A medical data index interpretation method and device, a storage medium and equipment

Abstract

The application discloses a medical data index interpretation method and device, a storage medium and equipment. The method comprises the following steps: firstly, medical concepts and entity information of target problem text to be interpreted input by a target user are extracted; then, based on the medical concepts and entity information in the target problem text, a preset rule engine and a weak classifier are used to construct a candidate subgraph corresponding to the target problem text; then, a template text for prompting interpretation is constructed according to the candidate subgraph; and finally, an index interpretation result corresponding to the target problem text can be determined according to the template text for prompting interpretation. In this way, the dependency relationship between multiple groups of indexes is explicitly established by firstly extracting the medical concepts and entity information in the target problem text and then constructing the candidate subgraph, and the auxiliary reasoning result of the preset rule engine and the weak classifier is fused for medical data index interpretation, so that the index interpretation difficulty of the large model can be greatly reduced, and the interpretation accuracy of the medical data index is improved.

Description

Technical Field

[0001] This application relates to the field of medical technology, and in particular to a method, apparatus, storage medium and device for interpreting medical data indicators. Background Technology

[0002] With the increasing prevalence of fast-paced lifestyles, people are paying more and more attention to their health. In the medical field, a core application is the interpretation and reasoning of medical data indicators. Traditional methods use a combination of rule engines and specific classification models to interpret indicators for specific intent scenarios in vertically segmented fields. However, it is difficult to combine common sense to handle user inquiries in general scenarios, and the scalability of the scenario is also greatly limited.

[0003] Currently, the dominant method for interpreting medical data indicators employs generative general-purpose models, with the Chat Generative Pre-trained Transformer (ChatGPT) being a typical example. Its basic principle is to rely on pre-trained corpora to integrate subject-specific knowledge, unifying different Natural Language Processing (NLP) tasks into a GPT task, and using an autoregressive paradigm to achieve end-to-end generation of interpretation results. However, using this end-to-end autoregressive generation method can easily generate text that appears grammatically correct but has issues with knowledge accuracy, leading to inaccurate interpretation results for medical data indicators. Summary of the Invention

[0004] The main objective of this application is to provide a method, apparatus, storage medium, and device for interpreting medical data indicators, which can effectively improve the accuracy of interpreting medical data indicators and thus enhance the user experience.

[0005] This application provides a method for interpreting medical data indicators, including:

[0006] Obtain the target question text input by the target user; and extract the medical concepts and entity information from the target question text;

[0007] Based on the medical concepts and entity information in the target question text, a candidate subgraph corresponding to the target question text is constructed using a preset rule engine and a weak classifier.

[0008] Based on the candidate subgraphs, construct template text for the prompt interpretation;

[0009] Based on the template text interpreted by the prompts, determine the interpretation result of the indicator corresponding to the target question text.

[0010] In one possible implementation, the step of constructing a candidate subgraph corresponding to the target question text based on medical concept and entity information in the target question text, using a preset rule engine and a weak classifier, includes:

[0011] Using the medical concepts and entity information in the target question text, an initial candidate subgraph corresponding to the target question text is constructed;

[0012] The initial candidate subgraph is subjected to entity disambiguation processing and numerical normalization processing to obtain the preprocessed candidate subgraph.

[0013] Using a preset rule engine and a weak classifier, the values ​​in the preprocessed candidate subgraph are classified, and the classification results and the preprocessed candidate subgraph are fused to obtain a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text.

[0014] In one possible implementation, extracting medical concept and entity information from the target question text includes:

[0015] Using the target question text, construct an information extraction template text that uses preset medical concepts and entities as slots;

[0016] The information extraction template text is input into a pre-built large language model (LLM) to predict medical concepts and entity information in the target question text.

[0017] The large language model LLM is obtained by training language rules and patterns using a large-scale language dataset through autoregressive generation.

[0018] In one possible implementation, constructing an initial candidate subgraph corresponding to the target question text using medical concept and entity information in the target question text includes:

[0019] Using a pre-defined entity linking technique, medical concepts and entity information in the target question text are linked to a knowledge graph; and with each candidate entity in the knowledge graph as the center, its triple information within N hops is returned to construct an initial candidate subgraph corresponding to the target question text; where N is a positive integer greater than 0.

[0020] In one possible implementation, the value of N is 2.

[0021] In one possible implementation, the entity disambiguation processing of the initial candidate subgraph includes:

[0022] Using the initial candidate subgraph, a template text for entity disambiguation is constructed;

[0023] The template text is input into a pre-built large language model (LLM) to perform entity disambiguation processing on the initial candidate subgraph.

[0024] In one possible implementation, the step of using a preset rule engine and a weak classifier to classify the values ​​in the preprocessed candidate subgraph, and then fusing the classification results with the preprocessed candidate subgraph to obtain a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text, includes:

[0025] Using a preset rule engine, the standard range and actual values ​​in the preprocessed candidate subgraphs are normalized and compared to obtain the comparison results;

[0026] The comparison result is used as auxiliary information to input a preset weak classifier to classify the values ​​in the preprocessed candidate sub-image, and the classification result is obtained.

