Child tumor management system and method based on multi-modal large model

Abstract

The invention discloses a children tumor management system and method based on a multi-modal large model, and relates to the technical field of children tumor management, and the system comprises a multi-modal data analysis module, a retrieval enhancement generation module, a patient end module and a doctor end module. The multi-modal data analysis module is used for processing clinical texts, medical images, pathological reports, gene sequencing, nutrition evaluation and psychological scales to generate structured data. And the retrieval enhancement generation module retrieves the first information fragment from the children tumor domain knowledge base according to the structured data, generates first traceable response content through a semantic rearrangement technology, and further generates visual information and text information related to the health condition of the patient and provides the visual information and the text information to the patient or family members. The multi-modal data are integrated to generate the structured information, artificial intelligence illusion is inhibited through traceable response, the health information is output in a popular form, the problems of information overload and patient management fragmentation are solved, and the diagnosis and treatment efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of pediatric cancer management technology, and in particular to a pediatric cancer management system and method based on a multimodal large model. Background Technology

[0002] Childhood malignant tumors are characterized by complex pathological subtyping, rapid progression, and narrow therapeutic windows. Furthermore, children in this stage of growth and development require extremely high precision in diagnosis and treatment, as well as excellent long-term survival. Current treatment models face the following challenges: scarcity and uneven distribution of professional resources; top pediatric oncology experts are concentrated in large cities, while primary healthcare institutions lack sufficient diagnostic capabilities, leading to delayed diagnosis and inappropriate referrals; information overload and complex decision-making; pediatric oncology involves heterogeneous data from multiple dimensions, including clinical, imaging, pathological, and genetic data, making it difficult for doctors to quickly integrate and formulate optimal treatment plans; fragmented patient management; disconnect between pre-diagnosis, treatment, long-term follow-up, and nutritional and psychological support, lacking integrated, full-cycle management tools.

[0003] In existing technologies, general-purpose large models or simple medical question-answering systems are being attempted to assist medical decision-making. These systems are typically based on natural language processing (NLP) techniques, capable of performing preliminary analysis and answering questions about text-based medical data. Some systems also attempt to integrate limited medical imaging data, providing supplementary diagnostic suggestions through image recognition technology. These existing technological solutions primarily provide knowledge-based question answering and simple decision support in specific medical scenarios through model fine-tuning or rule engines.

[0004] However, existing technologies have significant drawbacks. General-purpose large-scale models or simple medical question-answering systems suffer from the AI ​​illusion problem, outputting inaccurate or fabricated information, lacking professionalism, and unable to safely and effectively process multimodal medical data in the field of pediatric oncology. These systems lack in-depth optimization for the specific characteristics of pediatric oncology, making it difficult to achieve integrated analysis and fusion processing of multimodal data such as clinical texts, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales. Furthermore, existing systems cannot ensure the traceability of output and pose privacy risks when processing sensitive medical data, failing to meet medical-grade accuracy and security compliance requirements.

[0005] In view of the above-mentioned shortcomings in existing technologies, there is an urgent need for an intelligent system that can effectively process multimodal data, ensure output accuracy, and provide full-cycle management support to improve the efficiency and quality of pediatric tumor diagnosis and treatment. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of existing technologies, specifically by providing a pediatric tumor management system and method based on a multimodal large model, as detailed below: 1) In a first aspect, the present invention provides a pediatric tumor management system based on a multimodal large model, the specific technical solution of which is as follows: It includes a multimodal data parsing module, a retrieval enhancement generation module, and an information generation module; The multimodal data parsing module is used to process multimodal data from pediatric cancer patients and generate structured data, including clinical text, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales. The retrieval enhancement generation module is used to: retrieve relevant first information fragments from the pediatric oncology knowledge base based on structured data, and sort the first information fragments using semantic reordering technology to generate first traceable response content; The information generation module is used to: generate visual information and / or text information related to the health status of pediatric cancer patients based on the first traceable response content, and provide it to pediatric cancer patients or their families.

[0007] The beneficial effects of the pediatric tumor management system based on a multimodal large model provided by this invention are as follows: The multimodal data parsing module integrates multimodal data from pediatric oncology patients, including clinical texts, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales, generating unified structured data. This effectively integrates fragmented medical information and overcomes the problems of information overload and complex decision-making. The retrieval enhancement generation module, based on the structured data, retrieves relevant first-order information fragments from a pediatric oncology knowledge base and uses semantic reordering technology to finely sort these fragments, generating first-order traceable response content. This process significantly suppresses AI illusions, ensuring the accuracy and verifiability of the output suggestions. Based on the first-order traceable response content, the system generates visual and textual information related to the health status of pediatric oncology patients and provides it to the patients or their families in an easily understandable format. This solves the problem of fragmented patient management and achieves seamless integration of full-cycle management. Overall, this solution improves diagnostic efficiency and accuracy, compensates for the lack of professional resources in primary healthcare institutions, provides continuous and personalized support to families of sick children, and ensures medical-grade reliability through traceable output.

[0008] Based on the above scheme, the pediatric tumor management system based on a multimodal large model of the present invention can be further improved as follows.

[0009] Furthermore, it also includes a patient-side module; the patient-side module is used to send the interaction content with the target user to the search enhancement generation module, the target user being a pediatric cancer patient or a family member of a pediatric cancer patient; The retrieval enhancement generation module is also used to: retrieve relevant second information fragments from the pediatric oncology knowledge base based on structured data and interaction content with the target user, and sort the second information fragments using semantic reordering technology to generate second traceable response content; The patient-side module is also used to generate intelligent Q&A content, psychological screening and assessment information, peer care information, follow-up management suggestions, and nutritional guidance suggestions based on the second traceable response content, and provide them to the target users.

[0010] The beneficial effects of adopting the above-mentioned further solution are as follows: Currently, there are problems with insufficient doctor-patient communication and support in pediatric oncology diagnosis and treatment. Parents lack access to authoritative and easily understandable knowledge, children face psychological stress and lack age-appropriate psychological intervention methods, and the disconnect between various aspects of patient management leads to discontinuous support. This invention sends the interaction content with the target user to the retrieval enhancement generation module through the patient-side module. The retrieval enhancement generation module retrieves relevant second information fragments from the pediatric oncology knowledge base based on structured data and the interaction content with the target user, and uses semantic reordering technology to sort the second information fragments, generating second traceable response content. Based on the second traceable response content, the patient-side module generates intelligent question-and-answer content, psychological screening and assessment information, peer care information, follow-up management suggestions, and nutritional guidance suggestions, and provides them to the target user. This process effectively improves the accuracy and personalization of information delivery. The intelligent Q&A content answers parents' questions in plain language, the psychological screening and assessment information provides timely warnings of potential problems, the peer care information provides emotional support, and the follow-up management suggestions and nutritional guidance suggestions ensure the continuity of management throughout the entire cycle. This improves the medical experience, makes up for the shortcomings of the traditional model, and achieves comprehensive and reliable support for sick children and their families.

[0011] Furthermore, it also includes a doctor-side module; the doctor-side module is used to send the content of the interaction with the doctor to the retrieval enhancement generation module; The retrieval enhancement generation module is also specifically used to: retrieve relevant third-party information fragments from the pediatric oncology knowledge base based on structured data, interaction content with the target user, and interaction content with the doctor, and sort the third-party information fragments using semantic reordering technology to generate third-party traceable response content; The doctor-side module is also used to generate intelligent Q&A content, patient management overview, and follow-up record overview based on third-party traceable responses, and provide them to doctors.

[0012] The beneficial effects of adopting the above-mentioned further solution are as follows: Currently, pediatric oncology diagnosis and treatment faces the dilemma of uneven distribution of professional resources, insufficient diagnostic and treatment capabilities in primary healthcare institutions, and doctors need to process multi-dimensional heterogeneous information such as clinical texts, imaging reports, and genetic data, leading to low decision-making efficiency. This invention sends doctors' professional queries to a retrieval enhancement generation module through a doctor-side module. This module integrates structured data, patient interaction content, and doctor input, retrieving relevant third-party information fragments from a pediatric oncology knowledge base. After semantic reordering technology, it generates third-party traceable response content. Based on this content, the doctor-side module generates professional-grade intelligent question-and-answer content, a patient management overview integrating multimodal data, and a follow-up record overview including trend analysis. This mechanism provides doctors with evidence-based medicine support with sub-second response times, the patient management overview provides a multi-dimensional overview of the patient's condition, and the follow-up record overview automatically generates a visualized disease progression trajectory, effectively improving diagnostic accuracy and efficiency. Simultaneously, it assists in the downward flow of high-quality medical resources through standardized treatment pathways, enhancing the diagnostic and treatment capabilities of primary healthcare institutions.

