A multi-clinical stage disease auxiliary classification method and system based on federated learning

Abstract

This invention discloses a multi-clinical-staging disease auxiliary classification method and system based on federated learning, belonging to the fields of big data and medical technology. To improve the accuracy of the classification model and the security of privacy data, this invention collects case data to construct a case dataset. This dataset is input into a client for analysis and processing to obtain an optimized case dataset. The client trains an XGGridBoost model on the optimized dataset. The trained model parameters are then encrypted and compressed, and sent to a central server for decompression. The decompressed model parameters are then securely aggregated, and an approximate algorithm is used to determine the optimal split point. The central server sends the optimal split point to multiple clients. Upon receiving the split point, each client determines whether its local decision tree has reached its maximum depth and performs iterative training or outputs the trained model. This invention solves the problem of data silos and improves the security of privacy data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of big data and medical technology, and particularly relates to a multi-clinical staging disease auxiliary classification method and system based on federated learning. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, traditional handwritten medical records have been gradually replaced by electronic medical records. Based on electronic medical records, machine learning training is carried out, resulting in various disease auxiliary classification models. By inputting pathological characteristic values of patients, the disease stages are predicted to provide auxiliary suggestions for doctors. However, with the increase of medical record data and the improvement of data protection awareness, the existing multi-clinical staging disease auxiliary classification model training through traditional machine learning has some defects: the traditional machine learning model belongs to centralized learning, which needs to integrate patient data of various hospitals. In the integration process, the risk of privacy data leakage is increased. At the same time, due to the high privacy of medical data, some hospitals are unwilling to share data, which forms the problem of data islands. SUMMARY

[0003] The application aims to improve the accuracy of the classification model and the security of the privacy data, and proposes a multi-clinical staging disease auxiliary classification method and system based on federated learning.

[0004] To achieve the above-mentioned purpose, the application realizes the following technical solutions:

[0005] A multi-clinical staging disease auxiliary classification method based on federated learning includes a center server and multiple clients willing to participate in federated learning and submit federated learning tasks to the center server, and includes the following steps:

[0006] S1, collecting case data to construct a case data set, inputting the case data set into the client for analysis and processing to obtain an optimized case data set;

[0007] S2, the client trains the XGGridBoost model on the optimized case data set obtained in step S1 to obtain the trained model parameters;

[0008] The specific implementation method of step S2 includes the following steps:

[0009] S2.1, constructing an XGGridBoost initial model as Obj, and calculating the expression as:

[0010]

[0011] Wherein, T is the number of leaf nodes, j is the jth leaf node, λ and γ are hyperparameters that control the regularization of the regularization mode L1 and the regularization mode L2 in the regularization term respectively, G j is the sum of the first-order partial derivatives of the samples contained in the leaf node j, H j is the sum of the second-order partial derivatives of the samples contained in the leaf node j.

[0012] S2.2, the client trains the XGGridBoost model on the optimized case data set obtained in step S1, uses the XGGridBoost initial model constructed in step S2.1 to train, adopts cross entropy as the loss function, judges the prediction result, and calculates the expression:

[0013]

[0014] Wherein, p(x i ) is the true distribution probability of the ith disease, q(x i ) is the predicted distribution probability of the ith disease, n is the total number of disease stages, and Loss is the cross entropy loss function.

[0015] S3, the model parameters obtained in step S2 are encrypted and compressed to obtain processed model parameters sent to the central server.

[0016] The specific implementation method of step S3 includes the following steps:

[0017] S3.1, the trained model parameters obtained in step S2 are encrypted by using differential privacy technology, and the calculation expression is:

[0018] P r [M(x)∈S]≤exp(ε)P r [M(y)∈S]+δ

[0019] Wherein, P r [M(x)∈S] is the probability of obtaining result S by executing algorithm M on data set x containing trained model parameters, ε is the privacy parameter, x is the data set containing trained model parameters, y is the adjacent data set of x, δ is the additional parameter, and M is the algorithm with added noise.

