Report interpretation method and related device

Abstract

The invention provides a report interpretation method and a related device, and relates to the field of data interpretation. In the application, a plurality of target reports with an association relationship are acquired, report types of the target reports are identified, key information is extracted from the target reports by using the report type of each target report, a key information set is obtained, and a target retrieval result of at least one piece of target key information in the key information set is retrieved in a knowledge graph. Determining professional interpretation results corresponding to the plurality of target reports by utilizing the key information set, the target reasoning path and node paraphrases of all nodes in the target reasoning path, and performing content adjustment operation on the professional interpretation results according to common paraphrases of all the nodes in the target reasoning path to obtain the target reports. And obtaining a common interpretation result corresponding to the professional interpretation result. In the application, the general interpretation result of the target report is obtained through the report interpretation operation, so that the user can understand the report content.

Description

Technical Field

[0001] The present application relates to the field of data interpretation, and more specifically, to a report interpretation method and related device. Background Art

[0002] In actual scenarios, users will encounter various reports in their actual life or work scenarios, such as scientific research reports, experimental reports, project reports, pathological reports, etc.

[0003] After receiving a report, the user will perform a reading operation on the report. During the report reading process, due to reasons such as obscure terms in the report, the user is unable to understand the report content, reducing the user experience. Summary of the Invention

[0004] In view of this, the present application provides a report interpretation method and related device to obtain a popularized interpretation result of the report through the report interpretation operation, so as to enable the user to understand the report content.

[0005] To solve the above technical problems, the present application adopts the following technical solutions:

[0006] A report interpretation method, comprising:

[0007] Obtain a plurality of target reports having an association relationship;

[0008] Identify the report type of the target report, and use the report type of each target report to extract key information from the target report to obtain a key information set;

[0009] Retrieve a target retrieval result of at least one target key information in the key information set in the knowledge graph; the target retrieval result includes the target inference path of the target key information, the node interpretations of each node in the target inference path; the node interpretations include professional interpretations and popularized interpretations; the target inference path includes the logical relationships of each node in the knowledge graph that is associated with the target key information;

[0010] Use the key information set, the target inference path, and the node interpretations of each node in the target inference path to determine a professional interpretation result corresponding to the plurality of target reports;

[0011] According to the popularized interpretations of each node in the target inference path, perform a content adjustment operation on the professional interpretation result to obtain a popularized interpretation result corresponding to the professional interpretation result.

[0012] Optionally, identify the report type of the target report, and use the report type of each target report to extract key information from the target report to obtain a set of key information, including:

[0013] Perform text recognition and structure segmentation operations on the target report to obtain a segmentation result;

[0014] Use the segmentation result to analyze the structure layout of the target report;

[0015] Determine the report type of the target report according to the structure layout and content of the target report;

[0016] Obtain the key information to be extracted corresponding to the report type of each target report, and extract the key information from the target report;

[0017] Perform semantic alignment and conflict fusion operations on the extracted key information to obtain a set of key information.

[0018] Optionally, retrieve the target retrieval result of at least one target key information in the set of key information in the knowledge graph, including:

[0019] Determine at least one target key information from the set of key information;

[0020] Obtain the knowledge graph;

[0021] Generate the embedding vector of each node in the knowledge graph; the embedding vector includes the features of the node and the features of the neighbor nodes of the node;

[0022] Use the target key information to perform multi-hop path search in the embedding vector to obtain the target inference path of the target key information;

[0023] Obtain the node paraphrases of each node in the target inference path from the knowledge graph;

[0024] Use the target inference path of the target key information and the node paraphrases of each node in the target inference path to obtain the target retrieval result of the target key information.

[0025] Optionally, use the target key information to perform multi-hop path search in the embedding vector to obtain the target inference path of the target key information, including:

[0026] Use the target key information to perform multi-hop path search in the embedding vector to obtain multiple candidate paths;

[0027] Obtain the importance scores of each node and the edge weights in the candidate paths;

[0028] Determine the path score of the candidate path according to the importance scores of the respective nodes in the candidate path and the edge weights;

[0029] Select the target inference path of the target key information from the candidate paths according to the path score.

[0030] Optionally, determine the professional interpretation results corresponding to the multiple target reports by using the key information set, the target inference path, and the node interpretations of the respective nodes in the target inference path, including:

[0031] Perform a structuring process on the target inference path to obtain a structured path;

[0032] Construct a prompt according to the key information set, the structured path, and the node interpretations of the respective nodes in the target inference path;

[0033] Input the prompt into an interpretation model to obtain the professional interpretation results corresponding to the multiple target reports; the content output style of the interpretation model is a preset style, and the output content of the interpretation model adopts a preset structure.

[0034] Optionally, perform a content adjustment operation on the professional interpretation results according to the popularized interpretations of the respective nodes in the target inference path to obtain a popularized interpretation result corresponding to the professional interpretation results, including:

[0035] Perform a popularization process on the terms of the professional interpretation results according to the popularized interpretations of the respective nodes in the target inference path to obtain a first interpretation result corresponding to the professional interpretation results;

[0036] Perform a sentence style adjustment operation on the first interpretation result to obtain a second interpretation result;

[0037] Perform a cognitive reconstruction and word order rearrangement operation on the second interpretation result to obtain a popularized interpretation result.

[0038] Optionally, after obtaining the popularized interpretation result corresponding to the professional interpretation results, further include:

[0039] Output the popularized interpretation result corresponding to the professional interpretation results in a hierarchical display manner, and / or, in a preset style.

[0040] Optionally, after outputting the popularized interpretation result corresponding to the professional interpretation results, further include:

[0041] Adjust the weight of at least one edge in the target inference path according to the feedback information of the user on the popularized interpretation result.

[0042] A report interpretation device, comprising:

[0043] A report acquisition module, configured to acquire a plurality of target reports having an association relationship;

[0044] An information extraction module, configured to identify the report types of the target reports, and use the report types of each of the target reports to extract key information from the target reports to obtain a key information set;

[0045] A retrieval module, configured to retrieve a target retrieval result of at least one target key information in the key information set in a knowledge graph; the target retrieval result includes a target inference path of the target key information, and node interpretations of each node in the target inference path; the node interpretations include a professional interpretation and a popularized interpretation; the target inference path includes a logical relationship of each node in the knowledge graph that has an association relationship with the target key information;

[0046] An interpretation module, configured to use the key information set, the target inference path, and the node interpretations of each node in the target inference path to determine a professional interpretation result corresponding to the plurality of target reports;

[0047] A content adjustment module, configured to perform a content adjustment operation on the professional interpretation result according to the popularized interpretations of each node in the target inference path to obtain a popularized interpretation result corresponding to the professional interpretation result.

[0048] An electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0049] The memory is used to store a computer program;

[0050] The processor is configured to execute the computer program so that the electronic device can implement the above-mentioned report interpretation method.

[0051] The present application provides a method for interpreting reports and related devices. In the present application, multiple target reports with associated relationships are obtained, the report types of the target reports are identified, and using the report type of each target report, key information is extracted from the target reports to obtain a key information set. A target retrieval result of at least one target key information in the key information set is retrieved in a knowledge graph. Using the key information set, the target inference path, and the node interpretations of each node in the target inference path, a professional interpretation result corresponding to the multiple target reports is determined. According to the popularized interpretations of each node in the target inference path, a content adjustment operation is performed on the professional interpretation result to obtain a popularized interpretation result corresponding to the professional interpretation result. In the present application, through the report interpretation operation, a popularized interpretation result of the target report is obtained, so that the user can understand the report content. In addition, in the present application, when obtaining the professional interpretation result, the target retrieval result of at least one target key information in the key information set retrieved from the knowledge graph is used, that is, the content related to the target report in the knowledge graph is used to assist in generating the professional interpretation result, ensuring the correctness of the professional interpretation result. In addition, in the present application, the professional interpretation result of the target report is first obtained, and then the popularized interpretation result is further obtained based on the professional interpretation result, which can ensure that the basic data used for the popularization operation of the interpretation result is correct, ensuring the correctness of the popularized interpretation result. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the drawings in the following description are only the embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0053] Figure 1 It is a flowchart of a method for interpreting reports provided by an embodiment of the present application;

[0054] Figure 2 It is a flowchart of a method for determining a key information set provided by an embodiment of the present application;

[0055] Figure 3 It is a flowchart of a method for determining a target retrieval result provided by an embodiment of the present application;

[0056] Figure 4 It is a flowchart of a method for determining a professional interpretation result provided by an embodiment of the present application;

[0057] Figure 5 It is a flowchart of a method for determining a popularized interpretation result provided by an embodiment of the present application;

[0058] Figure 6 A structural schematic diagram of a report interpretation device provided by an embodiment of the present application;

[0059] Figure 7 A structural schematic diagram of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0060] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0061] In an actual scenario, users will come into contact with various reports in actual life or work scenarios, such as scientific research reports, experimental reports, project reports, pathological reports, etc.

