A privacy-preserving cross-hospital disease data analysis method

Abstract

The present invention relates to a privacy-protected cross-hospital disease data analysis method, which is as follows: Step 1: Acquire multi-site brain imaging data sets, Step 2: Construct a brain network based on a resting-state functional connectivity matrix, Step 3: Use contrastive learning within a site to extract the unique features of each site, Step 4: Use contrastive learning between sites to extract common features of different sites, Step 5: Minimize the mutual information of features extracted by a dual-channel feature extractor, Step 6: Use the extracted features to classify samples, Step 7: Construct a topological relationship diagram between multiple sites, Step 8: Aggregate a global model based on contribution, Step 9: The server transmits global model parameters to each site to replace local model parameters. This solution effectively alleviates the impact of cross-site non-independent and identically distributed medical data on model classification effects in federated learning, achieves optimal aggregation, protects sensitive data, and makes in-depth optimization on the heterogeneity of cross-site data.

Description

Technical Field

[0001] The present invention relates to a data analysis method, and in particular to a privacy-protected cross-hospital disease data analysis method, belonging to the technical field of data analysis. Background Art

[0002] In the healthcare sector, the high cost of collecting and annotating patient data and the limited data at a single site result in low diagnostic accuracy and versatility of deep models. The multi-site approach pools data to solve the "data island" problem and improve diagnostic accuracy and model versatility. However, data sharing may lead to privacy leakage, which poses challenges to multi-site joint training. Federated learning algorithms can protect patient privacy by only transmitting model parameters, but client drift occurs when processing heterogeneous data, the versatility and classification effect of the model cannot be effectively improved, and the global model cannot be effectively aggregated.

[0003] Compared with single-site methods, multi-site analysis methods can improve the performance of medical diagnosis models by integrating samples or knowledge collected from different hospitals or devices. However, the data sharing between different sites also poses a huge challenge to the privacy protection of medical data. With the increasing public concern about personal privacy data and the implementation of stricter privacy protection laws, there is an urgent need to develop accurate multi-site diagnostic models that protect medical privacy.

[0004] Federated learning enables collaborative model training between different sites while protecting user privacy. This is achieved by training the model on the local device without transmitting the original data to the central server. As the most widely adopted federated learning method, FedAvg uses local data to train the model and transmits the model parameters to the central server. The server aggregates the model parameters, calculates the average of the global model, and then sends the results back to the local device to update the model. Federated learning has received increasing attention in the field of medical data processing. For example, the Federated Conditional Mutual Learning (FedCM) method improves the diagnostic performance of Alzheimer's disease by considering the local performance of the site and the similarity between sites. The Federated Forest (FeDeFo) model calculates the hippocampal volume for AD classification by using sMRI images with data privacy protection. Although a federated learning method is disclosed in the prior art, the method implements a decentralized iterative optimization algorithm and changes the shared local model weights through a randomization mechanism. However, the above method is performed under the assumption that the data from different sites are independent and identically distributed (IID). However, in most multi-site medical data analysis problems, non-IID data is prevalent, which usually leads to client drift. Therefore, there is an urgent need to study robust federated multi-site methods that can enable aggregate models to bridge the gap caused by differences in data distribution among sites, thereby improving their predictive performance.

[0005] Recently, many studies have been devoted to alleviating the impact of non-IID problems between different sites on federated learning methods. For example, FedProx incorporates proximal terms to limit the model parameters obtained after local training to be not too far away from the global model parameters. SCAFFOLD introduces server and site control variables with model update direction information, aiming to alleviate the client drift problem by adding correction terms to the update equation of the local model. The Moon method aligns the representations of the global and local models in an optimized feature space. Although the above methods consider both global and local semantic alignment, they ignore the enhancement of feature distinguishability during the training of private models and the aggregation of global models, making it difficult to achieve an optimal federated diagnostic model.

[0006] As for federated aggregation, the typical federated algorithm FedAvg averages the aggregated model parameters according to the sample proportion. However, in the non-IID case, effective aggregation is often not achieved. Many research methods attempt to use key information reflecting the data differences between sites for model aggregation. For example, the model aggregation strategy based on the Hessian matrix takes into account the quality variation and assigns weights to each parameter of the local site model according to the model convergence speed.

