Training method of intelligent medical decision model based on digestive system department

Abstract

The invention relates to the technical field of medical model training, in particular to an intelligent medical decision model training method based on the digestive system department. The method comprises the following steps: obtaining digestive system department medical image data, digestive system department gene experiment time sequence data and digestive system department medical record basic data, and carrying out modeling pre-preprocessing and input layer training design to generate a decision model feature extraction input layer; performing feature vectorization processing on various corresponding medical features in the decision model feature extraction input layer through an embedded network layer to generate various digestive system department medical feature vector quantities; performing feature fusion and model joint training on the various digestive system department medical feature vectors to generate a digestive system department intelligent medical decision model; and performing feature dependence screening and deep reinforcement learning updating on the various digestive system department medical feature vectors to generate a digestive system department intelligent medical decision optimization model. The method can assist in making more accurate and efficient medical decisions.

Description

Technical Field

[0001] The present invention relates to the technical field of medical model training, and in particular to a training method for an intelligent medical decision-making model based on gastroenterology. Background Art

[0002] In recent years, intelligent medical decision-making models based on machine learning, deep learning and other technologies have gradually become effective tools for solving medical decision-making problems. Intelligent medical decision-making models based on big data and artificial intelligence analyze a large amount of clinical data, including symptoms, signs, laboratory test results, imaging data, and treatment responses, and use advanced deep learning algorithms and time series data analysis technology to conduct intelligent training for medical decision-making in gastroenterology. By dynamically tracking changes in patients' conditions and updating model parameters in real time, personalized and precise decision support is achieved to help doctors make corresponding personalized decision plans. However, most of the existing intelligent decision-making methods rely on rule engines or expert systems, and their reasoning process is relatively simple. They fail to fully consider the dynamic changes in patients' conditions and individual differences, and thus fail to effectively explore the dependencies between various medical features and the potential associations with decision results. Summary of the invention

[0003] Based on this, it is necessary for the present invention to provide a training method for an intelligent medical decision-making model based on gastroenterology to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above purpose, a training method for an intelligent medical decision-making model based on gastroenterology includes the following steps:

[0005] Step S1: Obtaining gastroenterology medical imaging data, gastroenterology gene experiment time series data, and gastroenterology medical record basic data from the gastroenterology medical information system and the electronic medical record system; performing modeling pre-preprocessing on the gastroenterology medical imaging data, gastroenterology gene experiment time series data, and gastroenterology medical record basic data, respectively, to generate standard medical imaging data, standard gene experiment time series data, and standard medical record basic data;

[0006] Step S2: Perform input layer training design on standard medical imaging data, standard gene experiment time series data and standard medical record basic data to generate a decision model feature extraction input layer; perform feature vectorization processing on various medical features corresponding to the decision model feature extraction input layer through an embedded network layer to generate various gastroenterology medical feature vectors; perform feature fusion on various gastroenterology medical feature vectors through the introduction of an attention mechanism fusion layer, and use a fully connected layer to perform joint training on an intelligent decision model to generate a gastroenterology intelligent medical decision model;

[0007] Step S3: performing feature dependency screening on various gastroenterology medical feature vectors to obtain gastroenterology medical feature dependency descriptors;

[0008] Step S4: Based on the gastroenterology medical feature dependency descriptor, the gastroenterology intelligent medical decision-making model is updated by deep reinforcement learning to generate a gastroenterology intelligent medical decision-making optimization model.

[0009] Further, step S1 includes the following steps:

[0010] Step S11: obtaining gastroenterology medical imaging data and gastroenterology gene experiment time series data from the gastroenterology medical information system, wherein the gastroenterology medical imaging data include gastrointestinal CT images, MRI images, ultrasound examination images and digestive endoscopy images, and the gastroenterology gene experiment time series data include blood test time series data, liver function experiment time series data, kidney function experiment time series data, intestinal microbiome experiment time series data and digestive tract gene experiment time series data;

[0011] Step S12: Obtaining basic medical record data of gastroenterology from the electronic medical record system, wherein the basic medical record data of gastroenterology includes basic information of the patient, diagnosis and treatment records, gastroenterology medical history, and gastroenterology medication history;

[0012] Step S13: performing image data preprocessing on the gastroenterology medical image data, including cropping, rotating, enlarging, and contrast enhancement, to generate standard medical image data;

[0013] Step S14: performing time series denoising and standardization processing on the gastroenterology gene experiment time series data to generate standard gene experiment time series data;

[0014] Step S15: extract information data and perform standardization processing on the basic data of gastroenterology medical records to generate standard basic data of medical records.

[0015] Further, step S2 includes the following steps:

[0016] Step S21: using a convolutional neural network to extract image lesion features from standard medical image data to extract gastroenterology medical image lesion features, and constructing a gastroenterology medical image feature extraction layer;

[0017] Step S22: extracting time series variation features of standard gene experiment time series data using a long short-term memory network to extract time series variation features of gastroenterology medical gene experiments, and constructing a gastroenterology medical gene time series feature extraction layer;

[0018] Step S23: using a multi-layer perceptron corresponding to a deep neural network to perform multi-level nonlinear feature extraction on the basic data of the standard medical records, so as to extract nonlinear high-level features of the gastroenterology medical records, and construct a gastroenterology medical records feature extraction layer;

[0019] Step S24: Performing input layer training design on the gastroenterology medical image feature extraction layer, the gastroenterology medical gene time series feature extraction layer, and the gastroenterology medical medical record feature extraction layer to generate a decision model feature extraction input layer; performing feature vectorization processing on the gastroenterology medical image lesion features, gastroenterology medical gene experiment time series change features, and gastroenterology medical medical record nonlinear high-level features corresponding to the decision model feature extraction input layer through an embedded network layer to generate various types of gastroenterology medical feature vectors;

[0020] Step S25: By introducing the attention mechanism fusion layer, various gastroenterology medical feature vectors are feature-fused, and the fully connected layer is used to jointly train the intelligent decision-making model to generate a gastroenterology intelligent medical decision-making model.

[0021] Furthermore, the feature extraction input layer of the decision model in step S24 is reconstructed through a cross-domain hierarchical connection design to eliminate the heterogeneity and feature information asymmetry between the gastroenterology medical image lesion characteristics, the gastroenterology medical gene experiment temporal change characteristics and the gastroenterology medical medical record nonlinear high-level features, wherein the gastroenterology medical image feature extraction layer is used as the feature input first layer, the gastroenterology medical gene temporal feature extraction layer is used as the feature input second layer, and the gastroenterology medical medical record feature extraction layer is used as the feature input third layer.

[0022] Furthermore, the step S24 of performing feature vectorization processing on the gastroenterology medical image lesion features, gastroenterology medical gene experiment temporal change features and gastroenterology medical record nonlinear high-level features corresponding to the decision model feature extraction input layer through the embedded network layer includes the following steps:

[0023] Through the embedded network layer, multi-level feature extraction and vectorization are performed on the corresponding gastroenterology medical image lesion features in the decision model feature extraction input layer, from low-level edge features to high-level structural features, to generate gastroenterology medical image feature vectors, wherein the gastroenterology medical image feature vectors are specifically feature vectors formed by the aggregation of lesion area texture features, lesion area morphological features, and lesion area pixel distribution features;

[0024] Through the embedded network layer, the time series change characteristics of the gastroenterology medical gene experiment corresponding to the decision model feature extraction input layer are dynamically extracted and vectorized to generate the gastroenterology medical gene experiment time series feature vector, wherein the gastroenterology medical gene experiment time series feature vector is specifically a feature vector formed by the aggregation of the activity of the gastroenterology medical gene at different time nodes and the experimental change trend characteristics;

[0025] Through the embedding network layer, the nonlinear high-level features of the gastroenterology medical records corresponding to the input layer of the decision model feature extraction are subjected to word embedding semantic feature extraction and vectorization, so as to extract the corresponding nonlinear and high-dimensional semantic features in the gastroenterology medical records text and generate the gastroenterology medical records text semantic feature vector, wherein the gastroenterology medical records text semantic feature vector is specifically a feature vector formed by the aggregation of the semantic features of the medical record disease description, pathological changes, and diagnosis and treatment process;

[0026] The gastroenterology medical image feature quanta, the gastroenterology medical gene experiment time series feature quanta and the gastroenterology medical record text semantic feature quanta are used as various types of gastroenterology medical feature quanta.

[0027] Further, step S25 includes the following steps:

[0028] Step S251: Performing medical feature contribution analysis on various gastroenterology medical features to quantum by introducing an attention mechanism fusion layer to obtain feature decision contribution corresponding to various gastroenterology medical features;

[0029] Step S252: performing feature contribution weighted fusion on various gastroenterology medical feature vectors based on the feature decision contribution corresponding to various gastroenterology medical features, so as to generate a gastroenterology medical global fusion feature vector;

[0030] Step S253: Obtain the clinical decision-making target of gastroenterology, and based on the clinical decision-making target of gastroenterology, use the fully connected layer to jointly train the intelligent decision-making model of gastroenterology global fusion feature vector to generate an intelligent medical decision-making model for gastroenterology.

