CN115565669B - Cancer survival analysis method based on GAN and multitask learning - Google Patents
Cancer survival analysis method based on GAN and multitask learning Download PDFInfo
- Publication number
- CN115565669B CN115565669B CN202211240631.1A CN202211240631A CN115565669B CN 115565669 B CN115565669 B CN 115565669B CN 202211240631 A CN202211240631 A CN 202211240631A CN 115565669 B CN115565669 B CN 115565669B
- Authority
- CN
- China
- Prior art keywords
- patient
- training
- survival
- cancer
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
- G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention belongs to the technical field of medical information, and particularly relates to a cancer survival analysis method based on GAN and multitask learning, which uses a GAN network for data enhancement, inputs the characteristics, survival time and the type of the ending cancer patient into the GAN network for training, and generates a large amount of non-deleted survival data; constructing a multi-task learning cancer survival analysis model based on soft parameter sharing, wherein a plurality of different tasks respectively predict the probability of different outcomes of a patient at each moment in a future period of time; the characteristics of the cancer patient to be analyzed are input into a constructed survival analysis model, and the probability of different future outcomes is output.
Description
Technical Field
The invention belongs to the technical field of medical information, and particularly relates to a cancer survival analysis method based on GAN and multitask learning.
Background
Accurate prognosis prediction for cancer patients is beneficial to doctors optimizing therapeutic measures, improving patient prognosis and reducing patient disease burden. Medically, prognosis generally refers to the use of a patient's characteristics to predict the probability of its occurrence of an outcome over a period of time. The outcome refers to death, recurrence or exacerbation. Survival assays are the assay methods often used in prognosis prediction of cancer. One key to survival analysis is the presence of deleted data, which indicates that the patient has not had a resultant event during the study. The survival analysis model does not directly predict the survival time of the patient, but predicts the probability distribution of the survival time of the patient.
Traditionally, cox Proportional Hazards (CPH) are often used for cancer survival analysis studies. CPH has two hypotheses: 1) Proportional risk assumption: the risk ratio between different patients is a constant value and does not change over time. 2) The log-linear assumption is: the characteristics of a patient are linearly related to the logarithm of the patient risk. However, it is difficult for real survival data to meet the linear scale risk condition. With the continuous development of deep learning in recent years, more and more scholars apply structures such as a fully connected neural network, a convolutional neural network, a cyclic neural network and a graph neural network to cancer survival analysis and research. In addition, some scholars apply methods such as semi-supervised, self-supervised, active learning, and multi-task learning to the field of cancer survival analysis.
Currently, existing methods for cancer survival analysis suffer from the following deficiencies. First: patients often experience deletion in cancer survival analysis and research, but the existing survival analysis methods cannot handle the situation of high deletion. Second,: cancer survival analysis methods using multitasking learning are all based on hard parameter sharing, which is mainly suited to the scenario where the links between tasks are tight. However, in cancer survival analysis, the variability between different tasks is great, and tasks may even be conflicting from task to task. Third,: existing survival analysis models are accurate for predicting short-term outcome occurrence of cancer patients, but have yet to be improved in their ability to predict long-term outcome occurrence.
Disclosure of Invention
In order to solve the technical problems in the prior art, the invention provides a cancer survival analysis method based on GAN and multitask learning, which aims to solve the problem that the existing survival analysis method cannot handle high deletion.
The technical scheme adopted by the invention is as follows:
a cancer survival analysis method based on GAN and multitasking learning, comprising the steps of:
step 1: acquiring survival data of a cancer patient, forming a survival data set of the cancer patient, and taking part of the survival data in the survival data set as a training set;
step 2: carrying out data enhancement on Survival data in a training set based on a trained survivinval-GAN model;
step 3: constructing a cancer survival analysis model based on multitask learning, and training the cancer survival analysis model based on the enhanced training set data; searching out the optimal super parameters of the cancer survival analysis model by using a grid search method and matching with five-fold cross validation, and retraining the cancer analysis model by using the optimal super parameters;
step 4: inputting the characteristics of the cancer patient to be analyzed into the constructed cancer survival analysis model to obtain the probability of different fates of the cancer patient at each moment in a future period of time.
The invention enhances the data based on the survivin-GAN model, so that a large amount of non-deleted Survival data can be generated, thereby expanding the sample size and enhancing the accuracy and the robustness of model prediction.
Preferably, the survival data of the cancer patient includes patient characteristics, time of observation, and outcome type at last follow-up time; if no outcome appears in the last follow-up time of the patient, the observation time is the deletion time of the patient; if the last follow-up time of the patient comes to a final result, the observed time is the survival time of the patient.
Preferably, the step 2 includes the steps of:
step 2.1: dividing the survival data of the cancer patients in the training set into two groups of deletion and ending occurrence according to whether the obtained survival data is ending or not, and respectively recording the numbers of the two groups;
step 2.2: training a survivinal-GAN model based on the Survival data with the outcomes;
step 2.3: searching out the optimal super parameters of the survivin-GAN model by using a grid search method and matching with five-fold cross validation, and retraining the survivin-GAN model by using the optimal super parameters;
step 2.4: randomly selecting K real survival times from the training set samples and respectively pairing the K real survival times with K different outcomes; sequentially inputting the K pairing results into a survivinval-GAN model to generate K Survival data with ending;
step 2.5: n2 self-increases by K, i.e., n2=n2+k, N2 represents the number of survival data with the occurrence of a ending;
because K Survival data are generated after each round of survivinal-GAN model training, the Survival data after one round of training is equal to K plus the Survival data when inputting; i.e., n2=n2+k.
