A blockchain-based method and system for sharing data in multi-center clinical trials

Abstract

This invention discloses a blockchain-based method and system for sharing multi-center clinical trial data, including: participating centers acquiring subject data and inclusion / exclusion criteria; deploying smart contracts according to the inclusion / exclusion criteria, and performing inclusion / exclusion verification based on a hierarchical verification tree to obtain verification data; performing deviation analysis on the verification data based on third-party benchmark data, classifying data with deviation scores below a global threshold as high-quality data and those above a global threshold as data to be screened; jointly evaluating the data to be screened based on the consistency of feature topology and deviation direction to obtain rare data; performing asymmetric soft mapping on high-quality data and rare data based on differentiated soft label decoupling to obtain a set of differentiated local soft labels; obtaining several shared groups and corresponding global soft labels through a rare-added grouping and aggregation method; using the global soft labels as supervision signals, generating a lightweight model through confidence-based fidelity distillation; inputting the data to be shared into the lightweight model, and outputting the sharing results.

Description

Technical Field

[0001] This invention relates to the field of multicenter clinical trial data sharing, and in particular to a blockchain-based method and system for sharing multicenter clinical trial data. Background Technology

[0002] Multicenter clinical trials are an important means of verifying the safety and efficacy of drugs and medical devices. Sharing clinical trial data is of great significance for improving statistical efficiency and accelerating the research and development process.

[0003] However, in existing technologies, the data quality of participating centers varies, and the inclusion and exclusion criteria are often checked independently without considering the logical dependencies between criteria, resulting in low verification efficiency and a lot of redundant operations. For outlier samples that deviate from the baseline data, existing methods usually directly remove them, failing to effectively identify rare cases with clinical value and wasting data resources. In the process of multi-center collaborative modeling, traditional federated learning schemes often use hard label aggregation, which is difficult to reflect the subtle differences in data distribution and pays insufficient attention to rare samples, resulting in limited model generalization ability. In the model distillation stage, existing technologies lack quantitative evaluation of teacher signal confidence and cross-center consistency, making it difficult for student models to fully learn global knowledge. Therefore, there is an urgent need for a multi-center clinical trial data sharing method that takes into account data quality verification, rare sample preservation, differential soft label aggregation, and credible evidence storage, so as to improve data utilization efficiency and model sharing performance. Summary of the Invention

[0004] The purpose of this invention is to provide a blockchain-based method and system for sharing multi-center clinical trial data.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: The participation center acquires subject data and inclusion / exclusion criteria, the subject data including clinical test indicators and clinical category labels; deploys a smart contract according to the inclusion / exclusion criteria, the smart contract verifies the subject data through layer-by-layer inclusion / exclusion verification based on a hierarchical verification tree, and the subject data that passes the verification is used as verification data. Based on third-party benchmark data, deviation analysis is performed on the verification data to obtain a deviation score. The verification data with a deviation score below the global threshold is regarded as high-quality data, and the verification data with a deviation score above the global threshold is regarded as data to be identified. The data to be identified is jointly evaluated based on the consistency between feature topology and deviation direction to obtain rare data. Based on the differentiated soft tag decoupling generation method, asymmetric soft mapping is performed on the high-quality data and the rare data to obtain a differentiated local soft tag set; based on the differentiated local soft tag set, several shared groups and corresponding global soft tags are obtained through a rare-bond-based grouping and aggregation method, and the global soft tags are stored on the blockchain for evidence. The participation center of the sharing group obtains the global soft label, uses the global soft label as a supervision signal, and generates a lightweight model through confidence-based fidelity distillation; the data to be shared is input into the lightweight model, and the sharing result is output.

