Federal learning optimization method and system for multi-source heterogeneous cardiovascular disease risk prediction

Abstract

The invention provides a federal learning optimization method and system for multi-source heterogeneous cardiovascular disease risk prediction, and relates to the technical field of data processing, and the method comprises the steps: processing original cardiovascular disease data, and obtaining standardized feature data; based on the standardized feature data, a differential training algorithm corresponding to the data structure of the standardized feature data is adopted for local calculation, and updated encrypted local parameters are generated; the updated encrypted local parameters are transmitted to a parameter server in a federated learning framework, the parameter server adopts a geometric mean algorithm based on secure multi-party calculation to perform secure aggregation and geometric mean integration on local parameter updates from different sources in an encryption state, and a preliminary global parameter set is generated. According to the method, data privacy protection is enhanced through a federated learning framework, the multi-source heterogeneous data fusion efficiency is improved, and parameter integration is optimized to enhance model performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a federated learning optimization method and system for multi-source heterogeneous cardiovascular disease risk prediction. BACKGROUND

[0002] In a cardiovascular disease research collaborative network scenario composed of multiple hospitals and medical institutions, each institution has a large amount of cardiovascular disease related data of patients, including clinical diagnosis records, examination data, image data, etc. These data are widely sourced and structurally different, belonging to typical multi-source heterogeneous data. Traditional schemes combined with traditional neural networks have exposed some technical defects when processing these data.

[0003] When inputting data into a traditional neural network for processing, the traditional neural network is not adaptable to complex data structures. For multi-source heterogeneous data, it is difficult to automatically extract effective features. For example, when facing cardiovascular disease data containing numerical, textual and image types at the same time, the traditional neural network often needs to manually extract and encode features in advance, which not only has a huge workload, but also the manually extracted features may not fully reflect the internal information of the data, resulting in the model being unable to fully learn the key patterns in the data. The training process of the traditional neural network has a large amount of calculation, and the processing efficiency is low for large-scale multi-source heterogeneous data. Especially when processing image data, the complex pixel matrix requires a large amount of computing resources and time to process, which makes the model training period long and difficult to quickly iterate and optimize to adapt to changing data and clinical needs. In addition, the amount of data from different sources is often severely unbalanced, and some sources with small amounts of data are easily ignored in the traditional neural network training process, resulting in insufficient learning of the model on the features of these data, reducing the adaptability and generalization ability of the model to the overall data. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a federated learning optimization method and system for multi-source heterogeneous cardiovascular disease risk prediction, which improves the collaborative utilization efficiency of multi-source heterogeneous data.

[0005] To solve the above technical problems, the technical solutions of the present application are as follows: In a first aspect, a federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction is provided, the method comprising: Step 1: processing the original cardiovascular disease data to obtain standardized feature data; Step 2: based on the standardized feature data, using a differentiated training algorithm corresponding to the data structure to perform local calculation and generate updated encrypted local parameters; Step 3: The updated encrypted local parameters are transmitted to the parameter server in the federated learning framework, and the parameter server uses a geometric mean algorithm based on secure multi-party computation to securely aggregate and geometrically average the local parameter updates from different sources in an encrypted state, generating a preliminary global parameter set; Step 4: Based on the preliminary global parameter set, the multi-source heterogeneous features are mapped to a unified geometric manifold space through a feature space geometric projection algorithm based on manifold learning, and feature distribution coordination and model weight redistribution are performed in this space, generating an optimized global parameter set; Step 5: The optimized global parameter set is loaded into a multi-layer perceptron risk assessment network based on attention mechanism, the input layer of the network receives the multi-source feature data to be evaluated, the feature data is converted into a high-dimensional vector by the embedding layer, the importance weighting and fusion of feature vectors of different sources and types are performed by the attention calculation layer, and finally the quantitative cardiovascular disease risk assessment results of the integrated multi-source data are generated by the output layer.

[0006] In a second aspect, a federated learning optimization system for multi-source heterogeneous cardiovascular disease risk prediction includes: A standardization module for processing original cardiovascular disease data to obtain standardized feature data; A calculation module for generating updated encrypted local parameters based on standardized feature data using differentiated training algorithms corresponding to their data structures; An integration module for transmitting updated encrypted local parameters to a parameter server in a federated learning framework, and the parameter server uses a geometric mean algorithm based on secure multi-party computation to securely aggregate and geometrically average local parameter updates from different sources in an encrypted state, generating a preliminary global parameter set; A coordination and optimization module for mapping multi-source heterogeneous features to a unified geometric manifold space through a feature space geometric projection algorithm based on manifold learning based on the preliminary global parameter set, and performing feature distribution coordination and model weight redistribution in this space to generate an optimized global parameter set; A risk assessment module for loading the optimized global parameter set into a multi-layer perceptron risk assessment network based on attention mechanism, receiving multi-source feature data to be evaluated through the input layer of the network, converting the feature data into a high-dimensional vector by the embedding layer, performing importance weighting and fusion of feature vectors of different sources and types through the attention calculation layer, and finally generating quantitative cardiovascular disease risk assessment results of integrated multi-source data through the output layer.

[0007] In a third aspect, a computing device includes: one or more processors; a memory device storing one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method.

[0008] In a fourth aspect, a computer-readable storage medium stores a program, when executed by a processor, implements the method.

[0009] The above scheme of the present application at least includes the following beneficial effects: By processing the original cardiovascular disease data, the original data of different sources and formats is converted into standardized feature data, so that the originally scattered and heterogeneous original data can form a unified standard feature data set. For the feature data of different data structures after standardization, a differentiated training algorithm adapted to the feature data is used for local calculation. Through the precise matching of the algorithm and the data structure, the core features of each type of data are maximized, and the pertinence and effectiveness of local data processing are improved. In the local parameter transmission and aggregation link, through the encryption processing of the local parameters, combined with the geometric mean algorithm based on secure multi-party computation, the parameter aggregation in the encrypted state is realized, and the original data of each data source does not need to be separated from the local, only the encrypted local parameters are transmitted. At the same time, the secure multi-party computation ensures that each participant cannot obtain the parameter information of the other party in the aggregation process, and realizes the effective integration of multi-source parameters. The feature space geometric projection algorithm based on manifold learning maps the multi-source heterogeneous features to a unified geometric manifold space, realizes the spatial alignment of the heterogeneous features through geometric projection, and further optimizes the adaptability between the features in the unified space, so that the features of different types and different sources can be efficiently fused to form a more correlated feature set, and the collaborative utilization efficiency of the multi-source heterogeneous data is improved. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 is a flowchart of a multi-source heterogeneous cardiovascular disease risk prediction federated learning optimization method provided by an embodiment of the present application.

[0011] Figure 2 is a schematic diagram of a multi-source heterogeneous cardiovascular disease risk prediction federated learning optimization system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0012] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in many forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0013] As Figure 1 shown, the embodiments of the present application propose a federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction, which comprises the following steps: Step 1, processing the original cardiovascular disease data to obtain standardized feature data; Step 2, based on the standardized feature data, using a differentiated training algorithm corresponding to its data structure for local calculation, generating updated encrypted local parameters; Step 3, transmitting the updated encrypted local parameters to the parameter server in the federated learning framework, using the geometric mean algorithm based on secure multi-party computation to securely aggregate and geometrically integrate the local parameter updates from different sources in an encrypted state, generating a preliminary global parameter set; Step 4, based on the preliminary global parameter set, mapping multi-source heterogeneous features to a unified geometric manifold space through a feature space geometric projection algorithm based on manifold learning, performing feature distribution coordination and model weight redistribution in this space, generating an optimized global parameter set; Step 5, loading the optimized global parameter set into a multi-layer perceptron risk assessment network based on the attention mechanism, receiving multi-source feature data to be evaluated through the input layer of the network, converting the feature data to a high-dimensional vector through the embedding layer, performing importance weighting and fusion on feature vectors of different sources and types through the attention calculation layer, then performing nonlinear transformation and deep feature extraction on the weighted and fused features through the multi-layer perceptron module, and finally generating a quantitative cardiovascular disease risk assessment result of the comprehensive multi-source data through the output layer.

[0014] In the embodiments of the present application, the original data has different dimensions, and standardization maps the features to a uniform scale, avoids large value features dominating model training, and unifies the data formats of different hospitals; different algorithms such as LSTM for time series data, CNN for image data, and BERT for text data are used to fit the structural characteristics of different types of data; based on secure multi-party computation, the parameters are aggregated in an encrypted state, and the parameter server can complete the calculation without decryption, which prevents the single-point risk of parameter leakage after decryption and avoids each institution obtaining the local parameters of others, protecting the training results and data privacy of each institution; through manifold learning, the internal geometric correlation of multi-source heterogeneous features is mined and mapped to a unified Riemannian manifold space, so that features that cannot be directly aligned can be optimized together. The attention calculation layer automatically identifies high-influence features in the data to be evaluated and gives them higher weights; the weights of secondary features are reduced to avoid the dilution of key information caused by traditional MLP treating all features equally, so that the evaluation result is more in line with the actual risk of patients; the risk of cardiovascular disease is influenced by the interaction of multiple features, and the multi-layer nonlinear transformation of MLP can deeply mine this complex correlation and accurately capture the implicit risk pattern; a specific risk score is generated instead of a qualitative high / low risk, and doctors can develop differentiated solutions based on the score to solve the problem of ambiguous traditional evaluation results that are difficult to guide precise intervention.

[0015] In a preferred embodiment of the present application, step 1 comprises: Step 100, data cleaning and missing value processing are performed on the original cardiovascular disease data to generate a normalized data set. Specifically, first, the source and type of the original cardiovascular disease data are sorted, including clinical diagnosis records, examination and test data, image data, time series monitoring data, and text reports; through the data unique identifier composed of patient ID, examination time, and examination item name, records with complete duplication or core information duplication are screened out; combined with cardiovascular clinical common sense, a reasonable range for each index is set, and data outside the range is preliminarily determined as abnormal; for image data, the pixel value distribution of normal images of the same type is compared to identify abnormal areas such as all black / all white caused by equipment failure; for abnormal values, first check the original collection records, correct if it is an input error, and keep the data and mark the clinical abnormality identifier if it cannot be corrected or is confirmed as a real clinical abnormality.

[0016] The missing proportion of each data field is calculated, and fields with a missing proportion exceeding 50% and a secondary clinical significance, such as patient occupation, are directly excluded; for fields with a missing proportion less than 50%, the data type is processed: for numerical data such as blood lipids, the mean or median of patients of the same age, gender, and disease type is used for filling, for time series data, the adjacent time point data is used for linear interpolation filling, for text data, the unrecorded or common description of the same type of case is used for filling, and for image local missing areas, the mean of the surrounding normal pixels is used for filling, finally forming a normalized data set without duplication, invalid abnormality, and key missing.

[0017] At step 101, based on the normalization data set, feature extraction operation is performed for different types of data features respectively, wherein sliding window feature extraction algorithm is adopted for time sequence waveform data to obtain time sequence feature vector; convolution feature extraction algorithm is adopted for image matrix data to generate image feature matrix; natural language processing feature extraction algorithm is adopted for text report data to obtain text feature; based on the time sequence feature vector, the image feature matrix and the text feature, a multi-modal feature set is generated, specifically, a window is set according to the sampling frequency (ECG 100Hz: 200-point window, 100-point step; dynamic blood pressure 1Hz: 60-point window, 30-point step), the data is segmented into continuous segments; the mean, variance, maximum, peak value number and waveform slope of each segment are extracted to form a time sequence feature vector, which is integrated into a time sequence feature vector set.

