Head and neck cancer patient symptom management method and digital intelligence system based on multi-source data fusion and AI driving

Abstract

The invention provides a head and neck cancer patient symptom management method and system based on multi-source data fusion and AI driving, and the method comprises the steps: integrating electronic medical records, wearable equipment, questionnaire survey, popular science videos and other multi-modal data, employing an attention mechanism to dynamically fuse features, and constructing unified knowledge representation. A dynamically extended symptom-risk-intervention knowledge graph is established based on a Bayesian network and a graph neural network (GNN), and cross-layer symptom recommendation is realized in combination with a collaborative filtering algorithm. The symptom severity is evaluated in real time through an XGBoost algorithm, three-color early warning is output, and finally the knowledge graph is continuously optimized through reinforcement learning based on patient feedback data. According to the method, accurate and personalized symptom management of the patient with the head and neck cancer is realized.

Description

Technical Field

[0001] The present invention relates to the field of medical and health management technology, and in particular to a symptom management method and digital intelligent system for head and neck cancer patients based on multi-source data fusion and AI-driven, which is suitable for personalized symptom monitoring, risk assessment and intervention recommendations for radiotherapy patients. Background Art

[0002] Surgery, radiotherapy, chemotherapy, and multimodal combined therapy are the main treatments for HNC. Due to the complex anatomy and the influence of treatment factors, patients will experience a variety of physical and psychological discomfort symptoms, such as difficulty chewing, difficulty swallowing, taste disturbance, oral mucositis, dry mouth, pain, fatigue, loss of appetite, anxiety, and depression. These symptoms persist year-round, not only causing malnutrition and weight loss, which delays the treatment process; they also aggravate the patient's pain, generate negative emotions such as anxiety and fear, and even interrupt the treatment plan, seriously affecting the patient's quality of life and treatment compliance. Therefore, it is very necessary to strengthen symptom management of patients with head and neck cancer.

[0003] Digital management platforms are mobile devices developed using smartphones, primarily for specific disease functions, for disease control, risk prediction and management, comprehensive disease assessment, and health promotion. Recent studies have confirmed that digital management platforms can accurately assess symptoms and help patients recover. The advantage of digital management platforms lies in their ability to complement traditional mobile applications, eliminating the need for patients to perform tedious installation and downloading. Digital management platforms are becoming an excellent auxiliary tool for nutritional management in patients undergoing radiotherapy and chemotherapy.

[0004] However, some platforms on the market now have the following problems: First, the functions are relatively scattered, and they may only provide simple records and reminders for some common symptoms, lacking in-depth exploration and professional management of the correlation between symptoms; second, the existing platforms do not currently have powerful predictive functions such as machine learning, and mainly focus on providing information and simple interactions; third, data is fragmented, lacking the ability to efficiently integrate and analyze multi-source data, and the correlation between data is difficult to fully explore; fourth, the response is delayed, and the content update of the platform usually relies on the manual operation of the developer, with a long update cycle, making it difficult to achieve factual knowledge update and optimization; fifth, the intervention is extensive: there is a lack of the ability to generate personalized care pathways based on evidence-based medicine.

[0005] Faced with the enormous demands of symptom assessment and the development of personalized nutritional intervention plans, artificial intelligence (AI) may be an effective means of addressing this challenge. AI's advantage lies in its ability to extract, analyze, and comprehensively manage data, such as graphics or narrative text. This enables big data in medicine, further enabling personalized analysis and precision treatment through empirical learning.

[0006] Head and neck cancer patients often experience multiple symptoms during radiotherapy, such as oral mucositis and dysphagia. Traditional management relies on manual assessment, which is subject to high subjectivity and delayed response. Existing electronic health record (EHR) systems lack the ability to integrate multimodal data and have low knowledge graph update efficiency, making precise intervention difficult. Therefore, an artificial intelligence system that integrates multi-source data and dynamically optimizes is urgently needed to improve the real-time and personalized level of symptom management. Summary of the Invention

[0007] In view of this, it is necessary to provide a symptom management method and system for head and neck cancer patients based on multi-source data fusion and AI-driven. By utilizing multi-source data fusion and artificial intelligence (AI) technology, through steps such as constructing a knowledge graph, real-time risk stratification, and continuous optimization, dynamic management and intervention of symptoms in head and neck cancer patients can be achieved, so as to solve the three core problems of low utilization of multi-source data, lagging symptom management, and insufficient personalization of intervention plans in existing technologies.

[0008] In a first aspect, embodiments of the present application provide a method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven, the method comprising:

[0009] Step S1, multimodal data collection: electronic medical record data is collected through a local database, sensor data is collected through wearable devices, patient data and effect evaluation data are collected through questionnaires, and head and neck cancer popular science video data is collected through online resources; the collected multimodal data is organized;

[0010] Step S2, multimodal data fusion: Standardize and extract knowledge elements from electronic medical records, wearable device data, patient questionnaires and effect evaluation data, and popular science video data. Apply an attention mechanism to dynamically assign feature weights, fuse the weighted knowledge elements, and form a unified knowledge representation.

[0011] Step S3: Construct a multi-layer symptom knowledge graph: Based on the unified knowledge representation data, Bayesian network modeling is performed to establish a symptom-risk-intervention association knowledge network, and a graph neural network (GNN) is used to achieve dynamic node expansion.

[0012] Step S4, collaborative recommendation: A clinical decision support system is constructed through the patient-side intelligent head and neck cancer radiotherapy patient self-management platform. Symptoms are recorded through the clinical decision support system, and human-computer interaction is performed based on a multi-layer symptom knowledge graph. Cross-layer recommendations are then performed based on collaborative filtering methods.

[0013] Step S5, real-time risk stratification: Calculate the symptom severity score using the XGBoost algorithm and output green / yellow / red warning signals and corresponding intervention paths;

[0014] Step S6, continuous optimization: update the knowledge graph through reinforcement learning based on patient feedback data.

[0015] Optionally, in an implementation of the first aspect of the present invention, in step S1, organizing the collected multimodal data includes:

[0016] Using a data warehouse to store and manage the collected multimodal data;

[0017] Use data pre-processing techniques to remove noise and outliers, including duplicate, irrelevant, inaccurate, or outdated data;

[0018] Classify and label the data.