[0062] Taking a pathological report as an example, after receiving the report, the user will perform a reading operation on the report. During the report reading process, due to reasons such as obscure terms in the report, the user is unable to understand the report content, reducing the user experience.

[0063] For this reason, the present application provides a report interpretation method and related device. In the present application, multiple target reports with an associated relationship are obtained, the report types of the target reports are identified, and using the report type of each target report, key information is extracted from the target reports to obtain a key information set. A target retrieval result of at least one target key information in the key information set is retrieved in the knowledge graph. Using the key information set, the target reasoning path, and the node interpretations of each node in the target reasoning path, a professional interpretation result corresponding to the multiple target reports is determined. According to the popularized interpretations of each node in the target reasoning path, a content adjustment operation is performed on the professional interpretation result to obtain a popularized interpretation result corresponding to the professional interpretation result. In the present application, through the report interpretation operation, a popularized interpretation result of the target report is obtained, so that the user can understand the report content. In addition, in the present application, when obtaining the professional interpretation result, the target retrieval result of at least one target key information in the key information set retrieved from the knowledge graph is used, that is, the content related to the target report in the knowledge graph is used to assist in generating the professional interpretation result, ensuring the correctness of the professional interpretation result. In addition, in the present application, the professional interpretation result of the target report is first obtained, and then the popularized interpretation result is further obtained based on the professional interpretation result, which can ensure that the basic data used for the popularization operation of the interpretation result is correct, ensuring the correctness of the popularized interpretation result.

[0064] In one implementation, referring to Figure 1 , one approach to interpreting the report may include:

[0065] S11. Obtain multiple target reports with associated relationships.

[0066] In this application, multiple target reports with related relationships can be input by the user. For example, if the user selects multiple reports on the mini program and clicks confirm, multiple target reports can be received. Taking the target report as a pathology report as an example, for the same user, the corresponding pathology reports are divided into at least the following types:

[0067] 1. Histopathology report: describe the morphology, grading, and diagnostic type;

[0068] 2. Molecular testing report: describing gene mutations, molecular marker detection, etc.;

[0069] 3. Integrated diagnostic report: summarizes tissue pathology and molecular testing information;

[0070] 4. Expert consultation report: including multidisciplinary joint consultation opinions.

[0071] In actual scenarios, a target report will be issued for each user after surgery. The target report can be multiple of the above reports. These reports are reports on the same lesion site of the same user, that is, there is a correlation between multiple target reports.

[0072] This application has different restrictions on the type of target report. It can be a text type such as .doc, .txt, .xml, or an image type such as .pdf (scanned or native), .jpg, .png, or a system interface type, such as structured interface data in HIS (Hospital Information System) / EMR (Electronic Medical Record System).

[0073] Compared with general physical examination reports, pathology reports are more complex in the following aspects:

[0074] 1. A wide range of report types (histopathology, molecular testing, integrated diagnosis, expert consultation, etc.);

[0075] 2. Inconsistent expression (sometimes formal terms, sometimes abbreviations or notes);

[0076] 3. Highly professional, with large variations in terminology status combinations;

[0077] 4. There may be multiple reports for the same patient, and the contents need to be combined and used.

[0078] Therefore, this application can adopt a report parsing method with semantic understanding, layout perception and strong generalization capabilities to extract unified and structured semantic information from pathology reports with different formats, complex content and variable structures, including diagnostic information, molecular markers and their status, tumor grade, etc., to provide standardized input for subsequent graph reasoning and language generation modules.

[0079] S12. Identify the report type of the target report, and extract key information from the target report using the report type of each target report to obtain a key information set.

[0080] In this application, by identifying the visual anchor points in the target report, the report type to which each part of the target report belongs is determined. The report type can be a tissue pathology report, a molecular detection report, an integrated diagnosis report, and an expert consultation report.

[0081] For target reporting, key information that can be extracted include:

[0082] 1) Immunohistochemistry results: such as "GFAP (Glial Fibrillary Acidic Protein) (brain tissue positive)" and "Ki-67 (approximately 1-2%)";

[0083] 2) Pathological diagnosis information: such as "astrocytoma", "anaplastic oligodendroglioma", etc.;

[0084] 3) Tumor grade information: such as "CNS WHO grade 2" or "CNS WHO grade 1-2";

[0085] 4) Histological features: such as "scattered multinucleated cells" and "active mitotic activity";

[0086] 5) Molecular markers and their status: such as "IDH1 gene R132, no mutation", "ATRX gene, no mutation", "chromosome 1p / 19q, no co-deletion", etc.

[0087] 6) Clinical recommendations or treatment plans: such as "reviewing cranial MRI every 6 months after surgery" or "considering stereotactic radiotherapy with a gamma knife";

[0088] 7) Report type identification: such as "histopathology report", "molecular detection report", "integrated diagnosis report", "expert consultation report", etc.

[0089] The extracted key information can be composed into a key information set.

[0090] S13. Retrieve a target retrieval result of at least one target key information in the key information set in the knowledge graph.

[0091] Among them, the knowledge graph refers to the knowledge graph related to the user's lesion type. For example, if the user's lesion site is the brain, the knowledge graph is the knowledge graph related to head and neck tumors or the knowledge graph related to central nervous system tumors.

[0092] The data sources for constructing the knowledge graph can be:

[0093] Authoritative knowledge bases: UMLS (Unified Medical Language System), MeSH (National Center for Biotechnology Information Medical Subject Headings), PubTator 3.0 (Public Text Annotator, a tool for mining biomedical literature resources driven by artificial intelligence), OncoKB (Precision Oncology Knowledge Base);

[0094] Structured information in clinical guidelines: such as "The 5th Edition World Health Organization Classification of Tumors of the Central Nervous System (WHO CNS2021)", "Clinical Practice Guidelines: Central Nervous System Tumors 2024.V3";

[0095] Doctor annotation: Supplement local diagnosis and treatment paths.

[0096] The knowledge graph adopts a heterogeneous graph structure (Heterogeneous Graph). In the knowledge graph, nodes represent medical entities (such as diseases, molecular markers, diagnostic attributes, etc.), and edges represent the relationships between medical entities (such as cause, prediction, recommendation, etc.).

[0097] The type attributes of the nodes can be specifically referred to Table 1:

[0098] Table 1

[0099]

[0100] For the knowledge graph, when configuring node attributes, each node comes with:

[0101] Standard name (such as: IDH1 gene R132);

[0102] Popular alternative expressions (such as: a gene mutation that determines the malignancy of glioma, with a better prognosis for the mutation);

[0103] Source citation (such as guideline chapter, literature identification ID);

[0104] Entity type label, etc.

[0105] The type attribute of the side can be specifically referred to Table 2:

[0106] Table 2

[0107]

[0108] After constructing the knowledge graph, it is possible to search in the knowledge graph to obtain a target retrieval result of at least one target key information in the key information set.

[0109] The target retrieval result includes the target inference path of the target key information and the node interpretations of each node in the target inference path.

[0110] Among them, the target inference path can also be called a path chain. The target inference path includes the logical relationships of each node in the knowledge graph that has an associated relationship with the target key information. Among them, taking the target key information as "IDH1 mutation" as an example, the target inference path can be "IDH1 mutation" → "slow progression" → "good prognosis" → "treatment recommendation". "Slow progression", "good prognosis", and "treatment recommendation" are all nodes in the knowledge graph that have an associated relationship with "IDH1 mutation", and the logical relationship among the three is: "IDH1 mutation" → "slow progression" → "good prognosis" → "treatment recommendation".

[0111] Each node in the target inference path has a corresponding node interpretation. The node interpretation includes a professional interpretation and a popularized interpretation. The professional interpretation refers to the interpretation of the node described using professional terms, and the popularized interpretation refers to the interpretation obtained by popularizing the professional interpretation.

