A method for predicting all-cause mortality risk in a peritoneal dialysis patient

Abstract

The present application relates to a peritoneal dialysis patient all-cause death risk prediction method, belonging to the technical field of death risk prediction, solving the problem of low prediction accuracy in the prior art. The method comprises: constructing a sample set in each data center; training a neural network model based on the sample sets of multiple data centers to obtain a risk prediction pre-training model; obtaining the feature representation of each sample based on the risk prediction pre-training model and clustering to obtain the clustering type of each sample; fine-tuning the risk prediction pre-training model based on the samples of each clustering type to obtain a death risk prediction model corresponding to each clustering type; inputting the time series test data and baseline data of the patient to be predicted into the risk prediction pre-training model to obtain the feature representation; obtaining the clustering type corresponding to the patient to be predicted according to the feature representation, and obtaining the prediction result of the patient to be predicted based on the risk prediction model corresponding to the clustering type. The death risk prediction result is obtained quickly and accurately.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of all-cause mortality risk prediction, and particularly relates to a peritoneal dialysis patient all-cause mortality risk prediction method. BACKGROUND

[0002] Peritoneal dialysis is one of the common renal replacement therapy methods for patients with end-stage renal disease, but patients face a high risk of all-cause mortality. Accurate prediction of all-cause mortality risk in peritoneal dialysis patients is of great significance for optimizing treatment plans and improving patient outcomes. Traditional risk prediction methods rely on data from a single medical institution, which has limited data volume and uneven data distribution, resulting in insufficient accuracy and generalization ability of the prediction model.

[0003] Although there are machine learning-based all-cause mortality risk prediction models, most of these models are based on a single data source and cannot fully utilize the advantages of multi-source data. In addition, traditional methods have shortcomings in data privacy protection, making it difficult to meet the strict privacy requirements of medical data. SUMMARY

[0004] In view of the above analysis, the embodiments of the present application aim to provide a peritoneal dialysis patient all-cause mortality risk prediction method to solve the problem of low prediction accuracy.

[0005] In one aspect, the present application provides a peritoneal dialysis patient all-cause mortality risk prediction method, comprising the following steps:

[0006] Constructing a sample set based on the patient's time series test data and baseline data in each data center;

[0007] Training a neural network model based on the sample sets of multiple data centers to obtain a risk prediction pre-training model;

[0008] Based on the risk prediction pre-training model, obtaining the feature representation of each sample, and clustering the feature representation to obtain the clustering type of each sample;

[0009] Fine-tuning the risk prediction pre-training model based on the samples of each clustering type to obtain a mortality risk prediction model corresponding to each clustering type;

[0010] Inputting the time series test data and baseline data of the patient to be predicted into the risk prediction pre-training model to obtain the feature representation of the patient to be predicted; obtaining the clustering type corresponding to the patient to be predicted according to the feature representation, and obtaining the prediction result of the patient to be predicted based on the risk prediction model corresponding to the clustering type.

[0011] Based on the further improvement of the above method, the clustering of the feature representation to obtain the clustering type of each sample comprises:

[0012] S31, randomly initializing K cluster centers;

[0013] S32, sending the K cluster centers to each data center; each data center determines the cluster type of each sample based on the distance between each sample in the local sample set and the K cluster centers, and calculates a local cluster center;

[0014] S33, updating the K cluster centers based on the local cluster center of each data center and the cluster weight of each data center;

[0015] S34, judging the update accuracy of the K cluster centers, if all are less than a preset threshold, ending the clustering to obtain the cluster type of each sample; otherwise, returning to step S32.

[0016] Based on the further improvement of the above method, the cluster weight of each data center is calculated in the following way, which includes:

[0017] Each data center calculates the local data clustering effect based on the local cluster center; and obtains a first weight of each data center based on the local data clustering effect of each data center;

[0018] Based on the local cluster center of each data center, K first pre-update cluster centers are obtained;

[0019] For each data center, K second pre-update cluster centers corresponding to the data center are obtained based on the local cluster center of other data centers;

[0020] Based on the K first pre-update cluster centers and the K second pre-update cluster centers corresponding to the data center, a second weight of the data center is calculated;

[0021] Based on the first weight and the second weight of each data center, the cluster weight of each data center is obtained.