[0018] The color image is converted to gray scale, the gray scale value is normalized, and the edge black border is cropped; a convolution module is constructed (coronary CT 5-8 layers, ultrasound 3-5 layers, convolution kernel CT uses 5x5, ultrasound uses 3x3, each layer is connected with ReLU activation + pooling after convolution); the preprocessed image is input, and the edge / texture (low-level) to core area (high-level) features are extracted layer by layer to generate an image feature matrix.

[0019] Traditional Chinese is converted to simplified Chinese, irrelevant symbols are removed, and medical dictionary segmentation and meaningless stop words are removed to retain medical terms / symptoms / examination items; the key words are converted to 100-200-dimensional vectors by Word2Vec / GloVe, the single report vector is averaged / summed, and the text feature set is integrated; the time sequence, image and text features are classified according to data source and type to ensure that the three types of features of the same patient correspond one by one to form a multi-modal feature set.

[0020] At step 102, the multi-modal feature set is input into a feature standardization processing unit, and a maximum and minimum value normalization algorithm is used to map each dimension feature to a unified numerical interval to generate a standardized feature vector, specifically, a feature standardization processing unit is constructed, and different types of features in the multi-modal feature set are processed separately; all feature dimensions are traversed, all sample data under each dimension are collected, each dimension of the time sequence feature, each channel dimension of the image feature matrix and each dimension of the text feature vector are calculated respectively; [0, 1] is selected as the standardized unified interval according to the subsequent model training requirements; the standardized time sequence feature vector, image feature matrix and text feature vector are recombined, and each patient corresponds to a group of standardized feature vectors containing multiple types of standardized features.

[0021] In step 103, based on the standardized feature vector, the feature dimension reduction processing is performed through a principal component analysis algorithm to obtain standardized feature data, specifically, the standardized feature vector is arranged by rows, each row represents the complete standardized feature of a patient, and each column represents a feature dimension; the mean of each feature dimension is calculated first, then the covariance of any two feature dimensions is calculated, and the covariance matrix is constructed by traversing all dimensions; the eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvalue and the corresponding eigenvector; the eigenvalues are sorted in descending order, and the variance contribution rate and the cumulative variance contribution rate of each eigenvalue are calculated; the cumulative variance contribution rate threshold is set to 85%-95%, the first k eigenvalues when the cumulative contribution rate reaches the threshold are selected, and the corresponding eigenvectors form a principal component matrix; the multiplication operation is performed on the standardized feature matrix and the principal component matrix, and the standardized feature vector of each patient is projected into a low-dimensional space to obtain a low-dimensional feature vector; and the low-dimensional feature vectors of all patients are integrated to form standardized feature data which retains key information and has reduced dimension.

[0022] In the embodiment of the present application, by eliminating noise and outliers, the interference of data acquisition error on subsequent analysis is reduced, the authenticity and accuracy of the data set are ensured, the missing values are processed to avoid the reduction of sample size caused by data missing, the data set is normalized to make the subsequent feature extraction, standardization and other operations smoother, and the sliding window, convolution, natural language processing and other adaptive extraction algorithms are adopted for different types of data such as time sequence waveform, image matrix and text report to capture the core features of various data; the generated time sequence feature vector, image feature matrix and text feature comprehensively capture the internal key information of various modal data from multiple dimensions such as time variation, spatial form and semantic description; by mapping each dimension feature to a unified numerical interval, the weight imbalance problem caused by different dimensions is solved, so that the features such as heart rate, blood lipid and image gray value have the same comparison basis in model training; the principal component analysis algorithm extracts the principal component with the maximum variance to remove the linear correlation information between features, reduces the feature dimension and simplifies the data structure.

[0023] In a preferred embodiment of the present application, the above step 2 comprises: In step 200, a corresponding training algorithm is selected according to the data structure characteristics of the standardized feature data, wherein the time sequence feature data adopts a time sequence feature extraction algorithm, the image feature data adopts a spatial feature extraction algorithm, and the text feature data adopts a semantic feature extraction algorithm; specifically, the standardized data from 24-hour electrocardiogram and dynamic blood pressure monitoring are classified as time sequence feature data, the features of coronary CT image and heart ultrasound image after standardization are classified as image feature data, and the text features of clinical medical records and examination reports after standardization are classified as text feature data.

[0024] For time series feature data, the long short-term memory network or the gated recurrent unit is selected as the time series feature extraction algorithm, the long-term dependence relationship in the time series data is remembered through the gating mechanism, and the time series characteristics of the cardiovascular time series data are adapted; for image feature data, the convolutional neural network is selected as the spatial feature extraction algorithm, the local spatial features of the image are captured through the convolution kernel sliding, the dimension is reduced and the key information is preserved through the pooling layer, and the spatial structure characteristics of the image matrix data are adapted; for text feature data, the pre-training model based on the Transformer is selected as the semantic feature extraction algorithm, the context semantic association of the words in the text is captured through the self-attention mechanism, and the semantic structure characteristics of the text report are adapted.

[0025] In step 201, the standardized feature data is loaded into the corresponding training algorithm as input for forward calculation process to generate intermediate calculation results of each data type. Specifically, the time series feature data is segmented into fixed-length input sequences according to time segments, such as 24-hour electrocardiogram data segmented into 24 sequence segments according to 1 hour; the image feature data is adjusted to a uniform size matrix and added with a channel dimension to meet the input requirements of the convolutional neural network; the text feature data is converted into a fixed-length word index sequence, and the insufficient length is supplemented with a filler and the length exceeding is truncated to adapt to the input format of the semantic feature extraction algorithm.

[0026] For time series feature data, the adapted input sequence is transmitted into the LSTM / GRU model, the time information in the sequence is first filtered and remembered through the input gate, the forgetting gate and the output gate, the feature vector containing the time series change rule is generated in the hidden layer, and then the intermediate calculation result corresponding to the time series feature is generated through the preliminary processing of the full connection layer, such as the heart rate abnormal probability prediction value of each time segment; for image feature data, the adapted feature matrix is transmitted into the CNN model, the first convolution kernel sliding extracts the basic spatial features of image edges, textures, etc., after the introduction of nonlinearity through the ReLU activation function, the key features are selected and the dimension is reduced through the max-pooling layer; the subsequent convolution layers and pooling layers are alternately performed, and higher-level spatial features are gradually extracted, and finally the multi-dimensional feature matrix is converted into a one-dimensional feature vector through the global average pooling layer to generate the intermediate calculation result corresponding to the image feature; for text feature data, the adapted word index sequence is transmitted into the BERT model, the word index is converted into a word vector through the embedding layer, the context semantic association between words is captured through the multi-layer self-attention layer, and finally the semantic feature vector of the entire text sequence is extracted through the pooling layer to generate the intermediate calculation result corresponding to the text feature; finally, the intermediate calculation results of the three types of data are collected.

[0027] Step 202: Based on the intermediate calculation results of each data type, the difference between the predicted output and the true annotation is calculated using the error backpropagation mechanism. The parameter configuration in the forward calculation process is adjusted through the gradient optimization algorithm to generate the local parameter update. Specifically, the true annotation of the time series feature data is the abnormal event determined by the clinician based on the 24-hour monitoring results; the true annotation of the image feature data is the diagnosis result of the radiologist on the image; and the true annotation of the text feature data is the final diagnosis conclusion given by the clinician. The intermediate calculation results of the three types of data (time series, image, and text) are compared with the corresponding true annotations. The cross-entropy loss function is used to calculate the prediction error of each type of data, or the mean squared error loss function is used to calculate the regression error. The errors of the three types of data are weighted and summed according to the proportion of data volume to obtain the overall error value.

[0028] Starting from the error calculation layer, the gradient of each layer's model parameters with respect to the overall error is calculated sequentially along the reverse path of the forward calculation, clarifying the direction and degree of each parameter's influence on the error. The Adam optimization algorithm is selected, and combined with a pre-set learning rate, it is dynamically decayed according to the number of training iterations. The parameters of each layer are adjusted according to the gradient direction, and the numerical difference before and after each parameter adjustment is recorded. All parameter adjustment differences are classified and organized according to the model layer structure and parameter type (weight, bias) to form a local parameter update quantity containing parameter adjustment information of each layer.

[0029] Step 203: Based on the pre-negotiated encryption security level, select a Paillier encryption scheme that satisfies the additive homomorphic property and generate a public-private key pair. Distribute the public key to each medical institution node. Convert each numerical parameter in the local parameter update quantity into a floating-point representation that meets the input requirements of the encryption scheme. Perform parallel encryption operations on each numerical parameter after conversion using the public key pair to generate the corresponding homomorphic ciphertext. Combine the ciphertexts of all parameters according to the original parameter structure and add integrity verification information to generate updated encrypted local parameters. Specifically, organize each medical institution node to pre-negotiate the encryption security level and determine the key length of the Paillier encryption scheme according to the medical data privacy protection requirements. The core node generates a Paillier public-private key pair through the secure encrypted transmission channel. The private key is stored separately by the core node and protected by multiple encryption methods, while the public key is distributed to all participating medical institution nodes to ensure that each node can use the public key for encryption operations.

[0030] Iterate through each numerical parameter in the local parameter update list, checking whether its data format conforms to the Paillier encryption scheme requirements. Since the Paillier encryption scheme has specific limitations on the precision and range of input data, convert all parameters from decimal fractions to floating-point representations with 6 decimal places. Start multi-threaded parallel encryption operations, with each thread corresponding to one numerical parameter. Use the acquired public key to perform encryption calculations on the converted parameters, generating Paillier homomorphic ciphertext for each parameter. After encryption, check the integrity of each ciphertext, discard parameters that failed to encrypt, and re-encrypt to ensure that all parameters can generate valid ciphertext.

[0031] Based on the original structure of the local parameter update, each layer is classified according to weight parameters and bias parameters. The valid ciphertext of all parameters is recombined to form a ciphertext set consistent with the original parameter update structure. The ciphertext set is hashed using the SHA-256 hash algorithm to generate an integrity check value. The ciphertext set and the integrity check value are packaged and integrated, and the node identifier of the medical institution is added to finally form an updated encrypted local parameter containing ciphertext, check information, and node identifier.

[0032] In this embodiment of the invention, appropriate extraction algorithms are matched to the different structural characteristics of time-series, image, and text data to ensure that the key features of each type of data can be accurately captured, providing a targeted algorithmic foundation for subsequent training. Standardized data is input into the corresponding algorithm to complete forward computation, effectively transforming the features of each type of data into intermediate results that can be used for error analysis, providing an accurate calculation basis for subsequent backpropagation to optimize parameters. Error backpropagation is used to locate parameter deviations, and gradient optimization is combined to adjust parameters, making the local parameters continuously conform to the characteristic patterns of the institution's data, thus improving the accuracy of the local parameters. Paillier additive homomorphic encryption is used to protect parameter privacy, supporting subsequent secure aggregation without decryption. Public and private key distribution, parallel encryption, and integrity verification not only prevent theft or tampering during parameter transmission and storage but also meet the requirements for medical data privacy protection.

[0033] In a preferred embodiment of the present invention, step 3 above includes: Step 300: The parameter server receives updated encrypted local parameters from each node and performs source authentication based on digital signature and integrity verification based on hash value on each received encrypted update packet to obtain a verified set of encrypted parameter updates. Specifically, firstly, the parameter server establishes a dedicated secure data receiving channel, which is encrypted using the SSL / TLS 1.3 protocol to ensure the link security when each medical institution node (hereinafter referred to as node) transmits encrypted local parameters; each node sends the updated encrypted local parameters (including ciphertext set, SHA-256 hash check value, node digital signature, and node identifier) ​​to the parameter server through this channel according to a preset transmission cycle.

