KR102190299B1 - Method, device and program for predicting the prognosis of gastric cancer using artificial neural networks - Google Patents

Abstract

A method for predicting the prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention includes: obtaining clinical data of a plurality of gastric cancer patients and survival period data after gastric cancer onset; Obtaining input data for training and output data for training from the clinical data and the survival period data; Learning an artificial neural network including an input layer, a hidden layer, and an output layer by using the training input data and the training output data; And generating a model for predicting the survival rate of gastric cancer patients by using the learned artificial neural network.

Description

Method, device and program for predicting the prognosis of gastric cancer using artificial neural networks

The present invention relates to a method, apparatus, and program for predicting the prognosis of gastric cancer using an artificial neural network.

Gastric cancer is a cancer that occurs as many times as the number 1 incidence and the 2nd cause of death among all cancers in Korea. Gastric cancer is staged according to the degree of invasion of the surrounding structures of the tumor, metastasis to regional lymph nodes, and metastasis to other organs, and accordingly, treatment and prognosis are different. The prognosis is getting worse as the stage increases, and the prognosis is so poor that the 5-year survival rate is only 3%, especially in the terminal stage (Stage 4).

However, even with the same stage, the prognosis may be heterogeneous depending on the person, which comes from various environmental and genetic causes. Therefore, various methods are being developed that allow the individual to accurately predict the prognosis of cancer.

For example, Korean Patent Registration No. 10-1415257 discloses a method for diagnosing the prognosis of gastric cancer by measuring the overexpression level of microRNA-196b RNA and HOXA10 (Homeobox A10) protein.

On the other hand, in Korean Patent Registration No. 10-1504818, A system for predicting the prognosis of gastric cancer by cluster analysis of several gene expression profiles is disclosed.

However, the existing gastric cancer prognosis prediction method has a problem that it is not possible to accurately predict the survival rate of gastric cancer patients because the gastric cancer prognosis is simply divided into a low-risk group, a medium-risk group, and a high-risk group.

An object of the present invention is to provide a method, apparatus, and program for predicting the prognosis of gastric cancer using an artificial neural network to solve various problems including the above problems.

However, these problems are exemplary, and the scope of the present invention is not limited thereby.

A method for predicting the prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention includes: obtaining clinical data of a plurality of gastric cancer patients and survival period data after gastric cancer onset; Obtaining input data for training and output data for training from the clinical data and the survival period data; Learning an artificial neural network including an input layer, a hidden layer, and an output layer by using the training input data and the training output data; And generating a model for predicting the survival rate of gastric cancer patients by using the learned artificial neural network.

In one embodiment, the input data for learning may include molecular genetic subtype data of the plurality of gastric cancer patients, and the subtype is MSI (microsatellite instable) subtype, MSS/EMT subtype, MSS/TP53+ subtype , MSS/TP53- subtype.

In an embodiment, the input layer may include four nodes to which the molecular genetic subtype data is input.

In one embodiment, the training of the artificial neural network using the training input data and the training output data comprises embedding each variable of the training input data into a vector of two or more dimensions to calculate an embedding layer. It may include steps.

In one embodiment, the step of obtaining input data for learning and output data for learning from the clinical data and the survival period data, respectively, comprises missing values using a k-nearest neighbor algorithm (knn). data, NaN) may be added.

In one embodiment, the hidden layer of the artificial neural network may include at least one RNN layer.

An apparatus for predicting prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention includes: a data acquisition unit for acquiring clinical data of a plurality of gastric cancer patients and survival period data after the onset of gastric cancer; And learning input data for learning and output data for learning from the clinical data and the survival period data, and learning an artificial neural network including an input layer, a hidden layer, and an output layer by using the training input data and the training output data. part; And a survival rate prediction model generation unit that generates a model for predicting the survival rate of gastric cancer patients using the learned artificial neural network.

In one embodiment, the input data for learning may include molecular genetic subtype data of the plurality of gastric cancer patients, and the subtype is MSI (microsatellite instable) subtype, MSS/EMT subtype, MSS/TP53+ subtype, MSS/ Contains the TP53-subtype.

In an embodiment, the input layer may include four nodes to which the molecular genetic subtype data is input.

In an embodiment, the artificial neural network learning unit may calculate an embedding layer by embedding each variable of the training input data into a vector of 2D or more.

In one embodiment, the artificial neural network learning unit may add missing data (NaN) of the training input data using a k-nearest neighbor algorithm (knn).

In one embodiment, the hidden layer of the artificial neural network may include at least one RNN layer.

Another embodiment of the present invention discloses a computer program stored in a medium in order to execute the method for predicting the prognosis of gastric cancer using an artificial neural network described above using a computer.

Other aspects, features, and advantages other than those described above will become apparent from the following drawings, claims, and detailed description of the invention.

According to the method, apparatus and program for predicting the prognosis of gastric cancer using an artificial neural network according to the present invention, it is possible to accurately predict the prognosis of a gastric cancer patient for each individual. In addition, since the prognosis of each treatment method can be simulated using the learned artificial neural network, a customized treatment method for each patient can be determined. Of course, the scope of the present invention is not limited by these effects.

