Prediction method of recurrence probability of specific disease based on clustering method

Abstract

The present invention discloses a method for predicting the recurrence probability of a specific disease based on a clustering method. The method comprises collecting patient disease data from different time periods, preprocessing the patient disease data, dividing the patient disease data according to EBV DNA levels and tumor response to obtain subgroup data, characterizing the subgroup data according to relative risk to obtain time-series disease features, permuting and combining the time-series disease features to obtain disease manifestation data, risk stratifying the disease manifestation data using supervised clustering to obtain risk groups, assessing the prognostic performance of the risk groups using the Concordance Index, comparing the prognostic performance with the TNM staging to obtain recurrence data, constructing a recurrence probability prediction model based on the recurrence data, inputting the data to be predicted into the recurrence probability prediction model, and outputting a prediction result. This method can not only improve the accuracy of predicting the recurrence probability of a specific disease, but also has good interpretability.

Description

Technical Field

[0001] The present invention relates to the field of prediction, and in particular to a method for predicting the recurrence probability of a specific disease based on a clustering method. Background Art

[0002] With the continuous advancement of medical technology and the accumulation of medical data, predicting the probability of recurrence of specific diseases has become a hot topic in current medical research. Traditional prediction methods rely heavily on physician experience and observation of individual patients. The accuracy and reliability of these methods are limited by the physician's subjective judgment and are difficult to rapidly analyze across large amounts of patient data.

[0003] Over the past half century, tremendous efforts have been made to elucidate the heterogeneity of NPC, but previous models did not consider changes in biomarkers during treatment, had relatively poor predictive effects, lacked external validation, and were unable to predict the benefit of AC.

[0004] Clustering, as an unsupervised learning technique, can identify patterns and structures in data without pre-labeling. In disease recurrence prediction, clustering can help identify patient subgroups that may have different recurrence risk profiles. With the accumulation of medical data and advances in computing power, data-driven personalized medicine has become possible. By analyzing multidimensional patient data, including genomic, transcriptomic, proteomic, metabolomic, and clinical data, the probability of disease recurrence can be more accurately predicted. Summary of the Invention

[0005] The object of the present invention is to provide a method for predicting the recurrence probability of a specific disease based on a clustering method.

[0006] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0007] The present invention comprises the following steps:

[0008] Collecting disease data of patients at different periods and preprocessing the disease data of patients;

[0009] Dividing the patient's disease data according to EBV DNA levels and tumor response to obtain subgroup data, and performing feature identification on the subgroup data according to relative risk to obtain temporal disease features;

[0010] Permuting and combining the time-series disease features to obtain disease manifestation data, and using supervised clustering to perform risk stratification on the disease manifestation data to obtain risk groups;

[0011] The prognostic performance of the risk group was evaluated by the Concordia Index, and the prognostic performance was compared with the TNM stage to obtain recurrence data;

[0012] A recurrence probability prediction model is constructed based on the recurrence data, the data to be predicted is input into the recurrence probability prediction model, and a prediction result is output.

[0013] Furthermore, the method for dividing the patient disease data according to EBV DNA level and tumor response includes:

[0014] The patient's tumor response is obtained through medical imaging and clinical examinations, and the EBV DNA level is obtained by extracting DNA from the patient's blood;

[0015] Calculate the disease factor:

[0016]

[0017] The disease factor at the tth moment is The DNA weight is λ1, the tumor weight is λ2, and the tumor response at the tth moment is The EBV DNA level at time t is Standard tumor response was The standard EBV DNA level is The EBV DNA level at time t-1 is The tumor response at time t-1 is

[0018] When the disease factor is greater than zero and less than 0.27, the patient's disease data is divided into the first subgroup data; when the disease factor is greater than 0.27 and less than 0.48, the patient's disease data is divided into the second subgroup data; when the disease factor is greater than 0.48 and less than 0.77, the patient's disease data is divided into the third subgroup data; when the disease factor is greater than 0.77 and less than 1, the patient's disease data is divided into the fourth subgroup data.

