A Patient Clustering and Survival Risk Prediction Method and System Based on Deep Probabilistic Models

Abstract

This invention provides a method and system for patient clustering and survival risk prediction based on a deep probabilistic model, belonging to the field of survival analysis technology. The invention first acquires user health data under right censoring conditions and preprocesses this data. Then, based on a feature extractor within a multi-task learning framework, it extracts feature representations of the preprocessed right-censored user health data. Based on these feature representations, a deep probabilistic model is used to obtain data distribution information, which is then reconstructed and optimized. Next, without considering parameter assumptions, a category-level risk function is estimated based on a survival prediction multi-task sub-network, and individual risk estimation results are obtained based on this risk function. Finally, based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, clustering prediction results and event time prediction results are obtained. This invention improves the accuracy of clustering and event time prediction results.

Description

Technical Field

[0001] This invention relates to the field of survival analysis technology, specifically to a method and system for patient clustering and survival risk prediction based on a deep probability model. Background Technology

[0002] Survival analysis is a common tool for predicting risk events. This method combines statistical and machine learning models to estimate the probability of a risk event occurring at a specific future time point *t*, also known as time-to-event prediction. In the healthcare field, survival analysis is primarily used to predict mortality or regression events. It can accurately identify a patient's clinical risk, discover explainable patient subgroups, and assist clinicians in adjusting their decision-making.

[0003] With the development of deep learning, traditional regression methods based on parametric assumptions, such as proportional hazards and accelerated time to failure, have been optimized by nonlinear modeling methods based on deep neural networks. This allows researchers to directly conduct risk estimation without considering parametric assumptions. Currently, survival analysis methods mainly focus on deep learning-based survival analysis research, including: 1) survival analysis research based on complex deep learning, which often uses advanced and complex deep learning models to improve the accuracy of time-event predictions; and 2) survival analysis research based on survival curve calibration, which calibrates survival estimation results by implicitly defining the survival distribution, thereby improving model reliability and enhancing model interpretability.

[0004] However, when applied to survival analysis research in the medical field, the lack of interpretability results in poor interpretability of the predicted survival analysis results. Nevertheless, it was found that patient subgroups can enhance the interpretability of deep learning models. In summary, current technologies cannot simultaneously perform class assignment and time-event prediction in survival analysis, nor can they simultaneously improve the overall performance of both. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a patient clustering and survival risk prediction method and system based on a deep probabilistic model, which solves the problem that existing technologies cannot simultaneously improve the comprehensive performance of category assignment and time-event prediction.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, this invention proposes a patient clustering and survival risk prediction method based on a deep probability model, the method comprising:

[0010] Obtain user health data under right censoring conditions, and preprocess the user health data under right censoring conditions;

[0011] A feature extractor based on a multi-task learning framework extracts feature representations of the preprocessed user health data under the right censoring condition; the feature extractor of the multi-task learning framework is an encoder of a deep probabilistic model.

[0012] Based on the feature representation, the deep probability model is used to obtain data distribution information, and the data distribution information is reconstructed and optimized; the data distribution information includes the prior probability distribution information of latent variables based on class membership in the feature representation;

[0013] Without considering parameter assumptions, a risk function at the category level is estimated based on a survival prediction multi-task subnetwork, and individual risk estimation results are obtained based on the risk function.

[0014] Based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, clustering prediction results and event time prediction results are obtained.

[0015] Preferably, the step of obtaining data distribution information based on the feature representation using the deep probability model, and reconstructing and optimizing the data distribution information includes:

[0016] S31. Based on the feature representation, the prior probability distribution of latent variables based on class membership prior is obtained using Gaussian mixture prior, and the posterior probability distribution of latent variables obtained by the encoder and reparameterization technique is made to approximate the prior probability distribution of latent variables based on KL divergence.

[0017] S32. Based on the posterior probability distribution of the latent variables, the original features of the feature representation are reconstructed using a feature reconstruction decoder.

[0018] Preferably, in step S31, based on the feature representation, a prior probability distribution of latent variables based on class membership prior is obtained using Gaussian mixture prior, and the posterior probability distribution of latent variables obtained by the encoder and reparameterization technique is approximated by the prior probability distribution of latent variables based on KL divergence, expressed by the formula:

[0019]

[0020]

[0021]

[0022] Among them, the variational posterior distribution Representing latent variables and categories The posterior distribution of membership degrees; the first term For refactoring terms, it indicates that they are based on latent variables. and parameterization Feature reconstruction decoder reconstructs original features ; Second item This prompts the model to utilize a survival prediction multi-task subnetwork parameterized with φ for group-level event-time prediction. ; Third item This indicates that the variational posterior distribution is obtained by using the KL divergence. Approximates the prior distribution of latent variables and class membership. ;in, , The prior probability distribution representing the membership degree of category c; latent variable ; and Let represent the mean and variance of the Gaussian mixture prior distribution determined by category c, respectively.

