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

Abstract

This application discloses a symptom management system and method for head and neck cancer patients based on multi-source data fusion and AI-driven approaches. The system includes: a data acquisition and fusion module for collecting multi-source data on symptoms of head and neck cancer patients from multiple dimensions, preprocessing the collected multi-source data, and fusing the preprocessed multi-source data to form fused symptom data; a model prediction module for inputting the fused symptom data into a trained deep learning model to obtain the symptom prediction results output by the trained deep learning model; and a symptom intervention and management module for intervening in and managing the symptoms of head and neck cancer patients based on the symptom prediction results, forming a closed-loop management system for the symptoms of head and neck cancer patients. This application addresses the problem of poor treatment outcomes in head and neck cancer patients due to the inability to promptly detect and intervene in their symptoms.

Description

Technical Field

[0001] This application relates to the field of head and neck cancer technology, specifically to a symptom management system and method for head and neck cancer patients based on multi-source data fusion and AI-driven approaches. Background Technology

[0002] Head and neck cancers include oral cancer, pharyngeal cancer, laryngeal cancer, nasal cancer, and salivary gland cancer, among others. The clinical symptoms vary depending on the specific disease. 1) Oral cancer: When a patient develops oral cancer, it can cause lesions in the oral mucosa and other tissues, leading to discomfort symptoms such as oral mucosal erythema or leukoplakia, oral mucosal ulcers, erosion, bleeding, severe pain, and loose teeth. 2) Pharyngeal cancer: When cancer cells invade the pharyngeal tissue, they can cause lesions in the pharyngeal tissue, leading to symptoms such as difficulty speaking, difficulty swallowing, neck pain, headache, tinnitus, and hearing loss. 3) Laryngeal cancer: When a patient has laryngeal cancer, cancer cells can invade the laryngeal tissue, causing cancerous lesions and resulting in discomfort symptoms such as pain when swallowing, ear pain, and hoarseness. 4) Nasal cancer: When cancer cells invade the nasal tissues of a patient and continue to grow in that area, they can form a mass, causing discomfort symptoms such as nasal congestion, nasal bleeding, and decreased vision. 5) Salivary gland cancer: When cancer cells grow in a patient's salivary glands, it may cause symptoms such as swelling around the jaw, facial numbness, muscle numbness, and facial and neck pain. These symptoms not only seriously affect the patient's quality of life, but may also lead to treatment interruption and affect the prognosis. Therefore, symptom management is particularly important for patients with head and neck cancer.

[0003] Current technologies cannot detect symptoms in head and neck cancer patients in a timely manner, nor can they intervene and manage these symptoms promptly, resulting in poor treatment outcomes and reduced quality of life for head and neck cancer patients. Summary of the Invention

[0004] The purpose of this application is to provide a symptom management system for head and neck cancer patients based on multi-source data fusion and AI-driven technology. This system can promptly detect symptoms in head and neck cancer patients and provide timely intervention and management, thereby improving the treatment outcomes and quality of life for head and neck cancer patients and solving the problems mentioned in the background art.

[0005] To achieve the above objectives, this application provides the following technical solution: The head and neck cancer patient symptom management system, based on multi-source data fusion and AI-driven technology, includes: a data acquisition and fusion module, used to collect multi-source data on head and neck cancer patient symptoms from multiple dimensions, preprocess the collected multi-source data, and fuse the preprocessed multi-source data to form fused data on head and neck cancer patient symptoms; a model prediction module, used to input the fused data on head and neck cancer patient symptoms into a trained deep learning model to obtain the predicted symptom results of the trained deep learning model, wherein the deep learning model is trained using historical multi-source data on head and neck cancer patient symptoms and corresponding historical head and neck cancer patient symptoms; and a symptom intervention and management module, used to intervene and manage the symptoms of head and neck cancer patients based on the predicted symptom results, forming a closed-loop management system for head and neck cancer patient symptoms.

[0006] Preferably, multi-source data on symptoms of head and neck cancer patients are collected from multiple dimensions, including: collecting electronic medical records and laboratory test data of head and neck cancer patients to obtain clinical data; collecting standardized scales and active symptom records of head and neck cancer patients to obtain reported data; collecting activity levels, physiological parameters, behavioral parameters, and environmental data of head and neck cancer patients to obtain behavioral data; and determining multi-source data on symptoms of head and neck cancer patients based on clinical data, reported data, and behavioral data.

[0007] Preferably, the electronic medical record includes the diagnostic staging, treatment plan information, pathology report information, and / or comorbidity information of the head and neck cancer patient; laboratory test data includes complete blood count, liver and kidney function, nutritional indicators, and / or imaging data; and / or standardized scales include a head and neck cancer symptom checklist, a quality of life scale, and / or a pain numerical rating scale; active symptom records include symptom information on pain, swallowing, and / or sleep recorded by the head and neck cancer patient; and / or activity levels include the number of steps and / or duration of activity of the head and neck cancer patient; physiological parameters include the resting heart rate, heart rate variability, and / or sleep quality of the head and neck cancer patient; behavioral parameters include the dietary intake, medication adherence, and / or oral care frequency of the head and neck cancer patient; and environmental data includes environmental noise and / or light exposure in the living environment of the head and neck cancer patient.

[0008] Preferably, the collected multi-source data on symptoms of head and neck cancer patients undergoes preprocessing, including: cleaning the multi-source data on symptoms of head and neck cancer patients to remove noise; identifying missing values ​​and outliers in the multi-source data on symptoms of head and neck cancer patients, evaluating the missing values ​​and outliers, identifying missing values ​​and outliers valuable for symptom management based on the evaluation results, performing numerical imputation on the multi-source data on symptoms of head and neck cancer patients based on the missing values ​​valuable for symptom management, and performing numerical replacement on the outliers valuable for symptom management; identifying missing values ​​and outliers of no value for symptom management based on the evaluation results, and deleting the missing values ​​and outliers of no value for symptom management from the multi-source data on symptoms of head and neck cancer patients; and obtaining preprocessed multi-source data on symptoms of head and neck cancer patients based on the cleaned, imputed, replaced, and deleted multi-source data on symptoms of head and neck cancer patients.

[0009] Preferably, the evaluation of missing values ​​and outliers in the multi-source data of head and neck cancer patient symptoms includes: if any missing value in the multi-source data of head and neck cancer patient symptoms is identified as a completely random missing value, then the missing value in the multi-source data of head and neck cancer patient symptoms is evaluated as having no value for the management of head and neck cancer patient symptoms; if any missing value in the multi-source data of head and neck cancer patient symptoms is identified as a non-random missing value, then the missing value in the multi-source data of head and neck cancer patient symptoms is evaluated as having value for the management of head and neck cancer patient symptoms; if any outlier in the multi-source data of head and neck cancer patient symptoms is identified as an error-type outlier, then the outlier in the multi-source data of head and neck cancer patient symptoms is evaluated as having no value for the management of head and neck cancer patient symptoms; if any outlier in the multi-source data of head and neck cancer patient symptoms is identified as a trend-type outlier, then the outlier in the multi-source data of head and neck cancer patient symptoms is evaluated as having value for the management of head and neck cancer patient symptoms.

