calculate the product of the syntactic dependency distance and a first coefficient, using this product as the time weighting factor. The first coefficient is a preset value and can be set according to actual needs.

[0104] When multiple syntactic dependency distances satisfy the given conditions, the weighting module can determine multiple time weighting factors corresponding to these syntactic dependency distances. The weighting module can determine the required time weighting factor in various ways. For example, it can select the maximum value as the required time weighting factor. Alternatively, the weighting module can calculate the product of multiple time weighting factors and use this product as the required time weighting factor.

[0105] The intensity channel is used to identify and match intensity entities (INT) in text information, such as severe, slight, stinging, unbearable, and dull pain.

[0106] The weighting module can have a built-in intensity lexicon and build an intensity level table based on the intensity lexicon, intensity levels and corresponding intensity weighting factors. Each intensity level corresponds to a different intensity entity. The weighting module can determine the corresponding intensity based on the intensity entity in the text information by querying the intensity level table.

[0107] The second preset condition can include an intensity level greater than or equal to a preset intensity level. The preset intensity level can be a preset value, such as level two, level three, etc., and its specific level can be set according to actual needs. For example, when the intensity is mild, the intensity level is level one, and the corresponding intensity weighting factor is 1. This intensity weighting factor does not amplify the coupling weight when determining the coupling weight. When the intensity is stinging or dull pain, the intensity level is level two, and the corresponding intensity weighting factor is 1.1. This intensity weighting factor amplifies the coupling weight when determining the coupling weight. When the intensity is severe or unbearable, the intensity level is level three, and the corresponding intensity weighting factor is 1.5. This intensity weighting factor amplifies the coupling weight when determining the coupling weight.

[0108] The dual-channel weighted layer enables feature extraction from the patient's input text. By simultaneously and separately identifying information related to time and intensity features through dual channels, it is easier to accurately determine the time and intensity weighting factors in the subsequent process. This helps to shorten the feature extraction time and speed up the feature extraction process, thereby reducing the patient's waiting time and device usage time, and improving the patient's user experience.

[0109] In one possible implementation, the dual-channel weighting layer further includes a time-intensity coupled attention layer; the weighting module can calculate the coupled weighting weights based on the time weighting factor and the intensity weighting factor through the coupled attention layer via a coupled amplification effect function.

[0110] The input to the time-intensity coupled attention layer is a time-weighted factor and an intensity-weighted factor, and the output is the coupling weight. The time-intensity coupled attention layer is existing technology, and its specific structure and working principle are not considered improvements to this solution and will not be elaborated here.

[0111] The coupling amplification effect function is used to calculate the coupling weighting based on the time weighting factor and the intensity weighting factor. In one possible implementation, the coupling amplification effect function may include formula (6):

[0112]

[0113] in, Indicates the coupled weighted weights. This represents the time-weighted factor intensity-weighted factor. This is the magnification factor. This represents the maximum value in the three-dimensional joint probability.

[0114] The coupled attention layer, based on time-weighted and intensity-weighted factors, performs nonlinear calculations using a pre-defined coupling amplification effect function to obtain the coupled weighting weights of the nonlinear coupling between the time-weighted and intensity-weighted factors. When both the time-weighted and intensity-weighted factors represent high-risk characteristics, an additional risk amplification value is added, reflecting the clinical characteristic of a synergistic increase in disease risk when acute course and severe symptoms coexist. If only a single dimension exhibits a high-risk characteristic, only the product of the two weights is retained as the coupled weighting weight. This allows for flexible adjustment of triage decisions based on the temporal and intensity characteristics of the disease, expanding its applicability and facilitating the analysis of disease occurrence probabilities based on different temporal and intensity characteristics and their combinations. When determining the coupled weighting weights, the maximum value in the three-dimensional joint probability is introduced to avoid simultaneously amplifying disease types with low matching degrees, thereby reducing the probability of misjudgment.

[0115] The dual-channel weighted layer also includes a symptom co-occurrence map.