[0006] Furthermore, the method for verifying the subject data by the smart contract through layer-by-layer in-out verification based on a hierarchical verification tree includes: The hierarchy of the inclusion and exclusion criteria is determined based on the logical dependencies between the inclusion and exclusion criteria. The inclusion and exclusion criteria without dependencies are set as the first level, the criteria that depend on the first level inclusion and exclusion criteria are set as the second level, and so on, to construct a hierarchical validation tree. The subject data is validated layer by layer according to the hierarchical validation tree, starting from the first level. Once all nodes in the current level pass the validation, the data proceeds to the next level. If any node in the current level fails the validation, all child nodes of the node that failed the validation are automatically skipped and no further validation is performed. The node that failed the validation and the skipped child nodes are marked as inapplicable. Subject data that passes the validation of all levels are used as validation data. Furthermore, the method for performing deviation analysis on the verification data based on third-party benchmark data includes: The layer depth k of the node where each verification data is located is determined based on the verification path depth in the hierarchical verification tree, and the corresponding field penalty coefficient is determined by an S-shaped curve: in, Let be the field penalty coefficient for the i-th validation data; The preset maximum penalty coefficient has a value range of 1.5 ≤ ≤3; This is the steepness coefficient, with a value range of 0.5 ≤ ≤2; k is the node level depth; For the depth of the transition, 1≤ ≤4; The deviation of each validation data point from the third-party benchmark dataset is calculated using the weighted Z-score method, with the field penalty coefficient used as the depth weight to obtain the deviation score. The Z-score method is used to calculate the deviation score of historical data in the third-party benchmark dataset. The maximum value of the deviation score sequence is used as the global threshold T. When the deviation score is less than the global threshold, the current verification data is determined to be high-quality data. When the deviation score is greater than or equal to the global threshold, the current verification data is determined to be data to be screened. Furthermore, the method for jointly evaluating the data to be identified based on the consistency of feature topology and deviation direction includes: Consistency in deviation direction includes: calculating the standardized deviation vector of each data point to be screened relative to a third-party benchmark dataset. Let be the standardized deviation component of the j-th feature dimension of the standardized deviation vector; The rarity coefficient of each data point to be screened is calculated based on the standardized deviation vector: in, Let M be the rarity coefficient of the i-th data to be identified; M is the total number of feature dimensions. The feature topology includes: in the feature space, taking each data point to be identified as the center, calculating the distance between each data point and all other data points using Euclidean distance, selecting K neighboring data points in ascending order of distance to form a local neighborhood, where 5≤K≤20; calculating the covariance matrix of the data points within the local neighborhood, and using the ratio of the largest to the smallest eigenvalue of the covariance matrix as the condition number N; and calculating the value coefficient of each data point to be identified based on the condition number. in, The value coefficient of the i-th piece of data to be identified; when ≥Preset rarity threshold P and When the data is greater than or equal to the preset reliability threshold Q, it is considered valuable rare data and marked; otherwise, it is discarded. Wherein, P is the 95th quantile of the deviation direction rarity distribution of historical normal samples in the third-party benchmark dataset, and P≥0.6; Q is the 5th quantile of the local structure reliability distribution of historical normal samples in the third-party benchmark dataset, and Q≥0.5. Furthermore, the method for performing asymmetric soft mapping on the high-quality data and the rare data based on the differential soft label decoupling generation method includes: The clinical test indicators obtained from high-quality data constitute the statistical feature vector of each high-quality data point; the high-quality data are grouped by clinical category, and the mean and standard deviation of each group of data in each clinical test indicator dimension are calculated as the standard distribution reference for the corresponding clinical category. Mahalanobis distance is used to calculate the similarity between the statistical feature vector of high-quality data and the standard distribution reference. The similarity is then mapped to a probability distribution to obtain conventional soft labels. Clinical test indicators from rare data are obtained to form the statistical feature vector of each rare data point. The covariance matrix is ​​calculated based on the statistical feature vectors of all rare data points, and the ratio of the largest to the smallest eigenvalue of the covariance matrix is ​​used as the rare data condition number. The interpolation coefficients for shrinkage estimation are determined based on the rare data condition number, and an adaptive covariance matrix is ​​constructed using these coefficients. Mahalanobis distance is calculated based on the adaptive covariance matrix to obtain the similarity between the statistical feature vector of each rare data point and the standard distribution reference. This similarity is mapped to a probability distribution, and rare labels are embedded in the probability distribution to obtain soft labels carrying rare labels. A differentiated set of local soft tags consists of regular soft tags and soft tags carrying rare markers; Furthermore, the generation of several shared groups and corresponding global soft tags through the grouping and aggregation method based on rare additive features includes: Based on the local soft tag set of the participation center, the central probability distribution vector of the participation center and the rare probability distribution vector of the soft tag carrying the rare tag are statistically analyzed. The central probability distribution vector and the rare probability distribution vector are concatenated to obtain the additive distribution vector. The clinical category distribution similarity between any two participating centers is calculated using total variation distance based on the additive distribution vector; the participating centers are divided into G consensus groups using graph clustering based on clinical category similarity; the local consensus degree between every two participating centers within a consensus group is calculated, and the expression for the local consensus degree of the i-th participating center is: in, Let g be the set of participating centers within the g-th consensus group. The total number of centers within the group. The similarity of clinical category distribution between the i-th participating center and the j-th participating center; Participating centers within a group are selected based on local consensus, and those reaching a consensus threshold are designated as shared centers. The consensus threshold ranges from 0.45 to 0.65. The weighted aggregation of the distribution vectors of the shared centers is then performed based on local consensus to generate a global soft label for each group. Furthermore, the method for generating a lightweight model through confidence-based fidelity distillation includes: Using the local model of the participation center as the student model and the global soft label of the shared group as the teacher signal, the distribution difference between the predicted probability distribution of the student model and the probability distribution of the global soft label is calculated by KL divergence and used as the original distillation loss. The normalized entropy value of each global soft tag is calculated using Shannon entropy. ; Obtain the standard deviation among all local soft tags in the shared centers under each global soft tag, as the cross-center variance. ; Generate confidence coefficients based on normalized entropy values ​​and cross-center variance: in, Let α be the confidence level of the i-th data point; α and β are weighting coefficients, and α + β = 1. The weighted distillation loss is obtained by multiplying the original distillation loss of each sample by the corresponding confidence coefficient. Retrieve the local soft labels of all shared centers under the global soft label carrying rare markers, and calculate the cross-center variance. Fidelity-preserving factor is generated based on cross-center variance: in, To predetermine the rare base bonus constant, 1.2 ≤ ≤1.8; This is the magnification factor, 0.5 ≤ ≤2; The weighted distillation loss of the global soft tag carrying rare markers is weighted a second time using a fidelity enhancement factor to obtain the enhanced fidelity loss; The weighted distillation loss of regular soft labels is added to the enhanced fidelity loss of rare soft labels to obtain the total distillation loss; the student model is iteratively trained using the total distillation loss to output a lightweight model. Furthermore, the output sharing results include: The data to be shared is input into the lightweight model, which outputs a prediction probability vector corresponding to the data to be shared. The prediction probability vector contains the probability value of the data to be shared belonging to each clinical category; the clinical categories include full-term birth, late preterm birth, and early preterm birth. Secondly, a blockchain-based multi-center clinical trial data sharing system includes: Data Acquisition and Verification Module: This module is used to participate in the acquisition of subject data and inclusion / exclusion criteria by the center. The subject data includes clinical test indicators and clinical category labels. The module deploys a smart contract based on the inclusion / exclusion criteria. The smart contract verifies the subject data through layer-by-layer inclusion / exclusion verification based on a hierarchical verification tree. The subject data that passes the verification is used as the verification data. The data filtering module is used to perform deviation analysis on the verification data based on third-party benchmark data to obtain a deviation score. The verification data with a deviation score below the global threshold is regarded as high-quality data, and the verification data with a deviation score above the global threshold is regarded as data to be identified. The data to be identified is jointly evaluated based on the consistency between feature topology and deviation direction to obtain rare data. Grouping and aggregation module: used to perform asymmetric soft mapping on the high-quality data and the rare data based on the differentiated soft tag decoupling generation method to obtain a differentiated local soft tag set; based on the differentiated local soft tag set, several shared groups and corresponding global soft tags are obtained through the grouping and aggregation method based on rare addition, and the global soft tags are stored on the blockchain for evidence; Distillation training module: The participation center of the sharing group obtains the global soft label, uses the global soft label as a supervision signal, and generates a lightweight model through confidence-based fidelity distillation; the data to be shared is input into the lightweight model, and the sharing result is output.