[0022] Based on the further improvement of the above method, each data center calculates the local data clustering effect based on the local cluster center using the following formula:

[0023] ;

[0024] Wherein, represents the clustering effect of the i-th data center, represents the cohesion degree of the j-th sample of the i-th data center, represents the separation degree of the j-th sample of the i-th data center, represents the number of samples of the i-th data center.

[0025] Based on the further improvement of the above method, the first weight is calculated using the following formula:

[0026] ;

[0027] wherein, denotes the first weight of the i-th data center, denotes the clustering effect of the i-th data center, and N denotes the number of data centers.

[0028] Based on the further improvement of the above method, the second weight of the data center is calculated based on the K first pre-updated clustering centers and the K second pre-updated clustering centers corresponding to the data center using the following formula:

[0029] ;

[0030] wherein, denotes the j-th first pre-updated clustering center, denotes the j-th second pre-updated clustering center corresponding to the i-th data center, denotes the j-th second pre-updated clustering center corresponding to the t-th data center, and N denotes the number of data centers, denotes the second weight of the i-th data center.

[0031] Based on the further improvement of the above method, the risk prediction model corresponding to each clustering type is obtained by fine-tuning the risk prediction pre-training model based on the samples of each clustering type, comprising:

[0032] For each clustering type, the third weight of each data center is determined based on the samples belonging to the clustering type in each data center;

[0033] The risk prediction model corresponding to each clustering type is obtained by fine-tuning the risk prediction pre-training model based on the third weight of each data center using federated learning.

[0034] Based on the further improvement of the above method, the third weight of each data center is determined based on the samples belonging to the clustering type in each data center using the following formula:

[0035] ;

[0036] wherein, denotes the sample quantity weight of the i-th data center, denotes the data quality weight of the i-th data center, and denotes an adjustment parameter, denotes the third weight of the i-th data center.

[0037] Based on the further improvement of the above method, the sample set of each data center includes a support set and a query set; the neural network model is trained based on the sample sets of multiple data centers to obtain a risk prediction pre-training model, including:

[0038] S20, constructing a neural network model, and sending the constructed neural network model to each data center;

[0039] S21, recording the parameters of the current neural network model as initial parameters;

[0040] S22, for each data center, training the neural network model based on the local support set to update the parameters of the local neural network model; calculating the loss of the updated local neural network model on the local query set to obtain a query loss;

[0041] S23, updating the initial parameters based on the query loss of each data center, and updating the parameters of the neural network model of each data center to the updated initial parameters;

[0042] S24, judging whether a stop condition is reached, if yes, stopping training to obtain a risk prediction pre-training model, otherwise, returning to step S21.

[0043] Based on the further improvement of the above method, the initial parameters are updated based on the query loss of each data center using the following formula:

[0044] ;

[0045] ;

[0046] wherein, represents the gradient of the total loss of N data centers, represents the total loss of N data centers, represents the query loss of the i-th data center, represents the training weight of the i-th data center, represents the learning rate, represents the initial parameters.

[0047] Compared with the prior art, the present application fully utilizes the data of multiple data sources by constructing a sample set in each data center, using the sample sets of multiple data centers to improve the accuracy of the trained model, and ensuring data security, obtaining a pre-training model, and performing clustering based on the feature representation of the sample, fine-tuning the pre-training model based on the sample of each clustering type, thereby training a corresponding death risk prediction model for each type to improve the accuracy of the model. Through the pre-training fine-tuning method, the performance of the model is not affected by the small number of a certain type, thereby improving the accuracy of the death risk prediction model corresponding to each clustering type. By obtaining the clustering type of the patient to be predicted, inputting the time series test data and baseline data of the patient to be predicted into the risk prediction model corresponding to the clustering type, the death risk prediction result of the patient to be predicted is quickly and accurately obtained.