[0023] Preferably, step S32, reconstructing the original features of the feature representation using a feature reconstruction decoder based on the posterior probability distribution of the latent variables, includes the following formula:

[0024]

[0025]

[0026] in, Represents the reconstruction of the original input features; =Bernoulli(x; L and W represent the number of Monte Carlo samplings and the dimension of the covariates, respectively; and ;in, and The parameters are respectively and The feature encoder and reconstruction decoder.

[0027] Preferably, obtaining the clustering prediction results and event time prediction results based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results includes:

[0028] Let (Z,X) and T be conditionally independent of patient category C, and C and X be conditionally independent of latent variable Z. Then the joint probability distribution is labeled as follows:

[0029]

[0030]

[0031] Among them, variational distribution , and Let z and c represent the posterior estimates of the membership degrees of the latent variable z and the category c, respectively. This indicates the use of parameterization. The decoder reconstructs the original features.

[0032] Secondly, this invention also proposes a patient clustering and survival risk prediction system based on a deep probabilistic model, the system comprising:

[0033] The data acquisition and preprocessing module is used to acquire user health data under right censoring conditions and preprocess the user health data under right censoring conditions.

[0034] The data feature representation extraction module is used to extract feature representations of the preprocessed user health data under the right censoring condition using a feature extractor based on a multi-task learning framework; the feature extractor of the multi-task learning framework is an encoder of a deep probabilistic model.

[0035] The data distribution information reconstruction module is used to obtain data distribution information based on the feature representation and the deep probability model, and to reconstruct and optimize the data distribution information; the data distribution information includes the prior probability distribution information of latent variables based on the category membership prior in the feature representation;

[0036] The individual risk estimation module is used to estimate the category-level risk function based on the survival prediction multi-task sub-network without considering parameter assumptions, and to obtain the individual risk estimation result based on the risk function.

[0037] The result acquisition module is used to obtain clustering prediction results and event time prediction results based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results.

[0038] Preferably, the data distribution information reconstruction module, based on the feature representation, uses the deep probability model to obtain data distribution information, and reconstructs and optimizes the data distribution information, including:

[0039] S31. Based on the feature representation, the prior probability distribution of latent variables based on class membership prior is obtained using Gaussian mixture prior, and the posterior probability distribution of latent variables obtained by the encoder and reparameterization technique is made to approximate the prior probability distribution of latent variables based on KL divergence.

[0040] S32. Based on the posterior probability distribution of the latent variables, the original features of the feature representation are reconstructed using a feature reconstruction decoder.

[0041] Preferably, in step S31, based on the feature representation, a prior probability distribution of latent variables based on class membership prior is obtained using Gaussian mixture prior, and the posterior probability distribution of latent variables obtained by the encoder and reparameterization technique is approximated by the prior probability distribution of latent variables based on KL divergence, expressed by the formula:

[0042]

[0043]

[0044]

[0045] Among them, the variational posterior distribution Representing latent variables and categories The posterior distribution of membership degrees; the first term For refactoring terms, it indicates that they are based on latent variables. and parameterization Feature reconstruction decoder reconstructs original features ; Second item This prompts the model to utilize a survival prediction multi-task subnetwork parameterized with φ for group-level event-time prediction. ; Third item This indicates that the variational posterior distribution is obtained by using the KL divergence. Approximates the prior distribution of latent variables and class membership. ;in, , The prior probability distribution representing the membership degree of category c; latent variable ; and Let represent the mean and variance of the Gaussian mixture prior distribution determined by category c, respectively.

[0046] Preferably, step S32, reconstructing the original features of the feature representation using a feature reconstruction decoder based on the posterior probability distribution of the latent variables, includes the following formula:

[0047]

[0048]

[0049] in, Represents the reconstruction of the original input features; =Bernoulli(x; L and W represent the number of Monte Carlo samplings and the dimension of the covariates, respectively; and ;in, and The parameters are respectively and The feature encoder and reconstruction decoder.

[0050] Preferably, the result acquisition module, based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, acquires clustering prediction results and event time prediction results, including:

[0051] Let (Z,X) and T be conditionally independent of patient category C, and C and X be conditionally independent of latent variable Z. Then the joint probability distribution is labeled as follows:

[0052]

[0053]

[0054] Among them, variational distribution , and Let z and c represent the posterior estimates of the membership degrees of the latent variable z and the category c, respectively. This indicates the use of parameterization. The decoder reconstructs the original features.

