A clinical decision support system based on thyroid atypical hyperplasia disease library and a construction method thereof

Abstract

The application discloses a clinical decision support system based on a thyroid atypical hyperplasia special disease library and a construction method thereof. The method comprises the following steps: collecting thyroid hyperplasia related multi-source heterogeneous data, performing feature extraction on the multi-source heterogeneous data, and constructing a multi-modal database based on the multi-source heterogeneous data after the feature extraction; based on the data contained in the multi-modal database, using an engine decision tree corresponding to a processing rule of thyroid hyperplasia and a data prediction model corresponding to a graph neural network, a corresponding processing strategy is obtained; based on the processing feedback data of thyroid hyperplasia, the multi-modal database and the processing strategy are updated through an incremental learning mechanism and a federated learning mechanism.

Description

Technical Field

[0001] This application relates to the field of multimodal data processing, and in particular to a clinical decision support system based on a thyroid atypical hyperplasia disease database and its construction method. Background Technology

[0002] Atypical thyroid hyperplasia is a borderline lesion in the diagnosis of thyroid nodules, referring to a lesion state with atypical features on cytology or histology but not yet meeting the diagnostic criteria for malignancy. Currently, the determination of the nature of thyroid atypical hyperplasia mainly relies on the comprehensive evaluation of multimodal data such as ultrasound imaging features, gene testing results, and pathology reports. However, clinical decision-making has the following problems: first, overtreatment, with some patients with benign atypical hyperplasia undergoing unnecessary surgical resection; second, delayed treatment, with some lesions with malignant potential not being intervened in a timely manner, leading to disease progression. Existing clinical decision support systems have significant limitations in thyroid specialty applications. On the one hand, knowledge-based systems, such as decision systems based on guideline rule bases, mainly rely on static clinical guidelines and expert rules for reasoning. However, medical knowledge is updated rapidly, and static rule bases cannot incorporate the latest evidence-based evidence in a timely manner, increasing the misdiagnosis rate of atypical hyperplasia. At the same time, such systems usually use a single data source and cannot effectively integrate multimodal information such as electronic medical records, ultrasound images, and gene data, resulting in one-sided decision-making basis. On the other hand, data-driven systems, such as deep learning-based prediction models, while exhibiting high prediction accuracy in tasks like image recognition, lack interpretability in their black-box decision-making process, making it difficult for clinicians to understand the model's reasoning. Furthermore, most clinical decision support systems lack specialized databases for thyroid atypical hyperplasia, struggle to handle complex situations such as conflicts between ultrasound features and genetic testing results, and suffer from insufficient decision reliability. Summary of the Invention

[0003] To address the aforementioned technical problems, this application provides a clinical decision-making construction method based on a thyroid atypical hyperplasia disease database, the method comprising: Collect multi-source heterogeneous data related to thyroid hyperplasia, extract features from the multi-source heterogeneous data, and construct a multimodal database based on the feature-extracted multi-source heterogeneous data; Based on the data contained in the multimodal database, the corresponding processing strategy is obtained by using the engine decision tree corresponding to the processing rules of thyroid hyperplasia and the data prediction model corresponding to the graph neural network. Based on the processing feedback data of thyroid hyperplasia, the multimodal database and the processing strategy are updated through incremental learning and federated learning mechanisms.

[0004] Optionally, the step of extracting features from the multi-source heterogeneous data and constructing a multimodal database based on the feature-extracted multi-source heterogeneous data includes: Based on the data format and content characteristics of the multi-source heterogeneous data, the multi-source heterogeneous data is divided into text data and image data; The text data is subjected to terminology standardization mapping, and based on the terminology-standardized text data, the semantic relationships between clinical entities are extracted to obtain structured text feature vectors. Image feature extraction is performed on the image data to obtain texture and morphological features. The extracted image features are then standardized by standard deviation and binned discretized to generate structured image feature vectors. The structured text feature vectors are fused with the structured image feature vectors to obtain multimodal feature matrix data, and a multimodal database is constructed based on the multimodal feature matrix data.

[0005] Optionally, based on the data contained in the multimodal database, the corresponding processing strategy is obtained by using the engine decision tree corresponding to the processing rules for thyroid hyperplasia and the data prediction model corresponding to the graph neural network, including: Patient data from the multimodal database is input into a preset rule engine. The rule engine analyzes the patient data based on the clinical guidelines rule base for thyroid hyperplasia and a localized rule base, and generates a first decision suggestion through the engine decision tree. Patient data from the multimodal database is input into a preset graph neural network model. The graph neural network model uses the features corresponding to the patient's symptoms as nodes and the relationships between clinical entities as edges. Through a multi-layer message passing mechanism, deep correlation features between multimodal data are extracted to generate risk prediction results. Using a multi-objective optimization algorithm, the first decision suggestion and the risk prediction result are dynamically weighted and fused to generate a corresponding processing strategy.

[0006] Optionally, the processing feedback data based on thyroid hyperplasia is used to update the multimodal database and the processing strategy through incremental learning and federated learning mechanisms, including: Collect feedback data during the clinical management of thyroid hyperplasia; Based on the feedback data, the pathological results are compared with the prediction results to determine the corresponding prediction deviation. When the prediction deviation exceeds a preset threshold, the incremental learning process is triggered to adjust the parameters of the data prediction model corresponding to the graph neural network. Based on the feedback data and the multimodal data, the graph neural network model is used to calculate the model gradient data, and the corresponding model parameters are updated based on the model gradient data. The updated data and features generated during incremental learning and federated learning are added to the multimodal database, and the multimodal database is dynamically updated.

[0007] Optionally, the step of calculating model gradient data using the graph neural network model based on the feedback data and the multimodal data, and updating the corresponding model parameters based on the model gradient data, includes: Based on locally collected feedback data and locally stored multimodal data, the model gradient data is calculated using the trained graph neural network model. The gradient data of the computational model is encrypted using a homomorphic encryption algorithm, and the encrypted gradient data of the computational model is sent to the federated learning aggregation server. The encrypted computational model gradient data is homomorphically encrypted and aggregated by the federated learning aggregation server, and differential privacy noise is added to obtain aggregated noisy gradient data. Decrypt the aggregated noise gradient data to obtain the updated model parameters corresponding to the model gradient data; The graph neural network model is updated based on the updated model parameters.

[0008] Optionally, the feedback data may include at least one of the following: records of doctors' adoption and adjustment of treatment strategies, postoperative recovery data and complication occurrences of patients, and comparison data between postoperative pathology results and preoperative prediction results.

[0009] Optionally, the method further includes: The decision results output by the clinical system are collected and compared with the actual clinical results to obtain the corresponding prediction bias. Determine whether the prediction deviation exceeds the preset threshold; If the prediction deviation exceeds the preset threshold, an incremental learning process is triggered to update the parameters of the data prediction model corresponding to the graph neural network. If the prediction deviation does not exceed the preset threshold, only feedback data is recorded, and the incremental learning process is not triggered.

[0010] Optionally, the step of dynamically weighting and fusing the first decision suggestion and the risk prediction result using a multi-objective optimization algorithm to generate a corresponding processing strategy includes: A multi-objective optimization algorithm is used to dynamically fuse the first decision suggestion and the risk prediction result to obtain multiple candidate fusion weight combinations. A target fusion weight combination is determined from multiple candidate fusion weight combinations, wherein the target fusion weight combination represents the optimal decision-making method among the multiple candidate fusion weight combinations; Based on the target fusion weight combination, a corresponding clinical treatment strategy is generated.

[0011] This application also provides a clinical decision support system based on a specialized database for thyroid atypical hyperplasia, including: A multimodal database construction module is used to collect multi-source heterogeneous data related to thyroid hyperplasia, extract features from the multi-source heterogeneous data, and construct a multimodal database based on the feature-extracted multi-source heterogeneous data. The dual-engine decision module is used to obtain the corresponding processing strategy based on the data contained in the multimodal database, using the engine decision tree corresponding to the processing rules of thyroid hyperplasia and the data prediction model corresponding to the graph neural network. The learning update module is used to update the multimodal database and the processing strategy based on the processing feedback data of thyroid hyperplasia through incremental learning and federated learning mechanisms.

[0012] The purpose of this application is to provide an electronic device, including a memory and a processor, wherein the memory stores an executable program, and the processor executes the executable program to implement the steps of the above method.

[0013] The beneficial effects of this application's embodiments are as follows: It deeply integrates and standardizes multi-source heterogeneous data related to thyroid atypical hyperplasia, such as electronic medical records, ultrasound images, gene testing, and pathology reports; it eliminates semantic ambiguity using SNOMED-CT terminology mapping; it eliminates differences in multi-center imaging equipment through Z-score standardization; and it constructs a multimodal knowledge graph containing patients, nodules, genes, pathological nodes, and semantic relationships, solving the technical problem of data fragmentation in traditional systems. It sets up a dual-engine decision-making architecture that combines a rule engine and a graph neural network, and uses the NSGA-II multi-objective optimization algorithm to dynamically fuse the outputs of both, obtaining the optimal trade-off processing strategy. It establishes a collaborative evolutionary mechanism based on incremental learning and federated learning, providing precise and continuously optimized clinical decision support for the diagnosis and treatment of thyroid hyperplasia. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a clinical decision-making construction method based on a thyroid atypical hyperplasia disease database, as described in this application. Figure 2 Examples of embodiments of this application Figure 1 A flowchart of one embodiment of step S100; Figure 3 Examples of embodiments of this application Figure 1 A flowchart of one embodiment of step S200; Figure 4 Examples of embodiments of this application Figure 1A flowchart of one embodiment of step S300; Figure 5 Examples of embodiments of this application Figure 4 A flowchart of an embodiment of step S330; Figure 6 This is another flowchart of the clinical decision-making construction method based on the thyroid atypical hyperplasia disease database in this application embodiment; Figure 7 Examples of embodiments of this application Figure 3 A flowchart of one embodiment of step S230; Figure 8 This is a structural block diagram of the clinical decision support system based on a thyroid atypical hyperplasia disease database, as described in this application embodiment. Figure 9 This is a structural block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0015] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0016] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0017] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0018] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0019] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0020] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0021] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0022] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0023] like Figure 1 As shown in the embodiments of this application, a clinical decision-making construction method based on a thyroid atypical hyperplasia disease database is provided, the method comprising: S100: Collect multi-source heterogeneous data related to thyroid hyperplasia, extract features from the multi-source heterogeneous data, and construct a multimodal database based on the feature-extracted multi-source heterogeneous data. This embodiment provides a clinical decision-making construction method based on a specialized database of thyroid atypical hyperplasia, which can be applied to clinical decision support systems based on multimodal data. For example, a multimodal data-based clinical decision support system could be a clinical decision support system for thyroid atypical hyperplasia. This method achieves multimodal fusion processing of thyroid hyperplasia-related data and decision-making on related processing methods through the following steps.