[0019] Furthermore, the method of performing feature recognition on the subgroup data according to the relative risk to obtain a temporal disease feature includes:

[0020] Edge detection algorithm was used to extract image features of subgroup data, and sliding window technology was used to segment the image in chronological order to obtain segmented time series data;

[0021] Calculate the center point of a segment of time series data:

[0022]

[0023] The window size is m, and the time series data of the a+wth segment is d a+w , the center point of the w-th segment of time series data is

[0024] Construct a dimension vector to formally represent the center point of the segment time series data. The expression is:

[0025]

[0026] The w-th d value of the i-th window is expressed as The wth segment time series data is d w , the w+1th segment time series data is d w+1 , the time series data of the w+m-1th segment is d w+m-1 ;

[0027] Calculate the similarity between vectors:

[0028]

[0029] The w+sth segment time series data of the i-th window is The w+sth segment time series data of the i+1th window is The center point of the w-th segment of the time series data in the i-th window is The center point of the w-th segment of the time series data in the i+1-th window is The difference between the time series data of the i-th window segment and the time series data of the z-th window segment is h i,i+1 , segment time series data and segment timing data The deviation is η i,i+1 , the similarity capacity parameter is ρ;

[0030] Calculate the change information between vectors:

[0031]

[0032] The time series data and segment timing data The change information degree is ε i,i+1 , the number of windows is p, segment time series data and segment timing data The change in

[0033] Calculate relative risk:

[0034]

[0035] The relative risk between the i-th window and the z-th window is The scaling factor is γ and the heritability coefficient is The initial error value is δ;

[0036] The window with the absolute value of relative risk greater than 0.11 is regarded as the disease window, and the segment time series data with the difference distance greater than 0.27 within the disease window is output as the disease feature.

[0037] Furthermore, the method of permuting and combining the disease characteristics to obtain disease manifestation data includes:

[0038] Calculate the association of disease features:

[0039]

[0040] The correlation coefficient between disease feature j and disease feature f is θ jf , the probability of disease feature j and disease feature f appearing at the same time is c(j,f), and the probability of disease feature j appearing is c j , the probability of the characteristic disease f appearing is c f , the first weight is β1, the second weight is β2;

[0041] Disease features are divided into associated feature sets according to the degree of association, and the associated feature sets are randomly combined and output as disease manifestation data.

[0042] Furthermore, a method for using supervised clustering to perform risk stratification on the disease manifestation data to obtain risk groups includes:

[0043] The disease manifestation data is input into a clustering model based on a graph autoencoder, which includes an encoder network, a graph encoding network, and an adaptive fusion network;

[0044] K-means clustering was used to perform initial clustering on the disease manifestation data to obtain the initial cluster centers;

[0045] Given the transfer function of the encoder network and the transfer function of the graph encoding network, the expressions are:

[0046]

[0047] The bias vector of the κth layer of the encoder network is φ κ , the transfer function of the κ-1 layer of the encoder network is G κ-1 , the weight matrix of the κth layer of the encoder network is F κ , the activation function is The normalized adjacency matrix is ​​Q * , the weight of the κ-1 layer of the graph encoding network is The transfer function of the κth layer of the graph coding network is S κ , the output of the κ-1 layer of the graph encoding network is S κ-1 , the transfer function of the κth layer of the encoder network is G κ ;

[0048] Calculate the propagation function of the first layer network:

[0049]

[0050] The learning rate is χ, and the weight matrix of the κth layer of the adaptive fusion network is The bias vector of the κth layer of the adaptive fusion network is The propagation function of the κth layer of the adaptive fusion network is W κ ;

[0051] Let κ = κ + 1, recalculate the transfer function of the encoder network, the transfer function of the graph encoding network, and the transfer function of the adaptive fusion network:

[0052]

[0053] The transfer function of the κ-1th layer of the graph coding network is W κ-1 , calculate the loss function of the encoder network:

[0054]

[0055] The number of disease manifestation data is P, the output norm function is ‖·‖, the disease manifestation data is Q, and the reconstructed disease manifestation data is V * , the loss function of the computational graph encoding network is:

[0056]

[0057] Calculate the encoder divergence:

[0058]

[0059] The encoder network disease manifestation data k g and the initial cluster center u h The probability distribution of The hth initial cluster center is u h , encoder network disease manifestation data k g and the initial cluster center u h The high confidence distribution of The g-th disease manifestation data is k g , the degrees of freedom are The adjustment factor is σ and the divergence of the encoder network is L gh ;

[0060] Compute the divergence of the graph encoding network:

[0061]

[0062] The heritability coefficient is ζ2, the deviation factor is λ, and the graph encoding network disease expression data kg and the initial cluster center u h The probability distribution of The divergence of the graph coding network is L gh ;

[0063] Calculate the loss function of the clustering model:

[0064]

[0065] The graph weight is α, the divergence weight is μ, and the cluster center is The cluster data is y, and the disease manifestation data is k g and the initial cluster center u h The probability distribution is r gh ;

[0066] Update the weight matrix, bias vector, and learning rate through backpropagation until the maximum number of iterations is reached, otherwise update the transfer function;

[0067] The clustering results are obtained through collaborative training and output as risk groups.

