A data analysis method in the process of building high-value medical data assets

Abstract

The present application provides a data analysis method in the process of building high-value medical data assets, which relates to the field of medical data analysis technology and includes the following steps: generating a gene feature subset, a clinical feature subset, and an image feature subset; encoding the gene feature subset in the time dimension to generate a gene expression time series vector; annotating the clinical feature subset with a timestamp to extract the time series characteristics of clinical events; registering the image feature subset in three-dimensional space to extract the spatiotemporal dynamic characteristics of the lesion area; obtaining the image lesion coordinates based on the spatiotemporal dynamic characteristics, and mapping the image lesion coordinates to the gene-clinical joint space; constructing a heterogeneous graph network containing gene nodes, clinical nodes, and image nodes, and calculating the edge weights between each node; in the heterogeneous graph network, pruning redundant edges to generate a spatiotemporal correlation map, and then obtaining a correlation mapping matrix. This method integrates heterogeneous data so that data originally scattered in different systems can establish effective connections.

Description

Technical Field

[0001] The present application relates to the field of medical data analysis technology, and in particular to a data analysis method in the process of constructing high-value medical data assets. Background Art

[0002] With the rapid development of medical informatization, medical data has seen explosive growth. This data holds immense value and can provide strong support for disease diagnosis and treatment plan development. However, because medical data often comes from heterogeneous sources, when doctors or hospitals want to study the relationship between certain disease types and one or more types of medical data, it is difficult to quickly and accurately extract and connect disease-related data from different systems, resulting in low analysis efficiency. Summary of the Invention

[0003] In view of the above-mentioned defects or deficiencies in the prior art, this application aims to provide a data analysis method in the process of building high-value medical data assets to establish a connection between multimodal medical data. The method includes the following steps:

[0004] Extract effective gene loci related to the target disease and generate a gene feature subset; extract clinical indicators associated with the target disease and generate a clinical feature subset; extract quantitative features of the lesion area and generate an imaging feature subset;

[0005] The gene feature subset is encoded in the time dimension to generate a gene expression time series vector; the clinical feature subset is timestamped to extract the clinical event time series features; the image feature subset is registered in three-dimensional space to extract the spatiotemporal dynamic features of the lesion area;

[0006] aligning the gene expression time series of the gene expression time series vector and the clinical event time series of the clinical event time series feature, and calculating the time alignment similarity;

[0007] Based on the spatiotemporal dynamic characteristics, the image lesion coordinates are obtained, the image lesion coordinates are mapped to the gene-clinical joint space, and the spatial correlation is calculated;

[0008] Constructing a heterogeneous graph network including gene nodes, clinical nodes, and imaging nodes, and calculating edge weights between nodes based on the temporal alignment similarity and the spatial correlation;

[0009] In the heterogeneous graph network, redundant edges corresponding to the edge weights lower than a first preset threshold are pruned to generate a spatiotemporal correlation graph, thereby obtaining a correlation mapping matrix.

[0010] According to the technical solution provided in this application, generating a spatiotemporal correlation map and then obtaining a correlation mapping matrix includes the following steps:

[0011] Converting the spatiotemporal correlation graph into a weighted adjacency matrix, wherein the matrix element values ​​of the weighted adjacency matrix are the weights of the corresponding edges;

[0012] Performing symmetric normalization processing on the weighted adjacency matrix to obtain a normalized matrix;

[0013] Performing singular value decomposition on the normalized matrix, retaining the first k principal components, and generating a low-dimensional mapping matrix as the association mapping matrix.

[0014] According to the technical solution provided by the present application, performing singular value decomposition on the normalized matrix, retaining the first k principal components, and generating a low-dimensional mapping matrix as the association mapping matrix includes the following steps:

[0015] Performing singular value decomposition on the normalized matrix, retaining the first k principal components, generating a low-dimensional mapping matrix, and generating feature clusters based on the low-dimensional mapping matrix;

[0016] The cluster labels of the feature clusters are fed back to the singular value decomposition as auxiliary information, and the truncation rank k value in the singular value decomposition process is iteratively optimized to obtain the association mapping matrix.

[0017] According to the technical solution provided in this application, the extraction of effective gene loci related to the target disease includes the following steps:

[0018] Using the chi-square test, the SNP sites whose occurrence frequency in the population exceeds the second preset threshold are screened to obtain a candidate SNP list;

[0019] Obtain a GWAS database, compare the candidate SNP list with the GWAS database, and retain SNP sites with a p-value less than a third preset threshold as valid gene sites; wherein the third preset threshold is less than 0.05.

[0020] According to the technical solution provided in this application, encoding the gene feature subset in the time dimension to generate a gene expression time series vector includes the following steps:

[0021] For moments with missing gene expression, cubic spline interpolation was used to fill in the gaps;

[0022] Multi-scale temporal features of gene expression are extracted by wavelet transform, wherein the multi-scale temporal features include high-frequency fluctuation components and low-frequency trend components.

[0023] According to the technical solution provided in this application, the alignment of the gene expression timing of the gene expression timing vector and the clinical event timing of the clinical event timing feature includes the following steps:

[0024] Performing dynamic time warping on the gene expression time series and the clinical event time series to generate a time-aligned path;

[0025] Calculating the Pearson correlation coefficient of the gene-clinical feature pairs along the time alignment path, and screening feature pairs with correlation coefficients greater than a fourth preset threshold as strongly associated feature pairs;

[0026] Based on the strongly correlated feature pairs, constructing a gene-clinical initial correlation matrix, wherein the matrix elements of the initial correlation matrix are correlation coefficient values;

[0027] Performing non-negative matrix decomposition on the initial correlation matrix to obtain a gene feature subspace and a clinical feature subspace;

[0028] The gene feature subspace and the clinical feature subspace are orthogonally projected and merged to generate the gene-clinical joint space.