[0019] Optionally, in an implementation of the first aspect of the present invention, step S2, multimodal data fusion: standardizing electronic medical records, wearable device data, patient questionnaires and effect evaluation data, and popular science video data and extracting knowledge elements, applying an attention mechanism to dynamically assign feature weights, and fusing the weighted knowledge elements to form a unified knowledge representation, including:

[0020] S2.1: Electronic medical records are encoded using international standards and structured metadata. Unstructured text is identified and terminology standardized using the BERT deep learning model. Private information is desensitized. Wearable devices, questionnaires, and evaluation forms are standardized in timestamp format, units, and Likert rating scales, and outliers are removed using cleaning techniques. Keyframe text is extracted from popular science videos using optical character recognition (OCR), speech-to-text data is converted, and semantic tags are annotated using a medical ontology library.

[0021] S2.2: For structured elements, symptoms, diagnosis codes, and medication records are extracted from electronic medical records; symptom scores are extracted from questionnaires; and disease keywords are extracted from videos. For unstructured elements, entities and relationships are extracted using the BERT deep learning model; and trend features are extracted from physiological parameters of wearable devices through time series analysis.

[0022] S2.3, using the attention mechanism to achieve cross-modal feature weighted fusion, based on the similarity between the query vector query and the feature vector key of each modality, the weight α is generated by softmax i , where the similarity S based on the attention mechanism att The formula is:

[0023]

[0024] Among them, Q is the query vector, K is the key, and V is the key value;

[0025] S2.4, weighted sum h = ∑α i e i , after splicing, the dimension is reduced by the fully connected layer, and the weight ratio is dynamically adjusted through the gating mechanism to align across modalities, where e i The corresponding feature sequence;

[0026] S2.5, the fused multi-dimensional vector is used as the multimodal fusion unified knowledge representation.

[0027] Optionally, in an implementation of the first aspect of the present invention, step S3, constructing a multi-layer symptom knowledge graph: performing Bayesian network modeling based on unified knowledge representation data, establishing a symptom-risk-intervention association knowledge network, and using a graph neural network (GNN) to achieve dynamic node expansion, including:

[0028] S3.1, using the multimodal fusion unified knowledge representation, perform symptom-risk-intervention association modeling and construct a Bayesian network, where nodes represent medical variables and edges represent conditional dependencies;

[0029] S3.2, initialize node probability distribution based on electronic medical record statistics;

[0030] S3.3, update the conditional probability table through variational inference to integrate new patient data;

[0031] S3.4, enabling causal reasoning through reverse diagnosis and risk prediction;

[0032] S3.5, monitoring new data, and when new data is generated, obtaining the new data;

[0033] S3.6, using a graph neural network (GNN) method to extract dynamically added medical entities from the newly added data and construct a hierarchical dynamic GNN architecture;

[0034] In S3.7, the initial embedding layer uses a Graph Attention Network to encode existing graph nodes. For the dynamic update layer, the initial embedding of new nodes is generated through neighbor sampling and aggregation. The LSTM module records the node state history, adapts to graph evolution, and realizes time-series state transfer to obtain the dynamic update layer.

[0035] S3.8, the probability distribution of BN output is input into GNN as node attributes to achieve cross-modal alignment.

[0036] Optionally, in an implementation of the first aspect of the present invention, S3.9 further includes processing medical uncertainty in combination with a Bayesian GNN, including:

[0037] Quantify the node classification confidence based on Dirichlet distribution and determine the uncertainty of evidence;

[0038] Determine probabilistic uncertainty by simulating the prediction variance under parameter perturbations using Monte Carlo Dropout.

[0039] S3.10, based on end-to-end training, performs joint optimization, including:

[0040] The loss function jointly optimizes the likelihood estimation of BN and the node classification loss of GNN. The formula of the joint loss function L is as follows:

[0041] L=λ1L que +λ2L dis ,

[0042] Among them, L que For the symptom query loss, the negative log-likelihood loss of the Bayesian network is used, L dis is the disease diagnosis loss, using cross entropy loss, λ1 and λ2 represent the dynamic weight coefficients of the corresponding loss function,

[0043] L que = -logP(X|Θ),

[0044] Among them, Θ is the conditional probability table, X is the symptom observation data;

[0045]

[0046] Among them, y i For head and neck cancer classification, Predict probability for GNN;

[0047] Adaptive adjustment is performed by dynamically balancing task contributions using gradient statistics;

[0048] S3.11 uses the AdamW optimizer, whose adaptive learning rate feature is suitable for handling the gradient difference of the joint loss:

[0049]

[0050] Among them, θ t+1 ,θ t Represent the gradient differences at t+1 and t times respectively, are the first-order and second-order moment estimates of the gradient, respectively.

[0051] Optionally, in an implementation of the first aspect of the present invention, step S4 includes constructing a clinical decision support system through a patient-side intelligent head and neck cancer radiotherapy patient self-management platform, recording symptoms through the clinical decision support system, performing human-computer interaction based on a multi-layer symptom knowledge graph, and performing cross-layer recommendations based on a collaborative filtering method, including:

[0052] Building a hybrid cloud architecture, including an intelligent patient self-management platform for head and neck cancer radiotherapy;

[0053] Information input is performed through the multimodal interactive terminal of the patient-side intelligent head and neck cancer radiotherapy patient self-management platform, including voice input, text input, scale selection, wearable device synchronization and image upload;

[0054] Embed the Bayesian-GNN multi-layer symptom graph generated in step S3;

[0055] A cross-layer patient-symptom matrix decomposition algorithm is used for cross-layer collaborative filtering recommendation, where cross-layer collaborative filtering recommendation includes horizontal same-level collaboration and vertical cross-level collaboration.

[0056] Optionally, in an implementation of the first aspect of the present invention, the cross-layer collaborative filtering recommendation using a cross-layer patient-symptom matrix decomposition algorithm includes:

[0057] Construct an m×n patient-symptom matrix:

[0058]

[0059] Where m represents the number of patients, n represents the number of symptoms, V ki Indicates the patient's matching degree to the symptoms;

[0060] According to the patient-symptom matrix V, the similarity S(i,j)1 between symptom i and symptom j is calculated using the following formula:

[0061]

[0062] Among them, N(a) represents the set of patients that match both symptom i and symptom j, N(i) and N(j) represent the sets of patients that match symptom i and symptom j respectively;

[0063] Calculate the similarity S(i,j)2 between symptom i and symptom j in the knowledge graph. The formula is as follows:

[0064]

[0065] Among them, x i 、x j Respectively represent the entity of symptoms in the multimodal fusion unified knowledge, y i 、y j Respectively represent the relationship between symptoms in the multimodal fusion unified knowledge, z i 、z j They represent the attributes of symptoms in the unified knowledge of multimodal fusion;

[0066] The similarity S(i,j)1 and the similarity S(i,j)2 are combined to obtain the total similarity S(i,j), which is expressed as:

[0067] S(i,j)=λS(i,j)1+(1-λ)S(i,j)2,

[0068] Among them, λ represents the weight of similarity;

[0069] Filter out the top N most matching symptom sets based on the total similarity S(i,j);

[0070] Based on the Bayesian-GNN multi-layer symptom graph generated in step S3 based on the first N symptom sets, horizontal same-level collaboration and vertical cross-level collaborative recommendation are performed.