[0116] A symptom co-occurrence graph is a graph structure used to describe the relationships between different symptoms. A symptom co-occurrence graph can include multiple nodes and multiple edges.

[0117] Nodes are used to represent different symptom entities.

[0118] An edge connects two nodes and represents the correlation between two connected symptoms. The correlation can be represented by a specific numerical value; the higher the value, the stronger the correlation between the two symptoms, and the higher the probability that two symptoms will occur simultaneously when a patient has a certain disease. For example, when a patient experiences an acute myocardial infarction, they may simultaneously experience chest pain, cold sweats, and chest tightness. There is a high correlation between chest pain and cold sweats, chest pain and chest tightness, and cold sweats and chest tightness. The specific numerical value of the correlation is a preset value and can be determined in various ways, such as obtaining it from historical data or setting it based on experience.

[0119] The weighting module can determine the co-occurrence enhancement factor based on the symptom co-occurrence map using a graph attention network; the weighting module can determine the fusion weighting vector based on the coupling weights, co-occurrence enhancement factors, and three-dimensional joint probabilities. Graph attention networks are existing technology, and their specific content and training methods are not considered improvements to this solution and will not be elaborated here.

[0120] The fusion weighted vector can include .in, , , , These can represent the three-dimensional joint probabilities of different diseases. It can represent a time-weighted factor. It can represent an intensity weighting factor. It can represent coupled weighted weights.

[0121] In one possible implementation, the weighting module can identify correlation as a co-occurrence enhancement factor. The fusion weighted vector may include... .in, , , , These can represent the three-dimensional joint probabilities of different disease types. It can represent a time-weighted factor. It can represent an intensity weighting factor. It can represent coupled weighted weights. Co-occurrence enhancing factor.

[0122] When constructing the fusion weighted vector, the co-occurrence enhancement factor is considered, taking into account the possibility that different symptoms may appear simultaneously for the same disease and the degree of correlation between symptoms. When inferring diseases associated with symptoms based on symptoms, the degree of correlation between different symptoms can also be taken into account, which helps to improve the accuracy of disease type determination.

[0123] In one possible implementation, the multi-task decision layer includes a disease prediction task layer, a department recommendation task layer, an urgency classification task layer, and a risk scoring task layer.

[0124] The disease prediction task layer is used to ultimately predict the types of diseases a patient may have and their corresponding probabilities.

[0125] In one possible implementation, the decision module can determine the final probability of a disease based on a fused weighted vector through the disease prediction task layer. The decision module can calculate the product of the three-dimensional joint probability, time-weighted factor, intensity-weighted factor, coupling weight, and co-occurrence enhancement factor corresponding to different disease types, calculate the sum of these products, and calculate the proportion of the product corresponding to different disease types to the total sum of products. This proportion is then determined as the final probability of the disease corresponding to each disease type.

[0126] In one possible implementation, the decision-making module can determine the recommended department based on the final probability of the disease through a department recommendation task layer. The decision-making module can construct a second data table based on disease types, departments, and their corresponding relationships from historical and / or test data. The decision-making module can select disease types with a final probability greater than a probability threshold and determine the corresponding department based on that disease type by querying the second data table. The probability threshold is a preset value, such as 50%, 70%, or other values, and its specific value can be set according to actual needs.

[0127] In one possible implementation, the decision module can determine the urgency level through an urgency-based task layer, based on the disease's final probability, coupling weighting, co-occurrence enhancement factor, and time weighting factor.

[0128] In one possible implementation, the decision-making module can construct a third data table based on urgency scores, urgency levels, and their corresponding relationships in historical and / or test data, where each urgency level corresponds to a range of urgency score values. The decision-making module can calculate the product of the final disease probability, coupling weighting, co-occurrence enhancement factor, and time weighting factor, confirm this product as the urgency score, and determine the urgency level by querying the third data table based on this product.

[0129] By calculating the final probability of a disease, the system outputs the most likely disease to the patient. After calculating the final probability, it can also determine the appropriate department for the patient and assess the urgency of the patient's need for medical attention, providing good guidance for the patient's medical treatment and improving the efficiency and user experience of the medical treatment.

