A disease risk prediction method, device and electronic equipment

Abstract

The application relates to the technical fields of medical artificial intelligence, medical health informatization and natural language processing, and specifically provides a disease risk prediction method and device and electronic equipment, which are used for improving the problem that the accuracy of a disease risk prediction mode is low. The method comprises the following steps: performing initial prediction on the disease type of a suspected patient according to the medical information of the suspected patient, obtaining an initial prediction type and a confidence score of the initial prediction type; performing entity extraction on the medical information according to the initial prediction type and the confidence score of the initial prediction type, to obtain a plurality of entity words, wherein the entity extraction is performed according to a corresponding standard dictionary set corresponding to the initial prediction type; and performing reasoning on the medical information according to the plurality of entity words through a pre-trained language model, to obtain disease risk information of the suspected patient, wherein the disease risk information comprises a disease risk type and a disease risk score.

Description

Technical Field

[0001] The present application relates to the technical fields of medical artificial intelligence, medical health informationization, and natural language processing, and specifically to a disease risk prediction method, device, and electronic equipment. Background Art

[0002] Currently, disease risk prediction is primarily based on analyzing patient medical information using structured data (such as laboratory indicators) based on a physician's personal clinical experience and general natural language processing (NLP) technology. However, this NLP technology, based on individual physician experience, typically relies on static dictionaries, which are difficult to adapt to the terminology of different medical institutions (such as hospital-defined terminology and dialect expressions). Consequently, current disease risk prediction methods have low accuracy. Summary of the Invention

[0003] The purpose of the embodiments of the present application is to provide a disease risk prediction method, device and electronic device for improving the problem of low accuracy of disease risk prediction methods.

[0004] An embodiment of the present application provides a disease risk prediction method, including: making an initial prediction of the disease type of the suspected patient based on the medical information of the suspected patient, obtaining the initial prediction type and the confidence score of the initial prediction type; determining a corresponding standard dictionary set based on the initial prediction type and the confidence score of the initial prediction type to perform entity extraction on the medical information to obtain multiple entity words; and inferring the medical information based on the multiple entity words through a pre-trained language model to obtain the disease risk information of the suspected patient, wherein the disease risk information includes: a disease risk type and a disease risk score. In the implementation process of the above solution, disease risk prediction is performed by combining the entity extraction of the standard dictionary set with the reasoning ability of the pre-trained language model. Since the initial prediction type determines the corresponding standard dictionary set to ensure the standardization of entity extraction and avoids prediction bias caused by terminology ambiguity or expression differences, it can more accurately capture the key medical features in medical information. Furthermore, the pre-trained language model, with its context understanding ability, not only reduces the burden of the pre-trained language model in processing redundant data, but also can capture subtle clues in complex cases, so as to mine the potential correlation between entity words, thereby enhancing the model's ability to distinguish similar diseases, and ultimately generating more reliable disease risk types and disease risk scores, effectively improving the recognition accuracy of disease risk types.

[0005] Optionally, in an embodiment of the present application, an initial prediction is made of the disease type of the suspected patient based on the medical information of the suspected patient, including: splitting the medical information into multiple structured fields; for each structured field in the multiple structured fields, deleting the preset disease diagnosis content from the structured field to obtain multiple processed fields; using a disease type prediction tool to make an initial prediction of the disease type of the suspected patient based on the multiple processed fields, to obtain the initial prediction type and the confidence score of the initial prediction type. In the implementation of the above scheme, by deleting the preset disease diagnosis content, the model's dependence on existing diagnostic results is avoided, and the deviation caused by interference from prior diagnostic information in the prediction process is reduced, so as to ensure that the prediction model is only based on objective medical data that is not mixed with clinical prior judgments for analysis, thereby improving the objectivity and credibility of the prediction results of the pre-trained language model.

[0006] Optionally, in an embodiment of the present application, entity extraction is performed on the medical information based on the initial prediction type and the confidence score of the initial prediction type, including: judging whether the confidence score of the initial prediction type is greater than a preset threshold score; if so, searching the dictionary database for the standard dictionary set corresponding to the initial prediction type, and performing entity extraction on the medical information based on the standard dictionary set to obtain multiple entity words, wherein the dictionary database stores standard dictionary sets corresponding to multiple disease types. In the implementation process of the above scheme, by setting the confidence score threshold of the initial prediction type, the system calls the corresponding standard dictionary set for entity extraction only when the prediction result has a high degree of confidence. This mechanism effectively improves the erroneous entity recognition that may be caused by low-confidence predictions, ensures that the extracted entity words are highly relevant to the disease type, and thus significantly improves the reliability of the entity extraction results.

[0007] Optionally, in an embodiment of the present application, after determining whether the confidence score of the initial prediction type is greater than a preset threshold score, the method further includes: if the confidence score of the initial prediction type is less than or equal to the preset threshold score, then using the trained entity extraction model to extract entities from the medical information to obtain multiple candidate entities, and screening the multiple candidate entities to obtain multiple entity words. During the implementation of the above solution, when the confidence score of the initial prediction type is low, the system instead calls the trained entity extraction model to perform full entity recognition on the medical information to avoid missing key entities due to the limitations of the standard dictionary set. In the implementation process of the above solution, the solution can cover a wider range of potential entities (such as rare disease-related words or non-standard expressions) through the generalization ability of the entity extraction model (such as the context-aware characteristics based on the pre-trained language model). At the same time, through the subsequent candidate entity screening mechanism (such as ranking and screening multiple candidate entities by importance based on the LDA topic model or TF-IDF weighting method to obtain multiple entity words), low-weight or irrelevant entities are filtered out, thereby reducing noise interference while ensuring coverage, and significantly improving the system's entity extraction robustness in low-confidence scenarios.

[0008] Optionally, in an embodiment of the present application, a pre-trained language model is used to infer medical information based on multiple entity words, including: searching for knowledge blocks in a vector database based on multiple entity words, where the vector database stores knowledge blocks for determining multiple disease types; and inferring the medical information based on the knowledge blocks and multiple entity words through a pre-trained language model to obtain disease risk information of suspected patients. In the implementation of the above scheme, the system can dynamically fuse medical field knowledge and semantic information of entity words in the reasoning process by combining the semantic understanding ability of the pre-trained language model and the structured knowledge blocks in the vector database. The synergistic effect of this dual information source reduces the model's dependence on surface text information, local language patterns, or single text features, and enhances the ability to capture complex medical logic, thereby significantly improving the accuracy of disease risk reasoning.