[0055] (III) Beneficial Effects

[0056] This invention provides a method and system for patient clustering and survival risk prediction based on a deep probabilistic model. Compared with existing technologies, it has the following advantages:

[0057] 1. This invention first acquires user health data under right-censoring conditions and preprocesses it. Then, based on a feature extractor (i.e., the encoder of a deep probabilistic model) within a multi-task learning framework, it extracts feature representations from the preprocessed right-censored user health data. Based on these feature representations, it uses a deep probabilistic model to obtain data distribution information and reconstructs and optimizes this information. Next, without considering parameter assumptions, it estimates a risk function at the category level based on a survival prediction multi-task sub-network and obtains individual risk estimation results based on this risk function. Finally, based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, it obtains clustering prediction results and time-event (event time) prediction results. This invention, based on optimized data distribution information and the variational evidence lower bound of individual risk estimation results, obtains accurate clustering and time-event prediction results, while simultaneously solving the problem that existing technologies cannot simultaneously improve the comprehensive performance of category assignment and time-event prediction.

[0058] 2. The deep survival analysis network based on the multi-task learning framework proposed in this invention combines the clustering survival prediction multi-task sub-network with the deep probabilistic model based on autoencoder to jointly optimize time-event prediction and clustering tasks, and simultaneously improve the comprehensive performance of category assignment and time-event prediction, thereby obtaining accurate clustering and time-event prediction results.

[0059] 3. This invention estimates both clustering and time-event prediction results without considering survival distribution assumptions, improving the scalability of the model in real-world data. Compared with other existing methods, the latent space in the proposed method only obtains distribution information based on covariates, while the time-event prediction network is directly connected to the encoder to obtain survival estimation results. Through joint modeling of the two, more accurate class-weighted time-event prediction results and posterior estimates of class membership can be obtained. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of a patient clustering and survival risk prediction method based on a deep probability model according to the present invention;

[0062] Figure 2 This is a flowchart illustrating an embodiment of the patient clustering and survival risk prediction method based on a deep probability model according to the present invention.

[0063] Figure 3 This is a structural diagram of the deep survival analysis network model based on a multi-task learning framework in an embodiment of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] This application provides a patient clustering and survival risk prediction method and system based on a deep probabilistic model, which solves the problem that existing technologies cannot simultaneously improve the comprehensive performance of category assignment and time-event prediction, and achieves the goal of high-precision survival analysis by performing category assignment and time-event prediction simultaneously.

[0066] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0067] To address the limitation of existing technologies in simultaneously improving the comprehensive performance of category assignment and time-event (event-time) prediction, and to achieve high-precision survival analysis for both categories assignment and time-event prediction, this application's technical solution preprocesses user health data under right censoring conditions. It then uses the encoder of a deep probabilistic model as the feature extractor of a multi-task network. Based on the preprocessed data, feature representations are obtained. By combining Gaussian mixture priors and a feature reconstruction decoder, the latent space embedding of input covariates is achieved to obtain data distribution information. Utilizing the survival prediction network within the multi-task sub-network, the group-level risk function is estimated without considering survival distribution assumptions. Accurate clustering and time-event prediction results are obtained by optimizing the combination of survival estimation and variational evidence lower bounds of data distribution. Compared to existing methods, this application's survival clustering and prediction method achieves more ideal results in both clustering and time-event prediction tasks.

[0068] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0069] A major challenge in survival analysis techniques is handling the censoring of patient-related health data, which aims to utilize the clinical characteristics of an individual patient. Time of occurrence or deletion of the event and deletion status Estimate the patient's risk probability h(t) and the cumulative distribution function F(t) of the probability of risk events such as death and readmission within a certain period of time.

[0070] The characteristic information of patients in the survival analysis dataset can be represented as: .in, The event of interest for the i-th patient is likely to not occur during the observation period, including cases where the patient survives to the end of follow-up, is lost to follow-up, or the cause of death is not the event of interest (right censoring). These are the most common censoring scenarios in survival analysis. The patient's class assignment is represented as c∈{1,...,N}. The technical solution of this application uses the survival dataset to carry out subsequent time-event prediction and clustering tasks.

[0071] Based on this, the technical solution proposed in this application is as follows:

[0072] Firstly, this invention proposes a patient clustering and survival risk prediction method based on a deep probabilistic model, see [link to relevant documentation]. Figure 1 The method includes:

[0073] S1. Obtain user health data under the right censoring condition, and preprocess the user health data under the right censoring condition;

[0074] S2. A feature extractor based on a multi-task learning framework extracts feature representations of the preprocessed user health data under the right censoring condition; the feature extractor of the multi-task learning framework is an encoder of a deep probabilistic model;

[0075] S3. Based on the feature representation, the data distribution information is obtained using the deep probability model, and the data distribution information is reconstructed and optimized; the data distribution information includes the prior probability distribution information of the latent variables based on the prior of the class membership in the feature representation;

[0076] S4. Without considering parameter assumptions, estimate the risk function of the category level based on the survival prediction multi-task sub-network, and obtain the individual risk estimation result based on the risk function;

[0077] S5. Based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, obtain the clustering prediction results and the time-event prediction results.

