Diabetes clinical test data intelligent clustering analysis system and method based on federal learning

Abstract

The invention relates to the technical field of data analysis, in particular to a diabetes clinical test data intelligent clustering analysis system and method based on federal learning. Comprising a data acquisition and preprocessing unit; a federal privacy protection unit; the dynamic clustering modeling unit is used for constructing a clustering model of self-adaptive diabetes data features, and realizing joint clustering analysis of multi-source heterogeneous data by adopting a hierarchical federal architecture and a dynamic parameter aggregation algorithm and combining a diabetes course time decay factor and a clinical feature weight adjustment strategy; a double-track verification optimization unit; and an intelligent decision support unit. According to the method, the incidence matrix of the diabetes disease course time decay factor and the clinical characteristics is introduced, the dynamic weight vector is constructed and applied to clustering distance calculation, so that the model can adapt to dynamic changes of the clinical characteristics in different disease course stages, the adaptability to multi-center heterogeneous data is improved, and the stability of a clustering result is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a federated learning-based intelligent clustering analysis system and method for diabetes clinical trial data. Background Art

[0002] Deep mining of diabetes clinical trial data has significant clinical value for disease classification, treatment optimization, and complication risk prediction. Collaborative analysis of multicenter data, in particular, can significantly improve the representativeness and reliability of results. However, cross-institutional data sharing is subject to strict privacy protection regulations, and traditional centralized data processing models are prone to information leakage risks. Furthermore, data from different institutions exhibit heterogeneous formats (e.g., differences in physiological indicator recording standards and electronic medical record structures) and dynamic clinical characteristics (e.g., the impact of disease progression on indicator correlations), making it difficult for conventional clustering methods to generate stable and clinically relevant analytical results.

[0003] In the existing technology, for example, Chinese patent CN202411073211.8 discloses a diabetes complication prediction method based on k-means clustering analysis. It only processes data from a single institution, does not build a privacy protection framework for cross-institutional data, and does not consider the dynamic adjustment of clustering weights due to disease course factors, making it difficult to adapt to multi-center heterogeneous data scenarios; its feature selection and clustering process rely on centralized data storage, which poses privacy and security risks, and the clustering results have not been systematically verified with clinical diagnosis and treatment rules, and its clinical applicability is limited. For example, Chinese patent CN202411009458.3 discloses a method for constructing a type 2 diabetes classification model and a classification system. Although the classification stability of single-institution data is improved by optimizing clustering indicators, it lacks a safe aggregation mechanism for multi-center data, and the model's generalization ability is limited by the single data source; it uses a fixed Euclidean distance to measure cluster similarity, and does not dynamically adjust weights based on the correlation of clinical characteristics. It cannot reflect the characteristic differences of diabetic patients at different stages of the disease course, and does not set up a dual-track verification mechanism to ensure the consistency of classification results with clinical guidelines, making it difficult to meet the analysis needs of multi-center clinical trials.

[0004] While the above-mentioned technical solution has design advantages, it also suffers from the following technical drawbacks: First, it fails to utilize a federated learning architecture and privacy-preserving computing technology, relying on single-institution data or centralized storage. This makes it impossible to securely aggregate multi-center data while retaining the original data locally, posing privacy risks and making it difficult to meet the compliance requirements for cross-institutional data collaborative analysis. Second, it lacks a dynamic feature weight adjustment mechanism, fails to integrate the disease course time decay factor with the adaptive weighting of clinical feature correlations, and uses a fixed distance metric to measure cluster similarity. This fails to reflect the dynamic changes in clinical features at different disease stages, lacks adaptability to multi-center heterogeneous data, and results are unstable. Third, it lacks a dual-track verification and iterative optimization mechanism for the clinical validity of clustering results. It fails to integrate statistical indicators and diagnosis and treatment rule verification, making it impossible to comprehensively assess clinical fit. Furthermore, it lacks a parameter adjustment strategy based on assessment bias, making it difficult to enhance the clinical guidance value of the results. Therefore, we propose a federated learning-based intelligent clustering analysis system and method for diabetes clinical trial data. Summary of the Invention

[0005] The purpose of the present invention is to provide a federated learning-based intelligent clustering analysis system and method for diabetes clinical trial data to solve the problems raised in the above background technology.

[0006] To solve the above technical problems, one of the objectives of the present invention is to provide an intelligent clustering analysis system for diabetes clinical trial data based on federated learning, comprising: The data collection and preprocessing unit is used to implement privacy-preserving collection and standardized preprocessing of multi-center diabetes data. Based on the edge computing architecture and federated learning communication protocol, it builds a cross-institutional data security aggregation channel and combines physiological indicator normalization rules with clinical feature encoding strategies to generate a homogeneous clustering input data set. Specifically, in the data acquisition and preprocessing unit, the medical device interface submodule and the electronic medical record docking submodule of the edge acquisition module operate in parallel: the former accesses the real-time monitoring data of blood glucose meters, insulin pumps and other devices through USB-C, Bluetooth and other interfaces, and the latter extracts patient medical records, test indicators and other data from the hospital information system based on the HL7FHIR protocol; the collected original data is encapsulated into encrypted fragments by the secure transmission module, and the breakpoints are resumed through the federal communication protocol, and the data is reassembled after arriving at the target node; the standardization processing module calls the clinical data conversion engine, performs Z-score standardization on continuous physiological indicators, performs one-hot encoding on classification features such as diabetes classification and complications, and fills missing values ​​with clinical correlation weighted interpolation method, and finally outputs the homogeneous clustering input data set for subsequent unit calls.

[0007] The federated privacy protection unit is used to ensure the privacy security of cross-institutional data interaction. It encrypts and aggregates clustering intermediate parameters through a secure multi-party computing protocol to achieve privacy protection for the local storage of original data. Specifically, in the federal privacy protection unit, the encryption calculation module first uses lightweight homomorphic encryption to generate ciphertext for the clustering intermediate parameters generated by the dynamic clustering modeling unit; the security aggregation module splits the ciphertext into sub-shares equal to the number of participating institutions based on the Shamir threshold secret sharing protocol. Each sub-share is accompanied by an HMAC-SHA256 authentication code and distributed to the corresponding institution through the federal communication protocol. After collecting at least a threshold number of valid sub-shares, the ciphertext is reconstructed; the verification and audit module is deployed on a trusted third-party node to verify the hash consistency of the sub-shares of each institution to ensure that the original data is not leaked during the aggregation process. The three form a privacy protection closed loop of "encryption → split distribution → aggregation → audit".

[0008] Dynamic clustering modeling unit, used to build a clustering model that adapts to diabetes data characteristics. It uses a hierarchical federated architecture and a dynamic parameter aggregation algorithm, combined with a diabetes course time decay factor and a clinical feature weight adjustment strategy to achieve joint clustering analysis of multi-source heterogeneous data. Specifically, the dynamic clustering modeling unit receives the homogeneous data set output by the data collection and preprocessing unit, and combines it with the aggregation parameters encrypted and transmitted by the federal privacy protection unit to build a hierarchical federal architecture: the edge layer nodes train the model based on local data and calculate the gradient, the regional layer dynamically adjusts the aggregation weight according to the sample size ratio and feature correlation, and the central layer summarizes the global gradient to update the model; at the same time, it incorporates the disease course time attenuation factor and the clinical feature correlation matrix, dynamically adjusts the feature weight and clustering distance metric, and outputs the clustering results to the dual-track verification optimization unit and the intelligent decision support unit.

[0009] The dual-track verification and optimization unit is used to ensure the clinical validity of clustering results. It generates a clustering quality assessment report through the linkage between the statistical indicator verification module and the diagnosis and treatment rule verification module, and the evaluation report generation module. Based on the indicator deviation in the clustering quality assessment report, it drives the iterative optimization of the clustering model in the dynamic clustering modeling unit. Specifically, in the dual-track verification and optimization unit, the data distribution verification submodule of the statistical indicator verification module verifies the rationality of the distribution of each cluster feature, the stability verification submodule evaluates cluster consistency through Bootstrap resampling, and the separation verification submodule calculates the silhouette coefficient and Calinski-Harabasz index to quantify the separation between clusters; the rule extraction submodule of the diagnosis and treatment rule verification module parses the clinical guidelines to generate executable rules, the guideline matching submodule compares the fit between clustering results and rules, and the anomaly identification submodule marks sample patterns that do not conform to the diagnosis and treatment logic; the evaluation report generation module integrates the two types of verification results, outputs a report containing statistical deviations and rule conflicts, and drives the dynamic clustering modeling unit to iteratively optimize parameters.

[0010] The intelligent decision support unit is used to convert clustering results into clinical decision recommendations, and generate personalized treatment plan recommendations and complication risk warnings based on the diabetes clinical guideline knowledge base and machine learning prediction model.

[0011] Specifically, in the intelligent decision support unit, the clustering result processing module extracts the key clinical features from the output results of the dynamic clustering modeling unit; the decision generation module calls the diabetes clinical guideline knowledge base to match diagnosis and treatment rules, combines the machine learning prediction model to evaluate the risk of complications, and generates personalized treatment plans and early warning recommendations; the result output module integrates plans and warnings, and outputs clinical decision recommendations in the form of visual reports, archivable PDFs, etc., to realize the transformation of clustering results into diagnosis and treatment actions.

[0012] As a further improvement of this technical solution, the data acquisition and preprocessing unit includes an edge acquisition module, a secure transmission module and a standardization processing module, wherein: The edge acquisition module is used to collect multi-source clinical data from diabetes clinical monitoring and treatment equipment and the hospital's electronic medical record system. It includes a medical device interface submodule and an electronic medical record docking submodule. The medical device interface submodule supports a USB-C wired interface and a Bluetooth 5.0 wireless interface for accessing real-time data from diabetes clinical monitoring and treatment equipment. The electronic medical record docking submodule uses the HL7FHIR standard protocol to extract patient basic information, medical records, and treatment history data from the hospital's electronic medical record system. The secure transmission module uses the TLS1.3 encryption protocol to build a transmission channel, performs shard encryption on the collected data, and each data shard is accompanied by a check code generated based on the SHA-256 algorithm. The federated learning communication protocol is used to achieve breakpoint-resume transmission of sharded data. The standardization processing module has a built-in clinical data conversion engine and performs hierarchical preprocessing on the collected data: Z-score standardization is used for continuous physiological indicators, one-hot encoding is used for complication types, medication regimens, diabetes classification, and treatment methods, and missing data is filled with weighted interpolation based on the correlation of clinical characteristics.