Step 2.6: judging whether N2 is smaller than N1, if not, directly ending; if yes, returning to the step 2.4 to continue execution until N2 is more than N1; where N1 represents the number of deleted data.
Preferably, the survivinal-GAN model includes a generator and a arbiter;
the generator comprises a full-connection network, wherein the number of layers of the full-connection layer of the full-connection network and the number of neurons of each layer are super parameters;
the discriminator is a multi-task full-connection network, and the number of layers of full-connection layers of the discriminator and the number of neurons of each layer are super parameters;
the arbiter comprises three tasks, the first task being for determining whether the input patient feature is true or the arbiter generated; the second task predicts a outcome type based on the survival data; the third task predicts a time to live based on the survival data.
Preferably, the training steps of the survivinal-GAN model are as follows:
setting super parameters of a generator: the dimension of the Embedding output, the dimension of random noise, the number of layers of the full-connection layer, the number of neurons of each layer, the learning rate and the optimizer;
setting super parameters of a discriminator: the number of layers of the full-connection layer, the number of neurons of each layer, the learning rate and the optimizer;
setting other super parameters: the training round number and the batch_size are the number of training samples grabbed by one training;
and (3) data splicing: randomly acquiring m noise data from standard normal distribution, encoding the input m tags of real living data by an encoding layer, and then splicing the tags with the noise data to obtain data C i ；
Calculating the total loss of the generator:
L G ＝L G1 +L G2 +L G3 ；
wherein: l (L) G Representing the total loss of the generator, L G1 、L G2 And L G3 All represent loss functions;
updating training parameters of a generator: updating training parameters of the generator based on the total loss function of the generator and a preset learning rate;
calculating the total loss of the discriminator:
L D ＝L D1 +L D2 +L D3 ；
wherein: l (L) D Representing the total loss of the discriminator, L D1 、L D2 And L D3 All represent loss functions;
updating training parameters of the discriminator: updating training parameters of the discriminator based on a total loss function of the discriminator and a preset learning rate;
end of training of generator and arbiter: judging whether the training round number reaches the appointed times, if so, finishing the training of the generator and the discriminator, and if not, continuing to execute the training of the discriminator and the generator until the appointed training times are met.
Preferably, the loss function L G1 For bringing the features generated by the generator and the real features closer together, expressed as:
wherein: MES is a mean square loss function expressed as
Representing the mean square loss of q and p of the input; MSE (G (C) i ),x i ) For the generated patient characteristics G (C i ) And true patient characteristics x i Mean square error of (a); MSE (D (G (C) i ))[1]1) the output D (G (C) i ))[1]Mean square error with 1;
the loss function L G2 The outcome for reconciling the patient characteristic predictions generated by the generator with the outcome of the input is expressed as:
wherein: cross Entropy cross entropy loss function is expressed as：
Where h is the predicted probability of K outcomes and class is the true outcome; cross Entropy (D (G (C) i ))[2]，e i ) The output D (G (C i ))[2]And true ending e i Cross entropy of (2);
the loss function L G3 The time-to-live for the patient feature predictions generated by the generator and the entered time-to-live are expressed as:
wherein: MSE (D (G (C) i ))[2]，s i ) The output D (G (C i ))[3]And true survival time s i Is a mean square error of (c).
Preferably, the loss function L D1 The patient characteristics used to enable the arbiter to identify the input are real or false, expressed as:
wherein MSE (D (x i )[1]1) for inputting the real patient characteristics x i When the first task of the discriminator outputs a mean square error of 1; MSE (D (G (C) i ))[1]0) patient characteristics G (C) generated for the input generator i ) When the first task output of the discriminator has a mean square error of 0;
the loss function L D2 For enabling the arbiter to accurately predict the outcome type of the patient, expressed as:
wherein CrossEntropy (D (x) i )[2]，e i ) To input real patient characteristics x i When the output of the second task of the discriminator is equal to e i Cross entropy loss of (2); cross Entropy (D (G (C) i ))[2]，e i ) Patient characteristics G (C) i ) When the output of the second task of the discriminator is equal to e i Cross entropy loss of (2);
the loss function L D3 For enabling the arbiter to accurately predict the survival time of the patient, expressed as:
where MSE (D (x i )[3]，s i ) To input real patient characteristics x i When the output of the third task and s of the discriminator i Mean square error of (a); MSE (D (G (C) i ))[3]，s i ) Patient characteristics G (C) i ) When the output of the third task and s of the discriminator i Is a mean square error of (c).
Preferably, the cancer survival analysis model includes an expert network, a task network, an attention network, and an auxiliary task network.