[0112] For example, if the professional interpretation is "slow tumor growth", the popularized interpretation is "Your tumor develops relatively mildly and has a low risk".

[0113] S14. Use the key information set, the target inference path, and the node interpretations of each node in the target inference path to determine the professional interpretation results corresponding to multiple target reports.

[0114] In this application, the professional interpretation results corresponding to multiple target reports refer to structured, editable, and semantically reasonable medical interpretation results. Among them, for the input multiple target reports, since the multiple target reports are all about the same disease, the finally obtained professional interpretation result is only one.

[0115] When specifically obtaining the professional interpretation result, it can be implemented using a model. The output result of the model is different from the general abstract or the Q&A output. The output result has the following characteristics:

[0116] 1. Have a fixed structure: including "diagnostic induction + molecular mechanism explanation + suggestion prompt";

[0117] 2. The content closely follows the inference results of the pathological atlas;

[0118] 3. Expressed in the style of a professional doctor, for machine reconstruction or manual editing.

[0119] S15. According to the popularized interpretations of each node in the target inference path, perform content adjustment operations on the professional interpretation results to obtain the popularized interpretation results corresponding to the professional interpretation results.

[0120] In this application, performing content adjustment operations on the professional interpretation results means converting the professional interpretation results into popularized interpretation results that can be understood by ordinary patients, emotionally acceptable, and logically easy to follow. The content adjustment operations can include:

[0121] 1. Popularized explanation of terms (deprofessionalization of medical terms): When explaining terms popularly, the popularized interpretations of each node in the target inference path can be used;

[0122] 2. Adjustment of emotional color (replacement with an encouraging tone);

[0123] 3. Reconstruction of logical structure (changing from "doctor writing to doctor" to "doctor speaking to patient");

[0124] 4. Natural tone, gentle semantics, and empathic expression.

[0125] Through the above content adjustment operations, the popularized interpretation results corresponding to the professional interpretation results are finally obtained. Compared with the professional interpretation results, the popularized interpretation results are easier for users to understand.

[0126] In this embodiment, multiple target reports with associated relationships are obtained, the report types of the target reports are identified, and using the report types of each target report, key information is extracted from the target reports to obtain a key information set. The target retrieval results of at least one target key information in the key information set are retrieved in the knowledge graph. Using the key information set, the target inference path, and the node interpretations of each node in the target inference path, the professional interpretation results corresponding to the multiple target reports are determined. According to the popular interpretations of each node in the target inference path, a content adjustment operation is performed on the professional interpretation results to obtain the popular interpretation results corresponding to the professional interpretation results. In this application, through the operation of interpreting the reports, the popular interpretation results of the target reports are obtained, so that users can understand the report content. In addition, in this application, when obtaining the professional interpretation results, the target retrieval results of at least one target key information in the key information set retrieved from the knowledge graph are used, that is, the content related to the target reports in the knowledge graph is used to assist in generating the professional interpretation results, ensuring the correctness of the professional interpretation results. In addition, in this application, the professional interpretation results of the target reports are obtained first, and then the popular interpretation results are further obtained based on the professional interpretation results, which can ensure that the basic data used for the popularization operation of the interpretation results is correct, ensuring the correctness of the popular interpretation results.

[0127] Based on any of the above embodiments, in another implementation, with reference to Figure 2 , identifying the report types of the target reports and using the report types of each target report to extract key information from the target reports to obtain a key information set may include:

[0128] S21. Perform text recognition and structure splitting operations on the target report to obtain a splitting result.

[0129] In this application, if the target report is in the PDF (Portable Document Format) / image format, Donut (OCR-free Document Understanding Transformer) or PaddleOCR (Paddle Optical Character Recognition, an optical character recognition tool based on PaddlePaddle) can be used to recognize paragraphs, tables, headers, and footers in the PDF / image, and the original position mapping is retained for report paragraph division to obtain a splitting result.

[0130] For text-type data or system interface-type data, paragraph division can be directly performed to obtain a splitting result.

[0131] S22. Analyze the structural layout of the target report using the segmentation results.

[0132] In this application, the visual anchor technology can be used to analyze the structural layout of the target report using the segmentation results.

[0133] Among them, the visual anchor technology can be implemented through the Document Intelligence Multimodal Pretrained Model (LayoutLMv3). After fine-tuning LayoutLMv3, LayoutLMv3 can be used to identify the visual anchors in the target report based on the segmentation results of each paragraph and determine the structural layout of the target report.

[0134] S23. Determine the report type of the target report according to the structural layout and content of the target report.

[0135] In this application, the classifier uses the structural layout and content of the target report to identify the report type of each target report through position awareness + field semantic matching (report types such as histopathological reports, molecular detection reports, etc.). Among them, position awareness + field semantic matching means:

[0136] Identify the report type by analyzing the structural layout in the document (such as the title position, table structure (such as table headers), etc.) and text content (such as keywords like "molecular detection report", "integrated diagnostic report", etc.).

[0137] Specifically, position awareness means determining the report type by the appearance position of the report name in the target report. For example, "molecular detection report" and "integrated diagnostic report" generally appear at the beginning of the report or in the table header. Field semantic matching means identifying the keywords in the target report and matching the keywords with the keywords in each report type to obtain the report type to which the target report belongs.

[0138] It should be noted that the integrated diagnostic report generally includes two parts: histopathology and molecular detection. Subsequently, the classifier returns both the report interval (page number, paragraph position) and the report type label.

[0139] S24. Obtain the key information to be extracted corresponding to the report type of each target report, and extract the key information from the target report.

[0140] In this application, the key information to be extracted corresponding to the report type of each target report is pre-configured. For example, for a report type of histopathological report, the corresponding key information to be extracted is pathological diagnosis information, immunohistochemical results, histological features, tumor grading information, etc.

[0141] Then, perform key information extraction operations on the target report according to the key information to be extracted corresponding to the report type of each target report to obtain the content of each key information.

[0142] When actually performing the key information extraction operation, a model such as DeepSeek-V3 or GLM-4 (GLM: General Language Model) can be used to perform imperative structured extraction on the paragraph text of the target report. Among them, the model supports fine-tuning with a small amount of data instructions to improve the ability to understand specialized disease terms.

[0143] The extraction granularity supported by DeepSeek-V3 or GLM-4 is as follows:

[0144] Terms: tumor name, gene, chromosome;

[0145] Attributes: mutation, methylation, expression status;

[0146] Grading: CNS WHO grading.

[0147] Prompt example for DeepSeek-V3 or GLM-4: Please extract the pathological diagnosis and molecular test results in the report, and the output results are represented in JSON (JavaScript Object Notation).

[0148] The output example of DeepSeek-V3 or GLM-4 is as follows:

[0149] {

[0150] "diagnosis": "astrocytoma",

[0151] "grade": "CNS WHO grade 2",

[0152] "marker": {

[0153] "IDH1 gene R132": "mutation",

[0154] "ATRX gene": "mutation",

[0155] "chromosome 1p / 19q": "non-codelletion"

[0156] }

[0157] }

[0158] S25. Perform semantic alignment and conflict fusion operations on the extracted key information to obtain a key information set.

[0159] In practical applications, the same patient may have multiple different types of pathological reports, and there may be cases of duplicate or conflicting information. To fuse these reports, unify the terminology, eliminate redundancy, and ensure the consistency and integrity of the data.

[0160] When performing fusion, semantic alignment operations need to be carried out on the same terminology. For example, "astrocytoma" appears in the tissue pathology report, and "IDH1 gene mutation" appears in the molecular detection report. It is necessary to unify the diagnosis / marker fields through entity name standardization + semantic similarity fusion.

[0161] Entity name standardization: Unify synonyms or variants that appear in different reports into standard terms. For example, "astrocytic tumor" and "astrocytoma" are unified as "astrocytoma".

[0162] Semantic similarity calculation: Use the pre-trained BioBERT (Biomedical Language Representation Model) model to calculate the similarity between terms and identify possible synonyms or related terms. For example, "IDH1 gene R132" and "IDH1 gene" have a high similarity and can be regarded as the same entity, and they are unified as "IDH1 gene".

[0163] The above semantic alignment operations are for the processing of non-conflicting terms. When there are conflicts between terms, conflict resolution strategies are used for conflict fusion operations.