[0013] Furthermore, the diabetes clinical monitoring and treatment equipment includes a blood glucose monitor, a glycosylated hemoglobin detector, a continuous glucose monitoring system (CGM) and an insulin pump.

[0014] As a further improvement of the present technical solution, in the standardization processing module, filling missing data using a weighted interpolation method based on the correlation of clinical characteristics includes the following steps: S130.1. For target clinical features with missing values, analyze their association with clinical features other than the target clinical feature, and screen for feature subsets with clinical diagnosis and treatment relevance (the association is based on the correlation of indicators clearly defined in diabetes clinical guidelines, such as the association between blood glucose level and glycosylated hemoglobin, and the association between disease duration and complication type); S130.2. Assign a corresponding weight to each associated feature based on the clinical importance of the associated feature subset in the diagnosis and treatment of diabetes; S130.3. Based on the measured data and corresponding weights of the associated features, calculate the filling value of the missing value of the target feature, and verify whether the filled data conforms to the clinical distribution law of the feature (for example, the blood glucose value filling result must fall within the normal or pathological reference range).

[0015] As a further improvement of this technical solution, the federal privacy protection unit includes an encryption calculation module, a security aggregation module, and a verification and audit module, wherein: The encryption calculation module uses a lightweight homomorphic encryption algorithm to encrypt the clustering intermediate parameters to generate a ciphertext parameter vector; The security aggregation module splits the ciphertext parameter vector into N sub-shares based on a threshold secret sharing protocol, where N is the number of participating institutions; and distributes each sub-share to the corresponding participating institutions through a federated learning communication protocol, where each sub-share is accompanied by a message authentication code generated based on HMAC-SHA256; The verification and audit module is deployed on a trusted third-party node to ensure that the aggregation process does not leak original data information by verifying the consistency of the sub-share hash digests submitted by each participating institution.

[0016] As a further improvement of the technical solution, in the dynamic clustering modeling unit, a hierarchical federation architecture and a dynamic parameter aggregation algorithm are adopted, including the following steps: S310.1. Build a three-tiered federation structure, including edge, regional, and central tiers; Furthermore, the edge layer is deployed in various medical institutions, responsible for processing and modeling local diabetes datasets; the regional layer is deployed in provincial medical data centers, responsible for parameter aggregation and management of the edge layer nodes under its jurisdiction; and the central layer is deployed in national research institutions, responsible for the optimization and distribution of global models.

[0017] S310.2. Partition the homogeneous clustering input dataset output by the data acquisition and preprocessing unit by medical institution to form a local diabetes dataset for each edge layer node; S310.3. The central layer node pre-trains the initial clustering model based on the historical public diabetes dataset to generate the initial gradient parameters of the regional layer. , and through the federated learning communication protocol Distribute to nodes at each regional layer; S310.4. The edge layer nodes train the initial clustering model based on the local diabetes dataset and calculate the gradient parameters. ,in Indicates the edge layer nodes; S310.5. The regional layer node receives the gradient parameters of the edge layer nodes under its jurisdiction and dynamically adjusts the aggregation weight based on the proportion of data volume of each edge layer. The calculation formula is: ; in, Edge layer sample size; is the sum of all edge layer sample sizes in the region; Edge layer Model gradient Current gradient with regional layer The cosine similarity of Indicates the current iteration round; is the adjustment coefficient, and ; S310.6. Regional layer nodes calculate the aggregate gradient based on the adjusted weights : ; And send the aggregated gradient to the central layer node; S310.7. The central layer node receives the aggregated gradients from the regional layer nodes and updates the global model: ; in, For regional layer The sample size percentage, is the global gradient; The central layer node will update the global model parameters Feedback is sent to each regional layer and edge layer, and the edge layer nodes are based on Update local model; S310.8, when entering the next iteration, each regional layer node will Update to the current gradient , repeat S310.4-S310.7 until the convergence conditions are met.

[0018] As a further improvement of the present technical solution, in the dynamic clustering modeling unit, the diabetes course time attenuation factor and the clinical feature weight adjustment strategy are combined, including the following steps: S320.1. Calculate the time decay factor of diabetes course based on the homogeneous clustered input data set output by the data acquisition and preprocessing unit ; ; in, The duration of the disease, is the time decay coefficient, and ;Control the decay rate of historical data; is the complication influence coefficient, which amplifies the effect of complications on the data weight, and ; A scoring function for complication severity was developed, and complication types were quantitatively scored based on the ICD-10 coding system generated by the standardized processing module; S320.2. Constructing a clinical feature correlation matrix ,in The total number of features of the homogeneous clustering input data set output by the data acquisition and preprocessing unit, Representation matrix yes OK A real matrix of columns; elements Representation characteristics and features The conditional probability of is calculated using the historical data set: ; in, Indicates simultaneous occurrence of features and The number of samples, Indicates the presence of features The historical data set is the homogeneous clustering input data set preprocessed by the standardization processing module; Indicates Under the conditions that occur, Probability of occurrence; S320.3. For target features , calculate its dynamic weight vector ; ; in, is the basic weight vector (preset by the standardization processing module based on diabetes clinical guidelines); Associative features The weight of Representation and characteristics Features with a correlation greater than the threshold , through the incidence matrix elements Weighted sum; S320.4. Dynamic weight vector Applied to cluster distance calculation, the improved weighted Euclidean distance formula is used: ; in, and are two sample points, and For the corresponding features The value of For samples and The weighted Euclidean distance between them; S320.5. Perform clustering iteration based on weighted Euclidean distance and update cluster centers ; ; in, For samples The weight of the feature is determined by the dynamic weight vector of the feature to which it belongs Decide; Indicates that all samples in the cluster are weighted Weighted sum; Represents the weighted sum of all samples in the cluster; S320.6, update the cluster center Feedback to the statistical indicator verification module to evaluate the clustering quality. If the evaluation indicator does not meet the preset threshold, the 、 parameters and re-execute S320.1-S320.5.

[0019] As a further improvement of this technical solution, the statistical indicator verification module includes a data distribution verification submodule, a stability verification submodule and a separation degree verification submodule, wherein: The data distribution verification submodule is used to calculate the statistical distribution indicators of the clustering results (such as Kolmogorov-Smirnov test value and chi-square test statistic); The stability verification submodule calculates the adjusted Rand index (ARI) and normalized mutual information (NMI) through the Bootstrap resampling technique to evaluate the consistency of clustering results under different sampling data sets; The separation verification submodule is used to calculate the silhouette coefficient and the Calinski-Harabasz index to quantify the compactness and separation of the clustering results; The diagnosis and treatment rule verification module includes a rule extraction submodule, a guideline matching submodule, and an abnormality identification submodule, wherein: The rule extraction submodule is used to parse the complication classification and diagnosis and treatment pathway clauses in the "Chinese Diabetes Prevention and Treatment Guidelines" and convert them into executable logical rules based on ICD-10 coding and clinical data characteristics; The guideline matching submodule is used to match the clustering results with the executable logic rules generated by the rule extraction submodule to generate a rule compliance score; The anomaly identification submodule identifies abnormal data patterns in the clustering results based on the executable logical rules generated by the rule extraction submodule; The evaluation report generation module is used to fuse the statistical indicators output by the statistical indicator verification module with the rule compliance score output by the guideline matching submodule, and generate a clustering quality evaluation report.

[0020] As a further improvement of this technical solution, in the dual-track verification optimization unit, based on the indicator deviation in the clustering quality assessment report, the clustering model of the dynamic clustering modeling unit is driven to iteratively optimize, including the following steps: S440.1. Extraction index deviation: Based on the clustering quality assessment report output by the assessment report generation module, extract indicator deviation items, including statistical indicator deviation and diagnosis and treatment rule deviation; S440.2. Attribution of deviation items: Statistical indicator deviation: Parameter problems in locating dynamic cluster modeling units (such as time decay coefficient (deviating from the range of 0.05-0.1), clinical characteristics correlation matrix The setting of the correlation threshold δ is unreasonable); Treatment rule deviation: Feature weight strategy for positioning dynamic clustering modeling units (such as basic weight vector Not in line with the guide presets, dynamic weight vector inadequate integration of clinically relevant features); S440.3. Generate parameter adjustment instructions: Adjusting for statistical bias: Value, modified feature correlation threshold (Based on clinical characteristics association matrix actual strength of association adjustment); For regular deviations: recalibrate the basis weight vector (According to the "Chinese Diabetes Prevention and Treatment Guidelines (2024 Edition)", the core feature weights were updated and the influence coefficient of complications was adjusted. ; S440.4. Send the adjustment instruction to the dynamic cluster modeling unit, driving the dynamic cluster modeling unit to re-execute S320.1-S320.6 and S310.4-S310.8; S440.5. After the dynamic clustering modeling unit outputs the updated clustering results, the dual-track verification optimization unit re-verifies through the statistical indicator verification module and the diagnosis and treatment rule verification module until the indicator deviation of the clustering quality assessment report is lower than the preset threshold or the maximum number of iterations is reached.

[0021] As a further improvement of this technical solution, the intelligent decision support unit includes a clustering result processing module, a decision generation module and a result output module, wherein: The clustering result processing module receives the clustering results output by the dynamic clustering modeling unit and extracts the key clinical features of the patient; The decision generation module generates personalized treatment plan recommendations and complication risk warnings based on the diabetes clinical guideline knowledge base and machine learning prediction model, combined with the extracted cluster features; The result output module integrates treatment plans and risk warnings to output clinical decision-making recommendations.