[0024] First, multi-source heterogeneous data related to thyroid hyperplasia can be collected through data interfaces. This can include different types of data, such as electronic medical record data collected from hospital information systems, including text data like patient examination results; thyroid ultrasound imaging data collected from imaging archives and communication systems, including nodule grayscale images and color Doppler blood flow images; gene testing report data collected from gene testing platforms; and postoperative pathology report data collected from pathology systems. The collected multi-source heterogeneous data can include complete clinical records, ultrasound images, gene testing results, and pathology reports.

[0025] For text-based data, natural language processing (NLP) techniques can be used to identify medical terms. Standardized mapping using the SNOMED-CT medical terminology system can then extract clinical entities and generate structured text feature vectors. SNOMED-CT (Systematized Nomenclature of Medicine -- Clinical Terms) is a multilingual clinical medical terminology system and a standard clinical terminology specification in electronic health record systems. In the proposed clinical decision support system based on a thyroid atypical hyperplasia database, SNOMED-CT is used for standardized terminology mapping of text-based data, converting unstructured clinical text into standardized terminology codes, laying the foundation for subsequent knowledge graph construction and graph neural network analysis. For example, from the pathology report "Fine-point biopsy of a nodule in the left lobe of the thyroid, pathological diagnosis of atypical cells of indeterminate significance," the entities "nodule," "left thyroid lobe," and "atypical cells," as well as the mapping relationships "nodule - located in - left thyroid lobe" and "nodule - diagnosed as - atypical cells," can be extracted.

[0026] For imaging data, ultrasound images are preprocessed and segmented to extract texture, morphological, and blood flow features. The extracted image features are then Z-score standardized to eliminate dimensional differences caused by different ultrasound devices. Finally, equal-frequency binning is used to discretize continuous features into multiple categories, generating structured image feature vectors. Z-score standardization (standard deviation standardization) is a data preprocessing method that transforms data of different dimensions and orders of magnitude into a uniform scale. This method standardizes the original data based on the mean and standard deviation, ensuring that the standardized data conforms to a standard normal distribution (mean = 0, standard deviation = 1). In the clinical decision support system for thyroid atypical hyperplasia proposed in this application, Z-score standardization is applied to the feature processing of imaging data to eliminate data differences caused by different ultrasound devices and operators, making the image features acquired from multiple centers comparable.

[0027] The structured text feature vectors and structured image feature vectors are horizontally concatenated to construct a multimodal feature matrix containing patient ID, nodule identifier, feature name, and feature value. Based on this feature matrix, knowledge graph nodes are generated, which may include, for example, patient nodes, nodule nodes, gene nodes, and pathology nodes. Associative edges are constructed between nodes based on entity relationships to form a thyroid hyperplasia-specific knowledge graph, which is stored in a graph database as a multimodal database.

[0028] S200, based on the data contained in the multimodal database, using the engine decision tree corresponding to the processing rules for thyroid hyperplasia and the data prediction model corresponding to the graph neural network, the corresponding processing strategy is obtained; Based on the constructed multimodal database, the following steps are performed for each patient's data to generate a personalized treatment strategy. Patient data can be input into a rule engine, which has a built-in decision tree constructed based on the ATA guidelines and thyroid treatment pathways. The rule engine is a software module that performs reasoning and decision-making based on predefined clinical rules and a logic tree structure. By formalizing the clinical guidelines for thyroid hyperplasia and localized treatment pathways into an executable set of rules, it performs step-by-step judgments on the input patient data to generate interpretable first-line decision recommendations. The role of the rule engine is to ensure the interpretability of decisions and adherence to clinical norms, providing physicians with clear decision-making basis. The ATA guidelines refer to the "Guidelines for the Management of Adult Thyroid Nodules and Differentiated Thyroid Cancer." These guidelines provide systematic recommendations based on evidence-based medicine for thyroid nodule risk assessment, fine-needle aspiration biopsy indications, surgical approach selection, postoperative management, and follow-up monitoring. For example, taking a 45-year-old female patient as an example, ultrasound showed a left lobe nodule with a diameter of 1.5cm, microcalcifications and an aspect ratio >1. The rule engine performed the following judgment: nodule size ≥1cm; ultrasound features suspicious (microcalcifications, aspect ratio >1); output the first decision recommendation: surgical intervention.

[0029] Patient data is input into a pre-trained graph neural network (GNN) corresponding to a data prediction model, i.e., a GNN model. GNNs are deep learning models used to process graph-structured data, capable of directly learning from graphs composed of nodes and edges, capturing complex dependencies and global structural information between nodes. In the clinical decision support system based on a thyroid atypical hyperplasia database of this application, the GNN model is used as a data engine to learn deep relational features from a knowledge graph constructed from multimodal data, generating risk prediction results. For example, the GNN model can use patients, nodules, genes, and pathology as nodes, and clinical relationships as edges, with message passing through a multi-layer graph convolutional network. For each layer, each node aggregates information from its neighboring nodes. The aggregation function uses mean pooling, and the update function fuses the node's previous layer representation with the aggregated message to generate the node's current layer representation. After three layers of message passing, the final representations of all nodes are read out, and the readout results are input into a fully connected layer, where a sigmoid activation function outputs the lesion risk prediction value.

[0030] S300, based on the processing feedback data of thyroid hyperplasia, the multimodal database and the processing strategy are updated through incremental learning and federated learning mechanisms.

[0031] In this embodiment, updated data and features generated during incremental learning and federated learning are added to the multimodal database. Complete data from newly acquired patients, including clinical records, ultrasound images, and pathology reports, are incrementally added. Simultaneously, new features learned during federated learning are associated and updated in the knowledge graph, enabling dynamic expansion and updating of the multimodal database. Incremental learning is a machine learning paradigm where the model continuously learns new knowledge from newly arriving data while retaining previously learned knowledge, without needing to retrain the entire model from scratch. This application utilizes incremental learning to trigger incremental updates based on clinical feedback data, allowing the system to quickly absorb new case patterns and optimize prediction accuracy. Federated learning is a distributed machine learning paradigm in which multiple participants collaboratively train a shared global model by exchanging intermediate results such as model parameters or gradients without sharing the original data. Through homomorphic encryption and differential privacy technology, it achieves cross-center model co-evolution while protecting the data privacy of all parties, enabling the system to absorb knowledge from multiple sources and improve the diagnostic accuracy of thyroid atypical hyperplasia. In conjunction with incremental learning, it jointly builds a self-evolving clinical decision support system that can quickly respond to new local data and benefit from global knowledge.

[0032] Specifically, during incremental learning and federated learning, all newly generated data and features are incorporated into the multimodal database. This includes: newly added complete patient data being added to the multimodal feature matrix after feature extraction; high-value samples identified in incremental learning having corresponding patient nodes, nodule nodes, and pathology nodes added to the knowledge graph, and establishing relational edges such as "patient-having-nodule" and "nodule-diagnosed-pathology"; and new features learned during federated learning updating the weights of corresponding edges in the knowledge graph. Through a feedback-driven continuous learning mechanism, the system achieves dynamic expansion of the multimodal database and continuous optimization of decision-making strategies, providing more accurate and reliable diagnostic and treatment support for thyroid hyperplasia in clinical practice.

[0033] The method described in this application deeply integrates and standardizes multi-source heterogeneous data related to thyroid atypical hyperplasia, including electronic medical records, ultrasound images, gene testing, and pathology reports. It uses SNOMED-CT terminology mapping to eliminate semantic ambiguity and Z-score standardization to eliminate differences in multi-center imaging equipment. This constructs a multimodal knowledge graph containing patients, nodules, genes, pathological nodes, and semantic relationships, solving the technical problem of data fragmentation in traditional systems. A dual-engine decision-making architecture combining a rule engine and a graph neural network is established, and the NSGA-II multi-objective optimization algorithm is used to dynamically fuse the outputs of both to obtain the optimal trade-off processing strategy. A co-evolutionary mechanism of incremental learning and federated learning based on clinical feedback is established, providing accurate and continuously optimized clinical decision support for the diagnosis and treatment of thyroid hyperplasia.

[0034] In one embodiment of this application, such as Figure 2 As shown, the step of extracting features from the multi-source heterogeneous data and constructing a multimodal database based on the feature-extracted multi-source heterogeneous data includes: S110, based on the data format and content characteristics of the multi-source heterogeneous data, the multi-source heterogeneous data is divided into text data and image data; S120, perform terminology standardization mapping on the text data, and extract semantic relationships between clinical entities based on the terminology standardization text data to obtain structured text feature vectors; S130, perform image feature extraction on the image data to obtain texture features and morphological features in the image data, and perform standard deviation standardization and binning discretization on the extracted image features to generate structured image feature vectors. S140, the structured text feature vector and the structured image feature vector are fused to obtain multimodal feature matrix data, and a multimodal database is constructed based on the multimodal feature matrix data.

[0035] This embodiment divides multi-source heterogeneous data into text and image categories, performs feature extraction and processing separately, and finally merges them to construct a multimodal database, providing a data foundation for subsequent dual-engine decision-making. Specifically, the system collects multi-source heterogeneous data related to thyroid hyperplasia through data interfaces. The collected data includes: electronic medical record data obtained from the hospital information system; thyroid ultrasound image data obtained from the image archiving and communication system, including grayscale images of nodules and color Doppler blood flow images; gene testing report data obtained from the gene testing platform; and biopsy and postoperative pathology report data obtained from the pathology system. These data come from different sources and have different formats, constituting multi-source heterogeneous data. The system first divides the collected multi-source heterogeneous data into text data and image data according to the data format and content characteristics. Text data includes descriptive text such as the progress notes in electronic medical records, mutation site descriptions in gene testing reports, and cytological feature descriptions and diagnostic conclusions in pathology reports. Image data includes grayscale images and color Doppler blood flow images generated from thyroid ultrasound examinations.

[0036] For text-based data, the system employs natural language processing (NLP) technology for standardized terminology mapping. Taking a pathology report as an example, the report states, "Ultrasound-guided fine-needle aspiration biopsy revealed a nodule in the left lobe of the thyroid gland; pathological diagnosis: Bethesda Class III, atypical cells of indeterminate significance." The system first identifies the medical terminology entities, including "left lobe thyroid nodule," "Bethesda Class III," and "atypical cells." Then, it uses the SNOMED-CT medical terminology system to standardize these terms, mapping "left lobe thyroid nodule" to the standard code "237488005," "Bethesda Class III" to "446580001," and "atypical cells" to "164870008." SNOMED-CT, as an internationally recognized clinical terminology standard, ensures semantic consistency across text data from different sources.

