Multi-dimensional data fusion processing method for medical text data

Abstract

The invention belongs to the technical field of deep learning, and particularly relates to a multi-dimensional data fusion processing method for medical text data, which comprises the following steps: acquiring multi-dimensional data, and inputting the multi-dimensional data into a trained fusion model to obtain a multi-dimensional data fusion result; the step of training the fusion model comprises the sub-steps of obtaining a multi-dimensional data set, and dividing the multi-dimensional data set into a training set and a test set by adopting a data splitting technology; performing feature extraction on the data in the training set, performing tensor conversion on the extracted features, and reconstructing the features after tensor conversion; inputting the reconstructed feature data into the fusion model for training, adjusting model parameters, and finishing model training when a model loss function is minimum; in a model definition stage, a DNN deep learning network architecture is used to ensure that a model can effectively learn and predict a target variable; through the constructed multi-dimensional data fusion processing method, the key information is directly extracted from the original medical text obtained in multiple inspection modes, and the working efficiency of the user is improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically relates to a multidimensional data fusion processing method for medical text data. Background Art

[0002] Data processing and analysis based on deep learning is a cutting-edge and rapidly developing direction in the field of data science today. As a branch of machine learning, deep learning has shown significant advantages in the processing and analysis of massive, high-dimensional and complex data with its powerful feature learning and representation capabilities. Using deep learning models and technologies to efficiently extract, clean, transform, model and analyze various types of data, thereby tapping the potential value in the data and providing strong support for applications such as decision support, predictive analysis, and pattern recognition. In terms of data processing, deep learning technology can automatically learn the complex features of data without the need for tedious manual feature engineering. For example, by processing image data through convolutional neural networks (CNN), features such as edges and textures in the image can be automatically learned; while recurrent neural networks (RNN) and long short-term memory networks (LSTM) can process time series data such as stock quotes and voice signals. In terms of data analysis, deep learning models can construct complex nonlinear relationships to more accurately capture the intrinsic connections between data. For example, autoencoders can be used for dimensionality reduction and denoising of data; generative adversarial networks (GANs) can generate new and realistic data samples for data enhancement or simulation experiments. Data processing and analysis based on deep learning are also widely used in natural language processing (NLP), recommendation systems, bioinformatics and other fields. For example, in NLP, deep learning models can realize tasks such as sentiment analysis, topic classification, and question-answering systems for text data; in recommendation systems, deep learning can build complex relationships between users and items to achieve personalized recommendations; in bioinformatics, deep learning can analyze data such as gene sequences and protein structures, providing strong support for disease prediction, drug development, etc.

[0003] In terms of clinical data, researchers have collected and analyzed biological indicators such as height, weight, bone age, sex hormone levels, and lifestyle data such as living habits and eating habits of children with CPP, combined with modern statistics and machine learning techniques, to explore the pathogenesis and influencing factors of precocious puberty. For example, some studies have found that the occurrence of CPP is related to factors such as high-protein diet, frequent consumption of nourishing foods, and family disharmony. In terms of family environment data, information such as family relationships, parents' education level, and family economic status is collected through questionnaires and family interviews to analyze the impact of these factors on the occurrence of CPP. Studies have shown that factors such as family disharmony and lack of parental companionship are closely related to the occurrence of CPP. In terms of social environment data, researchers are concerned about the impact of external factors such as environmental pollution and media influence on CPP. For example, some studies have found that environmental endocrine disruptors (EEDs) such as 4-nonylphenol (4-NP) and dichlorodiphenylethylene (P,P'-DDE) are related to the occurrence of CPP.

[0004] The current diagnosis of PP includes physical signs, basal hormone testing, wrist bone age, gonadal ultrasound, pituitary MRI, and GnRH stimulation test. Baseline hormone testing and gonadal ultrasound are used to initially evaluate sex hormone fluctuations and gonadal development. Pituitary MRI can exclude PP caused by pituitary or intracranial organic lesions. However, the reference standard for differentiating ICPP from PPP is the GnRH stimulation test, which is a time-consuming and invasive test. Repeated blood sampling is painful for children and may also lead to a lack of cooperation from their families. The cumbersome diagnostic procedures reduce the enthusiasm of parents to seek medical treatment for their children, resulting in delayed or absent treatment for ICPP patients.