1 is a flow chart showing a method for predicting prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention.
Figure 2 is a graph showing the correlation of survival rates according to the microsatellite instable (MSS) subtype, MSS/EMT subtype, MSS/TP53+ subtype, and MSS/TP53- subtype in the gastric cancer patient group studied by ACRG (Asian Cancer Research Group). .
3 is a diagram schematically showing the topology of an artificial neural network according to an embodiment of the present invention.
4 is a diagram schematically showing a part of a heatmap graph of an artificial neural network according to an embodiment of the present invention.
5 is a diagram illustrating a method of generating a model for predicting the survival rate of a t-th section of a gastric cancer patient according to a method for predicting a prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention.
6 is a diagram schematically showing a part of a heatmap graph of an artificial neural network sequentially trained by year according to an embodiment of the present invention.
7 is a graph comparing the AUC values of the ROC graph of the survival rate prediction model for each year during training.
8 is an ROC graph verifying the survival rate prediction model for each year with separate test data.
9 is a graph comparing the survival rate predicted by the survival rate prediction model and the actual survival rate.
10 to 14 are graphs showing a decision curve of a survival rate prediction model after 1, 2, 3, 4 and 5 years, respectively.
15 is a graph showing the learning effect of an artificial neural network and a simple artificial neural network for comparison according to an embodiment of the present invention.
16 is a schematic diagram of a method of retraining an artificial neural network using data from other regions.
FIG. 17 is a heat map graph comparing a retrained model by adding Singapore data to a recurrent survival network (RSN), which is a prognosis prediction model for gastric cancer using an artificial neural network according to an embodiment of the present invention, and a model trained only with Singapore data. to be.
18 is a graph comparing the effects of the original model, a retraining model that is retrained by adding Singapore data, and a Singapore model that is trained only with Singapore data.
19 is a diagram schematically showing the configuration of an apparatus for predicting prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention.

Best mode for carrying out the invention

Since the present invention can apply a variety of transformations and can have various embodiments, specific embodiments will be illustrated in the drawings and described in the detailed description. Effects and features of the present invention, and a method of achieving them will be apparent with reference to the embodiments described later in detail together with the drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various forms.

In the following embodiments, terms such as first and second are not used in a limiting meaning, but are used for the purpose of distinguishing one component from another component.

In the following examples, the singular expression includes the plural expression unless the context clearly indicates otherwise.

In the following embodiments, terms such as include or have means that the features or elements described in the specification are present, and do not preclude the possibility of adding one or more other features or elements in advance.

In the following embodiments, a'node' refers to an object of an abstract concept that can have a specific value and change its value through a specific algorithm, and can be connected to other nodes.

In the following embodiments, the term'input layer' is a set of one or more nodes having a specific variable assigned by a user, and the term'output layer' is one or more nodes having a result of the procedure according to a specific procedure determined by the user. Is a set of nodes, and the'hidden layer' refers to a set of one or more nodes that store intermediate results and temporary values that appear temporarily when performing a procedure specified by a user.

Links may exist between the nodes of the input layer and the nodes of the hidden layer, and between the nodes of the hidden layer and the nodes of the output layer, respectively, and these links may have a specific weight given by a user-defined procedure.

In the following examples, the term'prognosis' is a medical term for predicting a patient's survival rate, disease progression, and recovery.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings, and when describing with reference to the drawings, the same or corresponding constituent elements are assigned the same reference numerals, and redundant descriptions thereof will be omitted. .

1 is a flow chart showing a method for predicting prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention.

A method for predicting the prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention includes: obtaining clinical data of a plurality of gastric cancer patients and survival period data after gastric cancer onset; Obtaining input data for training and output data for training from the clinical data and the survival period data; Learning an artificial neural network including an input layer, a hidden layer, and an output layer by using the training input data and the training output data; And generating a model for predicting the survival rate of gastric cancer patients by using the learned artificial neural network.

Referring to FIG. 1, a step (S10) of acquiring clinical data of a plurality of gastric cancer patients and survival period data after onset of gastric cancer is performed.

In the present specification, the clinical data includes physical personal information such as the patient's age and sex, and a surgical record after the occurrence of gastric cancer, and a pathological record related to gastric cancer such as recurrence. The survival period data after the onset of gastric cancer is the period from the time of recognition of the occurrence of gastric cancer to death in the case of a patient who has already died, and the time period from the time of recognition of the occurrence of gastric cancer to the time of implementing the present invention in the case of a surviving patient. It can mean a period.

Clinical data and survival time after gastric cancer onset data may be obtained from one or more hospitals or local gastric cancer patients. The clinical data may be obtained from a medical image of a patient or from a specimen test result of a patient, but is not limited thereto. The present inventors obtained clinical data and survival period data from 1187 gastric cancer patients at Samsung Seoul Hospital who followed up for more than 5 years.

Thereafter, a step (S20) of acquiring input data for training and output data for training, respectively, from the clinical data and the survival period data is performed.

The input data for training refers to data to be input to a node of an input layer in order to learn an artificial neural network to be described later. [Table 1] shows variables that can be included in input data for learning and their classification.

variable Classification Molecular genetic subtype molecular subtype 1: MSI molecular subtype 2: MSS/EMT molecular subtype 3: MSS/TP53+ molecular subtype 4: MSS/TP53- RAS signature Real value Sex Male: 1, female: 2 age Real value HER2 0= negative; 1= positive WHO cancer classification 1=w/d adeno 2=m/d adeno 3=p/d adeno 4=signet ring 5= mucinous 6=papillary adeno 7=adenosquamous 8=undifferentiated ca 9=hepatoid adenoca 10=tubular adenoca 11=others (text) LAUREN pathologic classification 1=intestinal 2=diffuse 3=mixed Pathological findings: Perineural invasion PNI 0=(-), 1=(+) Pathological findings: Lymphatic vessel involvement inv 0=(-), 1=(+) TNM stage T N M Number of resected lymph nodes # LN dissected: integer value Number of lymph nodes invading cancer # of positive node (+): integer value Revised location Cardia Body Antrum Surgical method TG; total, ST: subtotal 1=TG, 2=STG Completion of adjuvant therapy 0=completed 1=not completed Type of adjuvant therapy (ADJ CTx. Description) CCRT LF_RT XP_RT XP Others Recurrence no=0 yes=1 Recurrence location First site of recurrence_liver 0=(-), 1=(+) First site of recurrence_peritoneal seeding 0=(-), 1=(+) First site of recurrence_ascites (clinically significant) 0=(-), 1=(+) First site of recurrence_intraabdominal_LN 0=(-), 1=(+) First site of recurrence_distant lymph node 0=(-), 1=(+) First site of recurrence_bone 0=(-), 1=(+) recurrence sites_others 0=(-), 1=(+)

It is obvious that not only the above variables but also various clinical variables such as K-ras mutation status, operation date, and recurrence time can be included in the input data for learning.