[0034] Next, the parameter server pre-stores the public keys of all legitimate nodes, separate from the Paillier public key, as dedicated public keys for node authentication. For each received encrypted update packet, the node's digital signature and node identifier are extracted from the packet. The digital signature is decrypted and verified. If the core information obtained after decryption is completely consistent, the source authentication is passed. If decryption fails or the original text does not match, it is marked as an illegal source packet and discarded directly. Subsequently, the parameter server extracts the SHA-256 hash checksum from the encrypted update packet and re-performs SHA-256 hash calculation on the ciphertext set (excluding checksum, signature, and other auxiliary information) in the update packet to obtain a real-time hash value. The real-time hash value is compared bit by bit with the pre-stored hash checksum from the packet. If they are completely identical, the integrity verification is passed. If they are inconsistent, it is marked as a tampered packet and the corresponding node is required to retransmit. This process continues until all encrypted local parameters from legitimate nodes are received and verified, ultimately forming a verified encrypted parameter update set.

[0035] Step 301: Based on the verified encrypted parameter update set, parse the metadata information contained in each encrypted update package to identify the data feature type and model structure type corresponding to each parameter update. Specifically, the metadata is auxiliary information added synchronously by each node when generating encrypted local parameters, containing three core fields: first, a data feature type label, indicating the original data type corresponding to the parameter update; second, a model structure type label, indicating the model structure used to train the parameter; and third, parameter dimension and level information, recording the model level and parameter dimension to which the parameter belongs. The parameter server performs metadata analysis on each encrypted update package in the verified encrypted parameter update set. According to the parsing operation, the binary encoded information of the metadata fields in the package is read through the preset metadata parsing module and converted into readable text identifiers. At the same time, combined with parameter dimension and hierarchical information, the matching between data feature type and model structure type is further verified. For example, if the data feature type is imaging-coronary CT, but the model structure type is labeled as LSTM time series model, it is determined that the type is mismatched, the update package is marked, and the corresponding node is contacted for verification and correction. If the match is consistent, the encrypted parameter update is bound to the combined label of data feature type and model structure type to form a parameter-type mapping table, which clearly identifies the core type attribute corresponding to each parameter update.

[0036] Step 302: Based on the identified data feature type and model structure type, an unsupervised clustering algorithm based on ciphertext similarity measurement is used to group the encrypted parameter updates, generating multiple parameter update groups based on type similarity. Specifically, since the parameters are in an encrypted state (Paillier ciphertext), the parameter values ​​cannot be directly parsed. Therefore, the ciphertext of each encrypted parameter is expanded into binary bits to form a fixed-length ciphertext vector. For matrix-type parameters, the ciphertext vectors of all elements of the matrix are concatenated in row-major order to form a ciphertext feature vector of uniform dimension.

[0037] The parameter types have been initially classified using metadata. First, they are classified according to the combination of data feature type and model structure type labels. Encrypted parameters with the same label are updated and grouped into the same pregroup. Then, for the encrypted feature vectors in each pregroup, a similarity measure based on Euclidean distance is used to calculate the Euclidean distance between any two encrypted vectors. Subsequently, the K-means unsupervised clustering algorithm is started. The number of clusters K is determined according to the actual difference of parameters in the pregroup. The clustering error corresponding to different K values ​​is calculated. The K value that tends to stabilize after the error drops sharply is selected. If there are two types of parameters for coronary CT and cardiac ultrasound, the K value is set to 2 to further subdivide the two types of parameters.

[0038] After clustering is completed, multiple parameter update groups based on type similarity are generated. The parameter server checks the grouping results by sampling, randomly selecting parameters in each group and comparing them with the specific data subtypes in their metadata. If the specific subtypes of the parameters in the same group are consistent, the grouping is deemed reasonable. If there is subtype mixing, the clustering parameters are readjusted and clustering is performed again until all parameters in all groups meet the requirements of type homology and feature similarity.

[0039] Step 303: Update the grouping for each parameter, and perform secure multi-party geometric mean calculation under encrypted conditions. The weight of each node is dynamically determined based on the data quality index and the number of valid samples for that node, generating encrypted aggregation results for each group. Specifically, the missing value rate, outlier rate, and clinical annotation consistency of the data for that type of node are statistically analyzed. The higher the consistency rate with the expert review results, the higher the quality. The three indicators are weighted and summed in a 4:3:3 ratio to obtain a data quality score of 0-100. The number of valid samples is the number of qualified patient samples after cleaning used by the node to train the parameters of that group. Subsequently, the weight of each node is calculated according to the formula: Node weight = (Data quality score × 0.6 + Normalized value of valid sample number × 0.4) / Sum of weights of all nodes. The normalized value of valid sample number is mapped to the 0-100 range through maximum and minimum value normalization.

[0040] Because the parameters are encrypted using Paillier and support additive homomorphism, the geometric mean calculation requires a logic of logarithmic transformation, addition, and exponential transformation. First, the parameter server performs a pseudo-logarithmic transformation on the encrypted local parameters of each node within the group. Utilizing Paillier homomorphism, the numerical value of the encrypted parameter is converted into an encrypted logarithmic form without decryption, achieved through preset homomorphic operation rules. Second, each encrypted logarithmic parameter is multiplied by the weight of that node, with the weight in plaintext. Homomorphic multiplication is used to multiply the encrypted parameter by the plaintext weight. Then, homomorphic addition is used to sum the weighted encrypted logarithmic parameters of all nodes to obtain the encrypted weighted logarithmic sum. Third, the encrypted weighted logarithmic sum is converted into an encrypted exponential form to obtain the encrypted geometric mean of the group's parameters, i.e., the group's encrypted aggregation result.

[0041] Step 304: Based on the encrypted aggregation results of each group, calculate the contribution weight of different group aggregation results to the global model, and perform weighted fusion of the results of each group according to the weight to generate a preliminary global parameter set. Specifically, the parameter server, in conjunction with experts, constructs a group contribution weight evaluation system to evaluate the importance of each group from three dimensions: first, clinical diagnostic value: assessing the clinical significance of the features corresponding to the group for cardiovascular disease risk assessment, directly related to the severity of the disease, with high clinical value; text-medical record group assists in supplementing medical history information, with secondary clinical value, scored from 0-100; second, feature discrimination: through... The training set validates the features corresponding to the grouping parameters and their ability to distinguish between diseased / non-diseased severe / mild cases, measured by the area under the ROC curve (AUC), mapped to a score of 0-100. Thirdly, data coverage: the coverage ratio of the data corresponding to this group across all nodes is calculated, with 100% coverage; only 30% of nodes have imaging-cardiac MRI data, with 30% coverage. Higher coverage indicates stronger universality of the grouping parameters for the global model, mapped to a score of 0-100. Subsequently, the contribution weight of each group is calculated as: Contribution Weight = Clinical Diagnostic Value × 0.5 + Feature Discrimination × 0.3 + Data Coverage × 0.2.

[0042] Next, the parameter server uses Paillier's private key, held only by the parameter server and stored with multiple layers of encryption, to decrypt the encrypted aggregation result of each group, obtaining the plaintext form of the group aggregation parameters. Then, the plaintext aggregation parameters of each group are multiplied by the contribution weight of that group to obtain the weighted group parameters. Finally, according to the structural hierarchy of the global model, all weighted group parameters are integrated into the corresponding layers of the model: time series group parameters correspond to the LSTM feature extraction layer, image group parameters correspond to the CNN feature extraction layer, and text group parameters correspond to the BERT feature extraction layer. Different group parameters within the same layer are fused according to their weight proportions. After fusion, a preliminary global parameter set is generated, which includes parameters of each layer of the model, parameter dimension descriptions, and generation timestamps.

[0043] In this embodiment of the invention, digital signatures confirm that parameters originate from legitimate medical institution nodes, preventing the infiltration of illegally forged parameters; hash value verification ensures that parameters have not been tampered with during transmission, providing a reliable and authentic encrypted parameter foundation for subsequent aggregation and ensuring the security of the starting point of global model aggregation; accurate identification of the data feature type and model structure type corresponding to the parameters provides a basis for subsequent aggregation by type, ensuring that the aggregation direction matches the characteristics of the data and model; grouping by ciphertext similarity in the encrypted state protects parameter privacy from leakage and groups parameters of the same type together, avoiding interference between heterogeneous parameters; encrypted state calculation ensures that parameter privacy is not decrypted and leaked; geometric mean suppresses extreme parameter interference and improves aggregation robustness; and the contribution weights of each group to the global model are fused, allowing key groups, such as imaging parameter groups which are more important for cardiovascular risk assessment, to play a greater role, ensuring that the initial global parameters can integrate the advantages of various types of parameters and fit the overall training objectives of the model.

[0044] In a preferred embodiment of the present invention, step 4 above includes: Step 400: Based on the preliminary global parameter set, the Riemannian manifold projection algorithm is used to map features from different data sources to a unified Riemannian manifold space, generating features aligned to the manifold space. Specifically, the parameters corresponding to time-series features, image features, and text features in the preliminary global parameter set are identified, and the original data format and dimensional information of each type of feature are extracted. For time-series features, the dynamic time warping algorithm is used to align time-series sequences of different lengths. For image features, bilinear interpolation is used to uniformly scale image matrices of different resolutions to a size of 224×224, and the matrices are flattened into one-dimensional vectors. For text features, L2 normalization is performed on the word embedding vectors.

[0045] Cardiovascular features exhibit nonlinear correlations. A symmetric positive definite Riemannian manifold is chosen as the unified space, as this manifold can effectively characterize the geometric structure of nonlinear data. A Riemannian manifold projection algorithm based on logarithmic and exponential mappings is employed. First, the preprocessed features of each type are converted into vectors in the tangent space of the Riemannian manifold through logarithmic mapping. Then, based on the global geometric structure of the manifold, the tangent space vectors are projected onto the unified coordinate system of the symmetric positive definite Riemannian manifold through exponential mapping. Finally, the overlap of different source features in the manifold space after projection is calculated to form a feature set aligned with the manifold space.

[0046] Step 401: Based on the features aligned in the manifold space, calculate the geometric difference between different feature distributions by geodesic distance to generate a multi-source feature distribution difference measurement result. Specifically, based on the features aligned in the manifold space, divide the distribution groups according to the dimensions of data source and feature type, and classify the features of different data sources under the same feature type into major categories to facilitate the calculation of differences within the group and differences within the major category. Next, considering the geometric characteristics of symmetric positive definite Riemannian manifolds, the log-Euclidean geodesic distance algorithm is selected to balance computational efficiency and difference characterization accuracy. For any two feature distributions, the weighted average of all feature points in the manifold space is used, with the weight being the number of samples corresponding to the feature. Then, logarithmic mapping is performed on the two geometric centers to convert them into vectors in the tangent space. Finally, the Euclidean distance between the two vectors in the tangent space is calculated, which is the geodesic distance between the two feature distributions on the Riemannian manifold.

[0047] For intra-group differences, the geodesic distance between different distribution groups within the same major category is calculated. For intra-category differences, the geodesic distance between the overall geometric centers of the time-series category, image category, and text category is calculated. Simultaneously, the distribution dispersion of each distribution group is calculated, represented by the mean of the geodesic distances from all feature points within that group to its geometric center, reflecting the degree of distribution concentration within a single group. Finally, a multi-source feature distribution difference measurement result is generated, and all geodesic distances and distribution dispersions are organized into a difference measurement matrix. The matrix rows and columns represent each distribution group / major category, and the cell values ​​are the corresponding geodesic distances. The dispersion values ​​of each distribution group are appended to the matrix. High-difference pairs and high-dispersion groups are also labeled to provide clear optimization targets for subsequent distribution coordination.