[0009] Optionally, in an embodiment of the present application, a knowledge block is searched in a vector database based on multiple entity words, including: searching the knowledge graph for all entity node words associated with each entity word in the multiple entity words and the attribute relationship of each entity node word; generating a text sentence from the attribute relationship of all entity node words and each entity node word, and converting the text sentence into a sentence embedding vector; searching the vector database for a knowledge block whose similarity with the sentence embedding vector is greater than a preset similarity threshold. In the implementation of the above scheme, by searching the knowledge graph for all entity node words and their attribute relationships associated with all entities, the system can prioritize focusing on semantically strongly related neighboring entities and the attribute relationship of each entity node word, and generate text sentences based on the attribute relationship, rather than a generalized global association. This local association screening mechanism significantly reduces the interference of irrelevant or weakly related entities, so that subsequent knowledge block retrieval is only performed on highly relevant candidate entities, thereby improving the accuracy and pertinence of knowledge block retrieval.

[0010] Optionally, in an embodiment of the present application, a pre-trained language model is used to infer medical information based on knowledge blocks and multiple entity words, including: filling the knowledge blocks and multiple entity words into a preset prompt word template to obtain a prompt engineering text; and reasoning the medical information based on the prompt engineering text through a multi-intelligent agent model. In the implementation process of the above scheme, by filling the knowledge blocks and entity words into the preset prompt word template, the system can clearly express the semantic boundaries and logical relationships of the input information in a structured manner. The preset template serves as a unified instruction framework, which constrains the generation space of the model and avoids the reasoning deviation caused by vague or ambiguous expressions in traditional free text input. Moreover, the multi-intelligent agent model can cross-verify the output results of different subtasks by reasoning on the prompt engineering text through division of labor and cooperation, reducing the subjective error of a single model, thereby significantly improving the accuracy of the overall reasoning and the consistency of the results.

[0011] Optionally, in an embodiment of the present application, a multi-intelligent agent model includes: a management agent intelligent agent, an inspection and analysis agent intelligent agent, and a result summary agent intelligent agent; the multi-intelligent agent model is used to reason about the medical information of the visit according to the prompt engineering text, including: the management agent intelligent agent extracts the inspection item tasks and the test item tasks in the medical information of the visit according to the prompt engineering text, and assigns the inspection item tasks and the test item tasks to the inspection and analysis agent intelligent agent; the inspection information and the test information of the medical information of the visit are analyzed according to the prompt engineering text by the inspection and analysis agent intelligent agent to obtain multiple analysis results; the result summary agent intelligent agent summarizes and infers the multiple analysis results to obtain disease risk information of suspected patients. In the implementation process of the above solution, by managing the intelligent agent specifically responsible for extracting inspection project tasks and testing project tasks from the prompt engineering text and assigning them to the inspection and analysis agent, the system can decouple complex medical reasoning tasks into specific executable subtasks. This modular division of labor mechanism avoids the ambiguity of traditional single models in task identification and allocation, ensuring that each subtask is handled by an intelligent agent with corresponding capabilities, thereby significantly improving the accuracy and execution efficiency of task decomposition.

[0012] An embodiment of the present application also provides a disease risk prediction device, including: a disease type initial prediction module, which is used to make an initial prediction of the disease type of a suspected patient based on the suspected patient's medical information, and obtain the initial prediction type and the confidence score of the initial prediction type; a medical information extraction module, which is used to determine the corresponding standard dictionary set and the confidence score of the initial prediction type based on the initial prediction type to perform entity extraction on the medical information and obtain multiple entity words; a disease risk information acquisition module, which is used to infer the medical information based on multiple entity words through a pre-trained language model to obtain the disease risk information of the suspected patient, and the disease risk information includes: disease risk type and disease risk score.

[0013] Optionally, in an embodiment of the present application, the disease type initial prediction module includes: a medical information splitting submodule, which is used to split the medical information of the visit into multiple structured fields; a disease diagnosis deletion submodule, which is used to delete the preset disease diagnosis content from each structured field in the multiple structured fields, and obtain the processed multiple fields; a prediction type acquisition submodule, which is used to make an initial prediction of the disease type of the suspected patient based on the processed multiple fields through a disease type prediction tool, and obtain the initial prediction type and the confidence score of the initial prediction type.

[0014] Optionally, in an embodiment of the present application, the medical information extraction module includes: a confidence score judgment submodule, used to judge whether the confidence score of the initial prediction type is greater than a preset threshold score; an information entity extraction submodule, used to search the dictionary database for a standard dictionary set corresponding to the initial prediction type if the confidence score of the initial prediction type is greater than the preset threshold score, and perform entity extraction on the medical information based on the standard dictionary set to obtain multiple entity words. The dictionary database stores standard dictionary sets corresponding to multiple disease types.

[0015] Optionally, in an embodiment of the present application, the medical information extraction module further includes: an entity extraction and screening submodule, which is used to use the trained entity extraction model to perform entity extraction on the medical information if the confidence score of the initial prediction type is less than or equal to the preset threshold score, to obtain multiple candidate entities, and to screen the multiple candidate entities to obtain the multiple entity words.

[0016] Optionally, in an embodiment of the present application, the disease risk information acquisition module includes: a knowledge block search sub-module, which is used to search for knowledge blocks in a vector database based on multiple entity words, and the vector database stores knowledge blocks for determining multiple disease types; a medical information reasoning sub-module, which is used to infer medical information based on knowledge blocks and multiple entity words through a pre-trained language model to obtain disease risk information of suspected patients.

[0017] Optionally, in an embodiment of the present application, the knowledge block search submodule includes: an entity word search unit, which is used to search for all entity node words associated with each entity word in a plurality of entity words and the attribute relationship of each entity node word in the knowledge graph; an embedding vector conversion unit, which is used to generate a text sentence from the attribute relationship of all entity node words and each entity node word, and convert the text sentence into a sentence embedding vector; a vector similarity comparison unit, which is used to search for knowledge blocks in the vector database whose similarity with the sentence embedding vector is greater than a preset similarity threshold.