[0007] To address these issues, we propose a federated learning approach with contrastive feature enhancement for robust multi-site brain disease diagnosis. This approach effectively alleviates the client drift problem caused by heterogeneous data while enhancing privacy protection across different medical sites or hospitals. Summary of the invention

[0008] The present invention is aimed at the technical problems existing in the prior art and provides a privacy-preserving cross-hospital disease data analysis method. This method only needs to transmit model parameters and gradient information between sites without transmitting patient information, thereby effectively protecting patient privacy. In addition, through comparative learning within and between sites and a federated aggregation strategy based on multi-site topological relationships, the impact of cross-site non-independent and identically distributed medical data on model classification effects in federated learning is effectively alleviated, and optimal aggregation is achieved. This solution not only protects sensitive data, but also makes in-depth optimization on the heterogeneity of cross-site data.

[0009] In order to achieve the above object, the technical solution of the present invention is as follows: a privacy-preserving cross-hospital disease data analysis method, the method comprising the following steps:

[0010] Step 1: Obtain a multi-site brain imaging dataset and preprocess it to obtain the processed resting-state functional connectivity matrix.

[0011] First, the raw rs-fMRI data were preprocessed including slice timing correction, noise signal regression, and temporal filtering. Afterwards, the exported rs-fMRI was normalized to the Montreal Neurological Institute (MNI) space by a nonlinear registration algorithm. Subsequently, the average time series of a set of brain regions were extracted based on the Anatomical Automatic Labeling (AAL) atlas containing 90 predefined regions of interest (ROIs). Finally, for each subject, we can generate a resting-state functional connectivity matrix represented by a 90×90 symmetric matrix, where each element in the matrix represents the Pearson correlation coefficient between a pair of ROIs;

[0012] Step 2: Construct a brain network based on the resting-state functional connectivity matrix.

[0013] The brain regions are regarded as nodes in the graph, and the elements X in the functional connectivity matrix are thresholded. ij Make a judgment, if brain area b i and brain area b j The Pearson correlation coefficient of is greater than the threshold, which means that brain area b i and brain area b j There is an edge connection to determine whether the edge in the graph exists, and the data in the i-th column of the functional connection matrix is ​​used as the brain area b i The node feature g i .

[0014] Step 3: Use intra-site comparative learning to extract the unique features of each site.

[0015] In order to enable the model to fully extract high-quality features from a specific site, the category information of the data is used to distinguish positive samples from negative samples, that is, samples with the same label are positive samples, while samples with different labels are negative samples, which makes the representation of the same category data concentrated in adjacent areas of the feature space. Specifically, we apply the contrast constraint within the site in the Local channel to extract the feature Z with enhanced representation loc , assuming that for the anchor sample feature Z i , the positive sample with the same label is Z p ∈P(i), the negative sample with a different label is Z n ∈N(i). The contrast loss function can be calculated by the following formula:

[0016]

[0017] Among them, P(i) represents the set of positive samples, N(i) represents the set of negative samples, sim(·,·) represents cosine similarity, Sig(·) is the sigmoid activation function, and τ is the temperature coefficient in contrastive learning.

[0018] Step 4: Use inter-site comparative learning to extract common features of different sites.

[0019] In order to alleviate the client drift caused by non-independent and identically distributed data, a graph comparison learning method between sites is used to improve the mutual information of the embedding representation of global and local graph structure information to achieve consistent global and local goals. Specifically,

[0020] First, a Laplacian-based graph structure embedding representation is obtained on the Global channel. Then, the feature Z extracted from the local model site channel and the global model site channel is maximized. cen and Z cen-g The mutual information of , the contrast loss is calculated by the following equation:

[0021]

[0022] Where S is the total number of samples, N is the number of nodes in the sample, and g cen k is the node representation of the kth node obtained from the Global channel of the private model, Z cen_g It is a graph representation obtained from the Global channel of the global model.