Preferably, the training steps of the cancer survival analysis model are as follows:
A. setting super parameters: setting the number of full-connection layers of a task network, an auxiliary task network, an expert network and an attention network, the number of neurons at each layer, the learning rate, an optimizer, the number of training rounds, the batch_size, the number of prediction moments and the weight of 4 loss functions;
B. presetting the value of batch_size as m, wherein the total number of the ending types of the patients is K; in the training process of each batch, the survival data of m patients are input into a cancer survival analysis model for training;
C. calculating the loss of the cancer survival analysis model:
total loss function L of cancer survival analysis model s Expressed as:
L s ＝λ 1 ·L s1 +λ 2 ·L s2 +λ 3 ·L s3 +λ 4 ·L s4 ；
wherein: lambda (lambda) 1 ，λ 2 ，λ 3 ，λ 4 Weights of 4 loss functions respectively are super parameters; l (L) s1 、L s2 、L s3 And L s4 All represent loss functions;
D. based on the total loss function L S Updating parameters theta of cancer survival analysis model by preset optimizer Adam and learning rate gamma S ：
θ s ＝Adam(L s ，θ s ，γ)；
E. And C, judging whether the training round number of the cancer survival analysis model accords with the designated times, if not, returning to the step B, and storing the cancer survival analysis model until the training round number accords with the designated times.
Preferably, the loss function L s1 Expressed as:
wherein:
characterizing a patient as x i Under the condition of (a) at time s i Occurrence e i Probability of ending P(s) i ，e i |x i )；
Is an indication function, and the condition is 1 if the indication function meets the condition, otherwise, the indication function is 0; f (F) j (s i |x i ) The expression of (2) is: f (F) j (s i |x i )＝P(s≤s i ，e i ＝j|x＝x i ) Represented by x in the patient characteristic i Under the condition of (1), patient outcome is j and occurs at time s i Previous probabilities;
loss ofFunction L s2 Expressed as:
wherein: a is that j，i，p The expression of (2) is:
the indicator function is to find patient pairs (i, p) that can be risk compared; the expression of the η function is: />
Loss function L s3 Expressed as:
wherein: the expression of the Sigmoid function is:
predicting the probability of the ith patient to develop an outcome j at time t for the model; />
The probability of ending j at time t for the ith patient in practice;
loss function L s4 Expressed as:
wherein:
the outcome type of non-deleted patients for assisting task network prediction +.>
And true ending type e i Cross entropy loss of (c).
Preferably, part of survival data in the survival data set is divided into test sets, the test sets are used for evaluating the performance of the trained cancer survival analysis model, and the evaluation index is C-index;
the specific calculation steps of the C-index are as follows:
a. pairing all patients pairwise;
b. if there is a situation in the pairing that the observation time of the patient A is smaller than that of the patient B and the patient A does not have a ending, the pairing is excluded; if there is a situation in which neither patient in the pairing has a outcome, the pairing is excluded; finally, useful pairing is obtained;
c. calculating the pairing number with the consistent predicted result and actual result in the useful pairing;
d. calculating a pairing value:
c-index = consistent log/useful log.
The beneficial effects of the invention include:
1. the invention enhances the data based on the survivin-GAN model, so that a large amount of non-deleted Survival data can be generated, thereby expanding the sample size and enhancing the accuracy and the robustness of model prediction.
2. Compared with the existing multi-task cancer survival analysis model based on hard parameter sharing, the multi-task cancer survival analysis model based on soft parameter sharing can better benefit the situation that the connection among a plurality of tasks for survival analysis is not tight, so that the prediction accuracy is higher.
3. An auxiliary task for distinguishing different outcomes is added on the basis of the original task, so that the accuracy of the cancer survival analysis model is improved.
4. The invention designs a loss function for calculating the difference between the predicted ending occurrence probability and the true ending occurrence probability, thereby improving the accuracy of model prediction.
Drawings
FIG. 1 is a schematic overall flow chart of the present invention.
FIG. 2 is a network structure diagram of the survivinal-GAN.
Fig. 3 is a network structure diagram of a cancer survival analysis model based on multitasking learning.
Detailed Description
For the purposes of making the objects, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is apparent that the described embodiments are only some embodiments of the present application, but not all embodiments. The components of the embodiments of the present application, which are generally described and illustrated in the figures herein, may be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present application, as provided in the accompanying drawings, is not intended to limit the scope of the application, as claimed, but is merely representative of selected embodiments of the application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present application without making any inventive effort, are intended to be within the scope of the present application.
It should be noted that, the outcome as described in the present application refers to death, disease recurrence and disease exacerbation; if the situation does not occur at the time of follow-up, no ending is considered to occur.
Of course, the above definition is not limiting, the end of the present invention can be dead, and the end is not seen at the follow-up; the specific definition of the ending can be determined according to the actual situation.
The invention is described in further detail below with reference to fig. 1 to 3:
a cancer survival analysis method based on GAN and multitasking learning, comprising the steps of:
step 1: acquiring survival data of a cancer patient, forming a survival data set of the cancer patient, and taking part of the survival data in the survival data set as a training set;
the survival data of the cancer patient includes patient characteristics, time of observation, and outcome type of last follow-up time; if no outcome appears in the last follow-up time of the patient, the observation time is the deletion time of the patient; if the last follow-up time of the patient comes to a final result, the observed time is the survival time of the patient.