[0022] A second object of the present invention is to provide a federated learning-based intelligent clustering analysis method for diabetes clinical trial data. The method comprises the following steps: S100, Data Collection and Preprocessing: Based on the edge computing architecture and federated learning communication protocol, multi-center diabetes clinical data are collected, standardized after secure transmission, and a homogeneous clustered input data set is generated; S200, federated privacy protection: A lightweight homomorphic encryption algorithm is used to encrypt the clustering intermediate parameters. The ciphertext parameters are split and distributed to each participating institution through a threshold secret sharing protocol. The privacy security of the aggregation process is verified by a trusted third party. S300, Dynamic Clustering Modeling: Constructing a three-tier federated architecture consisting of edge, regional, and central layers, and performing joint clustering of multi-source heterogeneous data based on a dynamic parameter aggregation algorithm. Combining the diabetes course time decay factor with a clinical feature weight adjustment strategy, the dynamic weight vector is calculated and applied to cluster distance calculation to generate initial clustering results. S400, Dual-Track Verification and Optimization: Generate a clustering quality assessment report by integrating statistical indicator verification and diagnosis and treatment rule verification. Based on the indicator deviations in the report, locate parameter problems and feature weight strategy deviations, and generate adjustment instructions to drive iterative optimization of the clustering model until the indicator deviation in the clustering quality assessment report falls below the preset threshold or the maximum number of iterations is reached. S500, Intelligent Decision Support: Extract key clinical features from the optimized clustering results, combine the diabetes clinical guideline knowledge base with the machine learning prediction model to generate personalized treatment plan recommendations and complication risk warnings, and output structured clinical decision recommendations.

[0023] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention builds a cross-institutional data security architecture based on federated learning, combining edge computing and privacy-preserving computing technologies. It achieves secure aggregation and joint analysis of multi-center diabetes data while retaining the original data locally. This effectively avoids the privacy leakage risks brought by centralized storage and meets the compliance requirements of cross-institutional data collaboration. 2. This invention introduces the diabetes course time attenuation factor and the clinical feature association matrix, constructs a dynamic weight vector, and applies it to cluster distance calculation. This enables the model to adapt to the dynamic changes in clinical features at different stages of the course of the disease, improves its adaptability to multi-center heterogeneous data, and enhances the stability of clustering results. 3. This invention uses a dual-track mechanism of statistical indicator verification (such as data distribution, stability, and separation) and diagnosis and treatment rule verification (such as associated clinical guideline clauses), combined with an iterative optimization strategy of model parameters based on evaluation bias, to comprehensively improve the clinical fit of clustering results and enhance their guiding value for the clinical diagnosis and treatment of diabetes. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the system framework of the present invention; Figure 2 Schematic diagram of the method steps of the present invention; The meaning of each number in the figure is: 100, data acquisition and preprocessing unit; 110, edge acquisition module; 111, medical equipment interface submodule; 112, electronic medical record docking submodule; 120, secure transmission module; 130, standardization processing module; 200, Federal Privacy Protection Unit; 210, Encryption Computing Module; 220, Security Aggregation Module; 230, Verification and Audit Module; 300, dynamic clustering modeling unit; 400, dual-track verification and optimization unit; 410, statistical indicator verification module; 411, data distribution verification submodule; 412, stability verification submodule; 413, separation verification submodule; 420, diagnosis and treatment rule verification module; 421, rule extraction submodule; 422, guideline matching submodule; 423, anomaly identification submodule; 430, evaluation report generation module; 500, intelligent decision support unit; 510, clustering result processing module; 520, decision generation module; 530, result output module. DETAILED DESCRIPTION

[0025] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0026] Example 1 like Figure 1 As shown, this embodiment provides an intelligent clustering analysis system for diabetes clinical trial data based on federated learning, including: The data collection and preprocessing unit 100 is used to implement privacy-preserving collection and standardized preprocessing of multi-center diabetes data. Based on the edge computing architecture and federated learning communication protocol, it builds a cross-institutional data security aggregation channel and combines physiological indicator normalization rules with clinical feature encoding strategies to generate a homogeneous clustering input data set. In this embodiment, the data acquisition and preprocessing unit 100 includes an edge acquisition module 110, a secure transmission module 120, and a standardization processing module 130, wherein: The edge acquisition module 110 is used to collect multi-source clinical data from diabetes clinical monitoring and treatment equipment and the hospital's electronic medical record system. It includes a medical device interface submodule 111 and an electronic medical record docking submodule 112. The medical device interface submodule 111 supports a USB-C wired interface and a Bluetooth 5.0 wireless interface for accessing real-time data from diabetes clinical monitoring and treatment equipment. The electronic medical record docking submodule 112 uses the HL7FHIR standard protocol to extract basic patient information, medical records, and treatment history data from the hospital's electronic medical record system. As a further explanation of this embodiment, the medical device interface submodule 111 in this embodiment specifically includes: For Bluetooth 5.0 wireless devices (such as blood glucose meters and continuous blood glucose monitoring systems), subscribe to the data channel broadcast by the device, parse the time series data frames, and convert them into structured data; For USB-C wired devices (such as insulin pumps), send query commands through the serial port, parse the reply frame, and extract the valid field encapsulation; A driver expansion interface is reserved to support new devices by loading device adapters, clarifying compatible device models (such as brand A blood glucose meter, brand B dynamic blood glucose meter), and achieving flexible device expansion.

[0027] As a further explanation of this embodiment, the electronic medical record docking submodule 112 of this embodiment extracts the patient's basic information, medical records, and treatment history data from the hospital's electronic medical record system, including the following steps: First, identity verification is completed through the open interface of the hospital information system (HIS) using token authentication (applying for an access token with a validity period of 1 hour); Then, the hierarchical interface is called to obtain data: the name, date of birth, and gender are extracted through the patient information interface; the diagnosis date and complications are extracted through the disease diagnosis interface (screening diabetes diagnosis codes, such as ICD-10:E11); and the measurement time, value, and unit are extracted through the test index interface (screening blood glucose and glycated hemoglobin codes, such as LOINC:41950-7). Then, the interface data is mapped to local models (such as patient model {patient ID, name, age, diabetes diagnosis date, complication list}, inspection model {indicator type, measurement timestamp, value, unit}), completing the structured conversion of medical record data.

[0028] The secure transmission module 120 uses the TLS1.3 encryption protocol to build a transmission channel, performs shard encryption on the collected data, and each data shard is accompanied by a checksum generated based on the SHA-256 algorithm. The federated learning communication protocol is used to achieve breakpoint-resume transmission of sharded data. As a further explanation of this embodiment, the secure transmission module 120 of this embodiment uses the TLS 1.3 encryption protocol to build a transmission channel, which includes the following process: First, configure the sharding parameters: the sharding size can be set to 1024KB to 8192KB (default 2048KB), and the formula , calculate the number of shards ; Then, TLS1.3 encryption (select TLS_AES_256_GCM_SHA384 suite) is used to generate the data packet; Then, resumable transmission is implemented: the ID of the transmitted fragment is recorded locally, and only the missing fragment is requested during retransmission; the server verifies the integrity through the checksum, and if it fails, it is retransmitted (up to 3 times), and the failed fragment is marked as "invalid" and logged.

[0029] The standardization processing module 130 has a built-in clinical data conversion engine, which performs hierarchical preprocessing on the collected data: Z-score standardization is used for continuous physiological indicators, one-hot encoding is used for complication types, medication regimens, diabetes classification, and treatment methods, and missing data is filled in with weighted interpolation based on the correlation of clinical characteristics.

[0030] As a further explanation of this embodiment, the standardization of the continuous indicator Z-score in the standardization processing module 130 of this embodiment includes the following steps: First, calculate the local center Internal continuous indicators (e.g. blood sugar) mean and standard deviation: ; ; in, No. The number of samples in each data center (or group); For the In the data center, Target physiological index values ​​of each sample (such as blood glucose, glycosylated hemoglobin); For the The mean value of the target physiological index in each data center; For the The standard deviation of the target physiological index in each data center; Then, the normalized value is calculated: ;in, For the In the data center, The standardized index value of each sample; Then, judge : If true, mark the "outlier", retain the original value and add a mark (it is optional to participate in the calculation during clustering).

[0031] As a further explanation of this embodiment, the standardization processing module 130 of this embodiment adopts one-hot encoding processing, including the following steps: First, enumerate the values ​​of the classification indicators (e.g., diabetes classification, complications / medication enumerated by clinical coding); Subsequently, the coding dimension is determined based on the number of enumeration values ​​(e.g., the diabetes classification dimension is 3); Then, map the values ​​into one-hot vectors (e.g., "type 2" - [0, 1, 0]); Finally, if a new classification appears, the dimension is dynamically expanded (a new "unknown" category is added, corresponding to all 0s or a new dimension).

[0032] Furthermore, clinical monitoring and treatment equipment for diabetes include blood glucose monitors, glycosylated hemoglobin detectors, continuous glucose monitoring systems (CGM) and insulin pumps.

[0033] In this embodiment, in the standardization processing module 130, filling missing data using a weighted interpolation method based on the correlation of clinical characteristics includes the following steps: S130.1. For target clinical features with missing values, analyze their association with clinical features other than the target clinical feature, and screen for feature subsets with clinical diagnosis and treatment relevance (the association is based on the correlation of indicators clearly defined in diabetes clinical guidelines, such as the association between blood glucose level and glycosylated hemoglobin, and the association between disease duration and complication type); S130.2. Assign a corresponding weight to each associated feature based on the clinical importance of the associated feature subset in the diagnosis and treatment of diabetes; S130.3. Based on the measured data and corresponding weights of the associated features, calculate the filling value of the missing value of the target feature, and verify whether the filled data conforms to the clinical distribution law of the feature (for example, the blood glucose value filling result must fall within the normal or pathological reference range).