[0018] Optionally, in an embodiment of the present application, the medical information reasoning submodule includes: a prompt text acquisition unit, used to fill the knowledge blocks and multiple entity words into a preset prompt word template to obtain a prompt engineering text; a medical information reasoning unit, used to reason about the medical information of the visit based on the prompt engineering text through a multi-intelligent agent model.

[0019] Optionally, in an embodiment of the present application, a multi-intelligent agent model includes: a management agent intelligent agent, an inspection and analysis agent intelligent agent, and a result summary agent intelligent agent; a medical information reasoning unit includes: a task extraction and assignment subunit, which is used to extract the inspection item tasks and the test item tasks in the medical information of the visit according to the prompt engineering text through the management agent intelligent agent, and assign the inspection item tasks and the test item tasks to the inspection and analysis agent intelligent agent; an analysis result acquisition subunit, which is used to analyze the inspection information and the test information of the medical information of the visit according to the prompt engineering text through the inspection and analysis agent intelligent agent to obtain multiple analysis results; a result summary reasoning subunit, which is used to summarize and conclude the multiple analysis results through the result summary agent intelligent agent to obtain disease risk information of suspected patients.

[0020] An embodiment of the present application further provides an electronic device, including: a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions execute the method described above when executed by the processor.

[0021] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method described above is executed.

[0022] An embodiment of the present application further provides a computer program product, including: a computer program or computer instructions, which executes the method described above when the computer program or computer instructions are executed by a processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only illustrate certain embodiments of the embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 Schematic diagram of the process of disease risk prediction method provided by the embodiment of the present application;

[0025] Figure 2 A schematic diagram of the structure of a disease risk prediction device provided in an embodiment of the present application is shown;

[0026] Figure 3 A schematic structural diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the embodiments of the present application only serve the purpose of illustration and description and are not intended to limit the scope of protection of the embodiments of the present application. In addition, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the embodiments of the present application illustrate the operations implemented according to some embodiments of the embodiments of the present application. It should be understood that the operations of the flowcharts can be implemented out of sequence, and steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art, guided by the contents of the embodiments of the present application, can add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0028] In addition, the described embodiments are only a portion of the embodiments of the present application, rather than all embodiments. The components of the embodiments of the present application generally described and shown in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed embodiments of the present application, but rather merely represents selected embodiments of the embodiments of the present application.

[0029] It is understandable that the "first" and "second" in the embodiments of the present application are used to distinguish similar objects. Those skilled in the art will understand that words such as "first" and "second" do not limit the quantity and execution order, and words such as "first" and "second" do not necessarily limit differences. In the description of the embodiments of the present application, the term "and / or" is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the previous and subsequent associated objects are in an "or" relationship. The term "multiple" refers to more than two (including two). Similarly, "multiple groups" refers to more than two (including two).

[0030] It should be noted that the disease risk prediction method provided in the embodiments of the present application can be executed by an electronic device. The electronic device here refers to a device terminal capable of executing a computer program or the aforementioned server. Examples of device terminals include smartphones, personal computers, tablet computers, personal digital assistants, or mobile internet devices. A server refers to a device that provides computing services over a network. Examples of servers include x86 servers and non-x86 servers. Non-x86 servers include mainframes, minicomputers, and UNIX servers.

[0031] Related technologies for predicting disease risk primarily rely on analyzing patient medical information using structured data (such as laboratory indicators) based on a physician's personal clinical experience and knowledge, along with general natural language processing (NLP) technology. These technologies then predict disease risk based on this information. However, this NLP technology, based on a physician's personal clinical experience and knowledge, typically relies on static dictionaries, which are difficult to adapt to the terminology of different medical institutions (such as hospital-defined terminology and dialect expressions). Consequently, current disease risk prediction methods have low accuracy.

[0032] For the above questions, please see Figure 1 The following is a flow chart of a disease risk prediction method provided by an embodiment of the present application; the main idea of ​​this disease risk prediction method is to predict disease risk by combining entity extraction from a standard dictionary set with the contextual understanding and reasoning capabilities of a pre-trained language model, thereby enhancing the model's ability to distinguish similar diseases and effectively improving the accuracy of identifying disease risk types. Implementations of the above-mentioned disease risk prediction method may include:

[0033] Step S110: performing an initial prediction of the disease type of the suspected patient based on the medical information of the suspected patient, and obtaining the initial prediction type and a confidence score of the initial prediction type.

[0034] Suspected patients refer to suspected individual patients whose disease risks need to be predicted. They are usually individuals who have not yet been diagnosed but have clinical manifestations of a certain disease or high-risk factors. In order to predict the disease risk information of the suspected patient, it is necessary to analyze and process the medical information of the suspected patient.

[0035] Medical information is structured data and / or unstructured data generated when a suspected patient visits a medical institution for treatment. It may include the medical information generated by the suspected patient during the current visit and the medical information generated in the past. Specifically, it may include at least one of: clinical symptom information, medical record information, examination information, testing information, genetic testing information, physical examination information, department information, health service personnel information, imaging data information, medication record information, etc.

[0036] The initial prediction type is a preliminary assessment of the suspected patient's possible disease type based on the preliminary analysis of the medical information. The purpose of this initial prediction type is to determine the standard set of fields that should be extracted from the medical information. It is understood that when the initial prediction type is obtained, a confidence score for the initial prediction type is also obtained.

[0037] Step S120: performing entity extraction on the medical information based on the initial prediction type and the corresponding standard dictionary set and the confidence score of the initial prediction type to obtain a plurality of entity words.

[0038] Entity extraction is the process of identifying and extracting standardized terms or key medical entity words related to disease types from medical records. These entity words are specific medical terms or keywords (such as symptoms, laboratory parameters, and test results) extracted from medical records. These terms have clear medical meanings and can serve as the basis for determining disease types.

[0039] Step S130: Inferring the medical information based on multiple entity words through a pre-trained language model to obtain the disease risk information of the suspected patient, the disease risk information including: disease risk type and disease risk score.

[0040] A pretrained language model is a pretrained large artificial intelligence language model, such as a deep learning language model trained on large-scale medical text. The language model can understand the semantics of complex medical texts and perform reasoning based on context.