[0023] Step 5: Minimize the mutual information of the features extracted by the two-channel feature extractor,

[0024] In order to make the features extracted by the dual-channel feature extractor more discriminative, Z loc and Z cen The Jensen-Shannon divergence of is constrained to be maximum, so as to minimize their mutual information. Specifically,

[0025] Z is calculated by the following formula loc and Z cen The inter-channel loss is:

[0026]

[0027] Among them, Z loc is the graph representation extracted from the Local channel, Z cen It is a graph representation obtained from the Global channel.

[0028] Step 6: Use the extracted features to classify the samples. First, extract the personalized features Z loc And the common feature Z cen The samples are concatenated and combined into sample feature Z, which is then input into a multi-layer perceptron for classification, mapping the samples from feature space to label space to achieve the classification of brain diseases.

[0029] Step 7: Build a topological relationship diagram between multiple sites.

[0030] In each round of communication, different sites upload the gradient information of the Global channel, and the server saves the model gradient information of different sites at different times. Then, the topological relationship between sites is captured by calculating the similarity of the model gradient time series, and the importance of different site models to the federated aggregation is dynamically evaluated through the PageRank algorithm. Finally, the contribution weight of each site is used as the weight in the federated aggregation. Specifically, in order to reduce the transmission cost, the gradient of each Global channel graph convolution layer is concatenated according to the following formula, and its 2-norm is calculated before uploading it to the server. The server retains the model gradient of the last t rounds of communication.

[0031]

[0032] in It is c i The gradient sequence of the stations in the last t rounds of communication, It is c i The 2-norm of the gradient of the Global channel model in the last t communications in the sites, is the gradient of the jth graph convolutional layer in the last t communications. Then, the similarity between the two private models is calculated by the DDTW method, and then the site c is obtained. i and c j The topological relationship between

[0033]

[0034] Step 8: Aggregate the global model based on the contribution of different sites.

[0035] Use PageRank to evaluate the contribution of each site based on the topological relationship between sites. Then, calculate the aggregate weight of each site using the power method:

[0036]

[0037] Where d is the damping factor 0≤d≤1, E is an all-1 matrix, T is the transfer matrix of the first-order Markov chain in the graph, W is the PageRank value matrix of the graph, and N c is the total number of sites,

[0038] Finally, the global model parameters are updated by the following equations:

[0039]

[0040] Among them, θ i represents the private model parameters of site i, θ g Represents the global model parameters.

[0041] Step 9: The server transmits the global model parameters to each site, replacing the local model parameters.

[0042] The server transmits the aggregated global model parameters to each site. Each site replaces the local model parameters with the global model parameters and repeats steps 3 to 6 in sequence until the model of each site reaches the target accuracy.

[0043] Compared with the prior art, the present invention has the following advantages: 1) The technical solution proposes a federated learning method for multi-site brain disease diagnosis. The method only needs to transmit model parameters and gradient information between sites, without transmitting patient information, thereby effectively protecting patient privacy. In addition, through the comparative learning within and between sites and the federated aggregation strategy based on the multi-site topological relationship, the impact of cross-site non-independent and identically distributed medical data on the model classification effect in federated learning is effectively alleviated, and the optimal aggregation is achieved. The federated learning method of this scheme not only protects sensitive data, but also makes in-depth optimization on the heterogeneity of cross-site data. 2) Through the dual-channel diagnostic model and comparative learning within and between sites, the discrimination ability of private models is enhanced. First, the comparative learning within the site considers the feature correlation between data of the same category, uses category information to distinguish positive and negative samples, and makes the representation of similar data in the feature space more concentrated. Specifically, this method improves the accuracy and consistency of the internal model of each site by clustering the features of similar data in adjacent areas of the feature space. Secondly, contrastive learning between sites enhances the extraction process of common features of different sites, improves the consistency of local targets between multiple sites, and ensures that the models between sites can effectively collaborate, thereby improving the overall diagnostic effect. 3) The scheme captures the topological relationship between different private models, adaptively evaluates the importance of each site in the aggregation process, and incorporates these evaluations into the global model aggregation optimization. Specifically, we analyze the model gradients of each site to determine the contribution of each site in the global model, dynamically adjust the aggregation weights, and achieve more accurate and stable model aggregation. 4) Experiments on a multi-site brain disease dataset show that the proposed method significantly improves the effectiveness of multi-site disease diagnosis while protecting data privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the overall framework of the federated learning strategy for multi-site brain disease diagnosis;

[0045] Figure 2 Workflow diagram of the brain disease diagnosis method based on federated learning. DETAILED DESCRIPTION

[0046] In order to deepen the understanding of the present invention, the present embodiment is described in detail below with reference to the accompanying drawings.