Among them, the molecular genetic subtypes of gastric cancer are the microsatellite instable (MSS) subtypes, microsatellite stable (MSS) subtypes, and epithelial-to-mesenchymal subtypes (EMTs) classified by measuring the microsatellite instability of gastric cancer samples. transition) and classified into MSS/EMT subtypes, MSS/epithelial subtypes, and MSS/TP53+ subtypes and MSS/TP53- subtypes classified by measuring the activity of TP53 (Tumor Protein 53) against MSS/epithelial subtypes.

According to an embodiment of the present invention, the input data for learning includes molecular genetic subtype data of a plurality of gastric cancer patients, and the subtype is MSI (microsatellite instable) subtype, MSS/EMT subtype, MSS/TP53+ subtype, MSS/TP53- subtype.

In order to classify the molecular genetic subtype of the new gastric cancer as described above, the present inventors have conducted a primary tumor sample of a patient who has undergone total gastrectomy or subtotal gastrectomy at Samsung Medical Center (SMC). n=300) was obtained.

First, gastric cancer samples can be classified into microsatellite instable (MSS) subtypes and microsatellite stable (MSS) subtypes according to the degree of microsatellite instability.

In the case of the MSI subtype, gene mutations occur a lot, and cancer progression is relatively slow. In addition, about 60% of patients with the MSI subtype corresponded to stage 1 or stage 2 of cancer, and the average survival period was the longest at 100.9 months, and the intestinals belong to a large proportion of the Lauren classification.

MSS subtypes can be classified into MSS/EMT subtypes and MSS/epithelial subtypes by measuring EMT (epithelial-tomesenchymal transition). If the EMT is measured to be similar to mesenchymal, it is classified as an MSS/EMT subtype, and if it is measured to be similar to an epithelium, it is classified as an MSS/epithelial subtype.

The MSS/EMT subtype mainly contains diffuse tumors in the Lauren classification, rarely shows genetic mutations, and has a poor prognosis in gastric cancer patients. In addition, this subtype is found at a lower age, and the cancer progresses rapidly and has the highest recurrence rate (63%).

In the case of the MSS/epithelial subtype, the activity of TP53 (Tumor Protein 53) can be measured and classified into the MSS/TP53+ subtype and the MSS/TP53- subtype. When the activity of TP53 is high, it is classified as an MSS/TP53+ subtype, and when the activity is low, it is classified as an MSS/TP53- subtype. The MSS/TP53+ and MSS/TP53- subtypes have moderate prognosis and recurrence rates among the four subtypes. Specifically, in the case of MSS/TP53+ subtype, there are many intestinal gastric cancer. The MSS/TP53- subtype is the most common subtype among the analyzed subjects, and the prognosis of gastric cancer patients is worse than that of the MSS/TP53+ subtype due to the loss of TP53 function.

Figure 2 is a graph showing the correlation of survival rates according to the microsatellite instable (MSS) subtype, MSS/EMT subtype, MSS/TP53+ subtype, and MSS/TP53- subtype in the gastric cancer patient group studied by ACRG (Asian Cancer Research Group). .

As shown in FIG. 2, the MSI subtype showed the best survival rate, followed by MSS/TP53+ and MSS/TP53-, and the MSS/EMT subtype showed the worst survival rate (log-rank, P = 0.0004). That is, the survival rate of gastric cancer patients is confirmed to be the highest in the MSI subtype and the lowest in the MSS/EMT subtype, so the survival rate can be predicted according to the subtype analysis through genetic analysis. Therefore, when learning an artificial neural network using the molecular genetic subtype of gastric cancer as an input value, the accuracy of predicting the prognosis of gastric cancer using the artificial neural network is increased.

Among the input data for learning illustrated in [Table 1], variables that can be categorized into three or more, such as WHO tumor type and molecular genetic subtype, can be converted into vectors using a one-hot encoding technique. When there are only two classifications, such as sex and cancer recurrence, the two classifications can be labeled as 0, 1 or 1, 2 and converted into a single value. Quantitative variables such as RAS signatures can be normalized and processed to change into a single number. Through this, input data for learning can be mathematically processed.

For example, assume that there is patient A having clinical data as shown in [Table 2] below. In [Table 2], only some of the variables shown in [Table 1] are described.

At this time, when the molecular genetic subtype information of patient A is input to the node of the artificial neural network where the input data for learning is input, for example, [MSI node, MSS/EMT node, MSS/TP53+ node, MSS/TP53- node] = [1, 0, 0, 0]. That is, the input layer of the artificial neural network may include four nodes for inputting the above-described molecular genetic subtype data.

On the contrary, if you label as MSI = 1, MSS/EMT = 2, MSS/TP53+ = 3, MSS/TP53- = 4 and enter as [subtype node] = [1] in the input layer, each classification in the hidden layer There is a loss of information about

On the other hand, if there are only two classifications, such as gender, when the gender information is input into the input data for learning, [Gender node] = 0 or 1 is input. That is, for a variable with only two classifications, one node is allocated. In such a case, it is possible to grasp the effect of gender on the patient's survival rate through the weight in the link connected to the node, so it is not necessary to allocate two nodes.