[0041] Disease risk information is a comprehensive assessment of the likelihood and severity of a suspected patient suffering from a certain disease, typically including a disease risk type and a disease risk score, where the disease risk score can range from 0 to 10. Of course, in actual practice, the aforementioned disease risk information may also include disease risk prompt information derived from inference using a pre-trained language model, such as follow-up examination recommendations, testing recommendations, treatment recommendations, expert consultation recommendations, and other prompt information for suspected patients. Examples of such follow-up examination recommendations and testing recommendations include "missing neostigmine test results, repeated electrical stimulation results, and myasthenia gravis-related serum antibody test results."

[0042] After obtaining the disease risk information of the suspected patient, the disease risk information of the suspected patient can also be stored in a training dataset. Then, the training dataset can be used to fine-tune the pre-trained language model. Specifically, fine-tuning can be used. For example, after desensitizing the disease risk information of the suspected patient (such as name, ID number, and mobile phone number, etc.), the desensitized information can be used to fine-tune the pre-trained language model, so that the pre-trained language model can more accurately identify key factors related to the disease risk type. In the process of fine-tuning the instructions of the pre-trained language model, it is also possible to fine-tune the instructions by combining a low learning rate and a small batch. For example, using a learning rate of LR=5e-6 can alleviate the situation of knowledge forgetting during the pre-training process, and using 50 data for a small batch operation. The small batch fine-tuning training combined with manually corrected high-quality data can effectively improve the situation of catastrophic forgetting.

[0043] In the implementation of the above solution, disease risk prediction is performed by combining the entity extraction of the standard dictionary set with the reasoning ability of the pre-trained language model. The standard dictionary set here ensures the standardization of entity extraction and avoids prediction bias caused by terminology ambiguity or expression differences. Therefore, it can more accurately capture the key medical features in medical information. Furthermore, the pre-trained language model, with its context understanding ability, not only reduces the burden of the pre-trained language model in processing redundant data, but also can capture subtle clues in complex cases, so as to explore the potential correlation between entity words, thereby enhancing the model's ability to distinguish similar diseases, and ultimately generating more reliable disease risk types and disease risk scores, effectively improving the recognition accuracy of disease risk types.

[0044] As an optional implementation of the above step S110, the implementation of the above-mentioned initial prediction of the disease type of the suspected patient based on the medical information of the suspected patient may include:

[0045] Step S111: Split the suspected patient's medical information into multiple structured fields.

[0046] The implementation of step S111 includes determining whether the suspected patient's medical information contains non-text content. If so, the non-text content may be converted to text. For example, if the medical information contains a computed tomography (CT) scan report, a text recognition neural network model or an image parsing agent can be used to identify text content such as the CT scan text from the CT scan report. Alternatively, a file extraction agent can be used to extract text content such as the file content from a PDF or Word file.

[0047] Then, the text content in the medical information is split according to the preset standardization rules to obtain multiple structured fields. The multiple structured fields here include: chief complaint content, current medical history, laboratory tests, imaging results, etc. For example, suppose the text content in the medical information is "More than 20 days ago, a patient developed ptosis of both upper eyelids without obvious cause, accompanied by dizziness, blurred vision, photophobia, slurred pronunciation, difficulty swallowing, a choking feeling but no choking, hoarseness, occasional dizziness, sudden and temporary loss of vision and then recovery, no headache, syncope, loss of consciousness, mild in the morning and severe in the evening, no limb weakness, no chest tightness, chest pain, precordial discomfort, no abdominal distension, abdominal pain, nausea and vomiting, no chills, fever, cough and sputum, and was treated in an external hospital. After treatment, the patient improved. In order to further clarify the diagnosis, the patient is now treated in our outpatient department and admitted to our department. Since the onset of the disease, the patient has been in poor spirits, has good sleep and diet, normal urination and defecation, and no change in weight." Data can be split based on preset symbols (such as commas or semicolons), semantic content units, or medical semantic structures. The specific splitting method can be determined based on preset standardization rules. After splitting medical information into structured fields, data standardization significantly reduces the complexity of unstructured text. Since structured fields can be directly parsed and processed by prediction tools, data cleaning and conversion time is reduced, thereby improving the efficiency of the overall prediction process.

[0048] Step S112: For each structured field in the plurality of structured fields, the preset disease diagnosis content is deleted from the structured field to obtain a plurality of processed fields.

[0049] For example, in the implementation of step S112 above, it has been discovered in practice that when filling out clinical symptom information or medical record information, doctors may include relevant content related to their own disease diagnoses, which may be inaccurate. Therefore, for each of the multiple structured fields, a preset rule (such as using regular expression rule matching) can be used to remove the preset disease diagnosis content (such as "suspected myasthenia gravis") from the structured field to prevent this disease diagnosis content from affecting the inference results of the pre-trained language model. Ultimately, the multiple processed fields obtained are all fields without disease diagnosis content, thereby avoiding interference with the predictions of the pre-trained language model and ensuring that the pre-trained language model is analyzed based on objective data. It can be understood that the multiple processed fields above remove the preset disease diagnosis content and retain more universal medical features (such as symptoms, examination indicators, etc.), enabling the prediction model to learn based on a wider data distribution. This process of removing the preset disease diagnosis bias enables the model to identify more potential disease patterns, improving its adaptability and generalization ability for unknown cases.

[0050] Step S113: using a disease type prediction tool to make an initial prediction of the disease type of the suspected patient based on the processed multiple fields, and obtaining an initial prediction type and a confidence score of the initial prediction type.

[0051] An implementation method of the above-mentioned step S113 is, for example: the disease type prediction tool of the general agent intelligent body can be used to make an initial prediction of the disease type of the suspected patient based on the processed multiple fields, and obtain the initial prediction type and the confidence score of the initial prediction type. Alternatively, the processed multiple fields are filled into the prompt word template for disease type prediction to obtain the prompt engineering text, and then the prompt engineering text is input into the trained disease type prediction model to obtain the initial prediction type and the confidence score of the initial prediction type output by the disease type prediction model. Among them, the disease type prediction tool of the above-mentioned general agent intelligent body can be completed by calling a pre-trained language large model (such as a model based on the Transformer architecture), or it can be developed through other machine learning models specifically for disease classification (such as random forests, support vector machines, etc.).