[0048] In the present application, the above technical solutions can also be combined with each other to realize more preferred combination solutions. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification or will be understood by implementing the present application. The purpose and other advantages of the present application can be realized and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0049] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and are used to explain the principles of the application, but are not used to limit the scope of the application.

[0050] Figure 1 The flowchart of the peritoneal dialysis patient all-cause death risk prediction method of the embodiments of the present application. DETAILED DESCRIPTION

[0051] The preferred embodiments of the present application will be specifically described below in conjunction with the drawings, wherein the drawings constitute a part of this application and are used to illustrate the principles of the embodiments of the present application, but are not used to limit the scope of the present application.

[0052] One specific embodiment of the present application discloses a peritoneal dialysis patient all-cause death risk prediction method, as shown in Figure 1 The method comprises the following steps:

[0053] S1, constructing a sample set based on the time series test data and baseline data of the patient in each data center;

[0054] S2, training a neural network model based on the sample sets of multiple data centers to obtain a risk prediction pre-training model;

[0055] S3, obtaining a feature representation of each sample based on the risk prediction pre-training model, and clustering the feature representation to obtain a clustering type of each sample;

[0056] S4, fine-tuning the risk prediction pre-training model based on the samples of each cluster type to obtain a death risk prediction model corresponding to each cluster type;

[0057] S5, inputting the time series test data and baseline data of the patient to be predicted into the risk prediction pre-training model to obtain a feature representation of the patient to be predicted; obtaining a cluster type corresponding to the patient to be predicted according to the feature representation, and obtaining a prediction result of the patient to be predicted based on the risk prediction model corresponding to the cluster type.

[0058] Compared with the prior art, the training method of the peritoneal dialysis patient all-cause death risk prediction model provided in the embodiment fully utilizes the data of multiple data sources by constructing sample sets in each data center and using sample sets of multiple data centers, improves the accuracy of the trained model, ensures data security, clusters based on the feature representation of the samples after obtaining the pre-training model, fine-tunes the pre-training model based on the samples of each cluster type, and thus trains a corresponding death risk prediction model for each type, thereby improving the accuracy of the model. Through the pre-training and fine-tuning method, the performance of the model will not be affected by the small number of a certain type, thereby improving the accuracy of the death risk prediction model corresponding to each cluster type. By obtaining the cluster type of the patient to be predicted, inputting the time series test data and baseline data of the patient to be predicted into the risk prediction model corresponding to the cluster type, the death risk prediction result of the patient to be predicted can be quickly and accurately obtained.

[0059] In implementation, the time series test data of the patient includes blood routine time series data, biochemical time series data, and vital sign time series data (such as blood pressure, heart rate, etc.), which dynamically change over time. The baseline data includes demographic information (such as age, gender, etc.) of the patient and basic disease information, which are relatively static data.

[0060] In each data center, patient data of the center is collected, and a sample set is constructed. The label of the sample is whether the patient is dead or not, 1 indicating death and 0 indicating non-death.

[0061] In implementation, if some time points of a certain data in the time series data are missing data, the missing data can be filled by median filling, mean filling, etc. If a certain data is missing, a default value is used instead.

[0062] In implementation, the sample set of each data center includes a support set and a query set, that is, a part of the samples in the sample set is divided into a support set, and a part of the samples is divided into a query set.

[0063] Training a neural network model based on the sample sets of multiple data centers to obtain a risk prediction pre-training model, comprising:

[0064] S20, construct a neural network model, and send the constructed neural network model to each data center;

[0065] S21, record the parameters of the current neural network model as initial parameters;

[0066] S22, for each data center, update the parameters of the neural network model based on the local support set, and calculate the loss of the updated local neural network model on the local query set to obtain a query loss;

[0067] S23, update the initial parameters based on the query loss of each data center, and update the parameters of the neural network model of each data center to the updated initial parameters;

[0068] S24, determine whether a stop condition is reached, if yes, stop training to obtain a risk prediction pre-training model, otherwise, return to step S21.

[0069] In implementation, first, a neural network model is constructed, the constructed neural network model is sent to each data center, and the current parameters of the neural network model are recorded as initial parameters .