[0130] In one possible implementation, the decision module can also calculate the output distribution entropy value, and trigger the generation of follow-up questions or transfer to the human channel when the entropy value meets the third preset condition.

[0131] The output distribution entropy value is used to describe the dispersion of the final probabilities of various diseases. Higher dispersion indicates a higher probability of one disease relative to others, with lower probabilities for other diseases and a larger numerical difference, leading to more accurate predictions of the final disease probabilities. Lower dispersion indicates smaller differences in the final probabilities of multiple diseases, making it difficult to accurately determine the patient's specific disease.

[0132] The decision module can calculate the output distribution entropy value based on the fifth algorithm. The fifth algorithm can include formula (5):

[0133]

[0134] in, This indicates the output distribution entropy value. Indicates the final probability of the disease. Indicates the calculation of multiple The average value.

[0135] The third preset condition may include a distribution entropy value greater than an entropy threshold. The entropy threshold can be a specific value, such as 1, 1.2, 1.3 or other values, and the specific value can be set according to actual needs.

[0136] Follow-up questions can be pre-set, such as "Please provide details about any other symptoms" or "Please provide details about the duration of your symptoms." Follow-up questions can be customized to meet specific needs.

[0137] In one possible implementation, let's take the case of patient Wang going to the hospital for treatment due to "sudden chest pain for 1 hour, which was severe and unbearable and accompanied by cold sweats".

[0138] When a patient inputs a symptom description, "sudden chest pain for 1 hour, severe and unbearable accompanied by cold sweats," through an input module (e.g., a hospital self-service terminal or mobile app), the input module first transmits the input text to the medical entity annotation and feature extraction module for preprocessing. A bidirectional long short-term memory network conditional random field model performs medical entity annotation on the text, outputting the annotation results as follows: Sudden B-TIM I-TIM Chest B-SYM Pain I-SYM 1B-TIM Small I-TIM Hour I-TIM Severe B-INT Intense I-INT Difficult I-TIM Unbearable I-TIM Accompanied B-MOD Sweating I-MOD Cold I-MOD I-MOD. After extracting contextual representations, the medical pre-trained model extracts four-dimensional features using an entity-aware tagging aggregation strategy: symptom feature v_sym is identified as the entity "chest pain"; time feature v_tim is extracted as the time tuple <1 hour, sudden onset>; intensity feature v_int is identified as level L3 (severe and unbearable); location feature v_loc is extracted as the default location; and modifier feature v_mod is identified as "cold sweats". The system simultaneously constructs a query vector Q={s=chest pain, t=<1 hour, sudden onset>, i_init=L3} and sends this query vector to the multi-dimensional symptom feature vector library module for a three-dimensional joint probabilistic query.

[0139] After receiving the query vector, the multidimensional symptom feature vector library module first queries the disease type component to obtain the conditional probability of "chest pain" against various diseases. , , Equal basic probability values.

[0140] The vector module identifies the time feature as "sudden" with a time value of 1 hour, and retrieves the result by querying the time matrix. , , Equal time correction factor.

[0141] The vector module identifies the intensity feature as L3, and the intensity matrix is ​​queried to obtain... , , Equal intensity correction factor. The three-dimensional joint probability calculation component calculates the three-dimensional joint probability based on the above data:

[0142]

[0143]

[0144]

[0145] After normalization, the three-dimensional combined probability of acute myocardial infarction was 0.39, the three-dimensional combined probability of aortic dissection was 0.33, and the three-dimensional combined probability of angina pectoris was 0.28. This three-dimensional combined probability was transmitted as the output to the dual-channel weighted layer.

[0146] After receiving the three-dimensional joint probability, the dual-channel weighted layer first performs time channel calculation.

[0147] The time channel detected "sudden" as a time modifier, with a time feature t = sudden and a time value of 1 hour. The time channel calculates the dependency distance between "sudden" and "chest pain" based on dependency parsing to derive the basic time weighting factor.