[0034] As a further explanation of this embodiment, the standardization processing module 130 of this embodiment uses the weighted interpolation method based on the correlation of clinical characteristics to fill in the missing data, including the following steps: First, for missing values ​​of target features, we screened the subset of related features (e.g., blood sugar related to glycosylated hemoglobin, disease course related to complications) by referring to the Chinese Diabetes Prevention and Treatment Guidelines (2024 Edition); Subsequently, weights were assigned according to the level of clinical evidence; Then, calculate the padding value: ;in, Fill the missing values ​​of the target feature with the result; is the number of features that are clinically associated with the target feature; For the The weight of the associated features; For the The measured value of the associated feature; Next, verify that the fill value is within the clinical range (e.g., fasting blood glucose needs to be between 3.9-6.1mmol / L); Finally, if the bounds are exceeded, halve the weakly associated features ( ) weights and recalculate. If it fails after 3 iterations, it is marked as "missing" and filled with the mean of the non-missing values ​​of the feature in the center.

[0035] It should be added that the interaction and fault tolerance of multiple modules of the data acquisition and pre-processing unit 100 in this embodiment includes the following process: First, the edge acquisition module 110 outputs the raw data (device + medical record JSON) and passes it to the secure transmission module 120; Subsequently, the secure transmission module 120 completes encryption fragmentation and breakpoint resuming, and reorganizes the data into the standardization processing module 130; Then, the standardization processing module 130 processes according to the process of "continuous Z-score-categorical coding-missing value interpolation" and outputs a homogeneous feature vector; Next, if the module execution fails (e.g., device timeout, shard verification failure), the degradation strategy is triggered: if device data is missing, historical data is reused; if transmission fails, it is marked as "to be completed"; if standardization anomalies occur, the original value is retained and marked; Finally, exceptions are recorded in logs (including timestamp, module, and error code) to support retrospective repair later.

[0036] The federal privacy protection unit 200 is used to ensure the privacy security of cross-institutional data interaction. It encrypts and aggregates clustering intermediate parameters through a secure multi-party computing protocol to achieve privacy protection for the local storage of original data. In this embodiment, the federal privacy protection unit 200 includes an encryption calculation module 210, a security aggregation module 220, and a verification and audit module 230, wherein: The encryption calculation module 210 uses a lightweight homomorphic encryption algorithm to encrypt the clustering intermediate parameters and generate a ciphertext parameter vector; As a further explanation of this embodiment, this embodiment uses a lightweight homomorphic encryption algorithm to encrypt the clustering intermediate parameters to generate a ciphertext parameter vector, including the following steps: First, the algorithm and key are initialized: a lightweight homomorphic encryption scheme (such as simplified Paillier encryption) is selected. A 2048-bit public key (PK) and private key (SK) are generated during system initialization. The private key is retained locally, and the public key is distributed to all participating institutions. Subsequently, the encryption object definition is performed: the intermediate clustering parameters are clearly defined as the cluster center feature vectors (e.g., blood glucose and disease course standardized values ​​of each center), and the sample-center distance matrix elements (standardized distance, dimensionless); Then, ciphertext generation: for each intermediate parameter , call the encryption function to generate ciphertext , assemble all ciphertexts into ciphertext parameter vectors (format: , is the number of parameters); Finally, exception handling: If the encryption times out (not completed within 10 seconds) or an operation error is reported, a retry is triggered (up to 3 times), an error log is recorded (including parameter values ​​and error codes), and the upper-level module is notified to terminate the encryption.

[0037] Furthermore, in the encryption calculation module 210 of the federated privacy protection unit 200, the Paillier lightweight homomorphic encryption algorithm is preferably used to protect the privacy of medical data and support the parameter aggregation calculation of the dynamic clustering modeling unit 300. The algorithm's additive homomorphic properties allow clustering feature operations (such as blood glucose mean calculation and disease course summation) to be performed directly on ciphertext, preventing the leakage of original data and adapting to the low-latency and high-privacy requirements of diabetes clinical trial data. When implementing the Paillier lightweight homomorphic encryption algorithm, during the edge node initialization phase, the encryption calculation module 210 generates a 2048-bit Paillier key pair. The public key is synchronized to the edge acquisition module 110 of the data acquisition and preprocessing unit 100, used to encrypt raw data such as real-time blood glucose levels collected by the medical device interface submodule 111. The private key is stored locally for subsequent decryption and audit scenarios. After acquiring the data, the edge acquisition module 110 performs encryption using the public key. The encrypted ciphertext is then transmitted to the encryption calculation module 210 via the TLS channel of the secure transmission module 120. The encryption calculation module 210 supports ciphertext addition operations. The sum of multi-center ciphertexts is directly fed to the secure aggregation module 220, providing the basis for parameter aggregation in dynamic clustering modeling. When the verification and audit module 230 triggers a data consistency check, the encryption calculation module 210 uses the private key to decrypt the data and verify that the decrypted result matches the hash value of the original data, thus ensuring the authenticity of the data input to the dynamic clustering modeling unit 300. At the same time, by pre-calculating cryptographically secure random numbers and using the Montgomery algorithm to accelerate modular exponentiation operations, it adapts to the latency requirements of real-time data processing on edge devices.

[0038] The security aggregation module 220 splits the ciphertext parameter vector into N sub-shares based on a threshold secret sharing protocol, where N is the number of participating institutions; and distributes each sub-share to the corresponding participating institutions through a federated learning communication protocol, where each sub-share is accompanied by a message authentication code generated based on HMAC-SHA256; As a further illustration of this embodiment, in this embodiment, the security aggregation module 220 splits the ciphertext parameter vector into N sub-shares based on the threshold secret sharing protocol, and distributes each sub-share to the corresponding participating institution through the federated learning communication protocol, including the following steps: First, threshold and algorithm configuration: Set the secret sharing threshold according to the number of participating institutions N. , select Shamir secret sharing algorithm; Then, the ciphertext is split: each ciphertext in the output ciphertext parameter vector , split into N sub-shares ( is the parameter index), ensuring that any Shares can be restored ; Then, authentication and distribution: for each sub-share (No. Parameters, The pre-shared 256-bit HMAC key is used to generate the authentication code. ;in, Represents the hash message authentication algorithm, based on the SHA256 hash function, which implements keyed hash verification; Represents a pre-shared symmetric key; Through the federal communication protocol, Distribute to An institution records the distribution timestamp; Next, share aggregation: each participating institution returns sub-shares , collect at least Valid shares (with MAC verification of integrity); Call Shamir recovery algorithm to reconstruct ciphertext , used for subsequent clustering operations; Finally, exception handling: If the splitting fails (the ciphertext format is incorrect), fall back to the encryption calculation module 210 to regenerate the ciphertext; if the distribution times out (no confirmation within 30 seconds), mark the mechanism and trigger retransmission (up to 2 times).

[0039] The verification and audit module 230 is deployed on a trusted third-party node to ensure that the aggregation process does not leak original data information by verifying the consistency of the sub-share hash digests submitted by each participating organization.

[0040] As a further explanation of this embodiment, the federal privacy protection unit 200 of this embodiment verifies the consistency of the sub-share hash digests submitted by each participating organization, including the following steps: First, trusted node deployment: Deployed as an independent trusted third-party node, establishing TLS1.3 encrypted channels with all institutions; Then, hash collection: each participating institution calculates the hash of its own share before sending its sub-share , uploaded to the audit node through a secure channel; Then, consistency verification: audit nodes compare the same parameters All If they are consistent, the aggregation is determined to be untampered and the process is allowed to continue; if they are inconsistent, the abnormal organization is marked and the sub-share is required to be retransmitted (up to 3 times); Next, the audit log records participating institutions, hash values, verification results, and timestamps to form a traceable audit log. Finally, abnormal response: If an institution has three abnormal hashes, it will be temporarily excluded from participating in this aggregation (recorded in the blacklist, triggering a compliance check).

[0041] Furthermore, the secret sharing threshold in this embodiment The total number of participating institutions N is defined as ,and ; Ensure that "any organization less than N cannot restore the ciphertext", and block the risk of collusion leakage from a mechanism perspective; the HMAC key uses a 256-bit length and is pre-distributed through an offline USBKey or TLS1.3 encrypted channel, forcing all participating organizations and verification and audit modules to use completely consistent keys, avoiding authentication failures caused by inconsistent keys.

[0042] Dynamic clustering modeling unit 300 is used to build a clustering model that is adaptive to diabetes data features. It uses a hierarchical federation architecture and a dynamic parameter aggregation algorithm, combined with a diabetes course time attenuation factor and a clinical feature weight adjustment strategy to achieve joint clustering analysis of multi-source heterogeneous data. In this embodiment, the dynamic clustering modeling unit 300 adopts a hierarchical federation architecture and a dynamic parameter aggregation algorithm, including the following steps: S310.1. Build a three-tiered federation structure, including edge, regional, and central tiers; Furthermore, the edge layer is deployed in various medical institutions, responsible for processing and modeling local diabetes datasets; the regional layer is deployed in provincial medical data centers, responsible for parameter aggregation and management of the edge layer nodes under its jurisdiction; and the central layer is deployed in national research institutions, responsible for the optimization and distribution of global models.

[0043] As a further illustration of this step, building a three-level federation architecture includes the following steps: First, the edge layer is deployed on the local servers of each medical institution and directly connected to the data acquisition and preprocessing unit 100, responsible for cleaning, feature extraction and model training of the local data set; Subsequently, the regional layer is deployed in the provincial medical data center server cluster, establishing an encrypted connection with the edge layer nodes under its jurisdiction through the federal communication protocol, and taking on the responsibilities of parameter aggregation and node management; Finally, the central layer is deployed on distributed servers in national research institutions, communicating with regional layer nodes through dedicated lines and is responsible for global model optimization and parameter distribution.