Assuming that survival data for a total of N patients is acquired, the survival data set D can be expressed as:
x is a patient characteristic and generally includes characteristics of the patient such as demographic information, tumor pathology, and treatment. s is the observation time, the cancer patient did not have any outcome at the last follow-up, s is the patient's deletion time; the patient had some outcome (e.g., death) at the last follow-up, s being the patient's survival time. e is the patient's outcome type at the last visit, and e=0 when no outcome occurred at the last visit. Assuming a total of K different outcomes, the range of values for e is: {0,1,2, … K }.
Step 2: carrying out data enhancement on Survival data in a training set based on a trained survivinval-GAN model;
the step 2 comprises the following steps:
step 2.1: dividing the survival data of cancer patients in the training set into two groups of deletion and ending occurrence according to whether ending occurs in the survival data, and respectively recording the numbers of the two groups;
step 2.2: training a survivinal-GAN model based on the Survival data with the outcomes;
step 2.3: searching out the optimal super parameters of the survivin-GAN model by using a grid search method and matching with five-fold cross validation, and retraining the survivin-GAN model by using the optimal super parameters;
step 2.4: randomly selecting K real survival times from the training set samples and respectively pairing the K real survival times with K different outcomes; sequentially inputting the K pairing results into a survivinal-GAN model to generate K Survival data;
step 2.5: n2 self-increases by K, i.e., n2=n2+k; n2 represents the number of survival data with ending;
step 2.6: judging whether N2 is smaller than N1, if not, directly ending; if yes, returning to the step 2.4 to continue execution until N2 is more than N1; where N1 represents the number of deleted data.
Referring to fig. 2, the survivinal-GAN model includes a generator and a arbiter;
the generator is composed of a full-connection network, and the number of layers of the full-connection layer and the number of neurons of each layer are super parameters; the label of survival data with the outcome consists of two parts: the ending type (e) and the time to live(s); the ending type and the survival time are respectively subjected to the Embedding, then are spliced with random noise (Z) and are input into a generator; the input of the discriminator is the false patient feature and the true patient feature of the generator output;
the discriminator is a multi-task full-connection network, and the number of layers of full-connection layers and the number of neurons in each layer are super parameters; the discriminators have three tasks respectively: the first task is to determine whether the input patient feature is true or generator-generated; the second task is to predict the outcome type using the survival data; the third task is to predict the time to live using the survival data. The outputs of these three tasks are represented using G1, G2 and G3, respectively.
The training steps of the survivinal-GAN model are as follows:
setting super parameters of a generator: the dimension of the Embedding output, the dimension of random noise, the number of layers of the full-connection layer and the number of neurons of each layer, the learning rate and the optimizers (Adam, SGD and the like);
setting super parameters of a discriminator: the number of layers of the full-connection layer and the number of neurons of each layer, the learning rate and the optimizer (Adam, SGD, etc.);
setting other super parameters: training round number (epoch) and batch_size, which is the number of training samples that are grabbed by one training;
and (3) data splicing: assume that the value of batch_size is m; randomly acquiring m noise data Z from standard normal distribution 1 ，Z 2 ，...Z m And m real survival data: (x 1 ，s 1 ，e 1 )，(x 2 ，s 2 ，e 2 )，...，(x m ，s m ，e m ). C for data after tag of real data is subjected to coding and spliced with noise data i A representation;
calculating the total loss of the generator:
L G ＝L G1 +L G2 +L G3 ；
wherein: l (L) G Representing the total loss of the generator, L G1 、L G2 And L G3 All represent loss functions;
loss function L G1 For bringing the features generated by the generator and the real features closer together, expressed as:
wherein: MES is expressed as a mean square loss function
Representing the mean square loss of q and p of the input; MSE (G (C) i )，x i ) For the generated patient characteristics G (C i ) And true patient characteristics x i Mean square error of (a); MSE (D (G (C) i ))[1]1) the output D (G (C) i ))[1]Mean square error with 1;
the loss function L G2 The outcome for reconciling the patient characteristic predictions generated by the generator with the outcome of the input is expressed as:
wherein: the expression of the cross entropy loss function is:
where h is the predicted probability of K outcomes and class is the true outcome; cross Entropy (D (G (C) i ))[2]，e i ) The output D (G (C i ))[2]And true ending e i Cross entropy of (2);
the loss function L G3 The time-to-live for the patient feature predictions generated by the generator and the entered time-to-live are expressed as:
wherein: MSE (D (G (C) i ))[2]，s i ) The output D (G (C i ))[3]And true survival time s i Is a mean square error of (c).