[0048] Step 402: Based on the multi-source feature distribution difference measurement results, the weight diffusion algorithm is used to calculate the importance weight of each feature dimension in the manifold space, generating a geometric feature importance weight distribution. Specifically, each feature dimension aligned with the manifold space is used as a node, such as the peak heart rate dimension in time series features, the coronary artery stenosis area dimension in image features, and the myocardial ischemia keyword dimension in text features. The geodesic distance between any two nodes on the manifold is calculated. The smaller the distance, the stronger the geometric correlation between the two feature dimensions. The connection strength between nodes is set based on the distance. The smaller the distance, the stronger the connection strength, forming a feature association graph that intuitively reflects the intrinsic correlation of each feature dimension.

[0049] Next, based on common sense in cardiovascular clinical diagnosis and treatment, core clinical feature dimensions are determined, and these core dimensions are assigned initial high weights. Feature dimensions with less clinical significance are assigned initial low weights, and neutral feature dimensions are assigned initial medium weights after data verification to confirm their importance. Then, using the nodes with initial high weights as diffusion sources, weights are diffused to adjacent nodes according to the connection strength of the feature association graph. For example, if the connection strength between core dimension A (weight 0.8) and node B is 0.9, then node B receives a weight increment of 0.8 × 0.9 × 0.3 (0.3 is the diffusion attenuation coefficient) from A. At the same time, if the connection strength between node B and node C is 0.7, then node B diffuses its updated weight (initial 0.5 + increment 0.216 = 0.716) to C, and so on, until the weight update magnitude of all nodes is less than a preset threshold, at which point diffusion stops. During the diffusion process, if a node receives weight increments from multiple sources simultaneously, the sum of the increments is added to its own weight.

[0050] Finally, the weight values ​​of all feature dimensions are input into the normalization processing module. The maximum-minimum normalization algorithm is used to map the weights to the 0-1 interval, ensuring that the sum of the weights of all feature dimensions is 1. Subsequently, a geometric feature importance weight distribution table is generated, which labels the name of each feature dimension, the corresponding data source type, the normalized weight value, and arranges them in descending order of weight, clearly presenting the hierarchy of core features (weight > 0.6) - important features (0.3 < weight ≤ 0.6) - minor features (weight ≤ 0.3).

[0051] Step 403: Based on the geometric feature importance weight distribution, perform parameter renormalization operation on the preliminary global parameter set based on the manifold geometry to generate a renormalized parameter set. Specifically, analyze the structure of the preliminary global parameter set, clarify the feature dimension corresponding to each parameter, and construct a parameter-feature-weight mapping table by matching parameter labels with feature dimension names. For each parameter, first extract its geometric attributes on the Riemannian manifold, and then adjust the parameter in combination with the importance weight of the corresponding feature. For parameters corresponding to high-weight features (weight > 0.6), appropriately increase the magnitude of the parameter, with the upper limit of the magnitude being 1.5 times the magnitude of the preliminary parameter, while keeping the direction of the parameter vector on the manifold unchanged. For parameters corresponding to low-weight features (weight ≤ 0.3), decrease the magnitude of the parameter, with the lower limit of the magnitude being 0.5 times the magnitude of the preliminary parameter. For parameters corresponding to medium-weight features, keep the magnitude basically unchanged, and only fine-tune the projection position of the parameter in the tangent space of the manifold to make it closer to the central region of the feature distribution.

[0052] 10% of the parameters are randomly selected from both the initial global parameter set and the renormalized parameter set, and input into the model validation module. Forward computation is performed using a small batch of cardiovascular validation data, and the prediction errors of the two sets of parameters are compared, such as the accuracy of classification tasks and the mean squared error of regression tasks. If the prediction error of the renormalized parameters is lower than that of the initial parameters, the renormalization is deemed effective. If the error increases, the influence coefficient of the weights on the parameters is adjusted, and renormalization is performed again until the error meets the requirements. All the renormalized parameters that have passed the validation are reorganized according to the structure of the initial global parameters, and a parameter renormalization log is added to record the adjustment range of each parameter, the corresponding feature weight, and the error change before and after the adjustment, forming a renormalized parameter set with a clear structure and adapted to the importance of features.

[0053] Step 404: Perform distribution consistency optimization on the renormalized parameter set to generate an optimized global parameter set. Specifically, based on the difference metric matrix and combined with the renormalized parameter set, recalculate the distribution center and dispersion of each data source feature on the Riemann manifold. Simultaneously, statistically analyze the average distribution center of the same feature type across all data sources, calculate the deviation between the distribution center of a single data source and the average center, identify high-bias distributions and high-dispersion distributions, and clarify the key targets for distribution consistency optimization. For high-bias distributions, employ a domain adaptive alignment algorithm, constructing a dual-module structure of a feature extractor and a domain classifier for feature extraction. The feature extractor extracts multi-source features based on renormalization parameters, while the domain classifier attempts to distinguish the data source to which the features belong. Through backpropagation, the feature extractor and the domain classifier are optimized simultaneously, enabling the feature extractor to learn domain-invariant features, i.e., feature patterns common to different data sources, and gradually reducing the distribution deviation of features from different data sources. For highly discrete distributions, a distribution smoothing algorithm is used to calculate the geodesic distance from all feature points in the distribution to their center. For feature points whose distance exceeds the dispersion threshold, their positions on the manifold are adjusted according to the principle of fine-tuning towards the center, with the fine-tuning amplitude being 10%-20% of the distance, while the corresponding renormalization parameters are updated simultaneously.

[0054] After each optimization round, align the high-bias distribution and smooth the high-dispersion distribution, recalculate the geodesic distance and dispersion of all feature distributions, and count the number of high-bias and high-dispersion distributions. If the number decreases by ≥30% compared to the previous round, continue to the next round of optimization. If the reduction is less than 10% or the preset number of iterations is reached, stop optimization. After each round of optimization, test the model performance with full validation data and record the model error on each data source. Organize the parameters after the final round of optimization according to both model level and feature type to form a structured parameter file. At the same time, attach a distribution consistency optimization report, including a comparison of feature distribution differences before and after optimization, a table of model error changes for each data source, and optimization algorithm parameter settings.

[0055] In this embodiment of the invention, multi-source heterogeneous features (time series / images / text) are mapped to a unified Riemannian manifold space to solve the spatial incompatibility problem of features from different data sources. By utilizing the geodesic distance unique to the manifold space, the geometric differences in the distribution of features from different sources are accurately quantified, avoiding the limitation of traditional distance calculations failing to reflect nonlinear feature differences. Based on the geometric correlation and differences of features in the manifold space, the importance weights of each feature dimension are dynamically allocated to ensure that features key to cardiovascular disease risk assessment receive higher weights, improving parameter targeting. Combining the manifold geometry and feature importance weights, preliminary global parameters are adjusted to adapt the parameters to the distribution patterns of multi-source features in the manifold space, improving the adaptability of parameters to multi-source heterogeneous features. This reduces the distribution gap of multi-source features in the manifold space, alleviates the problem of unbalanced distribution of data from different institutions, and reduces the model's adaptation bias towards some data sources.

[0056] In a preferred embodiment of the present invention, step 400 includes: Step 4001: Extract the feature representations of each data source in the preliminary global parameter set, calculate the statistical moments and distribution characteristics of each feature representation, and generate feature distribution descriptors. Specifically, for time-series data sources, such as 24-hour electrocardiograms from Hospital A and dynamic blood pressure data from Hospital B, extract the hidden layer output vector of the LSTM time-series model in the preliminary global parameters. This vector contains the dynamic features of the time-series data, such as heart rate fluctuation trends and blood pressure peak intervals. Each time-series sample corresponds to a 256-dimensional hidden layer feature representation. For image data sources, extract the output matrix of the last convolutional layer of the CNN image model, which contains the spatial features of the image. Flatten the matrix into a 512-dimensional vector as the feature representation of a single image. For text data sources, extract the [CLS] token embedding vector of the BERT text model, which contains the semantic features of the text, such as symptom descriptions and diagnostic conclusion associations. Each text sample corresponds to a 768-dimensional feature representation.

[0057] For each data source's feature representation, statistical moments are calculated for each feature dimension. The first moment (mean) reflects the overall level of the feature in that dimension, the second moment (variance) reflects the dispersion of the feature values, the third moment (skewness) reflects the asymmetry of the feature distribution, and the fourth moment (kurtosis) reflects the steepness of the feature distribution. Simultaneously, considering the characteristics of cardiovascular clinical data, the distribution type of each feature dimension is determined using the QQ plot test and the KS test. A structured descriptor document is constructed for each data source, containing three core pieces of information: first, basic feature information, indicating the data source type, feature dimension, and sample size; second, a set of statistical moments, organizing the specific values ​​of the first to fourth moments by dimension, such as the mean 75, variance 12.5, skewness 0.3, and kurtosis 3.2 for time-series feature dimension 1; and third, a description of distribution characteristics, recording the distribution type of each dimension, the confidence level of the distribution test, and the clinical significance association, forming a complete feature distribution descriptor.

[0058] Step 4002: Based on the feature distribution descriptor, the principal tangent space analysis method is used to construct the projection mapping relationship from the features of each data source to the common Riemannian manifold, generating an initial projection operator set. Specifically, considering the nonlinear correlation characteristics of cardiovascular multi-source features, a symmetric positive definite Riemannian manifold is selected as the common mapping space. This manifold can effectively characterize the geometric structure of medical data and supports smooth mapping of nonlinear features. Based on the feature dimensions of each data source, the dimension of the manifold is set to 1024 dimensions. For the feature representation of each data source, based on the statistical moments in its feature distribution descriptor... The process involves identifying the central feature point, calculating the mean of all sample feature representations across all dimensions to form a central feature vector, converting this vector into an SPD matrix, and using the outer product of the feature vectors as the reference center of the data source on the SPD manifold. A tangent space is constructed with the reference center as the origin, representing a linear approximation of the manifold at the reference center. Principal component analysis (PCA) is performed on all sample feature representations within the tangent space, selecting the top K principal components with a cumulative variance contribution rate of 90%. These principal components constitute the principal component basis of the tangent space.

[0059] For each data source, a projection operator is designed based on the principal component basis of its principal tangent space and the geometry of the common SPD manifold. The operator consists of two parts: a linear mapping matrix and a manifold transformation rule. The linear mapping matrix projects the feature representation of the data source (such as a 256-dimensional time-series vector) onto the principal component basis of its principal tangent space, obtaining a tangent space vector. The manifold transformation rule, through exponential mapping and the core operation of the Riemannian manifold, converts the tangent space vector into point coordinates on the common SPD manifold. Each data source corresponds to one initial projection operator, forming an initial projection operator set. 10% of the samples from each data source are selected and mapped to the common SPD manifold using the corresponding initial operator. The geodesic distance between the mapped samples and the reference center of the data source on the manifold is calculated. If the distance of more than 90% of the samples is less than a preset threshold, based on the characteristic fluctuation range of clinical data, the initial operator is deemed valid. If not, the cumulative variance contribution rate threshold of the principal component analysis is readjusted, such as increasing it to 95%, and the operator is reconstructed.

[0060] Step 4003: Optimize the initial projection operator set using the Riemann metric learning algorithm to minimize the distribution divergence of different source features in the manifold space, generating optimized projection operators. Specifically, considering the geometric characteristics of the SPD Riemann manifold, Wasserstein distance and KL divergence are selected as core evaluation metrics. For each data source's mapped feature distribution, calculate the Wasserstein distance and KL divergence with all other data source distributions to form a multi-source distribution divergence matrix. The core objective is to minimize the average divergence among all data source distributions, while constraining the deviation between the mapped feature distribution of each data source and its original distribution to not exceed 10%. In the objective function, the Wasserstein distance and KL divergence are weighted and summed at a weight of 6:4 to serve as the total divergence loss. The optimization process involves adjusting the parameters of the projection operator, the elements of the linear mapping matrix, and the smoothing coefficient of the exponential mapping to minimize the total divergence loss.