[0071] In a second aspect, an embodiment of the present application provides a symptom management system for head and neck cancer patients based on multi-source data fusion and AI-driven, which is applied to the symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven as described in the first aspect, including:

[0072] A multimodal data collection module is used to collect electronic medical record data through a local database, sensor data through wearable devices, patient data and effect evaluation data through questionnaires, and head and neck cancer popular science video data through online resources; and organize the collected multimodal data;

[0073] The multimodal data fusion module is used to standardize and extract knowledge elements from electronic medical records, wearable device data, patient questionnaires and effect evaluation data, and popular science video data. It uses the attention mechanism to dynamically assign feature weights and fuse the weighted knowledge elements to form a unified knowledge representation.

[0074] A multi-layer symptom knowledge graph construction module is used to perform Bayesian network modeling based on unified knowledge representation data, establish a symptom-risk-intervention association knowledge network, and use graph neural network (GNN) to achieve dynamic node expansion;

[0075] A multi-layer collaborative recommendation module is used to build a clinical decision-making support system through the patient-side intelligent head and neck cancer radiotherapy patient self-management platform. Symptoms are recorded through the clinical decision-making support system, and human-computer interaction is carried out based on the multi-layer symptom knowledge graph. Cross-layer recommendations are also made based on collaborative filtering methods.

[0076] A real-time risk stratification module, which uses the XGBoost algorithm to calculate symptom severity scores and output green / yellow / red warning signals and corresponding intervention pathways;

[0077] A continuous optimization module for updating the knowledge graph through reinforcement learning based on patient feedback data.

[0078] In a third aspect, an embodiment of the present application provides an electronic device, including:

[0079] processor;

[0080] a memory for storing processor-executable instructions;

[0081] Wherein, the processor is configured to implement the symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven as described in the first aspect when executing the instructions.

[0082] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium storing a program, wherein the program instructs a device to execute the symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven as described in the first aspect.

[0083] The present invention proposes a symptom management method and system for head and neck cancer patients based on multi-source data fusion and AI-driven. By integrating multimodal data such as electronic medical records, wearable devices, questionnaires and popular science videos, the attention mechanism is used to dynamically fuse features and construct a unified knowledge representation. Based on the Bayesian network and graph neural network (GNN), a dynamically expanded symptom-risk-intervention knowledge graph is established, and the collaborative filtering algorithm is combined to realize cross-layer symptom recommendation. The XGBoost algorithm is used to evaluate the severity of symptoms in real time and output a three-color warning. Finally, the knowledge graph is continuously optimized through reinforcement learning based on patient feedback data. This method realizes precise and personalized symptom management of head and neck cancer patients.

[0084] Beneficial effects:

[0085] (1) Multi-source data fusion: Combining multimodal data to improve the comprehensiveness and accuracy of symptom assessment.

[0086] (2) Dynamic knowledge graph: Symptom association modeling and dynamic expansion are achieved through Bayesian networks and GNNs to enhance medical reasoning capabilities.

[0087] (3) Intelligent recommendation and early warning: Provide personalized intervention suggestions and real-time risk stratification based on collaborative filtering and XGBoost algorithms to improve patient management efficiency.

[0088] (4) Continuous optimization: Use reinforcement learning to iteratively update the knowledge graph to improve the system's adaptability and clinical practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 A flowchart of a method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven implementation is provided in accordance with one embodiment of the present application.

[0090] Figure 2This is a functional diagram of the module of the self-management platform for head and neck cancer radiotherapy patients provided in one embodiment of the present application.

[0091] Figure 3 A questionnaire guide diagram provided in an embodiment of the present application.

[0092] Figure 4 A schematic diagram of human-machine collaboration in decision-making assistance provided in one embodiment of the present application.

[0093] Figure 5 A schematic diagram of a symptom management system module for head and neck cancer patients based on multi-source data fusion and AI-driven, provided in one embodiment of the present application.

[0094] Figure 6 A schematic diagram of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0095] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0096] It should be noted that, in the embodiments of the present application, "at least one" refers to one or more, and "more" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art in the art to which this application relates. The terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0097] It should be noted that, in the embodiments of the present application, words such as "first" and "second" are only used for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying an order. Features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a concrete way.

[0098] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0099] Example 1

[0100] Figure 1A flowchart of a method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven implementation is provided in accordance with one embodiment of the present application.

[0101] like Figure 1 As shown in the figure, a symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven, including:

[0102] Step S1, multimodal data collection: electronic medical record data is collected through the local database, sensor data is collected through wearable devices, patient data and effect evaluation data are collected through questionnaires, and head and neck cancer popular science video data is collected through online resources; the collected multimodal data is organized.

[0103] Specifically, electronic health records (EHRs) are a crucial source of medical data, containing basic patient information, medical history, diagnosis, treatment plans, and more. This data is typically stored in a hospital's EHR system, where it is structured and standardized, facilitating subsequent analysis and processing. Modern EHR systems provide detailed patient medical records, which can be used for subsequent multimodal data fusion and analysis.

[0104] Wearable devices can monitor patients' physiological parameters in real time, such as heart rate, blood pressure, blood oxygen saturation, and sleep quality. This data can provide patients with dynamic health monitoring information. Wearable devices transmit data to the cloud via Bluetooth or Wi-Fi technology for storage and analysis. Furthermore, these devices can collect information on the impact of environmental conditions on health, providing patients with a more comprehensive health assessment.

[0105] Questionnaires are an important means of obtaining data on patients' subjective experiences and quality of life. Questionnaires can be administered in either paper or electronic format and should be designed scientifically and rationally, avoiding leading questions to ensure the accuracy and reliability of the data. This data is often used to assess treatment effectiveness and provide personalized health advice to patients. Figure 2 This is a functional diagram of the self-management platform module for head and neck cancer radiotherapy patients provided in one embodiment of the present application. The patient inputs the questionnaire through the terminal. Figure 3 This is a questionnaire guide diagram provided in one embodiment of the present application. Figure 3 As shown, through questionnaires, patients are guided to complete their information and provide stratified management.

