Multi-source data processing method for thrombotic microvascular disease based on artificial intelligence

By constructing a heterogeneous GNN network and variational autoencoder, combined with deformable convolution and GRU-ODE-Bayes module, the problems of integrating multi-source TMA data and protecting privacy were solved, achieving efficient data processing and risk assessment, and improving the accuracy and efficiency of TMA diagnosis and treatment.

CN117789983BActive Publication Date: 2024-11-19NANCHANG HIGH-TECH IND COLLABORATIVE INNOVATION INST CHINESE ACAD OF SCI
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202311830972.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-28
Publication Date
2024-11-19
Estimated Expiration
2043-12-28

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively integrate and analyze multi-source data on thrombotic microvascular disease (TMA), leading to issues such as data inconsistency, privacy risks, and difficulties in data sharing, which affect the accuracy and efficiency of diagnosis and treatment.

Method used

We employ an AI-based approach, preprocessing, segmenting, and encrypting multi-source data to construct a heterogeneous GNN network and a variational autoencoder. This is combined with deformable convolutions and GRU-ODE-Bayes modules for data feature extraction and fusion. Finally, we utilize generative adversarial networks and reinforcement learning to optimize risk assessment.

Benefits of technology

It enables efficient compression and representation of TMA data, improving the accuracy and efficiency of diagnosis, reducing privacy risks, optimizing risk assessment strategies, and enhancing medical quality and prevention rates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117789983B_ABST
    Figure CN117789983B_ABST
Patent Text Reader

Abstract

The present invention discloses a multi-source data processing method for thrombotic microvascular disease based on artificial intelligence, obtains TMA multi-source data, and performs preprocessing to conform to a predetermined data format to form a TMA data set; reads the TMA data set and performs segmentation and encryption to protect the privacy of the TMA data, and obtains several TMA data subsets after segmentation and encryption; constructs an artificial intelligence module, the artificial intelligence module has at least two data extraction channels and at least one variational autoencoder unit based on a heterogeneous GNN network; configures the trained artificial intelligence module in a server, calculates the TMA disease assessment probability according to the real-time collection of TMA data, forms a risk assessment time series diagram, and pushes it. The present invention greatly improves data processing efficiency and accuracy, and solves many problems such as missing, noise, inconsistency and imbalance of the existing TMA multi-source data, as well as privacy protection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to TMA-related technologies, and in particular to an artificial intelligence-based multi-source data processing method for thrombotic microvascular diseases. Background Art

[0002] Thrombotic microangiopathy (TMA) is a group of serious diseases characterized by intravascular thrombosis, hemolytic anemia and thrombocytopenia, including thrombotic thrombocytopenic purpura (TTP), hemolytic uremic syndrome (HUS) and other related diseases. Among them, the number of patients with TTP secondary to other diseases and drugs is increasing, and the incidence is on the rise. The disease is very dangerous, with a poor prognosis and a short course. If not treated in time, the mortality rate is as high as 80~90%. Functional deficiency of metalloproteinase ADAMTS13 is the most important cause of TPP. The pathogenesis of TMA is complex, involving multiple factors such as genetics, immunity, infection, drugs, metabolism, etc., with diverse clinical manifestations, and challenging diagnosis and treatment. The prognosis of TMA is poor. If not diagnosed and treated in time, it can lead to renal failure, nervous system damage, cardiovascular complications and even death.

[0003] With the advancement of medical technology, the diagnostic and treatment methods of TMA are constantly increasing, generating a large amount of multi-source data, including clinical data, laboratory data, imaging data, genetic data, etc. These data are characterized by diversity, complexity, dynamics and high dimensionality. Traditional data processing methods are difficult to effectively integrate, analyze and utilize these data, limiting the level of accurate diagnosis and treatment of TMA. In response to the problems existing in the existing technology, technicians have proposed some machine learning methods, but there are still some problems, mainly including: TMA's multi-source data often have problems such as missing, noise, inconsistency, imbalance, etc., which affect the effect and stability of the training and testing of artificial intelligence models. The output results of the artificial intelligence model may have problems such as bias, error, inconsistency, etc., which affect the credibility and reliability of the artificial intelligence model. TMA's multi-source data involves patients' personal information, medical information, genetic information, etc., which are sensitive and private. In order to protect the rights and privacy of patients, it is necessary to comply with relevant laws, regulations and ethical principles, and reasonably collect, store, transmit and use TMA's multi-source data. At the same time, in order to promote the development and application of artificial intelligence technology, it is necessary to establish an effective data sharing mechanism and platform to achieve cross-institutional collaboration of TMA's multi-source data.

[0004] Therefore, further research and innovation are needed. Summary of the invention

[0005] The purpose of the invention is to provide an artificial intelligence-based multi-source data processing method for thrombotic microvascular disease to solve at least one technical problem existing in the prior art.

[0006] Technical solution, according to one aspect of the present application, a method for processing multi-source data of thrombotic microvascular diseases based on artificial intelligence comprises the following steps:

[0007] Step S1, obtaining TMA multi-source data, and preprocessing it to conform to a predetermined data format to form a TMA data set; the preprocessing includes cleaning, normalization, standardization and normalization;

[0008] Step S2, reading the TMA data set and performing segmentation and encryption to protect the privacy of the TMA data, and obtaining a plurality of segmented and encrypted TMA data subsets;

[0009] Step S3, constructing an artificial intelligence module, the artificial intelligence module has at least two data extraction channels and at least one variational autoencoder unit based on a heterogeneous GNN network, wherein one data extraction channel is an image processing channel, which is constructed using a deformable convolution module; the other data extraction channel is a detection data time series processing channel, which is constructed using a GRU-ODE-Bayes module; constructing a heterogeneous GNN network based on the extracted data features and training the variational autoencoder unit;

[0010] Step S4: The trained artificial intelligence module is configured in the server, and the TMA disease assessment probability is calculated based on the real-time collected TMA data, a risk assessment time series diagram is formed, and the data is pushed.

[0011] According to one aspect of the present application, step S1 further comprises:

[0012] Step S11, acquiring TMA data from a medical big data center or a pre-collected database;

[0013] Step S12: Clean the TMA data to remove invalid, abnormal and duplicate data;

[0014] Step S13, unify the data format and mark it, including patient ID, age, and each test indicator data of each diagnostic period, unify the data unit, accuracy and range, so that the data meets the requirements of comparability and computability;

[0015] Step S14: standardize and normalize the TMA data to eliminate the dimension and distribution differences of the data and ensure the normality, stability and robustness of the data.