[0061] Gradient descent is employed to adapt to nonlinear optimization of Riemannian manifolds. Feature samples from each data source are input in batches, with each batch containing 50 samples from each data source. These samples are mapped to the manifold space using an initial projection operator, and the total divergence loss for the current batch is calculated. The parameters of the projection operator are fine-tuned along the direction of loss descent. For example, if the divergence values ​​of the time-series and image data sources are too high, the linear mapping matrices of their projection operators are adjusted to make the mapped time-series features closer to the distribution area of ​​the image features, while ensuring that the deviations from the original distributions are within constraints. Every 10 iterations, the total divergence loss of all samples is calculated. If the decrease in loss value is less than 0.001, convergence is achieved, and optimization stops. If convergence is not achieved, iterations continue until the preset number of iterations or convergence conditions are reached. Finally, the divergence matrices of the multi-source distributions before and after optimization are compared. If the average divergence value decreases by ≥30%, and the deviations from the original distributions after mapping for each data source are all ≤10%, the optimization is considered effective. If not, the weights of the objective function are readjusted, and optimization is repeated. After successful optimization, the adjusted projection operators are saved, forming an optimized projection operator set.

[0062] Step 4004: Using the optimized projection operator, the feature representations of each data source are mapped from the original feature space to a unified Riemannian manifold space, generating a set of feature points in the manifold space. Specifically, for time-series feature representations, abnormal samples caused by equipment failures in the original data are first removed, and then the samples are sorted in chronological order. For image feature representations, L2 normalization is performed on the flattened vectors. For text feature representations, samples containing invalid semantics are filtered out to ensure that the input mapped feature representations are all high-quality data.

[0063] Processing is done sequentially according to data source type. For time-series data sources, the 256-dimensional feature representation of each sample is input into the optimized time-series projection operator. It is first projected onto the principal tangent space using a linear mapping matrix, and then converted to 1024-dimensional coordinates on the common SPD manifold using an exponential mapping. For image data sources, the 512-dimensional feature vector is input into the optimized image projection operator, and the same linear and exponential mappings are performed to obtain manifold feature points. For text data sources, the 768-dimensional embedding vector is input into the optimized text projection operator to generate manifold feature points. During the mapping process, multi-threaded parallel processing is used, with each data source... An independent thread is allocated; a three-dimensional index label is added to each feature point generated by mapping: first, a data source identifier, indicating the original source of the feature point; second, a sample identifier, associating the feature point with the original patient information; and third, a feature type identifier, indicating the core clinical significance of the feature point. Finally, all feature points mapped from all data sources are classified and organized according to the feature type identifier, with each category further subdivided according to the data source identifier, forming a structured manifold space feature point set. At the same time, a point set statistical report is generated, including the number of feature points from each data source, the coordinate range of the feature points in the manifold space, and the proportion of overlapping areas of different feature types.

[0064] Step 4005: Perform distribution alignment processing on the feature point set in the manifold space. Adjust the tangent space coordinates of the feature points using the parallel transport algorithm to generate manifold space aligned features. Specifically, calculate the geometric center of the feature point set in the manifold space. For the coordinates of all feature points, use the weighted average rule of the SPD manifold, with the weight being the clinical importance of the sample corresponding to each feature point, as the reference benchmark for distribution alignment, equivalent to the origin of the manifold space. For each feature point in the manifold space, first determine the geodesic between the point and the global reference point. The shortest path between two points on the Riemannian manifold is obtained by solving the geodesic equation using the gradient descent method, thus obtaining the continuous path coordinates from the feature point to the reference point. Then, based on the geodesic, determine the transformation relationship between the tangent space of the feature point and the tangent space of the global reference point. This transformation relationship is the parallel transport operator.

[0065] For each feature point, extract its feature vector in its own tangent space, which contains the local geometric information of the point; input the feature vector into the corresponding parallel transport operator to convert it into a vector in the tangent space of the global reference point; based on the converted vector, fine-tune the coordinates of the feature point in the manifold space. If the converted vector of the feature point deviates from the center direction of the tangent space of the reference point, then fine-tune the feature point along the geodesic towards the reference point. The fine-tuning range is 10%-20% of the magnitude of the converted vector, ensuring that the adjusted feature point is not only close to the reference point in the global space, but also that its local geometric information is consistent with the tangent space of the global reference point.

[0066] Finally, the geodesic distances between all aligned feature points and the global reference point are calculated, and the percentage of feature points with a distance less than 0.3 is counted. If the percentage is ≥90%, and the proportion of feature points from different data sources overlapping in the manifold space increases to more than 60%, then the alignment is considered valid. If not, the smoothing coefficient of the parallel transport operator is readjusted, and the alignment is re-aligned. After successful alignment, all adjusted feature points are integrated to form the aligned feature set in the manifold space.

[0067] In this embodiment of the invention, the statistical essence and distribution patterns of features from various data sources are captured, providing a feature distribution specification for subsequent projection mapping and avoiding mapping direction deviations caused by ambiguity in the understanding of feature distribution. Through principal tangent space analysis, a preliminary channel to a unified Riemannian manifold is built for features from different sources, solving the problem of heterogeneous features lacking unified mapping rules and forming a feasible initial mapping framework. By leveraging Riemannian metric learning, the distribution gap of multi-source features in the manifold space is reduced, decreasing the discreteness of features from source A concentrated in region A and features from source B concentrated in region B, thus improving the spatial consistency of multi-source features. Optimized operators are used to complete the mapping of heterogeneous features and unify the actual mapping of the manifold, allowing time-series, image, and text features that originally belonged to different spaces to form directly comparable feature points on the Riemannian manifold, breaking the spatial isolation of multi-source features. By fine-tuning the tangent space coordinates of feature points through the parallel transport algorithm, the coordinate deviation of multi-source features in the local manifold is eliminated, further improving the fine alignment of multi-source feature distributions and ensuring that the aligned features are not only in the same space but also share the same distribution trend.

[0068] In a preferred embodiment of the present invention, step 404 includes: Step 4041: Based on the renormalized parameter set, calculate the distribution statistics of each parameter dimension to generate parameter distribution features. Specifically, according to the model architecture corresponding to the cardiovascular multi-source heterogeneous data, classify the renormalized parameter set by parameter type-data source: time series model parameters, image model parameters, and text model parameters. For each type of parameter, decompose it by dimension. For example, the weights of the LSTM hidden layer are a 256×256-dimensional matrix, which is decomposed into 65536 independent dimensions; the CNN convolutional kernel is 3×3×64×128-dimensional, which is decomposed into 73728 independent dimensions. Clarify the parameter meaning and data source of each dimension.

[0069] For each independent parameter dimension, its values ​​across all data sources are iterated, and the first moment (mean), second moment (variance), third moment (skewness), and fourth moment (kurtosis) are calculated. Simultaneously, considering the value range of cardiovascular medical parameters, the similarity between the parameter distribution and the gamma / Poisson distribution is verified using the QQ plot test and the KS test. The gamma distribution is adapted to the spatial feature weights of image parameters, and the Poisson distribution is adapted to the semantic association weights of text parameters, thus determining the distribution type for each dimension. A distribution feature card is constructed for each parameter dimension, containing four parts of information: first, basic identifiers such as parameter type, data source, dimension number, and parameter meaning; second, statistical values ​​such as the specific values ​​of mean, variance, skewness, and kurtosis, e.g., mean 0.74, variance 0.008, skewness 0.6, and kurtosis 3.8; third, distribution attributes such as distribution type and confidence level of the fit test; and fourth, clinical association, i.e., the importance of the medical features corresponding to this dimension parameter. All distribution feature cards are categorized and organized according to parameter type and data source to form a complete parameter distribution feature document.

[0070] Step 4042: Based on the parameter distribution characteristics, the maximum mean difference algorithm is used to measure the degree of difference between parameter distributions from different sources, generating a distribution difference measurement result. Specifically, it focuses on the differences between parameters of the same type across data sources, excluding invalid comparisons of parameters of different types; it divides comparison groups according to parameter type, namely, time series comparison group, image comparison group, and text comparison group, and splits the comparison units within each group according to data source pairs; then, considering the nonlinear distribution characteristics of cardiovascular parameters, the radial basis function (RBF) is selected as the kernel function of MMD, which can effectively capture the differences in nonlinear distribution; the kernel function bandwidth is set according to the variance-to-mean ratio of the parameter dimension to ensure that the sensitivity of the kernel function to the differences in parameter values ​​is adapted to the fluctuation range of medical data; a difference threshold is set based on the clinically permissible parameter deviation range, and if the threshold is exceeded, it is judged as a high difference dimension.

[0071] Then, for each comparison unit, the parameter values ​​of the same dimension from the two data sources are extracted, and the difference between the two distributions is calculated using the MMD algorithm. The parameter values ​​are then input into the RBF kernel function to calculate the mean mapping of the two samples in the Reproducing Kernel Hilbert Space (RKHS). The distance between the mean mappings is then calculated using the norm, which is the MMD value of that data source pair. For each parameter dimension, its MMD value in different data source pairs is calculated separately. Finally, the MMD values ​​are organized according to parameter type and data source pair to form a multi-source parameter distribution difference matrix. The rows and columns of the matrix represent the data sources, and the cells represent the average MMD value of the corresponding data source pair. A list of high-difference dimensions is attached next to the matrix, indicating the dimension number, parameter type, and corresponding data source pair whose MMD value exceeds the threshold. At the same time, the proportion of high-difference dimensions for each data source is calculated, i.e., the number of high-difference dimensions / the total number of dimensions.

[0072] Step 4043: Based on the distribution difference measurement results, the parameter distribution is corrected to make the parameter distributions from different sources more consistent, generating a parameter set after distribution alignment. Specifically, for high difference dimensions (MMD values ​​exceeding the threshold), the differences are handled in two categories according to their causes. The first is the difference caused by data imbalance, which is corrected by weighted migration: using the parameter distribution of the large sample data source as the benchmark, the correction weight is set according to the sample size ratio of the small sample data source, and the small sample parameters are finely adjusted to the benchmark distribution, such as the small sample parameter value = original value × 0.9 + benchmark distribution mean × 0.1. The second is the difference caused by data type differences, which is corrected by domain adaptive correction: a domain classifier is constructed to distinguish the data source to which the parameters belong, and the parameters are optimized through backpropagation so that the corrected parameters can retain the key weights of critical illness data, such as the high weight of coronary artery stenosis areas, and also adapt to the distribution trend of general hospitals.

[0073] Next, for each high-variance dimension, correction is performed according to the corresponding strategy. For example, the original value of the time-series parameter dimension 123 for hospital F was 0.61, and the baseline distribution mean was 0.74. After correction, the value is 0.61 × 0.9 + 0.74 × 0.1 = 0.623. After correction, the MMD value of this dimension in the corresponding data source pair is recalculated. If the MMD value drops below the threshold, the dimension is considered to have met the correction standard. If it does not meet the standard, the correction weight is adjusted, such as increasing the weight of the baseline distribution mean from 10% to 15%, and re-correction is performed until the MMD values ​​of all high-variance dimensions meet the standard. For low-variance dimensions (MMD values ​​not yet met), the correction is performed on the other side. (Exceeding the threshold), only the dispersion of parameter values ​​is fine-tuned; finally, all corrected parameter dimensions are integrated and reorganized according to the original renormalized parameter set structure (time series-image-text model hierarchy); at the same time, a correction log is generated, recording the pre-correction MMD value, correction strategy, correction weight, and post-correction MMD value of each high-discrepancy dimension, as well as the fine-tuning magnitude of low-discrepancy dimensions; through sampling verification, 10% of parameter dimensions are randomly selected, and their distribution mean difference in all data sources is calculated. If the mean difference decreases by ≥40% compared to before correction, the alignment is deemed effective, confirming the parameter distribution alignment effect and forming a distribution-aligned parameter set.