[0047] Example 1: See Figure 1 , Figure 2 , a privacy-preserving cross-hospital disease data analysis method, the method comprising the following steps:

[0048] Step 1: Obtain a multi-site brain imaging dataset and preprocess it to obtain the processed resting-state functional connectivity matrix.

[0049] First, the raw rs-fMRI data were preprocessed including slice timing correction, noise signal regression, and temporal filtering. Afterwards, the exported rs-fMRI was normalized to the Montreal Neurological Institute (MNI) space using a nonlinear registration algorithm. Subsequently, the average time series of a set of brain regions were extracted based on the Anatomical Automatic Labeling (AAL) atlas containing 90 predefined regions of interest (ROIs). Finally, for each subject, we can generate a resting-state functional connectivity matrix represented by a 90×90 symmetric matrix, where each element in the matrix represents the Pearson correlation coefficient between a pair of ROIs,

[0050] Step 2: Construct a brain network based on the resting-state functional connectivity matrix.

[0051] The brain regions are regarded as nodes in the graph, and the elements X in the functional connectivity matrix are thresholded. ij Make a judgment, if brain area b i and brain area b j The Pearson correlation coefficient of is greater than the threshold, which means that brain area b i and brain area b j There is an edge connection to determine whether the edge in the graph exists, and the data in the i-th column of the functional connection matrix is ​​used as the brain area b i The node feature g i .

[0052] Step 3: Use intra-site comparative learning to extract the unique features of each site.

[0053] In order to enable the model to fully extract high-quality features from a specific site, the category information of the data is used to distinguish positive samples from negative samples, that is, samples with the same label are positive samples, while samples with different labels are negative samples, which makes the representation of the same category data concentrated in adjacent areas of the feature space. Specifically, we apply the contrast constraint within the site in the Local channel to extract the feature Z with enhanced representation loc , assuming that for the anchor sample feature Z i , the positive sample with the same label is Z p ∈P(i), the negative sample with a different label is Z n ∈N(i). The contrast loss function can be calculated by the following formula:

[0054]

[0055] Among them, P(i) represents the set of positive samples, N(i) represents the set of negative samples, sim(·,·) represents cosine similarity, Sig(·) is the sigmoid activation function, and τ is the temperature coefficient in contrastive learning.

[0056] Step 4: Use inter-site comparative learning to extract common features of different sites.

[0057] In order to alleviate the client drift caused by non-independent and identically distributed data, a graph contrast learning method between sites is used to improve the mutual information of the embedding representation of global and local graph structure information to achieve consistent global and local goals. Specifically, firstly, a Laplacian-based graph structure embedding representation is obtained on the Global channel, and then the feature Z extracted from the local model site channel and the global model site channel is maximized. cen and Z cen-g The mutual information of , the contrast loss is calculated by the following equation:

[0058]

[0059] Where S is the total number of samples, N is the number of nodes in the sample, and g cen k is the node representation of the kth node obtained from the Global channel of the private model, Z cen_g It is a graph representation obtained from the Global channel of the global model.

[0060] Step 5: Minimize the mutual information of the features extracted by the two-channel feature extractor,

[0061] In order to make the features extracted by the dual-channel feature extractor more discriminative, Z loc and Z cen The Jensen-Shannon divergence of is constrained to be maximum, so as to minimize their mutual information. Specifically,

[0062] Z is calculated by the following formula loc and Z cen The inter-channel loss is:

[0063]

[0064] Among them, Z loc is the graph representation extracted from the Local channel, Z cen It is a graph representation obtained from the Global channel.

[0065] Step 6: Use the extracted features to classify the samples. First, extract the personalized features Z loc And the common feature Z cen The samples are concatenated and combined into sample feature Z, which is then input into a multi-layer perceptron for classification, mapping the samples from feature space to label space to achieve the classification of brain diseases.