[0078] As can be seen, this embodiment first acquires user health data under right censoring conditions and preprocesses it. Then, based on the feature extractor (i.e., the encoder of the deep probabilistic model) of the multi-task learning framework, it extracts feature representations of the preprocessed right-censored user health data. Based on these feature representations, it uses the deep probabilistic model to obtain data distribution information and reconstructs and optimizes this information. Next, without considering parameter assumptions, it estimates the category-level risk function based on the survival prediction multi-task sub-network and obtains individual risk estimation results based on this risk function. Finally, based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, it obtains clustering prediction results and time-event prediction results. This embodiment obtains accurate clustering and time-event prediction results by optimizing the data distribution information and the variational evidence lower bound of the individual risk estimation results.

[0079] The following is in conjunction with the appendix Figure 1-3 The following details the implementation process of an embodiment of the present invention, including explanations of the specific steps S1-S5.

[0080] S1. Obtain user health data under the right censoring condition, and preprocess the user health data under the right censoring condition.

[0081] First, right-censored health data related to the patient's medical and health conditions is acquired. This mainly includes health data under right-censored conditions such as patients surviving to the end of follow-up, loss to follow-up during the follow-up period, and patients whose cause of death was not an event of interest. Patient health data is primarily acquired through publicly available online medical datasets, electronic medical record data obtained in cooperation with hospitals, and simulation data generated based on the characteristics of survival analysis. Then, the acquired user health data is preprocessed to format the data. Preprocessing of the user health data includes handling missing values, normalization of continuous variables, and vectorization of discrete data, specifically:

[0082] Handle missing values ​​in user health data. Read in the data fields to be processed one by one, determine if the data is null or abnormal, and fill in the missing values ​​using the mean imputation method if it is.

[0083] Continuous variable normalization is performed on user health data. The method for continuous variable normalization can be expressed by the formula:

[0084]

[0085] in, This represents the normalized user health data, where x is the original user health data, u is the sample mean, and σ is the sample standard deviation.

[0086] Discrete data vectorization is performed on user health data under right censoring conditions. The method for vectorizing discrete attribute data is as follows: using... one-hot The encoding method treats each value of a discrete feature as a state. If there are N different values ​​for a feature, the feature is abstracted into N different states. Among the N states, only one state bit is 1, and the other state bits are 0. That is, N states correspond to N strings of numbers, each number has only one 1 and the others are 0. Each string of numbers represents a state and is not repeated.

[0087] The above steps are used to obtain patient health data, which is then preprocessed to finally output patient health data in tabular format.

[0088] S2. A feature extractor based on a multi-task learning framework extracts feature representations of the preprocessed user health data under the right censoring condition; the feature extractor of the multi-task learning framework is an encoder of a deep probabilistic model.

[0089] In practice, to fully learn the feature information of preprocessed patient health data and thus assist in obtaining more accurate clustering and time-event prediction results, a preferred approach in this embodiment is to input the preprocessed patient health data into a feature extractor to learn data feature representations. This feature extractor is a deep probabilistic model (DPP). Figure 2 The right side of the middle represents the encoder of the deep probabilistic model, that is, the encoder of the deep probabilistic model is used as the survival prediction multi-task sub-network. Figure 2 The feature extractor (represented on the left in the middle section) learns the feature representation vector of patient health data from the survival prediction multi-task sub-network. In this embodiment, the deep probabilistic model and the survival prediction multi-task sub-network together form a multi-task learning framework, which can be found in the following documentation. Figure 2 and Figure 3 , Figure 3 This is a structural diagram of the deep survival analysis network model based on a multi-task learning framework in this embodiment. Figure 2 This is a flowchart of an embodiment of the patient clustering and survival risk prediction method based on a deep probability model in this example.

[0090] S3. Based on the feature representation, the data distribution information is obtained using the deep probability model, and the data distribution information is reconstructed and optimized; the data distribution information includes the prior probability distribution information of the latent variables based on the class membership prior in the feature representation.

[0091] S31. Based on the feature representation, the prior probability distribution of latent variables based on class membership prior is obtained using Gaussian mixture prior, and the posterior probability distribution of latent variables obtained by the encoder and reparameterization technique is made to approximate the prior probability distribution of latent variables based on KL divergence.

[0092] This paper combines Gaussian mixture prior and feature reconstruction decoder to embed the latent space of input features, thereby obtaining the prior probability distribution of latent variables based on class membership prior. KL divergence is then used to enable the encoder and reparameterization techniques to obtain the posterior probability distribution of the latent variables that approximates the prior distribution. The specific steps and formulas are as follows:

[0093] Let the prior probability distribution of the membership degree of category c be... , Latent variables ; and Let represent the mean and variance of the Gaussian mixture prior distribution determined by category c, respectively.

[0094] The feature representation obtained by the encoder is used to estimate the posterior distribution information of the mixture Gaussian. Then, KL divergence is used to make the posterior probability distribution of the latent variable close to the prior probability distribution of the latent variable, thereby optimizing the lower bound of variational evidence, as shown in the following formula (1).