[0005] Hepatocellular carcinoma (HCC) is the leading primary liver malignancy with a poor clinical prognosis. Microvascular invasion (MVI) is defined as the presence of cancer cell clusters within the intervascular spaces lined by endothelial cells and is visible under a microscope. However, preoperative prediction of MVI is challenging because MVI is a histopathological finding that can only be diagnosed using postoperative surgical specimens. Radiomics has the potential to transform digital medical images into numerous quantitative features that reveal pathophysiology. Radiomics models can accurately predict MVI based on CT images, therefore, constructing a radiomics-based deep learning prediction model may have the potential to predict MVI. Summary of the invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes a multidimensional data fusion processing method for medical text data, the method comprising: obtaining multidimensional data, inputting the multidimensional data into a trained fusion model, and obtaining a multidimensional data fusion result; training the fusion model comprises: obtaining a multidimensional data set, and using data splitting technology to divide the multidimensional data set into a training set and a test set; extracting features from the data in the training set, performing tensor conversion on the extracted features, and reconstructing the features after tensor conversion; inputting the reconstructed feature data into the fusion model for training, adjusting the model parameters, and completing the model training when the model loss function is minimized.

[0007] Beneficial effects of the present invention:

[0008] The present invention adopts a scientific method to comprehensively evaluate the training effect of the model. This includes visually displaying the change of the loss value (Loss) over time during the training process to evaluate the convergence and stability of the model. For classification tasks, evaluation indicators such as accuracy are used to more comprehensively demonstrate the model's predictive ability and performance on the problem of precocious puberty. In the model definition stage, the present invention uses the network architecture of DNN deep learning, selects appropriate network layers, activation functions, loss functions and optimizers, etc., to ensure that the model can effectively learn and predict the target variables. The model training stage is based on the reconstructed data, and through iterative training, the parameters are continuously updated according to the data and labels of the training set to minimize the prediction error and optimize the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is the overall flow chart of the present invention;

[0010] Figure 2 It is the overall framework structure diagram of the present invention;

[0011] Figure 3 It is a structural diagram of the DNN neural network model of the present invention;

[0012] Figure 4 It is the detection result diagram of the present invention. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] A multidimensional data fusion processing method for medical text data, such as Figure 1As shown, the method includes: obtaining multidimensional data, inputting the multidimensional data into the trained fusion model to obtain the multidimensional data fusion result; training the fusion model includes: obtaining a multidimensional data set, using data splitting technology to divide the multidimensional data set into a training set and a test set; extracting features from the data in the training set, performing tensor conversion on the extracted features, and reconstructing the features after tensor conversion; inputting the reconstructed feature data into the fusion model for training, adjusting the model parameters, and completing the model training when the model loss function is minimized. The specific implementation framework of the model is shown in Figure 2 shown.

[0015] In this embodiment, the data import phase is to systematically introduce the original data from local or online resources into the data processing and analysis environment, laying a solid foundation for subsequent data processing and analysis. Then, in order to simulate the model evaluation scenario in actual applications, following the standard practice of machine learning, the data splitting technology is used to carefully divide the data set into a training set and a validation set to ensure that the model still has good generalization ability on unknown data.

[0016] Then, the Tensor conversion is performed, which converts the data into a tensor format to support matrix operations and gradient updates. Next, the data reconstruction step finely divides the data into batches to support batch training of the model, thereby improving training efficiency and model stability.

[0017] In the model definition phase, the DNN deep learning network architecture is used to select appropriate network layers, activation functions, loss functions, and optimizers to ensure that the model can effectively learn and predict the target variables. The model training phase is based on the reconstructed data. Through iterative training, parameters are continuously updated according to the data and labels of the training set to minimize prediction errors and optimize model performance.

[0018] In the result presentation phase, a scientific method is used to comprehensively evaluate the training effect of the model. This includes visually displaying the change of the loss value (Loss) over time during the training process to evaluate the convergence and stability of the model. For classification tasks, evaluation indicators such as accuracy are used to more comprehensively demonstrate the model's predictive ability and performance on the problem of precocious puberty.