[0016] Step S15: Preliminarily classify the TMA data according to different data types and sources to form TMA data sets, each of which contains several samples, each of which contains several features and a label, and the label represents the disease status of the patient, and the disease status includes normal, mild, moderate or severe.

[0017] According to one aspect of the present application, step S2 is further:

[0018] Step S21: Use a multi-party secure computing protocol to divide the TMA data set into several data fragments, so that each data fragment only contains part of the data information and the original data cannot be restored alone;

[0019] Step S22: Encrypt the TMA data fragments into ciphertext using a homomorphic encryption algorithm, so that each ciphertext can only be decrypted by a specific key and cannot be stolen or tampered with by an unauthorized third party;

[0020] Step S23: Combining the segmented and encrypted TMA data segments into a plurality of TMA data subsets, each data subset containing different data segments and keys for different data analysis and processing tasks.

[0021] According to one aspect of the present application, step S3 further comprises:

[0022] Step S31, constructing an image processing channel using a deformable convolution module to extract the TMA image data for reflecting the chemical properties and change rules of blood and blood, and forming a first data vector;

[0023] Step S32: constructing a detection data time series processing channel using a GRU-ODE-Bayes module to extract the change trend, fluctuation range and periodicity in the TMA detection data, obtain the dynamic properties and change rules of blood and blood vessels, and form a second data vector;

[0024] Step S33, concatenating the first data vector and the second data vector into a comprehensive data vector, and constructing a heterogeneous GNN network for detecting data based on the comprehensive data vector;

[0025] Step S34: construct and use an artificial intelligence module based on a graph neural network and a variational autoencoder to fuse, compress and extract the comprehensive data vector to obtain a low-dimensional latent variable as a feature representation of the TMA;

[0026] According to one aspect of the present application, step S4 is further:

[0027] Step S41: deploy the artificial intelligence module on the cloud server and connect it to the data platform of each channel through the network interface;

[0028] Step S42: According to different data analysis and processing tasks, a corresponding TMA data subset is selected from the server, and the corresponding key is used to decrypt and restore the data fragment to obtain the original TMA data;

[0029] Step S43: the original TMA data is input into the artificial intelligence module for processing, and the feature representation of the TMA is output. Based on the feature representation of the TMA, the TMA disease assessment probability is given;

[0030] Step S44: Draw a risk assessment time series diagram based on the TMA disease assessment probability to display the patient's risk change trend and level, as well as the confidence interval of the risk;

[0031] Step S45: Push the risk assessment time sequence diagram and related data and information to the corresponding participants.

[0032] According to one aspect of the present application, the process of constructing the image processing channel in step S31 is further as follows:

[0033] Step S311, constructing and using a deformable convolution module to extract features from the TMA curve image, and adaptively adjusting the shape and position of the convolution kernel to adapt to different image structures and changes;

[0034] Step S312: construct and use a global average pooling layer to reduce the dimension and compress the output of the deformable convolution module, so as to average all pixel values ​​of each feature map to obtain a scalar, so as to reduce parameters and calculation amount and prevent overfitting;

[0035] Step S313: construct and use a fully connected layer to perform nonlinear transformation and mapping on the output of the global average pooling layer, map the input one-dimensional vector to an output vector of a specified dimension to form a first data vector; and introduce an activation function and a regularization term to improve the nonlinearity and generalization ability of the model.

[0036] According to one aspect of the present application, the process of constructing the detection data timing processing channel in step S32 is further as follows:

[0037] Step S321, construct and use the GRU-ODE-Bayes module to extract features from the TMA curve data to adaptively adjust the sampling frequency and accuracy of the time series data to adapt to different data densities and complexities;

[0038] Step S322: Perform nonlinear transformation and mapping on the output of the GRU-ODE-Bayes module through a fully connected layer to form a second data vector, map the input one-dimensional vector to an output vector of a specified dimension, and increase the complexity and flexibility of the model.

[0039] According to one aspect of the present application, step S33 is further:

[0040] Step S331: concatenate the first data vector and the second data vector into a comprehensive data vector, and use the comprehensive data vector as the input of the heterogeneous GNN network to represent data of different types and sources as a graph structure, and use the topology and attribute information of the graph to propagate and update features;

[0041] Step S332: construct and use a variational autoencoder to encode and decode the output of the heterogeneous GNN network to obtain a low-dimensional latent variable as the feature representation of the TMA, so as to compress the high-dimensional data into low-dimensional data while retaining important information of the data and improving the compactness and effectiveness of the features.

[0042] According to one aspect of the present application, in step S35, a data enhancement module based on a generative adversarial network is constructed to generate synthetic data using TMA data to expand the scale and diversity of the data set;

[0043] Step S351, construct and use a generative adversarial network, including a generator and a discriminator, wherein the input of the generator is a random noise vector, and the output is a synthesized TMA data, and the input of the discriminator is real or synthesized TMA data, and the output is the probability of the data being true or false;

[0044] Step S352: Use the minimum-maximum loss function to train the adversarial generative network so that the generator can generate high-quality synthetic data that is sufficient to deceive the discriminator, and the discriminator can accurately distinguish between real data and synthetic data;

[0045] Step S353: Mix the synthetic data generated by the generator with the real data to form an expanded TMA dataset, which is used to train the artificial intelligence module to improve the performance and accuracy of the model.

[0046] According to one aspect of the present application, there is also step S4A between step S44 and step S45, a risk assessment optimization module based on reinforcement learning, which uses the TMA disease assessment probability output by the artificial intelligence module to give the optimal risk assessment strategy and suggestions;

[0047] Step S4A1, constructing and using a reinforcement learning model, including an agent and an environment, wherein the input of the agent is the TMA disease assessment probability and the risk assessment time series diagram, and the output is the risk assessment strategy, the input of the environment is the risk assessment strategy and suggestions, and the output is the patient's feedback and reward;

[0048] Step S4A2: using a policy gradient-based algorithm to train a reinforcement learning model so that the agent can give the optimal risk assessment strategy and suggestions based on the TMA disease assessment probability and risk assessment time series diagram to maximize the patient's feedback and reward;

[0049] Step S4A3: Reinforce the risk assessment strategy and suggestions output by the learning model and the risk assessment time series Figure 1 Push it to the corresponding parties.

[0050] Beneficial effects: Technically, it can solve the problems of inconsistent data sources and privacy risks in the existing TMA hierarchical diagnosis, and provide a solution for processing multimodal medical data. In terms of economic and social effects, it improves the prevention and treatment rate of TMA, reduces the incidence of TMA, improves medical quality, reduces medical risks, and increases public awareness of TMA, achieving good social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the present invention.