[0095]

[0096]

[0097]

[0098] Among them, the first term in formula (1) The first term is the reconstruction term, used to generate reconstruction features; the second term... Encourage the model to utilize parameterization The time-event prediction subnetwork performs group-level time-event prediction; the third item This indicates that the variational posterior distribution containing latent variables is obtained by using KL divergence. Approximate to the prior distribution .

[0099] Specifically, the variational posterior distribution Representing latent variables and categories The posterior distribution of membership degrees; the first term For refactoring terms, it indicates that they are based on latent variables. and parameterization The feature reconstruction decoder reconstructs the original feature x; the second term This prompts the model to utilize a survival prediction multi-task subnetwork parameterized with φ for group-level event-time prediction. ; Third item This indicates that the variational posterior distribution is obtained by using the KL divergence. Approximates the prior distribution of latent variables and class membership. ;in, , The prior probability distribution representing the membership degree of category c; latent variable ; and Let represent the mean and variance of the Gaussian mixture prior distribution determined by category c, respectively.

[0100] S32. Based on the posterior probability distribution of the latent variables, the original features of the feature representation are reconstructed using a feature reconstruction decoder.

[0101] When reconstructing the original features based on the posterior information of latent variables and the feature reconstruction decoder, the specific process is as follows: Formula (2):

[0102]

[0103]

[0104] in, Represents the reconstruction of the original input features; =Bernoulli(x; ); L (L is usually 1) and W represent the number of Monte Carlo samplings and the dimension of the covariates, respectively; and , , . and The parameters are respectively and The feature encoder and reconstruction decoder are described. A reparameterization technique is employed to address the non-differentiability problem in the stochastic gradient descent process caused by Monte Carlo sampling. .

[0105] S4. Without considering parameter assumptions, estimate the risk function of the category level based on the survival prediction multi-task sub-network, and obtain the individual risk estimation result based on the risk function.

[0106] In the field of survival analysis technology, the definition For survival function, Indicates a given time period The probability that an event of interest does not occur is derived from the cumulative distribution function of the event's occurrence time. Conclusion:

[0107]

[0108] Risk function It is a conditional probability, representing the probability that the outcome event will occur within a subsequent time interval [t, t + ∆t) for an individual surviving at time t. It can be expressed by the formula:

[0109]

[0110] Where, the probability density function of F(t) is expressed as .

[0111] Use parameterization The survival prediction multi-task subnetwork directly predicts the survival risk at each time quantile under each category condition. Without considering parameter assumptions about the survival distribution, the softplus function is used to transform the output of the survival prediction multi-task subnetwork into different discrete-time quantiles within a certain time interval. Risk prediction results A timeline is predefined based on the total survival time. The C-index was evaluated at the three time quantiles of 0.25, 0.5, and 0.75. The weights assigned to each group depended on the sampled Bayesian posterior distribution of group membership. Therefore, the individual survival risk is obtained by weighted summation of the group-level survival risk results.

[0112] Based on cumulative risk function Acquired survival function (Corresponding formula (5)) and considering deletion ( The likelihood functions for the case where =0) (corresponding to formula (6)) are expressed as follows:

[0113]

[0114]

[0115]

[0116]

[0117] This embodiment estimates both clustering and time-event prediction results without considering survival distribution assumptions, which can improve the scalability of the model in real-world data. Compared with other existing methods, the latent space in the proposed method is only used to obtain distribution information based on covariates, while the time-event prediction network is directly connected to the encoder to obtain survival estimation results. Through joint modeling of the two, more accurate class-weighted time-event prediction results and posterior estimates of class membership can be obtained.

[0118] S5. Based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, obtain the clustering prediction results and the time-event prediction results.

[0119] Let (Z,X) and T be conditionally independent of patient category C, and C and X be conditionally independent of latent variable Z. The joint probability distribution is:

[0120]

[0121]

[0122] Variational distribution , and represents the posterior estimates of latent variables and class membership, respectively. This indicates the use of parameterization. The decoder reconstructs the original features. Since clustering involves Bayesian estimation of censored individuals, probability density functions are used respectively. and cumulative distribution function Calculate the individual in the censoring ( =0) and non-censoring ( Category membership degree in case =1)

[0123]

[0124]

[0125]

[0126] During model testing, if semi-supervised clustering based on individual survival time is considered, the clustering result is: If the individual's survival time is unknown, the clustering result is: .in:

[0127] 1) Based on the posterior distribution of latent variables Category Prior Individual risk estimation results at the group level Calculate the posterior probability estimate of the class assignment

[0128] This represents an estimate of the distribution of latent variables:

[0129]

[0130]

[0131]

[0132]

[0133]

[0134] Where J represents a vector And the dimension of z.

[0135] This indicates an optimization for category priors:

[0136]

[0137]

[0138]

[0139] ( () indicates optimization for variational posterior:

[0140]

[0141]

[0142]

[0143]

[0144] This represents the optimization for class posterior estimation:

[0145]

[0146]

[0147] 2) Posterior estimation based on category assignment Weighted group level individual risk Obtain individual risk estimation results .