[0068] Furthermore, the method for evaluating the prognostic performance of the risk group by the Concord Index comprises:

[0069] Machine learning algorithms are used to predict the survival probability and time of risk groups and calculate the concordance index of risk groups:

[0070]

[0071] The harmony index is The indicator function is K(·), and the survival time of patient a is T a , the survival time of patient m is T m , the number of comparison pairs is n, and the survival probability of patient a is The survival probability of patient m is

[0072] Patients who did not receive medical treatment were used as a control group, the risk group was compared with the control group, and the improved performance of the risk group was output as the prognostic performance.

[0073] Furthermore, the method of comparing the prognostic performance with the TNM staging to obtain recurrence data includes:

[0074] Obtain TNM stage based on patient disease data, and input prognostic performance and TNM stage into the nomination graph model;

[0075] The product limit method was used to draw the survival curves of TNM staging and nomination diagram models, and the prognostic performance was screened according to the survival curves to obtain the recurrence performance;

[0076] Compare the recurrence performance of patients at different times to obtain performance differences, calculate the correlation of performance differences, eliminate performance differences with a correlation less than 0.402, and obtain the optimal performance differences;

[0077] The preferred performance differences are output as recurrence data.

[0078] Furthermore, a method for constructing a recurrence probability prediction model based on the recurrence data includes:

[0079] Recurrence probability prediction models include random forest algorithm, time series analysis algorithm, and long short-term memory network algorithm;

[0080] The random forest algorithm randomly divided the recurrence data into a training group and a validation group according to a 5:2 ratio;

[0081] The time series analysis algorithm obtains the temporal variation pattern through the temporal variation trend of the training group; the long short-term memory network algorithm predicts the recurrence probability through the temporal variation pattern; the validation group is predicted to obtain the predicted recurrence probability, and the parameters of the long short-term memory network algorithm are adjusted according to the error between the predicted recurrence probability and the actual recurrence probability.

[0082] The beneficial effects of the present invention are:

[0083] The present invention is a method for predicting the recurrence probability of a specific disease based on a clustering method. Compared with the prior art, the present invention has the following technical effects:

[0084] The present invention can improve the accuracy of the prediction of the recurrence probability of a specific disease of the clustering method through preprocessing, data partitioning, feature identification, permutation and combination, risk stratification, acquisition of prognostic performance, acquisition of recurrence data and model construction steps, thereby improving the accuracy of the prediction of the recurrence probability of a specific disease of the clustering method, and optimizing the prediction of the recurrence probability of a specific disease of the clustering method, which can greatly save resources and improve work efficiency. It can realize the automatic prediction of the recurrence probability of a specific disease of the clustering method, and perform risk stratification on the prediction of the recurrence probability of a specific disease of the clustering method in real time. It is of great significance to the prediction of the recurrence probability of a specific disease of the clustering method, and can adapt to the prediction of the recurrence probability of a specific disease of clustering methods with different standards and the prediction needs of the recurrence probability of a specific disease of different clustering methods, and has a certain universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 Flow chart of the steps of the method for predicting the recurrence probability of a specific disease based on the clustering method of the present invention;

[0086] Figure 2 This is a schematic diagram of the structure of an electronic device in an embodiment of this specification. DETAILED DESCRIPTION

[0087] The present invention will be further described below through specific examples. The illustrative examples and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0088] The method for predicting the recurrence probability of a specific disease based on a clustering method of the present invention comprises the following steps:

[0089] like Figure 1 As shown, in this embodiment, the following steps are included:

[0090] Collecting disease data of patients at different periods and preprocessing the disease data of patients;

[0091] In the actual evaluation, disease data of NPC patients were collected before, during and after IC treatment;

[0092] Dividing the patient's disease data according to EBV DNA levels and tumor response to obtain subgroup data, and performing feature identification on the subgroup data according to relative risk to obtain temporal disease features;