[0052] Figure 2 It is a flow chart of step S1 of the present invention.

[0053] Figure 3 It is a flow chart of step S2 of the present invention.

[0054] Figure 4 It is a flow chart of step S3 of the present invention.

[0055] Figure 5 It is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0056] Common TMA tests include: blood routine test, which can show the number of red blood cells and platelets. In TMA, red blood cell reduction and thrombocytopenia are common. Peripheral blood smear: Check the morphology of red blood cells and look for red blood cell fragments, which are typical manifestations of TMA. Renal function test: Determine serum creatinine and urea nitrogen levels to evaluate renal function. Coagulation function test: Including PT, APTT, fibrinogen level, etc., to help rule out other coagulation disorders. ADAMTS13 enzyme activity test: Especially in suspected thrombotic thrombocytopenic purpura (TTP), low ADAMTS13 activity is its characteristic. Complement system test: Including C3, C4 levels and gene mutation detection of complement regulatory proteins, especially in suspected complement-related TMA (such as typical hemolytic uremic syndrome). Liver function test: Test liver function, because some types of TMA (such as HELLP syndrome) are associated with liver damage. Imaging examination: Such as abdominal ultrasound or CT scan to evaluate the degree of organ damage. Testing in special cases: Depending on the specific situation, more specialized tests may be required, such as genetic testing. The data obtained from the test becomes TMA data.

[0057] In actual practice, TMA has many inspection items, and the inspection units and instruments are different. It is difficult to align the data, resulting in duplicate medical treatment and increased medical costs. One of the purposes of promoting tiered medical care is to share medical test data, thereby reducing costs and improving test efficiency. However, in addition to the wide range of data sources, the problem of imbalance and inconsistency is prominent, which makes it difficult for doctors to make accurate judgments and requires secondary or even multiple tests, greatly increasing the medical burden. In addition, data from multiple sources need to be protected for privacy, which is also one of the reasons for the slow promotion of medical big data systems, and it also needs to be solved. Finally, there are many technical problems with the application of artificial intelligence in TMA. For example, the data has multiple imaging data and multiple test indicators for multiple organs. How to compare and process these data is one of the current technical difficulties.

[0058] To this end, the following solution is provided.

[0059] like Figure 1 As shown, according to one aspect of the present application, a method for processing multi-source data of thrombotic microvascular diseases based on artificial intelligence comprises the following steps:

[0060] Step S1, obtaining TMA multi-source data, and preprocessing it to conform to a predetermined data format to form a TMA data set; the preprocessing includes cleaning, normalization, standardization and normalization;

[0061] Step S2, reading the TMA data set and performing segmentation and encryption to protect the privacy of the TMA data, and obtaining a plurality of segmented and encrypted TMA data subsets;

[0062] Step S3, constructing an artificial intelligence module, the artificial intelligence module has at least two data extraction channels and at least one variational autoencoder unit based on a heterogeneous GNN network, wherein one data extraction channel is an image processing channel, which is constructed using a deformable convolution module; the other data extraction channel is a detection data time series processing channel, which is constructed using a GRU-ODE-Bayes module; constructing a heterogeneous GNN network based on the extracted data features and training the variational autoencoder unit;

[0063] Step S4: The trained artificial intelligence module is configured in the server, and the TMA disease assessment probability is calculated based on the real-time collected TMA data, a risk assessment time series diagram is formed, and the data is pushed.

[0064] By effectively processing and analyzing TMA multi-source data, valuable features and information are extracted, providing a scientific basis and method for TMA disease assessment. By adopting multi-party secure computing and homomorphic encryption technology, the privacy and security of TMA data are protected, the leakage and tampering of data are prevented, and the ethical and legal requirements of medical data are met. Through the variational autoencoder unit based on the heterogeneous GNN network, efficient compression and representation of TMA data are achieved, the dimension and complexity of the data are reduced, and the compactness and effectiveness of the data are improved. In some embodiments, synthetic data can also be generated using TMA data through a data enhancement module based on a generative adversarial network, which expands the scale and diversity of the data set and enhances the performance and accuracy of the model. It is also possible to use a risk assessment optimization module based on reinforcement learning, using the TMA disease assessment probability, to give the optimal risk assessment strategy and suggestions, and optimize the effect and efficiency of risk assessment. It is possible to achieve efficient compression and representation of TMA data, reduce the dimension and complexity of data, and improve the compactness and effectiveness of data.

[0065] According to one aspect of the present application, step S1 further comprises:

[0066] Step S11, obtaining TMA data from a medical big data center or a pre-collected database; accessing the medical big data center or the pre-collected database through the network or other means, screening out qualified data according to the definition and standard of TMA, and downloading or exporting it to a local or cloud storage space.

[0067] Step S12: Clean the TMA data to remove invalid, abnormal and duplicate data; perform quality check on the TMA data to find and delete missing values, erroneous values, abnormal values, duplicate values ​​and other data that affect data analysis, and retain valid, accurate and consistent data. The software or program for data cleaning can be Python, R, Excel, etc. The evaluation criteria and methods for data quality can be missing value ratio, abnormal value detection, duplicate value identification, etc.

[0068] Step S13, unify the data format and mark it, including patient ID, age, and each test indicator data of each diagnostic period, unify the data unit, accuracy and range, so that the data meets the requirements of comparability and computability;

[0069] Specifically, it can include: converting the format of TMA data to conform to the predetermined data format, such as CSV, JSON, XML, etc., to facilitate data reading and processing; labeling TMA data, adding necessary metadata to each data, such as patient ID, age, and each test indicator data in each diagnostic period, to facilitate data identification and analysis; unifying the unit, precision and range of TMA data so that the data has the same dimension and distribution, which is convenient for data comparison and calculation. The software or program for data format conversion can use Python, R, Excel, etc., and the rules and methods for data labeling can use data dictionaries, data templates, data encoding, etc., and the standards and methods for data units, precision and ranges, such as data conversion, data scaling, data truncation, etc.

[0070] Step S14, standardizing and normalizing the TMA data to eliminate the dimension and distribution differences of the data and ensure the normality, stability and robustness of the data;

[0071] In some embodiments, the TMA data can be standardized so that the mean of the data is M and the standard deviation is N, eliminating the influence of the dimension and scale of the data and improving the comparability of the data; the TMA data can be normalized so that the minimum value of the data is a and the maximum value is b, or the range of the data is within a certain interval, eliminating the influence of the distribution and discreteness of the data and improving the computability of the data. The software or program for data standardization and normalization includes Python, R, Excel, etc. The formulas and methods for data standardization and normalization can use Z-score, Min-Max, Sigmoid, etc.