[0148] This indicates an optimization of the risk estimation results:

[0149]

[0150]

[0151]

[0152]

[0153]

[0154]

[0155] This completes the entire process of a patient clustering and survival risk prediction method based on a deep probabilistic model, as described in this embodiment.

[0156] This embodiment validates the clustering and prediction accuracy on synthetic and semi-synthetic datasets through experiments, and further evaluates the time-event prediction accuracy using four real-world datasets. The results show that the model in this embodiment achieves optimal performance in both the time-event prediction task and the clustering task for populations with different survival distributions, solving the problem that previous survival analysis methods involving survival clustering tasks could not simultaneously improve performance in both survival clustering and time-event prediction.

[0157] Example 2:

[0158] Secondly, the present invention also provides a patient clustering and survival risk prediction system based on a deep probabilistic model, the system comprising:

[0159] The data acquisition and preprocessing module is used to acquire user health data under right censoring conditions and preprocess the user health data under right censoring conditions.

[0160] The data feature representation extraction module is used to extract feature representations of the preprocessed user health data under the right censoring condition using a feature extractor based on a multi-task learning framework; the feature extractor of the multi-task learning framework is an encoder of a deep probabilistic model.

[0161] The data distribution information reconstruction module is used to obtain data distribution information based on the feature representation and the deep probability model, and to reconstruct and optimize the data distribution information; the data distribution information includes the prior probability distribution information of latent variables based on the category membership prior in the feature representation;

[0162] The individual risk estimation module is used to estimate the category-level risk function based on the survival prediction multi-task sub-network without considering parameter assumptions, and to obtain the individual risk estimation result based on the risk function.

[0163] The result acquisition module is used to obtain clustering prediction results and event time prediction results based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results.

[0164] Optionally, the data distribution information reconstruction module, based on the feature representation, uses the deep probability model to obtain data distribution information and reconstructs and optimizes the data distribution information, including:

[0165] S31. Based on the feature representation, the prior probability distribution of latent variables based on class membership prior is obtained using Gaussian mixture prior, and the posterior probability distribution of latent variables obtained by the encoder and reparameterization technique is made to approximate the prior probability distribution of latent variables based on KL divergence.

[0166] S32. Based on the posterior probability distribution of the latent variables, the original features of the feature representation are reconstructed using a feature reconstruction decoder.

[0167] Optionally, in step S31, based on the feature representation, a prior probability distribution of latent variables based on class membership prior is obtained using Gaussian mixture prior, and the posterior probability distribution of latent variables obtained by the encoder and reparameterization technique is approximated by the prior probability distribution of latent variables based on KL divergence, expressed by the formula:

[0168]

[0169]

[0170]

[0171] Among them, the variational posterior distribution Representing latent variables and categories The posterior distribution of membership degrees; the first term For refactoring terms, it indicates that they are based on latent variables. and parameterization Feature reconstruction decoder reconstructs original features ; Second item This prompts the model to utilize a survival prediction multi-task subnetwork parameterized with φ for group-level event-time prediction. ; Third item This indicates that the variational posterior distribution is obtained by using the KL divergence. Approximates the prior distribution of latent variables and class membership. ;in, , The prior probability distribution representing the membership degree of category c; latent variable ; and Let represent the mean and variance of the Gaussian mixture prior distribution determined by category c, respectively.

[0172] Optionally, step S32, reconstructing the original features of the feature representation using a feature reconstruction decoder based on the posterior probability distribution of the latent variables, includes the following formula:

[0173]

[0174]

[0175] in, Represents the reconstruction of the original input features; =Bernoulli(x; L and W represent the number of Monte Carlo samplings and the dimension of the covariates, respectively; and ;in, and The parameters are respectively and The feature encoder and reconstruction decoder.

[0176] Optionally, the result acquisition module, based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, acquires clustering prediction results and event time prediction results, including:

[0177] Let (Z,X) and T be conditionally independent of patient category C, and C and X be conditionally independent of latent variable Z. Then the joint probability distribution is labeled as follows:

[0178]

[0179]

[0180] Among them, variational distribution , and Representing latent variables respectively Posterior estimate of membership degree of category c; This indicates the use of parameterization. The decoder reconstructs the original features.

[0181] It is understood that the patient clustering and survival risk prediction system based on deep probabilistic models provided in this embodiment of the invention corresponds to the aforementioned patient clustering and survival risk prediction method based on deep probabilistic models. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the patient clustering and survival risk prediction method based on deep probabilistic models, and will not be repeated here.