[0013] Furthermore, the pediatric oncology knowledge base stores authoritative treatment guidelines, expert clinical experience, and desensitized case data in the field of pediatric oncology.

[0014] The beneficial effects of adopting the above-mentioned further approach are as follows: The pediatric oncology knowledge base constructs a comprehensive professional knowledge foundation by integrating authoritative treatment guidelines, expert clinical experience, and anonymized case data. Authoritative treatment guidelines ensure that the system's output conforms to the latest medical evidence, expert clinical experience provides practical clinical insights, and anonymized case data provides real-world clinical validation. This multi-dimensional knowledge fusion effectively enhances the professionalism and applicability of the output content of the retrieval enhancement module, providing a reliable basis for generating traceable response content and ensuring the accuracy and clinical applicability of treatment recommendations.

[0015] 2) Secondly, the present invention also provides a method for childhood tumor management based on a multimodal large model, the specific technical solution of which is as follows: Multimodal data from pediatric cancer patients are processed to generate structured data, which includes clinical texts, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales. Based on structured data, relevant first information fragments are retrieved from a knowledge base in the field of pediatric oncology, and semantic reordering technology is used to sort the first information fragments to generate the first traceable response content. Based on the first traceable response content, generate visual and / or textual information related to the health status of pediatric cancer patients and provide it to pediatric cancer patients or their families.

[0016] Based on the above scheme, the pediatric tumor management method based on a multimodal large model of the present invention can be further improved as follows.

[0017] Furthermore, it also includes: Based on structured data and interaction content with the target user, relevant second information fragments are retrieved from the pediatric oncology knowledge base, and semantic reordering technology is used to sort the second information fragments to generate second traceable response content, wherein the target user is a pediatric cancer patient or a family member of a pediatric cancer patient. Based on the second traceable response content, intelligent Q&A content, psychological screening and assessment information, peer care information, follow-up management suggestions, and nutritional guidance suggestions are generated and provided to the target users.

[0018] Furthermore, it also includes: Based on structured data, interaction content with target users, and interaction content with doctors, relevant third-party information fragments are retrieved from the pediatric oncology knowledge base, and semantic reordering technology is used to sort the third-party information fragments to generate third-party traceable response content. Based on the third traceable response, intelligent question-and-answer content, patient management overview, and follow-up record overview are generated and provided to doctors.

[0019] Furthermore, the pediatric oncology knowledge base stores authoritative treatment guidelines, expert clinical experience, and desensitized case data in the field of pediatric oncology.

[0020] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so that the electronic device implements any of the above-mentioned methods for managing childhood tumors based on a multimodal large model.

[0021] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned methods for managing childhood tumors based on a multimodal large model.

[0022] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below: Figure 1 This is a schematic diagram of the structure of a pediatric tumor management system based on a multimodal large model according to an embodiment of the present invention; Figure 2This is a flowchart illustrating a method for childhood tumor management based on a multimodal large model, according to an embodiment of the present invention. Detailed Implementation

[0024] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0025] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0026] like Figure 1 As shown, an embodiment of the present invention provides a pediatric tumor management system based on a multimodal large model, including a multimodal data parsing module, a retrieval enhancement generation module, and an information generation module; The multimodal data parsing module is used to process multimodal data from pediatric cancer patients and generate structured data, including clinical text, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales. Clinical texts are the written content of medical records created by healthcare professionals, including descriptions of medical history, symptoms, physical examination findings, and treatment records. These texts exist in natural language and contain rich medical knowledge, but lack a standardized format, requiring conversion into a machine-readable structure using natural language processing (NLP) technology. Clinical texts often involve technical terminology and contextual dependencies; after parsing, standardized data fields such as symptom codes and drug dosages can be generated, supporting automated report generation and decision support.

[0027] Medical imaging refers to visual data acquired through medical imaging equipment, such as X-ray images, computed tomography (CT) scans, or magnetic resonance imaging (MRI) images. These images are stored in the form of digital matrices, with each pixel representing a specific physical characteristic, such as tissue density or signal intensity. Medical imaging is used for non-invasive observation of internal human structures. In pediatric oncology, it is often used to detect changes in tumor location, size, and morphology. Image processing techniques can extract quantitative features, such as tumor volume and texture indicators, and convert them into numerical data for further analysis.

[0028] A pathology report is a diagnostic document prepared by a pathologist after microscopic examination of tissue or cell samples, describing cell morphology, tissue structure, and abnormal findings. The report typically includes a diagnostic conclusion, grading information, and auxiliary test results, presented in free text or semi-structured format. Pathology reports provide authoritative evidence of tumor type and malignancy; after parsing, they can generate structured data elements such as cell type codes and staging levels, facilitating integration into electronic health records and clinical support systems.

[0029] Gene sequencing, specifically, is the process of determining the nucleotide sequence in DNA or RNA molecules using high-throughput sequencing technology, generating raw data representing genetic information. Gene sequencing data is stored in sequence file formats, such as FASTA or BAM files, containing information on millions of base pairs. In pediatric oncology, gene sequencing is used to identify genetic variations and expression profiles. Through bioinformatics analysis, these variations can be structured into records, including gene identifiers, mutation types, and functional impacts, supporting precision medicine and targeted therapy selection.

[0030] Nutritional assessment is the process of determining a patient's nutritional status by collecting and evaluating nutrition-related data, including dietary intake records, anthropometric measurements such as weight and height, and biochemical values ​​such as serum protein levels. Nutritional assessment aims to identify the risk of nutritional deficiencies or excesses and quantifies the status by calculating scores such as a nutritional risk index. The assessment results are structured into parameters such as energy requirements and nutrient intake to develop personalized nutritional intervention plans.

[0031] Psychological scales are standardized psychological measurement tools, typically presented in questionnaire form, used to assess an individual's emotions, behaviors, and mental state. A psychological scale contains a series of items, each with predefined response options. Statistical methods are used to score the responses, generating quantitative indicators such as anxiety scores or depression indices. In pediatric cancer management, psychological scales help monitor patients' mental health; after analysis, they can generate structured psychological state data for early intervention and support.

[0032] The specific implementation process for generating structured data is as follows: 1) For clinical texts, the multimodal data parsing module initiates a natural language processing flow, including word segmentation, part-of-speech tagging, and syntactic analysis, to identify medical entities such as symptoms, medications, and diagnoses within the text. Entity recognition technology, based on a pre-trained medical language model, maps unstructured descriptions to a standard medical terminology database; for example, mapping headache to ICD-10 code R51. Simultaneously, a relation extraction algorithm extracts attributes such as time, dosage, and frequency from the text and converts this information into structured data fields, such as symptom name, time of occurrence, and severity level. This process ensures that implicit knowledge within the text is explicitly represented as computable data.

[0033] 2) For medical images, the multimodal data parsing module applies image preprocessing techniques, including noise reduction, normalization, and segmentation, to enhance image quality and isolate regions of interest. The feature extraction algorithm then calculates the morphological features of the tumor, such as volume, boundary sharpness, and texture parameters, which are expressed as numerical vectors using mathematical formulas. For example, the technical formula for tumor volume is: ,in, Indicates tumor volume, Indicates the first The area of ​​the tumor region in each slice, Indicates the slice spacing. This represents the total number of slices. These vectors are stored as structured attributes for easy subsequent analysis and visualization.

[0034] 3) For pathology reports, the multimodal data parsing module performs text parsing and semantic analysis to identify key diagnostic elements such as cell type, tissue grade, and tumor stage. Through a rule engine and machine learning model, the free text in the report is converted into a predefined structured format. For example, the descriptive statement "70% malignant cells" is encoded as a numerical field representing the proportion of malignant cells as 70, and associated with a standard classification system such as TNM staging. This ensures the machine readability and consistency of pathology information.