Meanwhile, according to an embodiment of the present invention, the input data for learning may include survival rate or mortality data of gastric cancer patients. For example, if there are i data obtained from clinical data of each gastric cancer patient, the input data for learning may be i+2 data sets including survival rate data and mortality data for the previous year of the gastric cancer patient. This will be described later.

The output data for training refers to data to be input to the node of the output layer to learn the artificial neural network. The output data for learning may include information on whether the patient survives within a certain period after the onset of gastric cancer.

After the step of acquiring the training input data and the training output data, the step of training the artificial neural network using the training input data and the training output data (FIG. 1, S30) is performed.

3 is a diagram schematically showing the topology of an artificial neural network according to an embodiment of the present invention. The artificial neural network has an input layer with multiple nodes, one or more hidden layers and an output layer. The artificial neural network (Recurrent Survival Network, hereinafter also referred to as RSN) according to an embodiment of the present invention has eight hidden layers, but the present invention is not limited thereto.

The input layer of the artificial neural network has n _in nodes. 3 illustrates that the number of nodes in the input layer is 49, but is not limited thereto. In each node, a value obtained by mathematically processing the classification according to clinical variables of each gastric cancer patient is input. At this time, the input layer has a shape like an n _in × 1 matrix.

According to an embodiment, the training of the artificial neural network using input data for training and output data for training includes calculating an embedding layer by embedding each variable of the training input data into a vector of two or more dimensions. I can.

Each variable of the training input data input to each node of the input layer has a 1-dimensional real value, which can be embedding to make a vector of two or more dimensions. For example, 1, which is the value of the MSI node of patient A in [Table 2] above, can be converted into a 9-dimensional vector as follows using a known embedding method.

[MSI node embedding vector] = [0.1 0.3 -0.1 -0.5 -0.7 0.2 0.3 0.3 -0.1]

Similarly, if the value of each node is embedded in 9 dimensions, for example, the following matrix can be obtained.

The dimensions of the vectors and matrices and the values of each component are exemplarily determined to aid the understanding of the present invention, and do not limit the scope of the present invention.

In the embodiment of the present invention, each real variable was replaced with a 32-dimensional vector through embedding for every 49 nodes, and accordingly, an embedding layer in the form of a 49×32 matrix was calculated. Originally, embedding is a concept devised to quantify words in the field of natural language processing, but in the present invention, the numerical values of each real variable itself are also vectorized through embedding. By embedding each patient's input data for learning, the degree of similarity between each patient can be measured through vector operation, and the data can be processed in an integrated manner without loss of information inherent in each variable.

According to an embodiment of the present invention, the step of obtaining input data for training and output data for training from clinical data and survival period data, respectively, includes missing values (k-nearest neighbor algorithm: knn). missing data, NaN) may be added.

For example, assume that there are patients A, B, and C having clinical data as shown in [Table 3] below. In the case of patient C, HER2 was not tested, and the value corresponding to HER2 was missing (NaN).

At this time, whether the clinical data of patient C is closer to which of patient A and patient B is closer to, for example, the distance of each patient's input data vector for learning can be determined. In the exemplified case of [Table 3], since the clinical data of patient C is closer to patient B than patient A, 1 can be assigned to patient C's HER2 value.

In practice, the number of patients to be compared is large, so the above example is merely a simplification of the situation to explain the knn algorithm, and does not necessarily reflect the actual situation. At this time, since there are various known knn algorithms, detailed descriptions are omitted in this specification.

For example, if the item of input data for learning input to the input layer of the artificial neural network of the present invention is not included in the clinical data of another region or hospital, the missing item may be added using the knn algorithm. Therefore, it is possible to retrain the artificial neural network by adding data from other regions with missing data, which will be described later.

Meanwhile, the output layer of the artificial neural network has n _out nodes. 3 illustrates that there are two nodes in the output layer, but the present invention is not limited thereto.

The output layer may include a node representing the survival rate and mortality rate of the patient in N+1 year (or may be an arbitrary time unit such as a half year, quarter, month, day, etc., and N is an integer greater than or equal to 0). For example, when the artificial neural network is learned to predict the survival rate 2 years after the onset of gastric cancer of a gastric cancer patient, a value indicating whether a specific patient survives 2 years after the onset of gastric cancer may be input to the node of the output layer. For example, if a specific patient died 2 years before the onset of gastric cancer, [survival rate node, mortality node] = [0, 1], and when a specific patient died 2 years after the onset of gastric cancer, [survival rate node, mortality node] = [ 1, 0].

However, according to an embodiment of the present invention, a processing method capable of assigning a score by ranking without processing [survival rate node, mortality node] as [0, 1] when a patient dies is proposed. In an embodiment of the present invention, [survival rate node, mortality node] of a deceased patient may be [p, 1-p], where a non-zero score value may be assigned to p. In this case, the score may be given in proportion to the survival period of the deceased patient. Here, the survival period may be divided into at least monthly. For example, for patients who survived 3 months in N+1 year, the survival rate score at N+1 year was 3/12, and the value of [survival rate node, mortality node] was [3/12, 1-3/12]=[0.25]. , 0.75].

Meanwhile, the input layer may include a [N-year survival rate node, N-year mortality node]. That is, according to the present invention, input data for learning including clinical data information of gastric cancer patients and N-year survival rate and N-year mortality information are input to the input layer, and N+1 year survival rate and mortality information proportional to the patient's survival period are included. The artificial neural network is trained by inputting the learning output data to the output layer.

When learning artificial neural networks, hidden layers are used. Nodes of each hidden layer can be fully connected to each other with nodes of other adjacent hidden layers. In an embodiment of the present invention, an artificial neural network is trained using eight hidden layers including a recurrent neural network (RNN) layer using a long short term memory (LSTM) algorithm, but the number of hidden layers and types of algorithms, etc. Is not limited to this.