Updating training parameters of a generator: assuming that the learning rate is α, the SGD used by the optimizer. By theta G Representing parameters trained by the generator. Theta of each batch G The updating of (2) is as follows:
θ G ＝SGD(L G ，θ G ，α)；
calculating the total loss of the discriminator:
L D ＝L D1 +L D2 +L D3 ；
wherein: l (L) D Representing the total loss of the discriminator, L D1 、L D2 And L D3 All represent loss functions;
the loss function L D1 The patient characteristics used to enable the arbiter to identify the input are real or false, expressed as:
wherein MSE (D (x i )[1]1) for inputting the real patient characteristics x i When the first task of the discriminator outputs a mean square error of 1; MSE (D (G (C) i ))[1]0) is the inputPatient characteristics G (C) i ) When the first task output of the discriminator has a mean square error of 0;
the loss function L D2 For enabling the arbiter to accurately predict the outcome type of the patient, expressed as:
wherein CrossEntropy (D (x) i )[2]，e i ) To input real patient characteristics x i When the output of the second task of the discriminator is equal to e i Cross entropy loss of (2); cross Entropy (D (G (C) i ))[2]，e i ) Patient characteristics G (C) i ) When the output of the second task of the discriminator is equal to e i Cross entropy loss of (2);
the loss function L D3 For enabling the arbiter to accurately predict the survival time of the patient, expressed as:
where MSE (D (x i )[3]，s i ) To input real patient characteristics x i When the output of the third task and s of the discriminator i Mean square error of (a); MSE (D (G (C) i ))[3]，s i ) Patient characteristics G (C) i ) When the output of the third task and s of the discriminator i Is a mean square error of (c).
Updating training parameters of the discriminator: assuming a learning rate β, adam is used by the optimizer. By theta D Parameters representing the training of the arbiter; theta of each batch D The updating of (2) is as follows:
θ D ＝Adam(L D ，θ D ，β)；
end of training of generator and arbiter: judging whether the training round number reaches the appointed times, if so, finishing the training of the generator and the discriminator, and if not, continuing to execute the training of the discriminator and the generator until the appointed training times are met.
Step 3: constructing a cancer survival analysis model based on multitask learning, and training the cancer survival analysis model based on the enhanced training set data; searching out the optimal super parameters of the cancer survival analysis model by using a grid search method and matching with five-fold cross validation, and retraining the cancer analysis model by using the optimal super parameters;
the cancer survival analysis model comprises an expert network, a task network, an attention network and an auxiliary task network, wherein the expert network, the task network, the attention network and the auxiliary task network all belong to a fully-connected neural network, and the number of layers of the fully-connected layer and the number of neurons of each layer are super-parameters. A total of K outcomes, each outcome corresponding to an independent task network; the output of the K task networks is the probability that a cancer patient will have K different outcomes in the future, where T max For the longest lifetime of the patient in the training set. The auxiliary task network is a prediction of the outcome of the input patient features, which can help the model to better distinguish between different outcomes. The output of the attention mechanism network is the weight of k+2 expert networks. The outputs of the K+2 expert networks are multiplied by the outputs of the attention network and added to the task network and the auxiliary task network respectively. The K task networks and the 1 auxiliary task network share k+2 expert networks.
The training steps of the cancer survival analysis model are as follows:
A. setting super parameters: setting the number of full-connection layers of a task network, an auxiliary task network, an expert network and an attention network, the number of neurons at each layer, the learning rate, an optimizer (Adam, SGD and the like), the number of training rounds, batch_size, the number of prediction moments and the weights of 4 loss functions;
B. assuming that the value of batch_size is m, there are K total patient outcome types. Each batch of training requires that survival data of m patients be input into a cancer survival analysis model based on multitasking learning for training.
C. Calculating the loss of the cancer survival analysis model:
total loss function L of cancer survival analysis model s Expressed as:
L s ＝λ 1 ·L s1 +λ 2 ·L s2 +λ 3 ·L s3 +λ 4 ·L s4 ；
wherein: lambda (lambda) 1 ，λ 2 ，λ 3 ，λ 4 Weights of 4 loss functions respectively are super parameters; l (L) s1 、L s2 、L s3 And L s4 All represent loss functions;
loss function L s1 The role of (1) is to enable the model to learn a generic representation of the joint distribution of the ending occurrence time and ending event, L s1 Expressed as:
wherein:
characterizing a patient as x i Under the condition of (a) at time s i Occurrence e i Probability of ending P(s) i ，e i |x i )；
Is an indication function, and the condition is 1 if the indication function meets the condition, otherwise, the indication function is 0; f (F) j (s i |x i ) The expression of (2) is: f (F) j (s i |x i )＝P(s≤s i ，e i ＝j|x＝x i ) Represented by x in the patient characteristic i Under the condition of (1), patient outcome is j and occurs at time s i Previous probabilities;
loss function L s2 The effect of (1) is to make the survival time of the patient with higher ending occurrence probability predicted by the model smaller than that of the patient with lower ending occurrence rate, namely to improve the distinguishing capability of the model, L s2 Expressed as:
wherein: a is that j，i，p The expression of (2) is:
the indicator function is to find patient pairs (i, p) that can be risk compared; the expression of the η function is: />
Loss function L s3 The effect of (1) is to make the model predicted ending occurrence probability more similar to the true ending occurrence probability, namely to improve the calibration capability of the model, L s3 Expressed as:
wherein: the expression of the Sigmoid function is:
predicting the probability of the ith patient to develop an outcome j at time t for the model; />
The probability of ending j at time t for the ith patient in practice;
loss function L s4 The function of (1) is to enable the model to accurately predict the outcome of the patient, L s4 Expressed as:
wherein:
is taken as an auxiliaryNon-deleted patient outcome type assisting task network prediction +.>
And true ending type e i Cross entropy loss of (c).