[0010] Preferably, data fusion is performed on the preprocessed multi-source data of head and neck cancer patient symptoms, including: dividing the preprocessed multi-source data of head and neck cancer patient symptoms into a core layer, a key layer, and an auxiliary layer according to the clinical priority and data characteristics of each source data, and configuring dynamic clinical weights for the sub-data of each layer; performing data modality segmentation on the sub-data of each layer to obtain clinical modality data, report modality data, and behavioral modality data; extracting features from various modal data to obtain multiple modality features, including clinical modality features, report modality features, and behavioral modality features; calculating the similarity correlation degree between modality features within the same modality, and calculating the clinical correlation degree between modality features across modalities, and merging the data based on the similarity correlation degree and clinical correlation degree, as well as based on multiple modality features. A feature association network is constructed based on modal features. The clinical contribution value of each modal feature is calculated based on the feature association network and the dynamic clinical weights of the sub-data at each layer. Multiple high-value features are then selected from multiple modal features based on their clinical contribution values. The dynamic reliability value of each sub-data is obtained, and the symptom hypothesis reliability of each high-value feature is determined based on the dynamic reliability value and the dynamic clinical weights of each sub-data. Multiple high-value features are fused based on their symptom hypothesis reliability to obtain multiple first fused features. First fused features of the same modality are fused to obtain second fused features. Second fused features of different modalities are fused to obtain third fused features. The symptom fusion data of head and neck cancer patients includes the third fused features.

[0011] Preferably, the core layer includes electronic medical record sub-data and laboratory test sub-data; the key layer includes standardized scale sub-data, active symptom record sub-data, and physiological parameter sub-data; the auxiliary layer includes activity level sub-data, behavioral parameter sub-data, and environmental data sub-data; the clinical modality data includes the sub-data of the core layer and the physiological parameter sub-data of the key layer; the reporting modality data includes the standardized scale sub-data and active symptom record sub-data of the key layer; and the behavioral modality data includes the sub-data of the auxiliary layer.

[0012] Preferably, the model prediction module is further configured to: obtain multiple symptoms and multiple data features from the symptom fusion data of head and neck cancer patients, and construct a symptom-feature association graph based on the multiple symptoms and multiple data features; determine the correlation coefficient between adjacent nodes in the symptom-feature association graph according to the symptom fusion data of head and neck cancer patients, and obtain an adjacency matrix; input the graph node features and adjacency matrix in the symptom-feature association graph into a trained deep learning model; wherein, the deep learning model includes a first-layer multi-head attention graph network, a second-layer multi-head attention graph network, and a dynamic weight adaptive prediction head connected thereto, the dynamic weight adaptive prediction head being constructed based on a weight learning network and a risk prediction network, both of which are constructed based on a 3-layer fully connected network.

[0013] Preferably, the head and neck cancer patient symptom management system based on multi-source data fusion and AI-driven further includes a model training module, used for: collecting historical symptom data of head and neck cancer patients and constructing a training set, the training set including clinical historical data, historical report data of head and neck cancer patients, and historical behavioral data; extracting features from the training set, fusing the extracted features to obtain training fusion features; using historical symptoms and historical features in the training fusion features as nodes, calculating the correlation coefficient between nodes based on the historical data in the training set as edge weights and forming a historical adjacency matrix, and constructing a historical symptom-feature association graph; inputting the graph node features and historical adjacency matrix in the historical symptom-feature association graph into a first-layer multi-head attention graph attention network to obtain historical graph features; inputting the historical graph features into a second-layer multi-head attention graph attention network to obtain a historical symptom feature matrix; constructing a historical feature vector based on the historical features; inputting the historical symptom feature matrix into the risk prediction network of the dynamic weight adaptive prediction head, and inputting the historical feature vector into the weight learning network of the dynamic weight adaptive prediction head, the adaptive prediction head of the dynamic weight adaptive prediction head combining the output of the risk prediction network and the output of the weight learning network to determine the symptom prediction result of head and neck cancer patients.

[0014] A method for managing symptoms in head and neck cancer patients based on multi-source data fusion and AI-driven approaches includes: collecting multi-source data on symptoms of head and neck cancer patients from multiple dimensions; preprocessing the collected multi-source data; fusing the preprocessed multi-source data to form fused symptom data; inputting the fused symptom data into a trained deep learning model to obtain the symptom prediction results output by the trained deep learning model, wherein the deep learning model is trained using historical multi-source data on head and neck cancer patients' symptoms and corresponding historical symptom data; and intervening in and managing the symptoms of head and neck cancer patients based on the symptom prediction results, thus forming a closed-loop management system for head and neck cancer patient symptoms.

[0015] Compared with the prior art, the beneficial effects of this application are: This application collects multi-source data on symptoms of head and neck cancer patients from multiple dimensions, preprocesses and fuses the collected multi-source data to form fused data on symptoms of head and neck cancer patients. A constructed symptom prediction model (deep learning model) is used to analyze this fused data and automatically predict the symptoms of head and neck cancer patients. For example, it can predict the risk of severe oral mucositis, pain, and dysphagia in head and neck cancer patients within the next week. Based on the symptom prediction results, intervention and management of head and neck cancer patient symptoms are implemented, forming a closed-loop health management system for head and neck cancer patient symptoms. This system can promptly detect and intervene in head and neck cancer patient symptoms, improving treatment outcomes and quality of life for head and neck cancer patients. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the head and neck cancer patient symptom management system based on multi-source data fusion and AI-driven according to this application; Figure 2 This is a flowchart illustrating the symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven approach proposed in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] This application aims to address the problems of existing methods that fail to detect and manage symptoms in head and neck cancer patients in a timely manner, leading to poor treatment outcomes and reduced quality of life for these patients. It provides a head and neck cancer patient symptom management system based on multi-source data fusion and AI-driven approaches. Figure 1 As shown, the symptom management system for head and neck cancer patients based on multi-source data fusion and AI-driven includes: a data acquisition and fusion module 101, a model prediction module 102, and a symptom intervention management module 103.

[0019] The data acquisition and fusion module 101 is used to collect multi-source data on symptoms of head and neck cancer patients from multiple dimensions, preprocess the collected multi-source data on symptoms of head and neck cancer patients, and fuse the preprocessed multi-source data on symptoms of head and neck cancer patients to form fused data on symptoms of head and neck cancer patients.

[0020] In this embodiment, multi-source data on symptoms of head and neck cancer patients are collected from multiple dimensions, including: collecting electronic medical records and laboratory test data of head and neck cancer patients to obtain clinical data; collecting standardized scales and active symptom records of head and neck cancer patients to obtain reported data; collecting activity levels, physiological parameters, behavioral parameters, and environmental data of head and neck cancer patients to obtain behavioral data; and determining multi-source data on symptoms of head and neck cancer patients based on clinical data, reported data, and behavioral data.

[0021] The electronic medical records include the diagnosis and staging, treatment plan information, pathology report information and / or comorbidity information of head and neck cancer patients; laboratory test data include complete blood count, liver and kidney function, nutritional indicators and / or imaging data; and / or standardized scales include a head and neck cancer symptom checklist, a quality of life scale and / or a pain numerical rating scale; active symptom records include symptom information on pain, swallowing and / or sleep recorded by head and neck cancer patients; and / or activity levels include the number of steps and / or activity duration of head and neck cancer patients; physiological parameters include resting heart rate, heart rate variability and / or sleep quality of head and neck cancer patients; behavioral parameters include dietary intake, medication adherence and / or oral care frequency of head and neck cancer patients; and environmental data include environmental noise and / or light exposure in the living environment of head and neck cancer patients.