[0037] Building upon terminology standardization, the system further extracts semantic relationships between clinical entities. From the aforementioned pathology report, the system identifies a "located" relationship between the entity "nodule" and the entity "left thyroid lobe," meaning the nodule is located in the left thyroid lobe; a "diagnosed as" relationship between the entity "pathological diagnosis" and the entity "nodule," meaning the pathological diagnosis is nodule-like; and a "possesses characteristic" relationship between the entity "nodule" and the "atypical cell" feature, meaning the nodule possesses atypical cell characteristics. The system then structurally represents these relationships, generating a structured text feature vector containing entity type, entity code, and relationship type. For example, for this nodule, the generated feature vector includes: [entity: nodule (237488005)] — [relation: located at (704319004)] → [entity: left lobe of thyroid (91134007)], [entity: pathological diagnosis (129265001)] — [relation: diagnosed as (246090004)] → [entity: nodule (237488005)], [entity: nodule (237488005)] — [relation: has features (116676008)] → [feature: atypical cells (164870008)].

[0038] For imaging data, taking another patient as an example, their thyroid ultrasound image showed a 1.5 cm diameter nodule in the left lobe. The system first preprocesses the ultrasound image, including denoising, enhancement, and nodule region segmentation, extracting the area where the nodule is located. Nodule segmentation refers to the process of accurately delineating the boundary of the thyroid nodule from the ultrasound image, classifying each pixel in the image as either a "nodule" or "background," thereby obtaining the precise outline of the nodule. For example, the U-Net model based on deep learning can be used for automatic segmentation. This model is pre-trained on thousands of labeled images and can accurately identify nodule boundaries.

[0039] After segmentation, the system extracts multi-dimensional image features from the nodule region. Regarding texture features, the system calculates the gray-level co-occurrence matrix of the nodule, extracting feature parameters such as contrast, correlation, energy, and homogeneity. Simultaneously, wavelet transform is performed to extract wavelet coefficient statistics at different scales. The nodule's gray-level co-occurrence matrix shows a contrast of 0.78, a correlation of 0.32, an energy of 0.15, and a homogeneity of 0.41. Regarding morphological features, the system measures the nodule's maximum diameter as 1.5 cm and minimum diameter as 1.2 cm, calculating an aspect ratio of 1.25. Analysis of the nodule's edge indicates blurred boundary clarity and irregular morphology. The system also records the nodule's echogenicity as hypoechoic and the presence of microcalcifications. Regarding blood flow features, Doppler ultrasound shows abundant blood flow signals in the nodule, reaching Adler grade 3, with a resistance index of 0.78, a pulsatility index of 1.65, a maximum blood flow velocity of 45 cm / s, and a minimum blood flow velocity of 8 cm / s.

[0040] After extracting the aforementioned image features, the system performs Z-score standardization. Z-score standardization is a data preprocessing method that transforms data of different dimensions and orders of magnitude into a uniform scale. It standardizes the data based on the mean and standard deviation of the original data, ensuring that the standardized data conforms to a standard normal distribution.

[0041] After standardization, the system performs binning and discretization. Binning and discretization is a data processing method that converts continuous numerical features into discrete categorical features. It divides the range of the original continuous values ​​into several intervals and then maps each original value to its corresponding interval. For example, for the resistance index, a custom binning method based on clinical guidelines is used, dividing the nodule into three categories according to clinical research consensus thresholds: less than 0.55 for low resistance, 0.55 to 0.70 for medium resistance, and greater than 0.70 for high resistance. The nodule's resistance index before standardization is 0.78, belonging to the high resistance group, and its discretization code is 3. For nodule size, the clinical thresholds recommended by the ATA guidelines are used for binning: less than 1 cm for micro nodules, 1 to 2 cm for small nodules, 2 to 4 cm for medium nodules, and greater than 4 cm for large nodules. The nodule's diameter is 1.5 cm, belonging to the small nodule category, and its discretization code is 2. For aspect ratio, according to the TI-RADS classification threshold, it is divided into two categories: less than 1 and greater than or equal to 1. The aspect ratio of this nodule is 1.25, which belongs to the category greater than or equal to 1, and its discretization code is 1.

[0042] After completing the above processing, the system fuses the structured text feature vector with the structured image feature vector. For this patient, the structured text feature vector includes information extracted from the electronic medical record such as the patient's age of 45 years, female, and no family history; information extracted from the gene testing report such as BRAF gene wild type and TERT gene wild type; and Bethesda III diagnostic information extracted from the pathology report. The structured image feature vector includes texture feature vector [0.78, 0.32, 0.15, 0.41, 23.5, 18.7], morphological feature vector [1.5, 1.2, 1.25, blurred, irregular, hypoechoic, microcalcification], and blood flow feature vector [3, 0.78, 1.65, 45, 8].

[0043] The system horizontally concatenates these two feature vectors to construct a multimodal feature matrix containing patient identifier, nodule identifier, feature name, and feature value. Each row of this matrix corresponds to a nodule of a patient, and each column corresponds to a feature dimension. For this patient, one row of the feature matrix includes: Patient ID "P001", Nodule ID "N001", Age 45, Gender Female, BRAF status Wildtype, TERT status Wildtype, Bethesda grade 3, Nodule size 1.5 cm, Aspect ratio 1.25, Blurred borders, Irregular shape, Hypoechoic, Calcification and microcalcification, Blood flow grade 3, Resistance index 0.78, etc., totaling 32 feature dimensions.

[0044] Based on the constructed multimodal feature matrix, the system further generates knowledge graph nodes. Taking patient P001 as an example, the system creates a patient node with attributes including age 45 and female. It creates a nodule node N001 with attributes including size 1.5 cm, aspect ratio 1.25, blurred boundaries, and low echogenicity. It creates gene nodes BRAF and TERT with attributes of wild-type and wild-type, respectively. It creates a pathology node with the attribute of Bethesda III class. Then, based on the previously extracted semantic relationships, it constructs association edges between nodes: a "have" relationship edge between the patient node and the nodule node, a "expression" relationship edge between the nodule node and the gene node, and a "diagnosed as" relationship edge between the nodule node and the pathology node. Finally, a thyroid hyperplasia-specific knowledge graph containing patient, nodule, gene, and pathology nodes is formed and stored in a graph database as a multimodal database.

[0045] This embodiment transforms heterogeneous data scattered across multiple source systems into structured multimodal feature matrices and knowledge graphs using the methods described above, providing a high-quality data foundation for subsequent rule engine reasoning and graph neural network analysis.

[0046] In one embodiment of this application, such as Figure 3As shown, the process of obtaining a corresponding processing strategy based on the data contained in the multimodal database, using the engine decision tree corresponding to the processing rules for thyroid hyperplasia, and the data prediction model corresponding to the graph neural network, includes: S210, The patient data in the multimodal database is input into the preset rule engine. The rule engine analyzes the patient data based on the clinical guideline rule base for thyroid hyperplasia and the localized rule base, and generates a first decision suggestion through the engine decision tree. S220, The patient data in the multimodal database is input into a preset graph neural network model. The graph neural network model uses the features corresponding to the patient's symptoms as nodes and the relationships between clinical entities as edges. Through a multi-layer message passing mechanism, deep correlation features between multimodal data are extracted to generate risk prediction results. S230, using a multi-objective optimization algorithm, dynamically weights and fuses the first decision suggestion and the risk prediction result to generate a corresponding processing strategy.

[0047] In this embodiment, by working together with a rule engine and a graph neural network, the interpretability of clinical logic and the predictive accuracy of the data model are balanced. A multi-objective optimization algorithm is used to dynamically fuse the outputs of the two to generate personalized clinical treatment strategies.

[0048] Specifically, the system has completed the construction of a multimodal database containing complete data on multiple patients with thyroid nodules, including electronic medical records, ultrasound images, genetic testing reports, and pathology results. For example, if a new patient is enrolled and requires clinical decision support.

[0049] The patient was a 45-year-old female who presented with a thyroid nodule discovered during a routine physical examination. Structured data retrieved from a multimodal database revealed the following: Ultrasound examination showed a hypoechoic nodule in the left lobe of the thyroid gland, measuring 1.5 cm × 1.2 cm, with an aspect ratio of 1.25, indistinct borders, and internal microcalcifications. Color Doppler ultrasound showed abundant blood flow (Adler grade 3) and a resistance index of 0.78. Genetic testing showed a positive BRAF V600E gene mutation. Fine-needle aspiration biopsy indicated Bethesda III type, atypia cells of indeterminate significance. The patient had no family history of thyroid cancer and no history of neck radiation exposure.

[0050] The system first inputs patient data into a pre-defined rule engine. This rule engine has a built-in clinical guideline rule base based on the ATA guidelines (Guidelines for the Management of Adult Thyroid Nodules and Differentiated Thyroid Cancer), and also incorporates the hospital's localized rule base for thyroid treatment pathways. The rule engine uses a decision tree structure to parse the patient data level by level. After the rule engine starts, it makes judgments according to the hierarchical structure of the decision tree. For example, the first layer is the nodule size assessment node: the patient's nodule size is 1.5 cm, which is greater than or equal to the 1 cm biopsy threshold recommended by the ATA guidelines, so it proceeds to the next layer for judgment. The second layer is the ultrasound feature assessment node: the rule engine identifies several suspicious features in the patient's ultrasound image, including microcalcifications, aspect ratio greater than 1, blurred boundaries, hypoechoic areas, and abundant blood flow signals. According to the ATA guidelines, the presence of two or more suspicious features is considered suspicious on ultrasound, so the patient passes the ultrasound feature assessment and proceeds to the next layer for judgment. The third layer is the gene testing trigger node: the rule engine detects that the patient has undergone gene testing, and the test result is positive for the BRAF V600E gene mutation. Based on the rule in the localized rule base that "BRAF mutation positivity is highly associated with thyroid cancer", the recommended surgical pathway is directly triggered.

[0051] The rule engine outputs the first decision recommendation as "surgical intervention," along with a complete explanation of the decision path: nodule size ≥1cm → suspicious ultrasound features (microcalcifications, aspect ratio >1, blurred borders) → BRAF gene mutation positive → surgical treatment is recommended based on ATA guidelines and hospital rules. This explanation is output along with the decision recommendation, providing clinicians with a clear basis for decision-making.