Referring back to FIG. 1, after the step of learning the artificial neural network (S30), a step (S40) of generating a model for predicting the survival rate of a gastric cancer patient using the learned artificial neural network is performed. After the artificial neural network is trained, the weight corresponding to each node is optimized to predict the survival rate, so the patient's survival rate through the output value obtained by inputting the input data obtained from the clinical data of an arbitrary gastric cancer patient into the input layer of the artificial neural network. Can be predicted.

Mode for carrying out the invention

<Example 1>

The research team constructed an artificial neural network based on data on more than 19 clinical variables and the results of follow-up for more than 5 years from 1187 patients at Samsung Medical Center. 85% (1009 persons) of the total data was used as the training data, and the remaining 15% (178 subjects) were used as the test data, which were randomly re-extracted over 100 times. 15% of the training data were reclassified for cross-validation.

The optimized model was trained 30 times for each training data, and performance was evaluated through the test data.

4 is a diagram schematically showing a part of a heatmap graph of an artificial neural network according to an embodiment of the present invention.

Graph 100 shows a state in which input data for learning is input to the input layer. The horizontal axis of the heatmap of graph 100 is the serial number of each gastric cancer patient, and the vertical axis corresponds to each node of the artificial neural network input layer. In one embodiment, a total of 49 nodes in the input layer include 47 nodes and 2 survival and mortality nodes obtained from the clinical data shown in [Table 1]. The value corresponding to each node is displayed as the color density.

Thereafter, the data of each gastric cancer patient is expressed as a matrix through an embedding process. In one embodiment, the data of gastric cancer patients are each represented by a 49×32 matrix. (Graphs 101, 102, 103...) After that, learning of coefficients for each node is performed. The outcome of learning is labeled as survival or death, and finally expressed as the probability of survival through a softmax function. At this time, referring to graph 130, the first-year survival rate for a plurality of gastric cancer patients is indicated by two nodes. That is, the first-year survival rate prediction artificial neural network converges a total of 49 node values (shown in graph 110) to a total of two node values (shown in graph 130).

The hidden layer of the artificial neural network according to an embodiment of the present invention may include at least one cyclic neural network (RNN) layer, and the RNN layer may use a long short term memory (LSTM) algorithm.

5 is a diagram illustrating a method of generating a model for predicting the survival rate of a t-th section of a gastric cancer patient according to a method for predicting a prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention.

According to an embodiment of the present invention, generating a model for predicting the survival rate may include training the artificial neural network for each time interval. The time period may be various, such as a year, a half year, a quarter, and a month, but hereinafter, a year will be described as an example. For example, the artificial neural network may be learned to predict the annual survival rate of gastric cancer patients from clinical data of gastric cancer patients, for example, the survival rate from 1 year to 5 years after onset.

According to an embodiment of the present invention, the step of generating a model for predicting survival rate comprises using the clinical data and the t-th section survival period data of the plurality of gastric cancer patients to generate a t-th section survival rate prediction model (SM _t ). Generating; And the t-th section survival rate prediction model (SM _t+) using the t-th section survival rate prediction data (S _t ) obtained from the t-th section survival rate prediction model and the t+1 section survival period data of the plurality of gastric cancer patients. It may include the step of generating ₁ ).

According to an embodiment of the present invention, 1, 2, ... of gastric cancer patients. , the survival rate for each t, t+1th section is predicted (t: a natural number), and in this case, the survival rate prediction result data in the tth section is used to predict the survival rate in the t+1th section. That is, the survival rate prediction for each section is performed in an inductive and sequential manner.

Referring to FIG. 5, a survival rate prediction model after 1 year (SM ₁ ) and a survival rate prediction model after t year (SM _t ) are shown. At this time, when the initial clinical data (X ₁ ) and the initial survival rate (S ₀ ) are input, the 1-year survival rate prediction model (SM ₁ ), an input/output function capable of outputting the survival rate (S ₁ ) after 1 year, uses the artificial neural network. It is created by learning.

At this time, the input data for learning input to the input layer of the artificial neural network includes initial clinical data (X ₁ ) and an initial survival rate (S ₀ ). The initial clinical data (X ₁ ), which is an input to the survival rate prediction model after 1 year, may be clinical data at the initial examination. The initial survival rate (S ₀ ) may be set to 1, for example.

As the learning output data for predicting the survival rate after 1 year, the survival data after 1 year obtained from the survival period data of the patient are used. For example, if a patient D died 15 months after the onset of gastric cancer, it survived one year after the onset, so the output data for learning to be compared with the value to be output in [Survival rate node, mortality node] of the output layer becomes [1, 0]. . The artificial neural network is trained to predict the survival rate after one year of gastric cancer patients using the input data for learning and the output data for learning.

Next, the survival rate prediction model after 2 years, an input/output function that can output the survival rate after 2 years (S ₂ ) when the clinical data after 2 years (X ₂ ) and the predicted survival rate after 1 year (S ₁ ) are input. (SM ₂ ) is created by learning the artificial neural network. At this time, the input data for learning input to the input layer of the artificial neural network includes clinical data after 2 years (X ₂ ) and a predicted survival rate after 1 year (S ₁ ).

As the learning output data, data on whether or not to survive after 2 years obtained from the patient's survival period data is used. For example, if a patient D survived 15 months after the onset of gastric cancer, and died 2 years after the onset, the output data for learning to be compared with the value to be output to the [survival rate node, mortality node] of the output layer is [0, 1]. Can be.

However, according to an embodiment of the present invention, when a patient dies, a processing method capable of giving a score by ranking without processing output data for learning as [0, 1] as described above is proposed.

According to an embodiment, generating the t-th section survival rate prediction model SM _t may further include giving a score according to the survival period to the t-th section survival period data. That is, in the present embodiment, the training output data may be [p, 1-p], where a non-zero score value may be assigned to p. According to an embodiment, the score may be given in proportion to the survival period of the patient's t-th interval. Here, the survival period may be divided into at least monthly. For example, the scores according to the survival period for each section of patient D who survived 1 year and 3 months are as shown in [Table 4] below.

N (year) One 2 3 4 5 Score by section One 3/12 0 0 0

Therefore, in this case, when training the artificial neural network that predicts the survival rate after 2 years, the output data for learning to be compared with the value of [survival rate node, mortality node] of the output layer is [3/12, 1-3/12] = [0.75, 0.25].

According to this method, even in the case of data of patients whose follow-up period is less than 5 years due to death or the like (mid-cutting data, right-censored case), the survival rate is not counted as 0, and a ranked score is given as the survival period. As a result, the number of meaningful data used to generate the survival rate prediction model can be increased, resulting in improved accuracy of survival rate prediction.

The artificial neural network may be retrained to predict the survival rate after 2 years of gastric cancer patients by using the learning output data using the learning input data and the score.

As this process was repeated (t=t+1), the survival rate after t year (S _t ) when the clinical data after t year (X _t ) and the predicted survival rate after N-1 year (S _t-1 ) were entered. A survival rate prediction model (SM _t ) after t years, which is an input/output function that can output ), is created by learning an artificial neural network.

According to an embodiment of the present invention, since the survival rate after t year is predicted using the survival rate prediction result (S _t-1 ) after t-1 year reflecting the'prognosis of the patient at the time point after t-1 year', Each time the artificial neural network is trained for each year, the survival rate prediction performance improves.

In an embodiment of the present invention, a survival rate prediction model was generated using the LSTM algorithm. At this time, the survival probability at the discrete time t is determined as in Equation 1 below, and the hazard function may be determined as in Equation 2 below.

<Equation 1>

<Equation 2>

At this time, in the patient group having the input variable X _t and the survival data Y _t , when the parameter vector W _t = θ ₁ X + θ ₂ t is optimized at the first time point t, the LSTM layer remembers W _t . The survival rate model is then retrained to yield the parameter vector W _t+1 for the survival data Y _t+1 at the next time point t+1. For example, if a patient dies from disease 2 years after onset of gastric cancer, the survival rate model learns survival data Y ₁ = 1 in the first year and survival data Y ₂ = 0 in the second year. At this time, since the LSTM memorizes and optimizes the parameter W for each survival data value (Y), the RNN-based survival rate prediction model can calculate the survival rate at a specific time point.

However, since it is difficult to collect all the clinical data over time, the patient's clinical data X cannot be updated at all times. On the other hand, the purpose of the survival rate prediction model is to predict the survival rate in the long term, especially through information when a patient's gastric cancer is detected (ie, information at the first visit to the hospital). Therefore, the present inventors assumed that the clinical data of the patient is constant during the observation time, but it is dependent on the passage of time and there is a latent feature that indicates the patient's condition at a specific time. At this time, the latent factor and the risk function are defined as time-dependent values, such as <Equation 3> and <Equation 4>.

<Equation 3>

<Equation 4>

At this time, the time-dependent survival value is included in the patient's clinical data (X), and the final input data X _t becomes time-dependent data. The time-dependent survival value includes time (t) information and survival rate predicted value (S _t ) information obtained by a gradient descent equation ∂S.

Meanwhile, the gradient descent equation ∂S may be determined by the following <Equation 5> to <Equation 8>.

<Equation 5>

<Equation 6>

<Equation 7>

<Equation 8>

When the generation of the model for predicting the survival rate of a gastric cancer patient is completed through the above-described algorithm, the weight corresponding to the connection of each node of the artificial neural network is learned to be optimized for survival rate prediction. Therefore, the patient's survival rate can be predicted through the value output to the output layer by inputting the input data obtained from the clinical data of an arbitrary gastric cancer patient into the input layer of the artificial neural network. That is, according to the method for predicting the prognosis of gastric cancer using an artificial neural network according to the present invention, the prognosis of a gastric cancer patient can be accurately predicted for each individual.

6 is a diagram schematically showing a part of a heatmap graph of an artificial neural network sequentially trained by year according to an embodiment of the present invention. In an embodiment, as shown in FIG. 6, the present inventors generated a survival rate prediction model for each year by sequentially learning the artificial neural network for each year, and then evaluated the performance of each model.

7 is a graph comparing the AUC values of the ROC graph of the survival rate prediction model for each year during training. During 100 training sessions, the mean of AUC was 0.79±0.052 in the predictive model for survival after 1 year, 0.839±0.045 in the model after 2 years, 0.89±0.049 in the model after 3 years, 0.915±0.05 in the model after 4 years, and after 5 years. It was 0.92±0.049 in the model.

8 is an ROC graph verifying the survival rate prediction model for each year with separate test data. The value of AUC is 0.858 in the survival rate prediction model after 1 year, 0.869 in the model after 2 years, 0.879 in the model after 3 years, 0.912 in the model after 4 years, and 0.923 in the model after 5 years. You can see that the performance of is improved.

9 is a graph comparing the survival rate predicted by the survival rate prediction model and the actual survival rate. As a result of Kaplan-Meier survival analysis, it was confirmed that the survival rate prediction result and the actual survival rate were correlated with a 15% error range (dotted line) within the 95% confidence interval.

10 to 14 are graphs showing a decision curve of a survival rate prediction model after 1, 2, 3, 4 and 5 years, respectively. The AUC simply evaluates the accuracy of the prediction, but the judgment curve reflects the clinical results and calculates and visualizes each net benefit for the threshold probability that is the criterion for clinical judgment. Therefore, it is possible to evaluate the actual clinical value of the predictive model through the judgment curve. Referring to FIGS. 10 to 14, it can be seen that the net benefit exhibits a positive value in all critical probabilities, and in particular, it can be seen that the net benefit increases as the annual increase increases. That is, it can be seen that the survival rate prediction model of the present invention is helpful in clinical judgment.

15 is a graph showing the learning effect of an artificial neural network and a simple artificial neural network for comparison according to an embodiment of the present invention. In FIG. 15, an artificial neural network according to an embodiment of the present invention including an RNN layer is shown as RSN (Recurrent Survival Network), and a simple neural network for comparison is shown as Simple_NN.

FIG. 15A is a graph showing an error (cross_entropy) according to repetitive learning (nb_epoch). In the case of the simple neural network for comparison (Simple_NN), the absolute value of the error is large and the reduction rate is also small, but in the case of the artificial neural network (RSN) according to the present invention, it can be confirmed that the decrease in error due to repetitive learning is large.

FIG. 15B is a graph showing cross validation during each learning. As shown in the graph, in the case of a simple neural network (Simple_NN), even if iterative learning is performed, the decrease in validation loss is small, and the stability of the model is degraded as the deviation is large. On the other hand, in the case of RSN, the verification loss is very low and stable.

FIG. 15C is a ROC (receiver operating characteristic) graph showing the accuracy of survival rate prediction, and FIG. 15D is a graph showing the AUC (area under curve) distribution result of the ROC graph. The accuracy of survival rate prediction can be quantified by the AUC of the area under the ROC graph, and the accuracy is higher as the area is closer to 1. As a result of repeated trials by the present inventors, the RSN showed an average accuracy of 0.95 or higher, and the accuracy of the simple neural network (Simple_NN) was about 0.70. As a result of statistical verification through Mann_Whitney test using MedCalc program, the diagnostic accuracy of RNN was significantly higher than that of simple neural network (Simple_NN) with p<0.001.

<Example 2>

16 is a schematic diagram of a method of retraining an artificial neural network using data from other regions.

When the RSN artificial neural network is retrained using data from another region or hospital with a database of other clinical variables, an artificial neural network optimized for that region data can be newly constructed. At this time, the accuracy of survival prediction when the RSN model is retrained by adding data from other regions to the already learned RSN model, rather than training a new artificial neural network only with data from other regions with insufficient clinical data or different types of clinical data. Becomes higher.

In this study, data obtained from the Gastric Cancer Project Singapore cohort (hereinafter'Singapore data') were used as data from other regions. Singapore data shows that the number of available variables is 12 (molecular subtype, sex, age, pstage, peritoneal cytology, met site, p_node, lauren, pathology type, Lymphovascular invasion, recurrence, 5_FU adjuvant), 19 It had insufficient information to use the RSN model including variables.

The present inventors divided Singapore data into training data and test data, and compared the performance of an artificial neural network model that was retrained by RSN by adding Singapore training data to the existing data and a local artificial neural network model that was trained only with Singapore data.

17 is a heat map graph comparing a retrained model by adding Singapore data to an RSN and a model trained only with Singapore data.

Referring to FIG. 17A, the number of variables learned only with Singapore data is 12, and 29 nodes are generated in the input layer using the classification thereof. In the case of (b) of FIG. 8, a part of the Singapore data was added to the training data of the RSN for training, and 20 variables and 53 nodes were generated to retrain the artificial neural network. At this time, data (missing value, NaN) that is in the original data and not in Singapore data was added through the knn algorithm described above.

18 is a graph comparing the effects of the original model, a retraining model that is retrained by adding Singapore data, and a Singapore model that is trained only with Singapore data.

The original model (original) refers to an artificial neural network tested with existing data using existing data, and the model retrained with RSN (adaptive training set) is retrained through the existing data and Singapore data. It means an artificial neural network tested with data, and the Singapore training set means an artificial neural network that was trained and tested using only Singapore data.

Fig. 18A shows a decrease in errors due to repetitive learning. As shown in the figure, in the case of the re-learning model, it can be seen that the learning speed is fast as the reduction in error due to iterative learning is large. On the other hand, in the case of the Singapore model, the reduction of errors due to iterative learning is small.

18B is a graph showing the error reduction of the crossover test result during each learning. As shown in this graph, in the case of the Singapore model, the reduction of the validation loss is small even in iterative learning, and the stability of the model is low as the deviation is large, but the validation loss is very low and stable in the case of the retraining model.

On the other hand, in (a) and (b) of Figs. 18, since the retraining model is retrained using an artificial neural network that has already been learned, the learning speed is faster than the original model that starts learning from the beginning.

18C is an ROC graph showing the accuracy of prediction. As a result of repeating 100 times by the present inventors, in the case of the retraining model, the survival rate of the test data in Singapore was adjusted with an average probability of 0.95, but the accuracy of the Singapore model was around 0.80. As a result of statistical verification through the Mann Whitney test, p<0.001, the survival rate prediction accuracy of the retraining model was significantly higher than that of the Singapore model. The retraining model showed not much lower performance than the accuracy of the original model. That is, it can be seen that the prediction accuracy is not significantly lowered even if the missing value is added by the knn algorithm.

19 is a diagram schematically showing the configuration of an apparatus for predicting prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention.

The apparatus 10 for predicting the prognosis of gastric cancer shown in FIG. 19 shows only the components related to the present embodiment in order to prevent the features of the present embodiment from being blurred. Accordingly, it can be understood by those of ordinary skill in the art related to the present embodiment that other general-purpose components may be further included in addition to the components shown in FIG. 19.

The apparatus 10 for predicting prognosis of gastric cancer according to an embodiment of the present invention may correspond to at least one or more processors or may include at least one or more processors. Accordingly, the apparatus 10 for predicting prognosis of gastric cancer may be driven in a form included in another hardware device such as a microprocessor or a general-purpose computer system.

The present invention can be represented by functional block configurations and various processing steps. These functional blocks may be implemented with various numbers of hardware or/and software configurations that perform specific functions. For example, the present invention is an integrated circuit configuration such as memory, processing, logic, a look-up table, etc., capable of executing various functions by controlling one or more microprocessors or by other control devices. Can be hired. Similar to how the components of the present invention can be implemented with software programming or software elements, the present invention includes various algorithms implemented with a combination of data structures, processes, routines or other programming constructs, including C, C++ , Java, assembler, etc. may be implemented in a programming or scripting language. Functional aspects can be implemented with an algorithm running on one or more processors. In addition, the present invention may employ conventional techniques for electronic environment setting, signal processing, and/or data processing. Terms such as "mechanism", "element", "means", and "configuration" may be widely used, and components of the present invention are not limited to mechanical and physical configurations. The term may include a meaning of a series of routines of software in connection with a processor or the like.

Referring to FIG. 19, the apparatus 10 for predicting prognosis of gastric cancer includes a data acquisition unit 11, an artificial neural network learning unit 12, and a survival rate prediction model generation unit 13.

The data acquisition unit 11 captures medical data, such as clinical data, of a plurality of gastric cancer patients and acquires survival period data after gastric cancer onset. Clinical data may be acquired from a medical image of a patient or may be acquired from a specimen test result of a patient, but is not limited thereto.

The artificial neural network learning unit 12 acquires input data for learning and output data for learning from clinical data and survival period data of a plurality of gastric cancer patients, and includes an input layer, a hidden layer, and an output layer using the learning input data and the learning output data. Learning artificial neural networks.

The survival rate prediction model generation unit 13 generates a model for predicting the survival rate of gastric cancer patients using the learned artificial neural network. In this case, predicting the survival rate may mean that when clinical information of a gastric cancer patient is input, the survival rate of the patient is calculated through a predetermined algorithm.

According to an embodiment, the input data for learning may include molecular genetic subtype data of a plurality of gastric cancer patients, and the subtype is MSI (microsatellite instable) subtype, MSS/EMT subtype, MSS/TP53+ subtype, MSS Includes the /TP53-subtype. In this case, the input layer may include four nodes to which molecular genetic subtype data is input.

According to an embodiment, the artificial neural network learning unit 12 may calculate an embedding layer by embedding each variable of input data for training into a vector of two or more dimensions.

According to an embodiment, the artificial neural network learning unit 12 may add missing data (NaN) of input data for training using a k-nearest neighbor algorithm (knn).

According to an embodiment, the hidden layer of the artificial neural network may include at least one RNN layer.

Meanwhile, the method for predicting the prognosis of gastric cancer using an artificial neural network according to an embodiment of the present invention shown in FIG. 1 can be written as a program that can be executed on a computer, and the program is operated using a computer-readable recording medium. It can be implemented in a general purpose digital computer. The computer-readable recording medium includes a storage medium such as a magnetic storage medium (eg, ROM, floppy disk, hard disk, etc.), and an optical reading medium (eg, CD-ROM, DVD, etc.).

According to the method, apparatus and program for predicting the prognosis of gastric cancer using an artificial neural network according to the present invention, it is possible to accurately predict the prognosis of a gastric cancer patient for each individual. In addition, since the prognosis of each treatment method can be simulated using the learned artificial neural network, a customized treatment method for each patient can be determined.

The present invention has been described with reference to the embodiments shown in the drawings, but these are merely exemplary, and those of ordinary skill in the art will appreciate that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical scope of the present invention should be determined by the technical spirit of the appended claims.

The present invention relates to a method, apparatus, and program for predicting prognosis of gastric cancer using an artificial neural network, and can be used in the diagnosis and treatment apparatus industry.

Claims (13)

delete delete delete delete delete delete A data acquisition unit for acquiring clinical data of a plurality of gastric cancer patients and survival period data after onset of gastric cancer;
An artificial neural network learning unit that acquires input data for learning and output data for training from the clinical data and the survival period data, and learns an artificial neural network including an input layer, a hidden layer, and an output layer by using the training input data and the training output data. ; And
Including; a survival rate prediction model generating unit for generating a model for predicting the survival rate of gastric cancer patients using the learned artificial neural network,
The survival rate prediction model generation unit,
The t-th section survival rate prediction model (SM _t ) is generated using the clinical data and the t-th section survival period data of a plurality of gastric cancer patients, and the t-th section survival rate prediction data (S _t ) obtained from the t-th section survival rate prediction model And generating a t+1 th section survival rate prediction model (SM _t +1) using the t+1 th section survival period data of the plurality of gastric cancer patients,
When generating the t-th section survival rate prediction model (SM _t ), the t-th section survival period data is given a score according to the survival period, an apparatus for predicting prognosis of gastric cancer using an artificial neural network. The method of claim 7,
The learning input data includes molecular genetic subtype data of the plurality of gastric cancer patients, and the subtype is an artificial microsatellite instable (MSI) subtype, MSS/EMT subtype, MSS/TP53+ subtype, and MSS/TP53- subtype. A device for predicting the prognosis of gastric cancer using a neural network. The method of claim 8,
The input layer includes four nodes to which the molecular genetic subtype data is input, the apparatus for predicting prognosis of gastric cancer using an artificial neural network. The method of claim 7,
The artificial neural network learning unit calculates an embedding layer by embedding each variable of the training input data into a vector of two or more dimensions. An apparatus for predicting prognosis of gastric cancer using an artificial neural network. The method of claim 7,
The artificial neural network learning unit adds missing data (NaN) of the training input data using a k-nearest neighbor algorithm (knn), a prognosis apparatus for gastric cancer using an artificial neural network. The method of claim 7,
The hidden layer of the artificial neural network includes at least one cyclic neural network (RNN) layer. An apparatus for predicting prognosis of gastric cancer using an artificial neural network. delete

Images