[0031] Further, step S3 includes the following steps:

[0032] Step S31: performing feature correlation analysis on each pair of features in each type of gastroenterology medical feature vector to generate a gastroenterology medical feature correlation measurement matrix;

[0033] Step S32: performing high correlation feature clustering optimization on each gastroenterology medical sub-feature in each gastroenterology medical feature vector based on the gastroenterology medical feature correlation measurement matrix to obtain gastroenterology medical high correlation feature clustering clusters;

[0034] Step S33: using the inter-feature dependency strength calculation formula to perform feature dependency measurement calculation between each gastroenterology medical sub-feature within the gastroenterology medical high correlation feature cluster, so as to obtain the feature decision dependency strength corresponding to each gastroenterology medical sub-feature;

[0035] Step S34: performing dependency screening on each gastroenterology medical sub-feature in the gastroenterology medical high correlation feature cluster based on the feature decision dependency strength corresponding to each gastroenterology medical sub-feature, so as to obtain a gastroenterology medical optimized feature dependency subset;

[0036] Step S35: Perform structured description transformation on the gastroenterology medical optimized feature dependency subset to obtain a gastroenterology medical feature dependency descriptor.

[0037] Further, step S31 includes the following steps:

[0038] Combining each gastroenterology medical feature sub-feature in each gastroenterology medical feature vector to generate a gastroenterology medical feature combination pair;

[0039] The Pearson correlation coefficient is calculated between each pair of features of each gastroenterology medical feature combination pair to obtain the Pearson correlation coefficient between each gastroenterology medical feature pair;

[0040] Based on the Pearson correlation coefficients between each pair of gastroenterology medical features, a correlation measurement matrix is ​​constructed for the corresponding gastroenterology medical feature combination pairs to generate a gastroenterology medical feature correlation measurement matrix.

[0041] Furthermore, the calculation formula for the inter-feature dependency strength in step S33 is specifically:

[0042]

[0043] Where D ij is the feature decision dependency strength between the i-th gastroenterology sub-feature and the j-th gastroenterology sub-feature, X is the feature space corresponding to the gastroenterology high correlation feature cluster, x i is the i-th gastroenterology sub-feature, x j is the jth gastroenterology sub-feature, Cov(x i ,x j ) is the covariance between the i-th gastroenterology sub-feature and the j-th gastroenterology sub-feature, is the standard deviation corresponding to the i-th gastroenterology sub-feature, is the standard deviation corresponding to the j-th gastroenterology sub-feature, α is the dependency contribution coefficient of the gastroenterology sub-feature, exp is the exponential function, β is the distance attenuation rate of the gastroenterology sub-feature, dx is the integral differential, and η is the correction coefficient of the feature decision dependency strength.

[0044] Further, step S4 includes the following steps:

[0045] Step S41: reconstructing the temporal relationship of various medical features corresponding to the gastroenterology medical feature dependency descriptor to generate a gastroenterology medical feature temporal dependency association network;

[0046] Step S42: simulating the medical decision path according to the gastroenterology medical feature temporal dependency association network to obtain a gastroenterology medical decision simulation evolution path diagram;

[0047] Step S43: performing decision optimization selection on each medical decision path in the gastroenterology medical decision simulation evolution path diagram to generate a gastroenterology medical optimization decision path;

[0048] Step S44: performing deep cross-feedback update on the corresponding clinical medical decision in the gastroenterology intelligent medical decision model based on the gastroenterology medical optimization decision path to generate a gastroenterology intelligent decision feedback update model;

[0049] Step S45: Based on the gastroenterology intelligent decision-making feedback update model, the corresponding clinical medical decision in the gastroenterology intelligent medical decision-making model is updated by feedback loop reinforcement learning to generate a gastroenterology intelligent medical decision-making optimization model.

[0050] Beneficial effects of the present invention:

[0051] The training method of the intelligent medical decision-making model based on gastroenterology proposed in the present invention, compared with the prior art, has the beneficial effect of providing a comprehensive multidimensional data basis for subsequent medical analysis and research by obtaining gastroenterology medical imaging data and gene experiment time series data from the gastroenterology medical information system. This data set includes medical imaging data such as gastrointestinal CT images, MRI images, ultrasound examination images, and digestive endoscopy images, as well as blood test time series data, liver function experiment time series data, kidney function experiment time series data, intestinal microbiome experiment time series data, and digestive tract gene experiment time series data. Gene experiment data, the integration of these data helps to conduct in-depth analysis of research in the field of gastroenterology from different angles. In addition, by obtaining basic medical record data of gastroenterology from the electronic medical record system, the patient's personal health information, diagnosis and treatment records, medical history, and medication history are provided. These data contain the patient's basic personal information, such as age, gender, weight, etc., which can help doctors quickly understand the patient's basic situation, thereby supporting more accurate and efficient medical decision-making and health management. At the same time, by preprocessing the medical imaging data of gastroenterology, mainly including operations such as cropping, rotation, magnification and contrast enhancement, these processes can significantly improve the quality and analysis efficiency of medical imaging data, thereby enhancing doctors' ability to interpret images. By performing time series denoising and standardization on the time series data of gastroenterology gene experiments, these values ​​deviate from the true trend of change due to factors such as noise and instrument errors in the experimental process, making the data more stable and accurate, which helps researchers to dig out the potential laws and connections in the data. By extracting and standardizing the basic data of gastroenterology medical records, the original medical record data can be converted into structured, high-quality information. This process includes extracting the patient's symptoms, signs, past medical history, medication information, etc. from the unstructured medical records, and converting this information into a standardized format to make it easy for computers to process and analyze, which can help researchers effectively dig out the dependencies between various medical features and the potential associations with decision-making results at a macro level.Secondly, by designing the input layer training for the standardized medical imaging data, genetic experiment time series data and basic medical record data, and jointly training through feature vectorization, feature fusion and intelligent decision-making model, the intelligent medical decision-making model for gastroenterology is trained and generated. The key to this step is to be able to efficiently extract useful medical features from multi-source heterogeneous data and conduct comprehensive analysis through deep learning technology. Feature vectorization processing can convert different types of medical data into digital formats that can be understood by machine learning models, thereby providing actionable input for the decision-making model. On this basis, the introduction of the attention mechanism fusion layer can automatically adjust the importance of features according to the correlation of different features in the data, so that the model can pay more attention to the duel. The key features that have a greater impact on the decision-making process are selected, which avoids the problem that the dynamic changes in the patient's condition and individual differences are not fully considered when designing features manually in the traditional way. Through the joint training of the fully connected layer, the features from different data sources (such as imaging data, genetic data, and medical record data) can be effectively integrated to form a more comprehensive and integrated intelligent decision-making model. Such a model can not only process complex and multi-dimensional medical data, but also improve the accuracy and intelligence level of decision-making. Through this step, the generated intelligent decision-making model can adaptively extract important information from multi-dimensional data and conduct efficient learning and optimization, which provides a solid foundation for subsequent deep reinforcement learning and model optimization, and thus promotes the development of intelligent medical decision-making model training. Then, by performing feature dependency screening on various types of gastroenterology medical feature vector quanta, it aims to screen out features with strong dependencies from a large number of medical feature vectors and generate gastroenterology medical feature dependency descriptors. The core of this step is to reduce the impact of redundant features, improve the correlation between features, and effectively focus on the important features that affect decision-making. Through feature dependency screening, it is possible to identify which features have strong correlations and which features are redundant or irrelevant, thereby optimizing the feature space. There are certain correlations and dependencies between these features. Finally, the intelligent medical decision-making model of gastroenterology is updated through deep reinforcement learning based on the medical feature dependency descriptor of gastroenterology. The key to this step is to further improve the adaptability and optimization capabilities of the model through the reinforcement learning mechanism. Reinforcement learning is a learning method based on trial and error, which can learn the optimal decision-making strategy through interaction with the environment. In the intelligent medical decision-making model of gastroenterology, the introduction of reinforcement learning can enable the model to self-optimize and adjust in complex medical data, and improve the adaptability of the model to different scenarios and individual patients by continuously updating parameters and strategies. Reinforcement learning continuously adjusts the feedback of the model, so that the model can make more accurate and reasonable decisions when facing unknown or complex situations, thereby greatly improving the practicality and effectiveness of the intelligent decision-making model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0053] Figure 1 It is a schematic diagram of the steps of the training method of the intelligent medical decision-making model based on gastroenterology of the present invention;

[0054] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0055] Figure 3 for Figure 1 Detailed step flow chart of step S2 in FIG. DETAILED DESCRIPTION

[0056] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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 technicians in this field without creative work are within the scope of protection of the present invention.