[0007] The beneficial effects of this invention are: This invention is a blockchain-based method and system for sharing multi-center clinical trial data. Compared with existing technologies, this invention has the following technical advantages: This invention achieves layer-by-layer logical verification of ingress and exclusion criteria by constructing a hierarchical verification tree, automatically skipping failed sub-nodes based on dependencies, significantly improving verification efficiency. It utilizes deviation analysis and joint evaluation mechanisms to effectively identify rare clinically valuable data while retaining high-quality data, avoiding data resource waste. Differential soft label decoupling and rare-addition component aggregation reflect subtle differences in data distribution and strengthen the focus on rare samples, improving model generalization ability. Through confidence-based fidelity distillation to generate a lightweight model, it quantifies the credibility of teacher signals and cross-center consistency, ensuring the lightweight model fully learns global knowledge and achieves reliable and efficient sharing of multi-center data. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the steps of a blockchain-based multi-center clinical trial data sharing method according to the present invention. Detailed Implementation

[0009] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0010] This invention discloses a blockchain-based method and system for sharing multi-center clinical trial data, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: The participation center acquires subject data and inclusion / exclusion criteria, the subject data including clinical test indicators and clinical category labels; deploys a smart contract according to the inclusion / exclusion criteria, the smart contract verifies the subject data through layer-by-layer inclusion / exclusion verification based on a hierarchical verification tree, and the subject data that passes the verification is used as verification data. In this embodiment, five participating centers, designated as Center A, Center B, Center C, Center D, and Center E, shared clinical trial data on the assessment of neurological development in preterm infants. Each participating center extracted subject-source data from its local electronic data acquisition system. Clinical test indicators included gestational age, birth weight, 1-minute Apgar score, 5-minute Apgar score, umbilical artery blood gas pH, diagnostic markers for gestational hypertension and gestational diabetes, records of prenatal glucocorticoid use, and delivery mode coding, totaling nine dimensions. Clinical category labels, confirmed by clinical experts, were divided into three categories: full-term (gestational age not less than 37 weeks), late preterm (gestational age between 34 and 37 weeks), and early preterm (gestational age less than 34 weeks). The data volume for each center was 185, 162, 198, 143, and 176, respectively, totaling 864 data points. There are 12 inclusion and exclusion criteria, and a hierarchical validation tree is constructed based on logical dependencies. The first level consists of independent criteria, including subjects being at least 18 years old and having signed informed consent forms by their legal guardians or themselves. The second level depends on the first level, including three criteria: gestational age less than 37 weeks, singleton pregnancy, and no diagnosis of chorioamnionitis within 24 hours before delivery. The third level depends on the second level, including three criteria: no major congenital malformations, no severe perinatal asphyxia (i.e., an Apgar score of at least 4 points at 5 minutes), and no congenital metabolic diseases. The fourth level depends on the third level, including two criteria: the first neurobehavioral assessment completed within 72 hours after birth and no severe intracranial hemorrhage (i.e., grade III or higher intracranial hemorrhage). The validation tree has a depth of 4 levels. The smart contract is deployed on the consortium blockchain, with each participating center deploying an independent contract instance. Taking the data of subject number 37 from center A as an example, the gestational age is 32 weeks, it is a singleton pregnancy, the informed consent form has been signed, the 5-minute Apgar score is 6, and there are no congenital metabolic diseases; both items in the first level of verification are passed. Entering the second level, the gestational age is less than 37 weeks, the singleton pregnancy is passed, and there is no diagnosis of chorioamnionitis within 24 hours before delivery. Entering the third level, there are no major congenital malformations, no severe perinatal asphyxia, and no congenital metabolic diseases. Entering the fourth level, the first neurobehavioral assessment is completed within 72 hours after birth, and there is no severe intracranial hemorrhage. If all levels of verification are passed, this data is output as verification data to center A's local verification database. Taking the data of subject number 124 in center C as an example, the informed consent form was not signed; the verification of the informed consent form signing item in the first level failed, and all sub-level nodes that depend on this item, i.e., all nodes from the second to the fourth level, were automatically skipped and marked as inapplicable. The verification log recorded the failure of this item and the skipping of all its descendant nodes; this data did not enter the subsequent process and remained in the local isolation area for manual review; after the hierarchical verification tree verification, out of a total of 864 data from the 5 centers, 721 passed the verification and 143 failed the verification; the verification status, timestamp, and hash value of the verification result of each node were broadcast to the blockchain network for evidence storage after being signed by the private key of each center; Based on third-party benchmark data, deviation analysis is performed on the verification data to obtain a deviation score. The verification data with a deviation score below the global threshold is regarded as high-quality data, and the verification data with a deviation score above the global threshold is regarded as data to be identified. The data to be identified is jointly evaluated based on the consistency between feature topology and deviation direction to obtain rare data. In this embodiment, the third-party benchmark dataset is derived from the China Newborn Follow-up Network Database and the associated China Newborn Collaborative Network Outcome Database, which contain high-risk premature infant follow-up data that has passed quality control review. In this embodiment, 1,200 clinical test indicators consistent with the dimensions of the subject data are extracted from the associated database, including the same 9 clinical