[0027] The classification result and the preprocessed candidate subgraph are fused together, and virtual nodes are introduced into the preprocessed candidate subgraph to form a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text.

[0028] In one possible implementation, constructing the template text for the prompt interpretation based on the candidate subgraph includes:

[0029] Based on the candidate subgraphs, comparison results, and classification results, the path is compressed, and based on the compression results, a template text for prompt interpretation is constructed.

[0030] This application also provides a medical data indicator interpretation device, including:

[0031] The extraction unit is used to acquire the target question text input by the target user and extract the medical concepts and entity information from the target question text.

[0032] The first construction unit is used to construct a candidate subgraph corresponding to the target question text based on medical concept and entity information in the target question text, using a preset rule engine and a weak classifier;

[0033] The second construction unit is used to construct template text for prompt interpretation based on the candidate subgraphs;

[0034] The determining unit is used to determine the indicator interpretation result corresponding to the target question text based on the template text interpreted by the prompt.

[0035] In one possible implementation, the first building unit includes:

[0036] The first construction subunit is used to construct an initial candidate subgraph corresponding to the target question text using medical concepts and entity information in the target question text;

[0037] The processing subunit is used to perform entity disambiguation processing on the initial candidate subgraph and numerical normalization processing on the initial candidate subgraph to obtain a preprocessed candidate subgraph.

[0038] The fusion subunit is used to classify the values ​​in the preprocessed candidate subgraph using a preset rule engine and a weak classifier, and to fuse the obtained classification results with the preprocessed candidate subgraph to obtain a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text.

[0039] In one possible implementation, the extraction unit includes:

[0040] The second construction subunit is used to construct an information extraction template text that uses preset medical concepts and entities as slots using the target question text.

[0041] The prediction subunit is used to input the information extraction template text into a pre-built large language model (LLM) to predict medical concepts and entity information in the target question text.

[0042] The large language model LLM is obtained by training language rules and patterns using a large-scale language dataset through autoregressive generation.

[0043] In one possible implementation, the first building subunit is specifically used for:

[0044] Using a pre-defined entity linking technique, medical concepts and entity information in the target question text are linked to a knowledge graph; and with each candidate entity in the knowledge graph as the center, its triple information within N hops is returned to construct an initial candidate subgraph corresponding to the target question text; where N is a positive integer greater than 0.

[0045] In one possible implementation, the value of N is 2.

[0046] In one possible implementation, the processing subunit includes:

[0047] The third construction subunit is used to construct template text for entity disambiguation using the initial candidate subgraph;

[0048] The disambiguation subunit is used to input the template text into a pre-built large language model (LLM) to perform entity disambiguation processing on the initial candidate subgraph.

[0049] In one possible implementation, the fusion subunit includes:

[0050] The comparison subunit is used to use a preset rule engine to normalize and compare the standard range and actual value in the preprocessed candidate sub-graph to obtain the comparison result.

[0051] The classification subunit is used to input the comparison result as auxiliary information into a preset weak classifier to classify the values ​​in the preprocessed candidate sub-image and obtain the classification result.

[0052] The sub-unit is used to fuse the classification result and the preprocessed candidate subgraph, and to introduce virtual nodes into the preprocessed candidate subgraph to form a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text.

[0053] In one possible implementation, the second building unit is specifically used for:

[0054] Based on the candidate subgraphs, comparison results, and classification results, the path is compressed, and based on the compression results, a template text for prompt interpretation is constructed.

[0055] This application also provides a medical data indicator interpretation device, including: a processor, a memory, and a system bus;

[0056] The processor and the memory are connected via the system bus;

[0057] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described medical data indicator interpretation methods.

[0058] This application also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described methods for interpreting medical data indicators.

[0059] This application also provides a computer program product, which, when run on a terminal device, causes the terminal device to execute any of the above-described methods for interpreting medical data indicators.

[0060] This application provides a method, apparatus, storage medium, and device for interpreting medical data indicators. First, it acquires the target question text input by the target user; then, it extracts medical concepts and entity information from the target question text; next, based on the medical concepts and entity information in the target question text, it constructs a candidate subgraph corresponding to the target question text using a preset rule engine and a weak classifier; then, based on the candidate subgraph, it constructs a template text for prompting interpretation; finally, it determines the indicator interpretation result corresponding to the target question text based on the template text for prompting interpretation.

[0061] As can be seen, this application first extracts medical concepts and entity information from the target question text, then constructs candidate subgraphs corresponding to the target question text to explicitly establish the dependencies between multiple sets of indicators, and integrates the auxiliary reasoning results of the preset rule engine and weak classifier to interpret medical data indicators. This can significantly reduce the difficulty of interpreting large models, improve the accuracy of interpreting medical data indicators, and thus improve the user experience. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart illustrating a method for interpreting medical data indicators provided in this application embodiment;

[0064] Figure 2 A system block diagram for interpreting medical data indicators provided in the embodiments of this application;

[0065] Figure 3 Example diagram of the initial candidate subgraph provided in the embodiments of this application;

[0066] Figure 4 Example diagram of the preprocessed candidate subgraph provided in the embodiments of this application;

[0067] Figure 5 Example diagram of the fused candidate subgraph provided in the embodiments of this application;

[0068] Figure 6 This is a schematic diagram of the composition of a medical data indicator interpretation device provided in an embodiment of this application. Detailed Implementation

[0069] Traditional graph-based reasoning mainly uses graph neural networks. In recent years, generative general-purpose large models (such as ChatGPT) have been commonly used to improve reasoning performance. These models rely on pre-trained corpora to integrate subject knowledge, unifying different NLP tasks into GPT tasks and using an autoregressive paradigm to achieve end-to-end generation of interpretation results.