[0066] Step 7: Build a topological relationship diagram between multiple sites.

[0067] In each round of communication, different sites upload the gradient information of the Global channel, and the server saves the model gradient information of different sites at different times. Then, the topological relationship between sites is captured by calculating the similarity of the model gradient time series, and the importance of different site models to the federated aggregation is dynamically evaluated through the PageRank algorithm. Finally, the contribution weight of each site is used as the weight in the federated aggregation. Specifically, in order to reduce the transmission cost, the gradient of each Global channel graph convolution layer is concatenated according to the following formula, and its 2-norm is calculated before uploading it to the server. The server retains the model gradient of the last t rounds of communication.

[0068]

[0069] in It is c i The gradient sequence of the stations in the last t rounds of communication, It is c i The 2-norm of the gradient of the Global channel model in the last t communications in the sites, is the gradient of the jth graph convolutional layer in the last t communications. Then, the similarity between the two private models is calculated by the DDTW method, and then the site c is obtained. i and c j The topological relationship between

[0070]

[0071] Step 8: Aggregate the global model based on the contribution of different sites.

[0072] Use PageRank to evaluate the contribution of each site based on the topological relationship between sites. Then, calculate the aggregate weight of each site using the power method:

[0073]

[0074] Where d is the damping factor 0≤d≤1, E is an all-1 matrix, T is the transfer matrix of the first-order Markov chain in the graph, W is the PageRank value matrix of the graph, and N c is the total number of sites,

[0075] Finally, the global model parameters are updated by the following equations:

[0076]

[0077] Among them, θ i represents the private model parameters of site i, θ g Represents the global model parameters.

[0078] Step 9: The server transmits the global model parameters to each site, replacing the local model parameters.

[0079] The server transmits the aggregated global model parameters to each site. Each site replaces the local model parameters with the global model parameters and repeats steps 3 to 6 in sequence until the model of each site reaches the target accuracy.

[0080] Embodiment 2: A privacy-preserving cross-hospital disease data analysis system includes a feature extraction module, a classification module, a multi-site topology relationship building module, and a federation aggregation module. The overall framework structure is as follows: Figure 1 It consists of a private Center channel and a Local channel graph feature extractor F for each site. Center , F Local , a private multi-layer perceptron classifier C for each site and a federation aggregation strategy based on multi-site topological relationships, where each site’s private Center channel and Local channel graph feature extractor F Center , F Local They are used to extract the common features of samples from different sites and the individual features of samples from each site. The contrastive learning within the site is used to enhance the classification ability of the individual features, thereby improving the quality of individual feature extraction. The contrastive learning between sites is used to enhance the consistency of the common features of different sites, thereby aligning the local targets of different sites. The role of the classifier is to map the individual features and common features of the extracted samples into the label space, and to update the classifier through the cross entropy loss to realize the diagnosis function of brain diseases. The role of the federated aggregation strategy based on multi-site topological relationships is to capture the private model update trends of different sites, and dynamically evaluate the importance of each private model to the federated aggregation and use it for aggregation to achieve optimal federated aggregation.