[0029] According to the technical solution provided in this application, the extraction of effective gene loci related to the target disease and the generation of gene feature subsets include the following steps:

[0030] Performing random projection dimensionality reduction on the expression time series vectors of the screened effective gene loci to generate a feature space after dimensionality reduction;

[0031] Add Laplace noise to the feature space after dimensionality reduction to satisfy -Differential privacy constraints are applied to obtain a subset of genetic features.

[0032] According to the technical solution provided by this application, the method further includes the following steps:

[0033] When constructing the edge weights of a heterogeneous graph network, noise compensation is performed on a subset of gene features to obtain corrected weights, which are then used to generate spatiotemporal association maps.

[0034] According to the technical solution provided in this application, extracting clinical indicators associated with the target disease and generating a clinical feature subset includes the following steps:

[0035] Extract clinical indicators associated with the target disease and homomorphically encrypt the patient identifier to obtain a subset of clinical features;

[0036] The feature similarity between genetic features and encrypted clinical features is calculated in the encrypted domain, and a clinical feature subset is output.

[0037] According to the technical solution provided by this application, after homomorphically encrypting the patient identity identifier, the following steps are also included:

[0038] Correlation coefficients between gene signatures and subsets of clinical features were calculated in the encrypted domain;

[0039] In the orthogonal projection of the joint gene-clinical space, a subset of clinical features is locally decrypted, exposing only the standardized feature values ​​to participate in the space merging.

[0040] Compared with the existing technology, the beneficial effect of this application is that it integrates heterogeneous data such as genetic, clinical, and imaging data from different systems into a joint feature space, establishing a multi-dimensional association mapping matrix. This enables the data originally scattered across different systems to establish effective connections, greatly improving the efficiency of data association analysis and reducing the time cost of data extraction and integration. The multi-dimensional association mapping matrix can be used to fully explore the potential information in multimodal medical data, providing a more accurate and valuable diagnosis and treatment decision feature set for disease diagnosis, treatment plan formulation, etc., which will help to subsequently improve the accuracy and scientific nature of medical decision-making and achieve precision medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of the steps of the data analysis method in the process of building high-value medical data assets provided in this application. DETAILED DESCRIPTION

[0042] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0043] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] Example 1

[0045] As mentioned in the background technology, in response to the problems in the existing technology, this application proposes to provide a data analysis method in the process of building high-value medical data assets to establish connections between multimodal medical data, such as Figure 1 As shown, the method includes the following steps:

[0046] S1. Extract effective gene loci related to the target disease and generate a gene feature subset; extract clinical indicators associated with the target disease and generate a clinical feature subset; extract quantitative features of the lesion area and generate an imaging feature subset;

[0047] Furthermore, the extraction of effective gene loci related to the target disease includes the following steps:

[0048] Using the chi-square test, the SNP sites whose occurrence frequency in the population exceeds the second preset threshold are screened to obtain a candidate SNP list;

[0049] Obtain a GWAS database, compare the candidate SNP list with the GWAS database, and retain SNP sites with a p-value less than a third preset threshold as valid gene sites; wherein the third preset threshold is less than 0.05.

[0050] S2. Encode the gene feature subset in the time dimension to generate a gene expression time series vector; annotate the clinical feature subset with a timestamp to extract the clinical event time series features; perform three-dimensional spatial registration on the image feature subset to extract the spatiotemporal dynamic features of the lesion area;

[0051] Furthermore, encoding the gene feature subset in the time dimension to generate a gene expression time series vector comprises the following steps:

[0052] For moments with missing gene expression, cubic spline interpolation was used to fill in the gaps;

[0053] Multi-scale temporal features of gene expression are extracted by wavelet transform, wherein the multi-scale temporal features include high-frequency fluctuation components and low-frequency trend components.

[0054] S3. Aligning the gene expression time series of the gene expression time series vector with the clinical event time series of the clinical event time series feature, and calculating the time alignment similarity;

[0055] Furthermore, aligning the gene expression time series of the gene expression time series vector with the clinical event time series of the clinical event time series feature comprises the following steps:

[0056] Performing dynamic time warping on the gene expression time series and the clinical event time series to generate a time-aligned path;

[0057] Calculating the Pearson correlation coefficient of the gene-clinical feature pairs along the time alignment path, and screening feature pairs with correlation coefficients greater than a fourth preset threshold as strongly associated feature pairs;

[0058] Based on the strongly correlated feature pairs, constructing a gene-clinical initial correlation matrix, wherein the matrix elements of the initial correlation matrix are correlation coefficient values;

[0059] Performing non-negative matrix decomposition on the initial correlation matrix to obtain a gene feature subspace and a clinical feature subspace;

[0060] The gene feature subspace and the clinical feature subspace are orthogonally projected and merged to generate the gene-clinical joint space.

[0061] S4. Based on the spatiotemporal dynamic characteristics, obtaining the image lesion coordinates, mapping the image lesion coordinates to the gene-clinical joint space, and calculating the spatial correlation;

[0062] S5. Constructing a heterogeneous graph network including gene nodes, clinical nodes, and imaging nodes, and calculating edge weights between nodes based on the time alignment similarity and the spatial correlation;

[0063] S6. In the heterogeneous graph network, redundant edges corresponding to the edge weights that are lower than a first preset threshold are pruned to generate a spatiotemporal correlation graph, thereby obtaining a correlation mapping matrix.