test indicator dimensions as the subject data. For each piece of validation data, first determine its field penalty coefficient; taking data number 37 in center A as an example, this validation data passed all 4 levels of validation, and the node level depth is 4; preset the maximum penalty coefficient. Set the steepness coefficient to 2. Take 1, depth of transition The value is set to 2, and based on the S-curve, the field penalty coefficient for this data is determined to be 2.762. The deviation score was calculated using the weighted Z-score method: the mean and standard deviation of each dimension in the third-party benchmark dataset were calculated; the deviation score of data point 37 in center A was calculated as 7.83 by comparing the values ​​of each dimension with the benchmark mean and using the weighted Z-score method; the deviation score of 1200 historical data points in the third-party benchmark dataset was calculated using the same method, and the maximum value of the score sequence was 6.21; since 7.83 > 6.21, data point 37 in center A was determined to be data to be screened; the deviation score was calculated for all 721 verification data points, of which 467 data points with deviation scores below the global threshold were determined to be high-quality data, and 254 data points with deviation scores above the global threshold were data to be screened. A joint evaluation was conducted on 254 data points to be screened. Taking data point No. 37 from center A as an example, the standardized deviation vector relative to the third-party benchmark dataset was calculated. Based on the standardized deviation vector, the rarity coefficient was calculated to be 0.825. In the feature space, with this data point as the center, 15 nearest neighbors were selected using Euclidean distance to form a local neighborhood, and the condition number was calculated to be 3.32. The median of the condition number of historical normal samples in the third-party benchmark dataset was 2.85. Based on this median, the value coefficient was calculated, and the value coefficient of this data point was 0.312. The rarity threshold is taken as the 95th percentile of the distribution of the rarity coefficients of historical normal samples in the third-party benchmark dataset, and the 1140th value after sorting from smallest to largest is 0.78; the reliability threshold is taken as the 5th percentile of the distribution of the value coefficients of historical normal samples, and the 60th value is 0.52; the rarity coefficient 0.825 > 0.78, but the value coefficient 0.312 < 0.52, which does not meet both conditions at the same time, so it is removed. The above joint evaluation was performed on all 254 data entries to be screened. Ultimately, 196 entries did not meet the requirements and were removed. 58 entries met both threshold conditions and were marked as valuable rare data. Among the 58 rare data entries, there were 12 entries from Center A, 9 from Center B, 16 from Center C, 8 from Center D, and 13 from Center E. The 467 high-quality data entries and the 58 rare data entries proceeded to the next steps. Based on the differentiated soft tag decoupling generation method, asymmetric soft mapping is performed on the high-quality data and the rare data to obtain a differentiated local soft tag set; based on the differentiated local soft tag set, several shared groups and corresponding global soft tags are obtained through a rare-bond-based grouping and aggregation method, and the global soft tags are stored on the blockchain for evidence. In this embodiment, conventional soft labels are first generated for 467 high-quality data points; taking center A as an example, there are a total of 89 high-quality data points; grouped by clinical category: 33 full-term births, 31 late preterm births, and 25 early preterm births; for each category, the mean and standard deviation of 9 clinical test indicator dimensions are calculated as the standard distribution reference for that category; Taking a high-quality data point from the full-term birth category in Center A as an example, the feature vector is: gestational age 39 weeks, birth weight 3320 grams, 1-minute Apgar score 9, 5-minute Apgar score 9, pH value 7.32, no gestational hypertension, no gestational diabetes, no use of glucocorticoids, and vaginal delivery. The calculated Mahalanobis distance of the data point relative to the standard distribution of full-term births is 0.87, the Mahalanobis distance relative to the standard distribution of late preterm births is 4.21, and the Mahalanobis distance relative to the standard distribution of early preterm births is 8.56. The temperature parameter is set to 3.0, and the probability distribution is obtained after Softmax normalization: 0.82 for full-term births, 0.15 for late preterm births, and 0.03 for early preterm births. This is used as the regular soft label for this data point. Generate soft labels carrying rare markers for 58 rare data points: Stack all rare data points into a 58x9 feature matrix, calculate its covariance matrix, and the ratio of the largest eigenvalue to the smallest eigenvalue, i.e., the rare data condition number k, is 7.68; adjust parameters... Take 3.0, according to Obtain the shrinkage interpolation coefficients The value is 0.626. An adaptive covariance matrix is ​​constructed. Taking a rare data point from center A as an example, the Mahalanobis distance is calculated based on the adaptive covariance matrix. After Softmax normalization with a temperature parameter of 3.0, the probability distribution is 0.45 for full-term birth, 0.30 for late preterm birth, and 0.25 for early preterm birth. Rare identifier bits are added to form a soft label carrying a rare marker. Each center independently executes the above process to obtain a differentiated local soft label set for each center. Then, grouping and aggregation are performed: taking the differentiated local soft tag sets of each of the 5 participating centers as input, for each center, the average probability of each category of its regular soft tags is calculated as the center probability distribution vector, and the average probability of each category of its rare soft tags is calculated as the rare probability distribution vector. The two are concatenated to form an additive distribution vector with a dimension of 6. The additive distribution vectors of the 5 centers are as follows: center A is [0.58, 0.27, 0.15, 0.48, 0.15, 0.27, 0.15 ... Center B is [0.62, 0.25, 0.13, 0.52, 0.28, 0.20]; Center C is [0.45, 0.33, 0.22, 0.35, 0.38, 0.27]; Center D is [0.55, 0.29, 0.16, 0.50, 0.27, 0.23]; Center E is [0.48, 0.31, 0.21, 0.40, 0.35, 0.25]; Calculate the total variation distance between any two centers, and subtract the total variation distance from 1 to obtain the similarity. For example, the total variation distance between center A and center B is 0.08, and the similarity is 0.92; the