[0020] The algorithm with added noise uses Laplace algorithm, and the calculation expression is:

[0021]

[0022] Wherein, q'(x) is the result after adding Laplace noise, q(x) is the original data, Δf is the sensitivity of the query, and ε is the privacy budget parameter.

[0023] The calculation expression of the density function f(z) of Laplace noise is:

[0024]

[0025] Wherein, z is a random variable value, b is a scale parameter, μ is a position parameter, and |z-μ| is the absolute distance between z and μ;

[0026] S3.2, the model parameters processed by step S3.1 are combined with LZ77 algorithm and Huffman coding, first, the model parameters processed by encryption are compressed by using LZ77 algorithm, then the frequency of each compressed parameter is calculated, according to the frequency, the smallest two nodes are selected to be combined into a new node each time, and the weight value is added as the weight value of the new node, and the process is repeated until all nodes are combined into a root node, a Huffman tree is constructed, finally each compressed parameter is converted into binary code according to the generated Huffman coding, then all binary codes are connected to obtain the processed model parameters;

[0027] S3.3, the processed model parameters obtained in step S3.2 are sent to the central server;

[0028] S4, the central server decompresses the received processed model parameters, then aggregates the decompressed model parameters safely, and determines the best segmentation point using an approximation algorithm, then the central server sends the best segmentation point to multiple clients;

[0029] S5, after receiving the best segmentation point, the multiple clients judge whether the local decision tree reaches the maximum depth, if not, the next layer is constructed according to the obtained best segmentation point information, and the samples are redistributed for iterative model training, if yes, the decision tree construction is stopped to complete the training, and the trained model is output.

[0030] Further, the specific implementation method of step S1 includes the following steps:

[0031] S1.1, collect case data, then take the personal information, medical history information, physical examination, biochemical detection of patients as characteristic values, and take the diagnosis result of doctors as labels, then remove the error data existing in the medical record to construct a medical record data set F1;

[0032] S1.2, fill in the missing values of the medical record data set F1 obtained in step S1 to obtain a medical record data set F2 with supplementary data;

[0033] S1.3, delete the diagnosis-independent characteristic values of the medical record data set F2 with supplementary data obtained in step S1.2 to obtain a diagnosis-related data set F3;

[0034] S1.4, perform correlation analysis on the diagnostic related data set F3 obtained in step S1.3, and then delete feature values with low correlation in the diagnostic related data set F3 to obtain an optimized case data set F4.

[0035] Further, the specific implementation method of step S1.2 includes the following steps:

[0036] S1.2.1, for missing values of height and weight of the patient, respectively calculate the average values of height and weight of males and females to fill in the missing values;

[0037] S1.2.2, for missing values of waist circumference and hip circumference, use multiple linear regression to predict and fill in the missing values by taking hip circumference and waist circumference as dependent variables and the feature values of height and weight as independent variables;

[0038] S1.2.3, for missing values of biochemical tests, use an integrated learning algorithm based on decision tree to predict and fill in the missing values.

[0039] Further, the diagnostic result irrelevant feature values in step S1.3 include patient's ID number, mobile phone number, personal information feature values, and medication feature values after medical examination.

[0040] Further, the specific implementation method of step S1.4 includes the following steps:

[0041] S1.4.1, perform correlation analysis on the diagnostic related data set F3 obtained in step S1.3 based on chi-square test method, and the calculation expression of chi-square test method is:

[0042]

[0043] wherein f0 is an actual value, f e is an expected value, and x 2 is a chi-square statistic;

[0044] S1.4.2, perform correlation analysis on the diagnostic related data set F3 obtained in step S1.3 based on variance analysis method, and the calculation expression of variance analysis method is:

[0045]

[0046] wherein SSA is inter-group sum of squares, SSE is intra-group sum of squares, the corresponding SSA / (k-1) is inter-group variance, SSE / (n-k) is intra-group variance, and F is variance ratio;

[0047] S1.4.3, take the physical examination data and blood sample test results and urine sample test results of the patient after correlation analysis as a feature value set, and take the diagnostic result of the patient as a label set to constitute an optimized case data set F4.