[0064] Furthermore, the generation of the spatiotemporal correlation map and the acquisition of the correlation mapping matrix include the following steps:

[0065] Converting the spatiotemporal correlation graph into a weighted adjacency matrix, wherein the matrix element values ​​of the weighted adjacency matrix are the weights of the corresponding edges;

[0066] Performing symmetric normalization processing on the weighted adjacency matrix to obtain a normalized matrix;

[0067] Performing singular value decomposition on the normalized matrix, retaining the first k principal components, and generating a low-dimensional mapping matrix as the association mapping matrix.

[0068] Furthermore, performing singular value decomposition on the normalized matrix, retaining the first k principal components, and generating a low-dimensional mapping matrix as the correlation mapping matrix includes the following steps:

[0069] Performing singular value decomposition on the normalized matrix, retaining the first k principal components, generating a low-dimensional mapping matrix, and generating feature clusters based on the low-dimensional mapping matrix;

[0070] The cluster labels of the feature clusters are fed back to the singular value decomposition as auxiliary information, and the truncation rank k value in the singular value decomposition process is iteratively optimized to obtain the association mapping matrix.

[0071] In existing technologies, hospitals accumulate large amounts of multimodal data corresponding to patients. Often, different modalities of diagnosis and treatment data are stored in different systems. When doctors want to study the relationship between specific gene mutations and types of heart disease, it is difficult to quickly and accurately extract relevant data from different systems and establish connections, resulting in low analysis efficiency. In addition, when hospitals need to collaborate with external research institutions to study the effectiveness of disease treatments, they lack adequate protection for patient privacy and directly provide patients' original clinical data to the other party, which poses the risk of patient privacy leakage. At the same time, when doctors formulate treatment plans for new patients with traditional technologies, due to the lack of high-value mining of previous medical data, they cannot fully support clinical diagnosis and treatment decisions. They can only refer to simple clinical indicators and previous experience, which makes it difficult to formulate personalized treatment plans, affecting treatment effectiveness. Based on this, this solution is proposed.

[0072] Specifically, these different modal data are linked based on the patient's unique identifier (such as medical record number) and time information (such as the time of consultation and symptom onset). These associations are then converted into an association mapping matrix, where the elements represent the degree of association between different feature subsets, such as the degree of association between genetic features and clinical symptom features, and the degree of association between imaging features and clinical diagnostic features. For example, genetic features are used as one dimension, clinical features as another, and imaging features as a third. The elements in the association mapping matrix determine the interrelationships and connections between these dimensions. In this way, the previously dispersed multimodal features are integrated into a unified space, forming a structured joint feature space. In this space, features from different modalities no longer exist in isolation, but are interconnected and influence each other.

[0073] For example, a cardiology hospital has accumulated a large amount of patient genomic data, clinical text data (such as medical records and diagnostic reports), and medical imaging data such as cardiac ultrasound and electrocardiograms. The target disease is identified as coronary artery disease (CAD). Within a population of CAD patients, whole-genome sequencing data is used to screen for genetic loci associated with atherosclerosis (such as APOE and PCSK9). SNPs with a frequency greater than 5% (a second preset threshold) in the population are retained. For example, the rs7412 locus (APOE gene) is selected as a candidate gene feature. Clinical indicators related to CAD are extracted from electronic medical records. For structured data, left ventricular ejection fraction (LVEF) and BNP (brain natriuretic peptide) levels are directly recorded. For unstructured text, natural language processing (NLP) is used to identify key descriptors such as "chest pain duration >30 minutes" and "ST-segment elevation" and normalize them to ICD-10 codes (e.g., I21.0 acute anterior myocardial infarction). Three-dimensional reconstruction is then performed on cardiac CT images to extract quantitative features such as the coronary artery calcium score (CAC score) and the volume of the ischemic area. For example, a deep learning segmentation algorithm is used to calculate the degree of left anterior descending artery stenosis (e.g., 70% stenosis). Then, temporal encoding and time series alignment are performed: Gene time series vectors record APOE gene expression levels at one week, one month, and three months after surgery, with missing data imputed using cubic spline interpolation; clinical event time series annotate key time points (e.g., "arrhythmia developed on the third day after surgery"); and dynamic time warping alignment aligns the peak time of gene expression (e.g., a sudden increase in APOE expression one month after surgery) with the time of clinical indicator deterioration (e.g., a decrease in LVEF to 40% one month after surgery). Time offset similarity is calculated (e.g., DTW distance = 0.15), and a weighted adjacency matrix is ​​constructed. Three types of nodes are defined in the heterogeneous graph network: gene nodes (e.g., APOE, PCSK9, and other gene loci); clinical nodes (e.g., LVEF, BNP, and other indicators); and imaging nodes (e.g., coronary artery calcium score and myocardial ischemic volume). Edge weight calculation: Time alignment similarity: DTW correlation coefficient between APOE gene expression and LVEF change = 0.82; Spatial correlation: The coordinates of the myocardial ischemic area are mapped to the gene-clinical joint space, and the spatial correlation between the APOE gene and the ischemic volume is calculated (e.g., Pearson r = 0.75). Then, graph pruning and matrix decomposition are performed: redundant edges with weight < 0.6 (the first preset threshold) are pruned (e.g., the association weight between PCSK9 and BNP = 0.45 is removed); the remaining adjacency matrix is ​​symmetric normalized (formula: D - ¹ / ²AD -¹ / ²), singular value decomposition (SVD) retained the first k=3 principal components (cumulative contribution rate of 92%) to generate a three-dimensional mapping matrix. For example, the first principal component represents the "core pathway for atherosclerosis progression," which includes the APOE gene, LVEF, and calcification score. After the initial decomposition, the characteristic cluster silhouette coefficient of the APOE gene and calcification score was found to be 0.68 (<0.7 threshold), triggering k value optimization. Specifically, increasing k=5, the silhouette coefficient increased to 0.73 after re-decomposition. In the final association mapping matrix, the association weight between the APOE gene and calcification score increased from 0.71 to 0.89. For example, the APOE gene expression time series (1 week to 3 months after surgery) was aligned with the LVEF change time series; the gene expression peak (1 month after surgery) corresponded to the lowest LVEF value (40%); the optimal alignment path was generated, the correlation coefficient r = 0.85 (>0.7 threshold) was calculated, and the gene-clinical initial correlation matrix was constructed (APOE and LVEF correlation coefficient = 0.85, PCSK9 and BNP = 0.72); NMF decomposition obtained: gene subspace: weight vectors of APOE and PCSK9 genes (0.9, 0.1); clinical subspace: weight vectors of LVEF and BNP (0.8, 0.2); after orthogonal projection merging, the gene-clinical joint space coordinates of APOE-LVEF were (0.72, 0.15) reflects a strong correlation. Through the association mapping matrix, it can be obtained that among patients with high APOE gene expression, the group with LVEF < 50% and calcification score > 400 has a 3.2-fold increased risk of cardiac mortality. Based on this, doctors can intervene in high-risk patients in advance (such as intensive lipid-lowering treatment) to support diagnosis and treatment decisions. The gene-imaging joint space can reveal that coronary artery calcification progresses faster in APOEε4 allele carriers (p = 0.003), providing a basis for targeted treatment and having certain scientific research value.