[0070] However, using this end-to-end autoregressive generation method can easily produce text that appears grammatically correct but lacks accuracy in knowledge. One reason is that large models focus on a general and comprehensive scope, resulting in weaker modeling representations for specific knowledge. Furthermore, in the context of indicator interpretation, there are many numerical reasoning problems. Large models use natural language expressions in their reasoning, a paradigm that makes them tend to focus on the context of the text, especially local text. For complex reasoning and computation, multiple hops are needed to see relevant dependency information, which models are not adept at, leading to inaccurate interpretations of medical data indicators.

[0071] To address the aforementioned shortcomings, this application provides a method for interpreting medical data indicators. First, it acquires the target question text input by the target user; then, it extracts medical concepts and entity information from the target question text. Next, based on the medical concepts and entity information in the target question text, it constructs a candidate subgraph corresponding to the target question text using a preset rule engine and a weak classifier. Then, based on the candidate subgraph, it constructs a template text for prompting interpretation; finally, it determines the indicator interpretation result corresponding to the target question text based on this template text.

[0072] As can be seen, this application first extracts medical concepts and entity information from the target question text, then constructs candidate subgraphs corresponding to the target question text to explicitly establish the dependencies between multiple sets of indicators, and integrates the auxiliary reasoning results of the preset rule engine and weak classifier to interpret medical data indicators. This can significantly reduce the difficulty of interpreting large models, improve the accuracy of interpreting medical data indicators, and thus improve the user experience.

[0073] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0074] First Embodiment

[0075] See Figure 1This is a flowchart illustrating a method for interpreting medical data indicators provided in this embodiment. The method includes the following steps:

[0076] S101: Obtain the target question text input by the target user and extract the medical concepts and entity information from the target question text.

[0077] In this embodiment, the question text input by any user to intelligent interactive software or device (such as ChatGPT, smart speaker, etc.) using this embodiment to interpret medical data indicators is defined as the target question text to be interpreted, and the user is defined as the target user.

[0078] It should be noted that this embodiment does not limit the language type of the target question text. For example, the target question text can be Chinese text or English text, etc. Furthermore, the target question text can be directly input by the user into the intelligent interactive software, or it can be the text recognition result obtained after converting the user's input speech.

[0079] It is understood that this embodiment does not limit the length of the target question text. For example, the target question text can be a sentence text or a paragraph text. A sentence text refers to a collection of words, and a paragraph text refers to a collection of sentences. After obtaining the sentence text or paragraph text as the target question text to be interpreted, existing or future text processing methods can be used to analyze and process the target text in order to extract the medical concepts (referring to well-defined medical professional terms, also known as medical terminology) and entity information contained in the target question text, so as to determine a more accurate indicator interpretation result corresponding to the target question text by executing subsequent steps S102-S104.

[0080] It should be noted that, in order to effectively improve the accuracy of interpreting medical data indicators and enhance user experience, this application pre-constructs a static knowledge graph based on the normal range of medical data indicators and the concept types related to disease diagnosis. The graph content may include, but is not limited to, roles, genders, ages, examination and testing items, numerical ranges, diseases, and physical signs. Furthermore, the data sources for the graph are not limited; they can be from medical textbooks, journal articles, medical Wikipedia, etc. Thus, this application interprets medical data indicators based on this knowledge graph and a generative general-purpose model (such as a large language model (LLM)), integrating the auxiliary reasoning results of a pre-defined rule engine and a weak classifier. Figure 2 As shown, the detailed process will be introduced in detail later, which can greatly reduce the difficulty of interpreting large models and improve the accuracy of interpreting medical data indicators.

[0081] Specifically, one possible implementation is to first construct an information extraction template text (the specific content of which is not limited and the values ​​of each slot in the template text can be set according to the actual situation) using the target question text input by the target user after obtaining the target question text input by the target user. Then, the information extraction template text is input into a pre-built large language model LLM to predict the medical concept and entity information in the target question text.

[0082] In this implementation, the entity detection steps in Knowledge-based QA (KBQA) can be simulated. Medical concepts and entities on the knowledge graph are input to construct template text (i.e. prompt text) for information extraction. Then, LLM is used to extract medical concept and entity information from historical dialogues and current target user queries.