[0070] Specifically, the constructed neural network model comprises:

[0071] a multi-channel time series feature extraction module for extracting features of time series inspection data;

[0072] a baseline feature extraction module for extracting features of baseline data;

[0073] a feature fusion module for fusing the features of the time series inspection data and the features of the baseline data;

[0074] a prediction module for performing death risk prediction based on the fused features.

[0075] In implementation, the multi-channel time series feature extraction module comprises a plurality of channels, each channel is composed of a GRU unit, and each channel corresponds to an item of time series inspection data, such as blood routine data. The baseline feature extraction module adopts a learnable embedding matrix to map the baseline data to the same representation space as the features of the time series inspection data, i.e., the same feature dimension.

[0076] In implementation, the feature fusion module can adopt an attention mechanism to dynamically adjust the weights of different features, thereby improving the accuracy and interpretability of model prediction.

[0077] For each data center, first, the local neural network model is trained based on the support set in the local data set, and the parameters of the local neural network model are updated. In implementation, multi-step training can be performed, that is, the parameters of the local neural network model are updated multiple times, and the parameters updated in the last time are denoted as , denote the parameters of the local neural network model of the i-th data center.

[0078] After updating the parameters of the local neural network model, each data center calculates the query loss based on the updated local neural network model on the local query set.

[0079] In implementation, for the i-th data center, the sample on the local query set is input into the local neural network model (the model parameters are ), and the loss is calculated according to the output of the model. In implementation, the cross-entropy loss function can be used to calculate the loss of each sample, and the average loss of the samples in the query set is calculated to obtain the query loss .

[0080] After obtaining the query loss of each data center, the initial parameters are updated based on the query loss of each selected data center using the following formula:

[0081] ;

[0082] ;

[0083] wherein, denotes the gradient of the total loss of the N data centers, denotes the total loss of the N data centers, denotes the query loss of the i-th data center, denotes the training weight of the i-th data center, denotes the learning rate, denotes the initial parameters.

[0084] In implementation, the training weight of the data center can be determined according to the number and quality of the samples of the data center.

[0085] The parameters of the neural network model of each data center are updated to the updated initial parameters, so that the back propagation of the neural network model is completed based on multiple data centers.

[0086] After completing the back propagation once, if the current stop condition is reached, the training is ended, and the risk prediction pre-training model is obtained, and if not, the training is continued in S21 step until the stop condition of the model training is reached. In implementation, the stop condition can be that the preset training times are reached, or the total training loss reaches the preset accuracy.

[0087] The peritoneal transport type, complication condition, nutritional status and the like all affect the all-cause death risk of the peritoneal dialysis patient, different groups are different in patient distribution and statistical characteristics, and thus the predicted risk is different, in order to improve the accuracy of risk prediction, after obtaining the risk prediction pre-training model, the feature representation of each sample is obtained based on the risk prediction pre-training model, and the feature representation is clustered to obtain the cluster type of each sample.

[0088] In implementation, the feature representation of each sample is the feature output by the feature fusion module after inputting the time series test data and baseline data of the sample into the risk prediction pre-training model.

[0089] Specifically, the feature representation is clustered to obtain the cluster type of each sample, including:

[0090] S31, randomly initializing K cluster centers;

[0091] S32, sending the K cluster centers to each data center; each data center determines the cluster type of each sample based on the distance between each sample in the local sample set and the K cluster centers, and calculates the local cluster center;

[0092] S33, updating the K cluster centers based on the local cluster center of each data center and the cluster weight of each data center;

[0093] S34, judging the update accuracy of the K cluster centers, if all are less than a preset threshold, ending the clustering to obtain the cluster type of each sample; otherwise, returning to step S32.

[0094] In implementation, first, the K cluster centers are initialized, represented as Each cluster center represents a cluster type. The number of K can be set according to experience. The dimension of each cluster center is the same as that of the feature representation.

[0095] First, the K cluster centers are sent to each data center.

[0096] Each data center calculates the distance between each sample in the local sample set and the K cluster centers, and determines which cluster type each sample belongs to according to the distance. In implementation, the Euclidean distance can be used to calculate the distance between the sample and the cluster center. For each cluster type, the local cluster center of the data center is calculated according to the samples belonging to the type in the data center. In implementation, the centroid or mean of all samples belonging to the type can be calculated as the local cluster center corresponding to the type.