[0074] Step 4044: Perform smoothing optimization on the distributed aligned parameter set to generate a smoothed optimized parameter set. Specifically, traverse the distributed aligned parameter set, analyze the continuity of parameter values ​​according to parameter type-model level, and examine the changes in values ​​of adjacent dimensions for the LSTM weight matrix of the time series model. For example, if dimension 123 has a value of 0.623, dimension 124 has a value of 0.631, and dimension 125 has a value of 0.81, the difference between dimension 125 and its adjacent dimensions is 0.179, which far exceeds the average difference of 0.01. The following criteria are used to identify local anomalies: For CNN convolutional kernels in image models, examine the distribution of pixel weights within the same kernel. For example, in a 3×3 convolutional kernel, 8 weight values ​​are between 0.6 and 0.7, and 1 weight value is 0.92, which is considered an isolated anomaly. For BERT self-attention weights in text models, examine the fluctuation range of parameters within the same attention head. For example, if the variance of a parameter within a certain attention head is 0.02, which is much higher than the 0.005 of other attention heads, it is considered to have excessive fluctuation. Combining cardiovascular clinical knowledge, mark the dimensions of local anomalies and high fluctuation areas.

[0075] Next, for local outlier dimensions, neighborhood mean smoothing is used. For each outlier dimension, the average value of its five nearest neighbors is calculated, and the value of the outlier dimension is adjusted to 95% of this average. For example, if the neighborhood mean is 0.63, the outlier dimension is adjusted to 0.60. For high-fluctuation regions, Gaussian smoothing is used. A Gaussian window is constructed with each parameter dimension in the region as the center. The window size is set according to the number of dimensions in the region. The Gaussian weighted average of the parameters within the window is calculated, with the central dimension having the highest weight and decreasing towards the edges. The original parameter values ​​are replaced with the weighted average, which is consistent with the weight trend of the surrounding dimensions. For continuous parameter dimensions of the time series model, sliding window smoothing is used. The window size is 10, and the average value of the window is used to replace the value of the central dimension.

[0076] Then, an upper limit for the smoothing amplitude is set, with the adjustment amplitude of a single parameter dimension not exceeding 15% of the original value. For clinically critical parameter dimensions, the smoothing intensity is reduced individually, and the window weight is more significantly tilted towards the central dimension to ensure that the parameter characteristics that are crucial for disease diagnosis are still preserved after smoothing. After smoothing, it is checked whether the parameter values ​​are still within the reasonable medical range. Finally, the fluctuation coefficient of the smoothed parameter set and the standard deviation / mean of the values ​​of each parameter dimension are calculated. Compared with before smoothing, if the fluctuation coefficient decreases by ≥30%, the smoothing is considered effective. At the same time, the model performance is tested using a small batch of cardiovascular validation data. If the prediction error of the smoothed model does not decrease compared with before smoothing but the stability is improved, and the error fluctuation amplitude decreases by ≥25% in multiple tests, the smoothing optimization is confirmed to have met the standards. The smoothed parameters are integrated to form a smoothed and optimized parameter set, along with a smoothing report.

[0077] Step 4045: The smoothed and optimized parameter set is quality verified. The distribution consistency effect is confirmed by calculating parameter stability indices, ultimately generating an optimized global parameter set. Specifically, based on the clinical application needs of the cardiovascular model, three types of indices are set: First, the distribution consistency index: calculating the MMD value compliance rate and parameter distribution mean difference for the same parameter dimension across all data sources. The MMD value compliance rate is the number of dimensions with MMD values ​​below a threshold divided by the total number of dimensions, requiring a compliance rate ≥ 90%. The parameter distribution mean difference is the difference between the means of parameters of the same dimension from different data sources. The evaluation criteria are as follows: 1) Mean difference ≤ 0.05; 2) Parameter stability indicators: Calculate parameter fluctuation coefficient and parameter recalculation consistency. The parameter fluctuation coefficient is the standard deviation / mean of the parameter values ​​in each dimension, and the fluctuation coefficient should be ≤ 0.1. The parameter recalculation consistency refers to the repeated calculation of statistics for the same parameter dimension, with a result deviation ≤ 0.001; 3) Model performance indicators: Load the smoothed parameter set into the cardiovascular risk assessment model and test the model performance using full multi-source test data. The model classification accuracy should be ≥ 85%, prediction error ≤ 0.08, and performance difference between test sets from different data sources ≤ 5%.

[0078] First, iterate through all parameter dimensions, calculate the MMD value compliance rate, and calculate the mean difference between different data sources. If the compliance rate is ≥90% and the mean difference is ≤0.05, then the model passes the validation for this type of indicator. Next, calculate the global parameter fluctuation coefficient, and repeat the statistical calculation for 10% of the parameter dimensions. If the fluctuation coefficient is ≤0.1 and the recalculation deviation is ≤0.001, then the model passes the validation for this type of indicator. Finally, validate the model performance indicators by loading the parameter set into the model three times, testing with the full test data each time, and recording the accuracy and error each time. If the average accuracy of the three tests is higher than the target value, then the model passes the validation for this type of indicator. If the performance is ≥85%, the average error is ≤0.08, and the single performance difference is ≤5%, then the validation of this type of indicator is passed. If a certain indicator fails to meet the standard, the process is backtracked to step 4043 to recalibrate the high-difference dimension, or to step 4044 to smooth the fluctuation area until all indicators meet the standard. Finally, the smoothed optimization parameters that have passed all quality validations are integrated, and the final parameter file is generated according to the structure of model architecture-parameter type-data source adaptation annotation. At the same time, a global parameter quality report is generated, and the parameter file and the quality report are packaged to form the optimized global parameter set.

[0079] In this embodiment of the invention, the current distribution status of each dimension of the renormalized parameters is understood, and the starting point of the distribution difference of different source parameters is identified, providing a distribution benchmark for subsequent difference measurement and correction. The maximum mean difference algorithm is used to adapt to the nonlinear parameter distribution, quantitatively characterize the degree of difference in the distribution of different source parameters, and accurately locate the dimension of highly skewed parameters. The distribution gap between different source parameters is narrowed in a targeted manner, making parameters with large deviations tend to be consistent, improving the adaptability of parameters to the full dataset. Local outliers or abrupt fluctuations that may exist in the parameters after distribution alignment are eliminated, reducing parameter oscillations during model training and improving the model's convergence stability and the consistency of prediction results. The optimization effect is verified through parameter stability indicators to ensure that the optimized parameters truly achieve consistent and stable distribution.

[0080] In a preferred embodiment of the present invention, step 5 above includes: Step 500: Load the optimized global parameter set into the global risk assessment network in the parameter server to complete the parameter update and initialization of the global model, and obtain the initialized global risk assessment network. Specifically, first, clarify the core module structure of the attention-based multilayer perceptron (Attention-MLP) risk assessment network, including the input layer, embedding layer, layer normalization module, attention calculation layer, multilayer perceptron (MLP) module (containing 3-5 fully connected layers), dropout regularization module, and output layer. The parameter dimensions and connection relationships of each module need to match the optimized global parameter set. For example, the output dimension of the fully connected layer of the embedding layer is 256-dimensional, the weight matrix dimension of the attention layer is 256×256, and the output dimensions of each layer of the MLP are 128-dimensional, 64-dimensional, and 32-dimensional, respectively.

[0081] Next, from the optimized global parameter set, parameters are split into modules: embedding layer parameters, layer normalization parameters, attention calculation layer parameters, MLP module parameters, and output layer parameters. Through a dedicated interface of the parameter server, each module parameter is loaded into the corresponding layer of the global risk assessment network. During loading, the parameter dimension matching is verified in real time; if the input temporal feature is 100-dimensional, the weight dimension must be 100×256. After loading, 10 sets of cardiovascular multi-source feature samples with known risk levels are input, covering temporal, image, and text types. A forward computation is performed to check if the outputs of each network module are normal. For example, the embedding layer outputs a 256-dimensional vector, the attention layer outputs reasonable weights, the MLP outputs 32-dimensional deep features, and the output layer generates risk scores from 0 to 100, with no gradient overflow or parameter outliers during the computation. If the forward computation is normal and the output score matches the known risk level by ≥80%, the global risk assessment network is considered successfully initialized, resulting in an initialized network usable for evaluation.

[0082] Step 501: Utilizing the initialized global risk assessment network, the input layer receives multi-source characteristic data of cardiovascular diseases to be assessed from different sources. The input data undergoes format standardization and missing value processing to generate standardized input features. Specifically, the input layer adapts to different data formats through multiple interfaces, receiving time-series monitoring data from the hospital's HIS system, image data from the PACS system, and text data from the electronic medical record system. A unique identifier is added to each data set based on patient ID and examination time to ensure a one-to-one correspondence between data and patients. For time-series data, a dynamic time warping algorithm is used to warp data of different lengths... Columns are uniformly adjusted to a fixed length. Columns shorter than the required length are padded with the last sampling point. Columns exceeding the required length are truncated to key time periods, such as retaining the middle 2-hour segment of an electrocardiogram where heart rate fluctuations are significant. For image data, images of different resolutions are uniformly scaled to 224×224 pixels using bilinear interpolation, converted to grayscale images, and then flattened into a 50176-dimensional vector. For text data, a medical-specific word segmentation tool is used to segment medical record texts. After filtering out stop words, the words are converted into indices from a pre-trained vocabulary and uniformly adjusted to a fixed-length sequence of 512 indices. Columns shorter than the required length are padded with 0s, and columns exceeding the required length are truncated to diagnostically relevant segments.

[0083] The missing data fields are statistically analyzed. Numerical missing data is filled with the mean value of the corresponding features of patients of the same age and disease. For example, missing ECG sampling points of a 60-year-old coronary heart disease patient are filled with the average heart rate value of this group. Textual missing data is filled with the corresponding word index that is not recorded. All processed data are integrated according to the structure of time-series feature vectors, image feature vectors and text feature index sequences to form standardized input features with uniform format and no key missing data, ensuring that they can be directly input into the embedding layer.

[0084] Step 502: The standardized input features are input into the pre-initialized embedding layer. Each feature is mapped to a high-dimensional vector representation through fully connected operations corresponding to each feature type within the embedding layer, generating an initial high-dimensional feature vector set. Specifically, for time-series feature vectors, such as the flattened vector of a 10,000-dimensional electrocardiogram sequence, a time-specific fully connected layer is used. The input dimension is 10,000, and the output dimension is 256. The weights and biases of this layer are derived from the time-series embedding parameters in the optimized global parameter set. Through linear operations (feature vector × weight + bias) and the ReLU activation function, the low-dimensional time-series features are transformed into a 256-dimensional high-dimensional vector containing dynamic patterns. The image feature vector, a 50176-dimensional flattened vector, is fed into an image-specific fully connected layer with an input dimension of 50176 and an output dimension of 256. Utilizing the image embedding parameters in the global parameters, the spatial key features of the image are extracted and compressed into a 256-dimensional vector through linear transformation and activation. For the text feature index sequence (512-dimensional index), the index is first converted into a 512×128-dimensional word vector using a pre-trained medical word embedding matrix. Then, it is fed into a text-specific fully connected layer with an input dimension of 512×128 and an output dimension of 256. After global average pooling to compress the word vector to 128 dimensions, it is converted into a 256-dimensional high-dimensional text vector through the fully connected layer, preserving semantic association information.