[0083] For example, suppose the target user inputs the following question text: "My grandfather is 83 years old. His blood creatinine is 345 and his blood urea nitrogen is 23 mmol / L. The medical report indicates abnormalities. What disease is this, and what stage is it at?". When extracting the corresponding medical concepts and entities from the target question text, the concepts can be categorized into, but are not limited to, the following:

[0084] Role: A person's identity, such as father, mother, son, or daughter.

[0085] Gender: The character's gender, whether male or female.

[0086] Age: The patient's age.

[0087] Examinations and tests: Routine examinations and tests, such as brain CT, abdominal ultrasound, complete blood count, and urinalysis.

[0088] Example model input:

[0089] Dad's temperature is 38 degrees Celsius and his hemoglobin level is 76 g / L.

[0090] Example model output:

[0091] Role: Father.

[0092] Physical signs: body temperature.

[0093] Test results: Hemoglobin.

[0094] Thus, the user input is: "Grandpa is 83 years old. During a physical examination, his blood creatinine was 345 and his blood urea nitrogen was 23 mmol / L. The physical examination report indicated abnormalities. What disease is it and what stage is it at?"

[0095] The following are examples of medical concepts and entities extracted by the model:

[0096] Character: Grandpa.

[0097] Age: 83.

[0098] Test results: serum creatinine.

[0099] Test results: Blood urea nitrogen.

[0100] It's important to note that Large Language Models (LLMs) are deep learning-based language models that can generate new language expressions, such as text, sentences, and paragraphs, based on input text content. LLMs utilize large-scale language datasets and are trained on language rules and patterns through autoregressive generation, simulating human commands to generate language expressions. Specifically, when generating new text data, LLMs predict the probability of the next language unit based on previously generated content until complete text data is generated.

[0101] During model training, a sample text can be extracted from the training data to construct a sample prompt template text, which is then used as the model input. The model's output prediction results (such as extracted entities and medical concepts) are compared with the manually labeled results corresponding to the recorded sample text, and the model parameters are updated based on the differences. This process is repeated multiple times until preset conditions are met (such as reaching a preset number of training iterations or the optimization function value remaining essentially unchanged). At this point, updating the model parameters stops, completing the training of the Large Language Model (LLM) and generating a trained LLM.

[0102] In addition, validation texts can be used to effectively validate the large language model LLM. When the predicted results are inconsistent with the actual results corresponding to the validation texts, the large language model LLM can be adjusted and updated in a timely manner, which helps to improve the prediction accuracy and precision of the large language model LLM.

[0103] S102: Based on the medical concepts and entity information in the target question text, construct the candidate subgraph corresponding to the target question text using a preset rule engine and a weak classifier.

[0104] In this embodiment, after extracting the medical concepts and entity information in the target question text through step S101, in order to effectively improve the accuracy of interpreting medical data indicators and enhance user experience, a candidate subgraph corresponding to the target question text can be constructed based on the medical concepts and entity information in the target question text using a preset rule engine and a weak classifier, so as to explicitly establish the dependency relationship between multiple sets of indicators, in order to execute the subsequent step S103.

[0105] Specifically, one possible implementation is that the specific implementation process of step S102 may include the following steps S1021-S1023:

[0106] S1021: Construct an initial candidate subgraph corresponding to the target question text by utilizing medical concepts and entity information in the target question text.

[0107] In this embodiment, after extracting the medical concepts and entity information from the target question text in step S101, in order to effectively improve the accuracy of interpreting medical data indicators and enhance user experience, the medical concepts and entity information in the target question text can be used to construct an initial candidate subgraph corresponding to the target question text for subsequent step S1022.

[0108] Specifically, one possible implementation is to use a pre-defined entity linking technique to link medical concepts and entity information in the target question text to a knowledge graph; and use each candidate entity in the knowledge graph as the center to return its triple information within N (N is a positive integer greater than 0) hops to construct an initial candidate subgraph corresponding to the target question text, in order to execute the subsequent step S1022.

[0109] In this application, the content of the preset entity linking technology and the specific value of N are not limited. They can be selected and set according to the actual situation. In a preferred implementation, the value of N can be 2.

[0110] It should be noted that when using entity linking technology to link medical concepts and entities (the original descriptions of entities in natural language, i.e., spans) in the target question text to the knowledge graph, if a span has multiple candidate entities, all of them can be retained. The candidate entities are the entities in the graph after terminology normalization. Then, taking each candidate entity in the graph as the center, the triple information within N (N can usually be taken as 2) hops is returned to construct the initial candidate subgraph corresponding to the target question text.

[0111] For example: Based on the above example, assuming the target user's input question text is still: "My grandfather is 83 years old. His blood creatinine was 345 and his blood urea nitrogen was 23 mmol / L during a physical examination. The report indicates abnormalities. What disease is it, and to what extent?", then the corresponding initial candidate subgraph could be as follows: Figure 3 As shown, there is no one-to-many mapping from a span to a standard medical entity.

[0112] S1022: Perform entity disambiguation processing on the initial candidate subgraph and numerical normalization processing on the initial candidate subgraph to obtain the preprocessed candidate subgraph.