[0106] Online resources can provide a wealth of popular science videos. These videos typically cover the causes, treatments, and preventive measures of head and neck cancer, helping to improve patients' understanding of the disease and their ability to manage it. This data can be obtained by crawling online platforms or through collaborative efforts.

[0107] Specifically, the collected multimodal data needs to be organized, cleaned, standardized, and integrated to ensure compatibility and consistency. A data warehouse is used to store and manage the collected multimodal data. Data preprocessing techniques are used to remove noise and outliers, including duplicate, irrelevant, inaccurate, or outdated data. The data is then classified and labeled for subsequent analysis and mining.

[0108] After data organization is complete, data from different modalities can be fused to form a multimodal feature dataset for subsequent analysis and modeling. For example, combining electronic medical record data, sensor data, questionnaire data, and popular science video data can build a more comprehensive patient health model for disease diagnosis and treatment effect evaluation.

[0109] Through the above steps, multimodal data can be effectively collected and managed to provide support for subsequent medical research and clinical decision-making. This process not only improves data utilization but also provides patients with more accurate and personalized medical services.

[0110] Step S2, multimodal data fusion: standardize electronic medical records, wearable device data, patient questionnaires and effect evaluation data, and popular science video data and extract knowledge elements. Apply the attention mechanism to dynamically assign feature weights, fuse the weighted knowledge elements, and form a unified knowledge representation.

[0111] Specifically, the following steps are included:

[0112] S2.1: Electronic medical records are encoded using international standards and structured metadata. Unstructured text is identified and terminology standardized using the BERT deep learning model. Private information is desensitized. Wearable devices, questionnaires, and evaluation forms are standardized in timestamp format, units, and Likert rating scales, and outliers are removed using cleaning techniques. Keyframe text is extracted from popular science videos using optical character recognition (OCR), speech-to-text data is converted, and semantic tags are annotated using a medical ontology library.

[0113] S2.2: For structured elements, we extract diagnostic codes (ICDs) and medication records (NDCs) from electronic medical records; extract symptom scores from questionnaires; and extract disease keywords from videos. For unstructured elements, we use the BERT deep learning model to extract entities and relationships. We also extract trend features from physiological parameters of wearable devices through time series analysis.

[0114] S2.3, using the attention mechanism to achieve cross-modal feature weighted fusion, based on the similarity between the query vector query and the feature vector key of each modality, the weight α is generated by softmax i , where the similarity S based on the attention mechanism att The formula is:

[0115]

[0116] Among them, Q is the query vector, K is the key, and V is the key value;

[0117] S2.4, weighted sum h = ∑α i e i , after splicing, the dimension is reduced by the fully connected layer, and the weight ratio is dynamically adjusted through the gating mechanism to align across modalities, where e i The corresponding feature sequence;

[0118] S2.5, the fused multi-dimensional vector is used as the multimodal fusion unified knowledge representation.

[0119] Specifically, data from different sources must first be standardized to ensure consistency. Electronic medical records: Use international standard coding (such as ICD-10 disease classification) and metadata structured processing; unstructured text uses NLP technology (such as BERT or LSTM) for entity recognition and terminology standardization; private information needs to be desensitized. Wearable devices / questionnaires / evaluation forms: Unify timestamp formats (such as ISO8601), units (such as mmHg) and rating scales (such as Likert scales), and remove outliers through cleaning technology. Popular science videos: Extract keyframe text (OCR) and speech-to-text (ASR), and annotate semantic tags in combination with medical ontology libraries (such as SNOMEDCT).

[0120] Secondly, knowledge elements must be extracted. These are divided into structured and unstructured elements. For structured elements, diagnostic codes (ICDs) and medication records (NDCs) are extracted from electronic medical records; symptom scores are extracted from questionnaires; and disease keywords are extracted from videos. For unstructured elements, deep learning models (such as BiLSTM-CRF or BERT) are used to extract entities (disease names / drugs) and relationships. Trend features are extracted from wearable device physiological parameters through time series analysis.

[0121] Next, dynamic weight allocation and fusion are performed, and the attention mechanism is used to achieve cross-modal feature weighted fusion. Specifically, it includes: 1. Weight calculation step. Based on the similarity between the query vector (Query) and the feature vector (Key) of each modality The weight α is generated by Softmax i For example, physiological parameters (wearable devices) may be given higher weight when they are related to acute symptoms; subjective scores in questionnaires are given higher weight in long-term efficacy evaluation. 2. Feature fusion step. Weighted sum h = ∑α i e i Or after splicing, the dimensionality is reduced by the fully connected layer; cross-modal alignment can dynamically adjust the weight ratio through the gating mechanism.

[0122] Finally, unified knowledge representation is achieved. The fused multidimensional vectors can be applied in the following ways: Knowledge graph construction: Mapping entities and relationships into nodes and edges. Decision support: Integrating clinical pathway rule bases to generate treatment recommendations. Visual analysis: Highlighting the contributions of key modalities by weight. This process completes a closed loop from heterogeneous data processing to intelligent applications.

[0123] Step S3, construct a multi-layer symptom knowledge graph: perform Bayesian network modeling based on unified knowledge representation data, establish a symptom-risk-intervention association knowledge network, and use graph neural network (GNN) to achieve dynamic node expansion.

[0124] Specifically, the above operations include:

[0125] S3.1, using the multimodal fusion unified knowledge representation, perform symptom-risk-intervention association modeling and construct a Bayesian network, where nodes represent medical variables and edges represent conditional dependencies;

[0126] S3.2, initialize node probability distribution based on electronic medical record statistics;

[0127] S3.3, update the conditional probability table through variational inference to integrate new patient data;

[0128] S3.4, enabling causal reasoning through reverse diagnosis and risk prediction;

[0129] S3.5, monitoring new data, and when new data is generated, obtaining the new data;

[0130] S3.6, using a graph neural network (GNN) method to extract dynamically added medical entities from the newly added data and construct a hierarchical dynamic GNN architecture;

[0131] In S3.7, the initial embedding layer uses a Graph Attention Network to encode existing graph nodes. For the dynamic update layer, the initial embedding of new nodes is generated through neighbor sampling and aggregation. The LSTM module records the node state history, adapts to graph evolution, and realizes time-series state transfer to obtain the dynamic update layer.