[0182] In summary, compared with existing technologies, it has the following beneficial effects:

[0183] 1. This invention first acquires user health data under right-censoring conditions and preprocesses it. Then, based on a feature extractor (i.e., the encoder of a deep probabilistic model) within a multi-task learning framework, it extracts feature representations of the preprocessed right-censored user health data. Based on these feature representations, it uses a deep probabilistic model to obtain data distribution information and reconstructs and optimizes this information. Next, without considering parameter assumptions, it estimates a risk function at the class level based on a survival prediction multi-task sub-network and obtains individual risk estimation results based on this risk function. Finally, based on the optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, it obtains clustering prediction results and time-event prediction results. This invention, based on optimized data distribution information and the variational evidence lower bound of the individual risk estimation results, obtains accurate clustering and time-event prediction results, while simultaneously solving the problem that existing technologies cannot simultaneously improve the comprehensive performance of class assignment and time-event prediction.

[0184] 2. The deep survival analysis network based on the multi-task learning framework proposed in this invention combines the clustering survival prediction multi-task sub-network with the deep probabilistic model based on autoencoder to jointly optimize time-event prediction and clustering tasks, and simultaneously improve the comprehensive performance of category assignment and time-event prediction, thereby obtaining accurate clustering and time-event prediction results.

[0185] 3. This invention estimates both clustering and time-event prediction results without considering survival distribution assumptions, improving the scalability of the model in real-world data. Compared with other existing methods, the latent space in the proposed method only obtains distribution information based on covariates, while the time-event prediction network is directly connected to the encoder to obtain survival estimation results. Through joint modeling of the two, more accurate class-weighted time-event prediction results and posterior estimates of class membership can be obtained.

[0186] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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.

[0187] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep probabilistic model based patient clustering and survival risk prediction method, characterized in that, The method comprises: obtaining user health data under right-censoring condition and preprocessing the user health data under the right-censoring condition; extracting feature representation of the preprocessed user health data under the right-censoring condition based on a feature extractor of a multi-task learning framework; the feature extractor of the multi-task learning framework is an encoder of a deep probabilistic model; obtaining data distribution information based on the deep probabilistic model based on the feature representation, and reconstructing and optimizing the data distribution information; the data distribution information comprises implicit variable prior probability distribution information based on category membership prior in the feature representation; estimating a category-level risk function based on a survival prediction multi-task subnetwork without considering parameter assumptions, and obtaining individual risk estimation results based on the risk function; obtaining clustering prediction results and event time prediction results based on a variational evidence lower bound of the optimized data distribution information and the individual risk estimation results.

2. The method of claim 1, wherein, The method comprises: S31, obtaining implicit variable prior probability distribution based on category membership prior based on a Gaussian mixture prior based on the feature representation, and making posterior probability distribution of the implicit variable obtained by the encoder and the reparameterization technique approximate the implicit variable prior probability distribution based on KL divergence; S32, reconstructing original features of the feature representation based on the posterior probability distribution of the implicit variable by using a feature reconstruction decoder.

3. The method of claim 2, wherein, The S31, obtaining implicit variable prior probability distribution based on category membership prior based on a Gaussian mixture prior based on the feature representation, and making posterior probability distribution of the implicit variable obtained by the encoder and the reparameterization technique approximate the implicit variable prior probability distribution based on KL divergence, is expressed by a formula as follows: wherein the variational posterior distribution q(z, c|x, t) = q(z|x)q(c|t, x) represents the posterior distribution of the membership of the latent variable z and the class c; the first term E q(z,c∣x,t) logp(x|z; θ) is a reconstruction term, representing reconstructing the original feature x based on the latent variable z and the feature reconstruction decoder parameterized as θ; the second term encourages the model to utilize the survival prediction multi-task subnetwork parameterized as to perform event time prediction at the group class level The third term D KL (q(z, c|x, t)||p(z, c; μ c ,Σ c ,π)) represents utilizing the KL divergence to make the variational posterior distribution q(z, c|x, t) approach the latent variable and class membership prior distribution p(z, c) = p(z|c; μ c ,Σ c )p(c; π); wherein p(c; π) = π c , π∈R N represents the membership prior probability distribution of the class c; the latent variable z ~ p(z|c; {μ1, …, μ N}, {Σ1, …, Σ N}) = N(z; μ c ,Σ c ); μ c and Σ c respectively represent the mean and variance of the Gaussian mixture prior distribution determined by the class c.

4. The method of claim 2, wherein, The S32, reconstructing original features of the feature representation based on the posterior probability distribution of the implicit variable by using a feature reconstruction decoder, comprises the following formula: where E q (z|x) q (c|t,x)logp(x|z; θ) represents the reconstruction of the original input features; p(x|z; θ) = Bernoulli(x; μ θ ); L and W represent the number of Monte Carlo samples and the dimension of the covariate, respectively; and, where, f(x; γ) and f(z (l) ; θ) represent the feature encoder and the reconstruction decoder with parameters γ and θ, respectively.