[0085] Ensure that the outputs of the dedicated fully connected layers for time series, image, and text are all 256-dimensional floating-point vectors, with vector values ​​controlled between -1 and 1, using ReLU activation and parameter initialization constraints. For each patient's three types of high-dimensional vectors, combine them in the order of time series vector - image vector - text vector to form the patient's initial high-dimensional feature vector set, containing three 256-dimensional vectors. Randomly sample 10% of the vector set to check whether the vector dimensions are consistent and whether there are any abnormal values. If the abnormality rate is less than 1%, the initial high-dimensional feature vector set is considered to be valid.

[0086] Step 503: Input the initial high-dimensional feature vector set into the layer normalization module to perform scale unification processing based on layer normalization, eliminate the dimensional differences between different feature vectors, and generate scale-normalized high-dimensional feature vectors. Specifically, for each 256-dimensional vector in the initial high-dimensional feature vector set, layer normalization is performed according to the feature dimension, i.e., 256 dimensions, and the mean and variance of each vector in the 256 dimensions are calculated. The scaling factor (γ) and offset factor (β) of the layer normalization come from the optimized global parameter set, and the initial values ​​are set to 1 and 0, respectively. For each 256-dimensional high-dimensional vector, the values ​​of each dimension are first subtracted. The mean of the vector is divided by (variance + minimum value ε, ε = 1e-5, to avoid a denominator of 0) to obtain a standardized intermediate vector. The intermediate vector is multiplied by the scaling factor γ and then the offset factor β is added to obtain a scale-normalized high-dimensional feature vector. The mean (should be close to 0, deviation ≤ 0.1) and variance (should be close to 1, deviation ≤ 0.2) of all scale-normalized vectors are calculated. If the compliance rate is ≥ 95%, the scale is considered to be uniform and effective. For vectors that do not meet the standard, the normalization calculation process is re-examined, the scaling factor and offset factor are corrected, and the vectors are normalized again to finally generate scale-normalized high-dimensional feature vectors for all patients.

[0087] Step 504: The scale-normalized high-dimensional feature vectors are input into the attention calculation layer. The original attention weight matrix is ​​generated by calculating the similarity between the attention query vector of each feature vector and the attention key vectors of all feature vectors. Specifically, the scale-normalized high-dimensional feature vectors of each patient (three 256-dimensional vectors, corresponding to time series, image, and text) are input into the linear transformation module of the attention calculation layer. This module contains three independent fully connected layers (Q generation layer, K generation layer, and V generation layer), and the parameters are all from the optimized global parameter set. The Q generation layer converts the 256-dimensional vectors into 256-dimensional query vectors for actively matching other features. The K generation layer converts them into 256-dimensional key vectors for matching the query vectors. The V generation layer converts them into 256-dimensional value vectors for the feature carrier of subsequent weighted fusion. Each patient finally obtains three sets of Q, K, and V vectors (time series Q / K / V, image Q / K / V, and text Q / K / V).

[0088] The dot product similarity calculation method is used. For each query vector Q, the dot product is calculated with all key vectors K (including its own corresponding K). That is, Q is multiplied by the corresponding dimension values ​​of K and then summed to obtain the similarity score. In order to avoid the similarity score being too large and causing overflow in the subsequent softmax calculation, all dot product results are divided by the square root of 256 to obtain the scaled similarity score. The scaled similarity scores are then organized according to the correspondence between query vector and key vector to form the original weight matrix, such as a 3×3 matrix, where the rows represent the Q vector type, the columns represent the K vector type, and the cell values ​​are the similarity scores.

[0089] Step 505: Based on the original attention weight matrix distribution, a probability distribution transformation is performed to generate normalized attention weights that conform to the probability distribution characteristics. Specifically, each row of the original attention weight matrix is ​​taken as an independent input, and the probability distribution of that row is calculated using the softmax function. For each similarity score in the row, the maximum value of the row is subtracted first, then the exponent value is calculated, and finally each exponent value is divided by the sum of all exponent values ​​in the row to obtain the normalized weight of the row. For example, if the original weight of a row is [0.8, 0.5, 0.3], with a maximum value of 0.8, after subtraction, it becomes [0, -0.3, -0.5], the exponent value is approximately [1, 0.74, 0.61], the sum is approximately 2.35, and the normalized weight is approximately [0.425, 0.315, 0.26], ensuring that the sum of the weights in each row is 1, which conforms to the probability distribution characteristics.

[0090] Check that the row sums of all normalized weight matrices are all 1, allowing a calculation error of ±0.001. If there is a deviation in the row sums, readjust the numerical precision of the softmax calculation and convert again. At the same time, verify the weight allocation in conjunction with common sense in cardiovascular clinical practice. Label the verified weight matrices according to feature type, such as time series weight -0.425, image weight -0.315, text weight -0.26. Each patient corresponds to a set of normalized weights, which clearly reflects the preliminary importance ratio of each multi-source feature in risk assessment.

[0091] Step 506: The attention value vectors of all feature vectors are weighted and summed using the normalized attention weights to fuse feature information from different sources and types, generating an attention-weighted fused feature representation. Specifically, the normalized attention weight matrix for each patient is matched with the corresponding V vector key vector type. The columns of the weight matrix correspond to the K vector type, and the K vector type corresponds one-to-one with the V vector type, such as image K corresponding to image V. Therefore, each column of the weight matrix can be directly associated with a V vector of the same type; for example, normalized... The weights [0.425, 0.315, 0.26] are associated with time series V, image V, and text V, respectively. For each patient, the V vector of each feature type is multiplied by the corresponding normalized weight (each dimension of the vector is multiplied by the weight) to obtain a weighted V vector, such as time series V×0.425, image V×0.315, and text V×0.26. The three weighted V vectors are added one by one according to their dimensions to finally generate a 256-dimensional attention-weighted fusion feature representation, which integrates the key information of time series, image, and text.

[0092] By comparing the amount of feature information before and after fusion, the cosine similarity between the fused vector and each original V vector is calculated. If the similarity between the fused vector and the image V is ≥0.7 (image is the core feature), the similarity with the temporal V is ≥0.6, and the similarity with the text V is ≥0.5, then the fusion is considered effective.

[0093] Step 507: The attention-weighted fused feature representation is input into the multilayer perceptron module. Layer-by-layer nonlinear transformation is performed through multiple fully connected layers and nonlinear activation functions in the multilayer perceptron to extract deep hidden features. Specifically, the MLP module contains 3 fully connected layers and 2 ReLU activation functions. The input and output dimensions and parameters of each layer come from the optimized global parameter set. The first layer (input layer) receives 256-dimensional fused features and outputs 128-dimensional features, with a weight dimension of 256×128 and a bias dimension of 128. The second layer (hidden layer) receives 128-dimensional features and outputs 64-dimensional features, with a weight dimension of 128×64 and a bias dimension of 64. The third layer (output layer) receives 64-dimensional features and outputs 32-dimensional features, with a weight dimension of 64×32 and a bias dimension of 32. The ReLU activation function is used after the outputs of the first and second layers to introduce nonlinear transformation capabilities, avoiding the limitation of linear models in capturing complex risk associations.

[0094] The attention-weighted fusion features (256 dimensions) are input into the first layer of the MLP. A linear operation is performed, which multiplies the feature vector by the first-layer weights and then the first-layer bias. After passing through the ReLU activation function (setting negative numbers to 0 and retaining positive features), a 128-dimensional first-layer output feature is obtained. This feature initially extracts shallow associations among multi-source features, such as the basic risk association between hypertension and hyperglycemia. The 128-dimensional feature is input into the second layer, and the linear operation and ReLU activation are repeated to obtain a 64-dimensional second-layer output feature, which further explores intermediate associations, such as the synergistic risk association between hypertension, hyperglycemia, and obesity. The 64-dimensional feature is input into the third layer, where only linear operations are performed without activation, preserving the continuity of deep features, resulting in a 32-dimensional deep latent feature. This feature contains complex deep associations among multi-source features, such as the comprehensive risk association between hypertension, hyperglycemia, coronary artery stenosis, and smoking history.

[0095] The numerical distribution of the 32-dimensional deep features is statistically analyzed to ensure that there are no vectors with all zeros. If such vectors exist, it indicates that the previous layer feature extraction has failed, and the MLP parameters need to be re-examined, with the numerical range controlled between -2 and 2. The deep features are reduced to 2D space using visualization tools (such as t-SNE), and the feature clustering of patients with different risk levels is observed. High-risk patients should form independent clusters. If the clustering effect is obvious, the deep hidden feature extraction is considered effective.

[0096] Step 508: Regularize the deep latent features output by the multilayer perceptron using Dropout technology to reduce the risk of overfitting and generate regularized deep features. Specifically, the inactivation probability of the Dropout module is set to 0.3, determined based on the amount of medical data and model complexity. This avoids excessive inactivation leading to feature loss and effectively suppresses overfitting. Random inactivation is only enabled during the model training phase. During the evaluation phase, an inference mode is used, meaning that neurons are not randomly inactivated. Instead, the 32-dimensional deep feature vector output by the third layer of the MLP is multiplied by (1 - inactivation probability) = 0.7 to compensate for the impact of neuron inactivation during training on the feature amplitude, ensuring that the feature amplitude during the evaluation phase is consistent with that during the training phase.

[0097] Multiply each dimension of the 32-dimensional deep feature vector output by MLP by 0.7 to obtain the regularized deep feature vector. This process reduces the model's over-reliance on individual neurons in the MLP and lowers the risk of overfitting caused by local noise in the training data. Compare the feature vectors before and after regularization to ensure that the dimension remains unchanged (still 32-dimensional) and the numerical amplitude is reasonable. Validate using a test set. Input the features before and after regularization into the output layer respectively. If the error fluctuation of the risk assessment result after regularization is reduced by ≥20% on different hospital test sets, the regularization is deemed effective, and the regularized deep feature is generated.

[0098] Step 509: The regularized deep features are input into the output layer. A linear transformation of the output layer maps the features to the final cardiovascular disease risk prediction score, generating a quantitative cardiovascular disease risk assessment result. Specifically, the output layer is a single-output linear layer. The input is a 32-dimensional regularized deep feature, and the output is a single risk prediction score. The linear transformation has a weight dimension of 32×1 and a bias dimension of 1. All parameters are derived from the optimized global parameter set. This layer has no activation function, ensuring that the output score covers a reasonable risk range. The 32-dimensional regularized deep features are input into the output layer, and the original prediction score is obtained through a linear operation of feature vector × output layer weight + output layer bias. Combining this with cardiovascular clinical risk assessment standards, the original prediction score is normalized to 0-10 using a maximum-minimum normalization method. The risk score range is 0, and the mapping formula references the risk score range of historical clinical data to ensure that the mapped score matches the actual clinical risk level. For example, 0-20 is low risk, 21-50 is medium risk, and 51-100 is high risk. A structured assessment report is generated for each patient, containing three core pieces of information: first, the quantitative risk score and corresponding risk level; second, key feature basis, based on attention weight, highlighting the top 3 contributing features, such as the degree of coronary artery stenosis (weight 0.315), 24-hour heart rate variability (weight 0.425), and history of diabetes (weight 0.26); and third, clinical recommendations, such as further coronary angiography for high risk and lifestyle intervention for medium risk. The report format is compatible with the hospital's electronic medical record system and can be directly embedded into the patient's medical record.