[0052] Optionally, as an optional implementation of the above step S120, the implementation of performing entity extraction on the medical information according to the initial prediction type and the corresponding standard dictionary set and the confidence score of the initial prediction type may include:

[0053] Step S121: Determine whether the confidence score of the initial prediction type is greater than a preset threshold score.

[0054] It can be understood that in the process of the disease type prediction tool and / or disease type prediction model of the general agent intelligent body predicting the initial prediction type, the above-mentioned disease type prediction tool and / or disease type prediction model can also simultaneously output the confidence level corresponding to the initial prediction type, that is, the confidence score of the initial prediction type.

[0055] Step S122: If the confidence score of the initial prediction type is greater than a preset threshold score, a standard dictionary set corresponding to the initial prediction type is searched from a dictionary database, and entity extraction is performed on the medical information according to the standard dictionary set to obtain the multiple entity words. The dictionary database stores standard dictionary sets corresponding to multiple disease types.

[0056] For example, the implementation method of the above-mentioned step S122 is as follows: if the confidence score corresponding to the initial prediction type is greater than a preset threshold score (for example, 9.5), then the standard dictionary set corresponding to the initial prediction type can be searched from the dictionary database, and entity extraction is performed on the medical information according to the standard dictionary set, that is, multiple entity words matching the standard dictionary set are extracted, thereby obtaining the multiple entity words. The dictionary database stores standard dictionary sets corresponding to multiple disease types. Among them, the above-mentioned dictionary database is a database that stores standardized term sets associated with disease types, which is used to assist in entity extraction and term normalization. The dictionary database stores standard dictionary sets corresponding to multiple disease types (such as myasthenia gravis or cardiac amyloidosis, etc.). The standard dictionary set is a term set defined for a specific disease type, which is used to identify key entity words related to the disease from the medical information. The standard dictionary set may include: clinical characteristics, laboratory, etc. It is understandable that since the standard dictionary set can avoid interference from ambiguous words (such as synonyms and abbreviations) and ensure that the extraction results are consistent with the standard dictionary set, this standard dictionary set based on disease type matching is domain-specific and can directly filter out non-relevant terms, thereby effectively improving the accuracy and recall rate of entity word extraction.

[0057] Optionally, as an optional implementation of the above step S120, after determining whether the confidence score of the initial prediction type is greater than a preset threshold score, the method further includes:

[0058] Step S122: If the confidence score of the initial prediction type is less than or equal to the preset threshold score, the trained entity extraction model is used to extract entities from the medical information to obtain multiple candidate entities, and the multiple candidate entities are screened to obtain multiple entity words.

[0059] An implementation example of the above-mentioned step S122 is: if the confidence score corresponding to the initial prediction type is less than or equal to a preset threshold score (for example, 9.5), the trained entity extraction model can be used to perform entity extraction on the medical information to obtain multiple candidate entities, and the Dirichlet Allocation (LDA) topic model or the Term Frequency-Inverse Document Frequency (TF-IDF) named entity recognition (NER) model can be used to screen the multiple candidate entities to reduce the entity fields while screening out more important entity fields, thereby obtaining the above-mentioned multiple entity words.

[0060] Optionally, if the confidence score of the initial prediction type is less than or equal to a preset threshold score, the default disease type can also be determined as the above-mentioned initial prediction type. Specifically, the initial prediction type is myasthenia gravis as an example for detailed description. The standard dictionary set corresponding to myasthenia gravis can be searched from the dictionary database. The standard field set here may include: clinical symptom information field, medical record information field, examination information field, test information field, genetic testing information field, physical examination information field, department information field, health service personnel information field, imaging data information field, and medication record information field. The standard dictionary set is dynamically combined with the initial prediction results. Since the predefined rules of the dictionary database are strongly associated with the disease type, the blind traversal of the global term library in the traditional method is avoided, thereby narrowing the search scope of entity extraction and reducing redundant calculations.

[0061] As an optional implementation of the above step S130, the above implementation of reasoning about the medical information based on multiple entity words using the pre-trained language model may include:

[0062] Step S131: searching for knowledge blocks in a vector database based on text sentences generated from a plurality of entity words. The vector database stores knowledge blocks for determining a plurality of disease types.

[0063] A vector database is a database specifically designed to store and query vector disease data. By converting disease text and pathological medical terminology into high-dimensional vectors, it enables rapid retrieval and similarity matching. In medical scenarios, vector databases are used to store knowledge blocks related to disease types.

[0064] Knowledge blocks refer to structured medical knowledge fragments (or semantic blocks) pre-stored in a vector database, covering the diagnostic criteria, typical manifestations, differential diagnosis rules, etc. of specific diseases. Each knowledge block corresponds to a disease type and is efficiently retrieved through vectorization technology.

[0065] It is understandable that the disease-related knowledge blocks stored in the above-mentioned vector database can be dynamically matched through vector retrieval of entity words, rather than static rules or fixed model parameters. This mechanism enables the system to flexibly call the most relevant domain knowledge based on the input entity words, avoiding the lack of adaptability of traditional methods due to knowledge solidification, and improving the model's generalization ability for changing clinical scenarios.

[0066] Step S132: The pre-trained language model is used to infer the medical information based on the knowledge blocks and multiple entity words to obtain the disease risk information of the suspected patient.

[0067] Reasoning refers to the process of generating disease risk information by combining pre-trained language models with entity words and medical guidelines and logical rules in knowledge blocks. This process includes symptom matching, laboratory indicator verification, and test result analysis.

[0068] As an optional implementation of the above step S131, the above implementation of searching for knowledge blocks in a vector database based on multiple entity words may include:

[0069] Step S131a: Search the knowledge graph for all entity node words associated with each entity word in the plurality of entity words and the attribute relationship of each entity node word.

[0070] An example of an implementation of the above-mentioned step S131a is to search for all entity node words and attribute relationships of each entity node word associated with each entity word in a pre-constructed knowledge graph (KG); specifically, it can be to search for all entity node words whose similarity with the entity word is higher than a preset threshold, where the preset threshold can be set according to the specific situation (e.g., 70% or 80%), or to search for all entity node words whose distance from the entity word is within a preset number of hops, where the preset number of hops can be set according to the specific situation (e.g., 1 hop or multiple hops). The knowledge graph stores a pre-constructed network topology diagram with medical knowledge entity words as nodes and the relationships between the nodes. The medical knowledge entity words can include metadata extracted from a medical knowledge standard library text, a medical PDF file, or a word file.