[0035] 4) For gene sequencing, the multimodal data parsing module utilizes a bioinformatics pipeline to process raw sequence files, including sequence alignment, variant detection, and functional annotation. The variant recall algorithm identifies single nucleotide polymorphisms and insertions / deletions, generating tabular data containing gene names, variant locations, and clinical significance. For example, a variant record might include fields such as gene name TP53, mutation type missense mutation, and clinical significance (potentially pathogenic). This data is integrated into a structured variant list to support personalized treatment decisions.

[0036] 5) For nutritional assessment, the multimodal data parsing module analyzes input data such as food diaries, weight measurements, and blood test results, and applies nutritional models to calculate scores and risk levels. For example, the formula for calculating body mass index is: ,in, Indicates body mass index, Weight is expressed in kilograms. The height is indicated in meters. The module outputs standardized parameters such as nutritional risk score and recommended energy intake, which are structured into data table fields for generating personalized recommendations.

[0037] 6) For psychological scales, the multimodal data parsing module uses statistical scoring algorithms to process questionnaire responses. For example, it performs weighted summation on the items of the Rutter Behavior Scale to generate a psychological state index. The formula for calculating the score of the psychological state index is: ,in, Indicates the total score. Indicates the first The weight of each item, symbol Represents the response value, symbol This indicates the total number of items. Results are stored as structured scores, such as anxiety scores and depression levels, facilitating trend analysis and intervention triggering.

[0038] All parsed data is integrated into a unified data model, designed based on a relational database or NoSQL system, defining data tables, field types, and relationships. For example, the patient information table might contain fields such as patient identifier, diagnosis date, and treatment stage, and be linked to the imaging feature table and gene variation table via foreign keys. Data validation rules ensure integrity, such as checking numerical ranges and required fields. Finally, the multimodal data parsing module outputs a comprehensive structured dataset, stored in standard formats such as JSON or SQL tables, supporting efficient querying and response generation by the retrieval enhancement module.

[0039] The retrieval enhancement generation module is used to: retrieve relevant first information fragments from a pediatric oncology knowledge base based on structured data, and sort these first information fragments using semantic reordering technology to generate first traceable response content. The pediatric oncology knowledge base stores authoritative treatment guidelines, expert clinical experience, and desensitized case data in the field of pediatric oncology. The specific implementation process is as follows: The module receives structured data from the multimodal data parsing module as input. This structured data includes standardized fields such as symptom codes, tumor volume, gene mutation types, nutritional scores, and mental state indicators. The retrieval enhancement generation module first converts the structured data into query vectors, mapping numerical and categorical fields to a high-dimensional vector space using an embedding model. For example, for the symptom field, the module uses a pre-trained word embedding model to convert the text description into a vector representation, while numerical fields such as tumor volume are directly normalized and used as vector components. The formula for constructing the query vector is: , Represents the query vector. Indicates the first The weight of each field, Indicates the first Embedding vectors of each field, Indicates the total number of fields. Weight The weight of a field is dynamically adjusted based on its importance in the field of pediatric oncology; for example, the gene mutation field may be given higher weight.

[0040] Then, the retrieval enhancement generation module uses the query vector to perform a similarity search in a pediatric oncology knowledge base. This knowledge base stores authoritative treatment guidelines, expert clinical experience, and anonymized case data, all pre-processed into vector indexes. The module calculates the cosine similarity between the query vector and the vector of each information fragment in the knowledge base, using the following formula: ,in, Indicates the first in the knowledge base Vector representation of each information fragment Represents the vector dot product. Denotes the norm of the query vector. This represents the norm of the information fragment vector. The module retrieves information fragments with a similarity higher than a preset threshold as the initial first set of information fragments, which may include text paragraphs, data records, or image descriptions.

[0041] Then, the retrieval enhancement generation module applies semantic reordering to fine-tune the initial set of first information fragments. Semantic reordering uses a transformer-based neural network model, fine-tuned on pediatric oncology data, to evaluate the semantic relevance of each information fragment to the query. The model input includes the query vector and the text content of each information fragment, and outputs a relevance score. The relevance score is calculated based on an attention mechanism, using the following formula: , Indicates the first The relevance score of each information fragment. Indicates the first A semantic representation vector of an information fragment. Represents the dimension of a vector. This represents the normalized exponential function. The module sorts the first information segments in descending order based on their scores, ensuring that the most relevant and authoritative segments are at the top.

[0042] Finally, the retrieval enhancement generation module generates the first traceable response content. This module combines the sorted first information fragments into a coherent text or data response, attaching a source identifier to each fragment, including the document ID and version number from the knowledge base. For example, the response content might include treatment recommendation text and reference specific sections of authoritative treatment guidelines. The module ensures standardized response formats, such as using a JSON structure to include fields for response text and source information, thereby supporting user verification and auditing. The entire process achieves end-to-end traceability, making the first traceable response content both accurate and reliable.

[0043] The first traceable response content refers to the initial output generated by the retrieval enhancement generation module based on structured data and a knowledge base in the field of pediatric oncology. This output is presented in text or data structure form and includes source citations for each information fragment. The first traceable response content ensures that all suggestions or data are verifiable, allowing users to trace back to the original authoritative sources in the knowledge base, thereby verifying the accuracy and reliability of the response. In the pediatric oncology management system, the first traceable response content is used to provide diagnostic support or treatment recommendations while meeting medical compliance requirements.

[0044] Authoritative treatment guidelines refer to standardized medical practice documents published by professional medical organizations or institutions. These documents, based on the latest scientific evidence and expert consensus, provide recommended protocols for disease diagnosis, treatment, and management. In the field of pediatric oncology, authoritative treatment guidelines include content such as tumor classification standards, chemotherapy protocols, and follow-up procedures. These guidelines undergo rigorous review and updates to ensure the scientific rigor and consistency of clinical decision-making. Within the pediatric oncology knowledge base, authoritative treatment guidelines serve as a trusted knowledge source, supporting the accuracy of the output generated by the retrieval enhancement module.

[0045] Expert clinical experience refers to the knowledge and insights accumulated by senior medical professionals through long-term practice, including case management skills, complication management methods, and personalized treatment strategies. In the field of pediatric oncology, expert clinical experience encompasses aspects such as the management of rare tumors or the response to adverse reactions. This experience is typically documented through academic publications, case reports, or training materials. Within the pediatric oncology knowledge base, expert clinical experience provides practical supplementation, enhancing the system's adaptability to complex scenarios.

[0046] Anonymized case data refers to anonymized datasets obtained by removing personally identifiable information from authentic medical records. These data retain clinically relevant fields such as diagnosis, treatment response, and prognostic indicators. In the field of pediatric oncology, anonymized case data includes information such as patient age, tumor type, and treatment outcomes. This data undergoes privacy processing to ensure compliance with data protection regulations. Within the pediatric oncology knowledge base, anonymized case data provides real-world case references and supports a search enhancement module that generates responses based on historical patterns.

[0047] The information generation module is used to: generate visual and / or textual information related to the health status of pediatric cancer patients based on the first traceable response content, and provide it to the pediatric cancer patients or their families. The specific implementation process is as follows: The system receives the first traceable response content output from the retrieval enhancement generation module. This content contains sorted information fragments and their traceability identifiers, such as treatment recommendations, monitoring indicators, or risk warnings. The system parses the first traceable response content and extracts key health data elements, such as tumor size, complete blood count values, nutritional scores, or mental health scores. This data is stored in a structured format, including numerical values, categorical fields, and timestamps.

[0048] For generating visualized information, the system uses a data visualization engine to process the extracted data elements. The visualization engine selects the appropriate chart type based on the data type and user scenario. For example, for time-series data such as tumor volume from multiple follow-ups, the engine generates a line chart to display the trend. The line chart is drawn based on a set of data points, each including a time value and a measurement value. The chart generation process uses mathematical formulas to calculate coordinate positions; for example, in a two-dimensional coordinate system, the horizontal coordinate of a point is mapped from the time value, and the vertical coordinate is mapped from the normalized measurement value. The normalization formula is: , This represents the normalized vertical coordinate value. Represents the original measurement value. This represents the minimum measurement value in the dataset. This represents the maximum measurement value in the dataset. The engine uses visualization libraries such as D3.js or Chart.js to render charts and adds labels, legends, and color coding to enhance readability. For example, red may indicate a high-risk area, and green may indicate a safe range. Visualizations also include interactive elements, such as hover tips that display detailed values, ensuring that patients or their families can intuitively understand changes in their health status.