[0022] Specifically, electronic medical records and laboratory test data of head and neck cancer patients are collected to obtain clinical data of head and neck cancer patients. The electronic medical records include the diagnosis and staging, treatment plan, pathology report and comorbidities of head and neck cancer patients. The laboratory test data include blood routine, liver and kidney function, nutritional indicators and imaging data (CT, MRI, PET-CT). Standardized scales and active symptom records were collected from patients with head and neck cancer to obtain their reported data. The standardized scales were lists of head and neck cancer symptoms, quality of life scales, or pain numerical rating scales that were regularly pushed through an app or mini-program. The active symptom records were daily changes in pain, swallowing, and sleep recorded by patients at any time. Activity levels, physiological parameters, behavioral parameters, and environmental data of patients with head and neck cancer were collected. The behavioral data of patients with head and neck cancer included the number of steps and duration of activity, the physiological parameters included resting heart rate, heart rate variability, and sleep quality, the behavioral parameters included dietary intake, medication adherence, and oral care frequency, and the environmental data included the ambient noise and light levels in the living environment of patients with head and neck cancer. Based on clinical data, reported data, and behavioral data of head and neck cancer patients, multi-source data on symptoms of head and neck cancer patients were identified.

[0023] In this embodiment, the collected multi-source data on symptoms of head and neck cancer patients undergoes preprocessing, including: cleaning the multi-source data to remove noise; identifying missing and outlier values ​​in the multi-source data, evaluating these values, identifying valuable missing and outlier values ​​for symptom management based on the evaluation results, performing numerical imputation on the valuable missing values, and performing numerical replacement on the valuable outlier values; identifying and deleting valueless missing and outlier values ​​from the multi-source data; and obtaining preprocessed multi-source data on symptoms of head and neck cancer patients based on the cleaned, imputed, replaced, and deleted data.

[0024] Preferably, the evaluation of missing values ​​and outliers in the multi-source data of head and neck cancer patient symptoms includes: if any missing value in the multi-source data of head and neck cancer patient symptoms is identified as a completely random missing value, then the missing value in the multi-source data of head and neck cancer patient symptoms is evaluated as having no value for the management of head and neck cancer patient symptoms; if any missing value in the multi-source data of head and neck cancer patient symptoms is identified as a non-random missing value, then the missing value in the multi-source data of head and neck cancer patient symptoms is evaluated as having value for the management of head and neck cancer patient symptoms; if any outlier in the multi-source data of head and neck cancer patient symptoms is identified as an error-type outlier, then the outlier in the multi-source data of head and neck cancer patient symptoms is evaluated as having no value for the management of head and neck cancer patient symptoms; if any outlier in the multi-source data of head and neck cancer patient symptoms is identified as a trend-type outlier, then the outlier in the multi-source data of head and neck cancer patient symptoms is evaluated as having value for the management of head and neck cancer patient symptoms.

[0025] In this embodiment, the multi-source data on symptoms of head and neck cancer patients is cleaned to remove noise and reduce interference from noisy data in symptom management. Each data point in the multi-source data is examined to identify missing and outlier values, and these values ​​are evaluated. Specifically, the evaluation of missing and outlier values ​​is based on the missing value pattern and anomaly type to determine their value in symptom management for head and neck cancer patients, thus establishing the data evaluation results for missing and outlier values. If the missing values ​​in the multi-source data of head and neck cancer patients' symptoms are in a completely random missing pattern, then the missing values ​​in the multi-source data of head and neck cancer patients' symptoms are considered to have no value for the management of head and neck cancer patients' symptoms. If the missing values ​​in the multi-source data of head and neck cancer patients' symptoms are in a non-random missing pattern, then the missing values ​​in the multi-source data of head and neck cancer patients' symptoms are considered to have value for the management of head and neck cancer patients' symptoms. If the outliers in the multi-source data of head and neck cancer patients' symptoms are of the error type, then the outliers in the multi-source data of head and neck cancer patients' symptoms are considered to have no value for the management of head and neck cancer patients' symptoms. If the outliers in the multi-source data of head and neck cancer patients' symptoms are of the trend type, then the outliers in the multi-source data of head and neck cancer patients' symptoms are considered to have value for the management of head and neck cancer patients' symptoms.

[0026] Specifically, missing values ​​and outliers in multi-source data on symptoms of head and neck cancer patients were evaluated based on missing patterns and outlier types to determine their value in managing symptoms of head and neck cancer patients. The evaluation results for missing values ​​and outliers are shown in Table 1. Table 1: Results of Data Evaluation and Judgment for Missing Values ​​and Outliers

[0027] In this embodiment, the missing and outlier values ​​in the multi-source data of head and neck cancer patient symptoms are processed based on the data evaluation results of missing and outlier values. This includes: when the missing and outlier values ​​in the multi-source data of head and neck cancer patient symptoms are valuable for the management of head and neck cancer patient symptoms, interpolation is used to fill in the missing values ​​in the multi-source data of head and neck cancer patient symptoms, and the median is used to replace the outlier values ​​in the multi-source data of head and neck cancer patient symptoms; when the missing and outlier values ​​in the multi-source data of head and neck cancer patient symptoms are not valuable for the management of head and neck cancer patient symptoms, the missing and outlier values ​​in the multi-source data of head and neck cancer patient symptoms are deleted.

[0028] Specifically, based on the data evaluation results of missing and outlier values, missing and outlier values ​​in the multi-source data of symptoms of head and neck cancer patients were processed. The processing results of missing and outlier values ​​are shown in Table 2. Table 2: Results of handling missing and outlier values

[0029] Therefore, by preprocessing the multi-source data on symptoms of head and neck cancer patients, the data quality of the multi-source data on symptoms of head and neck cancer patients can be improved.

[0030] In this embodiment, data fusion is performed on preprocessed multi-source data of head and neck cancer patient symptoms, including: dividing the preprocessed multi-source data of head and neck cancer patient symptoms into a core layer, a key layer, and an auxiliary layer according to the clinical priority and data characteristics of each source data, and configuring dynamic clinical weights for the sub-data of each layer; performing data modality segmentation on the sub-data of each layer to obtain clinical modality data, report modality data, and behavioral modality data; extracting features from various modality data to obtain multiple modality features, including clinical modality features, report modality features, and behavioral modality features; calculating the similarity correlation degree between modality features within the same modality, and calculating the clinical correlation degree between modality features across modalities, and merging the data based on the similarity correlation degree and clinical correlation degree, as well as the multi-source data. A feature association network is constructed based on the modal features. The clinical contribution value of each modal feature is calculated based on the feature association network and the dynamic clinical weights of the sub-data at each layer. Multiple high-value features are then selected from the multiple modal features based on their clinical contribution values. The dynamic reliability value of each sub-data is obtained, and the symptom hypothesis reliability of each high-value feature is determined based on the dynamic reliability value and the dynamic clinical weights of each sub-data. Multiple high-value features are fused based on their symptom hypothesis reliability to obtain multiple first fused features. First fused features of the same modality are fused to obtain second fused features. Second fused features of different modalities are fused to obtain third fused features. The symptom fusion data of head and neck cancer patients includes the third fused features.

[0031] The core layer includes electronic medical record sub-data and laboratory test sub-data; the key layer includes standardized scale sub-data, active symptom record sub-data, and physiological parameter sub-data; the auxiliary layer includes activity level sub-data, behavioral parameter sub-data, and environmental data sub-data; the clinical modality data includes the sub-data from the core layer and the physiological parameter sub-data from the key layer; the reporting modality data includes the standardized scale sub-data and active symptom record sub-data from the key layer; and the behavioral modality data includes the sub-data from the auxiliary layer.

[0032] In this embodiment, before fusing the preprocessed multi-source data on symptoms of head and neck cancer patients, the method further includes: standardizing the multi-source data on symptoms of head and neck cancer patients, unifying the multi-source data on symptoms of head and neck cancer patients from different sources and in different formats into a standard format, eliminating the dimensional differences between the multi-source data on symptoms of head and neck cancer patients, and forming standardized multi-source data on symptoms of head and neck cancer patients; fusing the standardized multi-source data on symptoms of head and neck cancer patients to obtain fused data on symptoms of head and neck cancer patients.