similarity between center A and center C is 0.71; the similarity between center B and center C is 0.69; and the similarity between center D and center E is 0.89. Construct an undirected weighted graph using the similarity as the edge weight, and partition it using the spectral clustering algorithm, with a consensus group size of 2. The partitioning result is that consensus group 1 contains center A, center B, and center D, and consensus group 2 contains center C and center E. Taking consensus group 1 as an example, the set of centers within the group includes three centers: A, B, and D. The local consensus degree of center A is the average of its similarity to centers B and D, which is 0.90. The local consensus degree of center B is 0.89, and the local consensus degree of center D is 0.87. The consensus threshold is 0.55. All three centers reach the threshold and are considered as shared centers. The global soft label is the weighted average of the distribution vectors of each center according to their local consensus degree, calculated as [0.55, 0.28, 0.17, 0.48, 0.30, 0.22]. In consensus group 2, the local consensus degree of centers C and E is 0.78, both reaching the consensus threshold. The global soft labels are [0.46, 0.32, 0.22, 0.37, 0.37, 0.26]. The two sets of global soft labels, along with the group identifier, the list of participating centers, and the timestamp, are submitted to the blockchain network for notarization after being multi-signed by more than 2 / 3 of the centers within the group. The participation center of the sharing group obtains the global soft label, uses the global soft label as a supervision signal, and generates a lightweight model through confidence-based fidelity distillation; the data to be shared is input into the lightweight model, and the sharing result is output. In this embodiment, the three participating centers of consensus group 1, namely center A, center B and center D, read global soft tags from the blockchain; each center locally deploys a student model as a 3-layer fully connected network, with an input layer dimension of 9, hidden layers of 64 and 32 dimensions respectively, and an output layer of 3 dimensions, using ReLU as the activation function; Taking a regular global soft label of center A as an example, the teacher labels are [0.55, 0.28, 0.17], the student model predicted distribution is [0.52, 0.31, 0.17], and the KL divergence is calculated as the original distillation loss, with a loss value of 0.002; The information entropy of the global soft tag is calculated to be 0.986, the total number of categories is 3, and the normalized entropy value is 0.897; the cross-center variance is calculated as the mean of the standard deviations of the local soft tags of the three shared centers under this global soft tag, which is 0.058; the weighting coefficients α and β are 0.6 and 0.4, respectively, and the confidence coefficient is 0.571; the weighted distillation loss is the confidence coefficient multiplied by the original distillation loss, which is 0.00114. For the global soft tag portion carrying rare markers, i.e., the last three dimensions of the global soft tag [0.48, 0.30, 0.22], the local rare soft tags corresponding to the three shared centers are obtained, and the cross-center variance is calculated to be 0.084; the mean of the cross-center variance of the regular global soft tag is 0.052; the preset rare base enhancement constant is 1.5, the amplification factor is 1.0, and the fidelity enhancement factor is calculated to be 5.42; the enhanced fidelity loss is the fidelity enhancement factor multiplied by the weighted distillation loss, i.e., 0.00618; The total loss is obtained by adding the weighted distillation loss of regular soft labels to the enhancement fidelity loss of rare soft labels. The Adam optimizer is used with a learning rate of 0.001 and a batch size of 32. Iterative training is performed on 89 high-quality data points and 12 rare data points in center A. Training stops when the validation set loss no longer decreases for 10 consecutive rounds. After about 120 rounds, the lightweight model of center A is output. Centers B and D simultaneously perform the above distillation training process and each outputs its own local lightweight model. The sharing effect was evaluated using 180 independent test samples, with each center reserving 60 samples for each clinical category that were not used in the training. The data to be shared was input into a lightweight model, which outputs a three-dimensional probability vector. For example, the predicted probability vector for a certain data point to be shared is [0.78, 0.15, 0.07], indicating that the data has a 78% probability of being a full-term birth, a 15% probability of being a late preterm birth, and a 7% probability of being an early preterm birth. After obtaining the prediction results, participating centers in the sharing group can make clinical decisions based on the full-term birth category corresponding to the highest probability value. Based on all 180 test samples, the classification results of the method of this invention are as follows: Of the 60 full-term birth samples, 55 were correctly classified, 3 were misclassified as late preterm births, and 2 were misclassified as early preterm births; Of the 60 late preterm birth samples, 50 were correctly classified, 5 were misclassified as full-term births, and 5 were misclassified as early preterm births; Of the 60 early preterm birth samples, 51 were correctly classified, 4 were misclassified as full-term births, and 5 were misclassified as late preterm births; The overall classification accuracy rate was 86.7%, with an accuracy rate of 91.7% for full-term births, 83.3% for late preterm births, and 85.0% for early preterm births. A comparative experiment was conducted on the same 180 test samples: The control group used a traditional federated learning scheme, i.e., without hierarchical verification, without distinguishing between high-quality and rare data to directly generate hard labels, and without confidence-weighted fidelity distillation. The local model of each center was a 3-layer fully connected network with the same structure. The classification results of the control group were as follows: Among 60 full-term birth samples, 47 were correctly classified, 8 were misclassified as late preterm birth, and 5 were misclassified as early preterm birth; among 60 late preterm birth samples, 41 were correctly classified, 10 were misclassified as full-term birth, and 9 were misclassified as early preterm birth; among 60 early preterm birth samples, 40 were correctly classified, 9 were misclassified as full-term birth, and 11 were misclassified as late preterm birth; The overall classification accuracy of the control group was 71.1%, the accuracy for full-term birth was 78.3%, the accuracy for late preterm birth was 68.9%, and the accuracy for early preterm birth was 66.7%.