[0049] For text information generation, the system applies natural language generation (NLP) technology to convert the first traceable response content into easily understandable text descriptions. The NLP module, based on templates and a rule engine, combines structured data from the first traceable response content to generate personalized text. For example, if the first traceable response content includes a nutritional risk score and recommended intake, the module will populate a predefined text template, replacing variables with actual data. The text generation process uses conditional logic to ensure the language style is suitable for non-specialist users, such as avoiding medical terminology and using everyday language like "children need to eat more protein-rich foods." The module also integrates a sentiment analysis component to adjust the tone to provide supportive messages, such as encouraging statements or comforting tips. The text information is organized in paragraphs or lists and displayed through the patient-side interface, ensuring clarity and ease of follow.

[0050] The generated visual and textual information is integrated into a unified output and provided to pediatric cancer patients or their families through a patient-side module. The output format is based on web standards such as HTML and CSS, ensuring cross-device compatibility. For example, visual information is embedded as SVG graphics, and text information is displayed as a scrollable text area. The system also includes access control mechanisms, allowing only authorized users to view information, and logs access history for auditing purposes. The entire process ensures accurate, timely, and user-friendly information, supporting pediatric cancer patients or their families in managing their health in a home environment.

[0051] Visualized information related to the health status of pediatric cancer patients refers to health data presented in the form of graphs, charts, or images. This data originates from structured information in the first traceable response content, such as tumor markers, physiological parameters, or psychological state scores. Visualized information uses visual elements such as line graphs, bar charts, or heatmaps to display data trends, comparisons, or distributions, helping users intuitively understand complex health information. In pediatric cancer management systems, visualized information is used to display disease progression, treatment response, or risk levels, enabling patients or their families to quickly identify key changes and take appropriate action.

[0052] Textual information related to the health status of pediatric cancer patients refers to health conditions, advice, or guidance described in written language. This content is based on authoritative information snippets from the first traceable response and translated into easily understandable expressions. Textual information includes care instructions, dietary recommendations, psychological support messages, or daily reminders, using non-technical language to ensure accessibility. Within the pediatric cancer management system, textual information provides personalized support, helping patients or their families adhere to medical plans and improve health behaviors.

[0053] Optionally, the above technical solution also includes a patient-side module; the patient-side module is used to: send the interaction content with the target user to the search enhancement generation module, where the target user is a pediatric cancer patient or a family member of a pediatric cancer patient; The retrieval enhancement generation module is also used to: retrieve relevant second information fragments from the pediatric oncology knowledge base based on structured data and interaction content with the target user, and sort the second information fragments using semantic reordering technology to generate second traceable response content; the specific implementation process is as follows: The system performs semantic parsing and feature extraction on the interaction content with the target user. This step uses natural language understanding technology to convert the raw text input by the user into a structured intent representation. For example, when a user enters "What should I do if my child vomits after chemotherapy?", the system identifies the key entities "chemotherapy" and "vomiting" through named entity recognition and determines through intent classification that this is a request for care guidance. The system constructs an interaction feature vector to represent the semantic content of this interaction, and the formula for constructing this vector is: , Represents the interaction feature vector. This represents the original text content entered by the user. This refers to a neural network encoder specifically designed for encoding interactive content.

[0054] Then, the vector representation of the structured data is fused with the interaction feature vector to generate an enhanced query representation. The fusion process uses a weighted concatenation method, as shown in the formula: ,in, This represents the enhanced query vector. This represents a vector representation of the structured data from the multimodal data parsing module. Represents the interaction feature vector. and These are the weighting coefficients for both, and these coefficients are adaptively adjusted according to the query type. For example, for symptom consultation queries, the weighting coefficient for interactive content is... It will be set to a higher level.

[0055] The retrieval enhancement generation module uses enhanced query vectors to perform similarity searches within a pediatric oncology knowledge base. The module calculates the cosine similarity between the enhanced query vector and each information fragment vector in the knowledge base, selecting information fragments with similarity exceeding a threshold to form an initial second set of information fragments. The similarity calculation uses the following formula: , Indicates the first in the knowledge base Vector representation of each information fragment.

[0056] Then, the system applies semantic reordering technology to finely sort the initial set of second information fragments. Semantic reordering uses an attention-based ranking model that simultaneously considers the enhanced query vector and the semantic content of each information fragment, calculating a relevance score for each fragment. The score calculation formula is as follows: , Indicates the first The final relevance score of each information fragment, This represents the attention calculation function. This represents a multilayer perceptron used to map attention outputs to scalar scores. The system sorts the second information segments in descending order based on the scores, ensuring that the information segments that best match the user's needs and the current context are placed first.

[0057] Finally, the retrieval enhancement generation module generates a second traceable response. This module synthesizes the sorted second information fragments into a coherent natural language response, while attaching detailed source information to each fragment, including the source document identifier, version number, and location reference. For example, the response might include specific steps of the nursing recommendations, clearly indicating which chapter of the authoritative treatment guideline these recommendations originate from. The entire generation process ensures the accuracy and verifiability of the response, and the resulting second traceable response content will be presented to the target user through the patient-side module.

[0058] The interaction content with the target user refers to various forms of information exchanged through the system interface by pediatric cancer patients or their families, including but not limited to text questions, option selection, voice input, or form completion. This interaction content typically expresses the user's questions, needs, or status updates in the form of natural language or structured choices, such as inquiring about symptom management methods, reporting current physical condition, or selecting psychological assessment options. The interaction content with the target user reflects the user's immediate needs and concerns, and is an important basis for the system to provide personalized responses.

[0059] The second traceable response content refers to the second round of output generated by the retrieval enhancement generation module based on structured data and interactions with the target user. This output is organized in a user-friendly manner and includes complete traceability information. The second traceable response content provides answers, suggestions, or guidance based on the user's specific interactions, such as nursing plans, psychological support information, or nutritional advice, while ensuring that each piece of information can be traced back to an authoritative source in the pediatric oncology knowledge base. The second traceable response content is displayed to the user through the patient-side module, satisfying both information accuracy requirements and ensuring user understanding and usability.

[0060] The patient-side module is also used to: generate intelligent Q&A content, psychological screening and assessment information, peer care information, follow-up management suggestions, and nutritional guidance suggestions based on the second traceable response content, and provide them to the target user. The specific implementation process is as follows: The system receives a second traceable response content output from the retrieval enhancement generation module. This content contains sorted information fragments and their traceability identifiers, such as nursing instructions, psychological status assessment results, or nutritional parameters. The system parses the second traceable response content and extracts key data elements, such as symptom descriptions, psychological scores, follow-up time points, and nutritional indicators. This data is stored in a structured format, including text fields, numerical values, and timestamps.

[0061] For generating intelligent question-and-answer content, natural language generation (NLP) technology is applied to convert the second traceable response content into a user-friendly question-and-answer format. The NLP module, based on predefined templates and a rule engine, combines information fragments from the second traceable response content to generate a coherent text response. For example, if the second traceable response content includes methods for managing vomiting after chemotherapy, the module will populate a question-and-answer template, replacing variables with actual data, such as specific nursing procedures or medication recommendations. The generation process uses conditional logic to ensure the language is easy to understand, avoids medical jargon, and uses everyday language, such as suggesting trying small, frequent meals. The module also integrates context management to maintain the coherence of the question-and-answer process, such as adjusting the depth of the response based on the user's historical interactions. The intelligent question-and-answer content is output as text paragraphs or lists and displayed through the patient-side module's interface.

[0062] For generating psychological screening and assessment information, the system extracts psychological scale scores and assessment results, such as anxiety or depression scores, from the second traceable response content. The system uses statistical models to calculate risk levels and perform trend analysis. For example, for scores on the Rutter Behavior Scale, the system applies a weighted summation formula to calculate the total score: , This indicates the total score of the psychological screening assessment. Indicates the first The weight of each project Indicates the first The scores for each item This indicates the total number of items. The system generates an assessment report based on the total score, including risk warnings and suggestions. For example, if the total score exceeds a threshold, the report will indicate potential behavioral problems and recommend intervention measures. Psychological screening assessment information is presented in text and visual charts, such as score bar graphs, to help users intuitively understand their psychological state.