[0081] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. A privacy-preserving cross-hospital disease data analysis method, characterized in that: The method comprises the following steps: Step 1: Obtain a multi-site brain imaging dataset. Step 2: Construct a brain network based on the resting-state functional connectivity matrix. Step 3: Use intra-site comparative learning to extract the unique features of each site. Step 4: Use inter-site comparative learning to extract common features of different sites. Step 5: Minimize the mutual information of the features extracted by the two-channel feature extractor, Step 6: Use the extracted features to classify the samples. Step 7: Build a topological relationship diagram between multiple sites. Step 8: Aggregate the global model based on the contribution of different sites. Step 9: The server transmits the global model parameters to each site, replacing the local model parameters. Among them, step 1: obtain a multi-site brain imaging dataset and preprocess it to obtain the processed resting-state functional connectivity matrix, as follows: First, the raw rs-fMRI data were preprocessed including slice timing correction, noise signal regression and temporal filtering. After that, the exported rs-fMRI was normalized to the Montreal Neurological Institute (MNI) space by a nonlinear registration algorithm. Subsequently, the average time series of a set of brain regions were extracted based on the anatomically automatically labeled AAL atlas containing 90 predefined regions of interest (ROIs). Finally, for each subject, a resting-state functional connectivity matrix represented by a 90 × 90 symmetric matrix was generated, where each element in the matrix represents the Pearson correlation coefficient between a pair of ROIs. Step 2: Construct the brain network based on the resting state functional connectivity matrix, as follows: The brain regions are regarded as nodes in the graph, and the elements X in the functional connectivity matrix are thresholded. ij Make a judgment, if brain area b i and brain area b j The Pearson correlation coefficient of is greater than the threshold, which means that brain area b i and brain area b j There is an edge connection to determine whether the edge in the graph exists, and the data in the i-th column of the functional connection matrix is ​​used as the brain area b i The node feature g i , Step 3: Use intra-site comparative learning to extract the unique features of each site, as follows: Apply intra-site contrast constraints in the Local channel to extract features Z with enhanced representation loc , assuming that for the anchor sample feature Z i , the positive sample with the same label is Z p ∈P(i), the negative sample with a different label is Z n ∈N(i), the contrast loss function is calculated by the following formula: Among them, P(i) represents the positive sample set, N(i) represents the negative sample set, sim(·,·) represents the cosine similarity, Sig(·) is the Sigmoid activation function, τ is the temperature coefficient in contrastive learning, Step 4: Use inter-site comparative learning to extract common features of different sites, as follows: First, a Laplacian-based graph structure embedding representation is obtained on the Global channel. Then, the feature Z extracted from the local model site channel and the global model site channel is maximized. cen and Z cen-g The mutual information of , the contrast loss is calculated by the following equation: Where S is the total number of samples, N is the number of nodes in the sample, and g cen k is the node representation of the kth node obtained from the Global channel of the private model, Z cen_g is the graph representation obtained from the Global channel of the global model, Step 5: Minimize the mutual information of the features extracted by the two-channel feature extractor as follows: Z is calculated by the following formula loc and Z cen The inter-channel loss is: Among them, Z loc is the graph representation extracted from the Local channel, Z cen It is a graph representation obtained from the Global channel. Step 6: Use the extracted features to classify the samples. First, extract the personalized features Z loc And the common feature Z cen The samples are concatenated and combined into sample features Z, which are then input into a multi-layer perceptron for classification. The samples are mapped from feature space to label space to achieve the classification of brain diseases. Step 7: Build a topological relationship diagram between multiple sites. In each round of communication, different sites upload the gradient information of the Global channel. The server saves the model gradient information of different sites at different times. Then, the topological relationship between sites is captured by calculating the similarity of the model gradient time series, and the importance of different site models to the federated aggregation is dynamically evaluated through the PageRank algorithm. Finally, the contribution weight of each site is used as the weight in the federated aggregation. The gradient of each Global channel graph convolution layer is concatenated according to the following formula, and its 2-norm is calculated before uploading it to the server. The server retains the model gradient of the last t rounds of communication. in It is c i The gradient sequence of the stations in the last t rounds of communication, It is c i The 2-norm of the gradient of the Global channel model in the last t communications in the sites, is the gradient of the jth graph convolutional layer in the last t communications. Then, the similarity between the two private models is calculated by the DDTW method, and then the site c is obtained. i and c j The topological relationship between Step 8: Aggregate the global model based on the contribution of different sites. Use PageRank to evaluate the contribution of each site based on the topological relationship between sites, and then calculate the aggregate weight of each site using the power method: Where d is the damping factor 0≤d≤1, E is an all-1 matrix, T is the transfer matrix of the first-order Markov chain in the graph, W is the PageRank value matrix of the graph, and N c is the total number of sites, Finally, the global model parameters are updated by the following equations: Among them, θ i represents the private model parameters of site i, θ g Represents global model parameters.

2. A privacy-preserving cross-hospital disease data analysis method according to claim 1, characterized in that: Step 9: The server transmits the global model parameters to each site to replace the local model parameters. The server transmits the aggregated global model parameters to each site. Each site replaces the local model parameters with the global model parameters and repeats steps 3 to 6 in sequence until the models of each site reach the target accuracy.

Images