[0113] In this embodiment, an initial candidate subgraph corresponding to the target question text is constructed through step S1021 (e.g., Figure 3 Following this, to effectively improve the accuracy of interpreting medical data indicators and enhance user experience, existing or future entity disambiguation and numerical normalization methods can be utilized. Based on the obtained analysis results and relevant annotation information of the target Chinese text in the Chinese database, entity disambiguation and numerical normalization are performed on the initial candidate subgraphs to obtain preprocessed candidate subgraphs. A specific example can be seen as follows: Figure 4 As shown, this is used to execute the subsequent step S1023.

[0114] Specifically, one possible implementation is to construct an initial candidate subgraph corresponding to the target question text (e.g., Figure 3 Furthermore, the initial candidate subgraph can be used to construct a template text for entity disambiguation. This template text can then be input into a pre-built large language model (LLM) to perform entity disambiguation on the initial candidate subgraph (if the same span corresponds to multiple standard medical entities, the large model output will only retain the correct entities).

[0115] When both inspection and testing indicator items and inspection and testing indicator numerical items exist, since the inspection and testing indicator numerical items contain indicator names, only the inspection and testing indicator numerical items need to be entered; the indicator names and standard ranges do not need to be entered, as they are assumed to be known knowledge. Simultaneously, the corresponding numerical values ​​and their units are extracted from the original text based on numerical entity types. For example, serum creatinine value: 345, blood urea nitrogen: 23 mmol / L.

[0116] For example: Based on the examples in the above steps, suppose the target user's input question text is still: "Grandpa is 83 years old. His blood creatinine is 345 and blood urea nitrogen is 23 mmol / L. The medical report indicates abnormalities. What disease is it and what stage is it at?". Example of prompt input (using the multi-turn dialogue inheritance capability of the large model). When there is no explicit relationship on the edge, the default field is empty: For the above target question text, filter the triples of the following correlation index graph, and only retain the triples related to the original question. The output format is the same as the input format.

[0117] Candidate map information:

[0118] Grandfather - belongs to - male.

[0119] 83 years old - belongs to - over 80.

[0120] Grandpa - belongs to - 80 years and above.

[0121] Male - Examination and Tests - Serum Creatinine Level.

[0122] 80 or above - Examination and testing - serum creatinine level.

[0123] Male - Examination and Tests - Blood Urea Level.

[0124] 80 or above - check and test - blood urea level.

[0125] At the same time, the actual values ​​of the inspection and testing indicators are extracted, and if there is a unit, its unit is retained.

[0126] Example of model output (fully preserved because there are no one-to-many entity links):

[0127] Preserve spectral information:

[0128] Grandfather - belongs to - male.

[0129] 83 years old - belongs to - over 80.

[0130] Grandpa - belongs to - 80 years and above.

[0131] Male - Examination and Tests - Serum Creatinine Level.

[0132] 80 or above - Examination and testing - serum creatinine level.

[0133] Male - Examination and Tests - Blood Urea Level.

[0134] 80 or above - check and test - blood urea level.

[0135] The extracted relevant indicator values ​​are as follows:

[0136] Serum creatinine level: 345.

[0137] Blood urea nitrogen: 23 mmol / L.

[0138] During the numerical normalization process of the initial candidate subgraphs, when the extracted numerical class has units and the units are within the standard units, the values ​​are directly associated with the subgraph for expansion; otherwise, the unit closest to the numerical range is selected for unit conversion. When the extracted values ​​have no units (often lacking in colloquial expression), if a node has only one set of unit edges, it is directly used as its numerical unit. If multiple sets of edge mappings exist, numerical conversion is performed on all sets of edge mappings, providing multiple interpretation results (considering all possible cases). A specific example can be seen as follows... Figure 4 As shown.

[0139] S1023: Using a preset rule engine and a weak classifier, the values ​​in the preprocessed candidate subgraphs are classified, and the classification results and the preprocessed candidate subgraphs are fused to obtain the fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text.

[0140] In this embodiment, after obtaining the preprocessed candidate subgraph through step S1022, in order to effectively improve the accuracy of interpreting medical data indicators and improve user experience, a preset rule engine and a weak classifier can be used to classify the values ​​in the preprocessed candidate subgraph to obtain the probability of belonging to various diseases as the classification result. The obtained classification result and the preprocessed candidate subgraph are then fused to obtain the fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text, for use in subsequent step S103.

[0141] Specifically, one possible implementation is as follows: First, based on the standard range and actual values ​​(based on the same unit) in the graph, a preset rule engine (the specific content is not limited and can be selected according to the actual situation) is used to normalize and compare the standard range and actual values ​​in the preprocessed candidate subgraphs to obtain comparison results (and the values ​​are judged into five preset categories: significantly high, high, normal, low, and significantly low). Then, the comparison results are used as auxiliary information to input a preset weak classifier (the specific content is not limited and can be selected according to the actual situation, such as BERT as a preset weak classifier) ​​to classify the values ​​in the preprocessed candidate subgraphs to obtain classification results. Next, the classification results and the preprocessed candidate subgraphs can be fused, and virtual nodes are introduced into the preprocessed candidate subgraphs. Edges are used to represent the relative size of the values ​​and the probability of the disease to form a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target problem text for executing the subsequent step S103.