[0063] For generating peer-care information, the system generates age-appropriate interactive content based on emotional support data from the second traceable response content. For example, if the second traceable response content includes psychological counseling suggestions, the system uses a story generation algorithm to create a personalized fairy tale. Story generation is based on templates and variable substitution, such as mapping user-input emotional keywords like fear of injections to story elements. The module uses natural language processing technology to ensure the story language is vivid and engaging, for example, generating an adventure story about a brave little warrior. Peer-care information is output in text, audio, or animation format and provided through the patient-side module's interactive interface to enhance emotional support.

[0064] For generating follow-up management recommendations, the system extracts key data for the follow-up plan from the second traceable response content, such as the next check-up date, monitoring indicators, and precautions. The system uses a rule engine to generate personalized recommendations, such as recommending follow-up frequency based on treatment stage and historical data. The recommendation generation process involves time series analysis, such as calculating the next follow-up time point. , This indicates the recommended next follow-up time. Indicates the date of the last follow-up visit. This indicates the interval period based on the treatment protocol. Follow-up management recommendations are output as a text list or calendar view, including reminders and action items, such as "Please have a blood test next month."

[0065] For generating nutritional guidance recommendations, the system extracts nutritional assessment data from the second traceable response content, such as body mass index, dietary records, and deficiency markers. The system applies nutritional models to calculate personalized recommendations, such as recommending daily intake based on energy requirement formulas. , Indicates daily energy requirements. This represents the basal metabolic rate. This indicates the activity factor. Nutritional guidance recommendations are output in the form of text descriptions and recipes, such as recommending increased protein intake and providing specific food examples.

[0066] Finally, the system integrates all generated content into a unified output and provides it to the target users through the patient-side module. The output format is based on web standards such as HTML and CSS, ensuring accessibility across different devices. The system also includes personalized settings to adjust the display based on user preferences, such as prioritizing the display of psychological screening assessment information or nutritional guidance suggestions. The entire process ensures that the content is accurate, timely, and easy to understand, supporting target users in managing their health in the home environment.

[0067] Intelligent Q&A content refers to easily understandable natural language answers generated based on user queries and second-traceable responses. These answers are presented in a question-and-response format, providing immediate information support. Intelligent Q&A content covers medical knowledge, nursing guidance, and everyday advice, using non-technical language to ensure user comprehension. In pediatric oncology management systems, intelligent Q&A content helps patients or their families quickly obtain authoritative answers, such as how to manage treatment side effects or understand diagnostic results.

[0068] The psychological screening and assessment information refers to the assessment report generated by the system through parsing psychological scale data from the second traceable response content. This report includes score calculations, risk analysis, and intervention recommendations. The psychological screening and assessment information is based on standardized psychological measurement tools, such as behavioral scales or emotion questionnaires, and the output is presented in the form of text summaries and visual charts. In the pediatric oncology management system, psychological screening and assessment information is used to monitor patients' mental health status and provide early warnings and support resources.

[0069] Peer-to-child support information refers to emotional support and interactive content designed for pediatric patients. This content, based on the psychological guidance elements in the second traceable response content, generates stories, games, or encouraging messages that are suitable for children to understand. Peer-to-child support information uses age-appropriate language and themes, such as fairy tales or wish cards, aiming to alleviate anxiety and promote emotional expression. In the pediatric oncology management system, peer-to-child support information enhances patients' psychological resilience and treatment adherence through engaging interactions.

[0070] The follow-up management recommendations refer to personalized management plans generated based on follow-up data from the second traceable response, including schedules, examinations, and precautions. Based on treatment protocols and historical records, these recommendations provide structured action guidelines, such as next appointment dates or family monitoring indicators. Within the pediatric oncology management system, follow-up management recommendations help patients and their families systematically implement rehabilitation plans, ensuring continuity of medical care.

[0071] Nutritional guidance recommendations refer to dietary and nutritional plans generated based on the nutritional assessment results in the second traceable response content. These recommendations include food suggestions, intake levels, and supplementation strategies. Nutritional guidance recommendations utilize nutritional principles and individualized data, such as weight changes or dietary preferences, to provide specific and actionable guidance. In the pediatric cancer management system, nutritional guidance recommendations support patients in maintaining good nutritional status, promoting treatment recovery and overall health.

[0072] Optionally, the above technical solution also includes a doctor-side module; the doctor-side module is used to: send the interaction content with the doctor to the retrieval enhancement generation module; The retrieval enhancement generation module is also specifically used to: retrieve relevant third-party information fragments from the pediatric oncology knowledge base based on structured data, interaction content with the target user, and interaction content with the doctor; and use semantic reordering technology to sort the third-party information fragments to generate third-party traceable response content. The specific implementation process is as follows: The system receives structured data output from the multimodal data parsing module. This data includes standardized fields from parsed clinical text, medical image feature vectors, pathology report coding information, gene sequencing variation records, nutritional assessment parameters, and psychological scale scores. Simultaneously, the system receives interaction content from the patient-side module, such as questions posed or options selected by the user in natural language, and from the doctor-side module, such as professional queries or clinical decision requests entered by the doctor.

[0073] Semantic parsing and feature extraction are performed on interactions with the target user and interactions with the doctor, respectively. For interactions with the target user, the system uses natural language understanding technology to identify user intent and key entities, such as extracting symptom descriptions or sentiment keywords from the text, and generates a user interaction feature vector. For interactions with the doctor, the system applies a professional terminology processing model to parse the doctor's input, such as diagnostic codes or treatment protocol references, and generates a doctor interaction feature vector. The formula for constructing the user interaction feature vector is: , Represents the user interaction feature vector. A textual representation of the interaction content with the target user. This represents a neural network encoder specifically designed for encoding user interaction content. The formula for constructing the doctor's interaction feature vector is: , Represents the doctor interaction feature vector. Text representation of interactions with doctors. This refers to a neural network encoder specifically designed for encoding interactive content for doctors.

[0074] Then, the vector representations of the structured data, user interaction feature vectors, and doctor interaction feature vectors are fused to generate a comprehensive query vector. The fusion process uses a weighted summation method, with the following formula: , Represents the comprehensive query vector. Vector representation of structured data Represents the user interaction feature vector. Represents the doctor interaction feature vector. , and These are the weighting coefficients for the three elements, which are dynamically adjusted based on the context. For example, for complex clinical decisions, the weighting coefficients for the doctor's interaction content... It may be set higher to prioritize professional input.

[0075] The retrieval enhancement generation module uses a comprehensive query vector to perform a similarity search within a pediatric oncology knowledge base. The module calculates the cosine similarity between the comprehensive query vector and each information fragment vector in the knowledge base, selecting information fragments with similarity exceeding a preset threshold to form an initial third set of information fragments. The similarity calculation uses the following formula: , Indicates the first in the knowledge base Vector representation of each information fragment Represents the vector dot product. Denotes the norm of the comprehensive query vector. The norm of the information fragment vector.

[0076] Then, semantic reordering is applied to finely rank the initial set of third information fragments. The semantic reordering technique uses a transformer-based neural network model trained on pediatric oncology data to evaluate the semantic relevance of each information fragment to the comprehensive query vector. The model input includes the comprehensive query vector and the text content of each information fragment, and outputs a relevance score. The score is calculated based on an attention mechanism, using the following formula: , Indicates the first The relevance score of each information fragment. Indicates the first A semantic representation vector of an information fragment. Represents the dimension of a vector. This represents the normalized exponential function. The system sorts the third information segments in descending order based on their scores, ensuring that the most relevant and authoritative information segments are at the top.

[0077] Finally, the retrieval enhancement generation module generates third-party traceable response content. This module synthesizes the sorted third-party information fragments into a coherent natural language response or structured data output, attaching detailed source information to each fragment, including the source document identifier, version number, location reference, and update time. For example, the response content may include treatment recommendations, diagnostic support, or resource recommendations, clearly indicating that this content originates from specific chapters of authoritative treatment guidelines or case records of expert clinical experience. The entire generation process ensures the accuracy, verifiability, and professionalism of the response. The resulting third-party traceable response content will be provided to clinicians through the physician-side module to support decision-making and patient management.