[0164] Among them, in the conflict resolution strategy, according to factors such as report type and chronological order, determine the priority rules for choosing or discarding conflicting information. For example, preferentially adopt the conclusion of the integrated diagnostic report. If there is no integrated diagnostic report, the histological features are based on the tissue pathology report, and the molecular information is based on the molecular detection report.

[0165] After performing semantic alignment and conflict fusion operations, unified structured data is generated to obtain a key information set, which contains all key medical entities and their statuses for subsequent analysis.

[0166] In one example, the key information set is output in JSON structure, and its example is:

[0167] {

[0168] "patient_id": "P123456",

[0169] "reports":

[0170] {

[0171] "type": "tissue pathology report",

[0172] "diagnosis": "Astrocytoma",

[0173] "histo": "Scattered multinucleated cells",

[0174] "grade": "CNS WHO Grade 2"

[0175] },

[0176] {

[0177] "type": "Molecular test report",

[0178] "markers": {

[0179] "IDH1 gene R132": "Mutation",

[0180] "ATRX gene": "Mutation",

[0181] "Chromosome 1p / 19q": "Non-codelletion"

[0182] }

[0183] },

[0184] {

[0185] "type": "Integrated diagnosis report",

[0186] "summary": "IDH-mutant diffuse glioma",

[0187] "diagnosis": "Astrocytoma",

[0188] "grade": "CNS WHO Grade 2",

[0189] "markers": {

[0190] "IDH1 gene R132": "Mutation",

[0191] "ATRX gene": "Mutation",

[0192] "Chromosome 1p / 19q": "Non-codelletion"

[0193] }

[0194] }

[0196] }。

[0197] ​In this embodiment, in the obtained target report, the key information of the target report is extracted, and only the key information is analyzed in the subsequent operations, reducing the data processing volume of the subsequent operations and improving the efficiency.

[0198] Based on any of the above embodiments, in one implementation, referring to Figure 3 , the target retrieval results of at least one target key information in the key information set retrieved in the knowledge graph may include:

[0199] S31. Determine at least one target key information from the key information set.

[0200] In this application, when the key information set includes the key information of the integrated diagnostic report, the key information belonging to the integrated diagnostic report in the key information set is integrated and used as the target key information.

[0201] When the key information set does not include the key information of the integrated diagnostic report, the key information belonging to the tissue pathology report and the molecular detection report in the key information set can be integrated and used as the target key information.

[0202] It should be noted that other strategies can also be selected to determine at least one target key information from the key information set.

[0203] S32. Obtain the knowledge graph.

[0204] Among them, the explanation of the knowledge graph refers to the corresponding description above.

[0205] S33. Generate the embedding vectors of each node in the knowledge graph.

[0206] Among them, the embedding vector includes the features of the node and the features of the neighboring nodes of the node.

[0207] To improve the search efficiency, the embedding vectors of each node in the knowledge graph can be generated. In one implementation, a graph neural network (GNN) can be used to perform the operation of generating the embedding vectors. GNN is a neural network that can process graph-structured data and is suitable for scenarios modeling complex relationships between nodes. Specifically, GNN can be a graph attention network model, that is, the learning of node embeddings can be achieved through the graph attention network model.

[0208] In the knowledge graph, nodes represent medical entities and edges represent the relationships between medical entities. Through the GNN network, an embedding vector can be generated for each node. This vector not only contains the features of the node itself but also integrates the features of the neighboring nodes, thereby capturing the context relationships between medical entities. These embedding vectors can be used for subsequent path search, node classification, etc. tasks, improving the system's understanding ability of medical terms.

[0209] S34. Use the target key information to perform a multi-hop path search in the embedding vectors to obtain the target inference path of the target key information.

[0210] Specifically, when performing a multi-hop path search, a search method based on multi-hop path search + edge weight control can be used. During the search, starting from the input node corresponding to the target key information, the interpretable path is recursively deduced downward.

[0211] Among them, multi-hop path search refers to the path search process in a knowledge graph that starts from a starting node, passes through multiple intermediate nodes, and finally reaches the target node. This search can simulate the doctor's diagnostic reasoning process.

[0212] Edge weight control refers to setting the weights of the edges in the graph during the path search process to affect the path selection. The edge weights can be set according to factors such as the recommendation level of medical guidelines and the confidence of nodes.

[0213] In specific implementation, the multi-hop path search can adopt the breadth-first search (BFS, Breadth-First Search) algorithm. The breadth-first search combines the edge weight priority strategy to preferentially explore the target inference paths with higher weights.

[0214] In one implementation, step S34 includes:

[0215] 1) Use the target key information to perform a multi-hop path search in the embedding vectors to obtain multiple candidate paths.

[0216] During the specific search, the construction of node embeddings is the basis of path search, providing semantic information between nodes, and path search is to explore the inference path on this basis. Therefore, in specific implementation, it is necessary to first construct the embedding vectors of nodes through a graph neural network to obtain the context representation between medical entities. Then, these embedding vectors can be used to perform a multi-hop path search on the input node corresponding to the target key information, retrieve semantically similar nodes, and find multiple better candidate paths.

[0217] It should be noted that during the multi-hop path search, semantic rationality should be ensured: for example, "IDH1 mutation" cannot directly jump to "immunotherapy recommendation", and there should be "good prognosis" and "slow growth" in between.

[0218] 2) Obtain the importance scores of each node and the edge weights in the candidate paths.

[0219] In this application, after determining the candidate paths, it is necessary to select the optimal target inference path from multiple candidate paths in combination with the edge weight control strategy. When using the edge weight control strategy, it is necessary to know the importance scores of each node in the candidate path and the edge weights. Among them, the edge weights are determined based on the guideline recommendation level + node confidence, and the importance scores of the nodes can be configured according to the actual situation.

[0220] 3) Determine the path score of the candidate path according to the importance scores of each node in the candidate path and the edge weights.

[0221] Among them, after obtaining at least one candidate path through the above search steps, the same term may lead to multiple nodes. It is preferred to retain the authoritative path and the path with clear clinical significance. Specifically, when implementing, a path scoring function can be used to preferentially select the path that is more meaningful in medicine, avoid jumping reasoning, and ensure the rationality and authority of the inference path.

[0222] In one implementation, after obtaining at least one candidate path, use the path scoring function to complete the sorting operation of the candidate paths. The content of the path scoring function is:

[0223] ;

[0224] Among them, is the score of the candidate path, is the edge weight in the candidate path, and the edge weight reflects the importance of this edge in medical reasoning. is the importance score of the node in the candidate path, and this importance score takes into account the medical significance and confidence of the node.

[0225] The path score of each candidate path can be calculated through the path scoring function.

[0226] 4) Select the target inference path of the target key information from the candidate paths according to the path score.

[0227] In this application, select the candidate path corresponding to the maximum path score and use it as the target inference path of the target key information.

[0228] The breadth-first search in this application combines the edge weight first strategy. Compared with traditional medical information retrieval, there are the following specific differences:

[0229] Traditional medical information retrieval usually stays at the level of the definition of a single term and lacks an in-depth understanding of the causal relationship between terms. By constructing a knowledge graph containing causal relationships, the system can start from a term and perform multi-hop reasoning along the medical logic path to gradually reveal relevant diagnostic implications, prognostic information, and treatment recommendations. For example: the inference path from "IDH1 mutation" to treatment recommendations.

[0230] Starting term: "IDH1 mutation"

[0231] First hop: "IDH1 mutation" → "Slow tumor growth"

[0232] Second hop: "Slow tumor growth" → "Better prognosis"

[0233] Third hop: "Better prognosis" → "Radiotherapy and chemotherapy can be postponed."

[0234] S35. Obtain the node interpretations of each node in the target inference path from the knowledge graph.

[0235] In this application, the node interpretations of each node are pre-configured in the knowledge graph and can be directly obtained.

[0236] S36. Use the target inference path of the target key information and the node interpretations of each node in the target inference path to obtain the target retrieval result of the target key information.