[0011] In this embodiment, the method for verifying the subject data by the smart contract through layer-by-layer in-out verification based on a hierarchical verification tree includes: The hierarchy of the inclusion and exclusion criteria is determined based on the logical dependencies between the inclusion and exclusion criteria. The inclusion and exclusion criteria without dependencies are set as the first level, the criteria that depend on the first level inclusion and exclusion criteria are set as the second level, and so on, to construct a hierarchical validation tree. The subject data is validated layer by layer according to the hierarchical validation tree, starting from the first level. Once all nodes in the current level pass the validation, the data proceeds to the next level. If any node in the current level fails the validation, all child nodes of the node that failed the validation are automatically skipped and no further validation is performed. The node that failed the validation and the skipped child nodes are marked as inapplicable. Subject data that passes the validation of all levels are used as validation data. In this embodiment, the method for performing deviation analysis on the verification data based on third-party benchmark data includes: The layer depth k of the node where each verification data is located is determined based on the verification path depth in the hierarchical verification tree, and the corresponding field penalty coefficient is determined by an S-shaped curve: in, Let be the field penalty coefficient for the i-th validation data; The preset maximum penalty coefficient has a value range of 1.5 ≤ ≤3; This is the steepness coefficient, with a value range of 0.5 ≤ ≤2; k is the node level depth; For the depth of the transition, 1≤ ≤4; The deviation of each validation data point from the third-party benchmark dataset is calculated using the weighted Z-score method, with field penalty coefficients used as depth weights to obtain the deviation score. ; The Z-score method is used to calculate the deviation score of historical data in the third-party benchmark dataset. The maximum value of the deviation score sequence is used as the global threshold T. When the deviation score is less than the global threshold, the current verification data is determined to be high-quality data. When the deviation score is greater than or equal to the global threshold, the current verification data is determined to be data to be screened. In this embodiment, the method for jointly evaluating the data to be identified based on the consistency of feature topology and deviation direction includes: Consistency in deviation direction includes: calculating the standardized deviation vector of each data point to be screened relative to a third-party benchmark dataset. Let be the standardized deviation component of the j-th feature dimension of the standardized deviation vector; The rarity coefficient of each data point to be screened is calculated based on the standardized deviation vector: in, Let M be the rarity coefficient of the i-th data to be identified; M is the total number of feature dimensions. The feature topology includes: in the feature space, taking each data point to be identified as the center, calculating the distance between each data point and all other data points using Euclidean distance, selecting K neighboring data points in ascending order of distance to form a local neighborhood, where 5≤K≤20; calculating the covariance matrix of the data points within the local neighborhood, and using the ratio of the largest to the smallest eigenvalue of the covariance matrix as the condition number N; and calculating the value coefficient of each data point to be identified based on the condition number. in, Let be the value coefficient of the i-th piece of data to be identified; when ≥Preset rarity threshold P and When the data is greater than or equal to the preset reliability threshold Q, it is considered valuable rare data and marked; otherwise, it is discarded. Wherein, P is the 95th quantile of the deviation direction rarity distribution of historical normal samples in the third-party benchmark dataset, and P≥0.6; Q is the 5th quantile of the local structure reliability distribution of historical normal samples in the third-party benchmark dataset, and Q≥0.5. In this embodiment, the method for performing asymmetric soft mapping on the high-quality data and the rare data based on the differential soft tag decoupling generation method includes: The clinical test indicators obtained from high-quality data constitute the statistical feature vector of each high-quality data point; the high-quality data are grouped by clinical category, and the mean and standard deviation of each group of data in each clinical test indicator dimension are calculated as the standard distribution reference for the corresponding clinical category. Mahalanobis distance is used to calculate the similarity between the statistical feature vector of high-quality data and the standard distribution reference. The similarity is then mapped to a probability distribution to obtain conventional soft labels. Clinical test indicators from rare data are obtained to form the statistical feature vector of each rare data point. The covariance matrix is ​​calculated based on the statistical feature vectors of all rare data points, and the ratio of the largest to the smallest eigenvalue of the covariance matrix is ​​used as the rare data condition number. The interpolation coefficients for shrinkage estimation are determined based on the rare data condition number, and an adaptive covariance matrix is ​​constructed using these coefficients. Mahalanobis distance is calculated based on the adaptive covariance matrix to obtain the similarity between the statistical feature vector of each rare data point and the standard distribution reference. This similarity is mapped to a probability distribution, and rare labels are embedded in the probability distribution to obtain soft labels carrying rare labels. A differentiated set of local soft tags consists of regular soft tags and soft tags carrying rare markers; In this embodiment, the generation of several shared groups and corresponding global soft tags through a grouping and aggregation method based on rare additive features includes: Based on the local soft tag set of the participation center, the central probability distribution vector of the participation center and the rare probability distribution vector of the soft tag carrying the rare tag are statistically analyzed. The central probability distribution vector and the rare probability distribution vector are concatenated to obtain the additive distribution vector. The clinical category distribution similarity between any two participating centers is calculated using total variation distance based on the additive distribution vector; the participating centers are divided into G consensus groups using graph clustering based on clinical category similarity; the local consensus degree between every two participating centers within a consensus group is calculated, and the expression for the local consensus degree of the i-th participating center is: in, Let g be the set of participating centers within the g-th consensus group. The total number of centers within the group. The similarity of clinical category distribution between the i-th participating center and the j-th participating center; Participating centers within a group are selected based on local consensus, and those reaching a consensus threshold are designated as shared centers. The consensus threshold ranges from 0.45 to 0.65. The weighted aggregation of the distribution vectors of the shared centers is then performed based on local consensus to generate a global soft label for each group. In this embodiment, the method for generating a lightweight model through confidence-based fidelity distillation includes: Using the local model of the participation center as the student model and the global soft label of the shared group as the teacher signal, the distribution difference between the predicted probability distribution of the student model and the probability distribution of the global soft label is calculated by KL divergence and used as the original distillation loss. The normalized entropy value of each global soft tag is calculated using Shannon entropy. ; Obtain the standard deviation among all local soft tags in the shared centers under each global soft tag, as the cross-center variance. ; Generate confidence coefficients based on normalized entropy values ​​and cross-center variance: in, Let α be the confidence level of the i-th data point; α and β are weighting coefficients, and α + β = 1. The weighted distillation loss is obtained by multiplying the original distillation loss of each sample by the corresponding confidence coefficient. Retrieve the local soft labels of all shared centers under the global soft label carrying rare markers, and calculate the cross-center variance. Fidelity-preserving factor is generated based on cross-center variance: in, To predetermine the rare base bonus constant, 1.2 ≤ ≤1.8; This is the magnification factor, 0.5 ≤ ≤2; The weighted distillation loss of the global soft tag carrying rare markers is weighted a second time using a fidelity enhancement factor to obtain the enhanced fidelity loss; The weighted distillation loss of regular soft labels is added to the enhanced fidelity loss of rare soft labels to obtain the total distillation loss; the student model is iteratively trained using the total distillation loss to output a lightweight model. In this embodiment, the output sharing result includes: The data to be shared is input into the lightweight model, which outputs a prediction probability vector corresponding to the data to be shared. The prediction probability vector contains the probability value of the data to be shared belonging to each clinical category; the clinical categories include full-term birth, late preterm birth, and early preterm birth. Secondly, a blockchain-based multi-center clinical trial data sharing system includes: Data Acquisition and Verification Module: This module is used to participate in the acquisition of subject data and inclusion / exclusion criteria by the center. The subject data includes clinical test indicators and clinical category labels. The module deploys a smart contract based on the inclusion / exclusion criteria. The smart contract verifies the subject data through layer-by-layer inclusion / exclusion verification based on a hierarchical verification tree. The subject data that passes the verification is used as the verification data. The data filtering module is used to perform deviation analysis on the verification data based on third-party benchmark data to obtain a deviation score. The verification data with a deviation score below the global threshold is regarded as high-quality data, and the verification data with a deviation score above the global threshold is regarded as data to be identified. The data to be identified is jointly evaluated based on the consistency between feature topology and deviation direction to obtain rare data. Grouping and aggregation module: used to perform asymmetric soft mapping on the high-quality data and the rare data based on the differentiated soft tag decoupling generation method to obtain a differentiated local soft tag set; based on the differentiated local soft tag set, several shared groups and corresponding global soft tags are obtained through the grouping and aggregation method based on rare addition, and the global soft tags are stored on the blockchain for evidence; Distillation training module: The participation center of the sharing group obtains the global soft label, uses the global soft label as a supervision signal, and generates a lightweight model through confidence-based fidelity distillation; the data to be shared is input into the lightweight model, and the sharing result is output.