[0074] Example 2:

[0075] Based on Example 1, and considering the high sensitivity of medical data (traditional medical data analysis often treats data processing and privacy protection as separate, independent stages, resulting in either sacrificing data utility to achieve strong privacy protection or exposing sensitive information in exchange for analytical accuracy), this example proposes the following solution to simultaneously implement a privacy enhancement mechanism during the construction of the spatiotemporal correlation graph. Specifically:

[0076] In a preferred embodiment, extracting effective gene loci associated with the target disease and generating a gene feature subset comprises the following steps:

[0077] Performing random projection dimensionality reduction on the expression time series vectors of the screened effective gene loci to generate a feature space after dimensionality reduction;

[0078] Add Laplace noise to the feature space after dimensionality reduction to satisfy -Differential privacy constraints are applied to obtain a subset of genetic features.

[0079] In a preferred embodiment, extracting clinical indicators associated with the target disease and generating a clinical feature subset comprises the following steps:

[0080] Clinical indicators associated with the target disease are extracted, and patient identifiers are homomorphically encrypted to obtain a subset of clinical features.

[0081] The feature similarity between genetic features and encrypted clinical features is calculated in the encrypted domain, and a clinical feature subset is output.

[0082] Specifically, for the perturbed gene feature subset, a random projection transformation algorithm is used to transform the features of the effective gene loci. Noise is then injected according to the requirements of (ε, δ)-differential privacy. For example, random noise within a certain range is added to the gene expression data. The ciphertext features are encrypted using the Paillier homomorphic encryption algorithm to encrypt the highly sensitive patient identity features in the clinical indicator feature subset. Feature similarity comparison is performed in the encrypted domain. This ensures that even if the data is obtained, the patient's true identity cannot be known without the decryption key. The original form of the image feature subset is directly retained because the image features themselves are relatively less privacy-sensitive and retaining the original form facilitates subsequent analysis.

[0083] Furthermore, the privacy budget monitor dynamically adjusts the noise injection amount of different feature dimensions to ensure that the overall privacy budget consumption does not exceed the first preset budget. The following steps are included:

[0084] Based on the privacy budget monitor, the information entropy of each feature dimension is calculated in real time;

[0085] Specifically, the information entropy formula is used to analyze the distribution of genetic loci. If the distribution of genetic loci across a patient population is relatively uniform, meaning that different genetic types appear with similar frequencies, then the information entropy is high, indicating a high level of uncertainty within that feature dimension. Conversely, if the distribution of genetic loci is concentrated, the information entropy is low. For clinical indicator feature dimensions, such as patient age distribution, information entropy is high if the age range is wide and evenly distributed; if the majority of patients are concentrated in a certain age group, the information entropy is low. Similarly, for imaging feature dimensions, information entropy is calculated by analyzing the distribution and diversity of lesion features in the images.

[0086] Establish an inverse rule for allocating privacy budgets. Based on this inverse rule, allocate privacy budgets according to information entropy.

[0087] Specifically, because features with high information entropy inherently have high uncertainty, the need for privacy protection is relatively low. Features with low information entropy, on the other hand, have high certainty and, if leaked, could pose a greater privacy risk, thus requiring more privacy protection resources. Therefore, features with high information entropy are allocated a smaller privacy budget, while features with low information entropy are allocated a larger privacy budget. For example, within the gene sequence feature dimension, certain highly polymorphic gene regions have high information entropy and are allocated a smaller privacy budget; whereas relatively stable gene loci associated with key diseases have low information entropy and are allocated a larger privacy budget. In practice, the overall privacy budget is allocated according to the information entropy ratio of each feature dimension, using a pre-defined algorithm.

[0088] When it is detected that the privacy budget allocation fluctuation exceeds the second preset budget within N consecutive cycles, the feature dimension isolation mechanism is automatically triggered, and the dimension data is suspended from participating in feature evolution.

[0089] Specifically, the system continuously tracks fluctuations in privacy budget allocation. If it detects that the privacy budget allocation for a feature dimension fluctuates by more than a second preset budget (e.g., 0.1) for N consecutive cycles (assuming N = 5), the feature dimension isolation mechanism is automatically triggered.