[0048] Further, the specific implementation method of step S4 includes the following steps:

[0049] S4.1, the center server uses the Huffman tree constructed in step S3.2 to take out binary codes one by one and compare from the root node of the tree until the corresponding character is found, then perform LZ77 decoding, according to the window size and sliding distance used during compression, take out the matching data from the sliding window one by one and splice together to obtain the decompressed model parameters;

[0050] S4.2, the decompressed model parameters obtained in step S4.1 are securely aggregated, and an approximation algorithm is used to determine the best split point, first, candidate division points are proposed according to the percentiles of the feature distribution, then, the features are mapped to the candidate point division, the statistical information is aggregated, the sample statistical values G and H in each bucket are accumulated and counted, and finally the best split point is determined on the accumulated count, and the best split point is determined by using the scoring function Gain, and the expression is:

[0051]

[0052] Wherein, G L is all sample gradients in the left subtree, H L is all Hessian matrices in the left subtree, G R is all sample gradients in the right subtree, H R is all Hessian matrices in the right subtree, and λ is a regularization term coefficient and γ is a penalty term.

[0053] The maximum Gain value is selected as the best split point.

[0054] S4.3, the center server sends the best split point obtained in step S4.2 to multiple clients.

[0055] A computer system includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the steps of the above-mentioned method for auxiliary classification of multiple clinical stage diseases based on federated learning.

[0056] The beneficial effects of the present application are:

[0057] The method for auxiliary classification of multiple clinical stage diseases based on federated learning, in the federated learning training process, further includes: after the local training of the client is completed, the model parameters are compressed using the LHC lossless data compression algorithm to reduce the data transmission amount; and the parameters are decompressed at the center server. The compression algorithm is introduced to improve the training performance of federated learning.

[0058] The application discloses a multi-clinical staging disease auxiliary classification method based on federated learning.

[0059] The application discloses a multi-clinical staging disease auxiliary classification method based on federated learning. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The application discloses a multi-clinical staging disease auxiliary classification method based on federated learning.

[0061] Figure 2 The application discloses a multi-clinical staging disease auxiliary classification method based on federated learning. DETAILED DESCRIPTION

[0062] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application, that is, the specific embodiments described are only a part of the embodiments of the present application, but not all the specific embodiments. The components of the specific embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations, and the present application can also have other embodiments.

[0063] Therefore, the following detailed description of the specific embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only represents selected specific embodiments of the present application. Based on the specific embodiments of the present application, all other specific embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0064] In order to further understand the inventive content, characteristics and effects of the present application, the following specific embodiments are exemplified, and the accompanying drawings are Figure 1 -attached Figure 2 The detailed description is as follows: Specific embodiment one:

[0066] A multi-clinical staging disease auxiliary classification method based on federated learning, comprising a center server and a plurality of clients willing to participate in federated learning and submit federated learning tasks to the center server, comprising the following steps:

[0067] S1, collecting case data to construct a case data set, inputting the case data set into the client for analysis and processing to obtain an optimized case data set;

[0068] Further, the specific implementation method of step S1 comprises the following steps:

[0069] S1.1, collect case data, then take the patient's personal information, medical history information, physical examination, biochemical detection as characteristic values, and the doctor's diagnosis result as a label, then remove the error data existing in the medical record to construct a medical record data set F1;

[0070] S1.2, filling the missing values of the medical record data set F1 obtained in step S1 to obtain a medical record data set F2 with supplementary data;

[0071] Further, the specific implementation method of step S1.2 comprises the following steps:

[0072] S1.2.1, for the missing values of the patient's height and weight, the average values of height and weight of male and female are calculated respectively, and the missing values are filled;