[0132] S3.8, the probability distribution of BN output is input into GNN as node attributes to achieve cross-modal alignment.

[0133] Specifically, the first step is Bayesian network modeling based on unified knowledge representation, which includes: first, symptom-risk-intervention association modeling; second, GNN-driven dynamic graph expansion; and third, cross-modal alignment.

[0134] S3.9 also includes combining Bayesian GNN to deal with medical uncertainty, including:

[0135] The node classification confidence is quantified based on Dirichlet distribution to determine the uncertainty of evidence.

[0136] Determine probabilistic uncertainty by simulating the prediction variance under parameter perturbations using Monte Carlo Dropout.

[0137] Specifically, the Dirichlet distribution is a conjugate prior for categorical distributions and is naturally suitable for modeling epistemic uncertainty in multi-class classification problems. Its core idea is to regard the output of the neural network as the accumulation of evidence for each category and reflect the confidence level through the parameters of the Dirichlet distribution. The specific implementation is as follows:

[0138] 1. Evidence collection and parameterization. Replace the Softmax of the last layer of the neural network with a non-negative activation function (such as ReLU) to directly output the evidence strength e of each category k And transformed into the parameter a of Dirichlet distribution k =e k +1, where a k The larger the value, the more evidence there is for that category. 2. Calculation of uncertainty. Overall uncertainty: by using the differential entropy of the Dirichlet distribution or the total pseudo count S' = ∑a k If S' is small or the entropy is high, it indicates that the model has insufficient confidence in the classification of the sample. Evidence uncertainty: Use the covariance determinant or the degree of dispersion of the probability distribution between categories to identify out-of-domain samples. For example, when the input deviates from the training distribution, the Dirichlet parameter tends to be uniform, resulting in high entropy.

[0139] Specifically, Monte Carlo Dropout captures the accidental uncertainty caused by model parameter perturbations by generating samples of the predicted distribution by enabling the Dropout layer multiple times during the inference phase. The implementation steps include:

[0140] Perform T forward propagations on the same input and randomly discard some neurons each time to obtain different output probabilities. Compute the mean as the final prediction variance as a measure of uncertainty: represents the average probability, p t represents the input probability, where high variance areas correspond to high-risk predictions that are sensitive to model parameters.

[0141] Dirichlet captures epistemic uncertainty (the limitations of model knowledge), while MC Dropout reflects aleatory uncertainty (noise in the data). Combining the two allows for comprehensive risk assessment. For example, in radiotherapy side effect prediction, the former identifies unknown symptoms, while the latter quantifies individual differences. Combining the two approaches can construct a multi-level uncertainty assessment framework, such as in head and neck cancer patient management platforms. Dirichlet filters low-confidence symptom nodes to avoid over-recommendations, while MC Dropout generates personalized risk intervals to assist in dynamically adjusting intervention plans.

[0142] S3.10, based on end-to-end training, performs joint optimization, including:

[0143] The loss function jointly optimizes the likelihood estimation of BN and the node classification loss of GNN. The formula of the joint loss function L is as follows:

[0144] L=λ1L que +λ2L dis ,

[0145] Among them, L que For the symptom query loss, the negative log-likelihood loss of the Bayesian network is used, L dis is the disease diagnosis loss, using cross entropy loss, λ1 and λ2 represent the dynamic weight coefficients of the corresponding loss function,

[0146] L que = -logP(X|Θ),

[0147] Among them, Θ is the conditional probability table, X is the symptom observation data;

[0148]

[0149] Among them, y i For head and neck cancer classification, Predict probability for GNN;

[0150] Adaptive adjustment is performed by dynamically balancing task contributions using gradient statistics.

[0151] Specifically, in the management of head and neck cancer radiotherapy: Symptom inquiry stage L que : Estimating the probability of radiation mucositis; Diagnostic stage L dis :GNN combines imaging features to predict tumor progression risk; dynamic feedback: adjust λ1 / λ2 based on doctor feedback to optimize subsequent recommendations.

[0152] End-to-end training is implemented: BN and GNN share the underlying feature extraction layer (such as symptom embedding) and transmit symptom relevance through the graph attention mechanism (GAT); the output of BN (such as disease probability distribution) is used as an additional feature of the GNN node.

[0153] S3.11 uses the AdamW optimizer, whose adaptive learning rate feature is suitable for handling the gradient difference of the joint loss:

[0154]

[0155] Among them, θ t+1 ,θ t Represent the gradient differences in t+1 and t cycles respectively, are the first-order and second-order moment estimates of the gradient, respectively.

[0156] Step S4, building a clinical decision support system through the patient-side intelligent head and neck cancer radiotherapy patient self-management platform, recording symptoms through the clinical decision support system, conducting human-computer interaction based on the multi-layer symptom knowledge graph, and performing cross-layer recommendations based on the collaborative filtering method.

[0157] Specifically, the above steps include:

[0158] Construct a hybrid cloud architecture, wherein the hybrid cloud architecture includes a patient-side intelligent head and neck cancer radiotherapy patient self-management platform; for example, Figure 2 As shown in the figure, the self-management platform module for head and neck cancer radiotherapy patients includes the patient side and the medical side. Among them, patients use the patient side to manage symptoms, nutrition, and medical and patient management. Figure 4 A schematic diagram of human-machine interaction in decision-making assistance provided by an embodiment of the present application is shown in FIG. Figure 4 As shown, it is used to make decisions based on the input questions.

[0159] Specifically, information is input through the multimodal interactive terminal of the patient-side intelligent head and neck cancer radiotherapy patient self-management platform, including voice input, text input, scale selection, wearable device synchronization and image upload;

[0160] Embed the Bayesian-GNN multi-layer symptom graph generated in step S3;

[0161] A cross-layer patient-symptom matrix decomposition algorithm is used for cross-layer collaborative filtering recommendation, where cross-layer collaborative filtering recommendation includes horizontal same-level collaboration and vertical cross-level collaboration.