D. Updating parameters of the model. Assuming that the learning rate is gamma, the optimizer is Adam, and the parameters of the model are theta S Then every batch theta S The updating of (2) is as follows:
θ s ＝Adam(L s ，θ s ，γ)；
E. and C, judging whether the training round number of the cancer survival analysis model accords with the designated times, if not, returning to the step B, and storing the cancer survival analysis model until the training round number accords with the designated times.
Dividing part of survival data in the survival data set into test sets, and evaluating the performance of the trained cancer survival analysis model by using the test sets, wherein the evaluation index is C-index; the survival data set may be divided into training and testing sets in a 4:1 ratio.
The specific calculation steps of the C-index are as follows:
a. pairing all patients pairwise;
b. if there is a situation in the pairing that the observation time of the patient A is smaller than that of the patient B and the patient A does not have a ending, the pairing is excluded; if there is a situation in which neither patient in the pairing has a outcome, the pairing is excluded; finally, useful pairing is obtained;
c. calculating the pairing number with the consistent predicted result and actual result in the useful pairing;
d. calculating a pairing value:
c-index = consistent log/useful log.
Step 4: inputting the characteristics of the cancer patient to be analyzed into the constructed cancer survival analysis model to obtain the probability of different fates of the cancer patient at each moment in a future period of time.
The invention enhances the data based on the survivin-GAN model, so that a large amount of non-deleted Survival data can be generated, thereby expanding the sample size and enhancing the accuracy and the robustness of model prediction.
The invention uses the GAN network to enhance the data, and inputs the characteristics, the survival time and the ending type of the cancer patient with ending into the GAN network for training, thereby generating a large amount of non-deleted survival data; further, constructing a multi-task learning cancer survival analysis model based on soft parameter sharing, respectively predicting different ending probabilities of a patient at each moment in a future period of time by a plurality of different tasks, and simultaneously adding an auxiliary task for distinguishing different ending on the basis of an original task; then, a loss function is added, wherein the loss function is the product of the mean square error of the predicted ending occurrence probability and the actual ending occurrence probability at each moment and Sigmoid (current moment); finally, inputting the survival data subjected to data enhancement into a multi-task learning cancer survival analysis model based on soft parameter sharing for training, wherein the output of the model is the probability of different fatalities of a patient in a future period of time; if the maximum observation time of the patient in the training sample is s max Then the cancer survival method based on multitasking learning can predict the patient in the future s max Probability of different outcomes occurring within a range.
The foregoing examples merely represent specific embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, several variations and modifications can be made without departing from the technical solution of the present application, which fall within the protection scope of the present application.
Claims (2)
1. A cancer survival analysis method based on GAN and multitasking learning, comprising the steps of:
step 1: acquiring survival data of a cancer patient, forming a survival data set of the cancer patient, and taking part of the survival data in the survival data set as a training set;
step 2: carrying out data enhancement on Survival data in a training set based on a trained survivinval-GAN model;
step 3: constructing a cancer survival analysis model based on multitask learning, and training the cancer survival analysis model based on the enhanced training set data; searching out the optimal super parameters of the cancer survival analysis model by using a grid search method and matching with five-fold cross validation, and retraining the cancer analysis model by using the optimal super parameters;
step 4: inputting the characteristics of the cancer patient to be analyzed into the constructed cancer survival analysis model to obtain the probability of different fates of the cancer patient at each moment in a future period of time;
the step 2 comprises the following steps:
step 2.1: dividing the survival data of the cancer patients in the training set into two groups of deletion and ending occurrence according to whether the obtained survival data is ending or not, and respectively recording the numbers of the two groups;
step 2.2: training a survivinal-GAN model based on the Survival data with the outcomes;
step 2.3: searching out the optimal super parameters of the survivin-GAN model by using a grid search method and matching with five-fold cross validation, and retraining the survivin-GAN model by using the optimal super parameters;
step 2.4: randomly selecting K real survival times from the training set samples and respectively pairing the K real survival times with K different outcomes; sequentially inputting the K pairing results into a survivinval-GAN model to generate K Survival data with ending;
step 2.5: n2 self-increases by K, i.e., n2=n2+k, N2 represents the number of survival data with the occurrence of a ending;
step 2.6: judging whether N2 is smaller than N1, if not, directly ending; if yes, returning to the step 2.4 to continue execution until N2 is more than N1; wherein N1 represents the number of deleted data;
the survivinal-GAN model includes a generator and a arbiter;
the generator comprises a full-connection network, wherein the number of layers of the full-connection layer of the full-connection network and the number of neurons of each layer are super parameters;
the discriminator is a multi-task full-connection network, and the number of layers of full-connection layers of the discriminator and the number of neurons of each layer are super parameters;
the arbiter comprises three tasks, the first task being for determining whether the input patient feature is true or generator generated; the second task predicts a outcome type based on the survival data; a third task predicts a time to live based on the survival data;