[0071] Step S131b: Generate a text sentence from all entity node words and the attribute relationship of each entity node word, and convert the text sentence into a sentence embedding vector.

[0072] The implementation method of the above-mentioned step S131b is, for example: all the above-mentioned entity node words and the attribute relationship of each entity node word can be filled into the preset text template to obtain the above-mentioned text sentence, or all the above-mentioned entity node words and the attribute relationship of each entity node word can be directly spliced ​​in a preset order to obtain the above-mentioned text sentence, and then, the text sentence is converted into a sentence embedding vector through a vectorization model such as the M3E (Moka Massive Mixed Embedding) model or the BGE (BAAI General Embedding) model.

[0073] Step S131c: Search the vector database for knowledge blocks whose similarity to the sentence embedding vector is greater than a preset similarity threshold.

[0074] An example of an implementation method of the above-mentioned step S131c is as follows: after obtaining all entity node words associated with each entity word and the text sentences generated by the attribute relationship of each entity node word, all text sentences can be converted into sentence embedding vectors first, and then, the knowledge blocks whose similarity with the sentence embedding vector is greater than a preset similarity threshold are searched in the vector database. These knowledge blocks are also called recalled knowledge blocks. It can be understood that the efficient similarity retrieval characteristics of the above-mentioned vector database make it unnecessary to traverse the entire amount of data in the search process of knowledge blocks, and can also quickly locate target information through the approximate nearest neighbor algorithm in the vector space. Combined with the joint reasoning of entity words and knowledge blocks by the pre-trained language model, the system can complete high-precision analysis while reducing redundant calculations, thereby significantly reducing the computational overhead and resource usage of the overall reasoning process, while ensuring retrieval accuracy, effectively reducing the consumption of computing resources and storage space, and improving the real-time response capability and scalability of the system.

[0075] As an optional implementation of step S132, the implementation of reasoning about medical information based on knowledge blocks and multiple entity words using a pre-trained language model may include:

[0076] Step S132a: Fill the knowledge blocks and multiple entity words into the preset prompt word template to obtain the prompt project text.

[0077] It's understandable that the combination of the aforementioned knowledge chunking and entity words provides the model with explicit anchors of domain knowledge, while the pre-set prompt word templates strengthen the relevance of these anchors to medical information through a semantic guidance mechanism. When processing this structured input, the multi-agent model can focus on different dimensions of medical logic (such as symptom associations, diagnostic rules, and treatment pathways) and integrate multi-perspective analysis results through collaborative reasoning. This hierarchical, decoupled processing approach enables the system to more efficiently capture the implicit logical relationships of complex diseases and adapt to nonlinear, multi-factorial clinical scenarios.

[0078] For example, in an optional implementation of step S132a, assuming the preset prompt word template is "You are a medical expert, skilled in disease risk assessment based on medical guidelines and clinical data. Based on the input medical entity words (including {entities}) and knowledge blocks (including {knowledge_block}), analyze the patient's possible disease type and output disease risk information." Then, the knowledge blocks can be filled into the "{knowledge_block}" placeholder in the preset prompt word template, and the multiple entity words can be filled into the "{entities}" placeholder in the preset prompt word template, thereby ultimately obtaining the prompt engineering text. The design of the preset prompt word template transforms complex medical reasoning tasks into a reusable structured input pattern, eliminating the need to train separate models for specific disease types. The pre-storage of knowledge blocks and the dynamic injection of entity words further reduce the reliance on full medical data during real-time reasoning, reducing computing resource consumption. Furthermore, the multi-agent model achieves parallel processing through division of labor and collaboration, which not only improves reasoning efficiency but also avoids the performance bottleneck of a single model when facing high-dimensional input, thereby optimizing system scalability without increasing hardware costs.

[0079] Step S132b: Reasoning the medical information based on the prompt engineering text through the multi-intelligent agent model.

[0080] As an optional implementation of step S132b, the multi-agent model may include: a management agent, an inspection and analysis agent, and a result summary agent. The implementation of the multi-agent model for reasoning about medical information based on the prompt engineering text may include:

[0081] Step S132c: The management agent intelligent body extracts the inspection project tasks and the test project tasks in the medical information according to the prompt engineering text, and assigns the inspection project tasks and the test project tasks to the inspection and analysis agent intelligent body.

[0082] In an alternative embodiment of step S132c, for example, the aforementioned management agent, also known as the master agent, is primarily responsible for managing and assigning tasks to other agents. Thus, the management agent can extract examination and test items from the medical information (e.g., multi-time period data results for myocardial amyloidosis) based on the prompt engineering text and assign these examination and test items to the inspection and analysis agent.

[0083] Step S132d: The inspection and analysis agent analyzes the inspection information and test information of the medical information according to the prompt engineering text to obtain multiple analysis results.

[0084] For example, in step S132d, the inspection and analysis agent can act as an expert agent, analyzing the examination and test information of the medical information after understanding the prompting engineering text, to obtain multiple analysis results. The inspection and analysis agent independently analyzes the examination and test information. Its parallel processing capabilities enable separate modeling and analysis of medical data of different dimensions (such as imaging, laboratory parameters, and medical history). This divide-and-conquer strategy avoids the global modeling limitations of a single model for complex data, thereby enhancing the ability to deeply mine multi-source, heterogeneous medical information and increasing the comprehensiveness and detailed coverage of the analysis results.

[0085] Step S132e: The result summary agent is used to summarize and infer the multiple analysis results to obtain the disease risk information of the suspected patient.

[0086] An example of an implementation method of the above-mentioned step S132e is: the above-mentioned result summary agent intelligent agent can perform reasoning operations such as summarizing and drawing conclusions on the multiple analysis results output by the inspection and analysis agent intelligent agent, and finally obtain the disease risk information of the suspected patient. The above-mentioned result summary agent intelligent agent can effectively eliminate the local bias or misjudgment that may exist in a single intelligent agent by aggregating multiple analysis results and performing cross-validation and conclusion reasoning. This integration logic is based on the collaborative verification of multi-perspective analysis results rather than simple stacking, thereby enhancing the robustness and credibility of the final disease risk information. The above-mentioned mechanism significantly reduces the risk of conclusion distortion due to single-point analysis errors and improves the reliability of the overall system.