[0162] The cross-layer collaborative filtering recommendation using a cross-layer patient-symptom matrix decomposition algorithm includes:

[0163] Construct an m×n patient-symptom matrix:

[0164]

[0165] Where m represents the number of patients, n represents the number of symptoms, V ki Indicates the patient's matching degree to the symptoms;

[0166] According to the patient-symptom matrix V, the similarity S(i,j)1 between symptom i and symptom j is calculated using the following formula:

[0167]

[0168] Among them, N(a) represents the set of patients that match both symptom i and symptom j, N(i) and N(j) represent the sets of patients that match symptom i and symptom j respectively;

[0169] Calculate the similarity S(i,j)2 between symptom i and symptom j in the knowledge graph. The formula is as follows:

[0170]

[0171] Among them, x i 、x j Respectively represent the entity of symptoms in the multimodal fusion unified knowledge, y i 、y j Respectively represent the relationship between symptoms in the multimodal fusion unified knowledge, z i 、z j They represent the attributes of symptoms in the unified knowledge of multimodal fusion;

[0172] The similarity S(i,j)1 and the similarity S(i,j)2 are combined to obtain the total similarity S(i,j), which is expressed as:

[0173] S(i,j)=λS(i,j)1+(1-λ)S(i,j)2,

[0174] Among them, λ represents the weight of similarity;

[0175] Filter out the top N most matching symptom sets based on the total similarity S(i,j);

[0176] Based on the Bayesian-GNN multi-layer symptom graph generated in step S3 based on the first N symptom sets, horizontal same-level collaboration and vertical cross-level collaborative recommendation are performed.

[0177] Step S5, real-time risk stratification: Calculate the symptom severity score using the XGBoost algorithm, and output green / yellow / red warning signals and corresponding intervention paths.

[0178] Specifically, a hybrid cloud architecture is constructed, comprising private and public cloud layers, along with security mechanisms. The private cloud layer deploys a Bayesian-GNN knowledge graph engine (ensuring low latency for core algorithms); the public cloud layer handles non-sensitive tasks such as wearable device synchronization (leveraging elastic computing resources); and the security mechanism utilizes blockchain for evidence storage, ensuring traceability of data transmission, such as radiotherapy images.

[0179] The multimodal interactive terminal design includes: voice / text chief complaint: the NLP module integrating the Qwen model parses unstructured descriptions; scale selection: embedding the CTCAE5.0 standard electronic form (toxicity classification such as radiation mucositis); device synchronization: docking with Apple Watch and other devices to obtain real-time SpO2 data; image upload: supporting DICOM files to automatically segment the lesion area.

[0180] Bayesian-GNN graph embedding includes dynamic reasoning, for example, when a patient reports "painful swallowing"

[0181] During this time, the GAT network locates relevant symptom nodes (attention weight > 0.7); BN calculates the posterior probability of radiation esophagitis (if P > 80%, an alert is triggered). Incremental learning: New cases are embedded using GraphSAGE to update the graph.

[0182] The cross-layer collaborative filtering recommendation algorithm also includes matrix decomposition optimization and sparse processing: the BPR loss function is used for the patient-symptom matrix V, where

[0183] L BPR =-∑ (i,j,j') lnσ(V ij -V ij' ),

[0184] Among them, j is a positive sample (symptoms have appeared), and j' is a negative sample (no symptoms have appeared).

[0185] The hierarchical recommendation logic is shown in the following table:

[0186]

[0187] Specifically, the GAT attention mechanism is coupled with BN causal reasoning to improve interpretability; the knowledge graph semantic similarity is introduced in matrix decomposition to alleviate the cold start problem; and a balance between sensitive data processing and elastic computing is achieved through hybrid cloud.

[0188] Step S6, continuous optimization: update the knowledge graph through reinforcement learning based on patient feedback data.

[0189] Specifically, reinforcement learning can be used for framework design, where the state (S) is the confidence distribution of nodes in the current knowledge graph (derived from Dirichlet distribution) and the latest patient feedback (such as the degree of symptom relief); the action (A) is to adjust the weights of graph nodes, add associated edges, or modify the Bayesian network CPT table; and the reward (R) is dynamically calculated based on the subsequent improvement in patient efficacy (such as the proportion of pain score reduction after intervention for radiation mucositis).

[0190] The training process is implemented, for example, in the "state-action-reward" mechanism of the pediatric severe pneumonia classification model: each patient's treatment cycle is regarded as an episode, and the initial state is the current graph snapshot; the intelligent agent selects an action (such as increasing the edge weight of "dysphagia → nutritional intervention") and then executes the clinical recommendation; the next state is updated based on the patient's review data, and the reward is calculated by the efficacy indicator (such as the change in CTCAE grade); the Q function is iteratively updated through the Bellman equation to prioritize the high-reward path.

[0191] The above steps achieve the integration of key technologies, including hierarchical Bayesian optimization: assigning absolute weights to core symptoms (such as bleeding) based on the expert knowledge integration method of the DSVM-P model; federated learning to ensure privacy: after local training of the Q network, each hospital only uploads the gradient parameters to meet medical data security requirements; dynamic neural feedback: imitating the CHD diagnostic model and using the transformation matrix to filter irrelevant features to improve the purity of state representation.

[0192] Example 2

[0193] like Figure 5 As shown, the present application provides a symptom management system for head and neck cancer patients based on multi-source data fusion and AI-driven, which is applied to the symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven as described in Example 1, including: a multimodal data collection module 11, a multimodal data fusion module 12, a multi-layer symptom knowledge graph construction module 13, a multi-layer collaborative recommendation module 14, a real-time risk stratification module 15, and a continuous optimization module 16.

[0194] It can be understood that in this embodiment, the multimodal data collection module is used to collect electronic medical record data through a local database, collect sensor data through wearable devices, collect patient data and effect evaluation data through questionnaires, and collect head and neck cancer popular science video data through network resources; and organize the collected multimodal data.

[0195] It can be understood that in this embodiment, the multimodal data fusion module is used to standardize electronic medical records, wearable device data, patient questionnaires and effect evaluation data, and popular science video data and extract knowledge elements, apply the attention mechanism to dynamically assign feature weights, and fuse the weighted knowledge elements to form a unified knowledge representation.

[0196] It can be understood that in this embodiment, a multi-layer symptom knowledge graph construction module is used to perform Bayesian network modeling based on unified knowledge representation data, establish a symptom-risk-intervention association knowledge network, and use a graph neural network GNN to achieve dynamic node expansion.

[0197] It can be understood that in this embodiment, the multi-layer collaborative recommendation module is used to build a clinical decision support system through the patient-side intelligent head and neck cancer radiotherapy patient self-management platform, record symptoms through the clinical decision support system, conduct human-computer interaction based on the multi-layer symptom knowledge graph, and perform cross-layer recommendations based on the collaborative filtering method.