[0044] S310.2. Partition the homogeneous clustering input dataset output by the data acquisition and preprocessing unit 100 by medical institution to form a local diabetes dataset for each edge layer node; As a further explanation of this step, based on the homogeneous clustering input data set output by the data acquisition and preprocessing unit 100, it is partitioned according to the medical institution ID, and each edge layer node only retains the data of this institution to form an independent local diabetes data set (including preprocessing features such as patient basic information, physiological indicators, and medical records).

[0045] S310.3. The central layer node pre-trains the initial clustering model based on the historical public diabetes dataset to generate the initial gradient parameters of the regional layer. , and through the federated learning communication protocol Distribute to nodes at each regional layer; As a further illustration of this step, the training and distribution of the initial clustering model includes the following steps: First, the central node selects public medical databases (such as the diabetes subset in NHANES and MIMIC-III) as historical public datasets and uses the k-means algorithm to pre-train an initial clustering model (the number of clusters is preset to 3-5 based on clinical needs); Then, the initial gradient parameters of the regional layer are calculated (The dimension is consistent with the number of features, which is the first-order derivative vector of the model parameters); Finally, the federated learning communication protocol (compatible with the TLS 1.3 protocol of the secure transmission module 120) is used to transmit Encrypted and distributed to nodes at each regional layer.

[0046] S310.4. The edge layer nodes train the initial clustering model based on the local diabetes dataset and calculate the gradient parameters. ,in Indicates the edge layer nodes; As a further illustration of this step, the edge layer nodes are trained based on the local diabetes dataset using the initial model parameters and the gradient parameters are calculated by backpropagation. ( Number the edge layer nodes, such as 1, 2, ..., ), the gradient dimension is consistent with the model parameter dimension, and the calculation timestamp and sample size are recorded.

[0047] S310.5. The regional layer node receives the gradient parameters of the edge layer nodes under its jurisdiction and dynamically adjusts the aggregation weight based on the proportion of data volume of each edge layer. The calculation formula is: ; in, Edge layer sample size; is the sum of all edge layer sample sizes in the region; Edge layer Model gradient Current gradient with regional layer The cosine similarity of Indicates the current iteration round; is the adjustment coefficient, and ; As a further explanation of this step, the regional layer aggregation weight adjustment in this step includes the following steps: First, count the effective sample size of each edge layer node (The remaining number after excluding outliers and samples with missing values ​​exceeding 30%). Then, calculate the total number of samples in the region ( for all edge layer nodes in the region); Then, the model gradient of edge layer i is calculated Current gradient with regional layer Cosine similarity of : ; Finally, calculate the aggregation weight according to the formula of S310.5 above .

[0048] S310.6. Regional layer nodes calculate the aggregate gradient based on the adjusted weights ; ; And send the aggregated gradient to the central layer node; S310.7. The central layer node receives the aggregated gradients from the regional layer nodes and updates the global model: ; in, For regional layer The sample size percentage, is the global gradient; The central layer node will update the global model parameters Feedback is sent to each regional layer and edge layer, and the edge layer nodes are based on Update local model; S310.8, when entering the next round of iteration, each regional layer node will Update to the current gradient , repeat S310.4-S310.7 until the convergence conditions are met.

[0049] In this step, repeat steps S310.4-S310.7 until any of the following convergence conditions is met: Iteration rounds Reaching a preset maximum value (e.g. 50 rounds); Global gradient The modulus is less than the threshold (such as ), that is, the model converges.

[0050] In this embodiment, the dynamic cluster modeling unit 300 combines the diabetes course time attenuation factor with the clinical feature weight adjustment strategy, including the following steps: S320.1. Calculate the diabetes course time decay factor based on the homogeneous clustered input data set output by the data acquisition and preprocessing unit 100. ; ; in, The duration of the disease, is the time decay coefficient, and ;Control the decay rate of historical data; is the complication influence coefficient, which amplifies the effect of complications on the data weight, and ; A scoring function for the severity of complications is used to quantitatively score the types of complications based on the ICD-10 coding system generated by the standardization processing module 130; As a further explanation of this step, the calculation of the diabetes course time attenuation factor in this step includes the following steps: First, calculate the duration of the disease : The number of years from the date the patient was diagnosed with diabetes to the current date (accurate to 0.1 year, for example, if the diagnosis is 3 years and 2 months, then ( =3.2); Then, the complication severity scoring function was defined :Based on the ICD-10 codes generated by the standardization processing module 130, the severity of complications is graded (e.g. no complications =0, mild =1, moderate =2, severe =3); Finally, the attenuation factor is calculated according to the formula : .

[0051] S320.2. Constructing a clinical feature correlation matrix ,in is the total number of features of the homogeneous clustering input data set output by the data acquisition and preprocessing unit 100, Representation matrix yes OK A real matrix of columns; elements Representation characteristics and features The conditional probability of is calculated using the historical data set: ; in, Indicates simultaneous occurrence of features and The number of samples, Indicates the presence of features The historical data set is the homogeneous clustering input data set preprocessed by the standardization processing module 130; Indicates Under the conditions that occur, Probability of occurrence; As a further explanation of this step, the construction of the clinical feature association matrix in this step includes the following steps: First, determine the total number of features : is the number of homogeneous features output by the data acquisition and preprocessing unit 100 (such as blood sugar, glycosylated hemoglobin, course of disease, etc., Typically 10-20); Then, the pre-processed data of the past five years were selected as the historical data set to calculate the matrix elements ; Finally, the correlation threshold (e.g. features are considered as valid associations).

[0052] S320.3. For target features , calculate its dynamic weight vector ; ; in, is a basic weight vector (preset by the normalization processing module 130 based on diabetes clinical guidelines); Associative features The weight of Representation and characteristics Features with a correlation greater than the threshold , through the incidence matrix elements Weighted sum; As a further explanation of this step, the dynamic weight vector calculation in this step includes the following steps: First, preset the basic weight vector :According to the "Guidelines for the Prevention and Treatment of Diabetes in China (2024 Edition)", the importance of features (such as blood sugar , glycosylated hemoglobin ,make sure ); Then, the target features , filter related features (satisfy ); Finally, the dynamic weight is calculated according to the formula .

[0053] S320.4. Dynamic weight vector Applied to cluster distance calculation, the improved weighted Euclidean distance formula is used: ; in, and are two sample points, and For the corresponding features The value of For samples and The weighted Euclidean distance between them; S320.5. Perform clustering iteration based on weighted Euclidean distance and update cluster centers ; ; in, For samples The weight of the feature is determined by the dynamic weight vector of the feature to which it belongs Decide; Indicates that all samples in the cluster are weighted Weighted sum; Represents the sum of the weights of all samples in the cluster; if the number of samples in the cluster is 0, the center value of the previous round is retained.

[0054] S320.6, update the cluster center Feedback to the statistical indicator verification module 410, clustering quality evaluation, if the evaluation index does not meet the preset threshold, then adjust 、 parameters and re-execute S320.1-S320.5.

[0055] As a further explanation of this step, the clustering quality assessment and parameter adjustment in this step include the following steps: First, the updated cluster centers are sent to the statistical indicator verification module 410 to evaluate the cluster quality by the silhouette coefficient and the Calinski-Harabasz index; then, if the silhouette coefficient is less than 0.5 or the Calinski-Harabasz index is less than 100, 、 Adjust parameters within the range (such as Increase by 0.03, Finally, re-execute S320.1-S320.5 until the evaluation indicators meet the standards.

[0056] Furthermore, the dynamic clustering modeling unit 300 in this embodiment receives the homogeneous data set output by the data acquisition and preprocessing unit 100, and transmits the gradient parameters and cluster centers through the encrypted channel of the federal privacy protection unit 200; if the edge layer sample size is 0, the node is marked as "to be activated" and the current round of aggregation is skipped; if the gradient calculation has a numerical overflow (such as ), triggering gradient clipping (scaling the gradient to a modulus of 10) before participating in aggregation to ensure stable operation of the algorithm.

[0057] The dual-track validation and optimization unit 400 is used to ensure the clinical validity of the clustering results. It generates a clustering quality assessment report through the linkage between the statistical indicator verification module 410 and the diagnosis and treatment rule verification module 420, and the evaluation report generation module 430. It also drives the iterative optimization of the clustering model of the dynamic cluster modeling unit 300 based on the indicator deviation in the clustering quality assessment report. In this embodiment, the statistical indicator verification module 410 includes a data distribution verification submodule 411, a stability verification submodule 412, and a separation degree verification submodule 413, wherein: The data distribution verification submodule 411 is used to calculate the statistical distribution indicators of the clustering results (such as the Kolmogorov-Smirnov test value and the chi-square test statistic); As a further explanation of this embodiment, this embodiment calculates the statistical distribution index of the clustering result including the following steps: First, the clustering results (including the characteristic distribution of each cluster sample) output by the dynamic clustering modeling unit 300 are obtained; Then, the Kolmogorov-Smirnov test is performed on the continuous features (such as blood glucose and glycosylated hemoglobin) of each cluster sample, and the test statistic is calculated. (used to determine deviations from the normal distribution, ); Perform chi-square test on categorical features (such as complication type, diabetes type) and calculate chi-square statistic (Used to judge the difference between the actual distribution and the expected distribution. The larger the value, the more significant the deviation); Finally, record the Value (significance level, distribution was considered significantly abnormal).

[0058] The stability verification submodule 412 calculates the adjusted Rand index ARI and the normalized mutual information NMI through the Bootstrap resampling technique to evaluate the consistency of the clustering results under different sampling data sets; As a further explanation of this embodiment, this embodiment evaluates the consistency of clustering results under different sampling data sets, including the following steps: First, we use the Bootstrap resampling technique to generate 10 resampled datasets from the original dataset with replacement sampling (the sample size is the same as the original dataset); Subsequently, the same clustering process was performed on each resampled dataset to obtain 10 groups of clustering results. The adjusted Rand index (ARI) (value range [-1, 1], the closer to 1, the higher the consistency) and normalized mutual information (NMI) (value range [0, 1], the closer to 1, the higher the information overlap) of any two groups of results were calculated. Finally, the mean of ARI and NMI was taken as the stability score (mean < 0.5 was considered as insufficient stability).

