Pre-hospital emergency aid prediction method and system based on the fusion of large model and knowledge graph

Abstract

The present invention relates to a pre-hospital emergency aid prediction method that integrates a large model and a knowledge graph, comprising the following steps: performing real-time speech recognition on incoming telephone voices and converting them into text description information of the condition; querying a patient knowledge graph stored in a cloud data center based on the text description information of the condition; performing characterization learning on the text description information of the condition, performing vector representation on the patient knowledge graph, fusing the respectively obtained alarm description feature vectors and the patient's historical disease feature vectors to generate an emergency patient feature vector; performing knowledge graph embedding vector learning on a pre-constructed disease knowledge graph to generate a disease feature vector; and predicting the condition symptoms of emergency patients based on the feature vectors and disease feature vectors of the emergency patients. The present invention learns the feature vector descriptions of different diseases, performs an inner product operation on the patient feature vector and the disease feature vector, and generates a patient condition symptom prediction function, thereby achieving accurate prediction of the patient's condition.

Description

Technical Field

[0001] The present invention belongs to the field of medical health management, specifically a pre-hospital emergency auxiliary prediction method and system that integrates a large model and a knowledge graph. Background Art

[0002] Pre-hospital emergency care, also known as out-of-hospital emergency care, refers to the medical care a patient receives from the onset of illness until hospital admission. As a crucial component of emergency medicine, pre-hospital emergency care significantly reduces the mortality rate of patients with sudden injuries. The level of pre-hospital emergency care reflects a country's comprehensive capabilities in organizational management, medical care, and public welfare.

[0003] In this context, due to the complexity, hidden nature, and rapid progression of the conditions of elderly emergency patients, as well as the lack of professionalism of emergency dispatch and temporary medical personnel, it can be very difficult to identify sudden illnesses in elderly patients and provide timely symptomatic treatment. In addition, due to the uncontrollable progression of sudden illnesses, it is difficult to have more time to conduct an in-depth symptom analysis of the patient's condition at the scene and during emergency transportation, and to provide feasible priority treatment resources and assistance. This can easily lead to a series of problems such as emergency delays and inefficient emergency resource scheduling. At present, the existing pre-hospital emergency patient confirmation mainly relies on the patient's alarm description and the subjective experience of rescue personnel. This method is usually subjective and easily leads to risks such as misjudgment of symptoms and delayed treatment on the scene.

[0004] Therefore, there is an urgent need for a rapid emergency auxiliary prediction method that can adapt to unknown situations and meet the corresponding conditions. Summary of the Invention

[0005] In response to the problems existing in the existing methods, the present invention proposes a pre-hospital emergency auxiliary diagnosis method, device and system that integrates a cloud-based collaborative large model and knowledge graph.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a pre-hospital emergency aid prediction method integrating a large model and a knowledge graph, comprising the following steps:

[0007] Perform real-time speech recognition on incoming phone calls and convert them into text descriptions of the patient's condition; query the patient's feature knowledge graph stored in the cloud data center based on the text descriptions of the condition;

[0008] Perform characterization learning on the text description information of the condition to obtain an alarm description feature vector, perform vector representation on the patient feature knowledge graph to obtain a patient history disease feature vector, fuse the alarm description feature vector and the patient history disease feature vector to generate an emergency patient feature vector;

[0009] Perform knowledge graph embedding vector learning on the pre-built disease knowledge graph to generate disease feature vectors;

[0010] The symptoms of emergency patients are predicted based on their feature vectors and condition feature vectors.

[0011] The method includes performing real-time voice recognition on the incoming telephone voice and converting it into text description information of the condition; and querying the patient knowledge graph stored in the cloud data center based on the text description information of the condition, including the following steps:

[0012] First, the incoming telephone voice is recognized in real time and converted into text information. After the text information is cleaned, a text description of the condition is generated. The text description of the condition includes at least some of the following: the patient's name, ID card, description of the emergency situation, the patient's geographic location, the time of the alarm, and a contact person.

[0013] Secondly, the patient characteristic knowledge graph information stored in the cloud data center is queried based on the text description information of the disease condition; the patient characteristic knowledge graph is constructed based on the patient's historical data, with patients, historical diseases, disease symptoms as nodes, and the relationship between patients and their diseases, and between diseases and their disease symptoms as the relationship between nodes.