[0087] Optionally, after obtaining the disease risk information of the suspected patient, the disease risk type and disease risk score can also be parsed from the disease risk information, and then it is determined whether the disease risk score of the suspected patient is greater than a preset high-risk threshold. If the disease risk score of the suspected patient is greater than the preset high-risk threshold, risk warning information is generated based on the disease risk type and disease risk score in the disease risk information, and the risk warning information is sent to the doctor terminal device corresponding to the suspected patient. The risk warning information may include disease risk type, disease risk score, follow-up examination suggestions, test suggestions, treatment suggestions, expert consultation suggestions, etc. Similarly, if the disease risk score of the suspected patient is less than the preset high-risk threshold, the possibility that the suspected patient has the disease risk type can be ruled out. At this time, risk exclusion information can also be generated based on the disease risk type and disease risk score in the disease risk information, and then the risk exclusion information is sent to the doctor terminal device corresponding to the suspected patient.

[0088] See Figure 2 FIG2 is a schematic diagram showing the structure of a disease risk prediction device provided in an embodiment of the present application; an embodiment of the present application provides a disease risk prediction device 200, comprising:

[0089] The disease type initial prediction module 210 is used to make an initial prediction of the disease type of the suspected patient based on the medical information of the suspected patient, and obtain the initial prediction type and the confidence score of the initial prediction type.

[0090] The medical information extraction module 220 is used to perform entity extraction on the medical information based on the initial prediction type, which determines the corresponding standard dictionary set and the confidence score of the initial prediction type, to obtain multiple entity words.

[0091] The disease risk information acquisition module 230 is used to infer medical information based on multiple entity words using a pre-trained language model to obtain disease risk information of suspected patients. The disease risk information includes: disease risk type and disease risk score.

[0092] As an optional embodiment of the above device, the disease type initial prediction module includes:

[0093] The medical information splitting submodule is used to split the medical information of the visit into multiple structured fields.

[0094] The disease diagnosis deletion submodule is used to delete preset disease diagnosis content from each structured field in the multiple structured fields to obtain multiple processed fields.

[0095] The prediction type acquisition submodule is used to make an initial prediction of the disease type of the suspected patient based on the processed multiple fields through the disease type prediction tool to obtain the initial prediction type.

[0096] As an optional embodiment of the above device, the medical information extraction module includes:

[0097] The confidence score judgment submodule is used to judge whether the confidence score of the initial prediction type is greater than a preset threshold score.

[0098] The information entity extraction submodule is used to search the standard dictionary set corresponding to the initial prediction type from the dictionary database if the confidence score of the initial prediction type is greater than the preset threshold score, and perform entity extraction on the medical information based on the standard dictionary set to obtain multiple entity words. The dictionary database stores standard dictionary sets corresponding to multiple disease types.

[0099] As an optional embodiment of the above device, the medical information extraction module further includes:

[0100] The entity extraction and screening submodule is used to use the trained entity extraction model to perform entity extraction on the medical information if the confidence score of the initial prediction type is less than or equal to the preset threshold score, to obtain multiple candidate entities, and to screen the multiple candidate entities to obtain the multiple entity words.

[0101] As an optional embodiment of the above device, the disease risk information acquisition module includes:

[0102] The knowledge block search submodule is used to search for knowledge blocks in a vector database based on multiple entity words, and the vector database stores knowledge blocks for determining multiple disease types.

[0103] The medical information reasoning submodule is used to infer medical information based on knowledge blocks and multiple entity words through a pre-trained language model to obtain disease risk information of suspected patients.

[0104] As an optional implementation of the above device, the knowledge block search submodule includes:

[0105] The entity word search unit is used to search for all entity node words associated with each entity word in multiple entity words and the attribute relationship of each entity node word in the knowledge graph.

[0106] The embedding vector conversion unit is used to generate a text sentence from all entity node words and the attribute relationship of each entity node word, and convert the text sentence into a sentence embedding vector.

[0107] The vector similarity comparison unit is used to search the vector database for knowledge blocks whose similarity with the sentence embedding vector is greater than a preset similarity threshold.

[0108] As an optional embodiment of the above device, the medical information reasoning submodule includes:

[0109] The prompt text obtaining unit is used to fill the knowledge blocks and multiple entity words into the preset prompt word template to obtain the prompt project text.

[0110] The medical information reasoning unit is used to reason about the medical information of the patient based on the prompt engineering text through a multi-intelligent agent model.

[0111] As an optional embodiment of the above-mentioned device, the multi-agent model includes: a management agent, an inspection and analysis agent, and a result summary agent; the medical information reasoning unit includes:

[0112] The task extraction and assignment subunit is used to extract the inspection project tasks and test project tasks from the medical information according to the prompt engineering text through the management agent intelligent body, and assign the inspection project tasks and test project tasks to the inspection and analysis agent intelligent body.

[0113] The analysis result obtaining subunit is used to analyze the inspection information and test information of the medical information according to the prompt engineering text through the inspection and analysis agent intelligent body to obtain multiple analysis results.

[0114] The result summary reasoning subunit is used to summarize and reason about multiple analysis results through the result summary agent to obtain disease risk information of suspected patients.

[0115] It should be understood that this device corresponds to the aforementioned disease risk prediction method embodiment and is capable of executing each of the steps involved in the aforementioned method embodiment. The specific functions of this device can be found in the description above, and a detailed description is omitted here. The device includes at least one software functional module that can be stored in a memory in the form of software or firmware or embedded in the device's operating system (OS).

[0116] See Figure 3 The electronic device 300 provided in the embodiment of the present application includes a processor 310 and a memory 320, wherein the memory 320 stores machine-readable instructions executable by the processor 310, and when the machine-readable instructions are executed by the processor 310, the method described above is performed.

[0117] The embodiment of the present application further provides a computer-readable storage medium 330, on which a computer program is stored, and the computer program executes the above method when executed by the processor 310. The computer-readable storage medium 330 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0118] An embodiment of the present application further provides a computer program product, including: a computer program or computer instructions, which executes the method described above when the computer program or computer instructions are executed by a processor.