[0052] The system simultaneously inputs the same patient's data into a pre-set graph neural network model. This model is pre-trained on a multimodal database and employs a graph convolutional network architecture.

[0053] Specifically, the parameters involved in the detailed description of the Graph Neural Network (GNN) model are explained as follows: : Represents a node in the graph, where each node represents a clinical entity, such as a patient, nodule, gene, or pathological feature. Node set It includes all entities.

[0054] : Indicates a connecting node in the graph and Edges represent semantic relationships between entities, such as "belonging to," "expressing," and "associating." Edge set The interaction structure between entities is defined.

[0055] :node The initial feature vector is obtained by preprocessing (such as standardization and embedding) the original features of the corresponding entity (such as age, nodule size, gene mutation status, pathological grade, etc.) and serves as the starting point for model input.

[0056] The message passing formula for a graph neural network is shown below: For the Layers, nodes Aggregate neighbor information:

[0057]

[0058] in, : Indicates the index of the current network layer, with a total of 100 layers. Each layer updates the node representation via message passing.

[0059] :node In the The hidden layer representation vector of a layer. At layer 0, .

[0060] :node The set of neighboring nodes, that is, all nodes that are related to... Nodes that are directly connected by an edge.

[0061] : in the Layers, nodes The information vector obtained by aggregating from its neighboring nodes is called a "message".

[0062] : No. The aggregation function of a layer is used to represent the previous layer of neighboring nodes. Combined into a fixed-length message Common options include mean (averaging), max pooling (maximizing the value of each dimension), summation, or attention-weighted methods.

[0063] : No. The layer update function is used to update the representation of the node's previous layer. messages obtained from aggregation Merge to generate a new representation for the current layer. Commonly used update functions include gated recurrent units (GRU), long short-term memory networks (LSTM), and fully connected layers with nonlinear activation.

[0064] The final prediction formula of the graph neural network is shown below:

[0065] The total number of network layers, i.e., the number of layers passed through. The final node representation is obtained after this message passing.

[0066] : Read function, used to read the data after... The updated representation of all nodes in the layer This is integrated into a graph-level global representation. Common readout operations include summation, averaging, maximization, or more complex pooling methods. This step is typically used for graph-level tasks (such as graph classification), but for node-level tasks (such as predicting the malignancy risk of a specific node), the final representation of the target node can be used directly, skipping READOUT.

[0067] : The weight matrix of the output layer, used to linearly transform the global representation (or node representation) output by READOUT to the target dimension.

[0068] : The bias term of the output layer.

[0069] Activation function, specifically the sigmoid function, maps the result of a linear transformation to... An interval is used as a prediction probability (e.g., the probability of malignant risk).

[0070] The model's final prediction output, such as nodes. The probability of malignancy or the classification result at the graph level.

[0071] These parameters collectively define how GNNs learn node representations through multi-layer message passing and ultimately generate prediction results.

[0072] Specifically, the graph structure is constructed as follows: patients, nodules, genes, and pathological features are used as nodes, and the relationships between clinical entities are used as edges. For this patient case, the graph neural network constructs the following nodes: Patient node P001, with attributes including age 45 years, female, and no family history; Nodule node N001, with attributes including size 1.5 cm, aspect ratio 1.25, blurred boundaries, microcalcifications, hypoechoic, blood flow grade 3, and resistance index 0.78; Gene node BRAF, with the attribute of mutant V600E; Pathological node B001, with attributes of Bethesda III and atypical cells. The relationships between nodes include: the "having" relationship between patient node P001 and nodule node N001; the "expression" relationship between nodule node N001 and gene node BRAF; the "diagnosed as" relationship between nodule node N001 and pathology node B001; and the "having features" relationship between nodule node N001 and ultrasound features, including microcalcification features, aspect ratio greater than 1 features, and hypoechoic features.

[0073] The graph neural network initiates a multi-layer message passing mechanism. In the first layer, the nodule node N001 aggregates information from its neighboring nodes: demographic characteristics such as age and gender from the patient node, BRAF mutation status from the gene node, Bethesda classification from the pathology node, and various imaging features from the ultrasound feature node. The aggregation function employs an attention mechanism, dynamically assigning weights based on the importance of different neighboring nodes in determining the nodule's nature. Based on the attention parameters obtained during model pre-training, gene nodes and ultrasound feature nodes are given higher weights because they are key features for predicting malignancy.

[0074] After aggregation, the node merges the previous layer representation with the aggregated message through an update function to generate the first layer representation h_N001^(1). The update function uses a gated loop unit, which can selectively retain important information and filter out noise information.

[0075] In the second-layer network, the scope of message passing is further expanded. Nodule nodes not only aggregate information from their direct neighbors but can also obtain indirect association information through two-hop paths. For example, through the path "nodule—expression → BRAF gene—associated with → BRAF-related ultrasound features," the model learns the strong association patterns between BRAF mutations and features such as microcalcifications and hypoechoic foci in historical data; through the path "nodule—diagnosed as → Bethesda III—associated with → previously malignant transformation cases," the model learns the malignancy probability of Bethesda III under specific feature combinations.

[0076] After three layers of message passing, the system reads the final representation of the nodule node. The final representation of the nodule node is then input into a fully connected layer, and the sigmoid activation function outputs a predicted lesion risk value. The model outputs a 94% probability of malignancy for this patient, with a confidence level of 91%.

[0077] The graph neural network simultaneously generates interpretable information, which is visualized through attention weights. The features that contribute the most to the prediction results are, in descending order: BRAF gene mutation (weight 0.32), microcalcification (weight 0.18), aspect ratio greater than 1 (weight 0.15), blood flow grade 3 (weight 0.12), resistance index 0.78 (weight 0.10), and Bethesda class III (weight 0.08).

[0078] The system obtains the first decision suggestion from the rule engine as "surgical intervention" (quantifiable as a decision value of 1.0), while the risk prediction result from the graph neural network is a malignancy probability of 94% (decision value of 0.94). Both are highly consistent in direction, but still require quantification and fusion through multi-objective optimization to generate the final treatment strategy and determine the specific surgical plan.

[0079] The system activates the NSGA-II multi-objective optimization algorithm, with three clinical objectives as optimization targets: tumor radical resection rate (maximizing the probability of complete surgical resection of malignant lesions), functional preservation rate (maximizing the degree of thyroid function preservation after surgery), and complication risk (minimizing the probability of postoperative surgery-related complications).

[0080] For this patient, the multi-objective optimization algorithm searched for the optimal fusion weights on the Pareto front based on the patient's individual characteristics and clinical preferences. The patient was 45 years old, a high-incidence age group for thyroid cancer; the nodule was located in the middle of the left lobe of the thyroid gland, far from the isthmus and the entrance of the recurrent laryngeal nerve into the larynx, so surgery would have a relatively small impact on function; the patient had no underlying diseases such as hypertension or diabetes, and tolerated surgery well. Taking these factors into consideration, the algorithm determined the weight α of the rule engine to be 0.6 and the weight β of the graph neural network to be 0.4.

[0081] The formula for calculating the fusion decision value is: Decision value = α × Rule engine decision value + β × Graph neural network prediction value. Substituting the values, the calculation is: Decision value = 0.6 × 1.0 + 0.4 × 0.94 = 0.976.

[0082] The system's preset surgical threshold is 0.8. The fusion decision value of 0.976 significantly exceeds the threshold, confirming surgical intervention as the final treatment strategy. The system presents the treatment strategy to clinicians through a visual interface. The left side of the interface displays basic patient information, the middle left side displays the rule engine's decision path (showing each decision node in a tree diagram), the middle right side displays the graph neural network's contribution features (showing attention weights in a heatmap), and the right side displays the final recommended plan and multi-objective optimization results.

[0083] In one embodiment of this application, such as Figure 4 As shown, the processing feedback data based on thyroid hyperplasia updates the multimodal database and the processing strategy through incremental learning and federated learning mechanisms, including: S310, collecting feedback data during the clinical management of thyroid hyperplasia; S320, based on the feedback data, the pathological results are compared with the prediction results to determine the corresponding prediction deviation. When the prediction deviation exceeds a preset threshold, the incremental learning process is triggered to adjust the parameters of the data prediction model corresponding to the graph neural network. S330, based on the feedback data and the multimodal data, the graph neural network model is used to calculate the model gradient data, and the corresponding model parameters are updated based on the model gradient data; S340, the updated data and features generated during incremental learning and federated learning are added to the multimodal database, and the multimodal database is dynamically updated.

[0084] This embodiment collects clinical feedback data, triggers an incremental learning process, and coordinates with a federated learning mechanism to achieve dynamic updates of the multimodal database and continuous optimization of model parameters.

[0085] The system has completed the construction of a multimodal database and the deployment of a dual-engine decision model, and automatically collects various feedback data in the clinical treatment of thyroid hyperplasia through a feedback collector.

[0086] Taking a 58-year-old male patient as an example, ultrasound showed a 2.2 cm diameter nodule in the right lobe, with clear borders, regular shape, no calcification, and negative gene testing. The system predicted a 35% probability of malignancy and recommended "annual follow-up." The postoperative pathology report showed follicular thyroid carcinoma. The system collected feedback data: doctor's adoption record (adoption of the annual follow-up recommendation), postoperative recovery data (no complications), and pathological comparison results (malignancy was inconsistent with the preoperative prediction of 35%). The patient's postoperative pathological results showed a significant difference from the preoperative prediction.

[0087] The prediction bias assessment and incremental learning triggering are described below. The system compares the postoperative pathological results with the preoperative prediction results to calculate the prediction bias. For example, for a patient with a predicted malignancy probability of 35% and an actual malignancy probability of 100%, the absolute error Δ = |0.35-1| = 0.65, or 65%. The system's preset incremental learning triggering threshold is 10%. Since the patient's prediction bias of 65% far exceeds the threshold, the system automatically marks this case as a "high-value learning sample" and adds it to the incremental learning queue. The incremental learning module processes the accumulated high-value samples in batches weekly. The system merges these new samples with the existing historical data and constructs an incremental training set using a weighted sampling strategy. After batch training, the parameters of the graph neural network model are fine-tuned.