[0057] To achieve this, please refer to Figures 1 to 3 The present invention provides a training method for an intelligent medical decision-making model based on gastroenterology. In the embodiments of the present invention, please refer to Figure 1 FIG. 1 is a flow chart of the steps of the training method of the intelligent medical decision model based on gastroenterology of the present invention. In this example, the training method of the intelligent medical decision model based on gastroenterology includes the following steps:

[0058] Step S1: Obtaining gastroenterology medical imaging data, gastroenterology gene experiment time series data, and gastroenterology medical record basic data from the gastroenterology medical information system and the electronic medical record system; performing modeling pre-preprocessing on the gastroenterology medical imaging data, gastroenterology gene experiment time series data, and gastroenterology medical record basic data, respectively, to generate standard medical imaging data, standard gene experiment time series data, and standard medical record basic data;

[0059] In an embodiment of the present invention, the medical image data and gene experiment time series data of the Department of Gastroenterology are accessed and extracted through the medical information system of the Department of Gastroenterology. The medical image data include gastrointestinal CT images, MRI images, ultrasound examination images and digestive endoscopy images. For these image data, an image reading tool (such as a DICOM file reading tool) is used to decode and extract the image files to ensure that the extracted images conform to the standard DICOM format, and the gene experiment time series data include blood test time series data, liver function test time series data, kidney function test time series data, intestinal microbiome test time series data and digestive tract gene test time series data. These data are generally stored in a table or database format and are obtained from the database using an SQL database query tool, thereby obtaining the medical image data of the Department of Gastroenterology and the gene experiment time series data of the Department of Gastroenterology. The basic medical record data of the patient is obtained through the electronic medical record system (EMR system). The medical record data includes the basic information of the patient, diagnosis and treatment records, gastroenterology medical history and gastroenterology medication history. The API interface provided by the EMR system is used to query all relevant data of the patient through the patient's unique identifier (such as patient ID), thereby obtaining the basic medical record data of the Department of Gastroenterology. At the same time, standardized image preprocessing operations are performed on gastroenterology medical imaging data to ensure that the data quality is suitable for subsequent machine learning or image analysis. First, the image is cropped by using image processing software (such as ITK, SimpleITK or OpenCV) to remove irrelevant areas (such as the background in the image), and the image is rotated to make the image orientation consistent. Then, the image is scaled using magnification technology to ensure that the image resolution is high enough for more detailed analysis. At the same time, contrast enhancement operations are performed to highlight the details of gastrointestinal related structures by changing the brightness and contrast of the image, thereby generating standard medical imaging data. Genetic experiment time series data usually contains a large amount of time series data, which will be affected by noise and outliers. In order to denoise the original data using time series denoising techniques (such as wavelet transform, moving average or Savitzky-Golay filter), remove high-frequency noise, retain the main trend of the signal, and perform standardization. The Z-score standardization method or minimum-maximum standardization is used to convert data of different dimensions and different numerical ranges into a unified standard range (for example, normalize the data to the range of 0-1), thereby generating standard genetic experiment time series data.Then, by extracting and standardizing the electronic medical record data, natural language processing (NLP) technology is used to extract information from unstructured text data such as medical records, medical history, and medication history, and named entity recognition (NER) technology is used to extract key medical entities such as disease names, drug names, symptom descriptions, etc. These entities are then associated with other patient data to construct standardized structured data. For structured data (such as basic information, medical records, medical history, etc.), data cleaning technology is used to fill in missing values, correct erroneous data, and standardize all text fields (such as unifying the representation of drug names and standardizing disease classifications in medical records) to ultimately generate standard medical record basic data.

[0060] Step S2: Perform input layer training design on standard medical imaging data, standard gene experiment time series data and standard medical record basic data to generate a decision model feature extraction input layer; perform feature vectorization processing on various medical features corresponding to the decision model feature extraction input layer through an embedded network layer to generate various gastroenterology medical feature vectors; perform feature fusion on various gastroenterology medical feature vectors through the introduction of an attention mechanism fusion layer, and use a fully connected layer to perform joint training on an intelligent decision model to generate a gastroenterology intelligent medical decision model;

[0061] In an embodiment of the present invention, standard medical image data is processed to extract lesion features of gastroenterology medical images. The specific operation steps include using a convolutional neural network (CNN) to process the image data. The convolution layer in the network will automatically learn and extract low-level features (such as edges, textures, etc.) in the image. In the subsequent pooling layer, the feature dimension is reduced through dimensionality reduction operations, and significant features are highlighted. For gastroenterology images, special attention is paid to the features of related lesion areas such as the intestines and stomach. The level of feature extraction is deepened through a multi-layer convolutional network. During the network training process, the back-propagation algorithm is used to adjust the weights to minimize the feature extraction error. The extracted lesion features of gastroenterology medical images are used to construct a gastroenterology medical image feature extraction layer as input to the downstream decision model. The long short-term memory network (LSTM) is used to extract the time series features of the standard genetic experiment time series data, so that the time series data can be trained through the LSTM model. The LSTM network is particularly suitable for processing time series data and can capture time dependencies. In this step, the input of the LSTM model is the standard genetic experiment data, and the output is the corresponding time series feature vector. Bidirectional LSTM can be selected to extract forward and reverse time series features, and the appropriate number of layers and units can be set to improve the expressiveness of the model. During the model training process, back propagation is performed through the contrast loss function (such as mean square error or cross entropy loss) to update the network weights. The time series feature extraction layer of the genetic experiment is constructed and used to generate gastroenterology medical gene time series features. At the same time, the nonlinear features of gastroenterology medical records data are extracted by using the multilayer perceptron (MLP) in the deep neural network, and these numeralized medical record features are input into the multilayer perceptron. The MLP consists of multiple layers of fully connected neurons, and the output of each layer is used as the input of the next layer. The nonlinear features of different dimensions are learned layer by layer, and the input signal is transmitted to the output layer through weighted connections between the neurons in each layer. At the same time, the back propagation algorithm is used to optimize the network parameters. After multiple training iterations, the nonlinear high-level features of gastroenterology medical records are extracted, and the gastroenterology medical record feature extraction layer is constructed.The previously constructed gastroenterology medical image feature extraction layer, gastroenterology medical gene time series feature extraction layer and gastroenterology medical medical record feature extraction layer are trained and designed simultaneously to construct a decision model feature extraction input layer, wherein the gastroenterology medical image feature extraction layer is used as the first input layer, and the medical images (such as gastrointestinal endoscopy images, CT images, etc.) are subjected to feature extraction through a convolutional neural network (CNN) to obtain the features of the lesion area in the image, and the gastroenterology medical gene time series feature extraction layer is used as the second input layer to perform time series analysis on genetic experimental data (such as gene expression profiles, DNA sequencing data, etc.) through a long short-term memory network, and extract disease-related time series change features, and the gastroenterology medical record feature extraction layer is used as the third input layer to train and encode high-level nonlinear features extracted from medical record data through a deep neural network, and through the design of cross-domain hierarchical connection reconstruction, the features of these three types of different sources are integrated to eliminate the heterogeneity and information asymmetry between them, and ensure seamless connection between features, thereby generating a decision model feature extraction input layer. In addition, multimodal feature input training is performed by combining gastroenterology medical image features, genetic experiment time series features and medical record features to integrate the various features extracted in the previous steps (medical image lesion features, genetic experiment time series change features, medical record nonlinear high-level features) into a feature vector as the training data of the input layer. In order to make the features more adaptable to the needs of the model, these features are vectorized through the embedding network layer. The embedding layer converts each feature into a fixed-length vector representation through a lookup table, thereby generating various gastroenterology medical feature vectors. Then, by introducing the attention mechanism layer for feature fusion, the intelligent decision-making model is trained to assign different weights to each input feature vector by introducing the self-attention mechanism, so as to highlight important features according to the needs of the task. Specifically, the attention mechanism assigns weights to each feature by calculating the similarity between the input features. Features with larger weights will have a greater impact on the decision results, and the weighted features are processed by using the fully connected layer. The output of the fully connected layer will be a decision result that integrates medical image features, gene time series features and nonlinear features of medical records. The weighted fusion of these features can significantly improve the accuracy and robustness of the model in different clinical scenarios. All parameters of the network are optimized through the back propagation algorithm, so that the final intelligent decision-making model can effectively provide support for medical decision-making in the field of gastroenterology, and finally generate an intelligent medical decision-making model for gastroenterology.