[0097] After obtaining the local cluster center corresponding to each data center, the K cluster centers are updated based on the local cluster center of each data center and the weight of each data center.

[0098] In implementation, the data quantity and quality of each data center are different because of different weights.

[0099] The clustering weight of each data center is calculated in the following manner, including:

[0100] S331, each data center calculates a local data clustering effect based on a local clustering center; and a first weight of each data center is obtained based on the local data clustering effect of each data center;

[0101] S332, K first pre-updated clustering centers are obtained based on the local clustering center of each data center; and for each data center, K second pre-updated clustering centers corresponding to the data center are obtained based on the local clustering center of other data centers;

[0102] S333, a second weight of the data center is calculated based on the K first pre-updated clustering centers and the K second pre-updated clustering centers corresponding to the data center;

[0103] S334, a clustering weight of each data center is obtained based on the first weight and the second weight of each data center.

[0104] In implementation, each data center calculates a local data clustering effect based on a local clustering center in the following formula:

[0105] ;

[0106] wherein, represents a clustering effect of the i th data center, represents an internal cohesion degree of a j th sample of the i th data center, represents a separation degree of the j th sample of the i th data center, represents a sample quantity of the i th data center.

[0107] The clustering effect represents a similarity degree of samples of the same type and a separation degree between different types.

[0108] In implementation, the internal cohesion degree of the j th sample of the i th data center is an average value of distances of the j th sample to other samples in a clustering type to which the j th sample belongs.

[0109] The separation degree of the j th sample is calculated in the following manner: an average distance of the j th sample to samples of each clustering type other than the clustering type to which the j th sample belongs is calculated, and a minimum value of the average distances is selected as the separation degree of the j th sample.

[0110] From the above calculation formula, the clustering effect of the i-th data center is a value between -1 and +1. The closer to +1 indicates the better clustering effect, and the closer to -1 indicates the worse clustering effect. The data center with poor clustering effect has a greater impact on clustering.

[0111] Therefore, after obtaining the local data clustering effect of each data center, the first weight is calculated by using the following formula:

[0112] ;

[0113] wherein, represents the first weight of the i-th data center.

[0114] In implementation, K first pre-updated clustering centers are obtained based on the local clustering centers of each data center. In implementation, for K clusters, the mean of the corresponding clustering centers of the local clustering centers of all data centers is calculated, thereby obtaining K first updated clustering centers, denoted as .

[0115] For each data center, K second pre-updated clustering centers corresponding to the data center are obtained based on the local clustering centers of other data centers. For example, for the i-th data center, the local clustering centers of other data centers except the data center are averaged to obtain the second pre-updated clustering center corresponding to the i-th data center, denoted as . The difference between the first pre-updated clustering center and the second pre-updated clustering center represents the influence degree of the i-th data center on future clustering.

[0116] Therefore, the second weight of the data center is calculated by using the following formula based on the K first pre-updated clustering centers and the K second pre-updated clustering centers corresponding to the data center:

[0117] ;

[0118] wherein, represents the j-th first pre-updated clustering center, represents the j-th second pre-updated clustering center corresponding to the i-th data center, represents the j-th second pre-updated clustering center corresponding to the t-th data center, and N represents the number of data centers, represents the second weight of the i-th data center.

[0119] For the i-th data center, the first weight and the second weight are averaged to obtain the clustering weight corresponding thereto.

[0120] Then, the K cluster centers are updated based on the local cluster center of each data center and the cluster weight of each data center. That is, the local cluster centers of all data centers are weighted and averaged to obtain new K cluster centers, and the weights added are cluster weights.

[0121] In implementation, the cluster weight is not artificially set or fixed, but dynamically changes according to data, because it is more in line with the actual objective situation, and the clustering accuracy is higher.