[0088] Based on newly added local feedback data and historical multimodal data, the gradients of the graph neural network model are calculated separately. For example, the system trains a graph neural network model based on newly added local patient data and existing data, calculates the local gradient g_A, encrypts the gradient using the Paillier homomorphic encryption algorithm, and generates an encrypted gradient Encrypt(g_A), which is uploaded to the federated learning aggregation server deployed in the health data center through a secure channel. The aggregation server simultaneously receives encrypted gradients uploaded by different hospitals. Homomorphic encryption aggregation is performed in the encrypted state. To meet differential privacy protection requirements, the server calculates the noise scale σ=0.48 according to the preset privacy budget ε=1.0 and sensitivity Δf=0.1, generates a Gaussian noise vector N(0, σ²I), encrypts it, and multiplies it with the aggregated gradient to obtain the noisy aggregated gradient g_noisy = g_total × Encrypt(noise). The server returns g_noisy to each hospital. The aggregated gradient g_aggregated is obtained by decrypting using the local private key, and the global model parameters are updated with a learning rate of 0.01: θ_new = θ_old - 0.01 × g_aggregated. The updated model parameters are immediately deployed to the local clinical decision system for subsequent patient risk assessment.

[0089] In incremental learning and federated learning, all newly generated data and features are incorporated into the multimodal database, enabling dynamic expansion and updating of the database. Specifically, this includes: adding features to newly added case data after feature extraction; updating the knowledge graph of high-value samples, where high-value samples identified in incremental learning have corresponding patient nodes, nodule nodes, and pathology nodes added to the knowledge graph, and establishing relational edges such as "patient—having—nodule" and "nodule—diagnosed as—pathology"; and updating new feature associations, where newly discovered feature associations are shared during federated learning, updating the weights of corresponding edges in the knowledge graph to enhance the model's interpretability and reasoning ability.

[0090] This embodiment utilizes a feedback-driven continuous learning mechanism to dynamically expand the multimodal database and continuously optimize decision-making strategies. The model's performance is continuously improved, providing more accurate and reliable support for the diagnosis and treatment of thyroid atypical hyperplasia.

[0091] In one embodiment of this application, such as Figure 5 As shown, the step of calculating model gradient data using the graph neural network model based on the feedback data and the multimodal data, and updating the corresponding model parameters based on the model gradient data, includes: S3301 calculates model gradient data using a trained graph neural network model based on locally collected feedback data and locally stored multimodal data. S3302, The homomorphic encryption algorithm is used to encrypt the gradient data of the computational model, and the encrypted gradient data of the computational model is sent to the federated learning aggregation server. S3303, the encrypted computational model gradient data is homomorphically encrypted and aggregated through the federated learning aggregation server, and differential privacy noise is added to obtain aggregated noise gradient data. S3304, Decrypt the aggregated noise gradient data to obtain the updated model parameters corresponding to the model gradient data; S3305, Update the graph neural network model based on the updated model parameters.

[0092] This embodiment achieves the collaborative evolution of a multi-center model by using homomorphic encryption and differential privacy technology, while protecting the data privacy of each medical center.

[0093] For example, five hospitals (Hospital A, Hospital B, Hospital C, Hospital D, and Hospital E) jointly participate in federated learning collaboration. Each hospital has deployed the clinical decision support system of this invention and accumulated a local multimodal database. This embodiment takes Hospital A as an example to demonstrate in detail the complete process of model gradient calculation, encrypted upload, secure aggregation, decryption update, and model update.

[0094] First, Hospital A uses locally collected feedback data and locally stored multimodal data to calculate model gradient data using a trained graph neural network model. This graph neural network model consists of an input layer, three graph convolutional layers, a readout layer, and an output layer. The model uses patients, nodules, genes, and pathological features as nodes, and relationships between clinical entities as edges, extracting multimodal features through a message passing mechanism.

[0095] For example, the system randomly selects 256 patients from the local database as the local training batch for this week, including newly added high-value samples. The feature data of these patients is input into the model for forward propagation calculations to obtain the malignancy risk prediction value for each patient. The predicted value is compared with the patient's postoperative pathological gold standard y_i, and the cross-entropy loss function L = -Σ[y_i log( _i) + (1-y_i) log(1- _i)).

[0096] Using the backpropagation algorithm, the system calculates the partial derivatives of the loss function L with respect to the parameters of each layer of the model, obtaining the gradient vector g_A. The dimension of the gradient vector is consistent with the dimension of the model parameters, which is 520,000. Specifically, for the weight matrix W of the graph convolutional layer, the gradient... W represents the sensitivity of the loss to each weight parameter; for the bias term b, the gradient... b represents the sensitivity of the loss to the bias. The system concatenates the gradients of all parameters into a long vector, which serves as the local gradient g_A for this round of federated learning.

[0097] After completing the gradient calculation, Hospital A uses the Paillier homomorphic encryption algorithm to encrypt the local gradient g_A. Paillier encryption is a public-key encryption system with the homomorphic addition property, meaning that the addition of ciphertext equals the encryption result of the addition of plaintext. The system first generates a Paillier key pair. The key generation process is as follows: select two large prime numbers p and q, calculate n = p × q, λ = lcm(p-1, q-1); select a random integer g, satisfying that the order of g mod n² is a multiple of n; the public key is (n, g), and the private key is λ. In this example, p and q are chosen as 512-bit prime numbers, and the generated n is 1024 bits, ensuring that the encryption strength meets the security requirements of medical data. Since Paillier encryption requires the plaintext to be an integer, while the gradient g_A is a floating-point number, the system needs to quantize and encode the gradient. The specific method is as follows: multiply each element of the gradient vector g_A by 10^6, round it to the nearest integer, and convert the floating-point number to an integer representation; for negative values, perform offset processing through the modulo n² space to ensure that all integers fall within the range [0, n²).

[0098] After quantization and encoding, the system encrypts each gradient element using a public key. For example, the encryption algorithm is as follows: for a plaintext integer m, a random number r is selected, and the ciphertext c = g^m × r^n mod n² is calculated. The encrypted gradient vector Encrypt(g_A) consists of 520,000 ciphertext elements, each ciphertext being 2048 bits in size, for a total encrypted data size of approximately 133 megabytes. After encryption, Hospital A sends the encrypted gradient Encrypt(g_A) to the federated learning aggregation server via a secure transmission protocol. A TLS encrypted channel is used during transmission to ensure the confidentiality and integrity of the data during transmission.

[0099] The federated learning aggregation server is deployed in the health data center and is responsible for receiving and securely aggregating the encryption gradients from various hospitals. The server simultaneously receives encryption gradients uploaded by hospitals A, B, C, D, and E: Encrypt(g_A), Encrypt(g_B), Encrypt(g_C), Encrypt(g_D), and Encrypt(g_E). The server verifies that each encryption gradient is correctly formatted and originates from a trusted hospital. The server performs homomorphic encryption aggregation in ciphertext. Due to the homomorphic nature of Paillier encryption, ciphertext multiplication (in modulo n² space) corresponds to plaintext addition. The aggregation formula is: g_total = Encrypt(g_A) × Encrypt(g_B) × Encrypt(g_C) × Encrypt(g_D) × Encrypt(g_E) mod n². This operation is equivalent to calculating the sum of the five gradients in ciphertext.

[0100] During the aggregation process, the server cannot decrypt the gradient of any individual hospital, nor can it know the specific value of the aggregation result, ensuring the data privacy of each hospital. After aggregation, the server needs to add differential privacy noise to prevent inference of individual hospital data through differential attacks. The server calculates the number of participating hospitals in this round, k=5. For example, the data volume statistics for each hospital are: Hospital A 1401 cases, Hospital B 1520 cases, Hospital C 1280 cases, Hospital D 1610 cases, Hospital E 1450 cases, for a total of 7261 cases. The server dynamically adjusts the noise scale σ according to the preset privacy budget ε and sensitivity Δf. For example, the consortium sets the privacy budget ε=1.0, indicating that the federated learning process satisfies (1.0, 10...). -5 - Differential privacy protection. The sensitivity Δf is calculated based on the upper limit of the gradient norm and set to 0.1, meaning that the maximum influence of any single sample on the gradient does not exceed 0.1. The noise scale σ is calculated as: σ = Δf × √(2 ln(1.25 / δ)) / ε, where δ=10 -5 Substituting the values, we get σ = 0.1 × √(2 ln(1.25 × 10⁻⁶)). 5 )) / 1.0 ≈ 0.1 × √(2 × 11.5) ≈ 0.1 × 4.8 ≈ 0.48.

[0101] The server generates a Gaussian noise vector N(0, σ²I) with the same dimensions as the gradient, and a dimension of 520,000. The noise vector is then quantized and encoded in the same way as the gradient, converted to an integer representation, and encrypted using the Paillier public key to obtain the encrypted noise Encrypt(noise).

[0102] The server multiplies the encrypted noise by the aggregate gradient to obtain a noisy aggregate gradient: g_noisy = g_total × Encrypt(noise) mod n². According to the Paillier homomorphism property, this is equivalent to adding noise to the aggregate gradient. After adding the noise, the server returns the noisy aggregate gradient g_noisy to each participating hospital. The return process also uses TLS encryption.

[0103] Hospital A receives the noisy aggregated gradient g_noisy returned by the aggregation server. The system decrypts g_noisy using the locally stored Paillier private key. For example, the decryption process is as follows: For the ciphertext c, calculate the plaintext m = L(c^λmod n²) × μ mod n, where L(u)=(u-1) / n, and μ is a pre-calculated parameter. After decryption, an integer vector is obtained, with each element corresponding to the quantized value of the aggregated gradient. The system performs dequantization on the decrypted integer vector: dividing each integer by 10^6 to restore the aggregated gradient g_aggregated in floating-point representation. At this point, g_aggregated already contains the gradient information of the five hospitals and the added differential privacy noise. The system updates the global model parameters based on the aggregated gradient, with the update formula: θ_new= θ_old - η × g_aggregated, where η is the learning rate of the federated learning. In this example, the federation uniformly sets the learning rate η=0.01. The update process proceeds layer by layer: For the weight matrix W of the graph convolutional layer, the update formula is W_new = W_current -0.01 × W_aggregated; For the bias term b, the update formula is b_new = b_current - 0.01 × b_aggregated. The system iterates through all 520,000 parameters, completing one round of updates. After the update, the system saves the new model parameters θ_new and deploys them to the clinical decision system. The system also records the metadata of this round of federated learning: round number (round 12), number of participating hospitals (5), total number of cases (7261), learning rate (0.01), privacy budget (ε=1.0), etc., and stores them in the federated learning log.

[0104] Through the federated learning mechanism, hospitals have achieved cross-center knowledge flow and collaborative model evolution while protecting patient data privacy, thus jointly improving the accuracy and reliability of diagnosis and treatment of thyroid atypical hyperplasia.

[0105] In one embodiment of this application, the feedback data includes at least one of the following: records of doctors' adoption and adjustment of treatment strategies, postoperative recovery data and complication occurrence data of patients, and comparison data between postoperative pathology results and preoperative prediction results.