[0142] It should be noted that this application uses a rule engine to facilitate direct comparison of numerical values, compensating for the shortcomings of large models in terms of numerical computation capabilities. The use of a weak classifier serves two purposes: firstly, to improve efficiency; and secondly, this reference information provides an auxiliary layer for the end-to-end generation of large models. Considering the cost of training large models and the efficiency of iteration, small to medium-sized models in vertical domains (which themselves can be combined with rule engines for rapid intervention) can be used to classify and determine various diseases.

[0143] For example: Based on the above example, suppose the target user's input question text is still: "My grandfather is 83 years old. His blood creatinine was 345 and his blood urea nitrogen was 23 mmol / L during a physical examination. The report indicates abnormalities. What disease is it, and to what extent?", and the preprocessed candidate subgraphs are as follows: Figure 4 As shown, the corresponding fused candidate subgraph can be as follows: Figure 5 As shown.

[0144] S103: Based on the candidate subgraphs, construct the template text for the prompt interpretation.

[0145] In this embodiment, after obtaining the candidate sub-graph (e.g., through step S102) Figure 5 After (as shown), in order to effectively improve the accuracy of medical data indicator interpretation and improve user experience, a prompt interpretation template text can be constructed based on the candidate subgraph to execute subsequent step S104.

[0146] S104: Based on the template text provided in the prompts, determine the interpretation results of the indicators corresponding to the target question text.

[0147] In this embodiment, after constructing the template text for prompt interpretation through step S103, in order to effectively improve the accuracy of medical data indicator interpretation and enhance user experience, the template text for prompt interpretation can be further input into the Large Language Model (LLM) to more accurately determine the indicator interpretation result corresponding to the target question text, thereby effectively improving the target user experience.

[0148] One possible implementation is to compress the path based on candidate subgraphs, comparison results, and classification results, and then construct a template text for prompt interpretation based on the compression results to further improve the accuracy of the indicator interpretation results.

[0149] Specifically, when constructing a prompt based on candidate subgraphs, the paradigm can include the following text sequence. In addition to the candidate subgraph, the input example can include comparison and classification results. The paths are compressed, and a template text for prompt interpretation is constructed based on the compression result. The compression method can be, but is not limited to: merging identical nodes that are both tail and head entities; merging paths if they have the same beginning and end; and introducing virtual nodes (such as...). Figure 5 The disease diagnosis node in the data can be a path splitting point, recording only the head entity and not the tail entity.

[0150] For example: Based on the above example, suppose the target user's input question text is still: "My grandfather is 83 years old. His blood creatinine was 345 and his blood urea nitrogen was 23 mmol / L during a physical examination. The report indicates abnormalities. What disease is it, and to what extent?", and the obtained pre-fused candidate subgraphs can be as follows: Figure 5 As shown. An example of inputting a prompt could be:

[0151] Based on the patient's questions above, the following key information is known.

[0152] Grandpa is male, over 80 years old, and his blood creatinine level was 345 umol / L, which is significantly high.

[0153] Grandpa is male, over 80 years old, and his blood urea level was 23 mmol / L, which is significantly high according to the examination.

[0154] Diagnostic disease - 35% probability - heart failure.

[0155] Disease diagnosis - 80% probability - chronic kidney disease.

[0156] Disease diagnosis - 50% probability - chronic glomerulonephritis.

[0157] First, identify the purpose of the inquiry. Based on the final "judgment result," introduce the examination and testing items and their results. Next, provide the normal results or numerical ranges for the examination or testing items, along with the "judgment result," but do not output probability values. If multiple "judgment results" exist, sort them according to their existing probability values, without outputting the probability values. You are required to provide a detailed explanation of the "judgment result" using professional knowledge from medical textbooks and guidelines. The answer should be logically clear, comprehensive, and can be presented item by item, while also demonstrating humanistic concern.

[0158] Combining the above candidate subgraph examples and prompt examples, the large model can ultimately determine the following interpretation results for the metrics corresponding to the target question text:

[0159] "Based on the results of your grandfather's recent examination, both his serum creatinine and blood urea nitrogen levels are above the standard range, which may indicate impaired kidney function. Generally speaking, serum creatinine and blood urea nitrogen are commonly used indicators in clinical practice to reflect the state of kidney function."

[0160] Serum creatinine is a byproduct of muscle metabolism and is excreted through the kidneys. The normal range for serum creatinine varies depending on individual factors such as age, body type, sex, and muscle mass. Mild kidney damage does not affect creatinine excretion, and serum creatinine levels remain normal. Serum creatinine only rises when the glomerular filtration rate decreases by approximately half. In adult men, creatinine levels are generally between 53-106 umol / L. Grandpa's serum creatinine level was 345, significantly higher than the normal range, suggesting potential kidney dysfunction. This abnormality may be related to factors such as aging, medication use, and malnutrition.