[0073] S1.2.2, for the missing values of waist circumference and hip circumference, the hip circumference and waist circumference are respectively taken as dependent variables, and the characteristic values of height and weight are taken as independent variables, and multiple linear regression is used for prediction and filling;

[0074] Multiple linear regression is the correlation between one dependent variable and multiple independent variables, and its formula can be described as:

[0075] f(x 1， x 2，…， x n )=α1x1+α2x2+…+α n x n +β

[0076] Where, x 1， x 2，…， x nrespectively represent the 1st, 2nd,..., nth independent variables; in the present application, two independent variables are set, x1 is height, and x2 is weight; f(x 1， x 2，…， x n represents a dependent variable that is linearly synthesized under the influence of these independent variables; α1, α2,..., α n respectively represent coefficients to be fitted; β represents a constant to be fitted;

[0077] S1.2.3, for missing values of biochemical tests, a decision tree-based ensemble learning algorithm is used for prediction filling;

[0078] The biochemical test item has a low missing rate. A decision tree-based ensemble learning algorithm is used to train multiple decision tree models, and the importance scores of each feature in different decision tree models are obtained. Then, according to the importance scores, the feature vectors with relatively high importance are selected as input features, and the input features are used as independent variables. A regression model is used to predict the samples containing missing values, and the filling results of the missing values are obtained. After obtaining the prediction results of multiple decision tree models, the average value of the results is used as the final result, so as to fill the missing values;

[0079] S1.3, deleting the diagnosis result irrelevant feature values of the supplementary data set F2 obtained in step S1.2 to obtain a diagnosis related data set F3;

[0080] Further, the diagnosis result irrelevant feature values in step S1.3 include patient's certificate number, mobile phone number, personal information feature values, and medical advice and drug feature values after seeing a doctor;

[0081] S1.4, performing correlation analysis on the diagnosis related data set F3 obtained in step S1.3, then deleting the feature values with low correlation in the diagnosis related data set F3 to obtain an optimized case data set F4;

[0082] Further, the specific implementation method of step S1.4 includes the following steps:

[0083] S1.4.1, performing correlation degree analysis on the diagnosis related data set F3 obtained in step S1.3 based on the chi-square test method, and the calculation expression of the chi-square test method is:

[0084]

[0085] Wherein, f0 is the actual value, f e is the expected value, and x 2 is the chi-square statistic;

[0086] The numerator in the above formula represents the deviation of the actual value from the expected value, and the denominator represents the standardization process. The larger the chi-square value, the higher the correlation between the two variables;

[0087] S1.4.2, based on the analysis of variance method, the correlation degree analysis of the diagnostic correlation data set F3 obtained in step S1.3 is carried out, and the calculation expression of the analysis of variance method is:

[0088]

[0089] Wherein, SSA is the between-group sum of squares, SSE is the within-group sum of squares, the corresponding SSA / (k-1) is the between-group variance, SSE / (n-k) is the within-group variance, and F is the variance ratio;

[0090] The larger the F value, the greater the difference between the levels, and the greater the influence of the independent variable on the dependent variable.

[0091] S1.4.3, the physical examination data, blood sample test results and urine sample test results of the patients after correlation analysis are taken as a feature value set, and the diagnosis results of the patients are taken as a label set, to form an optimized case data set F4;

[0092] S2, the client trains the optimized case data set obtained in step S1 using the XGGridBoost model to obtain the trained model parameters;

[0093] Specifically, the XGGridBoost model is used for training, which combines XGBoost and Grid Search. By training a decision tree on the medical record data set, the residual error between the predicted value and the true value of each sample is calculated. These residuals are used to train the next tree. It trains the tree by continuously iterating until the residual error becomes smaller and smaller, and finally builds a gradient boosting model based on the decision tree. In the training process, Grid Search is introduced to specify the range and step of hyperparameters, cross-validation is performed through different combinations, and the combination with the best performance is selected as the final hyperparameters for training and prediction.