[0090] For example, during data processing analyzing the relationship between treatment outcomes and genetic and clinical indicators for heart disease patients, if the privacy budget allocation for a clinical indicator feature dimension fluctuates by more than 0.1 for five consecutive cycles, the data for that dimension will be suspended from participating in feature evolution. During this quarantine period, no adjustments to the noise injection level or other operations related to feature evolution will be performed on that dimension until the fluctuations return to normal or after further security assessment. A decision will then be made to reinclude the data for that dimension in feature evolution.

[0091] In a preferred embodiment, the method further comprises the following steps:

[0092] When constructing the edge weights of a heterogeneous graph network, noise compensation is performed on a subset of gene features to obtain corrected weights, which are then used to generate spatiotemporal association maps.

[0093] In a preferred embodiment, after homomorphically encrypting the patient identifier, the following steps are further included:

[0094] Correlation coefficients between gene signatures and subsets of clinical features were calculated in the encrypted domain;

[0095] In the orthogonal projection of the joint gene-clinical space, a subset of clinical features is locally decrypted, exposing only the standardized feature values ​​to participate in the space merging.

[0096] Furthermore, a trainable weight matrix is ​​obtained, and the trainable weight matrix is ​​combined with the association mapping matrix to obtain a feature association strength matrix; and a feature dimension risk score is comprehensively generated; the following steps are included:

[0097] Applying the trainable weight matrix to the association mapping matrix, assigning attention weights to different feature nodes, and generating feature importance scores, wherein the feature importance scores are used to measure the closeness between features;

[0098] Integrating the feature importance scores to obtain a feature association strength matrix;

[0099] Based on the feature dimensions of the feature correlation strength matrix, potential privacy attack behaviors are simulated and the possibility of privacy leakage of each feature dimension is evaluated to output the risk probability distribution corresponding to each feature dimension, thereby obtaining the feature dimension risk score.

[0100] Specifically, a trainable weight matrix can be obtained through machine learning algorithms, such as the training process of a neural network. This matrix is ​​applied to the association mapping matrix, assigning attention weights to different feature nodes. For example, in a disease diagnosis model, genetic feature A and clinical indicator B, which have a greater impact on the diagnosis, are given higher attention weights, generating feature importance scores. These scores are then integrated to form a feature association strength matrix. This matrix is ​​then used to simulate potential privacy attacks and assess the potential for privacy leakage in each feature dimension. For example, by simulating a hacker invading a data system, the team can analyze which feature dimensions are most susceptible to being cracked and thus leaking privacy. This analysis then outputs a risk probability distribution, resulting in a feature dimension risk score.

[0101] Furthermore, generating a dynamic feature weight matrix according to the feature association strength matrix and the feature dimension risk score includes the following steps:

[0102] The dynamic feature weight matrix is ​​obtained through the dual-objective reward function R = α·U(F)-β·P(F);

[0103] Wherein, U(F) is the clinical utility evaluation value obtained based on the feature subset F involved in the feature association strength matrix, P(F) is the privacy leakage risk evaluation value quantified based on the privacy leakage risk reflected by the feature dimension risk score, and α and β are dynamic adjustment coefficients. The dynamic adjustment coefficients in the dual-objective reward function are determined by the following method:

[0104] Establish the medical data value density function V(x)=log(1+IC(x)·SC(x)), where IC(x) is the information content index and SC(x) is the clinical significance index;

[0105] Specifically, for each piece of medical data x, we first need to determine its information content (IC(x)) and clinical significance (SC(x)). IC(x) can be measured by calculating the data's information entropy, for example. Higher information entropy indicates greater uncertainty and information content within the data. For example, for genetic testing data, complex polymorphisms at certain loci will result in higher IC(x) values. SC(x) needs to be evaluated in conjunction with clinical expert opinion and medical research findings. For example, ejection fraction data, a key indicator for diagnosing heart failure, will have a higher SC(x) value. By substituting IC(x) and SC(x) into the formula, we can determine the value density of each piece of data.

[0106] According to the security level requirements of real-time data access scenarios, the α / β ratio is dynamically adjusted through the fuzzy logic controller;

[0107] Specifically, the fuzzy logic controller monitors the security level requirements of data access scenarios in real time. Security levels can be divided based on factors such as the identity of the visitor (such as internal doctors, external researchers), the purpose of the visit (daily diagnosis and treatment, scientific research analysis), and so on. For example, when external researchers conduct large-scale data access for scientific research projects, the security level may be set to a medium level. The fuzzy logic controller dynamically adjusts the α / β ratio based on the security level according to pre-set fuzzy rules. If the security level is high, more emphasis is placed on privacy protection, and the ratio of β to α will be appropriately increased; if the security level is low, more emphasis is placed on the clinical utility of the data, and the ratio of α to β will be increased. According to different data usage scenarios, the relationship between clinical utility and privacy protection can be flexibly balanced to improve the security and effectiveness of data use.

[0108] When abnormal access behavior is detected, the β value multiplication mechanism is automatically triggered to enhance the privacy protection strength.

[0109] Specifically, when the system detects unusual access behavior, such as a large number of unauthorized access requests within a short period of time or access from suspicious IP addresses, it automatically triggers a β multiplication mechanism. For example, if β is originally set to 0.3, upon detecting the anomaly, β is increased to 0.6. This mechanism can rapidly enhance privacy protection in the face of potential data breaches, reducing the resulting losses and ensuring patient data security.

[0110] For example, suppose a heart disease hospital collects medical data from a group of patients and extracts a feature subset F, which includes genetic test results, electrocardiogram data, and clinical symptom descriptions. After evaluation by professional physicians and calculations using a data analysis model, the accuracy of feature subset F for determining the type of heart disease (e.g., coronary artery disease, cardiomyopathy, etc.) is estimated to be U(F) = 0.8. By analyzing sensitive patient information contained in feature subset F, such as family genetic history and personal identity information, the privacy leakage risk assessment value P(F) = 0.4 is estimated. Assuming that the current scenario is a scientific research project with high privacy protection requirements, α is set to 0.4 and β is set to 0.6. Based on the dual-objective reward function, R is calculated to be 0.08.