[0019] In this embodiment, the data consists of four parts: basic clinical data, basic biochemical data, ultrasound examination and left hand X-ray examination. Basic clinical data included 11 characteristics, namely age, height, weight, BMI, abnormal time of puberty (medical history), age of menarche, breast core, pigmentation, pubic development stage and left / right breast development stage; basic biochemical data included 9 characteristics, namely basal serum LH, basal serum FSH, LH / FSH, estradiol E2, growth hormone, testosterone, prolactin, IGF-1, IGF-binding protein-3; ultrasound examination included 19 characteristics, namely uterine length / width / thickness, endometrial thickness, left / right largest follicle diameter, number of left / right follicles larger than 4 mm, left / right ovarian length / width / thickness, left / right breast gland width / thickness and Tanner stage; left hand X-ray examination included 6 characteristics, including the number of ossified nuclei, bone age, the difference between biological age and bone age, whether sesamoid bones are visible, whether olecranon epiphysis is visible and whether trochlear epiphysis is visible.

[0020] Taking the problem of precocious puberty as an example, a total of 704 original case data were obtained in the study of precocious puberty. In order to ensure the quality and reliability of the data, the regression inclusion criteria of precocious puberty data were strictly followed when selecting the data, and the original data were screened and cleaned in detail. After carefully eliminating useless and missing data, 104 high-quality data were finally retained for subsequent data set construction. The data composition feature selection is shown in Table 1.

[0021] Table 1 High quality data

[0022]

[0023] In order to ensure that different types of data can minimize the deviation of correlation calculation when data is input into the model, comparative experiments of data with different components will be conducted. These experiments will focus on three key areas: examining input based only on basic biochemical data indicators; analyzing the input effect of non-basic biochemical data indicators; and inputting all indicators in combination to fully evaluate their impact on correlation calculation. Through these comparative experiments, the impact of data input on correlation analysis can be accurately grasped, thereby optimizing the accuracy of the model.

[0024] In this embodiment, in the study of hepatocellular carcinoma, a total of 213 original case data were obtained. In order to ensure the quality and reliability of the data, the regression inclusion criteria of precocious puberty data were strictly followed when selecting the data, and the original data were screened and cleaned in detail. After carefully eliminating useless and missing data, 190 high-quality data were finally retained for subsequent data set construction. The data composition includes age, sex, hepatitis B virus infection, Barcelona Clinic liver cancer stage, cirrhosis, Child-Pugh classification, AFP level (nanograms per milliliter), alkaline phosphatase level (unit per liter) and aspartate aminotransferase level (unit per liter).

[0025] The above data (104 cases of precocious puberty or 190 cases of hepatocellular carcinoma) were input into the gradient boosting tree (GBDT) classification for feature correlation extraction. GBDT (Gradient Boosting Decision Tree), whose full name is gradient boosting decision tree, is an iterative decision tree algorithm that constructs a set of weak learners (trees) and accumulates the results of multiple decision trees as the final prediction output. The sum of the results of all weak classifiers equals the predicted value. Each time, the current prediction is used as the benchmark, and the next weak classifier is used to fit the residual (the error between the predicted value and the true value). The weak classifier of GBDT uses a tree model.

[0026] The GBDT classification algorithm is conceptually no different from the GBDT regression algorithm. However, since the sample output is not a continuous value but a discrete category, it is impossible to directly fit the category output error from the output category.

[0027] In order to solve this problem, there are two main methods. One is to use an exponential loss function, in which case GBDT degenerates into an Adaboost algorithm. Another method is to use a method similar to the log-likelihood loss function of logistic regression. In other words, this application uses the difference between the predicted probability value and the true probability value of the category to fit the loss. As for the log-likelihood loss function, there is a difference between binary classification and multivariate classification. For this embodiment, a binary GBDT classification algorithm is used.

[0028] The loss function is:

[0029] L(y,f(x))=log(1+exp(-yf(x)))

[0030] Among them, y∈{-1,+1}. Then the negative gradient error at this time is

[0031]

[0032] For the generated decision tree, the optimal negative gradient fitting value for each leaf node is

[0033]

[0034] Feature extraction of data includes: using the GBDT algorithm to process the data in the data set, iteratively calculating the data by constructing multiple decision trees, and accumulating all iterative results to obtain the final result; extracting the features and target variables of the final result.