[0237] Among them, the target retrieval result can be output in JSON format. In one example, the target retrieval result is:

[0238] {

[0239] "term": "IDH1 mutation",

[0240] "diagnostic_path":

[0241] {

[0242] "node": "Slow tumor growth",

[0243] "type": "Clinical inference",

[0244] "plain_text": "Your tumor is developing relatively mildly with low risk"

[0245] },

[0246] {

[0247] "node": "Better prognosis",

[0248] "type": "Diagnostic attribute",

[0249] "plain_text": "Most patients of this type have good treatment effects"

[0250] },

[0251] {

[0252] "node": "Radiotherapy and chemotherapy can be postponed",

[0253] "type": "Treatment recommendation",

[0254] "plain_text": "Treatment can sometimes be deferred, with only regular follow-up."

[0255] }

[0257] }。

[0258] In this application, by summarizing the target inference path of the target key information and the node interpretations of each node in the target inference path, the target retrieval result of the target key information can be obtained.

[0259] In this embodiment, vector retrieval is performed in the knowledge graph to solve the problem of the lack of complex relationships between professional terms.

[0260] In this embodiment, multi-hop semantic path retrieval is adopted in the knowledge graph of medical causal logic to form an inference chain of inference path terms → diagnostic meaning → prognostic information → recommended treatment, which not only provides factual retrieval content, but also can simulate the doctor's reasoning process. Compared with traditional retrieval techniques, it can achieve interpretive knowledge guidance rather than simply information collage.

[0261] The target retrieval result in this application can be used for subsequent model generation. This method not only provides the definition of terms, but also reveals their significance in clinical diagnosis and their impact on treatment decisions. This kind of inference chain can be used as a prompt word to be input into the subsequent model to help it generate more clinically valuable interpretation content.

[0262] Based on any of the above embodiments, in one implementation, referring to Figure 4 , using the key information set, the target inference path, and the node interpretations of each node in the target inference path, the professional interpretation results corresponding to multiple target reports can be determined, which may include:

[0263] S41. Structurally process the target inference path to obtain a structured path.

[0264] In this application, an example of the target inference path is:

[0265] For the medical entity "IDH1 mutation", its corresponding target inference path is: "IDH1 mutation" → "Slow tumor growth" → "Good prognosis" → "Chemoradiotherapy can be deferred".

[0266] Structurally process the extracted target inference path to obtain a structured path.

[0267] ​S42. Construct a prompt according to the node interpretations of each node in the key information set, the structured path, and the target reasoning path.

[0268] In this application, the prompt is used to guide the large language model to generate content. Taking the target reasoning path of IDH mutation → slow tumor growth → better prognosis → deferral of radiotherapy and chemotherapy as an example, when embedding the structured path into the prompt, text prompt embedding or vector embedding can be used.

[0269] Among them, text prompt embedding means taking the structured path corresponding to the target reasoning path as a text prompt and directly using it as a part of the prompt. In the prompt, the nodes and edges in the target reasoning path can form a structured prompt pair.

[0270] Vector embedding means using a graph attention network to encode the nodes and edges of the structured path, generating vector representations, and taking these vectors as a part of the prompt to enhance the model's understanding of entity relationships. In addition, the vector representations obtained in this step can also be embedded into the intermediate layer of the model, and the intermediate layer can be, for example, an attention layer, to guide the model to use relevant medical knowledge for interpretation.

[0271] In this application, a report interpretation model is used for report interpretation operations. The report interpretation model can be a large language model with task prompt understanding capabilities such as DeepSeek-V3 and GLM-4. The output of the report interpretation model needs to have a clear paragraph distribution and semantic markings.

[0272] The input of the report interpretation model includes two parts:

[0273] 1. The key information set. Specific examples can be:

[0274] Core report terms (such as "astrocytoma", "IDH gene mutation", "ATRX gene mutation");

[0275] Report type (histopathological report / molecular test report / integrated diagnostic report / expert consultation report, etc.);

[0276] Tumor grading information (such as "CNS WHO grade 2");

[0277] Test results (mutation / no mutation / non-deletion status, etc.);

[0278] A JSON example of a key information set is:

[0279] {

[0280] "diagnosis": "astrocytoma",

[0281] "grade": "CNS WHO Grade 2",

[0282] "markers": {

[0283] "IDH1 gene R132": "mutation",

[0284] "ATRX gene": "mutation",

[0285] "Chromosome 1p / 19q": "non-codelletion"

[0286] }

[0287] }。

[0288] 2. Structure the target reasoning path to obtain the structured path.

[0289] Among them, the target reasoning path is "IDH1 mutation" → "slow growth" → "good prognosis" → "delaying radiotherapy and chemotherapy".

[0290] The structured path corresponding to the target reasoning path is:

[0292] { "term": "IDH1 gene mutation", "implication": "low-grade glioma, good prognosis"},

[0293] { "term": "Chromosome 1p / 19q non-codelletion", "implication": "non-oligodendroglioma, good prognosis"}

[0294] 。

[0295] 3. The node interpretations of each node in the target reasoning path.

[0296] Among them, the node interpretations include professional interpretations and popular interpretations. The popular interpretations are used to assist the large model in interpretation processing

[0297] In one example, the prompts of the interpretation model include the following three types of structural tasks:

[0298] TASK_DIAGNOSIS: Extract and summarize the diagnostic semantic paragraphs;

[0299] TASK_MOLECULAR: Explain the molecular markers and their clinical significance;

[0300] TASK_ADVICE: Give suggestions and next steps based on the atlas reasoning path.

[0301] ​Embed task type tags in each prompt to control the model to focus on different semantic segments:

[0302] For example:

[0303]

Task Type

[0304]

Prompt Content

[0305]

Atlas Prompt

[0306] The prompt also includes:

[0307]

Attention Content

[0308] The output requirement of the interpretation model is a medical-style semantic explanation

[0309] In one example, a small number of structured examples (Few-shot Schema) can be used for guided (example prompt) or dynamic task label generation of the prompt. The specific implementation of the Few-shot Schema guidance is:

[0310] Introduce similar cases to generate examples as imitation templates to control the output style, format, and content density of the model.

[0311] Example prompt:

[0312] Example input:

[0313]

Diagnosis

[0314]

Molecular

[0315]

Atlas

[0316] Example output:

[0317] The patient has oligodendroglioma, graded as CNS WHO grade 2. Molecular testing supports the classic oligodendroglial features, with a better prognosis. Follow-up observation is recommended.

[0318] Please generate a similar format according to the following input:

[0319]

Diagnosis

[0320] S43. Input the prompt into the interpretation model to obtain the professional interpretation results corresponding to multiple target reports.

[0321] Among them, the content output style of the interpretation model is the preset style, and the output content of the interpretation model adopts the preset structure.

[0322] In one implementation, the content output style of the interpretation model is: medical professional style (preset style);

[0323] The output content adopts the preset structure: three - part style (diagnosis → molecular explanation → treatment suggestion). In one example, the output format of the interpretation model is a structured paragraph:

[0324] {

[0325] "diagnosis_part": "The patient was diagnosed with astrocytoma, grade 2 of CNS WHO, belonging to the category of low - grade gliomas.",

[0326] "molecular_part": "It was detected that there was a mutation in the R132 site of the IDH1 gene, indicating that the tumor growth was relatively slow. The ATRX gene mutation and the non - co - deletion of chromosome 1p / 19q suggest that this tumor is more inclined to astrocytoma.",

[0327] "suggestion_part": "Currently, it is recommended to conduct regular re - examinations and evaluate whether to initiate radiotherapy and chemotherapy regimens in combination with the disease progression."

[0328] }

[0329] In this application, when the prompt is input into the interpretation model, the interpretation model can generate content containing professional explanations. For example: The patient has a mutation in the IDH1 gene, which is usually associated with slow tumor growth and a better prognosis. Therefore, it may be considered to delay radiotherapy and chemotherapy and adopt a wait - and - watch strategy.

[0330] The output of the interpretation model is the professional interpretation results corresponding to multiple target reports, and an example of the professional interpretation results is shown above.

[0331] The professional interpretation results in this application support doctors' rapid review and manual supplementation to ensure the accuracy of the professional interpretation results.

[0332] In this application, the professional interpretation results corresponding to multiple target reports are obtained through the interpretation model, providing basic data for subsequent popularization processing.

[0333] Based on any of the above embodiments, refer to Figure 5, according to the popular interpretations of each node in the target inference path, perform content adjustment operations on the professional interpretation results to obtain the popular interpretation results corresponding to the professional interpretation results, including:

[0334] S51. According to the popular interpretations of each node in the target inference path, perform the operation of popularizing terms on the professional interpretation results to obtain the first interpretation result corresponding to the professional interpretation result.