[0106] In this embodiment, the system can collect three main types of feedback data: the doctor's adoption and adjustment records of the treatment strategy, the patient's postoperative recovery data and the occurrence of complications, and the comparison data between postoperative pathology results and preoperative prediction results. Regarding the doctor's adoption, the system records details of the surgeon's adoption or modification of the system's suggested surgical methods, incision locations, and intraoperative precautions. For example, the doctor added a "recurrent laryngeal nerve exploration" step to the system's suggested "right thyroid lobe and isthmus resection + central lymph node dissection," and adjusted the incision location from "2cm above the suprasternal notch" to "2.5cm" to suit the patient's short neck. These adjustment reasons and operational records are fully saved. Regarding postoperative recovery data, the system automatically collects records of the patient's length of hospital stay, drainage volume, pain score, recovery of hoarseness, and the occurrence and management of complications such as hypocalcemia by connecting to the hospital's electronic medical record system. It also includes long-term follow-up data such as thyroid function, ultrasound examination, and quality of life scores from 3 to 12 months postoperatively. In terms of pathological comparison, the system accurately compares information such as tumor size, subtype, capsule invasion, vascular invasion, lymph node metastasis, and molecular markers in the postoperative pathology report with the preoperative risk prediction results. For example, in this case, the pathology confirmed the presence of minor capsule invasion, consistent with the preoperative high-risk prediction, but the absence of clear vascular invasion deviated from the preoperative predicted probability of 78%. Based on this feedback data, the system regularly performs incremental training on the risk prediction model, optimizes feature weights and risk thresholds, incorporates recurrent laryngeal nerve exploration into routine recommendations based on physician adoption statistics, adds the "posterior edge contact area of ​​the nodule" risk feature based on complication analysis, and uses manually corrected boundary regions by physicians as difficult case samples to fine-tune the segmentation model.

[0107] In one embodiment of this application, such as Figure 6 As shown, the method further includes: S400: Collect the decision results output by the clinical system and the actual clinical results, compare the decision results with the actual results, and obtain the corresponding prediction deviation; S500, determine whether the prediction deviation exceeds the preset threshold; S600, if the prediction deviation exceeds the preset threshold, an incremental learning process is triggered to update the parameters of the data prediction model corresponding to the graph neural network; S700: If the prediction deviation does not exceed the preset threshold, only the feedback data is recorded, and the incremental learning process is not triggered.

[0108] This embodiment compares the decision results output by the clinical system with the actual clinical results to dynamically determine whether incremental learning needs to be triggered, thereby enabling the system to evolve and ensuring continuous model optimization while avoiding unnecessary computational overhead.

[0109] The system has completed the construction of a multimodal database and the deployment of a dual-engine decision-making model, and continuously collects clinical feedback data. The system collects the decision results output by the clinical system and the corresponding actual clinical outcomes in real time through an interface. Taking treated patients as an example, the system records detailed predicted data and follow-up results for each patient.

[0110] For example, a 45-year-old female patient's ultrasound showed a 1.5 cm diameter nodule in her left lobe, with microcalcifications and an aspect ratio >1, and a positive BRAF V600E mutation. The system's dual-engine decision generation strategy was "surgery recommended," with a decision confidence level of 92% and a malignancy risk probability of 94% output by the graph neural network. After surgery, the postoperative pathology report confirmed papillary thyroid carcinoma, with a tumor size of 1.3 cm and no lymph node metastasis. The system automatically collected this actual result (malignancy) and established a correlation with the preoperative decision result. A 58-year-old male patient's ultrasound showed a 2.2 cm diameter nodule in his right lobe, with clear borders, regular shape, no calcification, and negative gene testing. The system decision was "annual follow-up," with a malignancy risk probability of 35% output by the graph neural network. The postoperative pathology report showed follicular thyroid carcinoma, a high degree of malignancy. The system collected feedback data: the doctor adopted the annual follow-up recommendation, but the actual result was malignancy.

[0111] The system compares the decision result with the actual result for each patient and calculates the prediction bias. For example, for classification decision and probability prediction, the system uses absolute error as the bias metric: Δ = |p_pred - p_actual|, where p_pred is the predicted malignancy probability and p_actual is the probability representation of the actual result (malignancy is 1, benign is 0). For the first patient, the predicted malignancy probability is 94%, the actual malignancy is (1), and the absolute error Δ = |0.94-1| = 0.06, or 6%. For the second patient, the predicted malignancy probability is 35%, the actual malignancy is (1), and the absolute error Δ = |0.35-1| = 0.65, or 65%. The system presets an incremental learning trigger threshold of 10%. The system performs threshold judgment on the prediction bias for each patient: First patient: bias 6% < 10%, not exceeding the threshold. Second patient: bias 65% > 10%, exceeding the threshold. Based on the judgment results, for the first case where the patient's deviation does not exceed the threshold, the system only records the feedback data and does not trigger incremental learning; for the second case where the patient's deviation exceeds the threshold, the system triggers the incremental learning process.

[0112] The system detected that the prediction deviation of the second case exceeded the threshold and automatically triggered the incremental learning process. The incremental learning module performs the following operations: for example, it moves the complete data of the second case from the temporary storage area into the incremental learning queue. The case data includes: electronic medical records, ultrasound images, gene testing results, biopsy pathology reports, postoperative pathology reports, system decision records, etc. The data includes: 58-year-old male, ultrasound features (clear boundaries, regular morphology, no calcification, no blood flow), negative gene testing, 35% system prediction, and postoperative pathology of follicular carcinoma. The incremental learning module calculates the loss contribution of each high-value sample. The feature data of Mr. Li is input into the current graph neural network model for forward propagation, and the output value is calculated. =0.35; Compared with the true label y=1, the cross-entropy loss L = -[1×log(0.35) + (1-1)×log(0.65)] = -log(0.35) = 1.05. This loss value is significantly higher than the average loss of 0.3, indicating that this case has high learning value for the current model. The incremental learning module marks the second case as a "high-value sample" and adds it to the incremental training batch.

[0113] For the first case where the prediction deviation does not exceed the threshold, the system only records the feedback data and does not trigger incremental learning. The recorded information includes: patient identification, consultation time, system decision result, actual result, calculated deviation value, deviation type (accurate prediction), doctor adoption status, follow-up results, etc.

[0114] Through a prediction bias-driven incremental learning triggering mechanism, the system achieves accurate identification and selective learning of high-value samples, ensuring continuous model evolution while avoiding unnecessary computational overhead, thus providing efficient and accurate decision support for the diagnosis and treatment of thyroid atypical hyperplasia.

[0115] In one embodiment of this application, such as Figure 7 As shown, the step of dynamically weighting and fusing the first decision suggestion and the risk prediction result using a multi-objective optimization algorithm to generate a corresponding processing strategy includes: S2301, A multi-objective optimization algorithm is used to dynamically fuse the first decision suggestion and the risk prediction result to obtain multiple candidate fusion weight combinations; S2302, determine a target fusion weight combination from multiple candidate fusion weight combinations, wherein the target fusion weight combination represents the optimal decision-making method among the multiple candidate fusion weight combinations; S2303, Based on the target fusion weight combination, generate the corresponding clinical treatment strategy.

[0116] This embodiment proposes a method that generates multiple candidate fusion weight combinations using the NSGA-II algorithm, selects the target fusion weight combination from these, and generates a personalized clinical treatment strategy based on this combination, achieving an optimal trade-off between the interpretability of the rule engine and the prediction accuracy of the graph neural network. The system has completed the construction of a multimodal database and the deployment of a dual-engine decision model.

[0117] For example, in a complex case requiring decision support, the system initiated a multi-objective optimization fusion process. A 62-year-old female patient presented with multiple nodules in both lobes of the thyroid gland on ultrasound. The left lobe nodule was 3.2 cm in diameter with indistinct borders, internal microcalcifications, and rich blood flow signal on color Doppler (Adler grade 3), with a resistance index of 0.82. The right lobe nodule was 1.8 cm in diameter with clear borders, regular shape, no calcifications, poor blood flow, and a resistance index of 0.54. Genetic testing showed a positive TERT promoter mutation and a negative BRAF result. Fine-needle aspiration biopsy report: left lobe nodule Bethesda V (suspected malignancy), right lobe nodule Bethesda II (benign). The patient had a 10-year history of hypertension, NYHA class II heart failure, no family history of thyroid cancer, and no history of neck radiation exposure.

[0118] The rule engine, based on the ATA guidelines and a local rule base, comprehensively evaluates bilateral lobe nodules before making a decision. Left lobe nodule assessment: Size ≥ 1 cm, suspicious ultrasound features (microcalcifications, indistinct borders, abundant blood flow), positive TERT mutation. According to the rule in the local rule base that "TERT mutation is associated with invasiveness," the first decision recommendation for the left lobe nodule is "surgical intervention," with a recommendation to consider total thyroidectomy. Right lobe nodule assessment: Size ≥ 1.5 cm (ATA guidelines threshold for low-susceptibility nodules via FNA), but atypical ultrasound features and Bethesda category II. The rule engine's first decision recommendation for the right lobe nodule is "annual follow-up."

[0119] The graph neural network model constructs a complex graph structure including two nodules, patients, genes, and pathology. Nodes include: Patient node P062 (attributes: 62 years old, female, hypertension, NYHA class II); Left lobe nodule node N001 (attributes: 3.2cm, indistinct borders, microcalcifications, rich blood flow, resistance index 0.82); Right lobe nodule node N002 (attributes: 1.8cm, clear borders, regular shape, no calcifications, resistance index 0.54); Gene node TERT (attributes: promoter mutation positive); Pathology nodes B001 (attributes: Bethesda V) and B002 (attributes: Bethesda II). Relationship edges are constructed between nodes: Patient—Has → Left lobe nodule, Patient—Has → Right lobe nodule, Left lobe nodule—Expresses → TERT, Left lobe nodule—Diagnosed as → B001, Right lobe nodule—Diagnosed as → B002.

[0120] After three layers of message passing, the graph neural network outputs: a 96% probability of malignancy for the left lobe nodule and a 12% probability of malignancy for the right lobe nodule. Simultaneously, the model notes the patient's age of 62 and NYHA Class II heart function, incorporating surgical risk information into the prediction and outputting a comprehensive risk warning.

[0121] The system initiates the NSGA-II multi-objective optimization algorithm, using three clinical objectives as optimization targets: Objective 1 (tumor eradication rate): Maximize the probability of complete surgical resection of malignant lesions. This objective has a positive association strength of 0.98 with total thyroidectomy, 0.90 with left lobectomy, 0.10 with right lobectomy, and 0.05 with conservative treatment.