[0111] Optionally, determining the contribution ratio of each feature dimension in the feature subset F (such as genetic test results, electrocardiogram data, and clinical symptom description) to the determination of the type of heart disease, and obtaining a dynamic feature weight matrix;

[0112] For example, through multiple experiments and data analysis, it is determined that the contribution ratio of genetic test results to the judgment of heart disease type is 0.5, the contribution ratio of electrocardiogram data is 0.3, and the contribution ratio of clinical symptom description is 0.2. The calculated R value (0.08) is distributed to each feature dimension according to these contribution ratios to obtain the initial dynamic feature weight matrix ;

[0113] Optionally, the original data of each feature dimension in the feature subset can be normalized so that its value range is between [0, 1], and the R value is distributed according to the normalized data ratio to obtain a dynamic feature weight matrix;

[0114] For example, if the normalized value of the gene test result is 0.6, the normalized value of the electrocardiogram data is 0.3, and the normalized value of the clinical symptom description is 0.1, then the weight of the gene test result is 0.048, the weight of the electrocardiogram data is 0.024, and the weight of the clinical symptom description is 0.008, and the dynamic feature weight matrix is ​​obtained. ;

[0115] Optionally, the hierarchical analysis method can be used to determine the relative importance weights of each feature dimension in the feature subset, construct a judgment matrix, and determine the relative importance comparison between each feature dimension through expert scoring, thereby obtaining a dynamic feature weight matrix.

[0116] For example, the importance of genetic test results compared to electrocardiogram data is 3 (indicating that genetic test results are slightly more important than electrocardiogram data), the importance of genetic test results compared to clinical symptom descriptions is 5 (indicating that genetic test results are significantly more important than clinical symptom descriptions), and the importance of electrocardiogram data compared to clinical symptom descriptions is 2 (indicating that electrocardiogram data is slightly more important than clinical symptom descriptions). The calculation yields a weight of 0.5396 for genetic test results, 0.2970 for electrocardiogram data, and 0.1634 for clinical symptom descriptions. The R value is then distributed according to these weights: the weight of genetic test results is 0.043168, the weight of electrocardiogram data is 0.02376, and the weight of clinical symptom descriptions is 0.013072, resulting in a dynamic feature weight matrix. .

[0117] Furthermore, based on the feature dimension risk score and the dynamic feature weight matrix, the sensitivity of each feature dimension is calculated, the privacy budget is dynamically allocated in combination with the privacy protection strength of each feature in the polymorphic feature space, and the perturbation feature is noise-recalibrated to generate an enhanced privacy feature set; the steps include:

[0118] By formula , calculate the sensitivity weight of each feature dimension; where, w i For the i The sensitivity weight of the feature dimension, , is the i-th feature dimension in the dynamic feature weight matrix, Score the risk of the feature dimension i; k The sensitivity adjustment coefficient is a predefined hyperparameter. It controls the sensitivity weight's response to the feature dimension's sensitivity score by adjusting the slope of the function. For example, when k is large, the sensitivity weight changes more steeply with the feature dimension's sensitivity score, resulting in higher discrimination. When k is small, the sensitivity weight changes more gradually, resulting in lower discrimination. k is typically set based on empirical experience (e.g., k = 1) or determined through cross-validation optimization.

[0119] By formula Dynamic allocation is performed to obtain the privacy budget allocated to each feature dimension, and the formula , a momentum factor is introduced to adjust the noise injection intensity so that the privacy budget of the current cycle refers to the historical allocation record; wherein, is the privacy budget allocated to the i-th feature dimension, which represents the amount of resources allocated for adding privacy protection operations, For the overall privacy budget, is the sum of the sensitivity weights of all feature dimensions, is the adjusted noise injection intensity at time period t, is the noise injection intensity of the previous time period, is the momentum coefficient.

[0120] For example, assuming that a risk score for a feature dimension has been obtained, for example, the risk score for the feature dimension of genetic test results is 0.7, the value of this dimension in the dynamic feature weight matrix is ​​0.9, and the sensitivity weight of this feature dimension is 0.9 / 0.7=1.29. Assuming that the overall privacy budget is 1.0, and the sum of the sensitivity weights of all feature dimensions is 5.0 (such as the three feature dimensions mentioned above, including genetic test results, electrocardiogram data, and clinical symptom descriptions), the privacy budget allocated to the feature dimension of genetic test results is calculated as , it is known that the noise injection intensity in the previous time period is 0.2, the momentum coefficient is 0.3, the currently allocated privacy budget is 0.258, and the last allocated privacy budget is 0.2. Substitute into the formula It can be obtained that the noise injection intensity of the feature dimension of the genetic test results in the current time period after adjustment is 0.2174. According to the privacy budget of each feature dimension calculated above and the adjusted noise injection intensity, the perturbed genetic features are noise recalibrated. For example, the genetic test result data is adjusted based on the original noise according to the new noise injection intensity of 0.2174. The electrocardiogram data and other features of the clinical symptom description are also subjected to corresponding privacy protection processing. These processed features are combined to generate an enhanced privacy feature set. This enhanced privacy feature set not only retains valuable information for the diagnosis and treatment of heart disease, but also protects the privacy of patients to a large extent. This embodiment can realize the dynamic adjustment of privacy protection intensity and optimize the privacy protection strategy in combination with historical data.