[0033] Specifically, based on the clinical priority and data characteristics of multi-source data, the standardized multi-source data of head and neck cancer patient symptoms are divided into a core layer, a key layer, and an auxiliary layer, and dynamic clinical weights are assigned to the sub-data in each layer. The core layer includes electronic medical record sub-data and laboratory test sub-data; the key layer includes standardized scale sub-data, active symptom record sub-data, and physiological parameter sub-data; the auxiliary layer includes activity level sub-data, behavioral parameter sub-data, and environmental data sub-data. The stratified sub-data are classified to obtain three modalities: clinical modality, report modality, and behavioral modality. The clinical modality includes all sub-data in the core layer and the physiological parameter sub-data in the key layer; the report modality includes standardized scale sub-data and active symptom record sub-data in the key layer; and the behavioral modality includes all sub-data in the auxiliary layer. Features are extracted for each modality to obtain the corresponding features. The features are divided into static features and time-series features. A pre-set algorithm calculates the similarity correlation between features within the same modality; a second pre-set algorithm calculates the clinical correlation between features across modalities; a feature association network is constructed based on the similarity correlation and clinical correlation, and a feature-based error judgment contribution is introduced to calculate the clinical contribution value of each feature; features with a clinical contribution value greater than or equal to a pre-set clinical contribution threshold are designated as high-value features, resulting in several high-value features; a third pre-set algorithm calculates the dynamic reliability value of each sub-data set; based on the dynamic reliability value and dynamic clinical weight of each sub-data set, a weight mapping mechanism is used to determine the symptom hypothesis confidence level of the high-value features; high-value features within various modalities are fused based on the symptom hypothesis confidence level to obtain the first fused feature corresponding to each modality; the first fused features within the same modality are fused to obtain the second fused feature; the second fused features from different modalities are fused to obtain the fused feature of multi-source symptom data of head and neck cancer patients.

[0034] Among them, dynamic clinical weights serve as adjustment parameters for subsequent calculation of feature contribution value.

[0035] If a high-value feature originates from a sub-data set, then the reliability of its symptom hypothesis is set as the product of the dynamic reliability value of that sub-data set and the dynamic clinical weight.

[0036] Furthermore, the electronic medical record sub-data records the diagnosis, stage, treatment plan, pathology report, and comorbidities of head and neck cancer patients; the laboratory test sub-data records blood routine, liver and kidney function, nutritional indicators, and imaging data of head and neck cancer patients; the standardized scale sub-data consists of a list of head and neck cancer symptoms, a quality of life scale, or a pain numerical rating scale that is regularly pushed through an app or mini-program; the active symptom recording sub-data consists of daily changes in pain, swallowing, and sleep patterns recorded by head and neck cancer patients at any time; the activity level sub-data records the number of steps and duration of activities of head and neck cancer patients; the physiological parameter sub-data records the resting heart rate, heart rate variability, and sleep quality of head and neck cancer patients; the behavioral parameter sub-data records the dietary intake, medication adherence, and oral care frequency of head and neck cancer patients; and the environmental data sub-data records the environmental noise and light levels in the living environment of head and neck cancer patients.

[0037] Specifically, the first preset algorithm includes: ; in, Indicates within the same mode The degree of similarity between features; Represents two features within the same mode; , This represents the specific values ​​of feature A in the i-th and j-th samples; , This represents the specific values ​​of feature B in the i-th and j-th samples; , These represent taking the maximum and minimum values, respectively. This represents the balance coefficient, with a value ranging from 0 to 1; Indicates zero-prevention items. It is defined independently of the weighting coefficients in the calculation of feature contribution value.

[0038] Specifically, the second preset algorithm calculates the clinical correlation degree of cross-modal features, including: ; in, Indicates the clinical correlation between C and D features across modalities; This represents the correlation coefficient between characteristics C and D in historical cases; Indicates the degree of clinical rule matching; The variable represents the influence coefficient of the scenario; |·| indicates taking the absolute value to ensure that the correlation is non-negative.

[0039] Specifically, the dynamic reliability value of each sub-data is calculated based on the third preset algorithm, including: ； Wherein, β and γ are proportionality coefficients and β+γ=1, θ is the balance coefficient; ZH represents the true value of the sub-data, determined through expert consistency assessment (such as the consistency rate of multi-expert review of pathology reports) or data consistency verification (such as the consistency between active symptom records and standardized scale scores); CV represents the cross-validation consistency of the sub-data, determined through clinical correlation verification between different sub-data (such as the efficacy correlation between blood routine and medication adherence); TE represents the time validity of the sub-data, determined based on the matching degree between the data collection time and clinical application needs (such as high validity within 7 days after the collection of laboratory test data); D represents the data density of the sub-data, determined based on the data collection frequency (such as high density for activity level sub-data collected ≥12 times per day).

[0040] Specifically, the clinical contribution value of each modality feature is calculated, including: ; in, This represents the classification performance value of the feature in symptom clustering; This represents the degree to which a feature contributes to misclassification in symptom classification. Values ​​indicating the degree of fluctuation of a feature at different treatment stages; This indicates the dynamic clinical weight of the sub-data to which this feature belongs; , , , Represents the weighting coefficient, and Clinical Contribution Value (CCV) Calculation Based on Feature Association Network: Classification Effect Value (CJ) Calculation: The feature association network is input into a spectral clustering algorithm, and the symptom type (e.g., pain, dysphagia, fatigue) is used as the clustering target to divide the data into categories. Clusters ( (where the number of symptom types is 0); for each feature node, calculate its silhouette coefficient in the cluster: ;in, Let i be the average distance from feature i to other features in the same cluster. is the average distance from feature i to the nearest heterogeneous feature; the average silhouette coefficient of this feature is used as the classification performance value; the higher the value, the better the classification performance. Acquisition method: Benchmark clustering: Initial clustering quality is calculated based on the complete network. (Adjusted Rand Index, ARI) is used as the baseline; Feature perturbation: For each feature node i, remove the node and all its connecting edges, and recalculate the perturbed clustering quality. Contribution quantification: The value range is [0,1], with higher values ​​indicating a greater contribution of the feature to misjudgment; if there is a conflict with clinical rules in the feature association network (e.g., "heart rate variability" has a high correlation with "pain score" but is not supported by clinical rules), then Increase Δ proportionally, for example: Δ = 0.2; Stochastic variability (Stab) calculation: For time-series characteristics, calculate their standard deviation during the treatment phase. For static features, Stab is always 1. ;in, The maximum possible fluctuation value of this feature, with a value range of [0,1]. The higher the value, the stronger the stability. Dynamic clinical weight (W) is obtained from the dynamic clinical weight of the sub-data to which the feature belongs.

[0041] In this embodiment, the first fusion feature is the result of preliminary fusion of high-value features and symptom hypothesis reliability within each modality. Specifically, it is based on the fusion of high-value features and symptom hypothesis reliability within each modality. For example: Clinical modality: fusion of its high-value features (such as staging features in electronic medical records, nutritional indicator features in laboratory tests, etc.) and corresponding symptom hypothesis reliability to generate the first fusion feature of the clinical modality; Reporting modality: fusion of high-value features (such as pain rating features) in standardized scale sub-data and active symptom record sub-data to generate the first fusion feature of the reporting modality; Behavioral modality: fusion of high-value features such as activity level and behavioral parameters (such as activity step count) to generate the first fusion feature of the behavioral modality.

[0042] In this embodiment, the second fusion feature is the result of modality-level optimization of the first fusion feature. Specifically, it is generated by fusing the first fusion features within the same modality. For example, for the first fusion feature of the clinical modality, a consistency verification rule (such as confidence adjustment by combining dynamic reliability values ​​and clinical rule matching degree) is applied to generate the second fusion feature of the clinical modality; similarly, the same optimization is performed on the first fusion features of the reporting modality and the behavioral modality to obtain their respective second fusion features.