[0093] In the actual evaluation, the first subgroup data were 2 weeks before the start of IC treatment, the second subgroup data were during IC treatment and within 2 weeks after the completion of IC treatment, the third subgroup data were 3-5 weeks during radiation therapy, and the fourth subgroup data were during radiation therapy and 4-12 weeks after the completion of radiation therapy;

[0094] The temporal disease characteristics were as follows: among all enrolled patients, 52.6% had stage III tumors and 47.4% had stage IV tumors. After IC treatment, more than 70% of patients received CCRT, and less than 30% of patients received IMRT but no CCRT. The 5-year PFS, OS, DMFS, and LRRFS rates in the training group were 68.4%, 78.9%, 73.3%, and 72.6%, respectively. The PFS, OS, DMFS, and LRRFS rates in the external cohort were 70%, 82.6%, 76.1%, and 77.5%, respectively.

[0095] Permuting and combining the time-series disease features to obtain disease manifestation data, and using supervised clustering to perform risk stratification on the disease manifestation data to obtain risk groups;

[0096] In the actual evaluation, there are 7 groups of disease manifestation data;

[0097] Risk stratification was as follows: among 984 patients, cfEBV DNA was undetectable in 79.7% after IC, detected in 87.5% during RT, and detected in 91% during the late RT period. In terms of tumor regression trajectory, 84.3% of patients achieved CR / PR after IC, 87.4% of patients continued to achieve CR / PR during the mid-RT period, and 93.2% of patients achieved CR / PR after RT, of which 71.8% achieved CR.

[0098] The prognostic performance of the risk group was evaluated by the Concordia Index, and the prognostic performance was compared with the TNM stage to obtain recurrence data;

[0099] In actual evaluation, the prognostic performance was a Concordance Index of 0.834 and a 95% IC; the recurrence data were: compared with the TNM stage combined with pre-treatment EBV DNA and the nomination graph model, the C index of the response group for predicting 5-year PFS was 0.825 vs, 0.603 vs, 0.626 vs, and 0.756 vs, respectively;

[0100] Constructing a recurrence probability prediction model based on the recurrence data, inputting the data to be predicted into the recurrence probability prediction model, and outputting a prediction result;

[0101] In practical evaluation, T4 disease, advanced N stage, primary tumor and metastatic lymph node volume, and higher pre-treatment cfEBV DNA load were associated with worse treatment efficacy.

[0102] In this embodiment, the method for dividing the patient disease data according to EBV DNA level and tumor response includes:

[0103] The patient's tumor response is obtained through medical imaging and clinical examinations, and the EBV DNA level is obtained by extracting DNA from the patient's blood;

[0104] Calculate the disease factor:

[0105]

[0106] The disease factor at the tth moment is The DNA weight is λ1, the tumor weight is λ2, and the tumor response at the tth moment is The EBV DNA level at time t is Standard tumor response was The standard EBV DNA level is The EBV DNA level at time t-1 is The tumor response at time t-1 is

[0107] When the disease factor is greater than zero and less than 0.27, the patient's disease data is divided into the first subgroup data; when the disease factor is greater than 0.27 and less than 0.48, the patient's disease data is divided into the second subgroup data; when the disease factor is greater than 0.48 and less than 0.77, the patient's disease data is divided into the third subgroup data; when the disease factor is greater than 0.77 and less than 1, the patient's disease data is divided into the fourth subgroup data.

[0108] In this embodiment, the method for performing feature recognition on the subgroup data according to relative risk to obtain time-series disease features includes:

[0109] Edge detection algorithm was used to extract image features of subgroup data, and sliding window technology was used to segment the image in chronological order to obtain segmented time series data;

[0110] Calculate the center point of a segment of time series data:

[0111]

[0112] The window size is m, and the time series data of the a+wth segment is d a+w , the center point of the w-th segment of time series data is

[0113] Construct a dimension vector to formally represent the center point of the segment time series data. The expression is:

[0114]

[0115] The w-th d value of the i-th window is expressed as The wth segment time series data is d w , the w+1th segment time series data is d w+1 , the time series data of the w+m-1th segment is d w+m-1 ;

[0116] Calculate the similarity between vectors:

[0117]

[0118] The w+sth segment time series data of the i-th window is The w+sth segment time series data of the i+1th window is The center point of the w-th segment of the time series data in the i-th window is The center point of the w-th segment of the time series data in the i+1-th window is The difference between the time series data of the i-th window segment and the time series data of the z-th window segment is h i,i+1 , segment time series data and segment timing data The deviation is η i,i+1, the similarity capacity parameter is ρ;