[0078] The third traceable response content refers to the third round of output generated by the retrieval enhancement generation module based on structured data, interactions with the target user, and interactions with physicians. This result is presented in a professional and structured format and includes complete traceability information. The third traceable response content provides comprehensive answers, decision support, or management suggestions for complex clinical scenarios, such as personalized treatment protocols, multimodal data integration and analysis, or referral guidance. It ensures that each information point can be traced back to authoritative sources in the pediatric oncology knowledge base, such as authoritative treatment guidelines, expert clinical experience, or anonymized case data. The third traceable response content is displayed to clinicians through the physician-side module, meeting medical-grade accuracy requirements while supporting auditing and continuous improvement.

[0079] The doctor-side module is also used to: generate intelligent Q&A content, patient management overview, and follow-up record overview based on third-party traceable responses, and provide them to doctors. The specific implementation process is as follows: The system receives third-party traceable response content from the retrieval enhancement generation module. This content includes sorted information fragments and their complete traceability identifiers, such as treatment recommendations, diagnostic support information, multimodal data integration results, or referral guidance. The system parses the third-party traceable response content and extracts key data elements, such as clinical indicators, treatment protocols, follow-up time points, abnormal event records, and patient status parameters. This data is stored in a structured format, including numerical fields, text descriptions, and timestamps.

[0080] For generating intelligent question-and-answer content, the system applies natural language generation technology to convert third-party traceable responses into a professional question-and-answer format. The natural language generation module, based on predefined medical templates and a rule engine, combines information fragments from the third-party traceable responses to generate coherent text responses. For example, if the third-party traceable response contains recommendations for targeted drugs for a specific gene mutation, the module will populate a question-and-answer template, replacing variables with actual data such as drug names, dosage instructions, and clinical evidence citations. The generation process uses conditional logic to ensure a professional language style that conforms to medical standards, such as using terminology like PD-1 inhibitors or adverse reaction monitoring. The module also integrates context management functionality, adjusting the depth and scope of the response based on the doctor's historical queries; for example, if the doctor has previously inquired about similar cases, the system will provide comparative analysis. The intelligent question-and-answer content is output in the form of text paragraphs or structured lists and displayed through the doctor's interface, ensuring accuracy and ease of reference.

[0081] For generating the patient management overview, the system integrates multi-source data from third-party traceable response content to create a comprehensive management view of patients. The system extracts basic information such as diagnostic results, treatment stages, key physiological indicators, and psychological state scores from the third-party traceable response content, and uses a data aggregation engine to combine these elements into a unified report. For example, the system calculates the patient's overall health score using the following formula: ,in, This represents the patient's overall health score. Indicates the total number of indicators. Indicates the first The weight of each indicator, Indicates the first Normalized values ​​of each indicator. Weights The system dynamically assigns weights based on clinical importance; for example, tumor volume metrics may be given higher weights. It also generates visualizations, such as dashboards or summary cards, displaying treatment progress, risk levels, and to-do lists. A patient management overview is output as an interactive webpage or report document, allowing physicians to quickly review patient status and make decisions.

[0082] For generating the follow-up record overview, the system uses follow-up data from the third traceable response content to generate a summary and analysis of historical records. The system extracts follow-up time-series data, such as tumor size from multiple examinations, complete blood count values, and adverse reaction records, and uses a time-series analysis model to calculate trends and anomalies. For example, the system applies a moving average method to smooth data fluctuations; the formula is: , Indicates a point in time The moving average, Indicates window size. Indicates a point in time The system uses the raw measurements as a basis to generate charts, such as line graphs showing changes in tumor volume, or tables listing key events like medication adjustment dates. The follow-up record overview also includes text summaries highlighting important changes and recommended actions; for example, if the tumor volume increases beyond a threshold, the report will suggest a follow-up examination. Output is presented in visual charts and structured text to ensure physicians have a comprehensive understanding of the patient's follow-up history.

[0083] Finally, all generated content is integrated into a unified output and provided to clinicians through a physician-side module. The output format is based on web standards such as HTML, CSS, and JavaScript, ensuring accessibility and interactivity across different devices. The system also includes access control mechanisms, allowing only authorized physicians to access information, and logs usage to support auditing. The entire process ensures that the content is based on authoritative data, supporting clinicians in improving efficiency and accuracy in diagnosis, treatment planning, and follow-up management.

[0084] The intelligent question-answering content refers to the professional natural language question-and-answer output generated by the system based on third-party traceable response content. This output is presented in the form of questions and responses, providing immediate clinical decision support. The intelligent question-answering content covers medical knowledge queries, treatment protocol interpretations, and case analyses, using standard medical terminology and evidence citations to ensure professionalism. In the pediatric oncology management system, intelligent question-answering content helps clinicians quickly obtain authoritative answers, such as querying the latest clinical trial evidence or understanding complex symptom associations.

[0085] The Patient Management Overview refers to a comprehensive patient management view generated based on third-party traceable response content. This view integrates multimodal data such as diagnostic information, treatment progress, and physiological indicators, providing a summary of the overall health status. The Patient Management Overview includes elements such as health scores, risk alerts, and treatment plan summaries, presented in a structured and visual format. In the pediatric oncology management system, the Patient Management Overview supports clinicians in comprehensively assessing patient status, optimizing resource allocation, and providing personalized care.

[0086] The follow-up record overview refers to a summary of follow-up history generated based on third-party traceable response content. This summary includes time-series data, event logs, and trend analysis, presented in charts and text. The follow-up record overview covers content such as changes in examination results, adverse reaction logs, and intervention effectiveness assessments, helping to identify abnormal patterns. In pediatric oncology management systems, the follow-up record overview enables clinicians to efficiently track patient progress, adjust treatment plans promptly, and ensure continuity of care.

[0087] The multimodal big data model of this invention is specifically embodied in a complete intelligent system architecture. This architecture deeply processes six categories of heterogeneous data generated throughout the entire cycle of pediatric oncology diagnosis and treatment through a multimodal data parsing module: symptom descriptions and diagnostic records in clinical texts, tumor morphological features in medical images, cytological analysis results in pathology reports, molecular variation information in gene sequencing, metabolic indicators in nutritional assessments, and behavioral and emotional data from psychological scales. The multimodal data parsing module relies on deep neural networks and specialized feature extraction algorithms to standardize and parse the data for each modality—performing 3D segmentation and feature quantification for medical images, medical entity recognition and relation extraction for clinical texts, and variation annotation and functional prediction for gene sequencing data. Ultimately, this unstructured raw data is uniformly transformed into structured data with clear semantic definitions and numerical representations, establishing semantic relationships between different modalities. Building upon this foundation, the retrieval enhancement generation module integrates the generated structured data with real-time user queries using multi-dimensional features. Through vectorized retrieval technology, it accurately retrieves relevant information fragments from a knowledge base in the field of pediatric oncology, which includes authoritative treatment guidelines, expert clinical experience databases, and desensitized case databases. Subsequently, it employs semantic reordering technology based on attention mechanisms to perform multiple rounds of fine-ranking of the initial screening results. Based on clinical relevance and evidence level, it assigns weights to information fragments, ultimately generating intelligent question-and-answer content, full-cycle management suggestions, and clinical decision support that simultaneously meet the requirements of medical accuracy, complete traceability, and personalized user needs. This forms a complete technological closed loop from multi-source data fusion and understanding to the generation and application of professional knowledge.

[0088] The technical solution of the present invention will be further described through another embodiment.

[0089] The purpose of this embodiment is to overcome the shortcomings of the prior art, improve the efficiency and accuracy of diagnosis and treatment, realize the downward flow of high-quality medical resources, and provide children and their families with comprehensive and personalized support from disease cognition, treatment to rehabilitation.