[0121] Furthermore, under the constraints of the dynamic feature weight matrix, a target feature subset is searched within the privacy-enhanced feature set, and a composite feature expression is generated from the target feature subset through genetic programming to obtain a diagnosis and treatment decision feature set; the method includes the following steps:

[0122] A multi-objective optimization algorithm is used to maximize the clinical utility of features and minimize privacy risks, and non-dominated solutions are searched in the feature space corresponding to the privacy-enhancing feature set to generate a Pareto frontier solution set.

[0123] Calculating the contribution of each feature to clinical utility and privacy risk under different feature subset combinations of the Pareto front solution set, and using the contribution value as the Shapley value; eliminating redundant features whose Shapley value is lower than a third preset threshold to obtain retained features;

[0124] The retained features are nonlinearly combined through genetic programming to generate a composite feature expression that satisfies the operator combination generated by symbolic regression, and the feature set corresponding to the composite feature expression is used as a diagnosis and treatment decision feature set.

[0125] Specifically, within a cardiology hospital's diagnostic and treatment feature set for heart failure patients, a multi-objective optimization algorithm, such as NSGA-II, was employed to search for non-dominated solutions within the feature space corresponding to the privacy-enhancing feature set, with the goal of maximizing the clinical utility of the features (e.g., accurately determining the severity of heart failure) while minimizing privacy risks. Some feature subsets may be highly effective in accurately determining the severity of heart failure, but carry a high privacy risk; others may be less so. The algorithm identifies non-dominated solutions that strike a good balance between clinical utility and privacy risks, generating a Pareto front solution set. The contribution (Shapley value) of each feature to clinical utility and privacy risk is calculated for different combinations of feature subsets within the Pareto front solution set. Assuming that feature A (e.g., a less relevant blood marker) has a Shapley value of 0.005 within a feature subset, which is lower than a third preset threshold (e.g., 0.01), it is deemed redundant and removed, resulting in the retained features. This streamlines the feature set and improves analysis efficiency. Genetic programming is then used to perform nonlinear combinations of the retained features, generating a composite feature expression that satisfies the operator combinations generated by symbolic regression.

[0126] For example, operators can combine subsets of gene sequence features with subsets of clinical indicator features. The expression levels of genes associated with heart disease can be combined with cardiac function indicators (such as ejection fraction) to generate a composite feature expression such as "Gene A expression * ejection fraction + Gene C expression / average heart rate." Correspondingly, the diagnostic and treatment decision feature set for heart diseases such as heart failure includes the expression levels of Gene A and Gene C, ejection fraction, and average heart rate. When processing diagnostic and treatment data for heart diseases like heart failure, matching features are extracted from a large amount of patient data, such as the expression levels of Gene A and Gene C, as well as ejection fraction and average heart rate. These features are combined and calculated according to the composite feature expression to generate a comprehensive feature value. This feature set integrates key information from multiple sources and contains richer and more relevant information than a single feature or a simple feature set. Doctors can use this diagnostic and treatment decision feature set, combined with machine learning models or other data analysis methods, to more accurately diagnose patients' conditions. For example, when judging the severity of heart failure in patients, comprehensive feature values ​​are input into the model. The model then outputs a judgment on the severity of the patient's condition based on the patterns and rules learned during training, assisting doctors in formulating personalized treatment plans, such as whether drug dosage needs to be adjusted or whether surgical treatment is appropriate.

[0127] In addition, diagnosis and treatment decision feature sets have other benefits. For example, heart failure is a chronic disease requiring long-term follow-up and management. Diagnosis and treatment decision feature sets can help doctors continuously monitor patient conditions. By comparing the latest examination data with historical diagnosis and treatment decision feature sets, doctors can evaluate treatment effectiveness and determine whether the condition is stable, worsening, or improving. By comparing changes in ejection fraction, gene expression, and cardiac function indicators over time, timely adjustments to treatment plans can be made. Diagnosis and treatment decision feature sets based on historical medical data are valuable research resources. Medical researchers can conduct large-scale data analysis and research to explore the pathogenesis of heart failure, the relationship between genetic factors and disease progression, and the long-term effects of different treatments. Analyzing large amounts of patient data can help discover new therapeutic targets, optimize existing treatment options, and promote medical advancements. The patterns and patterns found in diagnosis and treatment decision feature sets based on historical patient medical data are also valuable references for the diagnosis and treatment of new patients. When encountering a new heart failure patient, doctors can compare the new patient's characteristics with similar cases in the historical diagnosis and treatment decision feature sets to quickly identify potential diagnostic directions and treatment strategies, improving the efficiency and accuracy of diagnosis and treatment. In summary, the potential value of multimodal medical data has been explored while taking into account the protection of patient privacy.

[0128] This solution achieves an organic integration of value mining and privacy protection from the following perspectives: Endogenous privacy protection: In core links such as feature extraction (step S1), time alignment (step S3), and spatial mapping (step S4), privacy protection operations such as differential privacy perturbation and homomorphic encryption are directly embedded in the data processing process (such as adding Laplace noise to genetic features and encrypting clinical identifiers), rather than adding a protection layer afterwards; Dynamic balance mechanism: Through the privacy budget monitor, the information entropy and sensitivity of the feature dimension are evaluated in real time, and a dynamic inverse relationship between privacy budget allocation and data value density is established, so that high-value features receive more privacy resource investment; Risk-aware modeling: When constructing a heterogeneous graph network (step S5), the feature association strength matrix is ​​used to simultaneously evaluate clinical utility and privacy risk, so that the final generated diagnosis and treatment decision feature set is naturally privacy-safe.

[0129] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. The above is only the preferred implementation method of this application. It should be pointed out that due to the limitations of textual expression, there are objectively infinite specific structures. For ordinary technicians in this technical field, without departing from the principles of the present invention, they can also make several improvements, modifications or changes, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes or combinations, or the direct application of the inventive concept and technical solution to other occasions without improvement, should be regarded as the scope of protection of this application.