[0043] In this embodiment, the construction of the feature association network includes node definition and edge weight calculation.

[0044] Node definition: All extracted features (including static features and temporal features) are treated as network nodes. For example, the static feature of "diagnostic staging" in the clinical modality and the temporal feature of "activity steps" in the behavioral modality are both treated as independent nodes.

[0045] Edge weight calculation: If two features belong to the same modality (such as "pathology report" and "nutritional indicators" in the clinical modality), the edge weight is set to the similarity correlation calculated by the second preset algorithm. If the two features belong to different modalities (such as "imaging data" in the clinical modality and "medication adherence" in the behavioral modality), the edge weights are set to the clinical relevance calculated by the third preset algorithm. Network sparsification: Only retain weights with values ​​greater than a preset threshold. (For example Edges with a value of 0.5 are used to filter out weakly correlated noise. This ultimately results in an undirected weighted graph. ,in, For the set of feature nodes, Let be a set of edges, and let the edge weights be... This represents the strength of the association between features i and j. Network validation: Key edges (such as the association between "pain score" and "sleep quality") are reviewed by clinical experts to ensure that the network conforms to medical logic.

[0046] The working principle and beneficial effects of the above technical solution are as follows: Multi-source data is divided into a core layer, a key layer, and an auxiliary layer, and dynamic clinical weights are assigned to the sub-data in each layer. The electronic medical record sub-data and laboratory test sub-data in the core layer are crucial for diagnosis and treatment; the data in the key layer is important for understanding patient symptoms and physical condition; and the data in the auxiliary layer plays a supplementary role. By stratifying and assigning weights, the differences in importance of different data in clinical decision-making can be highlighted, allowing subsequent analysis to focus more on key information. The allocation of dynamic clinical weights considers the clinical priority and characteristics of the data, and can flexibly adjust the importance of each sub-data according to the actual clinical situation. At different treatment stages or for different patient conditions, the importance of certain data may change; dynamic weights can better adapt to this change and improve data efficiency. Based on the effectiveness of data fusion, the stratified sub-data is divided into clinical modalities, reporting modalities, and behavioral modalities. This classification method aligns with the essential characteristics and sources of the data. Clinical modalities encompass key data required for diagnosis and treatment, reporting modalities reflect patients' subjective feelings and self-reported information, and behavioral modalities reflect patients' daily lives and environmental factors. Clear modality division facilitates targeted processing and analysis of different types of data. Features are divided into static features and time-series features, enabling a more comprehensive capture of information in the data. Static features describe the patient's state at a specific moment, while time-series features reflect the trend of data changes over time. For disease diagnosis and prediction, time-series information is often crucial, and comprehensive feature extraction helps improve the model's understanding and predictive ability regarding disease progression.

[0047] The model prediction module 102 is used to input the fused data of symptoms of head and neck cancer patients into a trained deep learning model to obtain the symptom prediction results of head and neck cancer patients output by the trained deep learning model. The deep learning model is trained using multi-source data of symptoms of historical head and neck cancer patients and corresponding historical symptoms of head and neck cancer patients.

[0048] Specifically, the symptom fusion data of head and neck cancer patients is input into the symptom prediction model for head and neck cancer patients, that is, into a trained deep learning model. The symptom prediction model analyzes the symptom fusion data of head and neck cancer patients and automatically predicts the symptoms of head and neck cancer patients. For example, it predicts the risk of head and neck cancer patients developing severe oral mucositis, pain, and dysphagia in the next week and determines the symptom prediction results for head and neck cancer patients.

[0049] In this embodiment, the model prediction module 102 is further configured to: obtain multiple symptoms and multiple data features from the symptom fusion data of head and neck cancer patients, and construct a symptom-feature association graph based on the multiple symptoms and multiple data features; determine the correlation coefficient between each adjacent node in the symptom-feature association graph according to the symptom fusion data of head and neck cancer patients, and obtain the adjacency matrix; input the graph node features and adjacency matrix in the symptom-feature association graph into the trained deep learning model; wherein, the deep learning model includes a first-layer multi-head attention graph network, a second-layer multi-head attention graph network, and a dynamic weight adaptive prediction head connected thereto. The dynamic weight adaptive prediction head is constructed based on a weight learning network and a risk prediction network. Both the weight learning network and the risk prediction network are constructed based on a 3-layer fully connected network.

[0050] In this embodiment, the head and neck cancer patient symptom management system based on multi-source data fusion and AI-driven model training module is further included, which is used to: collect historical data of head and neck cancer patients' symptoms and construct a training set, the training set including clinical historical data, historical data of head and neck cancer patients' reports, and historical data of behavior; extract features from the training set, fuse the extracted features to obtain training fusion features; use historical symptoms and historical features in the training fusion features as nodes, calculate the correlation coefficient between nodes based on the historical data in the training set as edge weights and form a historical adjacency matrix to construct a historical symptom-feature association graph; input the graph node features and historical adjacency matrix in the historical symptom-feature association graph into a first-layer multi-head attention graph attention network to obtain historical graph features; input the historical graph features into a second-layer multi-head attention graph attention network to obtain a historical symptom feature matrix; construct a historical feature vector based on the historical features; input the historical symptom feature matrix into the risk prediction network of the dynamic weight adaptive prediction head, and input the historical feature vector into the weight learning network of the dynamic weight adaptive prediction head; the adaptive prediction head of the dynamic weight adaptive prediction head combines the output of the risk prediction network and the output of the weight learning network to determine the symptom prediction result of head and neck cancer patients.

[0051] In this embodiment, the deep learning model is trained. The specific training process may include: collecting historical symptom data of head and neck cancer patients and dividing it into a training set and a test set in a 7:3 ratio; training the deep learning model using the training set, enabling the model to autonomously learn symptom prediction behavior for head and neck cancer patients and automatically predict their symptoms, including the risk of severe oral mucositis, pain, and dysphagia within the next week, thus establishing a symptom prediction model for head and neck cancer patients; and using the test set to train the deep learning model for head and neck cancer patients. The patient symptom prediction model was tested, and its generalization performance was evaluated based on model evaluation metrics to determine whether the model could automatically predict the symptoms of head and neck cancer patients, thus determining the model test evaluation results. Based on these results, the model was optimized. If the model could not automatically predict the symptoms, its parameters were adjusted and iteratively optimized until it could automatically predict the symptoms, thus determining the optimal model.

[0052] The deep learning model is trained using a training set, enabling it to autonomously learn symptom prediction behaviors in head and neck cancer patients and automatically predict their symptoms. For example, it can predict the risk of severe oral mucositis, pain, and dysphagia in head and neck cancer patients within the next week. Specifically, features are extracted from the training set and fused to obtain training fusion features. These features include 256-dimensional structured data containing clinical data, reported data from head and neck cancer patients, and behavioral data, with a dimension of N×T×256. A node-specific feature generation mechanism generates differentiated feature vectors for each node in the symptom-feature association graph. Key symptoms and core features from the training set are used as nodes, and Pearson correlation coefficients between nodes are calculated based on historical data from the training set as edge weights to form an adjacency matrix and store graph relationships, thus constructing the symptom-feature association graph. A first-layer GAT multi-head attention graph network is constructed, with an input dimension of 40×256, to calculate the relationship between node i and its neighbors. The attention coefficient of node j is used to output a 40×256 dimensional spectral feature map. A second-layer GAT multi-head attention graph network is constructed for input dimension 40×256, retaining the features of 8 symptom nodes through a fixed index and outputting an 8×256 dimensional symptom feature matrix. A symptom association graph attention module is constructed based on the symptom-feature association graph, the first-layer GAT, and the second-layer GAT. A weight learning network is constructed based on a 3-layer fully connected network. A prediction output layer is constructed based on a 3-layer fully connected network, outputting a 3-dimensional risk level probability vector. A dynamic weight adaptive prediction head is constructed based on the weight learning network and the prediction output layer. Based on the symptom association graph attention module and the dynamic weight adaptive prediction head, a symptom prediction model for head and neck cancer patients is constructed.