[0122] Objective 2 (Function Preservation): To maximize the degree of thyroid function preservation after surgery. The association strength of this objective with conservative treatment was 0.90, with right lobectomy was 0.70, with left lobectomy was 0.50, and with total thyroidectomy was 0.20.

[0123] Objective 3 (Complication Risk): Minimize the probability of postoperative surgery-related complications. This objective has a strong association with conservative treatment (0.95, i.e., low risk), a strong association with right lobectomy (0.75), a strong association with left lobectomy (0.60), and a strong association with total thyroidectomy (0.30, i.e., high risk).

[0124] The decision variables are fused weight vectors, containing rule engine weights α and graph neural network weights β, satisfying α + β = 1 and α, β ≥ 0. Each weight combination corresponds to a fusion strategy, which in turn affects the selection of a specific surgical plan.

[0125] The NSGA-II algorithm begins iterative optimization, generating multiple candidate fusion weight combinations. The algorithm parameters are set as follows: population size 100, maximum number of iterations 200, crossover probability 0.9, mutation probability 0.1, crossover exponent 20, and mutation exponent 20. The algorithm initializes 100 random weight combinations as the first generation population. Each weight combination is a one-dimensional variable α (then β = 1 - α), taking values ​​in the range [0, 1]. The system calculates three objective function values ​​for each weight combination.

[0126] Taking α=0.7 (rule engine weight 0.7, GNN weight 0.3) as an example, the fused decision value tends to favor the rule engine's dominant opinion. Considering the patient's dual nodule situation, the system assesses that total thyroidectomy might be recommended under this weight (high risk of malignancy in the left lobe + presence in the right lobe but low risk). The calculated results are: tumor eradication rate 0.96%, functional preservation 0.25%, and complication risk 0.35 (i.e., the risk value, which needs to be minimized).

[0127] Taking α=0.3 (rule engine weight 0.3, GNN weight 0.7) as an example, the fused decision value tends to favor the more refined evaluation of the GNN. The GNN's judgment of a 12% risk for right lobe nodules may cause the system to favor preserving the right lobe. The calculated results are: tumor eradication rate 0.91, functional preservation 0.55, and complication risk 0.25.

[0128] Taking α=0.5 (balanced weight) as an example, the following results were obtained: tumor radical cure rate 0.94, functional preservation rate 0.40, and complication risk 0.30.

[0129] The algorithm generates a new generation of population through tournament selection, simulated binary crossover, and polynomial mutation. After 200 iterations, the algorithm converges to the Pareto front, generating 50 non-dominated solutions, i.e., 50 candidate fusion weight combinations. These combinations achieve different equilibria among the three objectives, constituting the Pareto optimal solution set.

[0130] The system performs three-dimensional visualization analysis on 50 candidate fusion weight combinations, and the Pareto front presents three typical regions: Region A (Aggressive): α is between 0.65 and 0.80, with a high tumor eradication rate (0.95-0.98), low functional preservation (0.20-0.30), and a high risk of complications (0.35-0.40). The corresponding strategy tends to be total thyroidectomy to completely remove the malignant nodule in the left lobe, but this sacrifices right lobe function and carries a higher surgical risk.

[0131] Region B (Balanced): α is between 0.45 and 0.60, with moderate tumor radicalization rates (0.92-0.94), moderate functional preservation (0.40-0.50), and moderate complication risk (0.28-0.32). The corresponding strategy tends to be left lobectomy + right lobe preservation, balancing radical cure and functional preservation.

[0132] Region C (Conservative): α is between 0.20 and 0.35, with a lower tumor eradication rate (0.88-0.91), higher functional preservation (0.55-0.65), and lower complication risk (0.22-0.26). The corresponding strategy tends to be left lobectomy plus close follow-up of the right lobe, maximizing functional preservation and minimizing surgical risks.

[0133] Based on these factors, the system employs multi-criteria decision analysis to assign weights to the three objectives: complication risk weight 0.45 (considering cardiac function), tumor eradication rate weight 0.35 (left lobe must be treated), and function preservation weight 0.20 (patient's wishes).

[0134] The system calculates a weighted score for each candidate combination: Score = 0.35 × Tumor eradication rate + 0.20 × Functional preservation + 0.45 × (1 - Complication risk). Scores are calculated for the 50 combinations at the Pareto front, and the combination with the highest score is selected as the target fusion weight combination.

[0135] The highest-scoring combination appeared at α=0.48 (rule engine weight 0.48, GNN weight 0.52), belonging to the balanced, conservative region. The corresponding objective function values ​​for this combination are: tumor eradication rate 0.93, functional preservation 0.52, and complication risk 0.27. The weighted score calculation is: 0.35×0.93 + 0.20×0.52 + 0.45×0.73 = 0.3255 + 0.104 + 0.3285 = 0.758, the highest among all combinations.

[0136] Based on the target fusion weight combination α=0.48, the system generates the corresponding clinical treatment strategy. The fusion decision value calculation formula is applied to bilateral lobular nodules: The left lobe nodule fusion decision value = 0.48 × rule engine decision value (left lobe surgery, quantified as 1.0) + 0.52 × GNN prediction value (0.96) = 0.48 + 0.499 = 0.979, which exceeds the surgery threshold of 0.8.

[0137] The decision value for fusion of the right lobe nodule = 0.48 × rule engine decision value (right lobe annual review, quantified as 0.3) + 0.52 × GNN prediction value (0.12) = 0.144 + 0.0624 = 0.206, which is lower than the surgical threshold of 0.8 and also lower than the puncture threshold of 0.5.

[0138] The system comprehensively assesses the situation and outputs the following final treatment strategy: Left thyroid lobectomy is recommended, preserving the right lobe; intraoperative frozen section pathology examination will be performed to determine whether the surgical scope needs to be expanded; postoperatively, the patient will be transferred to the cardiology intensive care unit for close monitoring of cardiac function; the endocrinology department will follow up on thyroid function and supplement levothyroxine if necessary; the right lobe nodule will be re-examined annually with ultrasound to dynamically monitor changes.

[0139] This embodiment uses the NSGA-II multi-objective optimization algorithm to select the most suitable target combination for the individual characteristics of the patient from multiple candidate fusion weight combinations, thereby generating a personalized treatment strategy.

[0140] This application also provides a clinical decision support system based on a specialized database for thyroid atypical hyperplasia, such as... Figure 8 As shown, it includes: A multimodal database construction module is used to collect multi-source heterogeneous data related to thyroid hyperplasia, extract features from the multi-source heterogeneous data, and construct a multimodal database based on the feature-extracted multi-source heterogeneous data. This embodiment provides a clinical decision support system based on multimodal data. Specifically, the clinical decision support system based on multimodal data can be a clinical decision support system based on a thyroid atypical hyperplasia disease database, including a multimodal database construction module, used to collect multi-source heterogeneous data related to thyroid hyperplasia, extract features from the multi-source heterogeneous data, and construct a multimodal database based on the feature-extracted multi-source heterogeneous data.

[0141] First, multi-source heterogeneous data related to thyroid hyperplasia can be collected through data interfaces. This can include different types of data, such as electronic medical record data collected from hospital information systems, including text data like patient examination results; thyroid ultrasound imaging data collected from imaging archives and communication systems, including nodule grayscale images and color Doppler blood flow images; gene testing report data collected from gene testing platforms; and postoperative pathology report data collected from pathology systems. The collected multi-source heterogeneous data can include complete clinical records, ultrasound images, gene testing results, and pathology reports.

[0142] For text-based data, natural language processing (NLP) techniques can be used to identify medical terms. Standardized mapping using the SNOMED-CT medical terminology system can then be performed to extract clinical entities and generate structured text feature vectors. SNOMED-CT (Systematized Nomenclature of Medicine -- Clinical Terms) is a multilingual clinical medical terminology system and a standard clinical terminology specification in electronic health record systems. In the clinical decision support system based on a thyroid atypical hyperplasia database proposed in this application, SNOMED-CT is used to standardize and map terms to text-based data, converting unstructured clinical text into standardized terminology codes, laying the foundation for subsequent knowledge graph construction and graph neural network analysis. For example, from the pathology report "Fine-point biopsy of a nodule in the left lobe of the thyroid, pathological diagnosis of atypical cells of indeterminate significance," the entities "nodule," "left thyroid lobe," and "atypical cells," as well as the mapping relationships "nodule - located in - left thyroid lobe" and "nodule - diagnosed as - atypical cells," are extracted.

[0143] For imaging data, ultrasound images are preprocessed and segmented to extract texture, morphological, and blood flow features. The extracted image features are then Z-score standardized to eliminate dimensional differences caused by different ultrasound devices. Finally, equal-frequency binning is used to discretize continuous features into multiple categories, generating structured image feature vectors. Z-score standardization (standard deviation standardization) is a data preprocessing method that transforms data of different dimensions and orders of magnitude into a uniform scale. This method standardizes the original data based on the mean and standard deviation, ensuring that the standardized data conforms to a standard normal distribution (mean = 0, standard deviation = 1). In the clinical decision support system for thyroid atypical hyperplasia proposed in this application, Z-score standardization is applied to the feature processing of imaging data to eliminate data differences caused by different ultrasound devices and operators, making the image features acquired from multiple centers comparable.

[0144] The structured text feature vectors and structured image feature vectors are horizontally concatenated to construct a multimodal feature matrix containing patient ID, nodule identifier, feature name, and feature value. Based on this feature matrix, knowledge graph nodes are generated, which may include, for example, patient nodes, nodule nodes, gene nodes, and pathology nodes. Associative edges are constructed between nodes based on entity relationships to form a thyroid hyperplasia-specific knowledge graph, which is stored in a graph database as a multimodal database.

[0145] The dual-engine decision module is used to obtain the corresponding processing strategy based on the data contained in the multimodal database, using the engine decision tree corresponding to the processing rules of thyroid hyperplasia and the data prediction model corresponding to the graph neural network. In this embodiment, the dual-engine decision module is used to obtain corresponding processing strategies based on the data contained in the multimodal database, utilizing the engine decision tree corresponding to the processing rules for thyroid hyperplasia and the data prediction model corresponding to the graph neural network. Patient data can be input into the rule engine of the dual-engine decision module. The rule engine has a built-in decision tree constructed based on the ATA guidelines and thyroid treatment pathways. The rule engine can be a software module that performs reasoning and decision-making based on predefined clinical rules and a logic tree structure. By formalizing the clinical guidelines for thyroid hyperplasia and the localized treatment pathways into an executable set of rules, it performs step-by-step judgments on the input patient data to generate interpretable first decision suggestions. The role of the rule engine is to ensure the interpretability of decisions and compliance with clinical norms, providing doctors with clear decision-making basis. The ATA guidelines refer to the "Guidelines for the Management of Adult Thyroid Nodules and Differentiated Thyroid Cancer". These guidelines provide systematic recommendations based on evidence-based medicine for risk assessment of thyroid nodules, indications for fine-needle aspiration biopsy, selection of surgical methods, postoperative management, and follow-up monitoring. For example, taking a 45-year-old female patient as an example, ultrasound showed a left lobe nodule with a diameter of 1.5cm, microcalcifications and an aspect ratio >1. The rule engine performed the following judgment: nodule size ≥1cm; ultrasound features suspicious (microcalcifications, aspect ratio >1); output the first decision recommendation: surgical intervention.