[0014] The text description information of the condition is characterized and learned to obtain an alarm description feature vector, the patient feature knowledge graph is vector-characterized to obtain a patient history disease feature vector, and the alarm description feature vector and the patient history disease feature vector are fused to generate an emergency patient feature vector, including the following steps:

[0015] First, the large language model is used to characterize the text description of the condition after the telephone voice is converted to generate the alarm description feature vector

[0016] Secondly, the patient feature knowledge graph is vectorized through the knowledge graph embedding vector learning method to generate the patient's historical disease feature vector

[0017] Then, the alarm description feature vector and the patient's historical disease feature vector Perform fusion to generate the fused emergency patient feature vector E u , the fusion uses the attention mechanism to aggregate the vectors of nodes and their neighboring nodes on the patient knowledge graph, which is achieved by the following formula:

[0018]

[0019] Among them, W1 and W2 are trainable weight matrices, and LeakyReLU is the activation function.

[0020] The method of performing knowledge graph embedding vector learning on the pre-built disease knowledge graph to generate a disease feature vector includes the following steps:

[0021] 1) Pre-construct a disease knowledge graph, using the triplet <disease name or disease symptom name, relationship, condition symptom> to describe the disease, where the disease name or disease symptom name and condition symptom are nodes, and the relationship is the relationship between the nodes;

[0022] 2) Perform knowledge graph embedding vector learning on the disease knowledge graph and vectorize each node on the disease knowledge graph:

[0023] Node v represents the name of the disease or the name of the disease symptom, and the embedding vector of v is represented as the disease feature vector E v , for E v The representation ability is further enhanced by aggregating the embedding vectors of its neighboring nodes:

[0024]

[0025] Among them, N(v) represents the neighbor set of node v on the disease knowledge graph, g(v,v′) represents the fusion weight function, and the embedding vector of node v′ is represented by E v′ ; Use the attention mechanism to implement g(v,v′). The implementation process is described as follows:

[0026]

[0027] Among them, |N(v)| and |N(v′)| represent the number of neighbor nodes of node v and v′ respectively. The matrix transpose represents the eigenvector of patient u, and d is the constant factor used in scaling the dot product of the vectors.

[0028] The method of predicting the condition of an emergency patient based on the emergency patient's feature vector and condition feature vector comprises the following steps:

[0029]

[0030] f(u,v) is a prediction function, which is used to judge the correlation between patient u and disease symptom v according to the size of f(u,v); E u is the feature vector of emergency patients, E v is the disease characteristic vector;

[0031] The disease condition is predicted based on the value of the prediction function f(u,v).

[0032] The pre-hospital emergency aid prediction system, which integrates large models and knowledge graphs, includes:

[0033] The speech recognition module is used to perform real-time speech recognition on incoming telephone voices and convert them into text descriptions of the patient's condition; and to query the patient knowledge graph stored in the cloud data center based on the text descriptions of the condition;

[0034] A disease characterization learning module is used to characterize and learn the text description information of the disease to obtain an alarm description feature vector, perform vector representation on the patient knowledge graph to obtain a patient history disease feature vector, and fuse the alarm description feature vector and the patient history disease feature vector to generate an emergency patient feature vector;

[0035] The disease feature learning module is used to learn the knowledge graph embedding vector of the pre-built disease knowledge graph and generate the disease feature vector;

[0036] The condition prediction module is used to predict the condition of emergency patients based on the emergency patients' feature vectors and condition feature vectors.

[0037] A pre-hospital emergency resource intelligent scheduling device includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement the pre-hospital emergency auxiliary prediction method that integrates the large model and knowledge graph when executing the computer program.