[0119] Calculate the change information between vectors:

[0120]

[0121] The time series data and segment timing data The change information degree is ε i,i+1 , the number of windows is p, segment time series data and segment timing data The change in

[0122] Calculate relative risk:

[0123]

[0124] The relative risk between the i-th window and the z-th window is The scaling factor is γ and the heritability coefficient is The initial error value is δ;

[0125] The window with the absolute value of relative risk greater than 0.11 is regarded as the disease window, and the segment time series data with the difference distance greater than 0.27 within the disease window is output as the disease feature.

[0126] In this embodiment, the method of obtaining disease manifestation data by permuting and combining the disease characteristics includes:

[0127] Calculate the association of disease features:

[0128]

[0129] The correlation coefficient between disease feature j and disease feature f is θ jf , the probability of disease feature j and disease feature f appearing at the same time is c(j,f), and the probability of disease feature j appearing is c j , the probability of the characteristic disease f appearing is c f , the first weight is β1, the second weight is β2;

[0130] Disease features are divided into associated feature sets according to the degree of association, and the associated feature sets are randomly combined and output as disease manifestation data.

[0131] In this embodiment, the method of using supervised clustering to perform risk stratification on the disease manifestation data to obtain risk groups includes:

[0132] The disease manifestation data is input into a clustering model based on a graph autoencoder, which includes an encoder network, a graph encoding network, and an adaptive fusion network;

[0133] K-means clustering was used to perform initial clustering on the disease manifestation data to obtain the initial cluster centers;

[0134] Given the transfer function of the encoder network and the transfer function of the graph encoding network, the expressions are:

[0135]

[0136] The bias vector of the κth layer of the encoder network is φ κ , the transfer function of the κ-1 layer of the encoder network is G κ-1 , the weight matrix of the κth layer of the encoder network is F κ , the activation function is The normalized adjacency matrix is ​​Q * , the weight of the κ-1 layer of the graph encoding network is The transfer function of the κth layer of the graph coding network is S κ , the output of the κ-1 layer of the graph encoding network is S κ-1 , the transfer function of the κth layer of the encoder network is G κ ;

[0137] Calculate the propagation function of the first layer network:

[0138]

[0139] The learning rate is χ, and the weight matrix of the κth layer of the adaptive fusion network is The bias vector of the κth layer of the adaptive fusion network is The propagation function of the κth layer of the adaptive fusion network is W κ ;

[0140] Let κ = κ + 1, recalculate the transfer function of the encoder network, the transfer function of the graph encoding network, and the transfer function of the adaptive fusion network:

[0141]

[0142] The transfer function of the κ-1th layer of the graph coding network is W κ-1 , calculate the loss function of the encoder network:

[0143]

[0144] The number of disease manifestation data is P, the output norm function is ‖·‖, the disease manifestation data is Q, and the reconstructed disease manifestation data is V * , the loss function of the computational graph encoding network is:

[0145]

[0146] Calculate the encoder divergence:

[0147]

[0148] The encoder network disease manifestation data k g and the initial cluster center u h The probability distribution of The hth initial cluster center is u h , encoder network disease manifestation data k g and the initial cluster center u h The high confidence distribution of The g-th disease manifestation data is k g , the degrees of freedom are The adjustment factor is σ and the divergence of the encoder network is L gh ;

[0149] Compute the divergence of the graph encoding network:

[0150]

[0151]

[0152] The heritability coefficient is ζ2, the deviation factor is λ, and the graph encoding network disease expression data k g and the initial cluster center u h The probability distribution of The divergence of the graph coding network is L gh ;

[0153] Calculate the loss function of the clustering model:

[0154]

[0155] The graph weight is α, the divergence weight is μ, and the cluster center is The cluster data is y, and the disease manifestation data is k g and the initial cluster center u h The probability distribution is r gh ;

[0156] Update the weight matrix, bias vector, and learning rate through backpropagation until the maximum number of iterations is reached, otherwise update the transfer function;

[0157] The clustering results are obtained through collaborative training and output as risk groups.