[0335] Among them, the operation of popularizing terms can adopt the method of rewriting with popular terms (De-jargonization). Specifically, when implementing, instruction-tuning models such as DeepSeek and GLM can be used. The instruction-tuning model uses the popular interpretations of each node in the target inference path to rewrite the terms in the professional interpretation results and adjust them into concise language.

[0336] The prompt example of the instruction-tuning model is:

[0337] Please rewrite the following medical terms into language that ordinary people can understand, avoid using professional terms, and retain the medical meaning:

[0338] Term: IDH1 gene R132 mutation

[0339] Hint: Appears in astrocytoma patients, and its popular interpretation is...

[0340] Output example:

[0341] "There is a change in a gene called IDH1, which usually means that the tumor grows relatively slowly and the condition is relatively stable."

[0342] After the instruction-tuning model performs the operation of popularizing terms on the professional interpretation results according to the popular interpretations of each node in the target inference path, the first interpretation result corresponding to the professional interpretation result can be obtained.

[0343] S52. Perform the operation of adjusting the sentence style on the first interpretation result to obtain the second interpretation result.

[0344] In this application, an emotion-aware tone transformation module can be used to perform the operation of adjusting the sentence style on the first interpretation result. The technical core of the emotion-aware tone transformation module is:

[0345] Based on emotion dimension injection + large model sentence restructuring + style memory template library to achieve tone optimization.

[0346] The control parameter dimensions of the emotion-aware tone transformation module include:

[0347] Emotional polarity (positive / neutral / negative);

[0348] Tone style (doctor's tone / empathetic tone / scientific explanation tone);

[0349] Focus information density (diagnosis / comfort / advice).

[0350] In this application, the emotional polarity can be selected as positive, the tone style can be selected as empathetic tone, and the focus information density can be selected as comfort.

[0351] The model in the emotional style regulation module can use Qwen2.5-instruct or InstructGLM (the instruction fine-tuned version of the GLM model) to reconstruct the sentence pattern, enhancing empathy and positive guidance.

[0352] The prompt example of the model is:

[0353] Original sentence: "This tumor is of low grade and has a good prognosis."

[0354] Transformed sentence: "Your tumor belongs to a type with a slower growth rate. Most patients can control the condition well, which is a positive sign."

[0355] By performing sentence style adjustment operations on the first interpretation result through the model in the above-mentioned emotional style regulation module, the second interpretation result can be obtained.

[0356] S53. Perform cognitive reconstruction and word order rearrangement operations on the second interpretation result to obtain a popularized interpretation result.

[0357] In this application, the Conceptual Reframing Engine is used to perform cognitive reconstruction and word order rearrangement operations on the second interpretation result.

[0358] The technical core of the Conceptual Reframing Engine is:

[0359] "Reconstruct" the structure of the professional draft into an expression order that is smoother for patients to understand.

[0360] The Conceptual Reframing Engine uses the structure rearrangement method of "phenomenon → explanation → advice" to make the information closer to the cognitive logic of patients;

[0361] Structure hint example:

[0362]

Input paragraph

[0363]

Output Paragraph

[0364] The cognitive restructuring and word order rearrangement module is used to perform cognitive restructuring and word order rearrangement operations on the second interpretation result to obtain a popularized interpretation result. An example of the popularized interpretation result is:

[0365] {

[0366] "patient_friendly_text": "Your test results show a relatively mild brain tumor...",

[0367] "key_terms":

[0368] {

[0369] "term": "IDH1 mutation",

[0370] "simplified_explanation": "This usually indicates slower tumor growth and a good prognosis"

[0371] }

[0372] ,

[0373] "suggestion_summary": "It is recommended to perform regular MRI examinations to observe the trend of changes"

[0374] }。

[0375] In this embodiment, the professional interpretation result is transformed into a popularized interpretation result that is professionally rigorous, understandable by ordinary patients, emotionally acceptable, and logically easy to follow, and the condition is explained to the patient in a positive and encouraging tone, facilitating user understanding and acceptance.

[0376] Based on any of the above embodiments, after obtaining the popularized interpretation result corresponding to the professional interpretation result, the popularized interpretation result corresponding to the professional interpretation result can also be output in a hierarchical display manner and / or a preset style.

[0377] In this embodiment, the popularized interpretation result adopts a structural rearrangement method of "phenomenon → explanation → suggestion", including:

[0378] The main explanatory text;

[0379] The key terms and their popular explanations;

[0380] The suggestion summary paragraph.

[0381] When outputting the popularized interpretation results, supplementary explanatory information from the knowledge graph can also be output, and the supplementary explanatory information can be viewed using the term click-to-view function.

[0382] When outputting information, the output information can be presented in a hierarchical display manner, and the hierarchical display manner is configured according to the logic of compressing from professional to popular expansion. In one example:

[0383] First layer: The directly readable main text segment (popular explanation, text with optimized emotion);

[0384] Second layer: Clickable term cards (hover to show an easy-to-understand explanation) showing popular interpretations + graph paths;

[0385] Third layer: Structured suggestion area (such as review frequency, possible treatment paths);

[0386] Fourth layer (optional): Doctor team notes, historical comparisons.

[0387] In another implementation, the style / role, etc. of the output information can also be selected.

[0388] Output styles available for users to choose:

[0389] Doctor's tone (formal, precise);

[0390] Sympathetic tone (gentle, encouraging);

[0391] Family member reminder style (auxiliary communication text);

[0392] Each style calls different preset prompts or output templates, and uses the corresponding model to adjust the style of the information to be output to obtain the final output information.

[0393] In one implementation, the output information also supports multi-language and cross-language output. When outputting in multiple languages or cross-languages, the output supports automatic switching between Chinese and English.

[0394] When converting languages, a large model can be used to generate the output information after language conversion, maintaining medical accuracy;

[0395] An example of language conversion using a large model is:

[0396] "IDH1 gene mutation" → "A genetic alteration associated with a milder disease condition".

[0397] In this application, the popularized interpretation results are presented to patients in a multi-modal and hierarchical structure, improving the understanding efficiency, emotional acceptance, and engagement. In one implementation, after outputting the popularized interpretation results corresponding to the professional interpretation results, the weights of at least one edge in the target inference path can also be adjusted according to the feedback information of the user on the popularized interpretation results.

[0398] In specific implementation, after information output, the user can provide corresponding feedback by using clicks & actions. Examples of providing corresponding feedback by using clicks & actions are:

[0399] Click on a term / expand the description → Record the points of interest;

[0400] Skip a term / close the module → Mark as "redundant / difficult to understand";

[0401] Like / don't understand button → Feedback on the model effect;

[0402] Submit a follow-up question / new question → Collect the need for secondary Q&A (for use in optimizing step 6).

[0403] In one implementation, the feedback information of the user obtained through the above clicks & actions can be saved using a behavior data structure. In one example, the behavior data structure is:

[0404] {

[0405] "user_id": "U001",

[0406] "interaction": {

[0407] "clicked_terms": ["IDH1", "ATRX"],

[0408] "skipped_sections": ["Suggestion section"],

[0409] "feedback": {

[0410] "understandable": true,

[0411] "emotion_tone": "soothing",

[0412] "followup_question": "Do I not need chemotherapy anymore?"

[0413] }

[0414] }

[0415] }。

[0416] Subsequently, a continuous learning mechanism is constructed to reverse-optimize the selection of plain explanation versions, the adjustment weight of emotional tone, the priority order of graph paths, the generation and expression of polysemous terms, and the adaptation of personalized styles by collecting users' cognitive feedback, click behaviors, language acceptance, and follow-up question behaviors.

[0417] In one implementation, the data used for optimization is the above-mentioned feedback information of users, including:

[0418] Collected user behavior data;

[0419] The record of term usage in the interpreted text;

[0420] Failed generated examples (such as paragraphs marked by users as "unreadable");

[0421] The content of follow-up questions raised by users.

[0422] In one implementation, evolutionary learning of polysemous expression of terms (evolution of plain expression) can be carried out based on users' feedback information.

[0423] In specific implementation, multiple plain expression versions are maintained for each term, and A / B testing is conducted on the multi-version expressions of a certain term to collect click / understanding distributions.