[0062] Step S3: performing feature dependency screening on various gastroenterology medical feature vectors to obtain gastroenterology medical feature dependency descriptors;

[0063] In an embodiment of the present invention, various types of medical feature vector quanta are extracted from a medical data set related to gastroenterology, and these feature vector quanta contain information on different gastroenterology diseases or clinical indicators. For each pair of feature vector quanta, the Pearson correlation coefficient, Spearman rank correlation coefficient or other correlation measurement methods are used to calculate the correlation value between each pair of features, and a feature correlation measurement matrix is ​​generated. After the feature correlation measurement matrix is ​​generated, the gastroenterology medical features are clustered and optimized based on the matrix, so as to use a clustering algorithm (such as K-means clustering, hierarchical clustering, etc.) to perform feature aggregation according to the correlation between the features. The core goal of clustering is to group features with strong correlation into the same group. At the same time, feature dependency measurement calculation is performed between each gastroenterology medical sub-feature in a high-correlation feature clustering cluster to quantify the dependency measurement value between each feature, and dependency screening is performed on the features in each feature clustering cluster based on the dependency measurement value, so as to screen out feature pairs with strong dependence by setting a dependency threshold, and retain these features. In this process, the screening of feature dependency can be achieved by By setting conditions, such as retaining only feature pairs with a dependency greater than 0.7, to reduce redundant features and improve model efficiency, the feature selection tools of Scikit-learn in Python (such as SelectKBest, VarianceThreshold, etc.) can be used to further optimize the feature subsets to ensure that the final selected feature subsets reduce the influence of irrelevant features while maintaining strong dependencies. Then, after obtaining the optimized feature dependency subsets, the structured description is transformed to generate standardized text or numerical descriptions by formatting each feature subset. The purpose of the structured description transformation is to present the optimized feature dependency in an interpretable form so that it can be further applied to the training and reasoning of the decision model. For example, the pandas library in Python is used to convert the feature dependency matrix into a tabular form, or a custom encoding rule is used to generate a description in a specific format. The obtained feature dependency descriptor provides a clear and structured feature dependency basis for the training of subsequent intelligent decision models, and finally the gastroenterology medical feature dependency descriptor is obtained.

[0064] Step S4: Based on the gastroenterology medical feature dependency descriptor, the gastroenterology intelligent medical decision-making model is updated by deep reinforcement learning to generate a gastroenterology intelligent medical decision-making optimization model.

[0065] In an embodiment of the present invention, various types of medical features corresponding to the medical feature dependency descriptors of gastroenterology are extracted previously, including but not limited to the patient's lesion texture, lesion pixel distribution, gene activity, medical record disease description, pathological changes and other features, and these data are processed using a time series analysis method to redefine the temporal relationship between the various medical features. A temporal dependency network of gastroenterology medical features can be constructed, in which nodes represent medical features and edges represent temporal dependencies between features. In the process of constructing the temporal relationship, the causal order between features needs to be considered. For example, gastrointestinal symptoms may precede abnormalities in certain biochemical indicators, or changes in certain test results may affect subsequent diagnostic decisions. The clinical decision-making process is simulated using a simulation algorithm in combination with the previously generated temporal dependency network, so as to construct a decision model based on temporal dependency to simulate the decision path under different medical feature change conditions. This process can use a graph-based traversal algorithm (such as depth-first search, breadth-first search, etc.) to simulate all medical decision paths. In this paper, a medical decision simulation evolution path diagram is generated. This diagram presents the key nodes (such as symptom changes, decision plans, etc.) on each decision path and its evolution process, and also applies optimization algorithms to evaluate and optimize each decision path. This process can use multi-objective optimization algorithms (such as genetic algorithms, particle swarm optimization algorithms, etc.) to weigh different decision paths and optimize the optimal decision plan. The optimization goals include factors such as the shortest decision time, the lowest patient risk, and the best decision effect. By continuously adjusting the decision nodes in the path and their selection, a medical optimization decision path is generated. By combining the previously obtained medical optimization decision path, the intelligent medical decision model of gastroenterology is cross-feedback updated through a deep learning algorithm. Specifically, the optimized decision path is input into the decision model, and a back propagation algorithm (such as a deep neural network training process) is used. According to the difference between the effect of each path and the actual clinical situation, error adjustment is performed to optimize the model's predictive ability to generate an intelligent decision feedback update model. Then, by combining the previously generated intelligent decision feedback to update the model, the reinforcement learning method is used to further optimize the intelligent medical decision-making model of gastroenterology. Specifically, the policy gradient method or Q-learning algorithm in reinforcement learning is used to dynamically adjust the decision path based on the patient's historical data and real-time decision feedback. Whenever the decision path generated by the model can bring better decision-making results or lower risks, positive feedback will be obtained, thereby adjusting the strategy to better adapt to clinical decision-making needs. Through long-term learning and feedback, the model is continuously optimized, and finally a gastroenterology intelligent medical decision optimization model is generated.

[0066] Further, as an embodiment of the present invention, refer to Figure 2 As shown, Figure 1Detailed step flow diagram of step S1 in FIG. 1 , in this embodiment, step S1 includes the following steps:

[0067] Step S11: obtaining gastroenterology medical imaging data and gastroenterology gene experiment time series data from the gastroenterology medical information system, wherein the gastroenterology medical imaging data include gastrointestinal CT images, MRI images, ultrasound examination images and digestive endoscopy images, and the gastroenterology gene experiment time series data include blood test time series data, liver function experiment time series data, kidney function experiment time series data, intestinal microbiome experiment time series data and digestive tract gene experiment time series data;

[0068] In an embodiment of the present invention, the gastroenterology medical image data and genetic experiment time series data are accessed and extracted through the gastroenterology medical information system. The medical image data include gastrointestinal CT images, MRI images, ultrasound examination images and digestive endoscopy images. For these image data, an image reading tool (such as a DICOM file reading tool) is used to decode and extract the image files to ensure that the extracted images conform to the standard DICOM format. The genetic experiment time series data includes blood test time series data, liver function experiment time series data, kidney function experiment time series data, intestinal microbiome experiment time series data and gastrointestinal gene experiment time series data. These data are generally stored in a table or database format and are obtained from the database using an SQL database query tool to ensure the integrity and continuity of the time series data and avoid affecting the analysis results due to data loss or errors. Finally, the gastroenterology medical image data and the gastroenterology genetic experiment time series data are obtained.

[0069] Step S12: Obtaining basic medical record data of gastroenterology from the electronic medical record system, wherein the basic medical record data of gastroenterology includes basic information of the patient, diagnosis and treatment records, gastroenterology medical history, and gastroenterology medication history;

[0070] In an embodiment of the present invention, the basic medical record data of the patient is obtained through an electronic medical record system (EMR system). The medical record data includes the patient's basic information, medical records, gastroenterology medical history, and gastroenterology medication history. The API interface provided by the EMR system is used to query all relevant data of the patient through the patient's unique identifier (such as patient ID). The data includes the patient's age, gender, medical history, medical records, and medication history. These data are stored in a structured form and can be transmitted using JSON, XML, and other formats to ultimately obtain the basic medical record data of gastroenterology.

[0071] Step S13: performing image data preprocessing on the gastroenterology medical image data, including cropping, rotating, enlarging, and contrast enhancement, to generate standard medical image data;

[0072] In an embodiment of the present invention, standardized image preprocessing operations are performed on gastroenterology medical image data to ensure that the data quality is suitable for subsequent machine learning or image analysis. First, the image is cropped by using image processing software (such as ITK, SimpleITK or OpenCV) to remove irrelevant areas (such as the background in the image), retain the gastrointestinal area, and rotate the image to make the image orientation consistent to avoid affecting the subsequent analysis results due to inconsistent image orientation. Then, the image is scaled using a magnification technique to ensure that the image resolution is high enough for more detailed analysis. At the same time, a contrast enhancement operation is performed to highlight the details of gastrointestinal related structures by changing the brightness and contrast of the image, enhance the visualization of the image, and finally generate standard medical image data.

[0073] Step S14: performing time series denoising and standardization processing on the gastroenterology gene experiment time series data to generate standard gene experiment time series data;

[0074] In an embodiment of the present invention, genetic experiment time series data usually contains a large amount of time series data, which may be affected by noise and outliers. The original data is denoised by using time series denoising technology (such as wavelet transform, moving average or Savitzky-Golay filter), high-frequency noise is removed, the main trend of the signal is retained, and standardized. The Z-score standardization method or minimum-maximum standardization is used to convert data of different dimensions and different numerical ranges into a unified standard range (for example, the data is normalized to the range of 0-1). This can eliminate the differences between different experimental data, improve the comparability and stability of the data in the machine learning model, and finally generate standard genetic experiment time series data.

[0075] Step S15: extract information data and perform standardization processing on the basic data of gastroenterology medical records to generate standard basic data of medical records.