[0035] Specifically, the features obtained by the GBDT algorithm are split into data, and the train_test_split function in the scikit-learn library is used to split the dataset into a training set and a test set, where the training set accounts for 70% of the data and the test set accounts for the remaining 30%. Then, features (i.e., all columns except the 'label' column) and target variables ('label' column) are extracted from the two datasets respectively. These features and target variables are converted to NumPy arrays, and then the target variables are converted to two-dimensional column vectors. Finally, the features of the training set (train_x), the targets of the training set (train_y), the features of the test set (test_x), and the targets of the test set (test_y) are generated, and the generated data is directly used for the training and evaluation of the machine learning model. Then, tensor conversion is performed, and the NumPy arrays are converted to PyTorch tensors using tensor conversion, and the data types of these tensors are ensured to be floating point numbers (float).

[0036] The tensor conversion expression for the feature map is as follows:

[0037] T = torch.from_numpy(A)

[0038] Here, torch from_numpy is a function in the PyTorch library, A is the original Numpy input, and T is the output Pytorch tensor.

[0039] In this embodiment, reconstructing the features after tensor conversion includes: taking the features and target data as data set objects, creating a data loader, extracting the data set objects in batches, and shuffling the order of the extracted data to obtain reconstructed features; wherein the data loader is created as:

[0040] texttrain=Dataloader(Dataset.train,batch_size,shuffle=True)

[0041] texttest=Dataloader(Dataset.test,batch_size×2,shuffle=False)

[0042] Among them, texttrain is the training set, Dataloader is the data loader function, Dataset.train is part of the training set, batch_size is the batch size of the training set, shuffle is the data order parameter, True means disrupting the data order, texttest is the test set, Dataset.test is part of the test set, and False means not disrupting the data order.

[0043] Specifically, in the data reconstruction step, TensorDataset is used to encapsulate the feature data train_x and the target data train_y (and the corresponding test set data test_x and test_y) into a dataset object (Datasetobjects), binding the features and the corresponding target values ​​together so that they can be obtained at the same time during iteration. Then create data loaders (DataLoaders), use the loaders to extract data from the dataset in batches, and support shuffling the data order (shuffle=True). At the same time, rely on randomization to reduce model overfitting. In the data loader train_dl of the training set, the batch_size parameter specifies the number of data samples contained in each batch. By setting shuffle=True, the data will be re-shuffled at the beginning of each training cycle (epoch). In the data loader test_dl of the test set, batch_size is set to twice the training set (batch*2). Increasing the batch size of the test set can help speed up the testing process, especially when memory allows. Finally, train_dl and test_dl can be used directly in PyTorch’s model training and evaluation loops to load data in batches and pass to the model.

[0044] In this embodiment, the fusion model is used to process the reconstructed features, including: obtaining target data, calculating the correlation between the reconstructed features and the target data using a gradient boosting tree, and converting the reconstructed features into a PyTorch tensor according to the calculation results; reconstructing the PyTorch tensor and the target data into a data set; and inputting the data in the data set into a DNN neural network to obtain a prediction result.

[0045] The fusion model includes determining the initial model and subsequent decision trees by minimizing the loss function, updating the model by the learning rate, and obtaining a strong learner composed of the weighted sum of all base learners; the fusion model is:

[0046]

[0047] Among them, F0(x) is the initial model, ρ m is the learning rate, h m(x) is the decision tree obtained after the mth round of iterative training, M is the number of decision trees, and m is the number of iterations.

[0048] Define the DNN neural network model architecture. Figure 3 As shown in the figure, in the DNN class, three fully connected layers (nn.Linear) are defined, where the input layer has 4 neurons, the two hidden layers have 64 neurons each, and the output layer has 1 neuron (used to output the predicted probability of binary classification). The data first passes through the first hidden layer, and then the ReLU activation function is applied; then it passes through the second hidden layer, and the ReLU activation function is also applied; finally, it passes through the output layer, and the Sigmoid activation function is applied to limit the output between 0 and 1, representing the predicted probability of the binary classification problem.

[0049] For DNN neural networks, an accuracy function is defined to calculate the accuracy of binary classification problems. This function compares the output of the model (prediction probability) with the true label. When the prediction probability is greater than 0.5, the prediction is considered to be a positive class (1), otherwise it is a negative class (0). Then, the degree of match between the prediction result and the true label is calculated and the accuracy is returned.

[0050]

[0051] in, is the predicted category, y i is the true category, I is the indicator function (returns 1 when the condition in the brackets is true, otherwise 0), and N is the total number of samples.