[0072] Step S15: Preliminarily classify the TMA data according to different data types and sources to form TMA data sets, each of which contains several samples, each of which contains several features and a label, and the label represents the disease status of the patient, and the disease status includes normal, mild, moderate or severe.

[0073] In a certain embodiment, it can be simply divided into two types of data. For example, according to the type and source of TMA data, TMA data is divided into two categories, one is TMA image data, including curve images of the patient's blood and blood vessels, reflecting the chemical properties and change laws of blood and blood vessels; the other is TMA detection data, including the detection index data of the patient's blood and blood vessels, reflecting the dynamic properties and change laws of blood and blood vessels. For each type of data, it is divided into a training set, a validation set and a test set for model training, validation and testing. For each data, it is represented as a sample, including several features and a label. The feature is the attribute value of the data, such as the pixel value of the image or the value of the detection index. The label is the category value of the data, such as the disease state of the patient, and the disease state includes normal, mild, moderate or severe. The software or program for data classification can use Python, R, Excel, etc. The basis and method of data classification include data type, data source, data distribution, etc. The rules and methods of data representation include data matrix, data vector, data encoding, etc.

[0074] In a certain embodiment, two TMA data sets are obtained, namely TMA_image_data.csv and TMA_test_data.csv, each of which contains 800 training samples, 100 validation samples and 100 test samples, and each sample contains 11 features and 1 label. The column labels are: patient ID, age, diagnosis period, feature 1, feature 2, ..., feature 10, feature 11, label.

[0075] In this embodiment, the relevant methods ensure the quality, quantity, validity, consistency, comparability, computability, normality, stability and robustness of the data, and provide reliable data support for subsequent data processing and analysis.

[0076] According to one aspect of the present application, step S2 is further:

[0077] Step S21: Use a multi-party secure computing protocol to divide the TMA data set into several data fragments, so that each data fragment only contains part of the data information and the original data cannot be restored alone;

[0078] In one embodiment, the detection data of each sample can be divided into two parts, one part contains the first 500 data, and the other part contains the last 500 data. Then, the two parts can be sent to two different cloud servers or two different partners. In this way, each recipient can only see half of the data information and cannot infer the complete data content. Therefore, it can be adjusted according to needs and security requirements, such as dividing the data into more fragments or using more complex segmentation algorithms.

[0079] Step S22: Encrypt the TMA data fragments into ciphertext using a homomorphic encryption algorithm, so that each ciphertext can only be decrypted by a specific key and cannot be stolen or tampered with by an unauthorized third party;

[0080] Each piece of data can be encrypted using an encryption algorithm that supports homomorphic addition and multiplication, such as Paillier encryption or Gentry's fully homomorphic encryption. Generate a key pair, including a public key and a private key. You can send the public key to the recipient and let them encrypt the data fragment with the public key. Keep the private key for the final decryption. This encryption method can ensure the confidentiality and integrity of the data, and it can also support some basic operations on the ciphertext, such as addition and multiplication, without decryption. Operations such as retrieval and comparison can be performed on the encrypted data to obtain the correct results without decrypting the data during the entire process.

[0081] Step S23: Combining the segmented and encrypted TMA data segments into a plurality of TMA data subsets, each data subset containing different data segments and keys for different data analysis and processing tasks.

[0082] Different data fragments and public keys can be combined to form a data subset according to the target task. For example, if you want to perform cluster analysis on the first 500 test data, you can form a data subset containing the first 500 test data and the corresponding public keys, and then send it to a cloud server or partner dedicated to clustering. If you want to perform feature selection on the last 500 test data, you can form another data subset containing the last 500 test data fragments and the corresponding public keys, and then send it to another cloud server or partner dedicated to feature selection. Data subsets can be flexibly organized and allocated according to different task requirements while ensuring data security and efficiency.

[0083] In a further embodiment, a differential privacy module can also be constructed to add random noise to TMA data to protect sensitive information such as personal information, medical information, and genetic information in the data, thereby achieving data security and legal use. Differential privacy can control the degree of privacy protection and data availability in a parameterized manner, thereby balancing the privacy and utility of data.

[0084] According to one aspect of the present application, step S3 further comprises:

[0085] Step S31, constructing an image processing channel using a deformable convolution module to extract the TMA image data for reflecting the chemical properties and change rules of blood and blood, and forming a first data vector;

[0086] Step S311, constructing and using a deformable convolution module to extract features from the TMA curve image, and adaptively adjusting the shape and position of the convolution kernel to adapt to different image structures and changes;

[0087] Step S312: construct and use a global average pooling layer to reduce the dimension and compress the output of the deformable convolution module, so as to average all pixel values ​​of each feature map to obtain a scalar, so as to reduce parameters and calculation amount and prevent overfitting;

[0088] Step S313: construct and use a fully connected layer to perform nonlinear transformation and mapping on the output of the global average pooling layer, map the input one-dimensional vector to an output vector of a specified dimension to form a first data vector; and introduce an activation function and a regularization term to improve the nonlinearity and generalization ability of the model.

[0089] In order to extract the chemical properties and change patterns of blood and blood vessels from TMA image data, a deformable convolution module is needed to build an image processing pipeline. The deformable convolution module can adaptively adjust the shape and position of the convolution kernel according to the content and structure of the image to capture features of different scales and directions. An instantiated deformable convolution module is as follows: Input image: A grayscale image of size 256x256, representing the changes in the TMA curve.

[0090] Deformable convolution module: contains three sub-layers, namely:

[0091] Deformable convolution layer: Use a 3x3 convolution kernel, a stride of 1, a padding of 1, and 64 output channels to perform a convolution operation on the input image and learn an offset for each convolution point so that the convolution kernel can be deformed to adapt to the characteristics of the image.

[0092] Batch Normalization Layer: Batch normalizes the output of the deformable convolutional layer to reduce internal covariate shift, speed up the training process, and improve the stability of the model.

[0093] Activation function layer: Use the ReLU function as the activation function to increase the nonlinearity of the model and improve the expressiveness of the model.

[0094] Output feature map: A tensor of size 256x256x64, representing the feature representation of the image after passing through the deformable convolution module.

[0095] Step S32: constructing a detection data time series processing channel using a GRU-ODE-Bayes module to extract the change trend, fluctuation range and periodicity in the TMA detection data, obtain the dynamic properties and change rules of blood and blood vessels, and form a second data vector;

[0096] Step S321, construct and use the GRU-ODE-Bayes module to extract features from the TMA curve data to adaptively adjust the sampling frequency and accuracy of the time series data to adapt to different data densities and complexities;

[0097] Step S322: Perform nonlinear transformation and mapping on the output of the GRU-ODE-Bayes module through a fully connected layer to form a second data vector, map the input one-dimensional vector to an output vector of a specified dimension, and increase the complexity and flexibility of the model.