[0158] In this embodiment, the method for evaluating the prognostic performance of the risk group by using the Concord Index comprises:

[0159] Machine learning algorithms are used to predict the survival probability and time of risk groups and calculate the concordance index of risk groups:

[0160]

[0161] The harmony index is The indicator function is K(·), and the survival time of patient a is T a , the survival time of patient m is T m , the number of comparison pairs is n, and the survival probability of patient a is The survival probability of patient m is

[0162] Patients who did not receive medical treatment were used as a control group, the risk group was compared with the control group, and the improved performance of the risk group was output as the prognostic performance.

[0163] In this embodiment, the method of comparing the prognostic performance with the TNM staging to obtain recurrence data includes:

[0164] Obtain TNM stage based on patient disease data, and input prognostic performance and TNM stage into the nomination graph model;

[0165] The product limit method was used to draw the survival curves of TNM staging and nomination diagram models, and the prognostic performance was screened according to the survival curves to obtain the recurrence performance;

[0166] Compare the recurrence performance of patients at different times to obtain performance differences, calculate the correlation of performance differences, eliminate performance differences with a correlation less than 0.402, and obtain the optimal performance differences;

[0167] The preferred performance differences are output as recurrence data.

[0168] In this embodiment, the method for constructing a recurrence probability prediction model based on the recurrence data includes:

[0169] Recurrence probability prediction models include random forest algorithm, time series analysis algorithm, and long short-term memory network algorithm;

[0170] The random forest algorithm randomly divided the recurrence data into a training group and a validation group according to a 5:2 ratio;

[0171] The time series analysis algorithm obtains the temporal variation pattern through the trend of the training group over time;

[0172] The long short-term memory network algorithm predicts the recurrence probability through the time series change law; the validation group is predicted to obtain the predicted recurrence probability, and the parameters of the long short-term memory network algorithm are adjusted according to the error between the predicted recurrence probability and the actual recurrence probability.