[0094] Further, the specific implementation method of step S2 includes the following steps:

[0095] S2.1, constructing an XGGridBoost initial model as Obj, and the calculation expression is:

[0096]

[0097] Wherein, T is the number of leaf nodes, j is the jth leaf node, λ and γ are hyperparameters for controlling the regularization of regularization mode L1 and regularization mode L2 in the regularization term, G j is the sum of the first-order partial derivatives of the samples contained in the leaf node j, which is a constant, and Hj The second-order partial derivative accumulation sum of the samples contained by the leaf node j is also a constant; the L1 regularization is to punish the complexity of the model by adding the sum of the absolute values of the weight values; and the L2 regularization is to limit the complexity of the model by adding a penalty term to the sum of the squares of the weight values. The first-order derivative g i of each leaf node is solved i , and the sum is obtained by traversing it to obtain G j and H j ;

[0098] S2.2, the client trains the optimized case data set obtained in step S1 using the XGGridBoost initial model constructed in step S2.1, uses cross-entropy as the loss function, judges the prediction result, and calculates the expression:

[0099]

[0100] where p(x i ) is the true distribution probability of the i-th disease, q(x i ) is the predicted distribution probability of the i-th disease, n is the total number of disease stages, and Loss is the cross-entropy loss function.

[0101] The cross-entropy loss function can measure the distribution degree of two different probabilities. The smaller the cross-entropy value, the better the prediction result of the model.

[0102] S3, encrypt and compress the trained model parameters obtained in step S2 to obtain processed model parameters and send them to the central server.

[0103] Further, the specific implementation method of step S3 includes the following steps:

[0104] S3.1, encrypt the trained model parameters obtained in step S2 using differential privacy technology, and calculate the expression:

[0105] P r [M(x)∈S]≤exp(ε)P r [M(y)∈S]+δ

[0106] where P r [M(x)∈S] is the probability of obtaining result S by executing algorithm M on data set x containing trained model parameters, ε is the privacy parameter, x is the data set containing trained model parameters, y is the adjacent data set of x, δ is the additional parameter, and M is the algorithm with added noise.

[0107] In differential privacy, the goal of privacy protection is tor [M(x)∈S] and P r [M(y)∈S] makes a deviation limit, if the deviation result is smaller, can guarantee that the private data will not be leaked;

[0108] The algorithm of adding noise adopts Laplace algorithm, and the calculation expression is:

[0109]

[0110] Wherein, q'(x) is the result after adding Laplace noise, q(x) is the original data, Δf is the sensitivity of the query, and ε is the privacy budget parameter;

[0111] The calculation expression of the density function f(z) of the Laplace noise is:

[0112]

[0113] Wherein, z is the value of the random variable, b is the scale parameter, μ is the position parameter, and |z-μ| is the absolute distance between z and μ;

[0114] S3.2, the model parameters after the encryption processing in step S3.1 are combined with the Huffman coding by using the LZ77 algorithm, first, the model parameters after the encryption processing are compressed by using the LZ77 algorithm, then the frequency of each compressed parameter is calculated, according to the frequency, the smallest two nodes are selected to be combined into a new node each time, and the weight value is added as the weight value of the new node, the process is repeated until all the nodes are combined into a root node, a Huffman tree is constructed, finally, each compressed parameter is converted into binary code according to the generated Huffman coding, then all the binary codes are connected to obtain the processed model parameters;