[0098] In order to extract the dynamic properties and changing rules of blood and blood vessels from TMA detection data, it is necessary to use the GRU-ODE-Bayes module to build a time series processing channel. The GRU-ODE-Bayes module can adaptively adjust the sampling frequency and accuracy of time series data according to the density and complexity of the data to capture the changing trend, fluctuation range and periodicity of the data. An instantiated GRU-ODE-Bayes module is shown below:

[0099] Input time series data: a matrix of size TxD, representing D-dimensional TMA detection data at T time moments, such as blood pressure, blood oxygen, blood sugar, etc. GRU-ODE-Bayes module: contains three sub-layers, namely:

[0100] GRU-ODE layer: Use a bidirectional GRU unit to encode the input time series data into a hidden state vector h(t), and use an ordinary differential equation (ODE) solver to perform continuous time modeling on the time series data according to the dynamic changes of h(t) to obtain a low-dimensional time series feature vector g(t).

[0101] Bayesian update layer: Use a Bayesian network to take the observation data y(t) and the time series feature vector g(t) at each moment as input, and update the hidden state vector h(t) through an attention mechanism to fuse the observation information and time series information.

[0102] Fully connected layer: Use a fully connected layer to map the final hidden state vector h(T) to an output vector z of a specified dimension as a feature representation of the time series data.

[0103] Output feature vector: A vector of size K, which represents the feature representation of the time series data after passing through the GRU-ODE-Bayes module.

[0104] Step S33, concatenating the first data vector and the second data vector into a comprehensive data vector, and constructing a heterogeneous GNN network for detecting data based on the comprehensive data vector;

[0105] Step S331: concatenate the first data vector and the second data vector into a comprehensive data vector, and use the comprehensive data vector as the input of the heterogeneous GNN network to represent data of different types and sources as a graph structure, and use the topology and attribute information of the graph to propagate and update features;

[0106] Step S332: construct and use a variational autoencoder to encode and decode the output of the heterogeneous GNN network to obtain a low-dimensional latent variable as the feature representation of the TMA, so as to compress the high-dimensional data into low-dimensional data while retaining important information of the data and improving the compactness and effectiveness of the features.

[0107] In order to fuse the first data vector and the second data vector into a comprehensive data vector and construct a heterogeneous graph neural network (GNN) for detection data based on the comprehensive data vector, the following method is required:

[0108] The first data vector and the second data vector are concatenated into a comprehensive data vector as the input of the heterogeneous GNN. The dimension of the comprehensive data vector is K+64, where K is the dimension of the second data vector and 64 is the dimension of the first data vector.

[0109] Data of different types and sources are represented as a graph structure, and the topology and attribute information of the graph are used to propagate and update features. Specifically, TMA detection data can be used as nodes of the graph, TMA image data can be used as edges of the graph, and the feature vectors of nodes and edges can be used as attributes of the graph. Then, a heterogeneous graph neural network can be used to transfer and aggregate information on the graph to obtain a feature representation of the graph.

[0110] Step S34, construct and use an artificial intelligence module based on graph neural network and variational autoencoder to fuse, compress and extract the comprehensive data vector to obtain a low-dimensional latent variable as the feature representation of TMA. In order to compress the high-dimensional comprehensive data vector into a low-dimensional latent variable as the feature representation of TMA, a variational autoencoder (VAE) is needed to encode and decode the output of the heterogeneous GNN. VAE can compress high-dimensional data into low-dimensional latent variables while retaining important information of the data, improving the compactness and effectiveness of the features.

[0111] An instantiated VAE is as follows:

[0112] Input data: A vector of size K+64, representing the output of the heterogeneous GNN, i.e., the synthetic data vector.

[0113] VAE: contains two sub-layers, namely:

[0114] Encoding layer: Use a fully connected layer to map the input data to a low-dimensional latent space to obtain a mean vector and a variance vector, which represent the distribution parameters of the latent variables respectively.

[0115] Decoding layer: Use a reparameterization technique to sample a latent variable from the latent space, and then use a fully connected layer to reconstruct the latent variable into an output data that is as close to the input data as possible.

[0116] Output data: A vector of size K+64, representing the output of VAE, that is, the reconstruction result of the latent variables.

[0117] Here, we only use a simple case to illustrate its working principle instead of introducing the actual operation process.

[0118] By using deformable convolution modules, GRU-ODE-Bayes modules, heterogeneous GNNs and VAEs, useful features can be extracted from TMA data of different types and sources, while considering the content, structure, timing and distribution of the data to obtain a low-dimensional latent variable as the feature representation of TMA. Such a feature representation can effectively reflect the inherent laws and differences of TMA data, which is conducive to subsequent data mining and knowledge discovery.

[0119] According to one aspect of the present application, in step S35, a data enhancement module based on a generative adversarial network is constructed to generate synthetic data using TMA data to expand the scale and diversity of the data set;

[0120] Step S351, construct and use a generative adversarial network, including a generator and a discriminator, wherein the input of the generator is a random noise vector, and the output is a synthesized TMA data, and the input of the discriminator is real or synthesized TMA data, and the output is the probability of the data being true or false;

[0121] Step S352: Use the minimum-maximum loss function to train the adversarial generative network so that the generator can generate high-quality synthetic data that is sufficient to deceive the discriminator, and the discriminator can accurately distinguish between real data and synthetic data;

[0122] Step S353: Mix the synthetic data generated by the generator with the real data to form an expanded TMA dataset, which is used to train the artificial intelligence module to improve the performance and accuracy of the model.

[0123] Generative Adversarial Network is an unsupervised learning method that implements the training process through the mutual constraints of two networks, one is a generative network and the other is a discriminative network. The task of the generative network is to generate instances that look natural and real and similar to the original data. The task of the discriminative network is to determine whether a given instance is "real". This can be seen as a zero-sum game, where the generative network and the discriminative network continuously improve each other until a Nash equilibrium is reached, that is, the generative network can generate high-quality synthetic data that is sufficient to deceive the discriminative network, and the discriminative network cannot distinguish between real data and synthetic data. An instantiated generative adversarial network is shown below:

[0124] Input data: A vector of size K+64, representing the feature representation of TMA, i.e. the output of VAE.