[0146] Patient data is input into a pre-trained graph neural network (GNN) corresponding to a data prediction model, i.e., a GNN model. GNNs are deep learning models used to process graph-structured data, capable of directly learning from graphs composed of nodes and edges, capturing complex dependencies and global structural information between nodes. In the clinical decision support system for thyroid atypical hyperplasia of this application, the GNN model is used as a data engine to learn deep relational features from a knowledge graph constructed from multimodal data, generating risk prediction results. For example, the GNN model can use patients, nodules, genes, and pathology as nodes, and clinical relationships as edges, with message passing through a multi-layer graph convolutional network. For each layer of the network, each node aggregates information from its neighboring nodes. The aggregation function uses mean pooling, and the update function fuses the node's previous layer representation with the aggregated message to generate the node's current layer representation. After three layers of message passing, the final representations of all nodes are read out, and the readout results are input into a fully connected layer, where a sigmoid activation function outputs the lesion risk prediction value.

[0147] The learning update module is used to update the multimodal database and the processing strategy based on the processing feedback data of thyroid hyperplasia through incremental learning and federated learning mechanisms.

[0148] In this embodiment, the learning update module is used to update the multimodal database and the processing strategy based on the processing feedback data of thyroid hyperplasia through incremental learning and federated learning mechanisms. Updated data and features generated during incremental learning and federated learning are added to the multimodal database. Complete data of newly acquired patients, including clinical records, ultrasound images, and pathology reports, are incrementally added. Simultaneously, new features learned during federated learning are associated and updated in the knowledge graph, achieving dynamic expansion and updating of the multimodal database. Incremental learning is a machine learning paradigm that allows a model to continuously learn new knowledge from newly arriving data while retaining previously learned knowledge, without retraining the entire model from scratch. This application utilizes incremental learning to trigger incremental updates based on clinical feedback data, enabling the system to quickly absorb new case patterns and optimize prediction accuracy. Federated learning is a distributed machine learning paradigm in which multiple participants collaboratively train a shared global model by exchanging intermediate results such as model parameters or gradients without sharing the original data. Through homomorphic encryption and differential privacy technology, it achieves cross-center model co-evolution while protecting the data privacy of all parties, enabling the system to absorb knowledge from multiple sources and improve the diagnostic accuracy of thyroid atypical hyperplasia. In conjunction with incremental learning, it jointly builds a self-evolving clinical decision support system that can quickly respond to new local data and benefit from global knowledge.

[0149] Specifically, during incremental learning and federated learning, all newly generated data and features are incorporated into the multimodal database. This includes: newly added complete patient data being added to the multimodal feature matrix after feature extraction; high-value samples identified in incremental learning having corresponding patient nodes, nodule nodes, and pathology nodes added to the knowledge graph, and establishing relational edges such as "patient-having-nodule" and "nodule-diagnosed-pathology"; and new features learned during federated learning updating the weights of corresponding edges in the knowledge graph. Through a feedback-driven continuous learning mechanism, the system achieves dynamic expansion of the multimodal database and continuous optimization of decision-making strategies, providing more accurate and reliable diagnostic and treatment support for thyroid hyperplasia in clinical practice.

[0150] Based on the aforementioned system of this application, multi-source heterogeneous data related to thyroid atypical hyperplasia, such as electronic medical records, ultrasound images, gene testing, and pathology reports, are deeply integrated and standardized. SNOMED-CT terminology mapping is used to eliminate semantic ambiguity, and Z-score standardization is used to eliminate differences in multi-center imaging equipment. A multimodal knowledge graph containing patients, nodules, genes, pathological nodes, and semantic relationships is constructed, solving the technical problem of data fragmentation in traditional systems. A dual-engine decision architecture of rule engine and graph neural network collaboration is set up, and the outputs of the two are dynamically fused using the NSGA-II multi-objective optimization algorithm to obtain the optimal trade-off processing strategy. A co-evolutionary mechanism of incremental learning and federated learning based on clinical feedback is established, providing accurate and continuously optimized clinical decision support for the diagnosis and treatment of thyroid hyperplasia.

[0151] Based on the same inventive concept, such as Figure 9 As shown, this embodiment also includes an electronic device, comprising: Memory, used to store executable programs; A processor for executing the executable program to implement the above method.

[0152] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A clinical decision-making construction method based on a specialized database of thyroid atypical hyperplasia, characterized in that, The method includes: Collect multi-source heterogeneous data related to thyroid hyperplasia, extract features from the multi-source heterogeneous data, and construct a multimodal database based on the feature-extracted multi-source heterogeneous data; Based on the data contained in the multimodal database, the corresponding processing strategy is obtained by using the engine decision tree corresponding to the processing rules of thyroid hyperplasia and the data prediction model corresponding to the graph neural network. Based on the processing feedback data of thyroid hyperplasia, the multimodal database and the processing strategy are updated through incremental learning and federated learning mechanisms.

2. The method according to claim 1, characterized in that, The step of extracting features from the multi-source heterogeneous data and constructing a multimodal database based on the feature-extracted multi-source heterogeneous data includes: Based on the data format and content characteristics of the multi-source heterogeneous data, the multi-source heterogeneous data is divided into text data and image data; The text data is subjected to terminology standardization mapping, and based on the terminology-standardized text data, the semantic relationships between clinical entities are extracted to obtain structured text feature vectors. Image feature extraction is performed on the image data to obtain texture and morphological features. The extracted image features are then standardized by standard deviation and binned discretized to generate structured image feature vectors. The structured text feature vectors are fused with the structured image feature vectors to obtain multimodal feature matrix data, and a multimodal database is constructed based on the multimodal feature matrix data.

3. The method according to claim 1, characterized in that, Based on the data contained in the multimodal database, the corresponding processing strategy is obtained by using the engine decision tree corresponding to the processing rules for thyroid hyperplasia and the data prediction model corresponding to the graph neural network, including: Patient data from the multimodal database is input into a preset rule engine. The rule engine analyzes the patient data based on the clinical guidelines rule base for thyroid hyperplasia and a localized rule base, and generates a first decision suggestion through the engine decision tree. Patient data from the multimodal database is input into a preset graph neural network model. The graph neural network model uses the features corresponding to the patient's symptoms as nodes and the relationships between clinical entities as edges. Through a multi-layer message passing mechanism, deep correlation features between multimodal data are extracted to generate risk prediction results. Using a multi-objective optimization algorithm, the first decision suggestion and the risk prediction result are dynamically weighted and fused to generate a corresponding processing strategy.

4. The method according to claim 1, characterized in that, The processing feedback data based on thyroid hyperplasia is used to update the multimodal database and the processing strategy through incremental learning and federated learning mechanisms, including: Collect feedback data during the clinical management of thyroid hyperplasia; Based on the feedback data, the pathological results are compared with the prediction results to determine the corresponding prediction deviation. When the prediction deviation exceeds a preset threshold, the incremental learning process is triggered to adjust the parameters of the data prediction model corresponding to the graph neural network. Based on the feedback data and the multimodal data, the graph neural network model is used to calculate the model gradient data, and the corresponding model parameters are updated based on the model gradient data. The updated data and features generated during incremental learning and federated learning are added to the multimodal database, and the multimodal database is dynamically updated.

5. The method according to claim 4, characterized in that, The step of calculating model gradient data using the graph neural network model based on the feedback data and the multimodal data, and updating the corresponding model parameters based on the model gradient data, includes: Based on locally collected feedback data and locally stored multimodal data, the model gradient data is calculated using the trained graph neural network model. The gradient data of the computational model is encrypted using a homomorphic encryption algorithm, and the encrypted gradient data of the computational model is sent to the federated learning aggregation server. The encrypted computational model gradient data is homomorphically encrypted and aggregated by the federated learning aggregation server, and differential privacy noise is added to obtain aggregated noisy gradient data. Decrypt the aggregated noise gradient data to obtain the updated model parameters corresponding to the model gradient data; The graph neural network model is updated based on the updated model parameters.

6. The method according to claim 4, characterized in that, in, The feedback data includes at least one of the following: the doctor's adoption and adjustment records of the treatment strategy, the patient's postoperative recovery data and the occurrence of complications, and the comparison data between postoperative pathology results and preoperative prediction results.

7. The method according to claim 4, characterized in that, The method further includes: The decision results output by the clinical system are collected and compared with the actual clinical results to obtain the corresponding prediction bias. Determine whether the prediction deviation exceeds the preset threshold; If the prediction deviation exceeds the preset threshold, an incremental learning process is triggered to update the parameters of the data prediction model corresponding to the graph neural network. If the prediction deviation does not exceed the preset threshold, only feedback data is recorded, and the incremental learning process is not triggered.

8. The method according to claim 3, characterized in that, The step of using a multi-objective optimization algorithm to dynamically weight and fuse the first decision suggestion and the risk prediction result to generate a corresponding processing strategy includes: A multi-objective optimization algorithm is used to dynamically fuse the first decision suggestion and the risk prediction result to obtain multiple candidate fusion weight combinations. A target fusion weight combination is determined from multiple candidate fusion weight combinations, wherein the target fusion weight combination represents the optimal decision-making method among the multiple candidate fusion weight combinations; Based on the target fusion weight combination, a corresponding clinical treatment strategy is generated.

9. A clinical decision support system based on a specialized database of thyroid atypical hyperplasia, characterized in that, include: A multimodal database construction module is used to collect multi-source heterogeneous data related to thyroid hyperplasia, extract features from the multi-source heterogeneous data, and construct a multimodal database based on the feature-extracted multi-source heterogeneous data. The dual-engine decision module is used to obtain the corresponding processing strategy based on the data contained in the multimodal database, using the engine decision tree corresponding to the processing rules of thyroid hyperplasia and the data prediction model corresponding to the graph neural network. The learning update module is used to update the multimodal database and the processing strategy based on the processing feedback data of thyroid hyperplasia through incremental learning and federated learning mechanisms.

Images