[0115] Specifically, the LHC lossless data compression algorithm is used as the compression algorithm. This algorithm combines the LZ77 algorithm with Huffman coding. First, initialization is performed by dividing the input model parameters into multiple non-overlapping blocks. Then, the input pointer is set to the beginning of the model parameters in each block, and an empty dictionary is created to store the model parameters that have already been traversed. Starting from the current position of the input model parameter, a substring is searched backwards, ensuring that the substring has appeared in the traversed model parameters. If found, the position and length of the substring in the dictionary are recorded; otherwise, the single parameter at that position is recorded in the dictionary. The recorded position and length, as well as the single parameter not found, are output. Then, the pointer is moved to the next unprocessed position, and the traversed parameters and the new parameter are added to the dictionary. After multiple iterations, the compression of the entire model parameter is completed. Then, Huffman coding is used, with shorter codes for frequently occurring characters and longer codes for less frequently occurring characters. Specifically, the frequency of each parameter is calculated, and a Huffman tree is constructed based on the frequency. Each time, the two nodes with the lowest frequencies are selected and merged into a new node, with their weights added together as the new node's weight. This process is repeated until all nodes are merged into a single root node. Finally, each parameter is converted to binary according to the generated Huffman code, and all the binary codes are concatenated to obtain the compressed data.

[0116] S3.3 Send the processed model parameters obtained in step S3.2 to the central server;

[0117] S4. The central server decompresses the received and processed model parameters, then performs secure aggregation on the decompressed model parameters, and uses an approximate algorithm to determine the optimal segmentation point. The central server then sends the optimal segmentation point to multiple clients.

[0118] Furthermore, the specific implementation method of step S4 includes the following steps:

[0119] S4.1 The central server uses the Huffman tree constructed in step S3.2 to extract the binary code one by one and compare it starting from the root node of the tree until the corresponding character is found. Then, LZ77 decoding is performed. According to the window size and sliding distance used during compression, the matching data is extracted from the sliding window one by one and concatenated together to obtain the decompression model parameters.

[0120] During the decompression process, the sliding window is continuously updated based on the already decompressed data and the data that has not yet been decompressed, and the compressed byte data is found in the window and decompressed based on the matching length and distance;

[0121] S4.2. Securely aggregate the decompressed model parameters obtained in step S4.1, and use an approximate algorithm to determine the optimal split point. First, candidate split points are proposed based on the percentiles of the feature distribution. Then, the features are mapped to the bins of the candidate point partitions. Statistical information is aggregated, and the sample statistics G and H within each bin are accumulated. Finally, the optimal split point is determined based on the accumulated statistics. The optimal split point is determined using the scoring function Gain, and the calculation expression is:

[0122]

[0123] Among them, G L H represents the gradient of all samples in the left subtree. L G represents all Hessian matrices in the left subtree. R H represents the gradient of all samples in the right subtree. R represents all Hessian matrices in the right subtree, where λ is the regularization coefficient and γ is the penalty term;

[0124] Select the largest Gain value as the optimal split point;

[0125] S4.3 The central server sends the optimal split point obtained in step S4.2 to multiple clients;

[0126] S5. After multiple clients receive the best split point, they determine whether the local decision tree has reached the maximum depth. If not, they use the obtained best split point information as the basis to build the next layer of the local decision tree and reallocate samples for iterative model training. If yes, they stop building the decision tree, complete the training, and output the trained model.

[0127] By following the steps above, the client and the central server collaboratively build a multi-clinical-stage disease auxiliary classification model, and each client can perform disease prediction locally using the classification model.

[0128] The advantage of this invention lies in the fact that traditional machine learning training requires integrating multiple datasets before model training. This process carries the risk of privacy data leakage, and some data providers may be unwilling to share data, creating data silos. This invention proposes a multi-clinical disease staging auxiliary classification method based on federated learning. By integrating a federated learning framework with machine learning, it solves the risks of privacy data leakage and the problem of data silos. Simultaneously, it employs a encrypted compression algorithm, effectively reducing data transmission volume and improving the performance of federated learning training. Specific Implementation Method Two:

[0130] A computer system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described federated learning-based multi-clinical-staging disease auxiliary classification method.

[0131] The computer device of the present invention may include a processor and a memory, such as a microcontroller containing a central processing unit. Furthermore, the processor executes the computer program stored in the memory to implement the steps of the aforementioned multi-clinical-staging disease auxiliary classification method based on federated learning.