[0099] In this embodiment of the invention, the optimized global parameter set is loaded into the global risk assessment network in the parameter server to avoid training inefficiency or model bias caused by initial parameter deviation; the format of multi-source data to be evaluated is unified, missing values ​​are filled in, data noise and format chaos are eliminated, and the consistency of evaluation data is improved; low-dimensional / heterogeneous features are mapped to high-dimensional vectors through fully connected operations, fully preserving the subtle differences of each feature and providing a rich information carrier for multi-source feature fusion; the dimensional differences of different high-dimensional feature vectors are eliminated to ensure that each feature contributes fairly in the attention layer and MLP; key features are automatically identified through feature vector similarity, solving the problems of traditional ML. The model addresses the issue of key information being diluted due to the equal treatment of all features; it transforms the original weights into weights that conform to a probability distribution, ensuring that important features receive a reasonable high proportion while secondary features have a lower proportion, thus ensuring more accurate weighting logic; it breaks down the integration barriers of multi-source heterogeneous features, allowing fused features to contain comprehensive information from time series, images, and text; it mines complex correlations in fused features through multi-layer nonlinear transformations to accurately capture deep patterns of cardiovascular disease risk; it uses Dropout regularization to reduce the model's over-reliance on noisy or redundant features in the training data, avoiding overfitting; and it provides doctors with precise decision-making basis by using specific risk scores rather than qualitative high / low risk.

[0100] like Figure 2 As shown, embodiments of the present invention also provide a federated learning optimization system for multi-source heterogeneous cardiovascular disease risk prediction, comprising: The standardization module is used to process raw cardiovascular disease data to obtain standardized feature data; The computation module is used to perform local computation based on standardized feature data and a differentiated training algorithm corresponding to its data structure to generate updated encrypted local parameters. An integration module is used to transmit updated encrypted local parameters to a parameter server within the federated learning framework. The parameter server uses a geometric mean algorithm based on secure multi-party computation to securely aggregate and geometrically average local parameter updates from different sources in an encrypted state, generating a preliminary global parameter set. The coordination and optimization module is used to map multi-source heterogeneous features to a unified geometric manifold space based on the initial global parameter set and a feature space geometric projection algorithm based on manifold learning. In this space, feature distribution coordination and model weight redistribution are performed to generate an optimized global parameter set. The risk assessment module loads the optimized global parameter set into a multilayer perceptron risk assessment network based on an attention mechanism. The input layer of this network receives multi-source feature data to be assessed, the embedding layer converts the feature data into high-dimensional vectors, the attention calculation layer performs importance weighting and fusion on feature vectors from different sources and types, the multilayer perceptron module performs nonlinear transformation and deep feature extraction on the weighted and fused features, and finally the output layer generates a quantitative cardiovascular disease risk assessment result based on comprehensive multi-source data.

[0101] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A federated learning optimization method for predicting multi-source heterogeneous cardiovascular disease risk, characterized in that, The method includes: Step 1: Process the raw cardiovascular disease data to obtain standardized feature data; Step 2: Based on standardized feature data, perform local computation using a differentiated training algorithm corresponding to its data structure to generate updated encrypted local parameters; Step 3: The updated encrypted local parameters are transmitted to the parameter server within the federated learning framework. The parameter server uses a geometric mean algorithm based on secure multi-party computation to securely aggregate and integrate the local parameter updates from different sources under encrypted conditions, generating a preliminary global parameter set. Step 4: Based on the preliminary global parameter set, the multi-source heterogeneous features are mapped to a unified geometric manifold space through a feature space geometric projection algorithm based on manifold learning. Feature distribution coordination and model weight redistribution are performed in this space to generate an optimized global parameter set. Step 5: The optimized global parameter set is loaded into the multilayer perceptron risk assessment network based on the attention mechanism. The input layer of the network receives the multi-source feature data to be evaluated. The embedding layer converts the feature data into high-dimensional vectors. The attention calculation layer performs importance weighting and fusion on the feature vectors from different sources and types. Then, the multilayer perceptron module performs nonlinear transformation and deep feature extraction on the weighted and fused features. Finally, the output layer generates a quantitative cardiovascular disease risk assessment result based on comprehensive multi-source data.

2. The federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction according to claim 1, characterized in that, Step 2 includes: The corresponding training algorithm is selected based on the data structure characteristics of the standardized feature data. The temporal feature data uses a temporal feature extraction algorithm, the image feature data uses a spatial feature extraction algorithm, and the text feature data uses a semantic feature extraction algorithm. Standardized feature data is used as input and loaded into the corresponding training algorithm for forward computation to generate intermediate computation results for each data type. Based on the intermediate calculation results of various data types, the difference between the predicted output and the real label is calculated using the error backpropagation mechanism. The parameter configuration in the forward calculation process is adjusted through the gradient optimization algorithm to generate the local parameter update amount. Based on the pre-negotiated encryption security level, a Paillier encryption scheme that satisfies the additive homomorphic property is selected and a public-private key pair is generated. The public key is distributed to each medical institution node. Each numerical parameter in the local parameter update is converted into a floating-point representation that meets the input requirements of the encryption scheme. Each converted numerical parameter is encrypted in parallel using the public key pair to generate the corresponding homomorphic ciphertext. The ciphertexts of all parameters are combined according to the original parameter structure, and integrity verification information is added to generate updated encrypted local parameters.

3. The federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction according to claim 3, characterized in that, Step 3 includes: The parameter server receives updated encrypted local parameters from each node and performs source authentication based on digital signature and integrity verification based on hash value on each received encrypted update packet to obtain a set of verified encrypted parameter updates. Based on the verified set of encrypted parameter updates, the metadata information contained in each encrypted update package is parsed to identify the data feature type and model structure type corresponding to each parameter update. Based on the identified data feature types and model structure types, an unsupervised clustering algorithm based on ciphertext similarity measurement is used to group the encrypted parameter updates, generating multiple parameter update groups based on type similarity. For each parameter update group, secure multi-party geometric mean calculation is performed in encrypted state, where the weight of each node is dynamically determined based on the data quality index and the number of valid samples of that node, generating encrypted aggregation results for each group; Based on the encrypted aggregation results of each group, the contribution weight of the aggregation results of different groups to the global model is calculated, and the results of each group are weighted and fused according to the weight to generate a preliminary global parameter set.

4. The federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction according to claim 4, characterized in that, Step 4 includes: Based on the preliminary global parameter set, the Riemann manifold projection algorithm is used to map features from different data sources to a unified Riemann manifold space, generating manifold space aligned features. Based on the features aligned to the manifold space, the geometric difference between different feature distributions is calculated by geodesic distance to generate a multi-source feature distribution difference measurement result. Based on the multi-source feature distribution difference measurement results, the weight diffusion algorithm is used to calculate the importance weight of each feature dimension in the manifold space, and generate the geometric feature importance weight distribution; Based on the weighted distribution of the importance of the geometric features, the preliminary global parameter set is subjected to a parameter renormalization operation based on the manifold geometry to generate a renormalized parameter set. The renormalized parameter set is subjected to distribution consistency optimization to generate an optimized global parameter set.

5. The federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction according to claim 5, characterized in that, Based on the preliminary global parameter set, the Riemannian manifold projection algorithm is used to map features from different data sources to a unified Riemannian manifold space, generating manifold-space aligned features, including: Extract feature representations from each data source in the preliminary global parameter set, calculate the statistical moments and distribution characteristics of each feature representation, and generate feature distribution descriptors; Based on the feature distribution descriptor, the principal tangent space analysis method is used to construct the projection mapping relationship from the features of each data source to the common Riemannian manifold, and an initial projection operator set is generated. The initial projection operator set is optimized by using the Riemann metric learning algorithm to minimize the distribution divergence of different source features in the manifold space, thereby generating optimized projection operators. The optimized projection operator is used to map the feature representations of each data source from the original feature space to a unified Riemannian manifold space, generating a set of feature points in the manifold space. The feature point set in the manifold space is distributed and aligned. The tangent space coordinates of the feature points are adjusted by the parallel transport algorithm to generate the manifold space aligned features.

6. The federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction according to claim 6, characterized in that, The renormalized parameter set is subjected to distribution consistency optimization to generate an optimized global parameter set, including: Based on the renormalized parameter set, the distribution statistics of each parameter dimension are calculated to generate parameter distribution characteristics; Based on the parameter distribution characteristics, the maximum mean difference algorithm is used to measure the degree of difference between different source parameter distributions, and a distribution difference measurement result is generated. Based on the distribution difference measurement results, the parameter distribution is corrected to make the distributions of different source parameters tend to be consistent, and a parameter set after distribution alignment is generated. The parameter set after distribution alignment is smoothed and optimized to generate a smoothed and optimized parameter set; The quality of the smoothed and optimized parameter set is verified, and the distribution consistency effect is confirmed by calculating the parameter stability index. Finally, the optimized global parameter set is generated.

7. The federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction according to claim 7, characterized in that, Step 5 includes: The optimized global parameter set is loaded into the global risk assessment network in the parameter server to complete the parameter update and initialization of the global model, and the initialized global risk assessment network is obtained. Using the initialized global risk assessment network, the input layer receives multi-source feature data of cardiovascular diseases to be assessed from different sources, performs format unification and missing value processing on the input data, and generates standardized input features. The standardized input features are input into the embedded layer that has been initialized with parameters. Each feature is mapped to a high-dimensional vector representation through the fully connected operation corresponding to each feature type in the embedded layer, generating an initial high-dimensional feature vector set. The initial high-dimensional feature vector set is input into the layer normalization module to perform scale unification processing based on layer normalization, thereby eliminating the dimensional differences between different feature vectors and generating scale-normalized high-dimensional feature vectors.

8. The federated learning optimization method for multi-source heterogeneous cardiovascular disease risk prediction according to claim 8, characterized in that, Step 5 also includes: The scale-normalized high-dimensional feature vectors are input into the attention calculation layer. The original attention weight matrix is ​​generated by calculating the similarity between the attention query vector of each feature vector and the attention key vector of all feature vectors. Based on the original attention weight matrix distribution, a probability distribution transformation is performed to generate normalized attention weights that conform to the characteristics of the probability distribution. The attention value vectors of all feature vectors are weighted and summed using the normalized attention weights to fuse feature information from different sources and types, generating a feature representation after attention weighted fusion. The attention-weighted fusion feature representation is input into the multilayer perceptron module, and deep hidden features are extracted by performing layer-by-layer nonlinear transformation through multiple fully connected layers and nonlinear activation functions in the multilayer perceptron. The deep latent features output by the multilayer perceptron are regularized using Dropout technology to reduce the risk of overfitting and generate regularized deep features. The regularized deep features are input into the output layer, and the features are mapped to the final cardiovascular disease risk prediction score through the linear transformation of the output layer, generating a quantitative cardiovascular disease risk assessment result.

9. A federated learning optimization system for predicting multi-source heterogeneous cardiovascular disease risk, the system implementing the method as described in any one of claims 1 to 9, characterized in that, include: The standardization module is used to process raw cardiovascular disease data to obtain standardized feature data; The computation module is used to perform local computation based on standardized feature data and a differentiated training algorithm corresponding to its data structure to generate updated encrypted local parameters. An integration module is used to transmit updated encrypted local parameters to a parameter server within the federated learning framework. The parameter server uses a geometric mean algorithm based on secure multi-party computation to securely aggregate and geometrically average local parameter updates from different sources in an encrypted state, generating a preliminary global parameter set. The coordination and optimization module is used to map multi-source heterogeneous features to a unified geometric manifold space based on the initial global parameter set and a feature space geometric projection algorithm based on manifold learning. In this space, feature distribution coordination and model weight redistribution are performed to generate an optimized global parameter set. The risk assessment module loads the optimized global parameter set into a multilayer perceptron risk assessment network based on an attention mechanism. The input layer of this network receives multi-source feature data to be assessed, the embedding layer converts the feature data into high-dimensional vectors, the attention calculation layer performs importance weighting and fusion on feature vectors from different sources and types, the multilayer perceptron module performs nonlinear transformation and deep feature extraction on the weighted and fused features, and finally the output layer generates a quantitative cardiovascular disease risk assessment result based on comprehensive multi-source data.

Images