[0161] Blood urea nitrogen (BUN) is a metabolic product of protein. It is filtered through the glomeruli in the bloodstream and excreted in the urine. Its levels are affected not only by glomerular function but also by renal tubular function. Elevated BUN can be seen in organic kidney damage, including glomerulonephritis, pyelonephritis, and space-occupying or destructive lesions in the kidney. Acute infectious diseases, high fever, and high-protein diets can also cause elevated BUN. Protein-degrading factors such as infection, intestinal bleeding, and hyperthyroidism can all increase BUN levels. The normal range for BUN in adults is 2.9–7.1 mmol / L, while my grandfather's BUN level was as high as 23 mmol / L, significantly higher than normal. High BUN may indicate impaired kidney function and the presence of diseases such as uremia.

[0162] Based on the above analysis, combined with Grandpa's age and examination results, it can be inferred that he may have kidney damage, such as chronic kidney disease, chronic glomerulonephritis, or heart failure. To confirm or further evaluate this issue, it is recommended that you accompany Grandpa to the hospital as soon as possible for further examination and diagnosis, such as kidney function pathology and imaging examinations, to determine the cause of the kidney damage. Treatment should focus on maintaining kidney function, including controlling underlying diseases such as hypertension and diabetes, maintaining a healthy lifestyle, and taking medication regularly as directed by the doctor. At the same time, attention should be paid to Grandpa's mental health and social needs, and his diet and lifestyle habits should be carefully managed to maintain a healthy lifestyle.

[0163] In summary, the medical data indicator interpretation method provided in this embodiment first obtains the target question text input by the target user; then extracts the medical concepts and entity information from the target question text; then, based on the medical concepts and entity information in the target question text, constructs a candidate subgraph corresponding to the target question text using a preset rule engine and a weak classifier; next, constructs a template text for prompt interpretation based on the fused candidate subgraph; and finally, determines the indicator interpretation result corresponding to the target question text based on the template text for prompt interpretation.

[0164] As can be seen, this application first extracts medical concepts and entity information from the target question text, then constructs candidate subgraphs corresponding to the target question text to explicitly establish the dependencies between multiple sets of indicators, and integrates the auxiliary reasoning results of the preset rule engine and weak classifier to interpret medical data indicators. This can significantly reduce the difficulty of interpreting large models, improve the accuracy of interpreting medical data indicators, and thus improve the user experience.

[0165] Second Embodiment

[0166] This embodiment will introduce a medical data indicator interpretation device; please refer to the above method embodiment for related content.

[0167] See Figure 6 This is a schematic diagram of the composition of a medical data indicator interpretation device provided in this embodiment. The device 600 includes:

[0168] Extraction unit 601 is used to acquire the target question text to be interpreted input by the target user; and to extract medical concepts and entity information from the target question text;

[0169] The first construction unit 602 is used to construct a candidate subgraph corresponding to the target question text based on medical concept and entity information in the target question text, using a preset rule engine and a weak classifier;

[0170] The second construction unit 603 is used to construct template text for prompt interpretation based on the candidate subgraph;

[0171] The determining unit 604 is used to determine the indicator interpretation result corresponding to the target question text based on the template text interpreted by the prompt.

[0172] In one implementation of this embodiment, the first construction unit 602 includes:

[0173] The first construction subunit is used to construct an initial candidate subgraph corresponding to the target question text using medical concepts and entity information in the target question text;

[0174] The processing subunit is used to perform entity disambiguation processing on the initial candidate subgraph and numerical normalization processing on the initial candidate subgraph to obtain a preprocessed candidate subgraph.

[0175] The fusion subunit is used to classify the values ​​in the preprocessed candidate subgraph using a preset rule engine and a weak classifier, and to fuse the obtained classification results with the preprocessed candidate subgraph to obtain a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text.

[0176] In one implementation of this embodiment, the extraction unit 601 includes:

[0177] The second construction subunit is used to construct an information extraction template text that uses preset medical concepts and entities as slots using the target question text.

[0178] The prediction subunit is used to input the information extraction template text into a pre-built large language model (LLM) to predict medical concepts and entity information in the target question text.

[0179] The large language model LLM is obtained by training language rules and patterns using a large-scale language dataset through autoregressive generation.

[0180] In one implementation of this embodiment, the first construction subunit is specifically used for:

[0181] Using a pre-defined entity linking technique, medical concepts and entity information in the target question text are linked to a knowledge graph; and with each candidate entity in the knowledge graph as the center, its triple information within N hops is returned to construct an initial candidate subgraph corresponding to the target question text; where N is a positive integer greater than 0.

[0182] In one implementation of this embodiment, the value of N is 2.

[0183] In one implementation of this embodiment, the processing subunit includes:

[0184] The third construction subunit is used to construct template text for entity disambiguation using the initial candidate subgraph;

[0185] The disambiguation subunit is used to input the template text into a pre-built large language model (LLM) to perform entity disambiguation processing on the initial candidate subgraph.

[0186] In one implementation of this embodiment, the fusion subunit includes:

[0187] The comparison subunit is used to use a preset rule engine to normalize and compare the standard range and actual value in the preprocessed candidate sub-graph to obtain the comparison result.