[0148] The time channel executes the following logic: it determines whether the time feature t is a sudden event and whether the time value is less than a preset threshold (1 hour). If the condition is met, the time weighting factor is activated. Time channel weight It was magnified to 1.8 times.

[0149] The intensity channel matches the intensity vocabulary and identifies "severe and unbearable" as level L3, while also detecting the modifier "cold sweats." The intensity channel follows this logic: it determines whether intensity level i reaches L3 and whether the symptoms belong to the vascular emergency-related symptom set. Since chest pain belongs to the vascular emergency-related symptom set, the intensity weighting factor is activated. Intensity channel weights It was magnified to 1.5 times.

[0150] The temporal intensity coupled attention layer calculates the combined weighted weights, first obtaining the maximum probability from the three-dimensional joint probability. (Acute myocardial infarction), and then the coupling weights are calculated through the coupling amplification effect function. This coupling amplification effect function calculates the coupling weights by fusing the time channel weights, intensity channel weights, and disease probability through the time-intensity coupling attention layer. =1.248, realizing nonlinear interaction between time and intensity characteristics, triggering a coupling amplification effect to enhance the risk signal when multiple dimensions all point to high risk.

[0151] In this embodiment, since both the time and intensity channels are activated and the disease probabilities point to acute myocardial infarction and aortic dissection, the coupling effect significantly amplifies the weight of high-risk diseases. The symptom co-occurrence graph attention unit constructs a symptom co-occurrence graph, with edge sets including relationships such as chest pain-cold sweats (related to myocardial infarction) and chest pain-sudden onset (related to vascular emergencies). The graph attention network aggregates adjacent symptom information to calculate the co-occurrence enhancement factor. The weighting module integrates attention weights, time channel weights, intensity channel weights, co-occurrence enhancement factors, and disease probability outputs into a weighted vector.

[0152] After receiving the fused weighted vector, the decision module executes the following triage decision process.

[0153] The disease prediction task layer determines the final probability of the disease based on the fusion weighted vector: acute myocardial infarction (probability 0.65), aortic dissection (probability 0.22), and angina pectoris (probability 0.13).

[0154] Based on the final probability of the disease, the department recommendation task layer determined that the department corresponding to acute myocardial infarction was cardiology, with a score of 0.65. The urgency grading task unit calculated an urgency score of 1.89 based on the final probability of the disease (0.65), the coupled attention weight (1.248), the co-occurrence enhancement factor (1.3), and the time weighting factor (1.8), and determined the urgency level to be Level I (immediate response).

[0155] In one possible implementation, the output of the decision module can be sent to a terminal used by a doctor or nurse, such as a mobile phone, computer, or tablet, so that the doctor or nurse can respond to the output of the decision module and guide the patient to seek medical treatment manually.

[0156] On the other hand, embodiments of the present invention also provide a medical triage semantic understanding system, including a control module for implementing a medical triage semantic understanding method. The control module includes: a vector module for determining and constructing a multi-dimensional symptom feature vector library based on a medical symptom knowledge graph, the multi-dimensional symptom feature vector library including a probability matrix, a time matrix, an intensity matrix, and a joint matrix; an input module for determining a query vector based on input text, and determining a three-dimensional joint probability based on the query vector through the multi-dimensional symptom feature vector library; a weighting module for determining a time weighting factor and an intensity weighting factor based on the input text, determining a coupling weighting weight based on the time weighting factor and the intensity weighting factor, determining a coupling weighting weight based on the coupling weighting weight and the three-dimensional joint probability, and determining a fusion weighting vector based on the coupling weighting weight, the time weighting factor, the intensity weighting factor, and the three-dimensional joint probability; and a decision module for determining a triage decision through a multi-task decision layer based on the fusion weighting vector.

[0157] Furthermore, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a medical triage semantic understanding method.