[0059] The separation verification submodule 413 is used to calculate the silhouette coefficient and the Calinski-Harabasz index to quantify the compactness and separation of the clustering results; As a further illustration of this embodiment, the calculation of the silhouette coefficient and the Calinski-Harabasz index in this embodiment includes the following steps: First, calculate the silhouette coefficient of each sample (the value range is [-1, 1], a positive number indicates that the sample is closer to the current cluster, and a negative number indicates that it is closer to other clusters), and take the mean of all samples as the overall silhouette coefficient; Then, the Calinski-Harabasz index was calculated (the larger the value, the more compact the clusters and the better the separation between clusters). The calculation formula is: ; Finally, the mean of the silhouette coefficient and the CH index were recorded as quantitative indicators of separation.

[0060] The diagnosis and treatment rule verification module 420 includes a rule extraction submodule 421, a guideline matching submodule 422, and an abnormality identification submodule 423, wherein: The rule extraction submodule 421 is used to parse the complication classification and diagnosis and treatment pathway clauses in the "Chinese Diabetes Prevention and Treatment Guidelines (2024 Edition)" and convert them into executable logical rules based on ICD-10 codes and clinical data features (such as blood glucose range and disease duration); As a further explanation of this embodiment, the parsing and conversion of the rule extraction submodule 421 in this embodiment includes the following steps: First, the "Guidelines for the Prevention and Treatment of Diabetes in China (2024 Edition)" were analyzed to extract core clauses (e.g., "HbA1c should be controlled at <7.0% in patients with type 2 diabetes and nephropathy" and "Patients with a disease duration of ≥10 years should be screened for retinopathy"). Then, convert the clauses into executable logic rules in the format of "IF [feature condition] THEN [expected result]", for example: Rule 1: IF diabetes type = type 2 AND complications = kidney disease (ICD-10: N08.3) THEN HbA1c should be ∈ [4.0%, 7.0%] Rule 2: If the disease duration is ≥ 10 years, then there should be a record of retinopathy screening (ICD-10: H35.0*) Finally, the above data is stored in the rule library (which supports dynamic addition and deletion of rules according to the guidelines).

[0061] The guideline matching submodule 422 is used to match the clustering results with the executable logic rules generated by the rule extraction submodule 421 to generate a rule compliance score; As a further illustration of this embodiment, the guideline matching submodule 422 of this embodiment compares the clustering results with the logical rules in the rule base one by one, calculates the rule compliance of each cluster sample (the proportion of samples that comply with the rules), and grades them according to "excellent (≥80%), good (60%-80%), and poor (<60%)", intuitively presenting the degree of fit between the clustering results and the clinical rules.

[0062] The anomaly identification submodule 423 identifies abnormal data patterns in the clustering results based on the executable logic rules generated by the rule extraction submodule 421; As a further explanation of this embodiment, in this embodiment, the abnormality identification submodule 423 identifies abnormal data patterns in the clustering results, including the following steps: First, based on the characteristic conditions in the rule base, set the abnormal threshold (e.g., in rule 1, glycosylated hemoglobin > 7.0% is abnormal, and in rule 2, a disease course of ≥ 10 years with no screening record is abnormal); Then, scan the clustering results and mark samples that do not conform to the rules as "abnormal data patterns", record the abnormal type (such as "exceeding the glycated hemoglobin standard" or "screening missing") and the cluster number; Finally, the proportion of abnormal samples in each cluster is counted (abnormal proportion = number of abnormal samples / total number of cluster samples) to form an abnormal list.

[0063] The evaluation report generation module 430 is used to fuse the statistical indicators output by the statistical indicator verification module 410 with the rule compliance score output by the guideline matching submodule 422 and generate a clustering quality evaluation report.

[0064] As a further explanation of this embodiment, the generation of the clustering quality assessment report in this embodiment includes the following steps: First, the output of the statistical indicator verification module 410 is integrated: including the Kolmogorov-Smirnov test D value of each cluster, the chi-square test value, mean of silhouette coefficient, CH index, ARI and NMI; Then, the output of the diagnosis and treatment rule verification module 420 is integrated: the rule compliance score of each cluster, the proportion of abnormal samples, and the abnormality type; Then, according to the three-part structured report of "Statistical Indicators - Rule Compliance - Abnormal Analysis", mark the abnormal indicators (such as distribution index, conformity <60%, and clusters with abnormality ratio >20%); Finally, the report is output (supporting PDF and structured JSON formats) and synchronized to the dynamic clustering modeling unit 300 and the system log.

[0065] In this embodiment, in the dual-track verification optimization unit 400, based on the indicator deviation in the cluster quality assessment report, the dynamic cluster modeling unit 300 is driven to iteratively optimize the cluster model, including the following steps: S440.1. Extraction index deviation: Extracting indicator deviation items based on the clustering quality assessment report output by the assessment report generation module 430, including statistical indicator deviation and diagnosis and treatment rule deviation; As a further illustration of this embodiment, the following deviation items are extracted from the evaluation report: Statistical index deviation: such as the mean value of silhouette coefficient <0.5 (insufficient separation), the mean value of ARI <0.6 (insufficient stability), Distribution indicators (distribution anomalies); Deviation from diagnosis and treatment rules: such as cluster conformity <60% (poor rule matching) and abnormality ratio >20% (many abnormal patterns).

[0066] S440.2. Attribution of deviation items: Statistical indicator deviation: locate the parameter problem of dynamic cluster modeling unit 300 (such as time decay coefficient (deviating from the range of 0.05-0.1), clinical characteristics correlation matrix If the silhouette coefficient is low, it is attributed to the unreasonable setting of the correlation threshold δ of the clinical feature correlation matrix M of the dynamic clustering modeling unit 300 (e.g., the threshold is too high, resulting in the omission of correlation features); if the ARI / NMI is low, it is attributed to the time decay coefficient Deviation from the range of 0.05-0.1 (such as Too large will result in too low a weight for historical data); Treatment rule deviation: Positioning the feature weight strategy of the dynamic clustering modeling unit 300 (such as the basic weight vector Not in line with the guide presets, dynamic weight vector If the degree of conformity is low, it is attributed to the basic weight vector Not in line with the guideline assumptions (e.g., insufficient weighting of core features); if the abnormality ratio is high, attribute it to the dynamic weight vector Inadequate integration of clinically relevant features (e.g., comorbidity features were underweighted).

[0067] S440.3. Generate parameter adjustment instructions: Adjusting for statistical bias: Value, modified feature correlation threshold (Based on clinical characteristics association matrix actual strength of association adjustment); For regular deviations: recalibrate the basis weight vector (According to the "Chinese Diabetes Prevention and Treatment Guidelines (2024 Edition)", the core feature weights were updated and the influence coefficient of complications was adjusted. ; As a further explanation of this embodiment, with respect to statistical deviation: if Deviation range, adjusted within 0.05-0.1 (such as the current =0.12, adjust it to 0.1); If the correlation threshold is unreasonable, adjust it within 0.1-0.3 (for example, adjust the original threshold of 0.3 to 0.2 to expand the range of correlation features); Targeting deviations from the rules: Recalibration (e.g., increase the weight of glycated hemoglobin from 0.2 to 0.3, based on the priority of the guidelines); adjust the influence coefficient of complications within 0.2-0.4 (As currently =0.2 to 0.3, increasing the weight of complications).

[0068] S440.4, adjust the instruction (including the adjusted 、 、 threshold, ) is sent to the dynamic clustering modeling unit 300, driving the dynamic clustering modeling unit 300 to re-execute S320.1-S320.6 and S310.4-S310.8; S440.5. After the dynamic clustering modeling unit 300 outputs the updated clustering results, the dual-track verification optimization unit 400 re-verifies through the statistical indicator verification module 410 and the diagnosis and treatment rule verification module 420 until the indicator deviation of the clustering quality assessment report is lower than the preset threshold or reaches the maximum number of iterations.

[0069] As a further illustration of this embodiment, after re-verification, if the evaluation report states: Statistical indicators: silhouette coefficient ≥ 0.5, ARI / NMI ≥ 0.6, The distribution index of ≥0.05 accounts for ≥80%; Rule indicators: All clusters must meet ≥70% compliance and anomaly ratio ≤10%; The iteration is terminated; otherwise, S440.1-S440.4 are repeated until the maximum number of iterations is reached (for example, 5 times is preset).

[0070] Furthermore, after the dual-track verification optimization unit 400 in this embodiment receives the clustering results of the dynamic clustering modeling unit 300, the statistical indicator verification module 410 and the diagnosis and treatment rule verification module 420 are processed in parallel, and the results are summarized in the evaluation report generation module 430; if a module fails to calculate (such as a rule library loading error, the statistical test data is empty), it is marked as "verification interrupted", and the previous round of report results are used to drive optimization, and the error log (including module name and timestamp) is recorded at the same time to support manual intervention and repair.

[0071] The intelligent decision support unit 500 is used to convert clustering results into clinical decision recommendations, and generate personalized treatment plan recommendations and complication risk warnings based on the diabetes clinical guideline knowledge base and machine learning prediction model.

[0072] In this embodiment, the intelligent decision support unit 500 includes a clustering result processing module 510, a decision generation module 520, and a result output module 530, wherein: The clustering result processing module 510 receives the clustering results output by the dynamic clustering modeling unit 300 and extracts the key clinical characteristics of the patient (such as type, blood sugar level, and disease course); As a further illustration of this embodiment, the clustering result processing module 510 in this embodiment serves as a key link in data integration. After receiving the clustering results output by the dynamic cluster modeling unit 300, it first analyzes the characteristic distribution of each cluster sample. For example, it identifies clinically significant cluster labels such as "high blood sugar fluctuation cluster" and "long-term complications cluster" in the clustering results. It then extracts key clinical indicators from the characteristics of each sample's cluster, including diabetes type (type 1 / type 2), current mean blood sugar level (fasting / postprandial), disease duration, complication type (based on ICD-10 codes), and medication history. These features are structured and integrated into a "patient feature vector," where categorical features retain their original codes (e.g., type is represented by 0 / 1), and continuous features retain their normalized values ​​(to facilitate subsequent model invocation). This ensures that the data passed to the decision generation module 520 is both complete and meets clinical analysis requirements.