[0122] After updating the K cluster centers, it is determined whether the end condition is reached, and whether the update accuracy is less than the preset threshold. If the update accuracy of the K cluster centers is less than the preset threshold, the clustering is ended, and the clustering type of each sample is obtained; otherwise, the step S32 is returned to continue clustering until the end condition is met. In implementation, the end condition can also be the number of iterations.

[0123] After clustering, each data center determines which cluster type its sample belongs to according to the last K cluster centers.

[0124] Based on the samples of each cluster type, the risk prediction pre-training model is fine-tuned to obtain a risk prediction model corresponding to each cluster type, specifically including:

[0125] S41, for each cluster type, determining a third weight of each data center based on the samples belonging to the cluster type in each data center;

[0126] S42, fine-tuning the risk prediction pre-training model based on the third weight of each data center to obtain a risk prediction model corresponding to each cluster type.

[0127] For the kth cluster type, the number and quality of samples of this type in each data center are different, because the influence weight of fine-tuning is different.

[0128] Specifically, the third weight of each data center is determined based on the samples belonging to the cluster type in each data center using the following formula:

[0129] ;

[0130] Wherein, represents the sample number weight of the ith data center, represents the data quality weight of the ith data center, and represents an adjustment parameter, represents the third weight of the ith data center.

[0131] In implementation, for the kth clustering type, the third weight of each data center is calculated first. The sample quantity weight of the ith data center is the proportion of the sample quantity of the ith data center belonging to the kth clustering type to the total sample quantity of the kth clustering type.

[0132] The data quality weight of the ith data center is the average accuracy of the samples of the ith data center belonging to the kth clustering type. The accuracy of a sample is , wherein, represents the prediction result of the pre-trained model, and y represents the sample label.

[0133] In implementation, each clustering type refines the risk prediction pre-trained model based on the corresponding samples to obtain the corresponding risk prediction model. Since data sharing is not allowed, the refinement process can refer to the above steps S21-S24, and the neural network model is the risk prediction pre-trained model. The samples of each data center belonging to the kth clustering type are also divided into a support set and a query set, and the training weight of each data center is its third weight, thereby obtaining the risk prediction model corresponding to each clustering type.

[0134] For the to-be-predicted patient, the time series test data and baseline data of the to-be-predicted patient are input into the pre-trained model to obtain the feature representation of the to-be-predicted patient. According to the feature representation, the clustering type corresponding to the to-be-predicted patient can be obtained. Then, the time series test data and baseline data of the to-be-predicted patient are input into the risk prediction model corresponding to the clustering type of the to-be-predicted patient, so as to quickly and accurately obtain the death risk prediction result of the to-be-predicted patient.

[0135] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing related hardware. The program can be stored in a computer readable storage medium. The computer readable storage medium includes a magnetic disk, an optical disk, a read-only memory, a random access memory, and the like.

[0136] The above description is only a preferred embodiment of the present application, and the protection scope of the present application is not limited thereto. Any changes or replacements within the technical scope disclosed by the present application can be easily thought of by those skilled in the art, and should be covered within the protection scope of the present application.