[0173] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the recurrence probability of a specific disease based on a clustering method, characterized in that: The following steps are involved: Collecting disease data of patients at different periods and preprocessing the disease data of patients; The patient's disease data is divided according to EBV DNA levels and tumor response to obtain subgroup data, and the subgroup data is characterized according to relative risks to obtain temporal disease characteristics; including: Edge detection algorithm was used to extract image features of subgroup data, and sliding window technology was used to segment the image in chronological order to obtain segmented time series data; Calculate the center point of a segment of time series data: The window size is , the time series data of the a+wth segment is , the center point of the w-th segment of time series data is ; Construct a dimension vector to formally represent the center point of the segment time series data. The expression is: The wth one in the i-th window The value is expressed as , the wth segment time series data is , the w+1th segment time series data is , No. The time series data for each segment is ; Calculate the similarity between vectors: The w+sth segment time series data of the i-th window is , No. The w+sth segment time series data of the window is , the center point of the w-th segment of the time series data in the i-th window is , No. The center point of the w-th segment of the time series data in the window is , the difference distance between the time series data of the i-th window segment and the time series data of the z-th window segment is , segment time series data and segment timing data The deviation is , similarity capacity parameter is ; Calculate the change information between vectors: The time series data and segment timing data The change information degree is , the number of windows is p, segment time series data and segment timing data The change in ; Calculate relative risk: The relative risk between the i-th window and the z-th window is , the scale factor is , the heritability coefficient is The initial error value is ; The window with the relative risk absolute value greater than 0.11 is regarded as the disease window, and the segment time series data with the difference distance greater than 0.27 within the disease window is output as the disease feature; The time series disease features are permuted and combined to obtain disease manifestation data, and supervised clustering is used to perform risk stratification on the disease manifestation data to obtain risk groups; including: The disease manifestation data is input into a clustering model based on a graph autoencoder, which includes an encoder network, a graph encoding network, and an adaptive fusion network; K-means clustering was used to perform initial clustering on the disease manifestation data to obtain the initial cluster centers; Given the transfer function of the encoder network and the transfer function of the graph encoding network, the expressions are: The encoder network The bias vector of the layer is , encoder network The transfer function of layer -1 is , encoder network The weight matrix of the layer is , the activation function is , the normalized adjacency matrix is , Graph Coding Network The weight of layer -1 is , Graph Coding Network The transfer function of the layer is , Graph Coding Network The output of layer -1 is , encoder network The transfer function of the layer is ; Calculate the propagation function of the first layer network: The learning rate is , Adaptive Fusion Network The weight matrix of the layer is , Adaptive Fusion Network The bias vector of the layer is , Adaptive Fusion Network The propagation function of the layer is ; make 1. Recalculate the transfer function of the encoder network, the transfer function of the graph encoding network, and the transfer function of the adaptive fusion network: The graph coding network The transfer function of layer -1 is , calculate the loss function of the encoder network: The number of disease manifestation data is P, and the output norm function is , the disease manifestation data is Q, and the reconstructed disease manifestation data is , the loss function of the computational graph encoding network is: Calculate the encoder divergence: The encoder network disease manifestation data and initial cluster centers The probability distribution of , the hth initial cluster center is , Encoder Network Disease Presentation Data and initial cluster centers The high confidence distribution of , the g-th disease manifestation data is , the degrees of freedom are , the adjustment factor is , the divergence of the encoder network is ; Compute the divergence of the graph encoding network: The heritability coefficient is , the deviation factor is , Graph Encoding Network Disease Manifestation Data and initial cluster centers The probability distribution of , the divergence of the graph coding network is ; Calculate the loss function of the clustering model: The graph weight is , the divergence weight is , the cluster center is , the clustering data is , disease manifestation data and initial cluster centers The probability distribution of ; Update the weight matrix, bias vector, and learning rate through backpropagation until the maximum number of iterations is reached, otherwise update the transfer function; Obtain clustering results through collaborative training and output the clustering results as risk groups; The prognostic performance of the risk group was assessed by the Concordia Index, and the prognostic performance was compared with the TNM stage to obtain recurrence data; including: Machine learning algorithms are used to predict the survival probability and time of risk groups and calculate the concordance index of risk groups: The harmony index is , the indicator function is ,patient The survival time is ,patient The survival time is , the number of comparison pairs is ,patient The survival probability is ,patient The survival probability is ; Patients who did not receive medical treatment were used as a control group, the risk group was compared with the control group, and the improved performance of the risk group was output as the prognostic performance; Obtain TNM stage based on patient disease data, and input prognostic performance and TNM stage into the nomination graph model; The product limit method was used to draw the survival curves of TNM staging and nomination diagram models, and the prognostic performance was screened according to the survival curves to obtain the recurrence performance; Compare the recurrence performance of patients at different times to obtain performance differences, calculate the correlation of performance differences, eliminate performance differences with a correlation less than 0.402, and obtain key performance differences; Output key performance differences as recurrence data; Constructing a recurrence probability prediction model based on the recurrence data, inputting the data to be predicted into the recurrence probability prediction model, and outputting a prediction result; comprising: Recurrence probability prediction models include random forest algorithm, time series analysis algorithm, and long short-term memory network algorithm; The random forest algorithm randomly divided the recurrence data into a training group and a validation group according to a 5:2 ratio; The time series analysis algorithm obtains the temporal variation pattern through the trend of the training group over time; The long short-term memory network algorithm predicts the recurrence probability based on the temporal variation pattern; the prediction of the validation group is performed to obtain the predicted recurrence probability, and the parameters of the long short-term memory network algorithm are adjusted according to the error between the predicted recurrence probability and the actual recurrence probability; The method of segmenting the patient disease data according to EBV DNA level and tumor response comprises: The patient's tumor response is obtained through medical imaging and clinical examinations, and the EBV DNA level is obtained by extracting DNA from the patient's blood; Calculate the disease factor: The disease factor at the tth moment is , the DNA weight is , the tumor weight is , the tumor response at the tth moment is , the EBV DNA level at the tth moment is , standard tumor response was , the standard EBV DNA level is , the EBV DNA level at time t-1 is , the tumor response at time t-1 is ; When the disease factor is greater than zero and less than 0.27, the patient's disease data is divided into the first subgroup data; when the disease factor is greater than 0.27 and less than 0.48, the patient's disease data is divided into the second subgroup data; when the disease factor is greater than 0.48 and less than 0.77, the patient's disease data is divided into the third subgroup data; when the disease factor is greater than 0.77 and less than 1, the patient's disease data is divided into the fourth subgroup data; The method of obtaining disease manifestation data by permuting and combining the disease characteristics comprises: Calculate the association of disease features: The correlation coefficient between disease feature j and disease feature f is , the probability of disease feature j and disease feature f appearing at the same time is , the probability of disease feature j appearing is , the probability of the characteristic disease f appearing is , the first weight is , the second weight is ; Disease features are divided into associated feature sets according to the degree of association, and the associated feature sets are randomly combined and output as disease manifestation data.

Images