[0158] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0159] 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. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0160] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0161] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A semantic understanding method for medical triage that integrates time and intensity features, characterized in that, The method includes: A multidimensional symptom feature vector library was constructed based on a medical symptom knowledge graph. Input text is generated based on the patient's voice; a query vector is determined based on the input text; target information is determined based on the query vector through the multidimensional symptom feature vector library; and a three-dimensional joint probability is determined based on the target information. Based on the input text, a time weighting factor and an intensity weighting factor are determined; based on the time weighting factor and the intensity weighting factor, a coupling weighting weight is determined; and based on the time weighting factor, the intensity weighting factor, the coupling weighting weight, and the three-dimensional joint probability, a fusion weighting vector is determined. Based on the fusion weighted vector, triage decisions are determined through a multi-task decision layer.

2. The method according to claim 1, characterized in that, The construction of a multidimensional symptom feature vector library based on a medical symptom knowledge graph includes: Based on the medical symptom knowledge graph, symptom entities, disease types, and the association information between the symptom entities and the disease types are determined, and the association information includes time features and intensity features. Based on the symptom entity and the disease type, a probability matrix is ​​constructed, the probability matrix including the symptom entity, the disease type, and the probability of the symptom entity occurring; Based on the disease type and the time characteristics, a time matrix is ​​constructed, which includes the disease type, the time characteristics, and a time correction factor; Based on the disease type and the intensity feature, an intensity matrix is ​​constructed, the intensity matrix including the disease type, the intensity feature and the intensity correction factor; Based on the probability matrix, the time matrix, and the intensity matrix, the joint matrix is ​​constructed, and the joint matrix includes the correspondence between the probability matrix, the time matrix, and the intensity matrix.

3. The method according to claim 1, characterized in that, The process of determining the query vector based on the input text includes: Obtain the text information after the input text has been converted; The text information is labeled with medical entities using a bidirectional long short-term memory network conditional random field model to obtain medical entities, which include symptom entities, time entities, intensity entities, and modifier entities. Medical features are determined based on the medical entities using a medical pre-trained model. These medical features include symptom features, time features, intensity features, and modifier entities. A query vector is constructed based on the aforementioned medical features.

4. The method according to claim 1, characterized in that, The determination of the time-weighted factor and intensity-weighted factor based on the input text includes: A dual-channel weighted layer is constructed, which includes a time channel and an intensity channel; The time modifiers are determined based on the input text through the time channel, and the syntactic dependency distance between the time modifiers and the symptom entities is calculated. When the syntactic dependency distance meets the first preset condition, the time weighting factor is determined. The intensity level of a symptom is determined based on the input text by intensity channel matching, and an intensity weighting factor is determined when the intensity level of the symptom meets a second preset condition.

5. The method according to claim 4, characterized in that, The dual-channel weighted layer also includes a time-intensity coupled attention layer; Through the coupled attention layer, the coupled weighting weight is calculated based on the time weighting factor and the intensity weighting factor using the coupled amplification effect function.

6. The method according to claim 5, characterized in that, The dual-channel weighted layer also includes a symptom co-occurrence map; Based on the symptom co-occurrence map, co-occurrence enhancement factors are identified using a graph attention network; The fusion weighting vector is determined based on the coupling weighting, the co-occurrence enhancement factor, and the three-dimensional joint probability.

7. The method according to claim 6, characterized in that, The multi-task decision-making layer includes a disease prediction task layer, a department recommendation task layer, an urgency classification task layer, and a risk scoring task layer. The final probability of the disease is determined through the disease prediction task layer based on the fusion weighted vector. Based on the final probability of the disease, the recommended department is determined through the department recommendation task layer. The urgency level is determined through the urgency classification task layer based on the final probability of the disease, the coupling weighting weight, the co-occurrence enhancement factor, and the time weighting factor.

8. The method according to claim 7, characterized in that, The step of determining triage decisions through a multi-task decision layer based on the fusion weighted vector also includes: Calculate the output distribution entropy value. When the entropy value meets the third preset condition, trigger the generation of follow-up questioning scripts or transfer to the human channel. Output key feature attributions, decision-making basis, and follow-up questions.

Images