[0053] In this embodiment, N is the number of samples, T is a fixed time step, and the variable-length sequence is processed by zero padding.

[0054] In this embodiment, a node-specific feature generation mechanism is used to generate a differentiated feature vector for each node in the symptom-feature association graph, specifically including: Key symptom nodes (8): Dedicated feature vectors are extracted from patient report data, with an input dimension of T×16 (e.g., NRS score time series). These vectors are then processed through a 2-layer fully connected network (input dimension 16, output dimension 256, activation function ReLU) to generate 256-dimensional symptom node features. Core feature nodes (32): The original 32-dimensional core feature vectors (diagnostic staging / blood routine / nutritional indicators from clinical data, NRS scores / active symptom records from report data, and activity steps / medication adherence / environmental noise from behavioral data) are directly used. These vectors are then concatenated into a 32×8=256-dimensional vector through an embedding layer (input dimension 1, output dimension 8). This vector is then processed through a 1-layer fully connected network (input dimension 256, output dimension 256, activation function Linear) to generate 256-dimensional features. The final output is a 40×256-dimensional atlas node feature matrix (8 symptom nodes + 32 core feature nodes), with each node feature generated independently.

[0055] In this embodiment, key symptoms include dysphagia, sore throat, hoarseness, enlarged neck mass, oral ulcers, difficulty breathing, earache, and limited mouth opening; core features include 32 features such as diagnostic staging of clinical data / complete blood count / nutritional indicators, NRS score / active symptom record of reported data, and activity steps / medication adherence / environmental noise of behavioral data; risk level (low risk, medium risk, high risk) is only used as a predictive output target and is not used as a graph node.

[0056] In this embodiment, the attention coefficient between node i and its neighbor node j is calculated: ;in, denoted by , where represents the attention coefficient between node i and its neighbor node j; 'a' represents the 512-dimensional attention vector; and 'W' represents the 256×256 weight matrix. , Represents the feature vectors of nodes i and j; This indicates splicing; the total number of nodes is 40 (8 key symptoms + 32 core features).

[0057] In this embodiment, a weight learning network is constructed based on a 3-layer fully connected network. Specifically, the input is the original 32-dimensional core feature vector (corresponding one-to-one with the 32 core features in the atlas, but independent of the atlas features); the first fully connected layer has an input dimension of 32, an output dimension of 64, an activation function of ReLU, and a dropout coefficient of 0.2; the second fully connected layer has an input dimension of 64, an output dimension of 32, an activation function of ReLU, and a dropout coefficient of 0.2; the third fully connected layer has an input dimension of 32, an output dimension of 8, an activation function of Softmax, and outputs the weights of 8 symptoms. — Dynamic weights are used only for symptom management priority ranking, and the calculation formula is: priority score. ;in, To predict the numerical representation of risk levels (low risk = 0.5, medium risk = 1.0, high risk = 2.0), ensure that the risk prediction logic is decoupled from priority adjustment.

[0058] In this embodiment, a prediction output layer is constructed based on a 3-layer fully connected network, specifically including: an input of an 8×256-dimensional symptom feature matrix; a first fully connected network with an input dimension of 256, an output dimension of 128, an activation function of ReLU, and a dropout coefficient of 0.3; a second fully connected network with an input dimension of 128, an output dimension of 64, an activation function of ReLU, and a dropout coefficient of 0.3; and a third fully connected network with an input dimension of 64, an output dimension of 3, an activation function of Softmax, and an output of a 3-dimensional risk level probability vector. , , (These correspond to low risk, medium risk, and high risk, respectively); Risk level determination rule: The level corresponding to the highest probability is taken as the final risk level.

[0059] In this embodiment, the total loss function is defined as: ;in, Indicates the total loss; 8 represents the total number of head and neck cancer patient samples in the training set; 8 represents the total number of predicted key symptoms of head and neck cancer; n represents the sample index; i represents the symptom index; 3 represents the natural logarithm (base e); 3 represents the total number of risk level categories; k represents the risk level index; This represents the actual label indicating the risk level of the i-th symptom in the n-th patient, classified as the k-th category. This represents the model's predicted probability of the risk level of the i-th symptom in the k-th category for the n-th patient.

[0060] Specifically, the input layer receives raw data from multiple sources, including clinical data (diagnostic staging / complete blood count / nutritional indicators, etc.), patient-reported data (NRS scores / active symptom records, etc.), and behavioral data (steps / medication adherence / environmental noise, etc.). After standardized preprocessing, the raw data is divided into two parallel paths: Path 1 (feature extraction and fusion): The raw data is input into the feature extraction module to generate training fusion features with dimensions of [dimension not specified]. ( (Fixed time step) is used for model training monitoring and validation, but does not participate in subsequent atlas construction; Path 2 (node ​​feature generation): Patient report data input symptom node generation submodule, extracting temporal features (dimensions) The system generates 256-dimensional feature vectors for 8 symptom nodes through a 2-layer fully connected network (input 16 → output 256), with an output dimension of 8×256. The original 32-dimensional core feature vectors are input into the core feature node generation submodule, and after passing through an embedding layer (32-dimensional → 256-dimensional) and a fully connected layer, 256-dimensional feature vectors for 32 core feature nodes are generated, with an output dimension of 32×256. The outputs of the two sub-paths are concatenated to form a 40×256-dimensional map node feature matrix (symptom nodes first, core feature nodes second), which serves as the input to the symptom-feature association map.

[0061] The graph construction layer calculates the Pearson correlation coefficients between 40 nodes based on historical data from the training set, forming a 40×40 adjacency matrix to store symptom-feature associations. This adjacency matrix, together with the 40×256-dimensional graph node feature matrix, is input into the symptom association graph attention module.

[0062] GAT processing layer: The first layer GAT receives 40×256-dimensional graph node features and a 40×40 adjacency matrix, calculates the attention coefficients between nodes through a multi-head attention mechanism, and outputs 40×256-dimensional graph features (the dimension remains unchanged, but the features are enhanced by relationships).

[0063] The second layer GAT receives the 40×256-dimensional features output by the first layer GAT, selects 8 symptom nodes through fixed indices (0-7), and outputs an 8×256-dimensional symptom feature matrix (dimensional compression, retaining key symptom information).

[0064] Parallel output layer: Weight learning branch: The original 32-dimensional core feature vector is directly input into a 3-layer fully connected network (32→64→32→8), and the output is an 8-dimensional dynamic weight vector. (8×1 dimension) used for symptom management priority ranking. Risk prediction branch: 8×256 dimensional symptom feature matrix input to a 3-layer fully connected network (256→128→64→3), outputting an 8×3 dimensional risk level probability matrix (each symptom corresponds to 3 risk level probabilities).

[0065] Dynamic weighted adaptive prediction head: Combines the 8-dimensional weight vector of the weight learning branch with the 8×3 probability matrix of the risk prediction branch to calculate the priority score. ;in, The risk level is represented numerically, but the risk prediction probability is not modified; the final output includes: Risk level prediction results: 3 risk probabilities for 8 symptoms (8×3 dimensions); Symptom management priority sequence: based on A priority list of 8 symptoms.