[0424] Example: "IDH1 mutation" →

[0425] "A type of tumor change";

[0426] "A signal of milder illness";

[0427] "A sign of slow development and good treatment response".

[0428] Subsequently, statistical indicators such as users' click-through rate, average reading duration, and whether follow-up questions are triggered in the feedback information can be obtained. Using these statistical indicators of users' click-through rate, average reading duration, and whether follow-up questions are triggered, the sorting strategy of plain term explanations is updated in a positive feedback manner.

[0429] In one implementation, an optimization scheme for graph inference paths can be realized using users' feedback information. The implementation principle is as follows:

[0430] Record the target inference path used when generating the result of plain interpretation (such as: IDH1 mutation → slow progression → delayed treatment). If the result of plain interpretation corresponding to this path is marked as "difficult to understand" by multiple users, lower the edge weight of this path, update the path scoring function, and improve the priority of paths with high understandability.

[0431] During specific implementation, when invoking the knowledge graph to generate popularized explanation results, the target reasoning path referred to by the large language model will be recorded. The target reasoning path is not a static definition lookup, but a "multi-hop semantic retrieval path" executed based on the knowledge graph structure.

[0432] This target reasoning path will be explicitly written into the prompt or implicitly passed into the model (such as as context embedding or control vector).

[0433] After the generation task of the popularized explanation result is completed, this target reasoning path will be bound and recorded in the generation task ID and stored in the log system for subsequent feedback learning.

[0434] When the user can perform operations in the usage interface of the popularized explanation result, such as marking "don't understand", "difficult to understand", skipping paragraphs, asking follow-up questions, etc. Analyze the relationship between these behaviors and the target reasoning path they use. Once a certain path repeatedly causes cognitive difficulties in the use of multiple users, the edge weights of some edges in this path (for example, "slow progress" → "delayed treatment") will be down-regulated. Then update the edge weight scores in the path scoring function, and this update will affect the future path search and sorting. Paths that are more intuitive, commonly used, or have a higher recommended priority in the guidelines will be ranked in the front. If this target reasoning path is repeatedly marked as "don't understand", "difficult to understand", skipped paragraphs, asked follow-up questions, etc., this path will be marked as "for doctors only" or "avoid presenting to patients".

[0435] In one example, the target reasoning path is: "TP53 mutation" → "gene repair disorder" → "disease heterogeneity" → "poor prognosis". This path was marked as "difficult to understand" or "unclear terms" by 78% of the users. At this time, the edge "gene repair disorder" → "disease heterogeneity" will be optimized, and its edge weight will be down-regulated from 0.9 to 0.2.

[0436] In this application, a self-learning mechanism after introducing user feedback is used to optimize the quality of the knowledge graph path to recommend more appropriate interpretation results for users.

[0437] In one implementation, the user preferences (emotional style, paragraph length, acceptance of medical knowledge) can be determined by combining the user behavior clustering results, and a user style preference portrait is constructed. An example of the user style preference portrait is: "tend to be positive, like popular long sentences, and accept more technical explanations".

[0438] The user style preference portrait can be used for personalized template recommendation or parameter configuration in the next generation task to provide users with popularized interpretation results that are more matched to the portrait.

[0439] In one implementation, the present application also supports model re - fine - tuning and reinforcement learning. Model re - fine - tuning and reinforcement learning can be, for example: updating the common - term library, fine - tuning the emotion template word library, retraining the edge weights of the knowledge graph, continuously enriching the fine - tuning data set of the large model using RLHF (Reinforcement Learning from Human Feedback) or SFT (Supervised Fine - Tuning), and regularly making lightweight updates to models such as DeepSeek and GLM to strengthen high - score outputs and optimize model outputs.

[0440] In the present application, a continuous learning mechanism is constructed. By collecting users' cognitive feedback, click behavior, language acceptance, and follow - up behavior, the parameters in the previous steps are optimized in reverse, so that the final popularized interpretation result has a higher matching degree with the user.

[0441] It should be noted that by using lightweight large - model fine - tuning and prompt optimization techniques, the model used in the present application is fine - tuned to achieve the purpose of accurately and directionally optimizing the large model with only a small number of annotated reports. Finally, the entire process from extraction to generation takes into account the patient's emotions and cognitive level, truly realizing "making it understandable for patients to hear and see", and achieving the goal of patient - centeredness.

[0442] Based on the above - mentioned embodiments of the report interpretation method, another embodiment of the present application provides a report interpretation device. Referring to Figure 6 , it may include:

[0443] A report acquisition module 11, configured to acquire multiple target reports with an associated relationship;

[0444] An information extraction module 12, configured to identify the report type of the target report, and use the report type of each target report to extract key information from the target report to obtain a key information set;

[0445] A retrieval module 13, configured to retrieve a target retrieval result of at least one target key information in the key information set in the knowledge graph; the target retrieval result includes the target reasoning path of the target key information, and the node interpretations of each node in the target reasoning path; the node interpretation includes a professional interpretation and a popularized interpretation; the target reasoning path includes the logical relationship of each node in the knowledge graph that has an associated relationship with the target key information;

[0446] An interpretation module 14, configured to determine professional interpretation results corresponding to multiple target reports by using the key information set, the target reasoning path, and the node interpretations of each node in the target reasoning path;

[0447] A content adjustment module 15, configured to perform a content adjustment operation on the professional interpretation result according to the popularized interpretations of the nodes in the target inference path, so as to obtain a popularized interpretation result corresponding to the professional interpretation result.

[0448] In one implementation, the information extraction module 12 includes:

[0449] A splitting sub-module, configured to perform text recognition and structure splitting operations on the target report to obtain a splitting result;

[0450] A layout determination sub-module, configured to analyze the structure layout of the target report by using the splitting result;

[0451] A type determination sub-module, configured to determine the report type of the target report according to the structure layout and content of the target report;

[0452] An extraction sub-module, configured to obtain the key information to be extracted corresponding to the report type of each target report, and extract the key information from the target report;

[0453] A processing sub-module, configured to perform semantic alignment and conflict fusion operations on the extracted key information to obtain a key information set.

[0454] In one implementation, the retrieval module 13 includes:

[0455] An information determination sub-module, configured to determine at least one target key information from the key information set;

[0456] A knowledge graph acquisition sub-module, configured to acquire a knowledge graph;

[0457] A vector generation sub-module, configured to generate embedding vectors for each node in the knowledge graph; the embedding vectors include the features of the node and the features of the neighbor nodes of the node;

[0458] A search sub-module, configured to perform a multi-hop path search in the embedding vectors by using the target key information to obtain a target inference path of the target key information;

[0459] An interpretation acquisition sub-module, configured to acquire the node interpretations of the nodes in the target inference path from the knowledge graph;

[0460] A result determination sub-module, configured to obtain a target retrieval result of the target key information by using the target inference path of the target key information and the node interpretations of the nodes in the target inference path.

[0461] In one implementation, the search sub-module includes:

[0462] A search unit, configured to perform a multi-hop path search in the embedding vectors by using the target key information to obtain multiple candidate paths;

[0463] A data acquisition unit for acquiring the importance scores of each node and the edge weights in the candidate paths.

[0464] A scoring determination unit for determining the path score of the candidate path according to the importance scores of each node and the edge weights in the candidate path.

[0465] A selection unit for selecting the target inference path of the target key information from the candidate paths according to the path score.

[0466] In one implementation, the interpretation module 14 includes:

[0467] A structured processing sub-module for performing structured processing on the target inference path to obtain a structured path.

[0468] A prompt word construction sub-module for constructing prompt words according to the key information set, the structured path, and the node interpretations of each node in the target inference path.

[0469] An interpretation sub-module for inputting the prompt words into an interpretation model to obtain professional interpretation results corresponding to multiple target reports; the content output style of the interpretation model is a preset style, and the output content of the interpretation model adopts a preset structure.

[0470] In one implementation, the content adjustment module 15 includes:

[0471] A first adjustment sub-module for performing a term popularization processing operation on the professional interpretation results according to the popularized interpretations of each node in the target inference path to obtain a first interpretation result corresponding to the professional interpretation results.

[0472] A second adjustment sub-module for performing a sentence style adjustment operation on the first interpretation result to obtain a second interpretation result.

[0473] A third adjustment sub-module for performing a cognitive reconstruction and word order rearrangement operation on the second interpretation result to obtain a popularized interpretation result.