the training steps of the survivinal-GAN model are as follows:
setting super parameters of a generator: the dimension of the Embedding output, the dimension of random noise, the number of layers of the full-connection layer, the number of neurons of each layer, the learning rate and the optimizer;
setting super parameters of a discriminator: the number of layers of the full-connection layer, the number of neurons of each layer, the learning rate and the optimizer;
setting other super parameters: the training round number and the batch_size are the number of training samples grabbed by one training;
and (3) data splicing: randomly acquiring m noise data from standard normal distribution, encoding the input m tags of real living data by an encoding layer, and then splicing the tags with the noise data to obtain data C i ；
Calculating the total loss of the generator:
L G ＝L Gi +L c2 +L c3 ；
wherein: l (L) G Representing the total loss of the generator, L G1 、L G2 And L G3 All represent loss functions;
updating training parameters of a generator: updating training parameters of the generator based on the total loss function of the generator and a preset learning rate;
calculating the total loss of the discriminator:
L D ＝L D1 +L D2 +L D3 ；
wherein: l (L) D Representing the total loss of the discriminator, L D1 、L D2 And L D3 All represent loss functions;
updating training parameters of the discriminator: updating training parameters of the discriminator based on a total loss function of the discriminator and a preset learning rate;
end of training of generator and arbiter: judging whether the training round number reaches the appointed times, if so, finishing the training of the generator and the discriminator, and if not, continuing to execute the training of the discriminator and the generator until the appointed training times are met;
the cancer survival analysis model comprises an expert network, a task network, an attention network and an auxiliary task network;
A. setting super parameters: setting the number of full-connection layers of a task network, an auxiliary task network, an expert network and an attention network, the number of neurons at each layer, the learning rate, an optimizer, the number of training rounds, the batch_size, the number of prediction moments and the weight of 4 loss functions;
B. presetting the value of batch_size as m, wherein the total number of the ending types of the patients is K; in the training process of each batch, the survival data of m patients are input into a cancer survival analysis model for training;
C. calculating the loss of the cancer survival analysis model:
total loss function L of cancer survival analysis model s Expressed as:
L s ＝λ 1 ·L s1 +λ 2 ·L s2 +λ 3 ·L s3 +λ 4 ·L s4 ；
wherein: lambda (lambda) 1 ,λ 2 ,λ 3 ,λ 4 Weights of 4 loss functions respectively are super parameters; l (L) s1 、L s2 、L s3 And L s4 All represent loss functions;
D. based on the total loss function L S Updating parameters theta of cancer survival analysis model by preset optimizer Adam and learning rate gamma S ：
θ s ＝Adam(L s ，θ s ，γ)；
E. B, judging whether the training round number of the cancer survival analysis model accords with the appointed times, if not, returning to the step B, and storing the cancer survival analysis model until the training round number accords with the appointed times;
the loss function L s1 Expressed as:
wherein:
characterizing a patient as x i Under the condition of (a) at time s i Occurrence e i Probability of ending P(s) i ,e i |s i )；/>
Is an indication function, and the condition is 1 if the indication function meets the condition, otherwise, the indication function is 0; f (F) j (s i |x i ) The expression of (2) is: f (F) j (s i |x i )＝P(s≤s i ,e i ＝j|x＝x i ) Represented by x in the patient characteristic i Under the condition of (1), patient outcome is j and occurs at time s i Previous probabilities;
loss function L s2 Expressed as:
wherein: a is that j,i,p The expression of (2) is:
the indicator function is to find patient pairs (i, p) that can be risk compared; the expression of the η function is: />
Loss function L s3 Expressed as:
wherein: the expression of the Sigmoid function is:
predicting the probability of the ith patient to develop an outcome j at time t for the model; />
The probability of ending j at time t for the ith patient in practice;
loss function L s4 Expressed as:
wherein:
the outcome type of non-deleted patients for assisting task network prediction +.>
And true ending type e i Cross entropy loss of (2);
the loss function L G1 Expressed as:
wherein: MES is a mean square loss function expressed as
Representing the mean square loss of q and p of the input; MSE (G (C) i ),x i ) For the generated patient characteristics G (C i ) And true patient characteristics x i Mean square error of (a); MSE (D (G (C) i ))[1]1) the output D (G (C) i ))[1]Mean square error with 1;
the loss function L G2 Expressed as:
wherein: the expression of the cross entropy loss function is:
where h is the predicted probability of K outcomes and class is the true outcome; cross Entropy (D (G (C) i ))[2],e i ) The output D (G (C i ))[2]And true ending e i Cross entropy of (2);
the loss function L G3 Expressed as:
wherein: MSE (D (G (C) i ))[2],s i ) The output D (G (C i ))[3]And true survival time s i Mean square error of (a);
the loss function L D1 Expressed as:
wherein MSE (D (x i )[1]1) for inputting the real patient characteristics x i When the first task of the discriminator outputs a mean square error of 1; MSE (D (G)(C i ))[1]0) patient characteristics G (C) generated for the input generator i ) When the first task output of the discriminator has a mean square error of 0;
the loss function L D2 Expressed as:
wherein CrossEntropy (D (x) i )[2],e i ) To input real patient characteristics x i When the output of the second task of the discriminator is equal to e i Cross entropy loss of (2); cross Entropy (D (G (C) i ))[2],e i ) Patient characteristics G (C) i ) When the output of the second task of the discriminator is equal to e i Cross entropy loss of (2);
the loss function L D3 Expressed as:
where MSE (D (x i )[3],s i ) To input real patient characteristics x i When the output of the third task and s of the discriminator i Mean square error of (a); MSE (D (G (C) i ))[3],s i ) Patient characteristics G (C) i ) When the output of the third task and s of the discriminator i Is a mean square error of (c).