[0076] In an embodiment of the present invention, information is extracted and standardized from electronic medical record data, so as to utilize natural language processing (NLP) technology to extract information from unstructured text data such as medical records, medical histories, and medication histories, and to extract key medical entities such as disease names, drug names, symptom descriptions, etc., through named entity recognition (NER) technology, and then these entities are associated with other patient data to construct standardized structured data. For structured data (such as basic information, medical records, medical histories, etc.), data cleaning technology is used to fill in missing values, correct erroneous data, and standardize all text fields (such as unifying the representation of drug names and standardizing disease classifications appearing in medical records). During the processing, data preprocessing tools (such as NLP libraries SpaCy or Stanford NLP, Pandas, etc.) are used to ensure data accuracy and consistency. All data are converted into a unified standard format (such as JSON or CSV), and finally standard medical record basic data is generated.

[0077] Further, as an embodiment of the present invention, refer to Figure 3 As shown, Figure 1 Detailed step flow diagram of step S2 in FIG. 1 , in this embodiment, step S2 includes the following steps:

[0078] Step S21: using a convolutional neural network to extract image lesion features from standard medical image data to extract gastroenterology medical image lesion features, and constructing a gastroenterology medical image feature extraction layer;

[0079] In an embodiment of the present invention, standard medical image data is processed to extract lesion features of gastroenterology medical images. The specific operation steps include using a convolutional neural network (CNN) to process the image data. The medical image data is standardized and preprocessed, such as normalization, denoising and data enhancement, to ensure the quality and consistency of the image data. The image data is feature extracted by using a pre-trained convolutional neural network (such as ResNet, VGG or Inception, etc.). The convolution layer in the network will automatically learn and extract low-level features (such as edges, textures, etc.) in the image. In the subsequent pooling layer, the feature dimension is reduced by dimensionality reduction operations, and significant features are highlighted. For gastroenterology images, special attention is paid to the features of related lesion areas such as the intestines and stomach. The feature extraction level is deepened by a multi-layer convolutional network. During the network training process, the back propagation algorithm is used to adjust the weights to minimize the feature extraction error. The extracted lesion features of gastroenterology medical images are used to construct a gastroenterology medical image feature extraction layer as input to the downstream decision model.

[0080] Step S22: extracting time series variation features of standard gene experiment time series data using a long short-term memory network to extract time series variation features of gastroenterology medical gene experiments, and constructing a gastroenterology medical gene time series feature extraction layer;

[0081] In an embodiment of the present invention, a long short-term memory network (LSTM) is used to extract time series features from standard genetic experiment time series data, so that the time series data is trained through an LSTM model. The LSTM network is particularly suitable for processing time series data and can capture time dependencies. In this step, the input of the LSTM model is standard genetic experiment data, and the output is the corresponding time series feature vector. To optimize the performance of the network, a bidirectional LSTM can be selected to extract forward and reverse time series features, and an appropriate number of layers and units are set to improve the expressive power of the model. During the model training process, back propagation is performed by contrasting loss functions (such as mean square error or cross entropy loss) to update the network weights. A time series feature extraction layer for genetic experiments is constructed and used to generate gastroenterology medical gene time series features.

[0082] Step S23: using a multi-layer perceptron corresponding to a deep neural network to perform multi-level nonlinear feature extraction on the basic data of the standard medical records, so as to extract nonlinear high-level features of the gastroenterology medical records, and construct a gastroenterology medical records feature extraction layer;

[0083] In an embodiment of the present invention, a multi-layer perceptron (MLP) in a deep neural network is used to extract nonlinear features of gastroenterology medical records data, so that these digitized medical record features are input into the multi-layer perceptron. The MLP is composed of multiple layers of fully connected neurons, and the output of each layer is used as the input of the next layer. The nonlinear features of different dimensions are learned layer by layer. During the training process of the model, an activation function (such as ReLU, Sigmoid or Tanh) is used to increase the nonlinear expression ability of the network. Between the neurons of each layer, the input signal is transmitted to the output layer through a weighted connection. At the same time, the back propagation algorithm is used to optimize the network parameters. After multiple training iterations, the nonlinear high-level features of the gastroenterology medical records are extracted, and a gastroenterology medical record feature extraction layer is constructed.

[0084] Step S24: Performing input layer training design on the gastroenterology medical image feature extraction layer, the gastroenterology medical gene time series feature extraction layer, and the gastroenterology medical medical record feature extraction layer to generate a decision model feature extraction input layer; performing feature vectorization processing on the gastroenterology medical image lesion features, gastroenterology medical gene experiment time series change features, and gastroenterology medical medical record nonlinear high-level features corresponding to the decision model feature extraction input layer through an embedded network layer to generate various types of gastroenterology medical feature vectors;

[0085] In an embodiment of the present invention, the previously constructed gastroenterology medical image feature extraction layer, gastroenterology medical gene temporal feature extraction layer and gastroenterology medical record feature extraction layer are simultaneously trained and designed to construct a decision model feature extraction input layer, wherein the gastroenterology medical image feature extraction layer is used as the first input layer, and the medical image (such as gastrointestinal endoscopy image, CT image, etc.) is subjected to feature extraction through a convolutional neural network (CNN) to obtain the features of the lesion area in the image, and the gastroenterology medical gene temporal feature extraction layer is used as the second input layer to extract the gene experimental data (such as gene expression spectrum, D NA sequencing data, etc.) for time series analysis, and extract the time series change characteristics related to the disease, and digest the internal medicine medical record feature extraction layer as the input third layer, so as to train and encode the high-level nonlinear features extracted from the medical record data through a deep neural network, and integrate the features of these three types of different sources through the design of cross-domain hierarchical connection reconstruction, eliminate the heterogeneity and information asymmetry between them, so that the model can simultaneously understand and integrate the correlation in imaging, gene and medical record data, and use a special hierarchical structure for docking to ensure seamless connection between features and avoid information loss or inconsistency, thereby generating a decision model feature extraction input layer. At the same time, multimodal feature input training is performed by combining gastroenterology medical image features, genetic experiment time series features and medical record features to integrate the various features extracted in the previous steps (medical image lesion features, genetic experiment time series change features, medical record nonlinear high-level features) into a feature vector as the training data of the input layer. In order to make the features more adaptable to the needs of the model, these features are vectorized through the embedding network layer. The embedding layer converts each feature into a vector representation of a fixed length through a lookup table to ensure that the sparsity of the feature information is effectively represented. All processed feature vectors will be input into the decision model. The input layer of the decision model will fuse these multimodal features to generate a unified input feature vector for intelligent decision-making. This step not only requires ensuring that various features can be effectively integrated, but also requires a reasonable design of the network structure to ensure that different modal data can cooperate with each other in the input layer to give play to their respective advantages, and finally generate various gastroenterology medical feature vectors.

[0086] Step S25: By introducing the attention mechanism fusion layer, various gastroenterology medical feature vectors are feature-fused, and the fully connected layer is used to jointly train the intelligent decision-making model to generate a gastroenterology intelligent medical decision-making model.

[0087] In an embodiment of the present invention, an attention mechanism layer is introduced to perform feature fusion, and then the intelligent decision-making model is trained. By introducing a self-attention mechanism, different weights are assigned to each input feature vector so as to highlight important features according to the needs of the task. Specifically, the attention mechanism assigns weights to each feature by calculating the similarity between the input features. Features with larger weights will have a greater impact on the decision results, and the weighted features are processed by using a fully connected layer. The training of the intelligent decision-making model is achieved through the combination of multiple layers of neurons. The output of the fully connected layer will be a decision result that integrates medical image features, gene time series features, and nonlinear features of medical records. The weighted fusion of these features can significantly improve the accuracy and robustness of the model in different clinical scenarios. All parameters of the network are optimized through the back propagation algorithm, so that the final intelligent decision-making model can effectively provide support for medical decision-making in the field of gastroenterology, and finally generate an intelligent medical decision-making model for gastroenterology.

[0088] Furthermore, the feature extraction input layer of the decision model in step S24 is reconstructed through a cross-domain hierarchical connection design to eliminate the heterogeneity and feature information asymmetry between the gastroenterology medical image lesion characteristics, the gastroenterology medical gene experiment temporal change characteristics and the gastroenterology medical medical record nonlinear high-level features, wherein the gastroenterology medical image feature extraction layer is used as the feature input first layer, the gastroenterology medical gene temporal feature extraction layer is used as the feature input second layer, and the gastroenterology medical medical record feature extraction layer is used as the feature input third layer.