[0132] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0133] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function (such as sound playback, image playback, etc.); the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0134] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0135] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A multi-staging disease auxiliary classification method based on federated learning, comprising a center server and a plurality of clients willing to participate in federated learning and submit federated learning tasks to the center server, characterized in that, Comprise the following steps: S1, collect case data to construct a case data set, input the case data set to the client, analyze and process to obtain an optimized case data set; S2, the client trains the XGGridBoost model of the optimized case data set obtained in step S1 to obtain the trained model parameters; The specific implementation method of step S2 comprises the following steps: S2.1, construct an XGGridBoost initial model as Obj, and the calculation expression is: wherein T is the number of leaf nodes, j is the jth leaf node, λ and γ are hyperparameters that control the regularization of the regularization terms L1 and L2, respectively, G j is the sum of the first-order partial derivatives of the samples contained in the leaf node j, H j is the sum of the second-order partial derivatives of the samples contained in the leaf node j. S2.2, the client trains the XGGridBoost model of the optimized case data set obtained in step S1, trains using the XGGridBoost initial model constructed in step S2.1, adopts cross entropy as the loss function, judges the prediction result, and the calculation expression is: where p(x i ) is the true distribution probability of the i-th disease, q(x i ) is the predicted distribution probability of the i-th disease, n is the total number of disease stages, and Loss is the cross-entropy loss function. S3, the trained model parameters obtained in step S2 are encrypted and compressed to obtain the processed model parameters, which are sent to the center server; The specific implementation method of step S3 comprises the following steps: S3.1, the trained model parameters obtained in step S2 are encrypted using differential privacy technology, and the calculation expression is: P r [M(x)∈S]≤exp(ε)P r [M(y)∈S]+δ where P r [M(x)∈S] is the probability of executing the algorithm M on the data set x containing the trained model parameters to obtain the result S, ε is a privacy parameter, x is a data set containing trained model parameters, y is a neighboring data set of x, δ is an additional parameter, and M is an algorithm with added noise. The algorithm for adding noise adopts Laplace algorithm, and the calculation expression is: where q ' (x) is the result after adding Laplace noise, q(x) is the original data, Δf is the sensitivity of the query, and ε is the privacy budget parameter. The calculation expression of the density function f(z) of Laplace noise is: Wherein, z is the value of random variable, b is the scale parameter, mu is the position parameter, and |z-mu| is the absolute distance between z and mu; S3.2, the model parameters processed in step S3.1 are combined using LZ77 algorithm and Huffman coding, first, the model parameters processed in step S3.1 are compressed using LZ77 algorithm, then the frequency of each compressed parameter is calculated, according to the frequency, the smallest two nodes are selected to be combined into a new node each time, and the weight value is added as the weight value of the new node, and the process is repeated until all nodes are combined into a root node, a Huffman tree is constructed, finally, each compressed parameter is converted into binary code according to the generated Huffman coding, then all binary codes are connected to obtain the processed model parameters; S3.3, the processed model parameters obtained in step S3.2 are sent to the center server; S4, the center server decompresses the received processed model parameters, then aggregates the decompressed model parameters safely, and determines the best segmentation point using an approximate algorithm, then the center server sends the best segmentation point to multiple clients; S5, after receiving the best segmentation point, the multiple clients judge whether the local decision tree reaches the maximum depth, if not, the next layer is constructed according to the obtained best segmentation point information, and the samples are redistributed for iterative model training, if yes, the decision tree construction is stopped to complete the training, and the trained model is output.