[0090] The technical solution of this invention is based on a large-scale intelligent system specifically designed for pediatric oncology. This system achieves its functionality through an architecture of a base, dual engines, and two-end services. The overall system architecture includes an application layer, an engine layer, and a data and deployment layer. These layers work collaboratively to ensure the system's integrity, security, and availability. At the application layer, the system provides dual-end application services, including a patient-side and a doctor-side. The patient-side functional module is designed for children and their parents, providing intelligent Q&A content presented in plain language, popular science education on psychological screening and assessment, C4C peer care interaction, pre-diagnosis and report generation, follow-up management and advice, and nutritional guidance. The doctor-side functional module is designed for clinicians, providing intelligent Q&A content presented in professional language, a patient management overview, standardized treatment pathway assistance, and a follow-up record overview. These modules interact through a user interface, ensuring the accuracy and applicability of information transmission. At the engine layer, the system deploys a core processing engine, including a retrieval enhancement generation module and a multimodal data parsing module. The retrieval enhancement and generation module is responsible for retrieving relevant information fragments from a domain-specific knowledge base and using semantic reordering technology to fine-tune the search results, generating traceable response content while effectively suppressing AI illusions and ensuring that the output is traceable and verifiable. The multimodal data parsing module, relying on a multimodal data intelligence platform, achieves integrated parsing and fusion processing of heterogeneous data such as clinical texts, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales, generating structured data.

[0091] At the data and deployment layer, the system constructs a knowledge base in the field of pediatric oncology. This knowledge base is not a simple tweak of a general model, but rather a deep integration of authoritative pediatric oncology treatment guidelines, expert clinical experience, and anonymized case data from top pediatric hospitals, forming the professional knowledge foundation of the system. Simultaneously, the system supports localized deployment environments, ensuring that all patient data remains within the hospital, achieving privacy protection and medical compliance requirements, and meeting the highest data security standards in the healthcare industry.

[0092] The system's workflow begins with the multimodal data parsing module processing the input multimodal data to generate structured data. The retrieval enhancement and generation module then retrieves relevant information fragments from a pediatric oncology knowledge base based on the structured data, interactions with the target user, or interactions with doctors. Semantic reordering techniques are then applied to rank these fragments, generating traceable response content. These semantic reordering techniques further utilize an attention-based model to calculate relevance scores, ensuring that the most relevant information is prioritized for injection into the main model.

[0093] Based on traceable response content, the system generates intelligent Q&A content, a patient management overview, a follow-up record overview, psychological screening and assessment information, peer care information, follow-up management suggestions, and nutritional guidance suggestions, which are then provided to users through the patient or doctor's end. These calculations support the generation of visual and textual information, ensuring that the output content is intuitive and professional.

[0094] The entire system ensures data security through localized deployment, with all processing completed within the medical institution to prevent data leakage. This architecture improves diagnostic and treatment efficiency, optimizes resource allocation, and provides closed-loop support for full-cycle management, offering reliable and personalized intelligent solutions for the field of pediatric oncology.

[0095] This invention provides a specific implementation of a pediatric tumor management system based on a multimodal large model, and demonstrates the system's workflow and technical implementation through the following three typical application scenarios.

[0096] Scenario 1: Initial diagnosis and referral at a primary care hospital: Under the guidance of doctors at primary care hospitals, parents of sick children fill in their child's basic information and upload various examination reports using the pre-consultation and report generation functions of the patient-side module. The multimodal data parsing module performs integrated parsing and fusion processing on the uploaded clinical text, medical images, and other multimodal data to generate structured data. The retrieval enhancement generation module retrieves relevant information fragments from a pediatric oncology knowledge base based on this structured data, sorts the information fragments using semantic reordering technology, and generates a structured pre-consultation report. This report includes a preliminary diagnostic impression and recommendations for further examinations. Primary care physicians log into the doctor-side module and view the pre-consultation report in the patient management overview. Simultaneously, doctors use the doctor-side intelligent question-and-answer function to input the child's key symptoms and examination indicators. Based on the input, the retrieval enhancement generation module retrieves the latest differential diagnosis and treatment recommendations from the pediatric oncology knowledge base and generates intelligent question-and-answer content within seconds. The system integrates pre-consultation information and doctor query results, and based on authoritative treatment guidelines and expert clinical experience, it suggests referral indications and recommends higher-level hospital specialists for referral, assisting primary care physicians in making accurate initial diagnosis and referral decisions.

[0097] Scenario 2: Diagnosis and follow-up at a specialized hospital: At specialized hospitals, attending physicians input the child's multiple follow-up records into the system. The multimodal data analysis module automatically aggregates key indicators from each follow-up visit, including tumor size, blood routine values, adverse reaction records, and other multimodal data, and performs integrated analysis and fusion processing to generate structured data. Based on this structured data, the system automatically generates individualized disease trend charts, intuitively displaying treatment response. This visualized information is presented through the follow-up record overview on the doctor's end module, assisting doctors in adjusting treatment plans. Doctors can use the intelligent question-and-answer function on the doctor's end to query the latest targeted drug clinical trial evidence for the child's specific gene mutation. The retrieval enhancement generation module retrieves relevant information fragments from the pediatric oncology knowledge base, uses semantic reordering technology to sort the information fragments, and generates intelligent question-and-answer content containing the latest clinical trial evidence, supporting personalized treatment.

[0098] Scenario 3: Home Management and Support for Patients Parents can use the intelligent question-and-answer function of the patient-side module at home to search for information in conversational language, such as "What should I do if my child vomits after chemotherapy?" The retrieval enhancement and generation module retrieves relevant information fragments from the pediatric oncology knowledge base based on the question content, uses semantic reordering technology to sort the information fragments, and generates easy-to-understand intelligent question-and-answer content to provide nursing guidance. Children regularly use the psychological screening and assessment function of the patient-side module to complete self-assessments such as the Rutter Behavior Scale. The system automatically scores the questionnaire responses, generates psychological screening and assessment information, and automatically warns of potential psychological problems based on the scores, reminding parents and doctors to intervene in a timely manner. Children can interact with the C4C peer support function of the patient-side module by inputting emotional expressions such as "I'm a little afraid of injections." The system generates peer support information based on the input, such as creating a fairy tale about a brave little warrior defeating the needle monster, providing emotional support and psychological guidance. Parents use the nutritional guidance function of the patient-side module to input the child's current weight, treatment stage, and appetite. The multimodal data parsing module analyzes and processes this nutritional assessment data to generate structured data. The retrieval enhancement module uses this structured data to retrieve relevant information fragments from a knowledge base in the field of pediatric oncology, generating personalized nutritional guidance recommendations, including dietary advice and nutritional supplementation plans.

[0099] As can be seen from the above specific implementation methods, the present invention deeply integrates cutting-edge artificial intelligence technologies of multimodal data parsing module, retrieval enhancement generation module, patient-side module and doctor-side module with clinical practice of pediatric oncology, forming an overall solution covering initial diagnosis, referral, diagnosis and treatment, follow-up and home management.

[0100] The beneficial effects of this invention are reflected in several aspects. Through the overall architecture and specific technical implementation of a large-scale intelligent system specifically for pediatric oncology, it brings substantial improvements to the diagnosis and treatment process, resource allocation, patient management, and data security. Based on the core processing capabilities of the multimodal data parsing module and the retrieval enhancement generation module, combined with dual-end services from the patient-side and doctor-side modules, and the professional foundation of a pediatric oncology knowledge base, the system achieves the following beneficial effects: First, it improves the accuracy and efficiency of diagnosis and treatment. The system provides clinicians with evidence-based medicine support with sub-second response times. Through a retrieval enhancement generation module, relevant information fragments are retrieved from a pediatric oncology knowledge base based on structured data, interactions with the target user, or interactions with the doctor. Semantic reordering technology is then used to sort these fragments, generating traceable response content. This process effectively suppresses AI illusions, ensuring that output suggestions are based on authoritative treatment guidelines, expert clinical experience, and anonymized case data. The multimodal data parsing module performs integrated analysis and fusion processing of multimodal data, including clinical text, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales, generating structured data and automatically producing pre-consultation reports and follow-up reports. This assists doctors in quickly identifying changes in key indicators, significantly improving decision-making efficiency and accuracy.

[0101] Second, the system facilitates the downward flow of high-quality resources. It empowers primary care physicians with standardized treatment pathways and expert experience through artificial intelligence. The patient-side module's pre-consultation and report generation functions allow parents to fill in basic information and upload examination reports; the enhanced generation module then generates structured pre-consultation reports based on these inputs. The physician-side module provides a patient management overview and intelligent Q&A content, allowing primary care physicians to access differential diagnosis and treatment recommendations based on the latest guidelines. By integrating pre-consultation information and physician queries, the system provides referral indications and recommends higher-level hospital specialties, thereby improving the identification and referral capabilities of primary care physicians, shortening the time to diagnosis for children, and alleviating the problem of uneven distribution of professional resources.