[0089] Furthermore, the step S24 of performing feature vectorization processing on the gastroenterology medical image lesion features, gastroenterology medical gene experiment temporal change features and gastroenterology medical record nonlinear high-level features corresponding to the decision model feature extraction input layer through the embedded network layer includes the following steps:

[0090] Through the embedded network layer, multi-level feature extraction and vectorization are performed on the corresponding gastroenterology medical image lesion features in the decision model feature extraction input layer, from low-level edge features to high-level structural features, to generate gastroenterology medical image feature vectors, wherein the gastroenterology medical image feature vectors are specifically feature vectors formed by the aggregation of lesion area texture features, lesion area morphological features, and lesion area pixel distribution features;

[0091] In an embodiment of the present invention, hierarchical features are extracted from the gastroenterology medical image lesion features corresponding to the input layer of the decision model feature extraction by using a corresponding embedded network layer. The network gradually extracts low-level edge features of the image through multiple convolutional layers and pooling layers, and further extracts high-level structural features through a fully connected layer. Specifically, when processing gastroenterology medical images, the network first identifies the edge information in the image, such as the contour of a tumor or lesion area. This information constitutes the low-level features of the image, and analyzes the tissue structure in the image through the network to identify the morphological features of the lesion area, including its size, shape, edge smoothness, etc. The network also extracts pixel distribution features of the lesion area, such as gray value distribution, texture information, etc. All these features are integrated into a feature vector, in which the texture features, morphological features, and pixel distribution features of the lesion area are weighted and merged in a certain way to ultimately generate a gastroenterology medical image feature vector.

[0092] Preferably, the temporal change characteristics of the gastroenterology medical gene experiment corresponding to the decision model feature extraction input layer are dynamically extracted and vectorized by embedding the network layer to generate the gastroenterology medical gene experiment temporal feature vector, wherein the gastroenterology medical gene experiment temporal feature vector is specifically a feature vector formed by aggregating the activity of the gastroenterology medical gene at different time nodes and the experimental change trend characteristics;

[0093] In an embodiment of the present invention, the embedded network layer dynamically extracts and vectorizes the time series features of the gastroenterology medical gene experiment data through a time series analysis method. In the specific operation, a long short-term memory (LSTM) network or a gated recurrent unit (GRU) network is used. These networks can effectively capture the dynamic changes of the time series data. The gastroenterology medical gene experiment data usually includes the expression level and activity of the gene at different time nodes. The LSTM network takes the experimental data of the gene at different time points as input to capture the activity fluctuation of the gene and its changing trend. Then the network analyzes the gene expression data at different time nodes to extract the time series change characteristics therein, such as rising and falling trends and periodic changes. Through the training of the time series model, the gastroenterology medical gene experiment time series feature vector quantum can fully reflect the change law of the gene at different time nodes. These features will be aggregated and form the final feature vector quantum, and finally the gastroenterology medical gene experiment time series feature vector quantum is generated.

[0094] Preferably, the word embedding semantic feature extraction and vectorization are performed on the nonlinear high-level features of the gastroenterology medical records corresponding to the feature extraction input layer of the decision model through the embedding network layer, so as to extract the corresponding nonlinear and high-dimensional semantic features in the gastroenterology medical records text, and generate the gastroenterology medical records text semantic feature vector, wherein the gastroenterology medical records text semantic feature vector is specifically a feature vector formed by the aggregation of the semantic features of the medical record disease description, pathological changes, and diagnosis and treatment process;

[0095] In an embodiment of the present invention, by relying on natural language processing (NLP) technology, especially word embedding (Word Embedding) and deep learning models, the gastroenterology medical record text is preprocessed by using the BERT (Bidirectional Encoder Representations from Transformers) model. Specifically, the medical record text is first segmented and converted into a series of vocabulary units. The BERT model vectorizes each vocabulary based on the contextual semantic relationship, and can capture the nonlinear and high-dimensional semantic features in the text, and extract the core information in the medical record text, such as disease description, pathological changes, diagnosis and treatment process, etc. This information usually exists in multiple parts of the text. Through the multi-level processing of the BERT model, a vector representation containing semantic information can be extracted. By aggregating the semantic vectors of each level, the corresponding semantic feature vector is generated to represent the key features in the medical record, including the description of the disease, pathological changes and treatment process, etc., and finally the semantic feature vector of the gastroenterology medical record text is generated.

[0096] Preferably, gastroenterology medical image feature quanta, gastroenterology medical gene experiment time series feature quanta and gastroenterology medical medical record text semantic feature quanta are used as various types of gastroenterology medical feature quanta.

[0097] In an embodiment of the present invention, previously extracted gastroenterology medical image feature vector quanta, gastroenterology medical gene experiment time series feature vector quanta and gastroenterology medical medical record text semantic feature vector quanta are integrated as a unified feature vector quanta. Each feature vector quanta carries key information under this modality. The image feature vector quanta provides morphological and texture information of lesions, the gene experiment time series feature vector quanta provides gene activity and time change information, and the medical record text semantic feature vector quanta provides disease description and diagnosis and treatment process information, thereby integrating these multi-dimensional feature information into a unified feature vector, and finally obtaining various types of gastroenterology medical feature vectors.

[0098] Further, step S25 includes the following steps:

[0099] Step S251: Performing medical feature contribution analysis on various gastroenterology medical features to quantum by introducing an attention mechanism fusion layer to obtain feature decision contribution corresponding to various gastroenterology medical features;

[0100] In an embodiment of the present invention, by representing various types of gastroenterology medical features as a set of feature vectors, relevant features can usually be extracted from medical data (such as medical records, laboratory test results, imaging data, etc.) through data preprocessing. These features include but are not limited to texture features of lesion areas, morphological features of lesion areas, pixel distribution features of lesion areas, activity of medical genes at different time nodes, experimental change trend features, medical record disease descriptions, pathological changes, semantic features of diagnosis and treatment processes, etc. In order to perform feature contribution analysis, an attention mechanism fusion layer is introduced to dynamically weight the contribution of each feature vector. This layer can automatically adjust the contribution of different features to the final output of the model by learning the influence weight of each feature on the decision. In specific implementation, a network structure based on additive attention mechanism (AdditiveAttention) or multiplicative attention mechanism (Multiplicative Attention) can be adopted, various types of feature vectors are input, and the importance score of each feature is calculated through the attention mechanism. The score reflects the contribution of the feature in the current medical decision-making task, and finally the feature decision contribution corresponding to various types of gastroenterology medical features is obtained.

[0101] Step S252: performing feature contribution weighted fusion on various gastroenterology medical feature vectors based on the feature decision contribution corresponding to various gastroenterology medical features, so as to generate a gastroenterology medical global fusion feature vector;

[0102] In the embodiment of the present invention, each gastroenterology medical feature vector is weighted according to the previously obtained feature decision contribution. Specifically, for each feature vector, it is multiplied element by element with the corresponding decision contribution to obtain a weighted feature vector. The weighted feature vector can be fused by weighted averaging, or multiple features can be combined by concatenation or addition. Specifically, assuming there are n feature vectors X1, X2, ..., X n and the corresponding decision contributions w1,w2,...,w n , the result after weighted fusion is X fused for This operation can be implemented using matrix operations in the deep learning framework. At this time, the contribution of each feature vector will be adjusted according to its decision contribution to generate a global fusion feature vector, and finally a global fusion feature vector for gastroenterology medicine is generated.

[0103] Step S253: Obtain the clinical decision-making target of gastroenterology, and based on the clinical decision-making target of gastroenterology, use the fully connected layer to jointly train the intelligent decision-making model of gastroenterology global fusion feature vector to generate an intelligent medical decision-making model for gastroenterology.

[0104] In the embodiment of the present invention, the corresponding clinical decision-making goal of gastroenterology is obtained, and the global fusion feature vector is trained through the fully connected layer to generate an intelligent decision-making model. In this step, it is first necessary to clarify the clinical decision-making goal of gastroenterology, which is usually based on the patient's condition and characteristics to predict the occurrence, development or decision-making effect of a certain disease. For example, the goal is to determine whether the patient has gastric cancer, predict the decision-making effect of gastric ulcer, etc., and the previously obtained global fusion feature vector is used as input, and the global fusion feature vector is trained through the fully connected layer. The fully connected layer can perform nonlinear mapping of features through multiple neural network layers to learn the relationship between features and clinical decision-making goals. During the training process, a standard loss function (such as cross entropy loss function or mean square error loss function) is used to measure the gap between the model output and the actual decision-making goal, and the model parameters are optimized through the back propagation algorithm. Through the joint training process, the model can automatically adjust the weights of each layer to optimize the final medical decision output. The training process can use an optimization algorithm based on gradient descent, such as Adam optimizer or SGD optimizer. After the training is completed, the generated intelligent medical decision model can make real-time decisions based on new patient data, thereby assisting doctors to make more accurate medical decision plans, and finally generating an intelligent medical decision model for gastroenterology.