[0038] The present invention has the following beneficial effects and advantages:

[0039] Existing methods for identifying and confirming the condition of pre-hospital emergency patients primarily rely on the patient's alarm description and the subjective experience of rescuers. Due to the complexity, hidden nature, and rapid progression of the conditions of elderly emergency patients, as well as the lack of professionalism of emergency dispatchers and temporary treatment personnel, problems such as inaccurate condition identification and delayed treatment may occur. To address this issue, the present invention has the following beneficial effects:

[0040] (1) This invention proposes to enhance learning of elderly patients’ alarm information with the help of a large language model to generate a more accurate description of the emergency patients’ condition characteristics;

[0041] (2) At the same time, the present invention proposes to construct a patient feature knowledge graph with the help of the patient's historical data (medical data, rehabilitation data, monitoring data, health records, etc.) stored on the cloud side, and generate a feature description of the patient's historical condition with the help of the embedding learning and attention mechanism of the knowledge graph, providing a more in-depth feature description for the patient's condition identification;

[0042] (3) Fusing the feature vector generated by the patient's emergency alarm information with the patient's historical disease feature knowledge graph vector to generate a more comprehensive and rich patient feature vector;

[0043] (4) In addition, the present invention also proposes to construct a disease knowledge graph, learn the feature vector descriptions of different diseases, perform inner product operations on patient feature vectors and disease feature vectors, and generate patient disease or symptom prediction functions, thereby realizing the accurate identification and prediction process of patient disease symptoms. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is the overall logical framework diagram for the implementation of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0046] Based on the deep fusion of a large language model and a knowledge graph, the present invention performs embedding vector representation learning on the real-time voice text of the patient's alarm and the patient's historical data stored in the cloud data center, and fuses the patient's language text feature vector generated by the large language model and the patient's historical condition feature vector generated by the knowledge graph embedding learning to form a cloud-based fused patient feature embedding vector.

[0047] The present invention constructs a disease feature knowledge graph, uses the attention mechanism to learn the embedding vectors of different diseases on the knowledge graph, and aggregates the embedding vectors of node neighbors based on the attention mechanism to generate disease feature vectors. By performing an inner product operation on the fused patient feature vector and the aggregated disease feature vector, a disease or symptom prediction function for emergency patients is generated.

[0048] like Figure 1 The figure shows the overall logical framework of the present invention, the core idea of ​​which is:

[0049] ① Alarm text information extraction and patient knowledge graph construction: Figure 1 As shown in the figure, first, when an elderly patient (assuming u represents the patient) calls the 120 emergency command and dispatch center, the phone voice is simultaneously connected to the backend for real-time speech recognition and text conversion, forming an online text description of the condition based on the phone call. The basic content collected by the phone call includes: patient name, ID card, description of the emergency situation, patient location, alarm time, contact person, etc. Secondly, based on the basic information provided by the elderly patient (such as name, ID card, etc.), the patient knowledge graph information stored in the Medical Prevention and Health Cloud Data Center is queried. We assume that the knowledge graph has been constructed through offline calculation.

[0050] Existing methods make it difficult to accurately identify the category and severity of an elderly patient's condition based on the voice description of the alarm caller. Unlike existing methods: (1) The present invention converts speech into text information through speech recognition, and then cleans the text information to remove unnecessary characters, formats and noise for subsequent processing; (2) Based on the patient's historical data (medical data, rehabilitation data, monitoring data, health records, etc.) stored in the cloud data center, a patient feature knowledge graph is constructed, which can quickly analyze the patient's historical condition characteristics and more accurately infer and judge the patient's possible condition category and severity. The patient feature knowledge graph is constructed based on the patient's historical data and is used to represent the patient's medical data, rehabilitation data, monitoring data, and health records; the patient feature knowledge graph is composed of patients, historical diseases (disease symptoms or classifications, etc.), disease characteristics (condition symptoms) as nodes, and the relationship between patients and their diseases (disease symptoms or classifications, etc.), and the relationship between diseases (symptoms or classifications, etc.) and their disease characteristics (condition symptoms) as the relationship between nodes.

[0051] ②Large language model and knowledge graph cloud-end fusion patient condition representation learning: Figure 1 As shown in Figure 1, the cloud-side data is mainly patient knowledge graph information, and the terminal-side data refers to the text information converted from the current alarm call. The innovation of this invention is: (1) First, the text data converted from the telephone voice is represented and learned by the large language model (LLM) to generate the alarm description feature vector Here, we propose to enhance the vector representation ability of the voice text of the telephone alarm with the help of a large language model to generate a richer, more complete and scientific description of the disease characteristics; (2) Secondly, with the help of the knowledge graph embedding vector learning method (here we use the classic TransR method) to vectorize the patient knowledge graph and generate the patient's historical disease feature vector (3) Then, the patient's alarm description feature vector generated based on the large language model is and patient history disease feature vectors generated based on knowledge graph embedding learning Perform fusion and generate the fused embedding vector, assuming it is E u , here we define the fusion method as:

[0052]

[0053] Where W1 and W2 are trainable weight matrices, and LeakyReLU is the activation function. The calculation process of is consistent with the implementation principle in formula (2), that is, the attention mechanism is used to aggregate the vectors of nodes and their neighboring nodes on the knowledge graph.