[0198] It can be understood that, in this embodiment, the real-time risk stratification module is used to calculate the symptom severity score through the XGBoost algorithm, and output green / yellow / red warning signals and corresponding intervention paths.

[0199] It can be understood that in this embodiment, the continuous optimization module is used to update the knowledge graph through reinforcement learning based on patient feedback data.

[0200] Figure 6 This is an electronic device provided by an embodiment of the present application. Figure 6 As shown, the electronic device includes at least the following parts: a processor 101 and a memory 100 , a communication interface 103 , and a bus 102 .

[0201] In the embodiment of the present application, the memory 100 is used to store instructions executable by the processor 101. The processor 101 is configured to execute the instructions to implement the following Figure 3 The device module shown is based on multi-source data fusion and AI-driven symptom management for head and neck cancer patients.

[0202] In an embodiment of the present application, a computer-readable storage medium includes instructions, and the instructions instruct a device to execute the method of the first aspect. For example, the instructions instruct the device to execute Figure 1 The method is shown in the process steps.

[0203] The program running in the electronic device involved in one embodiment of the present application can be a program that controls a central processing unit (CPU) and the like to realize the functions of the above-mentioned embodiment involved in one embodiment of the present invention (a program that enables a computer to function). Then, the information processed by these devices is temporarily stored in a random access memory (RAM) while being processed, and then stored in various ROMs such as read-only memory (Flash ROM) and hard disk drive (HDD), and is read, modified, and written by the CPU as needed.

[0204] It should be noted that a portion of the electronic device of the above embodiment may also be implemented by a computer. In this case, a program for implementing the control function may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer and executed.

[0205] It should be noted that the "computer" mentioned here refers to a computer built into an electronic device, employing hardware including an operating system (OS) and peripheral devices. Furthermore, "computer-readable recording medium" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard disks built into computers.

[0206] Furthermore, "computer-readable recording media" may include: media that dynamically store programs for a short period of time, such as communication lines when transmitting programs via networks such as the Internet or communication lines such as telephone lines; and media that store programs for a fixed period of time, such as volatile memory within computers acting as servers or clients in this context. Furthermore, the aforementioned program may be a program for implementing a portion of the aforementioned functions, or a program that can achieve the aforementioned functions by combining with a program already stored in a computer.

[0207] Furthermore, the electronic device in the above-described embodiments can also be implemented as a collection (device group) consisting of multiple devices. Each device constituting the device group may have a portion or all of the functions or functional blocks of the electronic device in the above-described embodiments. A device group only needs to have all the functions or functional blocks of the electronic device.

[0208] Those skilled in the art should recognize that the above embodiments are merely intended to illustrate the present application and are not intended to limit the present application. As long as they are within the spirit of the present application, appropriate changes and modifications to the above embodiments are within the scope of protection claimed in the present application.

Claims

1. A symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven, characterized by: The method comprises: Step S1, multimodal data collection: electronic medical record data is collected through a local database, sensor data is collected through wearable devices, patient data and effect evaluation data are collected through questionnaires, and head and neck cancer popular science video data is collected through online resources; the collected multimodal data is organized; Step S2, multimodal data fusion: Standardize and extract knowledge elements from electronic medical records, wearable device data, patient questionnaires and effect evaluation data, and popular science video data. Apply an attention mechanism to dynamically assign feature weights, fuse the weighted knowledge elements, and form a unified knowledge representation. Step S3: Construct a multi-layer symptom knowledge graph: Based on the unified knowledge representation data, Bayesian network modeling is performed to establish a symptom-risk-intervention association knowledge network, and a graph neural network (GNN) is used to achieve dynamic node expansion. Step S4, collaborative recommendation: A clinical decision support system is constructed through the patient-side intelligent head and neck cancer radiotherapy patient self-management platform. Symptoms are recorded through the clinical decision support system, and human-computer interaction is performed based on a multi-layer symptom knowledge graph. Cross-layer recommendations are then performed based on collaborative filtering methods. Step S5, real-time risk stratification: Calculate the symptom severity score using the XGBoost algorithm and output green / yellow / red warning signals and corresponding intervention paths; Step S6, continuous optimization: update the knowledge graph through reinforcement learning based on patient feedback data.

2. The method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven according to claim 1, characterized in that: In step S1, the collected multimodal data is sorted, including: Using a data warehouse to store and manage the collected multimodal data; Use data pre-processing techniques to remove noise and outliers, including duplicate, irrelevant, inaccurate, or outdated data; Classify and label the data.

3. The method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven according to claim 2, characterized in that: Step S2, multimodal data fusion: standardizes electronic medical records, wearable device data, patient questionnaires and effect evaluation data, and popular science video data and extracts knowledge elements. The attention mechanism is used to dynamically assign feature weights and fuse the weighted knowledge elements to form a unified knowledge representation, including: S2.1: Electronic medical records are encoded using international standards and structured metadata. Unstructured text is identified and terminology standardized using the BERT deep learning model. Private information is desensitized. Wearable devices, questionnaires, and evaluation forms are standardized in timestamp format, units, and Likert rating scales, and outliers are removed using cleaning techniques. Keyframe text is extracted from popular science videos using optical character recognition (OCR), speech-to-text data is converted, and semantic tags are annotated using a medical ontology library. S2.2: For structured elements, symptoms, diagnosis codes, and medication records are extracted from electronic medical records; symptom scores are extracted from questionnaires; and disease keywords are extracted from videos. For unstructured elements, entities and relationships are extracted using the BERT deep learning model; and trend features are extracted from physiological parameters of wearable devices through time series analysis. S2.3, using the attention mechanism to achieve cross-modal feature weighted fusion, based on the similarity between the query vector query and the feature vector key of each modality, the weight α is generated by softmax i , where the similarity S based on the attention mechanism att The formula is: Among them, Q is the query vector, K is the key, and V is the key value; S2.4, weighted sum h = ∑α i e i , after splicing, the dimension is reduced by the fully connected layer, and the weight ratio is dynamically adjusted through the gating mechanism to align across modalities, where e i The corresponding feature sequence; S2.5, the fused multi-dimensional vector is used as the multimodal fusion unified knowledge representation.