[0105] Further, step S3 includes the following steps:

[0106] Step S31: performing feature correlation analysis on each pair of features in each type of gastroenterology medical feature vector to generate a gastroenterology medical feature correlation measurement matrix;

[0107] In an embodiment of the present invention, various types of medical eigenvector quanta are extracted from a medical data set related to gastroenterology, and these eigenvector quanta contain information on different gastroenterology diseases or clinical indicators. For each pair of eigenvector quanta, the Pearson correlation coefficient, Spearman rank correlation coefficient or other correlation measurement methods are used to calculate the correlation value between each pair of features. In the specific operation process, a data analysis tool (such as the NumPy library and the SciPy library in Python, etc.) is used to perform numerical calculations on each pair of features, and a feature correlation measurement matrix is ​​generated. Each element in the matrix represents the correlation value between the corresponding features, and finally a gastroenterology medical feature correlation measurement matrix is ​​generated.

[0108] Step S32: performing high correlation feature clustering optimization on each gastroenterology medical sub-feature in each gastroenterology medical feature vector based on the gastroenterology medical feature correlation measurement matrix to obtain gastroenterology medical high correlation feature clustering clusters;

[0109] In an embodiment of the present invention, after the feature correlation measurement matrix is ​​generated, the gastroenterology medical features are clustered and optimized based on the matrix, so as to utilize clustering algorithms (such as K-means clustering, hierarchical clustering, etc.) to perform feature aggregation according to the correlation between features. The core goal of clustering is to classify features with strong correlation into the same group, thereby forming multiple high-correlation feature clustering clusters. In specific implementation, by setting a correlation threshold, the clustering algorithm will automatically adjust the number of clustering clusters according to the correlation strength of each pair of features. The KMeans function in the Scikit-learn library can be used to set appropriate clustering parameters (such as the number of clusters, distance metric, etc.) to complete the clustering process of the feature vector. The obtained clustering clusters represent different groups of gastroenterology medical features, each group of which contains highly correlated features, and finally obtain gastroenterology medical high-correlation feature clustering clusters.

[0110] Step S33: using the inter-feature dependency strength calculation formula to perform feature dependency measurement calculation between each gastroenterology medical sub-feature within the gastroenterology medical high correlation feature cluster, so as to obtain the feature decision dependency strength corresponding to each gastroenterology medical sub-feature;

[0111] In an embodiment of the present invention, a suitable inter-feature dependency strength calculation formula is formed by combining feature space, metric values ​​corresponding to gastroenterology medical sub-features, covariances corresponding between gastroenterology medical sub-features, standard deviations corresponding to gastroenterology medical sub-features, gastroenterology medical sub-feature dependency contribution coefficients, gastroenterology medical sub-feature distance attenuation rates, integral differentials and related parameters to perform feature dependency metric calculations on each gastroenterology medical sub-feature within a high-correlation feature cluster to quantify the dependency relationship between the features, and ultimately obtain the feature decision dependency strength corresponding to each gastroenterology medical sub-feature.

[0112] Step S34: performing dependency screening on each gastroenterology medical sub-feature in the gastroenterology medical high correlation feature cluster based on the feature decision dependency strength corresponding to each gastroenterology medical sub-feature, so as to obtain a gastroenterology medical optimized feature dependency subset;

[0113] In an embodiment of the present invention, after the feature dependency measurement is completed, the features in each feature cluster are dependency screened based on the dependency measurement value, so as to set a dependency threshold to screen out feature pairs with strong dependence and retain these features. In this process, the feature dependency can be screened by setting conditions, such as only retaining feature pairs with a dependency greater than 0.7 to reduce redundant features and improve model efficiency. The feature selection tools of Scikit-learn in Python (such as SelectKBest, VarianceThreshold, etc.) can be used to further optimize the feature subset to ensure that the final selected feature subset reduces the influence of irrelevant features while maintaining strong dependence, and finally obtains the optimized feature dependency subset for gastroenterology medicine.

[0114] Step S35: Perform structured description transformation on the gastroenterology medical optimized feature dependency subset to obtain a gastroenterology medical feature dependency descriptor.

[0115] In an embodiment of the present invention, after obtaining the optimized feature dependency subset, a structured description conversion is performed on it, so as to generate a standardized text or digital description by formatting the description of each feature subset. The purpose of the structured description conversion is to present the optimized feature dependency relationship in an interpretable form so as to further apply it to the training and reasoning of the decision model. In the implementation process, natural language processing (NLP) technology or database table structure can be used to convert the feature dependency relationship into an easy-to-understand description. For example, the pandas library in Python is used to convert the feature dependency matrix into a tabular form, or a custom encoding rule is used to generate a description in a specific format. The obtained feature dependency descriptor provides a clear and structured feature dependency relationship basis for the subsequent training of the intelligent decision model, and finally the gastroenterology medical feature dependency descriptor is obtained.

[0116] Further, step S31 includes the following steps:

[0117] Combining each gastroenterology medical feature sub-feature in each gastroenterology medical feature vector to generate a gastroenterology medical feature combination pair;

[0118] In an embodiment of the present invention, each gastroenterology medical sub-feature in each type of gastroenterology medical feature vector is combined with each other. In this step, all gastroenterology related features are first extracted from the gastroenterology medical feature library, such as texture features of lesion areas, morphological features of lesion areas, pixel distribution features of lesion areas, activity of medical genes at different time nodes, experimental change trend features, medical record disease descriptions, pathological changes, semantic features of diagnosis and treatment process, etc., and appropriate algorithms or programming tools (such as itertools.combinations in Python or other custom algorithms) are selected to combine feature pairs. Each pair of feature combinations should ensure its medical relevance and operability, for example, the relationship between liver function gene indicators and gastrointestinal diseases, the relationship between gastrointestinal symptoms and eating habits, etc. These combinations should cover multiple dimensions to comprehensively reflect the interactive relationship between different factors in gastroenterology, and finally generate gastroenterology medical feature combination pairs.

[0119] Preferably, the Pearson correlation coefficient is calculated between each pair of features of each gastroenterology medical feature combination pair to obtain the Pearson correlation coefficient between each gastroenterology medical feature pair;

[0120] In the embodiment of the present invention, the Pearson correlation coefficient is calculated between each pair of features of each gastroenterology medical feature combination. In this step, for each pair of generated feature combinations, its Pearson correlation coefficient needs to be calculated. The Pearson correlation coefficient is a statistical method for evaluating the strength and direction of the linear correlation between two variables, and its corresponding calculation formula is: Where X k and Y k are the values ​​of the two features, and is the mean of the feature. The Pearson correlation coefficient can be directly calculated using programming tools such as the scipy.stats.pearsonr method in Python. In this process, for each pair of feature combinations, it is necessary to check whether the data meets the calculation conditions of the Pearson correlation coefficient (such as whether the data is a continuous variable, whether it is normally distributed, etc.) to ensure the validity of the calculation results, and finally obtain the Pearson correlation coefficient between each pair of gastroenterology medical features.

[0121] Preferably, a correlation measurement matrix is ​​constructed for corresponding gastroenterology medical feature combination pairs based on the Pearson correlation coefficients between various gastroenterology medical feature pairs to generate a gastroenterology medical feature correlation measurement matrix.

[0122] In an embodiment of the present invention, a correlation measurement matrix is ​​constructed for the corresponding gastroenterology medical feature combinations by combining the previously quantified Pearson correlation coefficient, so that the Pearson correlation coefficient calculated between each pair of features is filled into the correlation measurement matrix, each row and each column of the matrix corresponds to a gastroenterology medical feature, and each element in the matrix represents the correlation measurement value between two features. When constructing the matrix, tools such as numpy or pandas can be used to store the results of the Pearson correlation coefficient matrix as a two-dimensional array or DataFrame format. This correlation measurement matrix is ​​a symmetric matrix, and the elements on its diagonal should be 1 (indicating the correlation between the feature itself). This matrix can help subsequent steps in establishing an intelligent decision-making model to identify which features have a strong correlation, thereby improving the training efficiency and accuracy of the model, and finally generating a gastroenterology medical feature correlation measurement matrix.

[0123] Furthermore, the calculation formula for the inter-feature dependency strength in step S33 is specifically:

[0124]

[0125] Where D ij is the feature decision dependency strength between the i-th gastroenterology sub-feature and the j-th gastroenterology sub-feature, X is the feature space corresponding to the gastroenterology high correlation feature cluster, x i is the i-th gastroenterology sub-feature, x j is the jth gastroenterology sub-feature, Cov(x i ,x j ) is the covariance between the i-th gastroenterology sub-feature and the j-th gastroenterology sub-feature, is the standard deviation corresponding to the i-th gastroenterology sub-feature, is the standard deviation corresponding to the j-th gastroenterology sub-feature, α is the dependency contribution coefficient of the gastroenterology sub-feature, exp is the exponential function, β is the distance attenuation rate of the gastroenterology sub-feature, dx is the integral differential, and η is the correction coefficient of the feature decision dependency strength.