[0102] Third, a closed-loop management system is constructed to cover the entire lifecycle. The system seamlessly integrates pre-consultation, diagnosis, treatment, follow-up, nutrition, and psychological aspects, forming a management platform covering multiple dimensions of physiology, psychology, and nutrition. The multimodal data parsing module aggregates key indicators from each follow-up visit, such as tumor size, blood routine values, and adverse reaction records, generating structured data. The retrieval enhancement generation module uses this data to generate an overview of follow-up records, nutritional guidance suggestions, and psychological screening and assessment information. This provides continuous, personalized, and comprehensive management for children, ensuring consistency across all stages.

[0103] Fourth, it greatly improves the medical experience. The system provides humanized interactive services through the patient-side module, generating intelligent Q&A content in plain language to answer parents' questions. The psychological screening and assessment module automatically scores questionnaire responses, generating psychological screening and assessment information. Based on the scores, the system provides early warnings of potential psychological problems, prompting timely intervention. The C4C peer care module generates peer care information, such as fairy tales or wish cards, transforming complex medical knowledge and psychological guidance into interactive forms that children can understand, providing emotional support and psychological counseling, and creating a warm and caring medical environment.

[0104] Fifth, ensure data security and compliance. The system supports localized deployment environments, with all patient data processed within the medical institution, without leaving the hospital, achieving zero privacy leaks. The multimodal data parsing module and the retrieval enhancement generation module run locally, ensuring that data processing meets the highest data security and compliance requirements in the medical industry. The knowledge base in the field of pediatric oncology is built based on anonymized case data, further protecting patient privacy and enabling the system to be safely promoted and applied in a strictly regulated medical environment.

[0105] Overall, this invention forms a complete and reliable technical solution by integrating a domain-deeply customized large model architecture, multimodal data fusion and full-cycle management integration, an adaptive interactive system for dual user groups, medical-grade reliability assurance based on retrieval enhancement generation modules and semantic reordering technology, and technical implementation of children's emotional support, providing comprehensive support for the diagnosis, treatment and management of pediatric tumors.

[0106] like Figure 2 As shown in the figure, an embodiment of the present invention provides a method for childhood tumor management based on a multimodal large model, comprising: S1. Process multimodal data of pediatric cancer patients to generate structured data, including clinical text, medical images, pathology reports, gene sequencing, nutritional assessments and psychological scales; S2. Based on structured data, retrieve relevant first information fragments from the pediatric oncology knowledge base, and sort the first information fragments using semantic reordering technology to generate first traceable response content; S3. Based on the first traceable response content, generate visual and / or textual information related to the health status of pediatric cancer patients and provide it to pediatric cancer patients or their families.

[0107] Optionally, the above technical solution also includes: Based on structured data and interaction content with the target user, relevant second information fragments are retrieved from the pediatric oncology knowledge base, and semantic reordering technology is used to sort the second information fragments to generate second traceable response content, wherein the target user is a pediatric cancer patient or a family member of a pediatric cancer patient. Based on the second traceable response content, intelligent Q&A content, psychological screening and assessment information, peer care information, follow-up management suggestions, and nutritional guidance suggestions are generated and provided to the target users.

[0108] Optionally, the above technical solution also includes: Based on structured data, interaction content with target users, and interaction content with doctors, relevant third-party information fragments are retrieved from the pediatric oncology knowledge base, and semantic reordering technology is used to sort the third-party information fragments to generate third-party traceable response content. Based on the third traceable response, intelligent question-and-answer content, patient management overview, and follow-up record overview are generated and provided to doctors.

[0109] Optionally, in the above technical solution, the pediatric oncology knowledge base stores authoritative diagnosis and treatment guidelines, expert clinical experience, and desensitized case data in the field of pediatric oncology.

[0110] It should be noted that the beneficial effects of the pediatric tumor management method based on a multimodal large model provided in the above embodiments are the same as those of the pediatric tumor management method based on a multimodal large model described above, and will not be repeated here. Furthermore, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation processes are detailed in the system embodiments, and will not be repeated here.

[0111] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned methods for childhood tumor management based on a multimodal large model.

[0112] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned methods for childhood tumor management based on a multimodal large model.

[0113] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A pediatric tumor management system based on a multimodal large model, characterized in that, It includes a multimodal data parsing module, a retrieval enhancement generation module, and an information generation module; The multimodal data parsing module is used to process multimodal data of pediatric cancer patients to generate structured data, including clinical text, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales. The retrieval enhancement generation module is used to: retrieve relevant first information fragments from the pediatric oncology knowledge base based on the structured data, and sort the first information fragments using semantic reordering technology to generate first traceable response content; The information generation module is used to: generate visual information and / or text information related to the health status of the child cancer patient based on the first traceable response content, and provide it to the child cancer patient or the child cancer patient's family.

2. The pediatric tumor management system based on a multimodal large model according to claim 1, characterized in that, It also includes a patient-side module; the patient-side module is used to: send the interaction content with the target user to the search enhancement generation module, the target user being the pediatric cancer patient or the family member of the pediatric cancer patient; The retrieval enhancement generation module is also used to: retrieve relevant second information fragments from the pediatric oncology knowledge base based on the structured data and the interaction content with the target user, and sort the second information fragments using semantic reordering technology to generate second traceable response content; The patient-side module is also used to: generate intelligent question-and-answer content, psychological screening and assessment information, peer care information, follow-up management suggestions, and nutritional guidance suggestions based on the second traceable response content, and provide them to the target user.

3. A pediatric tumor management system based on a multimodal large model according to claim 2, characterized in that, It also includes a doctor-side module; the doctor-side module is used to: send the interaction content with the doctor to the retrieval enhancement generation module; The retrieval enhancement generation module is also specifically used to: retrieve relevant third information fragments from the pediatric oncology knowledge base based on the structured data, the interaction content with the target user, and the interaction content with the doctor, and sort the third information fragments using semantic reordering technology to generate third traceable response content; The doctor-side module is also used to: generate intelligent question-and-answer content, patient management overview, and follow-up record overview based on the third traceable response, and provide them to the doctor.

4. A pediatric tumor management system based on a multimodal large model according to any one of claims 1 to 3, characterized in that, The pediatric oncology knowledge base stores authoritative treatment guidelines, expert clinical experience, and desensitized case data in the field of pediatric oncology.

5. A method for managing childhood tumors based on a multimodal large model, characterized in that, include: Multimodal data of pediatric cancer patients are processed to generate structured data, which includes clinical text, medical images, pathology reports, gene sequencing, nutritional assessments, and psychological scales. Based on the structured data, relevant first information fragments are retrieved from the pediatric oncology knowledge base, and the first information fragments are sorted using semantic reordering technology to generate first traceable response content. Based on the first traceable response content, visual information and / or text information related to the health status of the child cancer patient are generated and provided to the child cancer patient or the child cancer patient's family.

6. A method for childhood tumor management based on a multimodal large model according to claim 5, characterized in that, Also includes: Based on the structured data and the interaction content with the target user, relevant second information fragments are retrieved from the pediatric oncology knowledge base, and the second information fragments are sorted using semantic reordering technology to generate second traceable response content, wherein the target user is the pediatric oncology patient or the pediatric oncology patient's family member; Based on the second traceable response content, intelligent question-and-answer content, psychological screening and assessment information, peer care information, follow-up management suggestions, and nutritional guidance suggestions are generated and provided to the target user.

7. A method for childhood tumor management based on a multimodal large model according to claim 6, characterized in that, Also includes: Based on the structured data, the interaction content with the target user, and the interaction content with the doctor, relevant third information fragments are retrieved from the pediatric oncology knowledge base, and the third information fragments are sorted using semantic reordering technology to generate third traceable response content. Based on the third traceable response, intelligent question-and-answer content, patient management overview, and follow-up record overview are generated and provided to doctors.

8. A method for childhood cancer management based on a multimodal large model according to any one of claims 5 to 7, characterized in that, The pediatric oncology knowledge base stores authoritative treatment guidelines, expert clinical experience, and desensitized case data in the field of pediatric oncology.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for childhood tumor management based on a multimodal large model as described in any one of claims 5 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for childhood tumor management based on a multimodal large model as described in any one of claims 5 to 8.

Images