2. The method of claim 1, wherein the method is a federated learning based multi-staging disease auxiliary classification method. The specific implementation method of step S1 comprises the following steps: S1.1, collect case data, then take the patient's personal information, medical history information, physical examination, biochemical detection as characteristic value, and take the doctor's diagnosis result as label, then remove the error data in the medical record to construct a medical record data set F1; S1.2, the medical record data set F1 obtained in step S1 is filled with missing values to obtain a medical record data set F2 with supplementary data; S1.3, the medical record data set F2 with supplementary data obtained in step S1.2 is subjected to diagnosis result irrelevant feature value deletion to obtain a diagnosis related data set F3; S1.4, the diagnosis related data set F3 obtained in step S1.3 is subjected to correlation analysis, and then the feature values with low correlation in the diagnosis related data set F3 are deleted to obtain an optimized case data set F4.

3. The federated learning-based multi-staging disease auxiliary classification method according to claim 1 or 2, characterized in that: The specific implementation method of step S1.2 includes the following steps: S1.2.1, for the missing values of the height and weight of the patient, the average height and weight of males and females are calculated respectively to fill in the missing values; S1.2.2, for the missing values of waist circumference and hip circumference, the hip circumference and waist circumference are taken as dependent variables, and the feature values of height and weight are taken as independent variables, and multiple linear regression is used for prediction and filling; S1.2.3, for the missing values of biochemical detection, an integrated learning algorithm based on decision tree is used for prediction and filling.

4. The federated learning-based multi-staging disease auxiliary classification method according to claim 3, characterized in that: The diagnosis result irrelevant feature values in step S1.3 include the patient's certificate number, mobile phone number, personal information feature values, and medication feature values after seeing a doctor.

5. The federated learning based multi-staging disease auxiliary classification method according to claim 4, characterized in that: The specific implementation method of step S1.4 includes the following steps: S1.4.1, the diagnosis related data set F3 obtained in step S1.3 is subjected to correlation degree analysis based on the chi-square test method, and the calculation expression of the chi-square test method is: where f0is the actual value, f e is the expected value, x 2 is the chi-square statistic; S1.4.2, the diagnosis related data set F3 obtained in step S1.3 is subjected to correlation degree analysis based on the analysis of variance method, and the calculation expression of the analysis of variance method is: Wherein, SSA is the between-group sum of squares, SSE is the within-group sum of squares, the corresponding SSA / (k-1) is the between-group variance, SSE / (n-k) is the within-group variance, and F is the variance ratio; S1.4.3, the physical examination data and blood sample test results and urine sample test results of the patient after correlation analysis are taken as a feature value set, and the diagnosis result of the patient is taken as a label set to form an optimized case data set F4.

6. The federated learning based multi-staging disease auxiliary classification method according to claim 1, characterized in that: The specific implementation method of step S4 includes the following steps: S4.1, the center server uses the Huffman tree constructed in step S3.2 to take out binary codes one by one and compare from the root node of the tree until the corresponding character is found, then performs LZ77 decoding, according to the window size and sliding distance used during compression, takes out the matching data from the sliding window one by one, and splices them together to obtain the decompressed model parameters; S4.2, the decompressed model parameters obtained in step S4.1 are subjected to security aggregation, and an approximate algorithm is used to determine the best segmentation point, first, according to the percentile of the feature distribution, a candidate division point is proposed, then, the features are mapped to the candidate point division of the bucket, the statistical information is aggregated, the sample statistical values G and H in each bucket are accumulated and counted, finally, the best segmentation point is determined on the accumulated statistics, and the best segmentation point is determined by using the scoring function Gain, the calculation expression is: where G L is the gradient of all samples in the left sub-tree, H L is the Hessian matrix of all samples in the left sub-tree, G R is the gradient of all samples in the right sub-tree, H R is the Hessian matrix of all samples in the right sub-tree, λ is the coefficient of the regularization term, and γ is the penalty term. The maximum Gain value is selected as the best segmentation point; S4.3, the center server sends the best segmentation point obtained in step S4.2 to multiple clients.

7. A computer system comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the federated learning-based multi-clinical stage disease auxiliary classification method in any one of claims 1-6 when executing the computer program.

Images