[0054] The innovation proposed by the present invention here is that the method of generating embedded vectors by fusing cloud data with a large language model and a knowledge graph can enhance the ability to learn the representation of the patient's condition, improve the accuracy of the diagnosis and analysis of the condition of emergency patients under data sparsity and cold start conditions, and provide possible reference information for pre-hospital and in-hospital emergency treatment.

[0055] ③ Disease feature vector representation learning: First, assume that the feature knowledge graphs of various diseases have been built in the cloud data center, where the knowledge graph triples are represented using typical triples.<name,include,symptom> In this embodiment, the triples <disease name or symptom name, relationship, symptom> are used for description; for example: <arrhythmia, characteristic manifestation, lack of energy>, <neurasthenia, characteristic manifestation, insomnia and forgetfulness>. Here, the knowledge graph embedding vector learning is performed on the disease knowledge graph (also using the TransR method), and each node on the knowledge graph is vectorized. Assuming that the disease is represented by v, the embedding vector representation of v is assumed to be E v . v The representation ability is further enhanced by aggregating the embedding vectors of its neighboring nodes:

[0056]

[0057] Where N(v) represents the neighbor set of node v on the disease knowledge graph, and g(v,v′) represents the fusion weight function. Here, the attention mechanism is used to implement g(v,v′). The implementation process is described as follows:

[0058]

[0059] Among them, |N(v)| and |N(v′)| represent the number of neighbor nodes of node v and v′ respectively. The matrix transpose of the eigenvector of patient u is represented. d is the constant factor in the dot product of the scaled vector. The eigenvector of disease v can be calculated by formulas (2) and (3). Assuming that E v express.

[0060] ④ Emergency patient condition prediction: Through ② and ③, the feature vector E of the emergency patient is generated respectively u and disease feature vector E v Next, the condition of the emergency patient can be predicted based on the generated feature vector. Assuming that the prediction function is represented by f(u,v), the calculation process of the prediction function is given as follows:

[0061]

[0062] f(u,v) is the prediction function, which means judging the correlation between patient u and disease (or disease symptoms) v according to the size of f(u,v); E u is the feature vector of emergency patients, E v is the disease feature vector;

[0063] The disease symptoms are predicted based on the value of the prediction function f(u,v). For example, when v1 = "Disease 1 or Disease Symptom 1", v2 = "Disease 2 or Disease Symptom 2", assuming that f(u,v1) = 0.4 and f(u,v2) = 0.5 are calculated according to the above formula, the probability of patient u suffering from disease or disease symptom 2 is higher.

[0064] In summary, based on Figure 1 Steps ①, ②, ③, and ④ shown here complete the overall implementation process description of the present invention's solution. The present invention integrates and analyzes the historical medical information and current alarm description information of the current emergency patient, and leverages the representation learning capabilities of the large language model and the representation learning and reasoning capabilities of the knowledge graph to intelligently predict and analyze the patient's possible diseases or symptoms. This will provide necessary and timely auxiliary decision support for pre-hospital emergency care and in-hospital comprehensive treatment, shorten the time required to assess the condition of emergency patients, and improve the timeliness of pre-hospital and in-hospital treatment for emergency patients.