[0073] The decision generation module 520 generates personalized treatment plan recommendations and complication risk warnings based on the diabetes clinical guideline knowledge base and the machine learning prediction model, combined with the extracted cluster features; To further illustrate this embodiment, the decision generation module 520 in this embodiment is the core of intelligent decision-making. Its operation relies on a constructed diabetes clinical guideline knowledge base and an adapted machine learning prediction model. The clinical guideline knowledge base is based on the "Chinese Diabetes Prevention and Treatment Guidelines (2024 Edition)" and structuredly stores clauses such as diagnosis and treatment pathways, medication recommendations, and complication screening. For example, clauses such as "Type 2 diabetes patients with HbA1c ≥ 7.0% should initiate drug treatment" and "Those with a disease course of ≥ 5 years should be screened annually for diabetic nephropathy" are broken down into callable rules (formatted as "IF feature condition THEN recommended measure") and associated with corresponding evidence levels (e.g., Class I evidence, Class II evidence), enabling precise matching based on patient characteristics. The machine learning prediction model is trained based on historical diagnosis and treatment data (using highly interpretable algorithms such as logistic regression and random forest). Its input is the patient feature vector extracted by the clustering result processing module 510, and its output is the complication risk level (e.g., "low / medium / high risk") and treatment suitability (e.g., the probability that a particular medication regimen is suitable for a given patient).

[0074] Furthermore, during the actual decision-making process, the decision generation module 520 first matches the patient's feature vector with rules in the clinical guideline knowledge base. For example, if a patient belongs to the "long-term complication cluster" with a disease duration of 10 years or longer, a "retinopathy screening" recommendation is automatically triggered; if fasting blood glucose levels are consistently ≥7.0 mmol / L, a "metformin initial treatment" clause is matched. Simultaneously, a machine learning prediction model is used to assess complication risk. Combining the patient's blood glucose fluctuation characteristics and BMI, it outputs warning results such as "risk of diabetic nephropathy in the next year" and "risk of cardiovascular events," and annotates key risk-related features (e.g., "blood glucose fluctuation amplitude is positively correlated with risk"). Ultimately, treatment recommendations and risk warnings are integrated into a "decision candidate set." The treatment plan includes the drug type, starting dose range (adjusted based on guideline-recommended weight and renal function), and monitoring frequency recommendations. The risk warning specifies the risk level and corresponding intervention measures (e.g., "moderate risk requires urine microalbumin recheck every three months").

[0075] The result output module 530 integrates the treatment plan and risk warning and outputs clinical decision-making recommendations.

[0076] To further illustrate this embodiment, the result output module 530 in this embodiment is responsible for converting the "decision candidate set" generated by the decision generation module 520 into a clinically readable output format. It first prioritizes the information in the candidate set—for example, placing "high-risk complication warnings" first. Treatment options are ranked by "guideline evidence level + model fit" (with Class I evidence and high fit preferred). Subsequently, the report is presented in the form of a structured report, including basic patient information (de-identified), cluster labels, a summary of key features, recommended treatment options (including medications, dosages, and precautions), risk warning levels and their rationale, and follow-up recommendations (such as blood glucose monitoring frequency and recheck items). The report supports two output formats: a clinician-oriented visual interface (highlighting the comparison between risk warnings and options) and an archivable PDF document (containing the complete decision logic chain, such as "Based on cluster feature X matching guideline item Y, the model predicts risk Z"), ensuring that the output facilitates rapid decision-making while meeting the regulatory requirements of medical records.

[0077] As a further illustration of this embodiment, Furthermore, the interactive logic and process of the multiple modules in this embodiment are as follows: Clustering result processing module 510 transmits the structured feature vector to decision generation module 520, which then calls the knowledge base and model to return a set of decision candidates. Result output module 530 completes information integration and formatting. If an anomaly occurs in any link (e.g., a feature vector lacks key indicators, or a knowledge base call fails), it is automatically marked as "decision-restricted" and basic suggestions are output (e.g., "Refer to the conventional diagnosis and treatment pathway, supplement examination item A, and then regenerate a decision"), while also recording the anomaly details (including timestamp and module name) to support subsequent tracing and repair.

[0078] Example 2 like Figure 2 As shown, this embodiment also provides a method for intelligent clustering analysis of diabetes clinical trial data based on federated learning. The method is based on the above-mentioned intelligent clustering analysis system for diabetes clinical trial data based on federated learning, and includes the following steps: S100, Data Collection and Preprocessing: Based on the edge computing architecture and federated learning communication protocol, multi-center diabetes clinical data are collected, standardized after secure transmission, and a homogeneous clustered input data set is generated; S200, federated privacy protection: A lightweight homomorphic encryption algorithm is used to encrypt the clustering intermediate parameters. The ciphertext parameters are split and distributed to each participating institution through a threshold secret sharing protocol. The privacy security of the aggregation process is verified by a trusted third party. S300, Dynamic Clustering Modeling: Constructing a three-tier federated architecture consisting of edge, regional, and central layers, and performing joint clustering of multi-source heterogeneous data based on a dynamic parameter aggregation algorithm. Combining the diabetes course time decay factor with a clinical feature weight adjustment strategy, the dynamic weight vector is calculated and applied to cluster distance calculation to generate initial clustering results. S400, Dual-Track Verification and Optimization: Generate a clustering quality assessment report by integrating statistical indicator verification and diagnosis and treatment rule verification. Based on the indicator deviations in the report, locate parameter problems and feature weight strategy deviations, and generate adjustment instructions to drive iterative optimization of the clustering model until the indicator deviation in the clustering quality assessment report falls below the preset threshold or the maximum number of iterations is reached. S500, Intelligent Decision Support: Extract key clinical features from the optimized clustering results, combine the diabetes clinical guideline knowledge base with the machine learning prediction model to generate personalized treatment plan recommendations and complication risk warnings, and output structured clinical decision recommendations.

[0079] Those skilled in the art will appreciate that the process of implementing all or part of the steps of the above embodiments may be accomplished by hardware, or by instructing related hardware through a program.

[0080] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent clustering analysis system for diabetes clinical trial data based on federated learning, characterized by: include: A data collection and preprocessing unit (100) is used to realize privacy-preserving collection and standardized preprocessing of multi-center diabetes data, build a cross-institutional data security aggregation channel based on edge computing architecture and federated learning communication protocol, and combine physiological indicator normalization rules and clinical feature coding strategies to generate a homogeneous clustering input data set; The federal privacy protection unit (200) is used to ensure the privacy security of cross-institutional data interaction, encrypt and aggregate clustering intermediate parameters through a secure multi-party computing protocol, and achieve privacy protection of the original data stored locally; A dynamic clustering modeling unit (300) is used to construct a clustering model that is adaptive to diabetes data features, using a hierarchical federation architecture and a dynamic parameter aggregation algorithm, combined with a diabetes course time attenuation factor and a clinical feature weight adjustment strategy, to achieve joint clustering analysis of multi-source heterogeneous data; A dual-track verification and optimization unit (400) is used to ensure the clinical effectiveness of the clustering results, by linking the statistical indicator verification module (410) with the diagnosis and treatment rule verification module (420), and generating a clustering quality evaluation report through the evaluation report generation module (430); and based on the indicator deviation in the clustering quality evaluation report, driving the iterative optimization of the clustering model of the dynamic cluster modeling unit (300); The intelligent decision support unit (500) is used to convert the clustering results into clinical decision recommendations, and generate personalized treatment plan recommendations and complication risk warnings based on the diabetes clinical guideline knowledge base and machine learning prediction model.

2. The diabetes clinical trial data intelligent clustering analysis system based on federated learning according to claim 1, characterized in that: The data acquisition and pre-processing unit (100) comprises an edge acquisition module (110), a secure transmission module (120) and a standardization processing module (130), wherein: The edge acquisition module (110) is used to collect multi-source clinical data from diabetes clinical monitoring and treatment equipment and a hospital electronic medical record system, and includes a medical device interface submodule (111) and an electronic medical record docking submodule (112), wherein: the medical device interface submodule (111) supports a USB-C wired interface and a Bluetooth 5.0 wireless interface for accessing real-time data from diabetes clinical monitoring and treatment equipment; the electronic medical record docking submodule (112) uses the HL7FHIR standard protocol to extract patient basic information, medical records, and treatment history data from the hospital electronic medical record system; The secure transmission module (120) uses the TLS1.3 encryption protocol to build a transmission channel, performs shard encryption processing on the collected data, and each data shard is accompanied by a check code generated based on the SHA-256 algorithm, and achieves breakpoint-resume transmission of the sharded data through the federated learning communication protocol; The standardization processing module (130) has a built-in clinical data conversion engine, which performs hierarchical preprocessing on the collected data: Z-score standardization is used for continuous physiological indicators, one-hot encoding is used for complication types, medication regimens, diabetes classification, and treatment methods, and missing data is filled in by weighted interpolation based on the correlation of clinical characteristics.

3. The intelligent clustering analysis system for diabetes clinical trial data based on federated learning according to claim 2, characterized in that: In the standardization processing module (130), the missing data are filled using a weighted interpolation method based on the correlation of clinical characteristics, including the following steps: S130.

1. For target clinical features with missing values, analyze their association with clinical features other than the target clinical features, and screen out feature subsets that are associated with clinical diagnosis and treatment; S130.

2. Assign a corresponding weight to each associated feature based on the clinical importance of the associated feature subset in the diagnosis and treatment of diabetes; S130.

3. Based on the measured data and corresponding weights of the associated features, calculate the filling value of the missing value of the target feature, and verify whether the filled data conforms to the clinical distribution law of the feature.