Claims

1. A blockchain-based method for sharing multi-center clinical trial data, characterized in that, Includes the following steps: The participation center acquires subject data and inclusion / exclusion criteria, the subject data including clinical test indicators and clinical category labels; deploys a smart contract according to the inclusion / exclusion criteria, the smart contract verifies the subject data through layer-by-layer inclusion / exclusion verification based on a hierarchical verification tree, and the subject data that passes the verification is used as verification data. Based on third-party benchmark data, deviation analysis is performed on the verification data to obtain a deviation score. The verification data with a deviation score below the global threshold is regarded as high-quality data, and the verification data with a deviation score above the global threshold is regarded as data to be identified. The data to be identified is jointly evaluated based on the consistency of feature topology and deviation direction to obtain rare data. Based on the differentiated soft tag decoupling generation method, asymmetric soft mapping is performed on the high-quality data and the rare data to obtain a differentiated local soft tag set; based on the differentiated local soft tag set, several shared groups and corresponding global soft tags are obtained through a rare-bond-based grouping and aggregation method, and the global soft tags are stored on the blockchain for evidence. The participation center of the sharing group obtains the global soft label, uses the global soft label as a supervision signal, and generates a lightweight model through confidence-based fidelity distillation; the data to be shared is input into the lightweight model, and the sharing result is output.

2. The blockchain-based multi-center clinical trial data sharing method according to claim 1, characterized in that, The smart contract uses a method for validating the subject data through a layer-by-layer in-out validation based on a hierarchical validation tree, including: The hierarchy of the inclusion and exclusion criteria is determined based on the logical dependencies between the inclusion and exclusion criteria. The inclusion and exclusion criteria without dependencies are set as the first level, the criteria that depend on the first level inclusion and exclusion criteria are set as the second level, and so on, to construct a hierarchical validation tree. The subject data is validated layer by layer according to the hierarchical validation tree, starting from the first level. Once all nodes in the current level pass the validation, the data proceeds to the next level. If any node in the current level fails the validation, all sub-level nodes that depend on the failed node are automatically skipped and no further validation is performed. The failed node and the skipped sub-nodes are marked as inapplicable. Subject data that passes the validation of all levels are used as validation data.

3. The method for sharing multi-center clinical trial data based on blockchain according to claim 1, characterized in that, The method for performing deviation analysis on the verification data based on third-party benchmark data includes: The layer depth k of the node where each verification data is located is determined based on the verification path depth in the hierarchical verification tree, and the corresponding field penalty coefficient is determined by an S-shaped curve: ； in, Let be the field penalty coefficient for the i-th validation data; The preset maximum penalty coefficient has a value range of 1.5 ≤ ≤3; This is the steepness coefficient, with a value range of 0.5 ≤ ≤2; k is the node level depth; For the depth of the transition, 1≤ ≤4; The deviation of each validation data point from the third-party benchmark dataset is calculated using the weighted Z-score method, with the field penalty coefficient used as the depth weight to obtain the deviation score. The Z-score method is used to calculate the deviation score of historical data in the third-party benchmark dataset. The maximum value of the deviation score sequence is used as the global threshold T. When the deviation score is less than the global threshold, the current verification data is determined to be high-quality data. When the deviation score is greater than or equal to the global threshold, the current verification data is determined to be data to be screened.

4. The blockchain-based multi-center clinical trial data sharing method according to claim 1, characterized in that, The method for jointly evaluating the data to be identified based on the consistency of feature topology and deviation direction includes: Consistency in deviation direction includes: calculating the standardized deviation vector of each data point to be screened relative to a third-party benchmark dataset. Let be the standardized deviation component of the j-th feature dimension of the standardized deviation vector; The rarity coefficient of each data point to be screened is calculated based on the standardized deviation vector: ； in, Let M be the rarity coefficient of the i-th data to be identified; M is the total number of feature dimensions. The feature topology includes: in the feature space, taking each data point to be identified as the center, calculating the distance between each data point and all other data points using Euclidean distance, selecting K neighboring data points in ascending order of distance to form a local neighborhood, where 5≤K≤20; calculating the covariance matrix of the data points within the local neighborhood, and using the ratio of the largest to the smallest eigenvalue of the covariance matrix as the condition number N; and calculating the value coefficient of the data point to be identified based on the condition number. ； in The value coefficient of the i-th piece of data to be identified; when ≥Preset rarity threshold P and When the data is greater than or equal to the preset reliability threshold Q, it is considered valuable rare data and marked; otherwise, it is discarded. Here, P is the 95th percentile of the deviation direction rarity distribution of historical normal samples in the third-party benchmark dataset, and P ≥ 0.6; Q is the 5th percentile of the local structure reliability distribution of historical normal samples in the third-party benchmark dataset, and Q ≥ 0.