[0066] The feature dimension transition is shown in Table 1: Table 1

[0067] Node feature generation and GAT processing are completely decoupled from training and feature fusion. The risk prediction branch relies only on the symptom features output by GAT, and the weight branch relies only on the original core features, ensuring that the risk prediction logic and priority adjustment do not interfere with each other. All dimensional transformations are implemented through fully connected layers, and the input / output dimensions remain consistent during the training and inference phases.

[0068] The working principle and beneficial effects of the above technical solution are as follows: Key symptoms and core features in the training set are used as nodes. An adjacency matrix is ​​constructed based on the Pearson correlation coefficient to store the graph relationships. This graph structure clearly represents the association between symptoms and features, helping the model capture potential patterns and regularities in the data. Through the graph, the model can learn the mutual influence between different symptoms and features, thus more accurately predicting patient symptoms. Differentiated feature vectors are generated for each node in the symptom-feature association graph. Different nodes represent different symptoms or features, and the differentiated feature vectors better reflect the unique properties of each node, enabling the model to analyze and process each node more accurately when processing graph data, improving the model's expressive power. A two-layer GAT multi-head attention graph network is constructed. The first layer of GAT has an input dimension of 40×256. By calculating the attention coefficient between node i and its neighboring node j, the weights between different nodes can be adaptively allocated, highlighting important nodes and features, thereby better capturing local and global information in the graph. The second-layer GAT retains the features of eight symptom nodes through a fixed index and outputs an 8×256-dimensional symptom feature matrix, further focusing on key symptom nodes and providing more targeted feature representations for subsequent predictions. A symptom association graph attention module is constructed based on the symptom-feature association graph, the first-layer GAT, and the second-layer GAT. A weight learning network and a prediction output layer are constructed based on a three-layer fully connected network, and a dynamic weight adaptive prediction head is further built. This modular design makes the model structure clear, with each module having a defined function, facilitating development and debugging. It also has good scalability; if improvements to the model or the addition of new functions are needed, each module can be easily modified and expanded. The model can automatically predict the risk of severe oral mucositis, pain, and dysphagia in head and neck cancer patients within the next week, outputting a 3-dimensional risk level probability vector. This has significant reference value for clinicians, helping them to understand the possible symptoms in advance, take timely preventative and treatment measures, and improve patients' treatment outcomes and quality of life.

[0069] Among them, the symptom intervention management module 103 is used to intervene and manage the symptoms of head and neck cancer patients based on the symptom prediction results, forming a closed-loop management of the symptoms of head and neck cancer patients.

[0070] In this embodiment, intervention and management of symptoms in head and neck cancer patients based on symptom prediction results include: generating a symptom health management plan for head and neck cancer patients based on the symptom prediction results, and managing the symptoms of head and neck cancer patients based on the symptom health management plan; wherein, managing the symptoms of head and neck cancer patients based on the symptom health management plan includes: issuing an early warning when it is predicted that the risk of severe oral mucositis in head and neck cancer patients is high within the next week, strengthening the daily mouth rinsing task of head and neck cancer patients, tracking the symptom health management status of head and neck cancer patients in real time, and adjusting the symptom health management plan of head and neck cancer patients based on the feedback of real-time tracking, thus forming a closed-loop management of the symptoms of head and neck cancer patients.

[0071] Specifically, based on the symptom prediction results of head and neck cancer patients, a symptom health management plan is generated for them. The symptom health management plan is then used to manage the symptoms of head and neck cancer patients. Among these plans, when it is predicted that the risk of severe oral mucositis in a head and neck cancer patient is high within the next week, an early warning is issued in a timely manner, and the daily mouth rinsing task of the head and neck cancer patient is strengthened. The patient uses analgesic mouthwash to rinse three times a day, and the diet is changed to warm or cool liquid. The symptom health management status of the head and neck cancer patient is tracked in real time, and the symptom health management plan is adjusted based on the feedback from the real-time tracking, forming a closed-loop management of the symptoms of head and neck cancer patients.

[0072] This application also provides a method for symptom management in head and neck cancer patients based on multi-source data fusion and AI-driven approaches, such as... Figure 2 As shown, it includes the following steps: Step S201: Collect multi-source data on symptoms of head and neck cancer patients from multiple dimensions, preprocess the collected multi-source data on symptoms of head and neck cancer patients, and fuse the preprocessed multi-source data on symptoms of head and neck cancer patients to form fused data on symptoms of head and neck cancer patients.

[0073] Step S202: Input the fused data of symptoms of head and neck cancer patients into the trained deep learning model to obtain the predicted results of symptoms of head and neck cancer patients output by the trained deep learning model. The deep learning model is trained by multi-source data of symptoms of historical head and neck cancer patients and the corresponding symptoms of historical head and neck cancer patients.

[0074] Step S203: Based on the symptom prediction results of head and neck cancer patients, intervene and manage the symptoms of head and neck cancer patients to form a closed-loop management of the symptoms of head and neck cancer patients.

[0075] The implementation of each step in the above-mentioned method for managing symptoms of head and neck cancer patients based on multi-source data fusion and AI is described in the above-mentioned device for managing symptoms of head and neck cancer patients based on multi-source data fusion and AI, and will not be detailed here.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0077] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A symptom management system for head and neck cancer patients based on multi-source data fusion and AI driving, characterized in that, include: The data acquisition and fusion module is used to collect multi-source data on symptoms of head and neck cancer patients from multiple dimensions, preprocess the collected multi-source data on symptoms of head and neck cancer patients, and fuse the preprocessed multi-source data on symptoms of head and neck cancer patients to form fused data on symptoms of head and neck cancer patients. The model prediction module is used to input the fused data of symptoms of the head and neck cancer patients into a trained deep learning model to obtain the symptom prediction results of the head and neck cancer patients output by the trained deep learning model. The deep learning model is trained by multi-source data of symptoms of historical head and neck cancer patients and corresponding historical head and neck cancer patient symptoms. The symptom intervention management module is used to intervene and manage the symptoms of head and neck cancer patients based on the symptom prediction results, forming a closed-loop management of the symptoms of head and neck cancer patients.

2. The multi-source data fusion and AI-driven head and neck cancer patient symptom management system according to claim 1, wherein, Multi-source data on symptoms of head and neck cancer patients were collected from multiple dimensions, including: Collect electronic medical records and laboratory test data of patients with head and neck cancer to obtain clinical data of these patients. Standardized scales and active symptom records of head and neck cancer patients were collected to obtain reported data from them. Activity levels, physiological parameters, behavioral parameters, and environmental data of head and neck cancer patients were collected to obtain behavioral data of head and neck cancer patients; Based on clinical data, reported data, and behavioral data of head and neck cancer patients, multi-source data on symptoms of head and neck cancer patients were identified.

3. The multi-source data fusion and AI-driven head and neck cancer patient symptom management system according to claim 1, wherein, in, The electronic medical record includes the diagnosis and staging of head and neck cancer patients, treatment plan information, pathology report information and / or comorbidity information, and the laboratory test data includes complete blood count, liver and kidney function, nutritional indicators and / or imaging data. And / or, The standardized scales include a head and neck cancer symptom checklist, a quality of life scale, and / or a pain numerical rating scale. The active symptom record includes symptom information on pain, swallowing, and / or sleep recorded by head and neck cancer patients. And / or, The activity level includes the number of steps and / or duration of activity for patients with head and neck cancer; the physiological parameters include the resting heart rate, heart rate variability and / or sleep quality for patients with head and neck cancer; the behavioral parameters include the dietary intake, medication adherence and / or oral care frequency for patients with head and neck cancer; and the environmental data include the ambient noise and / or light intensity in the living environment of patients with head and neck cancer.