[0188] The classification subunit is used to input the comparison result as auxiliary information into a preset weak classifier to classify the values ​​in the preprocessed candidate sub-image and obtain the classification result.

[0189] The sub-unit is used to fuse the classification result and the preprocessed candidate subgraph, and to introduce virtual nodes into the preprocessed candidate subgraph to form a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text.

[0190] In one implementation of this embodiment, the second construction unit 603 is specifically used for:

[0191] Based on the candidate subgraphs, comparison results, and classification results, the path is compressed, and based on the compression results, a template text for prompt interpretation is constructed.

[0192] Furthermore, this application embodiment also provides a medical data indicator interpretation device, including: a processor, a memory, and a system bus;

[0193] The processor and the memory are connected via the system bus;

[0194] The memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform any of the above-described implementations of the medical data indicator interpretation method.

[0195] Furthermore, embodiments of this application also provide a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described methods for interpreting medical data indicators.

[0196] Furthermore, this application also provides a computer program product, which, when run on a terminal device, causes the terminal device to execute any of the above-described methods for interpreting medical data indicators.

[0197] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0198] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0199] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0200] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for interpreting medical data indicators, characterized in that, include: Obtain the target question text input by the target user and to be interpreted; And extract medical concepts and entity information from the target question text; Using a pre-defined entity linking technique, medical concepts and entity information in the target question text are linked to a knowledge graph; and with each candidate entity in the knowledge graph as the center, its triple information within N hops is returned to construct an initial candidate subgraph corresponding to the target question text; where N is a positive integer greater than 0. The initial candidate subgraph is subjected to entity disambiguation processing and numerical normalization processing to obtain the preprocessed candidate subgraph. Based on the standard range and actual values ​​in the graph, a preset rule engine is used to normalize and compare the standard range and actual values ​​in the preprocessed candidate sub-graphs to obtain the comparison results. The comparison results are used as auxiliary information to input a preset weak classifier to classify the values ​​in the preprocessed candidate sub-images, and the classification results are obtained. The classification results and the preprocessed candidate subgraphs are fused together. Virtual nodes are introduced into the preprocessed candidate subgraphs, and edges are used to represent the relative magnitude of values ​​and the probability of diseases to form a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text. Based on the candidate subgraphs, construct template text for the prompt interpretation; Based on the template text interpreted by the prompts, determine the interpretation result of the indicator corresponding to the target question text.

2. The method according to claim 1, characterized in that, The extraction of medical concepts and entity information from the target question text includes: Using the target question text, construct an information extraction template text that uses preset medical concepts and entities as slots; The information extraction template text is input into a pre-built large language model (LLM) to predict medical concepts and entity information in the target question text. The large language model LLM is obtained by training language rules and patterns using a large-scale language dataset through autoregressive generation.

3. The method according to claim 1, characterized in that, The value of N is 2.

4. The method according to claim 1, characterized in that, The entity disambiguation processing of the initial candidate subgraph includes: Using the initial candidate subgraph, a template text for entity disambiguation is constructed; The template text is input into a pre-built large language model (LLM) to perform entity disambiguation processing on the initial candidate subgraph.

5. The method according to claim 1, characterized in that, The step of constructing the template text for prompt interpretation based on the candidate subgraph includes: Based on the candidate subgraphs, comparison results, and classification results, the path is compressed, and based on the compression results, a template text for prompt interpretation is constructed.

6. A medical data indicator interpretation device, characterized in that, include: The extraction unit is used to obtain the target question text to be interpreted, input by the target user; And extract medical concepts and entity information from the target question text; The first construction unit is used to construct a candidate subgraph corresponding to the target question text based on medical concept and entity information in the target question text, using a preset rule engine and a weak classifier; The second construction unit is used to construct template text for prompt interpretation based on the candidate subgraphs; The determining unit is used to determine the indicator interpretation result corresponding to the target question text based on the template text interpreted by the prompt; The first building unit includes: The first construction subunit is used to link medical concepts and entity information in the target question text to a knowledge graph using a preset entity linking technology; and to construct an initial candidate subgraph corresponding to the target question text by returning the triple information within N hops of each candidate entity in the knowledge graph as the center; where N is a positive integer greater than 0. The processing subunit is used to perform entity disambiguation processing on the initial candidate subgraph and numerical normalization processing on the initial candidate subgraph to obtain a preprocessed candidate subgraph. The comparison subunit is used to normalize and compare the standard range and actual value in the preprocessed candidate sub-map based on the standard range and actual value in the map, using a preset rule engine, to obtain the comparison result; The classification subunit is used to input the comparison result as auxiliary information into a preset weak classifier to classify the values ​​in the preprocessed candidate sub-image and obtain the classification result. The sub-unit is used to fuse the classification results and the preprocessed candidate subgraphs. Virtual nodes are introduced into the preprocessed candidate subgraphs, and edges are used to represent the relative magnitude of values ​​and the probability of diseases, forming a fused candidate subgraph, which serves as the candidate subgraph corresponding to the target question text.

7. A medical data indicator interpretation device, characterized in that, include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the method described in any one of claims 1-5.

Images