Claims

1. A pre-hospital emergency aid prediction method based on the fusion of a large model and a knowledge graph, characterized by: The following steps are involved: Perform real-time speech recognition on incoming phone calls and convert them into text descriptions of the patient's condition; query the patient's feature knowledge graph stored in the cloud data center based on the text descriptions of the condition; Perform characterization learning on the text description information of the condition to obtain an alarm description feature vector, perform vector representation on the patient feature knowledge graph to obtain a patient history disease feature vector, fuse the alarm description feature vector and the patient history disease feature vector to generate an emergency patient feature vector; Perform knowledge graph embedding vector learning on the pre-built disease knowledge graph to generate disease feature vectors; Predicting the symptoms of emergency patients based on their feature vectors and condition feature vectors; The method includes performing real-time voice recognition on the incoming telephone voice and converting it into text description information of the condition; and querying the patient knowledge graph stored in the cloud data center based on the text description information of the condition, including the following steps: First, the incoming telephone voice is recognized in real time and converted into text information. After the text information is cleaned, a text description of the condition is generated. The text description of the condition includes at least some of the following: the patient's name, ID card, description of the emergency situation, the patient's geographic location, the time of the alarm, and a contact person. Secondly, the patient characteristic knowledge graph information stored in the cloud data center is queried based on the text description information of the disease condition; the patient characteristic knowledge graph is constructed based on the patient's historical data, with patients, historical diseases, disease symptoms as nodes, and the relationship between patients and their diseases, and between diseases and their disease symptoms as the relationship between nodes; The text description information of the condition is characterized and learned to obtain an alarm description feature vector, the patient feature knowledge graph is vector-characterized to obtain a patient history disease feature vector, and the alarm description feature vector and the patient history disease feature vector are fused to generate an emergency patient feature vector, including the following steps: First, the large language model is used to characterize the text description of the condition after the telephone voice is converted to generate the alarm description feature vector Secondly, the patient feature knowledge graph is vectorized through the knowledge graph embedding vector learning method to generate the patient's historical disease feature vector Then, the alarm description feature vector and the patient's historical disease feature vector Perform fusion to generate the fused emergency patient feature vector E u , the fusion uses the attention mechanism to aggregate the vectors of nodes and their neighboring nodes on the patient knowledge graph, which is achieved by the following formula: Among them, W1 and W2 are trainable weight matrices, and LeakyReLU is the activation function; The method of performing knowledge graph embedding vector learning on the pre-built disease knowledge graph to generate a disease feature vector includes the following steps: 1) Pre-construct a disease knowledge graph, using the triplet <disease name or disease symptom name, relationship, condition symptom> to describe the disease, where the disease name or disease symptom name and condition symptom are nodes, and the relationship is the relationship between the nodes; 2) Perform knowledge graph embedding vector learning on the disease knowledge graph and vectorize each node on the disease knowledge graph: Node v represents the name of the disease or the name of the disease symptom, and the embedding vector of v is represented as the disease feature vector E v , for E v The representation ability is further enhanced by aggregating the embedding vectors of its neighboring nodes: Among them, N(v) represents the neighbor set of node v on the disease knowledge graph, g(v,v′) represents the fusion weight function, and the embedding vector of node v′ is represented by E v′ ; Use the attention mechanism to implement g(v,v′). The implementation process is described as follows: Among them, N(v) and N(v′) represent the number of neighbor nodes of node v and v′ respectively. represents the matrix transpose of the eigenvector of patient u, and d is the constant factor in the dot product of the scaled vectors; The method of predicting the condition of an emergency patient based on the emergency patient's feature vector and condition feature vector comprises the following steps: f(u,v) is a prediction function, which is used to judge the correlation between patient u and disease or disease symptom v according to the size of f(u,v); E u is the feature vector of emergency patients, E v is the disease characteristic vector; The disease condition is predicted based on the value of the prediction function f(u,v).

2. A pre-hospital emergency aid prediction system based on the pre-hospital emergency aid prediction method based on the fusion of the large model and knowledge graph according to claim 1, characterized in that: include: The speech recognition module is used to perform real-time speech recognition on incoming telephone voices and convert them into text descriptions of the patient's condition; and to query the patient knowledge graph stored in the cloud data center based on the text descriptions of the condition; A disease characterization learning module is used to characterize and learn the text description information of the disease to obtain an alarm description feature vector, perform vector representation on the patient knowledge graph to obtain a patient history disease feature vector, and fuse the alarm description feature vector and the patient history disease feature vector to generate an emergency patient feature vector; The disease feature learning module is used to learn the knowledge graph embedding vector of the pre-built disease knowledge graph and generate the disease feature vector; The condition prediction module is used to predict the condition of emergency patients based on the emergency patients' feature vectors and condition feature vectors.

3. A pre-hospital emergency resource intelligent scheduling device, comprising a memory and a processor; the memory is used to store a computer program; the processor is used to implement the pre-hospital emergency auxiliary prediction method that integrates the large model and knowledge graph as described in claim 1 when executing the computer program.

Images