4. The multi-source data fusion and AI-driven head and neck cancer patient symptom management system according to any one of claims 1-3, characterized in that, The preprocessing of the collected multi-source symptom data from head and neck cancer patients includes: The multi-source data on symptoms of head and neck cancer patients were cleaned to remove noise from the data. The missing values ​​and outliers in the multi-source data of symptoms of head and neck cancer patients are identified and evaluated. Based on the evaluation results, missing values ​​and outliers that are valuable for the management of symptoms of head and neck cancer patients are identified. Based on the missing values ​​that are valuable for the management of symptoms of head and neck cancer patients, numerical imputation is performed on the multi-source data of symptoms of head and neck cancer patients. Based on the outliers that are valuable for the management of symptoms of head and neck cancer patients, numerical replacement is performed. Based on the evaluation results, missing values ​​and outliers that are of no value for symptom management in head and neck cancer patients were identified, and these missing values ​​and outliers that are of no value for symptom management in head and neck cancer patients were removed from the multi-source data on symptoms of head and neck cancer patients. The preprocessed multi-source data of symptoms of head and neck cancer patients was obtained based on the multi-source data of symptoms of head and neck cancer patients after cleaning, numerical filling, numerical replacement and numerical deletion processing.

5. The multi-source data fusion and AI-driven head and neck cancer patient symptom management system according to claim 4, wherein, The evaluation of missing and outlier values ​​in multi-source data on symptoms of head and neck cancer patients includes: If any missing value in the multi-source data of symptoms of head and neck cancer patients is identified as a missing value with a completely random missing pattern, then the missing value in the multi-source data of symptoms of head and neck cancer patients is assessed as having no value for the management of symptoms of head and neck cancer patients. If any missing value in the multi-source data of symptoms of head and neck cancer patients is identified as a missing value with a non-random missing pattern, then the missing value in the multi-source data of symptoms of head and neck cancer patients is assessed as having value for the management of symptoms of head and neck cancer patients. If any outlier in the multi-source data of symptoms of a head and neck cancer patient is identified as an error-type outlier, then the outlier in the multi-source data of symptoms of a head and neck cancer patient is assessed as having no value for the management of symptoms of head and neck cancer patients. If any outlier in the multi-source data of symptoms of a head and neck cancer patient is identified as a trend-type outlier, then the outlier in the multi-source data of symptoms of a head and neck cancer patient is assessed as having value for the management of symptoms of head and neck cancer patients.

6. The multi-source data fusion and AI-driven head and neck cancer patient symptom management system according to claim 1, wherein, The data fusion of preprocessed multi-source data on symptoms of head and neck cancer patients includes: Based on the clinical priority of multi-source data and the data characteristics of each source data, the preprocessed multi-source data of head and neck cancer patients' symptoms are divided into a core layer, a key layer, and an auxiliary layer, and dynamic clinical weights are configured for the sub-data of each layer. The sub-data of each layer are divided into data modalities to obtain clinical modality data, report modality data, and behavioral modality data; Feature extraction is performed on various modal data to obtain multiple modal features, including clinical modal features, report modal features, and behavioral modal features; Calculate the similarity correlation degree between modal features within the same modality, and calculate the clinical correlation degree between modal features across modalities, and construct a feature association network based on the similarity correlation degree, the clinical correlation degree, and multiple modal features; The clinical contribution value of each modality feature is calculated based on the dynamic clinical weights of the sub-data in the feature association network and each layer, and multiple high-value features are selected from multiple modality features based on the clinical contribution value of each modality feature. The dynamic reliability values ​​of each sub-data set are obtained, and the symptom hypothesis reliability of each high-value feature is determined based on the dynamic reliability values ​​and dynamic clinical weights of each sub-data set. Based on the symptom hypothesis confidence of each high-value feature, feature fusion is performed on the multiple high-value features to obtain multiple first fused features; The first fusion feature of the same modality is fused to obtain the second fusion feature; The second fusion features of different modalities are fused to obtain the third fusion feature, wherein the symptom fusion data of the head and neck cancer patients includes the third fusion feature.

7. The head and neck cancer patient symptom management system based on multi-source data fusion and AI-driven according to claim 6, characterized in that, in, The core layer includes electronic medical record sub-data and laboratory test sub-data; the key layer includes standardized scale sub-data, active symptom record sub-data, and physiological parameter sub-data; the auxiliary layer includes activity level sub-data, behavioral parameter sub-data, and environmental data sub-data. The clinical modality data includes sub-data from the core layer and sub-data from physiological parameters in the key layer; the reporting modality data includes sub-data from standardized scales and active symptom records in the key layer; and the behavioral modality data includes sub-data from the auxiliary layer.

8. The head and neck cancer patient symptom management system based on multi-source data fusion and AI-driven according to claim 1, characterized in that, The model prediction module is also used for: Multiple symptoms and multiple data features are obtained from the symptom fusion data of the head and neck cancer patients, and a symptom-feature association map is constructed based on the multiple symptoms and multiple data features; Based on the symptom fusion data of the head and neck cancer patients, the correlation coefficients between adjacent nodes in the symptom-feature association map are determined, and the adjacency matrix is ​​obtained; The graph node features in the symptom-feature association graph and the adjacency matrix are input into the trained deep learning model; The deep learning model includes a first-layer multi-head attention graph network, a second-layer multi-head attention graph network, and a dynamic weight adaptive prediction head, all connected together. The dynamic weight adaptive prediction head is constructed based on a weight learning network and a risk prediction network, both of which are constructed based on a 3-layer fully connected network.

9. The head and neck cancer patient symptom management system based on multi-source data fusion and AI-driven according to claim 8, characterized in that, It also includes a model training module, used for: Historical symptom data of head and neck cancer patients were collected and a training set was constructed. The training set included clinical historical data, historical report data of head and neck cancer patients, and historical behavioral data. Feature extraction is performed on the training set, and the extracted features are fused to obtain training fused features; The historical symptoms and historical features in the training fusion features are used as nodes. The correlation coefficient between nodes is calculated based on the historical data in the training set as the edge weight and a historical adjacency matrix is ​​formed to construct a historical symptom-feature association graph. The graph node features in the historical symptom-feature association graph and the historical adjacency matrix are input into the first layer multi-head attention graph attention network to obtain historical graph features; The historical graph features are input into the second-layer multi-head attention graph attention network to obtain the historical symptom feature matrix; Construct a historical feature vector based on the aforementioned historical features; The historical symptom feature matrix is ​​input into the risk prediction network of the dynamic weight adaptive prediction head, and the historical feature vector is input into the weight learning network of the dynamic weight adaptive prediction head. The adaptive prediction head of the dynamic weight adaptive prediction head combines the output of the risk prediction network and the output of the weight learning network to determine the symptom prediction result of the head and neck cancer patient.

10. A symptom management method for head and neck cancer patients based on multi-source data fusion and AI-driven approach, characterized in that... include: Multi-source data on symptoms of head and neck cancer patients were collected from multiple dimensions, and the collected multi-source data on symptoms of head and neck cancer patients were preprocessed. The preprocessed multi-source data on symptoms of head and neck cancer patients were then fused to form fused data on symptoms of head and neck cancer patients. The fused data of symptoms of head and neck cancer patients is input into a trained deep learning model to obtain the symptom prediction results of head and neck cancer patients output by the trained deep learning model. The deep learning model is trained by multi-source data of symptoms of historical head and neck cancer patients and corresponding historical symptoms of head and neck cancer patients. Based on the symptom prediction results of head and neck cancer patients, intervention and management of symptoms are carried out to form a closed-loop management of symptoms of head and neck cancer patients.

Images