[0052] The get_model function is used to initialize the DNN model and an Adam optimizer, where the optimizer parameters are all the parameters of the model (model.parameters()), and the learning rate (lr) is defined outside the function. Define a binary cross entropy loss function (nn.BCELoss()) as the loss function. Its expression is:

[0053]

[0054] Among them, y is the true label value, which is a vector or tensor composed of 0 and 1, indicating the true category of each sample. is the predicted value of the model, which is a vector or tensor composed of probability values, indicating the probability that the model predicts each sample to be a positive example. N is the number of samples.

[0055] After successfully building the model, the Adam optimizer was used with the momentum parameter set to 0.9, the initial learning rate to 0.001, the weight decay to 1e-4, and the cosine annealing scheme for learning rate decay. The batch size was set to 4. The binary cross entropy loss function was used. The training was performed for 200 epochs without early stopping. During the training process, if the performance on the validation set improved, the model weights were saved after each epoch to avoid underfitting or overfitting. The model evaluation results under different data implementation examples are shown in Figure 2. Figure 4 shown.

[0056] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multidimensional data fusion processing method for medical text data, characterized in that: The processing steps include: obtaining multidimensional data of medical texts, inputting the multidimensional data of medical texts into a trained fusion model to obtain a multidimensional data fusion result; training the fusion model includes: obtaining medical texts, constructing a multidimensional data set based on the medical texts, and using data splitting technology to divide the multidimensional data set into a training set and a test set; extracting features from the data in the training set to obtain a feature map; performing tensor conversion on the feature map and reconstructing the features after tensor conversion; inputting the reconstructed feature data into the fusion model for training, adjusting the model parameters, and completing the model training when the model loss function is minimized.

2. A multidimensional data fusion processing method for medical text data according to claim 1, characterized in that: Multidimensional data include clinical data, biochemical data, imaging examination results, and laboratory test results.

3. A multidimensional data fusion processing method for medical text data according to claim 1, characterized in that: Feature extraction of data includes: using the GBDT algorithm to process the data in the data set, iteratively calculating the data by constructing multiple decision trees, and accumulating all iterative results to obtain the final result; extracting the features and target variables of the final result.

4. The multidimensional data fusion processing method for medical text data according to claim 1 is characterized in that: The tensor conversion expression for the feature map is as follows: T = torch.from_numpy(A) Here, torch from_numpy is a function in the PyTorch library, A is the original Numpy input, and T is the output Pytorch tensor.

5. The multidimensional data fusion processing method for medical text data according to claim 1 is characterized in that: Reconstructing the features after tensor conversion includes: taking the features and target data as dataset objects, creating a data loader, extracting the dataset objects in batches, and shuffling the extracted data to obtain reconstructed features; the data loader is created as: texttrain=Dataloader(Dataset.train,batch_size,shuffle=True) texttest=Dataloader(Dataset.test,batch_size×2,shuffle=False) Among them, texttrain is the training set, Dataloader is the data loader function, Dataset.train is a part of the training set, batch_size is the batch size of the training set, shuffle is the data order parameter, True means shuffling the data order, False means not shuffling the data order. texttest is the test set, and Dataset.test is a part of the test set.

6. A multidimensional data fusion processing method for medical text data according to claim 1, characterized in that: The processing of reconstructed features using the fusion model includes: obtaining target data, using the gradient boosting tree to calculate the correlation between the reconstructed features and the target data, and converting the reconstructed features into PyTorch tensors according to the calculation results; reconstructing the PyTorch tensor and the target data into a data set; and inputting the data in the data set into the DNN neural network to obtain the prediction results.

7. A multidimensional data fusion processing method for medical text data according to claim 6, characterized in that: The fusion model includes determining the initial model and subsequent decision trees by minimizing the loss function, updating the model by the learning rate, and obtaining a strong learner composed of the weighted sum of all base learners; the fusion model is: Among them, F0(x) is the initial model, ρ m is the learning rate, h m (x) is the decision tree obtained after the mth round of iterative training, M is the number of decision trees, and m is the number of iterations.

8. The multidimensional data fusion processing method for medical text data according to claim 6 is characterized in that: The DNN neural network consists of three fully connected layers, two hidden layers and one output layer. The input layer of each fully connected layer has 4 neurons, the two hidden layers have 64 neurons each, and the output layer has 1 neuron.

9. The multidimensional data fusion processing method for medical text data according to claim 1, characterized in that: The loss function of the model adopts the binary cross entropy loss function, which is expressed as: Among them, y is the true label value, is the predicted value of the model and N is the number of samples.

Images