Claims

1. A data analysis method in the process of building high-value medical data assets, characterized in that: The following steps are involved: Extract effective gene loci related to the target disease and generate gene feature subsets; Extract clinical indicators associated with the target disease and generate a clinical feature subset; extract quantitative features of the lesion area and generate an imaging feature subset; The gene feature subset is encoded in the time dimension to generate a gene expression time series vector; the clinical feature subset is timestamped to extract the clinical event time series features; the image feature subset is registered in three-dimensional space to extract the spatiotemporal dynamic features of the lesion area; aligning the gene expression time series of the gene expression time series vector and the clinical event time series of the clinical event time series feature, and calculating the time alignment similarity; Based on the spatiotemporal dynamic characteristics, the image lesion coordinates are obtained, the image lesion coordinates are mapped to the gene-clinical joint space, and the spatial correlation is calculated; Constructing a heterogeneous graph network including gene nodes, clinical nodes, and imaging nodes, and calculating edge weights between nodes based on the time alignment similarity and the spatial correlation; In the heterogeneous graph network, redundant edges corresponding to the edge weights lower than a first preset threshold are pruned to generate a spatiotemporal correlation graph, thereby obtaining a correlation mapping matrix.

2. The data analysis method for constructing high-value medical data assets according to claim 1, characterized in that: The generating of the spatiotemporal correlation map and obtaining the correlation mapping matrix comprises the following steps: Converting the spatiotemporal correlation graph into a weighted adjacency matrix, wherein the matrix element values ​​of the weighted adjacency matrix are the weights of the corresponding edges; Performing symmetric normalization processing on the weighted adjacency matrix to obtain a normalized matrix; Performing singular value decomposition on the normalized matrix, retaining the first k principal components, and generating a low-dimensional mapping matrix as the association mapping matrix.

3. The data analysis method for constructing high-value medical data assets according to claim 2, characterized in that: The step of performing singular value decomposition on the normalized matrix, retaining the first k principal components, and generating a low-dimensional mapping matrix as the association mapping matrix comprises the following steps: Performing singular value decomposition on the normalized matrix, retaining the first k principal components, generating a low-dimensional mapping matrix, and generating feature clusters based on the low-dimensional mapping matrix; The cluster labels of the feature clusters are fed back to the singular value decomposition as auxiliary information, and the truncation rank k value in the singular value decomposition process is iteratively optimized to obtain the association mapping matrix.

4. The data analysis method for constructing high-value medical data assets according to claim 1, characterized in that: The method of extracting effective gene loci related to the target disease comprises the following steps: Using the chi-square test, the SNP sites whose occurrence frequency in the population exceeds the second preset threshold are screened to obtain a candidate SNP list; Obtain a GWAS database, compare the candidate SNP list with the GWAS database, and retain SNP sites with a p-value less than a third preset threshold as valid gene sites; wherein the third preset threshold is less than 0.

05.

5. The data analysis method for constructing high-value medical data assets according to claim 1, characterized in that: The step of encoding the gene feature subset in a time dimension to generate a gene expression time series vector comprises the following steps: For moments with missing gene expression, cubic spline interpolation was used to fill in the gaps; Multi-scale temporal features of gene expression are extracted by wavelet transform, wherein the multi-scale temporal features include high-frequency fluctuation components and low-frequency trend components.

6. The data analysis method for constructing high-value medical data assets according to claim 1, characterized in that: The step of aligning the gene expression timing of the gene expression timing vector with the clinical event timing of the clinical event timing feature comprises the following steps: Performing dynamic time warping on the gene expression time series and the clinical event time series to generate a time-aligned path; Calculating the Pearson correlation coefficient of the gene-clinical feature pairs along the time alignment path, and screening feature pairs with correlation coefficients greater than a fourth preset threshold as strongly associated feature pairs; Based on the strongly correlated feature pairs, constructing a gene-clinical initial correlation matrix, wherein the matrix elements of the initial correlation matrix are correlation coefficient values; Performing non-negative matrix decomposition on the initial correlation matrix to obtain a gene feature subspace and a clinical feature subspace; The gene feature subspace and the clinical feature subspace are orthogonally projected and merged to generate the gene-clinical joint space.

7. The data analysis method for constructing high-value medical data assets according to claim 4, characterized in that: The method of extracting effective gene loci related to the target disease and generating a gene feature subset includes the following steps: Performing random projection dimensionality reduction on the expression time series vectors of the screened effective gene loci to generate a feature space after dimensionality reduction; Add Laplace noise to the feature space after dimensionality reduction to satisfy -Differential privacy constraints are applied to obtain a subset of genetic features.

8. The data analysis method for constructing high-value medical data assets according to claim 7, characterized in that: The method further comprises the following steps: When constructing the edge weights of a heterogeneous graph network, noise compensation is performed on a subset of gene features to obtain corrected weights, which are then used to generate spatiotemporal association maps.

9. The data analysis method for constructing high-value medical data assets according to claim 1, characterized in that: The step of extracting clinical indicators associated with the target disease and generating a clinical feature subset comprises the following steps: Extract clinical indicators associated with the target disease and homomorphically encrypt the patient identifier to obtain a subset of clinical features; The feature similarity between genetic features and encrypted clinical features is calculated in the encrypted domain, and a clinical feature subset is output.

10. The data analysis method in the process of constructing high-value medical data assets according to claim 9, characterized in that: After homomorphically encrypting the patient identifier, the following steps are further included: Correlation coefficients between gene signatures and subsets of clinical features were calculated in the encrypted domain; In the orthogonal projection of the joint gene-clinical space, a subset of clinical features is locally decrypted, exposing only the standardized feature values ​​to participate in the space merging.

Images