Claims

1. A method of predicting the risk of all-cause mortality in a peritoneal dialysis patient, characterized in that, The method comprises the following steps: collecting time-series test data and baseline data of patients in each data center, and constructing a sample set based on the time-series test data and baseline data of the patients; training a neural network model based on the sample sets of multiple data centers to obtain a risk prediction pre-training model; obtaining a feature representation of each sample based on the risk prediction pre-training model, and clustering the feature representation to obtain a clustering type of each sample; fine-tuning the risk prediction pre-training model based on samples of each clustering type to obtain a death risk prediction model corresponding to each clustering type; inputting time-series test data and baseline data of a patient to be predicted into the risk prediction pre-training model to obtain a feature representation of the patient to be predicted; obtaining a clustering type corresponding to the patient to be predicted according to the feature representation, and obtaining a prediction result of the patient to be predicted based on the risk prediction model corresponding to the clustering type; clustering the feature representation to obtain a clustering type of each sample, comprising: S31, randomly initializing K clustering centers; S32, sending the K clustering centers to each data center; each data center determines the clustering type of each sample in the local sample set based on the distance between each sample and the K clustering centers, and calculates a local clustering center; S33, updating the K clustering centers based on the local clustering center of each data center and the clustering weight of each data center; S34, judging the update accuracy of the K clustering centers, if all are less than a preset threshold, ending the clustering to obtain the clustering type of each sample; otherwise, returning to step S32; The clustering weight of each data center is calculated in the following manner: each data center calculates a local data clustering effect based on the local clustering center; and obtains a first weight of each data center based on the local data clustering effect of each data center; obtaining K first pre-update clustering centers based on the local clustering center of each data center; for each data center, obtaining K second pre-update clustering centers corresponding to the data center based on the local clustering centers of other data centers; calculating a second weight of the data center based on the K first pre-update clustering centers and the K second pre-update clustering centers corresponding to the data center; obtaining the clustering weight of each data center based on the first weight and the second weight of each data center; each data center calculates a local data clustering effect based on the local clustering center in the following formula: wherein S i represents the clustering effect of the i-th data center, a ij represents the cohesion of the j-th sample of the i-th data center, b ij represents the separation of the j-th sample of the i-th data center, n i represents the number of samples of the i-th data center; the first weight is calculated in the following formula: where ω 1i represents the first weight of the i-th data center, S i represents the clustering effect of the i-th data center, and N represents the number of data centers; the second weight of the data center is calculated in the following formula based on the K first pre-update clustering centers and the K second pre-update clustering centers corresponding to the data center: wherein, denotes the jth first pre-updated cluster center, denotes the jth second pre-updated cluster center corresponding to the ith data center, denotes the jth second pre-updated cluster center corresponding to the tth data center, N denotes the number of data centers, ω i2 denotes the second weight of the ith data center; The sample set of each data center includes a support set and a query set; training a neural network model based on the sample sets of multiple data centers to obtain a risk prediction pre-training model, comprising: S20, constructing a neural network model, and sending the constructed neural network model to each data center; S21, recording the parameters of the current neural network model as initial parameters; S22, for each data center, training the neural network model based on the local support set to update the parameters of the local neural network model; calculating the loss of the updated local neural network model on the local query set to obtain a query loss; S23, updating the initial parameters based on the query loss of each data center, and updating the parameters of the neural network model of each data center to the updated initial parameters; S24, determining whether a stop condition is reached, if yes, stopping training to obtain a risk prediction pre-training model, otherwise, returning to step S21.

2. The method of prediction of the risk of all-cause mortality in a peritoneal dialysis patient according to claim 1, characterized in that, Based on the samples of each clustering type, the risk prediction pre-training model is fine-tuned to obtain a risk prediction model corresponding to each clustering type, comprising: For each clustering type, determining a third weight of each data center based on the samples belonging to the clustering type in each data center; Based on the third weight of each data center, the risk prediction pre-training model is fine-tuned by federated learning to obtain a risk prediction model corresponding to each clustering type.

3. The method of prediction of the risk of all-cause mortality in a peritoneal dialysis patient according to claim 2, characterized in that, The third weight of each data center is determined based on the samples belonging to the clustering type in each data center by using the following formula: ω 3i = a1w 1i + a2w 2i ; wherein w 1i represents the sample quantity weight of the i-th data center, w 2i represents the data quality weight of the i-th data center, and a1 and a2 represent adjustment parameters, w 3i represents the third weight of the i-th data center, the sample quantity weight of the i-th data center is the proportion of the sample quantity of the i-th data center belonging to the cluster type in the total sample quantity of the cluster type, and the data quality weight of the i-th data center is the average accuracy of the sample of the i-th data center belonging to the cluster type.

4. The method of predicting the risk of all-cause mortality of a peritoneal dialysis patient according to claim 1, characterized in that, The initial parameters are updated based on the query loss of each data center by using the following formula: wherein, represents the gradient of the total loss of N data centers, represents the total loss of N data centers, Loss(θ i ) represents the query loss of the i-th data center, ω i represents the training weight of the i-th data center, β represents a learning rate, represents an initial parameter, θ i represents the parameter of the local neural network model of the i-th data center.

Images