[0119] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similarities between the various embodiments can be referred to in conjunction with each other. For device embodiments, since they are generally similar to method embodiments, their description is relatively simple, and for relevant details, reference can be made to the description of the method embodiments.

[0120] In the several embodiments provided in the embodiments of the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are only schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible implementation architectures, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and a module, a program segment or a part of a code contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also be different from the order of occurrence marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, which is mainly based on the functions involved.

[0121] In addition, the functional modules of each embodiment in the embodiments of the present application can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part. In addition, in the description of this specification, the descriptions with reference to the terms "one embodiment", "some embodiments", "example", "specific example", "some examples", etc. mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiments of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in an appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0122] The above description is only an optional implementation method of the embodiment of the present application, but the protection scope of the embodiment of the present application is not limited to this. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in the embodiment of the present application, and they should all be covered by the protection scope of the embodiment of the present application.

Claims

1. A disease risk prediction method, characterized in that: include: Performing an initial prediction of the disease type of the suspected patient based on the medical information of the suspected patient, and obtaining the initial prediction type and a confidence score of the initial prediction type; Determining a corresponding standard dictionary set according to the initial prediction type and a confidence score of the initial prediction type, performing entity extraction on the medical information to obtain a plurality of entity words; Inferring the medical information based on the multiple entity words using a pre-trained language model to obtain disease risk information of the suspected patient, wherein the disease risk information includes: disease risk type and disease risk score; Among them, the initial prediction of the disease type of the suspected patient based on the medical information of the suspected patient includes: splitting the medical information into multiple structured fields; for each structured field in the multiple structured fields, deleting the preset disease diagnosis content from the structured field to obtain multiple processed fields; using a disease type prediction tool of a general agent intelligent body to perform an initial prediction of the disease type of the suspected patient based on the multiple processed fields, to obtain the initial prediction type and the confidence score of the initial prediction type, and the disease type prediction tool of the general agent intelligent body is used to call a model of a Transformer architecture to complete the initial prediction, or call a machine learning model for disease classification to complete the initial prediction.

2. The disease risk prediction method according to claim 1, characterized in that: The step of performing entity extraction on the medical information based on the determination of a corresponding standard dictionary set and a confidence score of the initial prediction type according to the initial prediction type includes: Determining whether the confidence score of the initial prediction type is greater than a preset threshold score; If so, the standard dictionary set corresponding to the initial prediction type is searched from the dictionary database, and entity extraction is performed on the medical information according to the standard dictionary set to obtain the multiple entity words. The dictionary database stores standard dictionary sets corresponding to multiple disease types.

3. The disease risk prediction method according to claim 2, characterized in that: After determining whether the confidence score of the initial prediction type is greater than a preset threshold score, the method further includes: If the confidence score of the initial prediction type is less than or equal to the preset threshold score, the trained entity extraction model is used to perform entity extraction on the medical information to obtain multiple candidate entities, and the multiple candidate entities are screened to obtain the multiple entity words.

4. The disease risk prediction method according to claim 1, characterized in that: The reasoning of the medical information based on the multiple entity words using the pre-trained language model includes: searching a knowledge block in a vector database according to the plurality of entity words, wherein the vector database stores knowledge blocks for determining a plurality of disease types; The pre-trained language model is used to infer the medical information based on the knowledge blocks and the multiple entity words to obtain the disease risk information of the suspected patient.

5. The disease risk prediction method according to claim 4, characterized in that: The searching for knowledge blocks in a vector database according to the multiple entity words includes: Searching the knowledge graph for all entity node words associated with each entity word in the plurality of entity words and the attribute relationship of each entity node word; Generating a text sentence from all entity node words and the attribute relationship of each entity node word, and converting the text sentence into a sentence embedding vector; Search the vector database for a knowledge block whose similarity to the sentence embedding vector is greater than a preset similarity threshold.

6. The disease risk prediction method according to claim 4, characterized in that: The reasoning of the medical information based on the knowledge blocks and the multiple entity words by using a pre-trained language model includes: Filling the knowledge blocks and the multiple entity words into a preset prompt word template to obtain a prompt engineering text; The medical information of the consultation is inferred based on the prompt engineering text through a multi-intelligent agent model.

7. The disease risk prediction method according to claim 6, characterized in that: The multi-agent model includes: a management agent, an inspection and analysis agent, and a result summary agent; the multi-agent model is used to reason about the medical information according to the prompt engineering text, including: Extracting the inspection item tasks and the test item tasks from the medical information according to the prompt engineering text through the management agent intelligent body, and assigning the inspection item tasks and the test item tasks to the inspection and analysis agent intelligent body; The inspection and analysis agent analyzes the inspection information and test information of the medical information according to the prompt engineering text to obtain multiple analysis results; The result summarizing agent intelligent body summarizes the multiple analysis results and makes conclusion inferences to obtain the disease risk information of the suspected patient.

8. A disease risk prediction device, characterized in that: include: A disease type initial prediction module is used to make an initial prediction of the disease type of the suspected patient based on the medical information of the suspected patient, and obtain the initial prediction type and the confidence score of the initial prediction type; a medical information extraction module for performing entity extraction on the medical information according to a standard dictionary set corresponding to the initial prediction type and a confidence score of the initial prediction type to obtain a plurality of entity words; a disease risk information acquisition module, configured to infer the medical information based on the plurality of entity words using a pre-trained language model to obtain disease risk information of the suspected patient, wherein the disease risk information includes: disease risk type and disease risk score; Among them, the initial prediction of the disease type of the suspected patient based on the medical information of the suspected patient includes: splitting the medical information into multiple structured fields; for each structured field in the multiple structured fields, deleting the preset disease diagnosis content from the structured field to obtain multiple processed fields; using a disease type prediction tool of a general agent intelligent body to perform an initial prediction of the disease type of the suspected patient based on the multiple processed fields, to obtain the initial prediction type and the confidence score of the initial prediction type, and the disease type prediction tool of the general agent intelligent body is used to call a model of a Transformer architecture to complete the initial prediction, or call a machine learning model for disease classification to complete the initial prediction.

9. An electronic device, characterized in that: include: A processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions are executed by the processor to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is executed.

11. A computer program product, characterized in that include: A computer program or computer instruction, wherein the computer program or the computer instruction is executed by a processor to perform the method according to any one of claims 1 to 7.

Images