5.

5. The blockchain-based multi-center clinical trial data sharing method according to claim 1, characterized in that, The method for performing asymmetric soft mapping on the high-quality data and the rare data based on the differential soft label decoupling generation method includes: The clinical test indicators obtained from high-quality data constitute the statistical feature vector of each high-quality data point; the high-quality data are grouped by clinical category, and the mean and standard deviation of each group of data in each clinical test indicator dimension are calculated as the standard distribution reference for the corresponding clinical category. Mahalanobis distance is used to calculate the similarity between the statistical feature vector of high-quality data and the standard distribution reference. The similarity is then mapped to a probability distribution to obtain conventional soft labels. Clinical test indicators from rare data are obtained to form statistical feature vectors for the rare data. A covariance matrix is ​​calculated based on the statistical feature vectors of all rare data, and the ratio of the largest to the smallest eigenvalue of the covariance matrix is ​​used as the rare data condition number. Interpolation coefficients for shrinkage estimation are determined based on the rare data condition number, and an adaptive covariance matrix is ​​constructed using these coefficients. Mahalanobis distance is calculated based on the adaptive covariance matrix to obtain the similarity between the statistical feature vector of each rare data point and a standard distribution reference. This similarity is mapped to a probability distribution, and rare labels are embedded in the probability distribution to obtain soft labels carrying rare labels. A differentiated set of local soft tags consists of regular soft tags and soft tags carrying rare markers.

6. The blockchain-based multi-center clinical trial data sharing method according to claim 1, characterized in that, The method of generating several shared groups and corresponding global soft tags through a grouping and aggregation method based on rare additive features includes: Based on the local soft tag set of the participation center, the central probability distribution vector of the participation center and the rare probability distribution vector of the soft tag carrying the rare tag are statistically analyzed. The central probability distribution vector and the rare probability distribution vector are concatenated to obtain the additive distribution vector. The clinical category distribution similarity between any two participating centers is calculated using total variation distance based on the additive distribution vector; the participating centers are then divided into multiple consensus groups using graph clustering based on the clinical category similarity; the local consensus degree between any two participating centers within a consensus group is calculated, and the expression for the local consensus degree of the i-th participating center is: ； in, Let g be the set of participating centers within the g-th consensus group. The total number of centers within the group. The similarity of clinical category distribution between the i-th participating center and the j-th participating center; The participating centers within the group are screened based on the local consensus degree, and those participating centers that reach the consensus threshold are designated as shared centers. The consensus threshold ranges from 0.45 to 0.

65. The weighted aggregation of the additive distribution vector of the shared centers is then performed based on the local consensus degree to generate a global soft label for each group.

7. The blockchain-based multi-center clinical trial data sharing method according to claim 1, characterized in that, The method for generating lightweight models through confidence-based fidelity distillation includes: Using the local model of the participation center as the student model and the global soft label of the shared group as the teacher signal, the distribution difference between the predicted probability distribution of the student model and the probability distribution of the global soft label is calculated by KL divergence and used as the original distillation loss. The normalized entropy value of each global soft tag is calculated using Shannon entropy. ; Obtain the standard deviation among all local soft tags in the shared centers under each global soft tag, as the cross-center variance. ; Generate confidence coefficients based on normalized entropy values ​​and cross-center variance: ； in, Let be the confidence level of the i-th global soft label; α and β are weighting coefficients, α+β=1; The weighted distillation loss is obtained by multiplying the original distillation loss of each sample by the corresponding confidence coefficient. Retrieve the local soft labels of all shared centers under the global soft label carrying rare markers, and calculate the cross-center variance. Fidelity-preserving factor is generated based on cross-center variance: ； in, To predetermine the rare base bonus constant, 1.2 ≤ ≤1.8; This is the magnification factor, 0.5 ≤ ≤2; The weighted distillation loss of the global soft tag carrying rare markers is weighted a second time using a fidelity enhancement factor to obtain the enhanced fidelity loss; The weighted distillation loss of regular soft labels is added to the enhanced fidelity loss of rare soft labels to obtain the total distillation loss; the student model is iteratively trained using the total distillation loss to output a lightweight model.

8. The blockchain-based multi-center clinical trial data sharing method according to claim 1, characterized in that, The output sharing results include: The data to be shared is input into the lightweight model, which outputs a prediction probability vector corresponding to the data to be shared. The prediction probability vector contains the probability value of the data to be shared belonging to each clinical category; the clinical categories include full-term birth, late preterm birth, and early preterm birth.

9. A blockchain-based multi-center clinical trial data sharing system for performing the method described in any one of claims 1-8, characterized in that, include: Data Acquisition and Verification Module: This module is used to participate in the acquisition of subject data and inclusion / exclusion criteria by the center. The subject data includes clinical test indicators and clinical category labels. The module deploys a smart contract based on the inclusion / exclusion criteria. The smart contract verifies the subject data through layer-by-layer inclusion / exclusion verification based on a hierarchical verification tree. The subject data that passes the verification is used as the verification data. Data filtering module: used to perform deviation analysis on the verification data based on third-party benchmark data, obtain deviation score, and classify the verification data with deviation score below the global threshold as high-quality data, and the verification data with deviation score above the global threshold as data to be identified; The data to be identified is jointly evaluated based on the consistency of feature topology and deviation direction to obtain rare data. Grouping and aggregation module: used to perform asymmetric soft mapping on the high-quality data and the rare data based on the differentiated soft tag decoupling generation method to obtain a differentiated local soft tag set; based on the differentiated local soft tag set, several shared groups and corresponding global soft tags are obtained through the grouping and aggregation method based on rare addition, and the global soft tags are stored on the blockchain for evidence; Distillation training module: The participation center of the sharing group obtains the global soft label, uses the global soft label as a supervision signal, and generates a lightweight model through confidence-based fidelity distillation; the data to be shared is input into the lightweight model, and the sharing result is output.

Images