4. The intelligent clustering analysis system for diabetes clinical trial data based on federated learning according to claim 3, characterized in that: The federal privacy protection unit (200) includes an encryption calculation module (210), a security aggregation module (220) and a verification and audit module (230), wherein: The encryption calculation module (210) uses a lightweight homomorphic encryption algorithm to encrypt the clustering intermediate parameters to generate a ciphertext parameter vector; The security aggregation module (220) splits the ciphertext parameter vector into N sub-shares based on a threshold secret sharing protocol, where N is the number of participating institutions; and distributes each sub-share to the corresponding participating institution through a federated learning communication protocol, where each sub-share is accompanied by a message authentication code generated based on HMAC-SHA256; The verification and audit module (230) is deployed on a trusted third-party node to ensure that the aggregation process does not leak original data information by verifying the consistency of the sub-share hash digests submitted by each participating organization.

5. The intelligent clustering analysis system for diabetes clinical trial data based on federated learning according to claim 4, characterized in that: In the dynamic clustering modeling unit (300), a hierarchical federation architecture and a dynamic parameter aggregation algorithm are adopted, including the following steps: S310.

1. Build a three-tiered federation structure, including edge, regional, and central tiers; S310.2, partitioning the homogeneous clustering input data set output by the data acquisition and preprocessing unit (100) by medical institution to form a local diabetes data set for each edge layer node; S310.

3. The central layer node pre-trains the initial clustering model based on the historical public diabetes dataset to generate the initial gradient parameters of the regional layer. , and through the federated learning communication protocol Distribute to nodes at each regional layer; S310.

4. The edge layer nodes train the initial clustering model based on the local diabetes dataset and calculate the gradient parameters. ,in Indicates the edge layer nodes; S310.

5. The regional layer node receives the gradient parameters of the edge layer nodes under its jurisdiction and dynamically adjusts the aggregation weight based on the proportion of data volume of each edge layer. , the calculation formula is: ； in, Edge layer sample size; is the sum of all edge layer sample sizes in the region; Edge layer Model gradient Current gradient with regional layer The cosine similarity of Indicates the current iteration round; is the adjustment coefficient; S310.

6. Regional layer nodes calculate the aggregate gradient based on the adjusted weights : ； And send the aggregated gradient to the central layer node; S310.

7. The central layer node receives the aggregated gradients from the regional layer nodes and performs global model updates: ； in, For regional layer The sample size percentage, is the global gradient; The central layer node will update the global model parameters Feedback is sent to each regional layer and edge layer, and the edge layer nodes are based on Update local model; S310.8, when entering the next iteration, each regional layer node will Update to the current gradient , repeat S310.4-S310.7 until the convergence conditions are met.

6. The intelligent clustering analysis system for diabetes clinical trial data based on federated learning according to claim 5, characterized in that: In the dynamic clustering modeling unit (300), the diabetes course time attenuation factor and the clinical feature weight adjustment strategy are combined, including the following steps: S320.

1. Calculating the time decay factor of the diabetes course based on the homogeneous clustered input data set output by the data acquisition and preprocessing unit (100) ; ； in, The duration of the disease, is the time decay coefficient; it controls the decay speed of historical data; is the complication influence coefficient, which amplifies the effect of complications on the data weight, and ; A scoring function for complication severity was developed, and complication types were quantitatively scored based on the ICD-10 coding system generated by the standardized processing module ( 130 ); S320.

2. Constructing a clinical feature correlation matrix ,in is the total number of features of the homogeneous clustering input data set output by the data acquisition and preprocessing unit (100), Representation matrix yes OK A real matrix of columns; elements Representation characteristics and features The conditional probability of is calculated using the historical data set: ； in, Indicates simultaneous occurrence of features and The number of samples, Indicates the presence of features The historical data set is the homogeneous clustering input data set pre-processed by the standardization processing module (130); Indicates Under the conditions that occur, Probability of occurrence; S320.

3. For target features , calculate its dynamic weight vector ; ； in, is the basic weight vector; Associative features The weight of Representation and characteristics Features with a correlation greater than the threshold , through the incidence matrix elements Weighted sum; S320.

4. Dynamic weight vector Applied to cluster distance calculation, the improved weighted Euclidean distance formula is used: ； in, and are two sample points, and For the corresponding features The value of For samples and The weighted Euclidean distance between them; S320.

5. Perform clustering iteration based on weighted Euclidean distance and update cluster centers ; S320.6, update the cluster center Feedback is sent to the statistical indicator verification module (410) to perform clustering quality evaluation. If the evaluation indicator does not meet the preset threshold, the 、 parameters and re-execute S320.1-S320.

5.

7. The intelligent clustering analysis system for diabetes clinical trial data based on federated learning according to claim 6, characterized in that: The statistical indicator verification module (410) includes a data distribution verification submodule (411), a stability verification submodule (412) and a separation degree verification submodule (413), wherein: The data distribution verification submodule (411) is used to calculate the statistical distribution index of the clustering results; The stability verification submodule (412) calculates the adjusted Rand index (ARI) and the normalized mutual information (NMI) by using the Bootstrap resampling technique to evaluate the consistency of the clustering results under different sampling data sets; The separation verification submodule (413) is used to calculate the silhouette coefficient and the Calinski-Harabasz index to quantify the compactness and separation of the clustering results; The diagnosis and treatment rule verification module (420) includes a rule extraction submodule (421), a guideline matching submodule (422) and an abnormality identification submodule (423), wherein: The rule extraction submodule (421) is used to parse the complication classification and diagnosis and treatment pathway clauses in the "Guidelines for the Prevention and Treatment of Diabetes in China (2024 Edition)" and convert them into executable logical rules based on ICD-10 coding and clinical data characteristics; The guideline matching submodule (422) is used to match the clustering results with the executable logic rules generated by the rule extraction submodule (421) to generate a rule compliance score; The anomaly identification submodule (423) identifies abnormal data patterns in the clustering results based on the executable logic rules generated by the rule extraction submodule (421); The evaluation report generation module (430) is used to fuse the statistical indicators output by the statistical indicator verification module (410) and the rule compliance score output by the guideline matching submodule (422), and generate a clustering quality evaluation report.

8. The intelligent clustering analysis system for diabetes clinical trial data based on federated learning according to claim 7, characterized in that: In the dual-track verification optimization unit (400), based on the indicator deviation in the clustering quality evaluation report, the clustering model iterative optimization of the dynamic clustering modeling unit (300) is driven, including the following steps: S440.

1. Extraction index deviation: Extracting indicator deviation items based on the clustering quality assessment report output by the assessment report generation module (430), including statistical indicator deviation and diagnosis and treatment rule deviation; S440.

2. Attribution of deviation items: Statistical indicator deviation: Parameter issues in locating dynamic cluster modeling units (300); Diagnosis and treatment rule deviation: feature weight strategy for positioning dynamic clustering modeling unit (300); S440.

3. Generate parameter adjustment instructions: Adjusting for statistical bias: Value, modified feature correlation threshold ; For regular deviations: recalibrate the basis weight vector , Adjust the influence coefficient of complications ; S440.4, sending the adjustment instruction to the dynamic clustering modeling unit (300), driving the dynamic clustering modeling unit (300) to re-execute S320.1-S320.6 and S310.4-S310.8; S440.

5. After the dynamic clustering modeling unit (300) outputs the updated clustering results, the dual-track verification optimization unit (400) re-verifies through the statistical indicator verification module (410) and the diagnosis and treatment rule verification module (420) until the indicator deviation of the clustering quality assessment report is lower than the preset threshold or the maximum number of iterations is reached.

9. The intelligent clustering analysis system for diabetes clinical trial data based on federated learning according to claim 8, characterized in that: The intelligent decision support unit (500) comprises a clustering result processing module (510), a decision generation module (520) and a result output module (530), wherein: The clustering result processing module (510) receives the clustering result output by the dynamic clustering modeling unit (300) and extracts the key clinical features of the patient; The decision generation module (520) generates personalized treatment plan recommendations and complication risk warnings based on the diabetes clinical guideline knowledge base and the machine learning prediction model, combined with the extracted clustering features; The result output module (530) integrates the treatment plan and risk warning and outputs clinical decision-making suggestions.

10. A method for intelligent clustering analysis of diabetes clinical trial data based on federated learning, based on the intelligent clustering analysis system for diabetes clinical trial data based on federated learning according to any one of claims 1 to 9, characterized in that: The steps include: S100, Data Collection and Preprocessing: Based on the edge computing architecture and federated learning communication protocol, multi-center diabetes clinical data are collected, standardized after secure transmission, and a homogeneous clustered input data set is generated; S200, federated privacy protection: A lightweight homomorphic encryption algorithm is used to encrypt the clustering intermediate parameters. The ciphertext parameters are split and distributed to each participating institution through a threshold secret sharing protocol. The privacy security of the aggregation process is verified by a trusted third party. S300, Dynamic Clustering Modeling: Constructing a three-tier federated architecture consisting of edge, regional, and central layers, and performing joint clustering of multi-source heterogeneous data based on a dynamic parameter aggregation algorithm. Combining the diabetes course time decay factor with a clinical feature weight adjustment strategy, the dynamic weight vector is calculated and applied to cluster distance calculation to generate initial clustering results. S400, Dual-Track Verification and Optimization: Generate a clustering quality assessment report by integrating statistical indicator verification and diagnosis and treatment rule verification. Based on the indicator deviations in the report, locate parameter problems and feature weight strategy deviations, and generate adjustment instructions to drive iterative optimization of the clustering model until the indicator deviation in the clustering quality assessment report falls below the preset threshold or the maximum number of iterations is reached. S500, Intelligent Decision Support: Extract key clinical features from the optimized clustering results, combine the diabetes clinical guideline knowledge base with the machine learning prediction model to generate personalized treatment plan recommendations and complication risk warnings, and output structured clinical decision recommendations.

Images