[0125] Generative Adversarial Network: It contains two sub-layers:

[0126] Generator: Use a fully connected layer to map a random noise vector to a vector of size K+64 as the synthetic TMA data with the same dimension and format as the input data.

[0127] Discriminator: Use a fully connected layer to map the input data or synthetic data to a scalar, which represents the probability of the data being true or false. If it is 1, it means that it is 100% real data, and if the output is 0, it means that it cannot be real data.

[0128] Output data: A vector of size K+64, representing the output of the adversarial generation network, i.e., the synthesized TMA data.

[0129] Using the minimum-maximum loss function, the adversarial generative network is trained so that the generator can maximize the discriminator's error and the discriminator can minimize its own error. The form of the loss function is as follows:

[0130] ;

[0131] Where x represents the real TMA data, z represents the random noise vector, G(z) represents the generated TMA data, D(x) represents the true or false probability of the real data, D(G(z)) represents the true or false probability of the generated data, E represents the expectation, pdata(x) represents the distribution of the real data, and pz(z) represents the distribution of the noise vector.

[0132] By using a generative adversarial network, we can learn the distribution and characteristics of TMA data, generate high-quality synthetic data, and mix it with real data to form a larger and richer TMA dataset for training artificial intelligence modules and improving the performance and accuracy of the model. Such data enhancement methods can effectively solve the problems of insufficient and unbalanced data and improve the robustness and stability of the model.

[0133] According to one aspect of the present application, step S4 is further:

[0134] Step S41: deploy the AI ​​module on the cloud server and connect it to the data platform of each channel through the network interface; use a self-built or commercial telecom cloud to deploy the AI ​​module and provide corresponding APIs to communicate and interact with the data platform of each channel. Use some encryption and authentication technologies, such as SSL, OAuth, etc., to protect data transmission and access.

[0135] Step S42: According to different data analysis and processing tasks, select the corresponding TMA data subset from the server, and use the corresponding key to decrypt and restore the data fragments to obtain the original TMA data; according to different needs, data can be flexibly selected and allocated, and the integrity and correctness of the data can also be guaranteed to avoid data loss and damage. Some commonly used databases and data warehouses, such as MySQL, MongoDB, Hadoop, etc., can be used to store and manage TMA data subsets, and provide corresponding query and retrieval functions to obtain the required data from the server. Some commonly used encryption and decryption technologies, such as AES, RSA, etc., can also be used to encrypt and decrypt data fragments to restore the original data from the data subset.

[0136] Step S43, the original TMA data is input into the artificial intelligence module for processing, and the feature representation of TMA is output. Based on the feature representation of TMA, the probability of TMA disease assessment is given; the deep learning ability of the artificial intelligence module can be used to extract useful features and rules from the TMA data, thereby giving the probability of TMA disease assessment as a prediction and evaluation of the patient's health status. Some commonly used deep learning frameworks, such as TensorFlow, PyTorch, Keras, etc., can be used to implement the artificial intelligence module, and provide corresponding input and output interfaces to interact with the data in the server. Some commonly used evaluation indicators, such as accuracy, recall, AUC, etc., can also be used to evaluate the performance and accuracy of the artificial intelligence module.

[0137] Step S44: Draw a risk assessment time series diagram based on the TMA disease assessment probability to display the patient's risk change trend and level, as well as the confidence interval of the risk; some commonly used visualization tools, such as Matplotlib, Seaborn, Plotly, etc., can be used to draw the risk assessment time series diagram, and provide corresponding parameters and options to adjust the style and content of the graph. Some commonly used statistical methods, such as confidence intervals, hypothesis tests, etc., can also be used to calculate and display the uncertainty and credibility of the risk.

[0138] Step S45: Push the risk assessment time sequence diagram and related data and information to the corresponding participants.

[0139] According to one aspect of the present application, there is also step S4A between step S44 and step S45, a risk assessment optimization module based on reinforcement learning, which uses the TMA disease assessment probability output by the artificial intelligence module to give the optimal risk assessment strategy and suggestions;

[0140] Step S4A1, constructing and using a reinforcement learning model, including an agent and an environment, wherein the input of the agent is the TMA disease assessment probability and the risk assessment time series diagram, and the output is the risk assessment strategy, the input of the environment is the risk assessment strategy and suggestions, and the output is the patient's feedback and reward;

[0141] Step S4A2: using a policy gradient-based algorithm to train a reinforcement learning model so that the agent can give the optimal risk assessment strategy and suggestions based on the TMA disease assessment probability and risk assessment time series diagram to maximize the patient's feedback and reward;

[0142] Step S4A3: Reinforce the risk assessment strategy and suggestions output by the learning model and the risk assessment time series Figure 1 Push it to the corresponding parties.

[0143] In one embodiment, the details are as follows:

[0144] #Introduce related libraries and modules

[0145] import numpy as np

[0146] import tensorflow as tf

[0147] from ai_module import AI_model # Artificial intelligence module, which can output TMA disease assessment probability

[0148] from risk_graph import Risk_graph # Risk assessment time series graph, which can display the patient's risk assessment status

[0149] # Define the parameters and hyperparameters of the reinforcement learning model

[0150] state_size = 2 # The dimension of the state, including TMA disease assessment probability and risk assessment time series diagram

[0151] action_size = 3 # The dimension of the action, including three risk assessment strategies and suggestions: high risk, medium risk, and low risk

[0152] hidden_size = 10 # Number of neurons in the hidden layer

[0153] learning_rate = 0.01 # learning rate

[0154] gamma = 0.95 # discount factor

[0155] epsilon = 0.1 # exploration rate

[0156] batch_size = 32 # batch size

[0157] # Define the structure and functions of the reinforcement learning model

[0158] class RL_model(object):

[0159] def __init__(self):

[0160] # Define the input layer to receive the state vector

[0161] self.state_input = tf.placeholder(tf.float32, [None, state_size])

[0162] # Define hidden layer, use fully connected layer and activation function

[0163] self.hidden_layer = tf.layers.dense(self.state_input, hidden_size, activation=tf.nn.relu)

[0164] # Define the output layer and output the strategy function, that is, the probability distribution of each action

[0165] self.policy_output = tf.layers.dense(self.hidden_layer,action_size, activation=tf.nn.softmax)

[0166] # Define the loss function and use cross entropy loss

[0167] self.action_input = tf.placeholder(tf.int32, [None]) # Receive the action to be executed

[0168] self.reward_input = tf.placeholder(tf.float32, [None]) # Receive the reward

[0169] self.action_one_hot = tf.one_hot(self.action_input, action_size) # Convert the action to one-hot encoding