2. The GAN and multitasking based cancer survival analysis method of claim 1, wherein said cancer patient's survival data includes patient characteristics, time of observation and outcome type at last follow-up time; if no outcome appears in the last follow-up time of the patient, the observation time is the deletion time of the patient; if the last follow-up time of the patient comes to a final result, the observed time is the survival time of the patient.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211240631.1A CN115565669B (en) | 2022-10-11 | 2022-10-11 | Cancer survival analysis method based on GAN and multitask learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211240631.1A CN115565669B (en) | 2022-10-11 | 2022-10-11 | Cancer survival analysis method based on GAN and multitask learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115565669A CN115565669A (en) | 2023-01-03 |
| CN115565669B true CN115565669B (en) | 2023-05-16 |
Family
ID=84744408
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211240631.1A Active CN115565669B (en) | 2022-10-11 | 2022-10-11 | Cancer survival analysis method based on GAN and multitask learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115565669B (en) |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110660478A (en) * | 2019-09-18 | 2020-01-07 | 西安交通大学 | Cancer image prediction and discrimination method and system based on transfer learning |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TW200532523A (en) * | 2004-02-27 | 2005-10-01 | Aureon Biosciences Corp | Methods and systems for predicting occurrence of an event |
| CN111640510A (en) * | 2020-04-09 | 2020-09-08 | 之江实验室 | Disease prognosis prediction system based on deep semi-supervised multitask learning survival analysis |
| CN112687327B (en) * | 2020-12-28 | 2024-04-12 | 中山依数科技有限公司 | Cancer survival analysis system based on multitasking and multi-mode |
-
2022
- 2022-10-11 CN CN202211240631.1A patent/CN115565669B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110660478A (en) * | 2019-09-18 | 2020-01-07 | 西安交通大学 | Cancer image prediction and discrimination method and system based on transfer learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115565669A (en) | 2023-01-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhao et al. | A novel combination method for conflicting evidence based on inconsistent measurements | |
| CN104572583B (en) | Method and system for data densification | |
| CN111243736B (en) | Survival risk assessment method and system | |
| Obajemu et al. | A new fuzzy modeling framework for integrated risk prognosis and therapy of bladder cancer patients | |
| CN111917785A (en) | Industrial internet security situation prediction method based on DE-GWO-SVR | |
| CN110838364A (en) | Crohn disease prediction method and device based on deep learning hybrid model | |
| CN113889262A (en) | Model-based data prediction method and device, computer equipment and storage medium | |
| Shmueli et al. | Neural network analysis of travel behaviour | |
| CN115982141A (en) | Characteristic optimization method for time series data prediction | |
| CN114239397A (en) | Soft measurement modeling method based on dynamic feature extraction and local weighted deep learning | |
| CN115565669B (en) | Cancer survival analysis method based on GAN and multitask learning | |
| Guzmán-Torres et al. | Non-destructive Tests for Estimating the Tensile Strength in Concrete with Deep Learning | |
| Termritthikun et al. | Neural architecture search and multi-objective evolutionary algorithms for anomaly detection | |
| Martsenyuk et al. | On development of machine learning models with aim of medical differential diagnostics of the comorbid states | |
| CN111882045B (en) | Brain time-space network decomposition method and system based on micronerve structure search | |
| CN114678083A (en) | Training method and prediction method of chemical genetic toxicity prediction model | |
| Prahmana et al. | Knearst Algorithm Analysis–Neighbor Breast Cancer Prediction Coimbra | |
| CN114519508A (en) | Credit risk assessment method based on time sequence deep learning and legal document information | |
| CN112634993A (en) | Prediction model and screening method for activation activity of estrogen receptor of chemicals | |
| Zhou et al. | An improved information volume of mass function based on plausibility transformation method | |
| Bhatia | Deep Learning-Based Approach for Thyroid Dysfunction Prediction | |
| Srivathsa | Knowledge discovery in medical mining by using genetic algorithms and artificial neural networks | |
| CN117609731A (en) | Gramineae graduation forward prediction method and visual analysis system based on causal inference | |
| Abdel-Kader | Prediction of Lung Cancer Using Supervised Machine Learning | |
| Stainton et al. | On the application of quantization for mobile optimized convolutional neural networks as a predictor of realtime ageing biomarkers |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |