Crohn disease prediction method based on deep learning

Abstract

The invention relates to the technical field of deep learning, in particular to a Crohn disease prediction method based on deep learning, which comprises the following steps: collecting intestinal flora distribution data of a Crohn disease patient and a healthy control group, and obtaining flora composition information of different parts of an intestinal tract; the age of the Crohn disease patient is also collected; feature extraction is carried out on the spatial distribution data of the intestinal flora, the intestinal tract is divided into M voxels, each voxel contains N bacteria, the spatial distribution data of the intestinal flora is expressed as an M * N matrix X, and sliding is carried out at different positions of the intestinal tract space through a convolution kernel; constructing a short-term long memory network to process the fused feature vectors to form a prediction model, adding a full connection layer to an output layer of the prediction model, and mapping the output to a probability value; according to the method, the spatial distribution characteristics of the intestinal flora are deeply excavated, and the relationship between the spatial distribution characteristics of the intestinal flora and the age and the Crohn's disease is found in combination with the age, so that a more reliable prediction basis is provided for doctors.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a Crohn's disease prediction method based on deep learning. Background Art

[0002] Currently, most methods for predicting Crohn's disease only make preliminary judgments based on the patient's clinical symptoms and routine blood test indicators. These indicators often only show obvious abnormalities after the disease has developed to a certain extent, and their sensitivity and specificity for early prediction are insufficient. Moreover, the potential association between the spatial distribution of intestinal flora, a key factor, and Crohn's disease has not been fully explored, and the demand for early and accurate prediction of the disease cannot be met. Therefore, the present invention proposes a deep learning-based Crohn's disease prediction method that explores the potential association between the spatial distribution of intestinal flora, a key factor, and Crohn's disease. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a Crohn's disease prediction method based on deep learning that mines the potential association between the spatial distribution of intestinal flora, a key factor, and Crohn's disease.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: A deep learning-based Crohn's disease prediction method, comprising: Data collection: Intestinal microbial distribution data were collected from Crohn's disease patients and healthy controls to obtain information on the composition of microbiota in different parts of the intestine, including the species, abundance, and spatial location of the microbiota. The age of Crohn's disease patients was also collected. The distribution data were normalized. Feature extraction and fusion: Convolutional neural networks are used to extract features from the spatial distribution data of intestinal flora. The intestine is divided into M voxels, each containing N types of bacteria. The spatial distribution data of intestinal flora is represented as an M×N matrix X. The convolution kernel slides at different positions in the intestinal space to extract local spatial features and convert them into planar features. The formula for the convolution operation is: =σ( × +b); in, The features extracted at the plane coordinate position (x, y) are represents the element at row i+k and column j+l in the input matrix X, is the weight matrix of the convolution kernel, b is the bias term, K and L are the sizes of the convolution kernel in the row and column directions, k and l are the sliding position indexes of the convolution kernel in the row and column directions, respectively; σ is the activation function; the patient's age is represented as a vector Y, and the feature map obtained by the convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z; Z and Y are fused to form a feature vector =[Z;Y]; Prediction model construction, constructing short-term long memory network to process the fused feature vector Form a prediction model and add a fully connected layer to the output layer of the prediction model to map the output to a probability value.

[0005] Preferably, a short-term long memory network is constructed to process the fused feature vector Forming a prediction model further includes: Forget Gate: =σ( ×[ ]+ ); Input Gate: =σ( ×[ ]+ ); Candidate memory cells: =tanh( ×[ ]+ ); Memory unit update: = × + × ; Output Gate: =σ( ×[ ]+ ); Hidden state update: = ×tanh( ); in, 、 、 、 is the corresponding weight matrix, 、 、 、 is the bias term; [ ] means to change the hidden state of the previous moment and the fused feature vector at the current moment Stitched together.

[0006] Preferably, model optimization is also included: a binary cross entropy loss function is used to measure the difference between the model prediction result and the true label, and the formula is: Loss=- Where N is the number of samples; is the true label of the nth sample, 1 indicates Crohn's disease, and 0 indicates no disease; is the probability that the nth sample suffers from Crohn's disease as predicted by the model.

[0007] Preferably, the parameters of the prediction model are updated using the Adam optimization algorithm according to the result of the loss function.

[0008] Preferably, the formula of the fully connected layer is: =σ( × + ); Where p is the predicted probability that the subject suffers from Crohn's disease, is the weight matrix of the fully connected layer, is the bias term, T is the hidden state of the last moment of LSTM, and σ is the sigmoid function, which compresses the output value to between [0, 1].

[0009] Preferably, the collected data is cleaned to remove outliers and missing values, and then the spatial distribution data of the intestinal flora are normalized so that its value is between [0,1].

[0010] Preferably, σ in the formula of the convolution operation is a ReLU function.

[0011] Preferably, the patient's age is represented as a vector Y with a length of P, and the feature map obtained by convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z with a length of Q; Z and Y are directly spliced and fused to form a feature vector =[Z;Y], and the length of the fused feature vector Q+P is obtained by adding them together.

[0012] Preferably, the patient's age is represented as a vector Y, and the feature map obtained by convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z; the Z and Y features are weighted and fused to form a feature vector =[Z;Y]= ×Z+ ×Y; where is the spatial feature weight of intestinal flora, is the age weight.

[0013] Preferably, + =1.

[0014] The beneficial effects of the present invention are: by deeply exploring the spatial distribution characteristics of the intestinal flora and combining them with age, this is because Crohn's disease is highly prevalent in people aged 15-40 years old. Therefore, as another important factor, the purpose is to be able to find the relationship between the spatial distribution characteristics of the intestinal flora and age and Crohn's disease. For example, the presence of a certain flora on a certain voxel is highly correlated with Crohn's disease. Therefore, in subsequent examinations, long-term attention can be paid to the changes in a certain position and a certain flora corresponding to a certain voxel in the population, that is, early prediction of Crohn's disease, improving the accuracy and advanceness of the prediction, and having advantages in early detection compared to traditional methods based only on clinical symptoms and routine blood indicators; and by sliding the convolution kernel at different positions in the intestinal space, extracting local spatial features and then converting them into planar features, the complex spatial features of the intestine can be simplified into planar features, which is consistent with the characteristics of the intestinal surface; furthermore, considering the distribution patterns of intestinal flora at different spatial positions and their dynamic changes over time, deep learning can capture more complex and implicit disease-related information, providing clinicians with a more reliable prediction basis, thereby achieving prediction and prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The figure is a flowchart of a Crohn's disease prediction method based on deep learning according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0016] To illustrate the technical content, achieved objectives and effects of the present invention in detail, the following description is given in conjunction with the embodiments and accompanying drawings.

[0017] Please refer to Figure 1 , a deep learning-based Crohn's disease prediction method, including: Data collection: Intestinal microbial distribution data were collected from Crohn's disease patients and healthy controls to obtain information on the composition of microbiota in different parts of the intestine, including the species, abundance, and spatial location of the microbiota. The age of Crohn's disease patients was also collected. The distribution data were normalized. Feature extraction and fusion: Convolutional neural networks are used to extract features from the spatial distribution data of intestinal flora. The intestine is divided into M voxels, each containing N types of bacteria. The spatial distribution data of intestinal flora is represented as an M×N matrix X. The convolution kernel slides at different positions in the intestinal space to extract local spatial features and convert them into planar features. The formula for the convolution operation is: =σ( × +b); in, The features extracted at the plane coordinate position (x, y) are represents the element at row i+k and column j+l in the input matrix X, is the weight matrix of the convolution kernel, b is the bias term, K and L are the sizes of the convolution kernel in the row and column directions, k and l are the sliding position indexes of the convolution kernel in the row and column directions, respectively; σ is the activation function; the patient's age is represented as a vector Y, and the feature map obtained by the convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z; Z and Y are fused to form a feature vector =[Z;Y]; Prediction model construction, constructing short-term long memory network to process the fused feature vector Form a prediction model and add a fully connected layer to the output layer of the prediction model to map the output to a probability value.

[0018] As can be seen from the above description, by deeply exploring the spatial distribution characteristics of the intestinal flora and combining them with age, this is another important factor. Since Crohn's disease is highly prevalent in people aged 15-40, the goal is to be able to identify the relationship between the spatial distribution characteristics of the intestinal flora and age and Crohn's disease. For example, the presence of a certain bacterial species in a certain voxel is highly correlated with Crohn's disease. Therefore, in subsequent examinations, long-term attention can be paid to the changes in a certain bacterial species at a certain location corresponding to a certain voxel in the population, that is, early prediction of Crohn's disease, improving the accuracy and advanceness of predictions. Compared with traditional methods based solely on clinical symptoms and routine blood indicators, this has advantages in early detection. In addition, by sliding the convolution kernel across different locations in the intestinal space, extracting local spatial features and then converting them into planar features, the complex spatial characteristics of the intestine can be simplified into planar features, which is consistent with the characteristics of the intestinal surface. Furthermore, considering the distribution patterns of the intestinal flora at different spatial locations and their dynamic changes over time, deep learning can capture more complex and implicit disease-related information, providing clinicians with more reliable predictive basis, thereby achieving prediction and prevention.

[0019] Furthermore, a short-term long memory network is constructed to process the fused feature vector Forming a prediction model further includes: Forget Gate: =σ( ×[ ]+ ); Input Gate: =σ( ×[ ]+ ); Candidate memory cells: =tanh( ×[ ]+ ); Memory unit update: = × + × ; Output Gate: =σ( ×[ ]+ ); Hidden state update: = ×tanh( ); in, 、 、 、 is the corresponding weight matrix, 、 、 、 is the bias term; [ ] means to change the hidden state of the previous moment and the fused feature vector at the current moment Stitched together.

[0020] From the above description, we can see that the forget gate determines which previous information is discarded, the input gate determines which new information is updated to the memory unit, and the output gate determines how to output the hidden state at the current moment based on the information in the memory unit, thereby realizing dynamic processing of the fused feature sequence and exploring the potential relationship between the spatial distribution of intestinal flora and other factors changing over time and the risk of Crohn's disease. That is, the time factor is supplemented. The time factor here is different from the age factor. The age factor refers to the high-incidence interval, while the time factor here mainly refers to the worsening of Crohn's disease over time. It is important to distinguish them and not to confuse them.

[0021] Furthermore, it also includes model optimization: the binary cross entropy loss function is used to measure the difference between the model prediction results and the true label. The formula is: Loss=- Where N is the number of samples; is the true label of the nth sample, 1 indicates Crohn's disease, and 0 indicates no disease; is the probability that the nth sample suffers from Crohn's disease as predicted by the model.

[0022] From the above description, we can see that through the loss function, the model prediction probability and the true label are compared, guiding the model to continuously adjust the parameters to make the prediction results as close to the actual situation as possible and reduce the prediction error.

[0023] Furthermore, the Adam optimization algorithm is used to update the parameters of the prediction model according to the results of the loss function.

[0024] From the above description, we can see that the Adam optimization algorithm, combined with the advantages of momentum and RMSprop, can adaptively adjust the learning rate, allowing the model to converge to a better solution faster during training.

[0025] Furthermore, the formula of the fully connected layer is: =σ( × + ); Where p is the predicted probability that the subject suffers from Crohn's disease, is the weight matrix of the fully connected layer, is the bias term, T is the hidden state of the last moment of LSTM, and σ is the sigmoid function, which compresses the output value to between [0, 1].

[0026] Furthermore, the collected data were cleaned to remove outliers and missing values, and then the spatial distribution data of intestinal flora were normalized so that its value was between [0,1].

[0027] Furthermore, σ in the formula of the convolution operation is the ReLU function.

[0028] From the above description, we can see that the ReLU function is used to introduce nonlinearity, so that the model can learn more complex feature patterns.

[0029] Furthermore, the patient's age is represented as a vector Y with a length of P, and the feature map obtained after the convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z with a length of Q; Z and Y are directly spliced and fused to form a feature vector =[Z;Y], and the length of the fused feature vector Q+P is obtained by adding them together.

[0030] As can be seen from the above description, Q+P represents the fused feature vector that encompasses all features of the original gut microbiome spatial features and other patient information. This completeness ensures that the model can simultaneously utilize two different types of information for learning and prediction, avoiding degradation in predictive performance due to information loss.

[0031] Furthermore, the patient's age is represented as a vector Y, and the feature map obtained after the convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z; the Z and Y features are weighted and fused to form a feature vector =[Z;Y]= ×Z+ ×Y; where is the spatial feature weight of intestinal flora, is the age weight.

[0032] Further, + =1.

[0033] Example 1 A deep learning-based Crohn's disease prediction method, comprising: Data collection: Intestinal flora distribution data of Crohn's disease patients and healthy controls were collected to obtain information on the flora composition of different parts of the intestine, including the type, abundance and spatial location data of the flora; the age of Crohn's disease patients was also collected; the collected data were cleaned to remove outliers and missing values, and then the spatial distribution data of the intestinal flora were normalized so that its value was between [0,1].

[0034] Feature extraction and fusion: Convolutional neural networks are used to extract features from the spatial distribution data of intestinal flora. The intestine is divided into M voxels, each containing N types of bacteria. The spatial distribution data of intestinal flora is represented as an M×N matrix X. The convolution kernel slides at different positions in the intestinal space to extract local spatial features and convert them into planar features. The formula for the convolution operation is: =σ( × +b); in, The features extracted at the plane coordinate position (x, y) are Represents the element at the i+kth row and j+lth column in the input matrix X. This element is the value of a specific position in the input matrix covered by the convolution kernel during the sliding process. is the weight matrix of the convolution kernel, b is the bias term, K and L are the sizes of the convolution kernel in the row and column directions, respectively, k and l represent the sliding position index of the convolution kernel in the row and column directions, respectively; σ is the activation function (ReLU function); the patient's age is represented as a vector Y with a length of P, and the feature map obtained after the convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z with a length of Q; Z and Y are directly spliced and fused to form a feature vector =[Z;Y], and the length of the fused feature vector Q+P is obtained by adding them together.

[0035] Prediction model construction, constructing short-term long memory network to process the fused feature vector A prediction model is formed, which includes: Forget Gate: =σ( ×[ ]+ ); Input Gate: =σ( ×[ ]+ ); Candidate memory cells: =tanh( ×[ ]+ ); Memory unit update: = × + × ; Output Gate: =σ( ×[ ]+ ); Hidden state update: = ×tanh( ); in, 、 、 、 is the corresponding weight matrix, 、 、 、 is the bias term; [ ] means to change the hidden state of the previous moment and the fused feature vector at the current moment Stitched together.

[0036] Add a fully connected layer to the output layer of the prediction model to map the output to a probability value.

[0037] The formula for the fully connected layer is: =σ( × + ); Where p is the predicted probability that the subject suffers from Crohn's disease, is the weight matrix of the fully connected layer, is the bias term, T is the hidden state of the last moment of LSTM, σ is the sigmoid function (to be distinguished from the ReLU function above, the sigmoid function transmits probability), which compresses the output value to between [0,1].

[0038] Model optimization: The binary cross entropy loss function is used to measure the difference between the model prediction results and the true label. The formula is: Loss=- Where N is the number of samples; is the true label of the nth sample, 1 indicates Crohn's disease, and 0 indicates no disease; is the probability that the nth sample suffers from Crohn's disease as predicted by the model.

[0039] The Adam optimization algorithm is used to update the parameters of the prediction model according to the results of the loss function.

[0040] Example 2 A Crohn's disease prediction method based on deep learning is described below. The differences from Example 1 are as follows: The patient's age is represented as a vector Y, and the feature map obtained by convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z; the Z and Y features are weighted and fused to form a feature vector =[Z;Y]= ×Z+ ×Y; where is the spatial feature weight of intestinal flora, is the age weight.

[0041] + =1.

[0042] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's description and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A Crohn's disease prediction method based on deep learning, characterized in that: include: Data collection: Intestinal microbial distribution data were collected from Crohn's disease patients and healthy controls to obtain information on the composition of microbiota in different parts of the intestine, including the species, abundance, and spatial location of the microbiota. The age of Crohn's disease patients was also collected. The distribution data were normalized. Feature extraction and fusion: Convolutional neural networks are used to extract features from the spatial distribution data of intestinal flora. The intestine is divided into M voxels, each containing N types of bacteria. The spatial distribution data of intestinal flora is represented as an M×N matrix X. The convolution kernel slides at different positions in the intestinal space to extract local spatial features and convert them into planar features. The formula for the convolution operation is: =σ( × +b); in, The features extracted at the plane coordinate position (x, y) are represents the element at row i+k and column j+l in the input matrix X, is the weight matrix of the convolution kernel, b is the bias term, K and L are the sizes of the convolution kernel in the row and column directions, k and l are the sliding position indexes of the convolution kernel in the row and column directions, respectively; σ is the activation function; the patient's age is represented as a vector Y, and the feature map obtained by the convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z; Z and Y are fused to form a feature vector =[Z;Y]; Prediction model construction, constructing short-term long memory network to process the fused feature vector Form a prediction model and add a fully connected layer to the output layer of the prediction model to map the output to a probability value.

2. The Crohn's disease prediction method based on deep learning according to claim 1, characterized in that: Construct a short-term long memory network to process the fused feature vector Forming a prediction model further includes: Forget Gate: =σ( ×[ ]+ ); Input Gate: =σ( ×[ ]+ ); Candidate memory cells: =tanh( ×[ ]+ ); Memory unit update: = × + × ; Output Gate: =σ( ×[ ]+ ); Hidden state update: = ×tanh( ); in, 、 、 、 is the corresponding weight matrix, 、 、 、 is the bias term; [ ] means to change the hidden state of the previous moment and the fused feature vector at the current moment Stitched together.

3. The Crohn's disease prediction method based on deep learning according to claim 2, characterized in that: Model optimization is also included: a binary cross entropy loss function is used to measure the difference between the model prediction results and the true label. The formula is: Loss=- Where N is the number of samples; is the true label of the nth sample, 1 indicates Crohn's disease, and 0 indicates no disease; is the probability that the nth sample suffers from Crohn's disease as predicted by the model.

4. The Crohn's disease prediction method based on deep learning according to claim 3, characterized in that: The Adam optimization algorithm is used to update the parameters of the prediction model according to the results of the loss function.

5. The Crohn's disease prediction method based on deep learning according to claim 1, characterized in that: The formula for the fully connected layer is: =σ( × + ); Where p is the predicted probability that the subject suffers from Crohn's disease, is the weight matrix of the fully connected layer, is the bias term, T is the hidden state of the last moment of LSTM, and σ is the sigmoid function, which compresses the output value to between [0, 1].

6. The Crohn's disease prediction method based on deep learning according to claim 1, characterized in that: The collected data were cleaned to remove outliers and missing values, and the spatial distribution data of intestinal flora were normalized so that their values were between [0,1].

7. The Crohn's disease prediction method based on deep learning according to claim 1, characterized in that: The σ in the formula of the convolution operation is the ReLU function.

8. The Crohn's disease prediction method based on deep learning according to claim 1, characterized in that: The patient's age is represented as a vector Y with a length of P. The feature map obtained by convolution of the intestinal flora spatial feature matrix is flattened into a vector Z with a length of Q. Z and Y are directly concatenated to form a feature vector =[Z;Y], and the length of the fused feature vector Q+P is obtained by adding them together.

9. The Crohn's disease prediction method based on deep learning according to claim 1, characterized in that: The patient's age is represented as a vector Y, and the feature map obtained by convolution operation of the intestinal flora spatial feature matrix is flattened into a vector Z; the Z and Y features are weighted and fused to form a feature vector =[Z;Y]= ×Z+ ×Y; where is the spatial feature weight of intestinal flora, is the age weight.

10. The Crohn's disease prediction method based on deep learning according to claim 9, characterized in that: + =1。

Images