5. The method of claim 1, wherein, The method comprises: Supposing that (Z, X) and T are conditionally independent of patient category C, C and X are conditionally independent of implicit variable Z, the joint probability distribution is denoted as: wherein, the variational distribution q(z, c|x, t) = q(z|x)q(c|t, x), q(z|x) and q(c|t, x) represent posterior estimation of membership of the implicit variable z and the category c respectively; p(x|z; θ) represents reconstruction of the original features by using a decoder parameterized as θ.

6. A deep probabilistic model based patient clustering and survival risk prediction system, characterized in that, The system comprises: a data acquisition and preprocessing module configured to obtain user health data under right-censoring condition and preprocess the user health data under the right-censoring condition; a data feature representation extraction module configured to extract feature representation of the preprocessed user health data under the right-censoring condition based on a feature extractor of a multi-task learning framework; the feature extractor of the multi-task learning framework is an encoder of a deep probabilistic model; a data feature representation extraction module configured to extract feature representation of the preprocessed user health data under the right-censoring condition based on a feature extractor of a multi-task learning framework; the feature extractor of the multi-task learning framework is an encoder of a deep probabilistic model; The data distribution information reconstruction module is configured to obtain data distribution information based on the feature representation by using the deep probabilistic model, and to reconstruct and optimize the data distribution information; the data distribution information comprises latent variable prior probability distribution information based on category membership prior in the feature representation; The individual risk estimation module is configured to estimate a category-level risk function based on the survival prediction multi-task subnetwork without considering parameter assumptions, and to obtain an individual risk estimation result based on the risk function; The result obtaining module is configured to obtain a clustering prediction result and an event time prediction result based on a variational evidence lower bound of the optimized data distribution information and the individual risk estimation result.

7. The system of claim 6, wherein, The data distribution information reconstruction module is configured to obtain data distribution information based on the feature representation by using the deep probabilistic model, and to reconstruct and optimize the data distribution information; the data distribution information comprises latent variable prior probability distribution information based on category membership prior in the feature representation; S31, based on the feature representation, latent variable prior probability distribution based on category membership prior is obtained by using Gaussian mixture prior, and the posterior probability distribution of the latent variable obtained by the encoder and the reparameterization technique is caused to approximate the latent variable prior probability distribution based on KL divergence; S32, based on the posterior probability distribution of the latent variable, the original feature of the feature representation is reconstructed by using a feature reconstruction decoder.

8. The system of claim 7, wherein, The S31, based on the feature representation, latent variable prior probability distribution based on category membership prior is obtained by using Gaussian mixture prior, and the posterior probability distribution of the latent variable obtained by the encoder and the reparameterization technique is caused to approximate the latent variable prior probability distribution based on KL divergence, which is expressed by a formula as follows: where q(z, c|x, t) = q(z|x)q(c|t, x) represents the posterior distribution of the membership of the latent variable z and the class c; the first term E q(z,c∣x,t) logp(x|z; θ) is the reconstruction term, representing reconstructing the original feature x based on the latent variable z and the feature reconstruction decoder parameterized as θ; the second term encourages the model to utilize the survival prediction multi-task subnetwork parameterized as to perform event time prediction at the group class level the third term D KL (q(z, c|x, t)||p(z, c; μ c ,Σ c ,π)) represents utilizing the KL divergence to make the variational posterior distribution q(z, c|x, t) approach the prior distribution of the latent variable and the class membership p(z, c) = p(z|c; μ c ,Σ c )p(c; π); where p(c; π) = π c , π ∈ R N represents the prior probability distribution of the membership of the class c; the latent variable z ~ p(z|c; {μ1, …, μ N}, {Σ1, …, Σ N}) = N(z; μ c ,Σ c ); μ c and Σ c respectively represent the mean and variance of the Gaussian mixture prior distribution determined by the class c.

9. The system of claim 7, wherein, The S32, based on the posterior probability distribution of the latent variable, the original feature of the feature representation is reconstructed by using a feature reconstruction decoder, which comprises the following formula: where E q (z|x) q (c|t,x)logp(x|z; θ) represents the reconstruction of the original input features; p(x|z; θ) = Bernoulli(x; μ θ ); L and W represent the number of Monte Carlo samples and the dimension of the covariate, respectively; and, where, f(x; γ) and f(z (l) ; θ) represent the feature encoder and the reconstruction decoder with parameters γ and θ, respectively.

10. The system of claim 6, wherein, The result obtaining module is configured to obtain a clustering prediction result and an event time prediction result based on a variational evidence lower bound of the optimized data distribution information and the individual risk estimation result; It is assumed that (Z, X) and T are conditionally independent of the patient category C, and C and X are conditionally independent of the latent variable Z, and the joint probability distribution is denoted as: wherein the variational distribution q(z, c|x, t) = q(z|x)q(c|t, x), q(z|x) and q(c|t, x) represent the posterior estimates of the membership of the latent variable z and the category c respectively; p(x|z; θ) represents the reconstruction of the original feature by using the decoder parameterized as θ.

Images