4. The method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven according to claim 3, characterized in that: Step S3, constructing a multi-layer symptom knowledge graph: performing Bayesian network modeling based on unified knowledge representation data, establishing a symptom-risk-intervention association knowledge network, and using a graph neural network (GNN) to achieve dynamic node expansion, including: S3.1, using the multimodal fusion unified knowledge representation, perform symptom-risk-intervention association modeling and construct a Bayesian network, where nodes represent medical variables and edges represent conditional dependencies; S3.2, initialize node probability distribution based on electronic medical record statistics; S3.3, update the conditional probability table through variational inference to integrate new patient data; S3.4, enabling causal reasoning through reverse diagnosis and risk prediction; S3.5, monitoring new data, and when new data is generated, obtaining the new data; S3.6, using a graph neural network (GNN) method to extract dynamically added medical entities from the newly added data and construct a hierarchical dynamic GNN architecture; In S3.7, the initial embedding layer uses a Graph Attention Network to encode existing graph nodes. For the dynamic update layer, the initial embedding of new nodes is generated through neighbor sampling and aggregation. The LSTM module records the node state history, adapts to graph evolution, and realizes time-series state transfer to obtain the dynamic update layer. S3.8, the probability distribution of BN output is input into GNN as node attributes to achieve cross-modal alignment.

5. The method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven according to claim 4, characterized in that: S3.9 also includes combining Bayesian GNN to deal with medical uncertainty, including: Quantify the node classification confidence based on Dirichlet distribution and determine the uncertainty of evidence; Determine probabilistic uncertainty by simulating the prediction variance under parameter perturbations using Monte Carlo Dropout. S3.10, based on end-to-end training, performs joint optimization, including: The loss function jointly optimizes the likelihood estimation of BN and the node classification loss of GNN. The formula of the joint loss function L is as follows: L=λ1L que +λ2L dis , Among them, L que For the symptom query loss, the negative log-likelihood loss of the Bayesian network is used, L dis is the disease diagnosis loss, using cross entropy loss, λ1 and λ2 represent the dynamic weight coefficients of the corresponding loss function, L que =-logP(X|Θ), Among them, Θ is the conditional probability table, X is the symptom observation data; Among them, y i For head and neck cancer classification, Predict probability for GNN; Adaptive adjustment is performed by dynamically balancing task contributions using gradient statistics; S3.11 uses the AdamW optimizer, whose adaptive learning rate feature is suitable for handling the gradient difference of the joint loss: Among them, θ t+1 ,θ t Represent the gradient differences at t+1 and t times respectively, are the first-order and second-order moment estimates of the gradient, respectively.

6. The method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven according to claim 1, characterized in that: Step S4, constructing a clinical decision support system through the patient-side intelligent head and neck cancer radiotherapy patient self-management platform, recording symptoms through the clinical decision support system, conducting human-computer interaction based on a multi-layer symptom knowledge graph, and performing cross-layer recommendations based on a collaborative filtering method, includes: Building a hybrid cloud architecture, including an intelligent patient self-management platform for head and neck cancer radiotherapy; Information input is performed through the multimodal interactive terminal of the patient-side intelligent head and neck cancer radiotherapy patient self-management platform, including voice input, text input, scale selection, wearable device synchronization and image upload; Embed the Bayesian-GNN multi-layer symptom graph generated in step S3; A cross-layer patient-symptom matrix decomposition algorithm is used for cross-layer collaborative filtering recommendation, where cross-layer collaborative filtering recommendation includes horizontal same-level collaboration and vertical cross-level collaboration.

7. The method for symptom management of head and neck cancer patients based on multi-source data fusion and AI-driven according to claim 6, characterized in that: The cross-layer collaborative filtering recommendation using a cross-layer patient-symptom matrix decomposition algorithm includes: Construct an m×n patient-symptom matrix: Where m represents the number of patients, n represents the number of symptoms, V ki Indicates the patient's matching degree to the symptoms; According to the patient-symptom matrix V, the similarity S(i,j)1 between symptom i and symptom j is calculated using the following formula: Among them, N(a) represents the set of patients that match both symptom i and symptom j, N(i) and N(j) represent the sets of patients that match symptom i and symptom j respectively; Calculate the similarity S(i,j)2 between symptom i and symptom j in the knowledge graph. The formula is as follows: Among them, x i 、x j Respectively represent the entity of symptoms in the multimodal fusion unified knowledge, y i 、y j Respectively represent the relationship between symptoms in the multimodal fusion unified knowledge, z i 、z j They represent the attributes of symptoms in the unified knowledge of multimodal fusion; The similarity S(i,j)1 and the similarity S(i,j)2 are combined to obtain the total similarity S(i,j), which is expressed as: S(i,j)=λS(i,j)1+(1-λ)S(i,j)2, Among them, λ represents the weight of similarity; Filter out the top N most matching symptom sets based on the total similarity S(i,j); Based on the Bayesian-GNN multi-layer symptom graph generated in step S3 based on the first N symptom sets, horizontal same-level collaboration and vertical cross-level collaborative recommendation are performed.

8. A symptom management system for head and neck cancer patients based on multi-source data fusion and AI-driven, applied to the symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven according to any one of claims 1 to 7, characterized in that: include: A multimodal data collection module is used to collect electronic medical record data through a local database, sensor data through wearable devices, patient data and effect evaluation data through questionnaires, and head and neck cancer popular science video data through online resources; and organize the collected multimodal data; The multimodal data fusion module is used to standardize and extract knowledge elements from electronic medical records, wearable device data, patient questionnaires and effect evaluation data, and popular science video data. It uses the attention mechanism to dynamically assign feature weights and fuse the weighted knowledge elements to form a unified knowledge representation. A multi-layer symptom knowledge graph construction module is used to perform Bayesian network modeling based on unified knowledge representation data, establish a symptom-risk-intervention association knowledge network, and use graph neural network (GNN) to achieve dynamic node expansion; A multi-layer collaborative recommendation module is used to build a clinical decision-making support system through the patient-side intelligent head and neck cancer radiotherapy patient self-management platform. Symptoms are recorded through the clinical decision-making support system, and human-computer interaction is carried out based on the multi-layer symptom knowledge graph. Cross-layer recommendations are also made based on collaborative filtering methods. A real-time risk stratification module, which uses the XGBoost algorithm to calculate symptom severity scores and output green / yellow / red warning signals and corresponding intervention pathways; A continuous optimization module for updating the knowledge graph through reinforcement learning based on patient feedback data.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement the symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven as described in any one of claims 1 to 7 when executing the instructions.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, and the program instructs the device to execute the symptom management method for head and neck cancer patients based on multi-source data fusion and AI driving as described in any one of claims 1 to 7.

Images