[0474] In one implementation, it further includes:

[0475] A display module for outputting the popularized interpretation result corresponding to the professional interpretation result in a hierarchical display manner and / or a preset style.

[0476] In one implementation, it further includes:

[0477] An optimization module for adjusting the weight of at least one edge in the target inference path according to the feedback information of the user on the popularized interpretation result.

[0478] In this embodiment, multiple target reports with associated relationships are obtained, the report types of the target reports are identified, and using the report types of each target report, key information is extracted from the target reports to obtain a key information set. At least one target retrieval result of the target key information in the key information set is retrieved in the knowledge graph. Using the key information set, the target inference path, and the node interpretations of each node in the target inference path, a professional interpretation result corresponding to the multiple target reports is determined. According to the popular interpretations of each node in the target inference path, a content adjustment operation is performed on the professional interpretation result to obtain a popular interpretation result corresponding to the professional interpretation result. In this application, the popular interpretation result of the target report is obtained through the report interpretation operation, so that the user can understand the report content. In addition, in this application, when obtaining the professional interpretation result, at least one target retrieval result of the target key information in the key information set retrieved from the knowledge graph is used, that is, the content related to the target report in the knowledge graph is used to assist in generating the professional interpretation result, ensuring the correctness of the professional interpretation result. In addition, in this application, the professional interpretation result of the target report is obtained first, and then the popular interpretation result is further obtained based on the professional interpretation result, which can ensure that the basic data used for the popularization operation of the interpretation result is correct, ensuring the correctness of the popular interpretation result.

[0479] It should be noted that for the working processes of each module, sub-module, and unit in this application, please refer to the corresponding descriptions in the above embodiments and will not be elaborated here.

[0480] This application embodiment also provides an electronic device, including at least one processor and a memory connected to the processor, where:

[0481] The memory is used to store a computer program;

[0482] The processor is used to execute the computer program so that the electronic device can implement the above-mentioned report interpretation method.

[0483] Reference Figure 7 As shown, it shows a schematic structural diagram of an electronic device suitable for implementing the electronic device in this application embodiment. The electronic device in this application embodiment may include, but is not limited to, fixed terminals such as mobile phones, laptop computers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), desktop computers, and the like. Figure 7 The shown electronic device is only an example and should not bring any limitations to the functions and usage scopes of this application embodiment.

[0484] Such as Figure 7As shown, the electronic device may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 601, which may perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 602 or the program loaded from the storage device 608 into the random access memory (RAM) 603. When the electronic device is powered on, various programs and data required for the operation of the electronic device are also stored in the RAM 603. The processing device 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604. The input / output (I / O) interface 605 is also connected to the bus 604.

[0485] Generally, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a memory card, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device to communicate with other devices wirelessly or wirelesly to exchange data. Although Figure 7 an electronic device with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. Instead, more or fewer devices may be implemented or had.

[0486] An embodiment of the present application also provides a computer program product including computer-readable instructions. When the computer-readable instructions run on an electronic device, the electronic device is enabled to implement any one of the report interpretation methods provided by the embodiments of the present application.

[0487] An embodiment of the present application also provides a computer-readable storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can be enabled to implement any one of the report interpretation methods provided by the embodiments of the present application.

[0488] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but rather will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for interpreting reports, characterized in that, Including: Obtain multiple target reports with associated relationships; Identify the report types of the target reports, and use the report types of each target report to extract key information from the target reports to obtain a key information set; Retrieve the target retrieval results of at least one target key information in the key information set in the knowledge graph; the target retrieval results include the target inference path of the target key information, and the node interpretations of each node in the target inference path; the node interpretations include professional interpretations and popularized interpretations; the target inference path includes the logical relationships of each node in the knowledge graph that is associated with the target key information; Use the key information set, the target inference path, and the node interpretations of each node in the target inference path to determine the professional interpretation results corresponding to the multiple target reports; According to the popularized interpretations of each node in the target inference path, perform a content adjustment operation on the professional interpretation results to obtain the popularized interpretation results corresponding to the professional interpretation results.

2. The report interpretation method according to claim 1, wherein Identify the report types of the target reports, and use the report types of each target report to extract key information from the target reports to obtain a key information set, including: Perform text recognition and structure segmentation operations on the target reports to obtain a segmentation result; Use the segmentation result to analyze the structure layout of the target reports; Determine the report types of the target reports according to the structure layout and content of the target reports; Obtain the key information to be extracted corresponding to the report type of each target report, and extract the key information from the target reports; Perform semantic alignment and conflict fusion operations on the extracted key information to obtain a key information set.

3. The report interpretation method according to claim 1, characterized in that Retrieve the target retrieval results of at least one target key information in the key information set in the knowledge graph, including: Determine at least one target key information from the key information set; Obtain a knowledge graph; Generate the embedding vectors of each node in the knowledge graph; the embedding vectors include the features of the node and the features of the neighbor nodes of the node; Use the target key information to perform multi-hop path search in the embedding vectors to obtain the target inference path of the target key information; Obtain the node interpretations of each node in the target inference path from the knowledge graph; Use the target inference path of the target key information and the node interpretations of each node in the target inference path to obtain the target retrieval results of the target key information.

4. The report interpretation method according to claim 3, wherein Use the target key information to perform multi-hop path search in the embedding vectors to obtain the target inference path of the target key information, including: Use the target key information to perform multi-hop path search in the embedding vectors to obtain multiple candidate paths; Obtain the importance scores of each node in the candidate paths and the edge weights; Determine the path scores of the candidate paths according to the importance scores of each node in the candidate paths and the edge weights; Select the target inference path of the target key information from the candidate paths according to the path score.

5. The report interpretation method according to claim 1, characterized in that Using the key information set, the target inference path, and the node interpretations of each node in the target inference path, determine the professional interpretation results corresponding to multiple target reports, including: Perform structural processing on the target inference path to obtain a structured path; Construct a prompt word according to the key information set, the structured path, and the node interpretations of each node in the target inference path; Input the prompt word into the interpretation model to obtain the professional interpretation results corresponding to the multiple target reports; the content output style of the interpretation model is a preset style, and the output content of the interpretation model adopts a preset structure.

6. The report interpretation method according to claim 1, wherein According to the popularized interpretations of each node in the target inference path, perform content adjustment operations on the professional interpretation results to obtain the popularized interpretation results corresponding to the professional interpretation results, including: According to the popularized interpretations of each node in the target inference path, perform term popularization operations on the professional interpretation results to obtain the first interpretation results corresponding to the professional interpretation results; Perform sentence style adjustment operations on the first interpretation results to obtain the second interpretation results; Perform cognitive reconstruction and word order rearrangement operations on the second interpretation results to obtain the popularized interpretation results.

7. The report interpretation method according to claim 1, wherein After obtaining the popularized interpretation results corresponding to the professional interpretation results, it also includes: Output the popularized interpretation results corresponding to the professional interpretation results in a hierarchical display manner and / or a preset style.

8. The report interpretation method according to claim 7, wherein After outputting the popularized interpretation results corresponding to the professional interpretation results, it also includes: Adjust the weight of at least one edge in the target inference path according to the feedback information of the user on the popularized interpretation results.

9. A report interpretation device, characterized in that, Include: A report acquisition module for acquiring multiple target reports with an associated relationship; An information extraction module for identifying the report types of the target reports and using the report types of each target report to extract key information from the target reports to obtain a key information set; A retrieval module for retrieving the target retrieval results of at least one target key information in the key information set in the knowledge graph; the target retrieval results include the target inference path of the target key information, the node interpretations of each node in the target inference path; the node interpretations include professional interpretations and popularized interpretations; the target inference path includes the logical relationships of each node in the knowledge graph that is associated with the target key information; An interpretation module for using the key information set, the target inference path, and the node interpretations of each node in the target inference path to determine the professional interpretation results corresponding to multiple target reports; A content adjustment module for performing content adjustment operations on the professional interpretation results according to the popularized interpretations of each node in the target inference path to obtain the popularized interpretation results corresponding to the professional interpretation results.

10. An electronic device, characterized in that, Include at least one processor and a memory connected to the processor, where: The memory is used to store computer programs; The processor is configured to execute the computer program so that the electronic device can implement the report interpretation method according to any one of claims 1 to 8.

Images