[0170] self.log_policy = tf.log(self.policy_output) # Take the logarithm

[0171] self.log_action = tf.reduce_sum(self.log_policy * self.action_one_hot, axis=1) # Calculate the logarithmic probability of executing the action

[0172] self.loss = -tf.reduce_mean(self.log_action * self.reward_input) # Calculate loss function

[0173] # Define the optimizer, using the gradient ascent method

[0174] self.optimizer = tf.train.AdamOptimizer(learning_rate).minimize(self.loss)

[0175] # Define session and initialize variables

[0176] self.sess = tf.Session()

[0177] self.sess.run(tf.global_variables_initializer())

[0178] # Define a memory to store the state, action, and reward at each step

[0179] self.states = []

[0180] self.actions = []

[0181] self.rewards = []

[0182] def choose_action(self, state):

[0183] # Choose an action based on the current state using the epsilon - greedy strategy

[0184] policy = self.sess.run(self.policy_output, feed_dict={self.state_input: [state]}) # Calculate the policy function for the current state

[0185] policy = np.squeeze(policy) # Remove the extra dimension

[0186] if np.random.rand() < epsilon: # Randomly choose an action with probability epsilon

[0187] action = np.random.choice(action_size)

[0188] else: # Choose the action with the highest probability with probability 1 - epsilon

[0189] action = np.argmax(policy)

[0190] return action

[0191] def store_transition(self, state, action, reward):

[0192] # Store the state, action, and reward at each step

[0193] self.states.append(state)

[0194] self.actions.append(action)

[0195] self.rewards.append(reward)

[0196] def learn(self):

[0197] # Learning and updating policy functions

[0198] def learn(self):

[0199] # Calculate the cumulative discounted return at each step

[0200] discounted_rewards = []

[0201] cumulative_reward = 0

[0202] for reward in reversed(self.rewards): # Traverse the reward list from back to front

[0203] cumulative_reward = reward + gamma * cumulative_reward # Calculate the cumulative discount reward

[0204] discounted_rewards.insert(0, cumulative_reward) # Insert the cumulative discounted reward to the beginning of the list

[0205] # Normalized cumulative discounted return

[0206] discounted_rewards = np.array(discounted_rewards)

[0207] discounted_rewards = (discounted_rewards - np.mean(discounted_rewards)) / (np.std(discounted_rewards) + 1e-8) # Prevent division by zero

[0208] # Update the parameters of the strategy function

[0209] self.sess.run(self.optimizer, feed_dict={

[0210] self.state_input: self.states, # Input state list

[0211] self.action_input: self.actions, # Input action list

[0212] self.reward_input: discounted_rewards # Input the cumulative discounted rewards list

[0213] })

[0214] # Clear the memory

[0215] self.states = []

[0216] self.actions = []

[0217] self.rewards = []

[0218] # Create an instance of a reinforcement learning model

[0219] rl_model = RL_model()

[0220] # Define a loop to play multiple games

[0221] for episode in range(1000):

[0222] # Initialize the state, starting from any patient

[0223] state = np.random.choice(patients) # patients is a list containing all patient data

[0224] # Initialize the flag to indicate whether the game is over

[0225] done = False

[0226] # Initialize the total reward of the game

[0227] total_reward = 0

[0228] # Define a loop and play a game

[0229] while not done:

[0230] # Use the artificial intelligence module to output the TMA disease assessment probability based on the patient's data

[0231] probability = AI_model.predict(state) # AI_model is a trained artificial intelligence module that can output the probability of TMA disease assessment

[0232] # Use the risk assessment timing diagram to display the patient's risk assessment status

[0233] graph = Risk_graph.plot(state) # Risk_graph is a graphical tool that can display the patient's risk assessment status

[0234] # Concatenate the TMA disease assessment probability and risk assessment time series diagram into a state vector

[0235] state = np.concatenate([probability, graph])

[0236] # Use the reinforcement learning model to select a risk assessment strategy and recommendations based on the current state

[0237] action = rl_model.choose_action(state)

[0238] # Execute risk assessment strategies and recommendations, and observe the next state and reward

[0239] next_state, reward, done = env.step(action) # env is a simulation environment that can return the next state and reward, as well as a sign of whether the game is over, based on risk assessment strategies and suggestions

[0240] # Store the current state, action and reward

[0241] rl_model.store_transition(state, action, reward)

[0242] # Update Status

[0243] state = next_state

[0244] # Total accumulated rewards

[0245] total_reward += reward

[0246] # After a game is over, learn and update the strategy function

[0247] rl_model.learn()

[0248] # Print game information

[0249] print("Episode: {}, Total reward: {}".format(episode, total_reward)).

[0250] It should be noted that the results output by this application cannot be used as medical diagnosis data, but only for the analysis and processing of medical test data. Doctors are required to give the final results based on actual conditions.

[0251] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A multi-source data processing method for thrombotic microvascular disease based on artificial intelligence, characterized in that: The steps include: Step S1, obtaining TMA multi-source data, and preprocessing it to conform to a predetermined data format to form a TMA data set; the preprocessing includes cleaning, normalization, standardization and normalization; Step S2, reading the TMA data set and performing segmentation and encryption to protect the privacy of the TMA data, and obtaining a plurality of segmented and encrypted TMA data subsets; Step S3, constructing an artificial intelligence module, the artificial intelligence module has at least two data extraction channels and at least one variational autoencoder unit based on a heterogeneous GNN network, wherein one data extraction channel is an image processing channel, which is constructed using a deformable convolution module; the other data extraction channel is a detection data time series processing channel, which is constructed using a GRU-ODE-Bayes module; constructing a heterogeneous GNN network based on the extracted data features and training the variational autoencoder unit; Step S4: The trained artificial intelligence module is configured in the server, and the TMA disease assessment probability is calculated according to the real-time collected TMA data, a risk assessment time sequence diagram is formed, and the data is pushed; Step S2 is further as follows: Step S21: Use a multi-party secure computing protocol to divide the TMA data set into several data fragments, so that each data fragment only contains part of the data information and the original data cannot be restored alone; Step S22: Encrypt the TMA data fragments into ciphertext using a homomorphic encryption algorithm, so that each ciphertext can only be decrypted by a specific key and cannot be stolen or tampered with by an unauthorized third party; Step S23, combining the segmented and encrypted TMA data segments into a plurality of TMA data subsets, each data subset containing different data segments and keys for different data analysis and processing tasks; The step S3 is further as follows: Step S31, constructing an image processing channel using a deformable convolution module to extract the TMA image data for reflecting the chemical properties and change rules of blood and blood, and forming a first data vector; Step S32: constructing a detection data time series processing channel using a GRU-ODE-Bayes module to extract the change trend, fluctuation range and periodicity in the TMA detection data, obtain the dynamic properties and change rules of blood and blood vessels, and form a second data vector; Step S33, concatenating the first data vector and the second data vector into a comprehensive data vector, and constructing a heterogeneous GNN network for detecting data based on the comprehensive data vector; Step S34: construct and use an artificial intelligence module based on graph neural network and variational autoencoder to fuse, compress and extract the comprehensive data vector to obtain a low-dimensional latent variable as the feature representation of TMA.

2. The method for processing multi-source data of thrombotic microvascular disease based on artificial intelligence according to claim 1, characterized in that: Step S1 is further as follows: Step S11, acquiring TMA data from a medical big data center or a pre-collected database; Step S12: Clean the TMA data to remove invalid, abnormal and duplicate data; Step S13, unify the data format and mark it, including patient ID, age, and each test indicator data of each diagnostic period, unify the data unit, accuracy and range, so that the data meets the requirements of comparability and computability; Step S14: standardize and normalize the TMA data to eliminate the dimension and distribution differences of the data and ensure the normality, stability and robustness of the data. Step S15: Preliminarily classify the TMA data according to different data types and sources to form TMA data sets, each of which contains several samples, each of which contains several features and a label, and the label represents the disease status of the patient, and the disease status includes normal, mild, moderate or severe.

3. The method for processing multi-source data of thrombotic microvascular disease based on artificial intelligence according to claim 2, characterized in that: The step S4 is further as follows: Step S41: deploy the artificial intelligence module on the cloud server and connect it to the data platform of each channel through the network interface; Step S42: According to different data analysis and processing tasks, a corresponding TMA data subset is selected from the server, and the corresponding key is used to decrypt and restore the data fragment to obtain the original TMA data; Step S43: the original TMA data is input into the artificial intelligence module for processing, and the feature representation of the TMA is output. Based on the feature representation of the TMA, the TMA disease assessment probability is given; Step S44: Draw a risk assessment time series diagram based on the TMA disease assessment probability to display the patient's risk change trend and level, as well as the confidence interval of the risk; Step S45: Push the risk assessment time sequence diagram and related data and information to the corresponding participants.

4. The method for processing multi-source data of thrombotic microvascular disease based on artificial intelligence according to claim 3, characterized in that: The process of constructing the image processing channel in step S31 is further as follows: Step S311, constructing and using a deformable convolution module to extract features from the TMA curve image, and adaptively adjusting the shape and position of the convolution kernel to adapt to different image structures and changes; Step S312: construct and use a global average pooling layer to reduce the dimension and compress the output of the deformable convolution module, so as to average all pixel values ​​of each feature map to obtain a scalar, so as to reduce parameters and calculation amount and prevent overfitting; Step S313: construct and use a fully connected layer to perform nonlinear transformation and mapping on the output of the global average pooling layer, map the input one-dimensional vector to an output vector of a specified dimension to form a first data vector; and introduce an activation function and a regularization term to improve the nonlinearity and generalization ability of the model.

5. The method for processing multi-source data of thrombotic microvascular disease based on artificial intelligence according to claim 4, characterized in that: The process of constructing the detection data timing processing channel in step S32 is further as follows: Step S321, construct and use the GRU-ODE-Bayes module to extract features from the TMA curve data to adaptively adjust the sampling frequency and accuracy of the time series data to adapt to different data densities and complexities; Step S322: Perform nonlinear transformation and mapping on the output of the GRU-ODE-Bayes module through a fully connected layer to form a second data vector, map the input one-dimensional vector to an output vector of a specified dimension, and increase the complexity and flexibility of the model.

6. The method for processing multi-source data of thrombotic microvascular disease based on artificial intelligence according to claim 5, characterized in that: The step S33 is further as follows: Step S331: concatenate the first data vector and the second data vector into a comprehensive data vector, and use the comprehensive data vector as the input of the heterogeneous GNN network to represent data of different types and sources as a graph structure, and use the topology and attribute information of the graph to propagate and update features; Step S332: construct and use a variational autoencoder to encode and decode the output of the heterogeneous GNN network to obtain a low-dimensional latent variable as the feature representation of the TMA, so as to compress the high-dimensional data into low-dimensional data while retaining important information of the data and improving the compactness and effectiveness of the features.

7. The method for processing multi-source data of thrombotic microvascular disease based on artificial intelligence according to claim 1, characterized in that: The method further includes step S35, constructing a data enhancement module based on a generative adversarial network to generate synthetic data using the TMA data to expand the scale and diversity of the data set; Step S351, construct and use a generative adversarial network, including a generator and a discriminator, wherein the input of the generator is a random noise vector, and the output is a synthesized TMA data, and the input of the discriminator is real or synthesized TMA data, and the output is the probability of the data being true or false; Step S352: Use the minimum-maximum loss function to train the adversarial generative network so that the generator can generate high-quality synthetic data that is sufficient to deceive the discriminator, and the discriminator can accurately distinguish between real data and synthetic data; Step S353: Mix the synthetic data generated by the generator with the real data to form an expanded TMA dataset, which is used to train the artificial intelligence module to improve the performance and accuracy of the model.

8. The method for processing multi-source data of thrombotic microvascular disease based on artificial intelligence according to claim 3, characterized in that: There is also step S4A between step S44 and step S45, a risk assessment optimization module based on reinforcement learning, which uses the TMA disease assessment probability output by the artificial intelligence module to give the optimal risk assessment strategy and suggestions; Step S4A1, constructing and using a reinforcement learning model, including an agent and an environment, wherein the input of the agent is the TMA disease assessment probability and the risk assessment time series diagram, and the output is the risk assessment strategy, the input of the environment is the risk assessment strategy and suggestions, and the output is the patient's feedback and reward; Step S4A2: using a policy gradient-based algorithm to train a reinforcement learning model so that the agent can give the optimal risk assessment strategy and suggestions based on the TMA disease assessment probability and risk assessment time series diagram to maximize the patient's feedback and reward; Step S4A3: Push the risk assessment strategies and suggestions output by the reinforcement learning model together with the risk assessment time sequence diagram to the corresponding participants.

Citation Information

Patent Citations

  • Cardiovascular disease risk prediction network model based on multiple parameters and construction method thereof

    CN111292853A

  • Method and device for constructing and identifying septicopyemia model based on GRU-ODE-Bayes and storage medium

    CN115238778A

  • Diabetes risk early warning method based on big data analysis

    CN117253614A