[0126] The present invention obtains a calculation formula for the strength of dependency between features by using a specific mathematical model and after verification, which is used to perform feature dependency measurement calculation between each gastroenterology medical sub-feature in the gastroenterology medical high correlation feature clustering cluster. The calculation formula for the strength of dependency between features measures the linear relationship between features through covariance, and captures the correlation between different features. The larger the covariance, the stronger the correlation between features. The covariance between features helps to reveal the mutual influence of features, especially in the medical field, where different clinical features show obvious correlation in the decision-making process. The standard deviation in the formula makes the calculation result of the strength of dependency not affected by the dimension of the feature, which is very important for processing medical features with different dimensions or scales. For example, the dimensions of blood sugar level and body weight are different. The use of standard deviation ensures the scale invariance of the strength of dependency and ensures the fairness and comparability of the results. Secondly, the exponential decay term in the formula can better model the nonlinear relationship between features by attenuating the distance between features. In medical data, the relationship between features is often nonlinear. This term helps to capture complex nonlinear dependencies, especially in clinical decision-making. Many symptoms and indicators are not simply linearly related, but have more complex dependency patterns. By using the distance decay term, the model is very sensitive to local differences between features. When two features are very close, their dependence strength is large, which is crucial for accurately capturing subtle differences between features. For example, some medical features change slightly in some cases, but this slight change has important clinical significance. The adjustable parameters of the dependency measure provide space for model adjustment, allowing customized adjustments to the contribution of inter-feature dependency and the distance decay rate. These adjustments allow the model to be optimized according to specific gastroenterology medical data to better adapt to the feature distribution of different data sets. By calculating the feature decision dependency strength, strong dependencies between features can be identified. This information is crucial for feature screening in subsequent steps. Features with high dependencies may be retained, while features with low dependencies can be eliminated, thereby optimizing feature subsets, improving data processing efficiency, and reducing the interference of redundant features. This formula can not only help identify relevant features, but also reveal the dependency strength between features, which helps to generate a more reliable medical feature set and provide strong data support for clinical decision-making. In summary, this formula fully considers the feature decision dependency strength D corresponding to the i-th gastroenterology medical sub-feature and the j-th gastroenterology medical sub-feature. ij , the feature space X corresponding to the cluster of high correlation features of gastroenterology, the i-th gastroenterology sub-feature x i , the jth gastroenterology sub-feature x j , the covariance Cov(x between the i-th gastroenterology sub-feature and the j-th gastroenterology sub-feature i ,xj ), the standard deviation corresponding to the i-th gastroenterology sub-characteristic The standard deviation of the j-th gastroenterology sub-characteristic The gastroenterology sub-feature dependency contribution coefficient α, exponential function exp, gastroenterology sub-feature distance decay rate β, integral differential dx, feature decision dependency strength correction coefficient η, according to the feature decision dependency strength D corresponding to the i-th gastroenterology sub-feature and the j-th gastroenterology sub-feature ij The correlation between the above parameters constitutes a functional relationship This formula can realize the feature dependency measurement calculation process between each gastroenterology sub-feature within the gastroenterology high-correlation feature cluster. At the same time, by introducing the correction coefficient η of the feature decision dependency strength, it can be adjusted according to the errors occurring in the calculation process, thereby improving the accuracy and applicability of the formula for calculating the dependency strength between features.

[0127] Further, step S4 includes the following steps:

[0128] Step S41: reconstructing the temporal relationship of various medical features corresponding to the gastroenterology medical feature dependency descriptor to generate a gastroenterology medical feature temporal dependency association network;

[0129] In an embodiment of the present invention, various types of medical features corresponding to the dependency descriptors of gastroenterology medical features are previously extracted, including but not limited to the patient's lesion texture, lesion pixel distribution, gene activity, medical record disease description, pathological changes and other features, and these data are processed using a time series analysis method to redetermine the temporal relationship between the various medical features. For example, algorithms based on graph models, such as dynamic Bayesian networks (DBNs), temporal association rule mining (Temporal Association Rule Mining) and other technologies can be used to identify the dependencies of different medical features in the time dimension. Through these technologies, a temporal dependency network of gastroenterology medical features can be constructed, in which nodes represent medical features and edges represent temporal dependencies between features. In the process of constructing the temporal relationship, the causal order between features needs to be considered. For example, gastrointestinal symptoms may precede abnormalities in certain biochemical indicators, or changes in certain test results may affect subsequent diagnostic decisions. Finally, a temporal dependency association network of gastroenterology medical features is reconstructed and generated.

[0130] Step S42: simulating the medical decision path according to the gastroenterology medical feature temporal dependency association network to obtain a gastroenterology medical decision simulation evolution path diagram;

[0131] In an embodiment of the present invention, a simulation algorithm is used to simulate the clinical decision-making process by combining the previously generated gastroenterology medical feature temporal dependency association network, so as to simulate the decision path under different medical feature change conditions by constructing a decision model based on temporal dependency. This process can utilize a graph-based traversal algorithm (such as depth-first search, breadth-first search, etc.) to simulate all medical decision paths and generate a medical decision simulation evolution path diagram. This diagram helps clinical experts understand the entire decision-making process by presenting the key nodes (such as symptom changes, decision plans, etc.) and their evolution process on each decision path. The generation of each path not only reflects the causal relationship in the time series, but also reveals the possible results under different decision paths, and finally obtains a gastroenterology medical decision simulation evolution path diagram.

[0132] Step S43: performing decision optimization selection on each medical decision path in the gastroenterology medical decision simulation evolution path diagram to generate a gastroenterology medical optimization decision path;

[0133] In an embodiment of the present invention, by simulating the evolutionary path diagram of gastroenterology medical decisions previously generated, an optimization algorithm is applied to evaluate and optimize each decision path. This process can utilize a multi-objective optimization algorithm (such as a genetic algorithm, a particle swarm optimization algorithm, etc.) to weigh different decision paths and optimize the optimal decision plan. The optimization objectives include factors such as the shortest decision time, the lowest patient risk, and the best decision effect. By continuously adjusting the decision nodes in the path and their selection, a medically optimized decision path is generated. These optimized paths represent the decision-making process that can most effectively improve patient decision effects and reduce adverse reactions under given conditions, and ultimately generate a gastroenterology medically optimized decision path.

[0134] Step S44: performing deep cross-feedback update on the corresponding clinical medical decision in the gastroenterology intelligent medical decision model based on the gastroenterology medical optimization decision path to generate a gastroenterology intelligent decision feedback update model;

[0135] In an embodiment of the present invention, by combining the previously obtained gastroenterology medical optimization decision path, the gastroenterology intelligent medical decision model is cross-feedback updated through a deep learning algorithm. Specifically, by inputting the optimized decision path into the decision model, using a back propagation algorithm (such as a deep neural network training process), and based on the difference between the effect of each path and the actual clinical situation, error adjustment is performed to optimize the model's predictive ability. Through continuous training, the model can gradually capture more accurate decision patterns and optimization rules. In this process, each feedback update of the model will be corrected according to clinical data and actual decision results, so that it can better adapt to the clinical environment and improve the accuracy of future decisions, ultimately generating a gastroenterology intelligent decision feedback update model.

[0136] Step S45: Based on the gastroenterology intelligent decision-making feedback update model, the corresponding clinical medical decision in the gastroenterology intelligent medical decision-making model is updated by feedback loop reinforcement learning to generate a gastroenterology intelligent medical decision-making optimization model.

[0137] In an embodiment of the present invention, by combining the previously generated gastroenterology intelligent decision-making feedback update model, the gastroenterology intelligent medical decision-making model is further optimized by using a reinforcement learning method. Specifically, the policy gradient method or Q learning algorithm in reinforcement learning is used to dynamically adjust the decision path based on the patient's historical data and real-time decision feedback. Whenever the decision path generated by the model can bring better decision-making results or lower risks, positive feedback will be obtained, thereby adjusting the strategy to better adapt to clinical decision-making needs. Through long-term learning and feedback, the model is continuously optimized to form a gastroenterology intelligent medical decision optimization model. This model can not only improve the accuracy of decision-making, but also adjust the individual differences of patients through an adaptive mechanism, thereby achieving precise and personalized medical decision support, and finally generating a gastroenterology intelligent medical decision optimization model.

[0138] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A training method for an intelligent medical decision-making model based on gastroenterology, characterized in that: The following steps are involved: Step S1: Obtaining gastroenterology medical imaging data, gastroenterology gene experiment time series data, and gastroenterology medical record basic data from the gastroenterology medical information system and the electronic medical record system; performing modeling pre-preprocessing on the gastroenterology medical imaging data, gastroenterology gene experiment time series data, and gastroenterology medical record basic data, respectively, to generate standard medical imaging data, standard gene experiment time series data, and standard medical record basic data; Step S2: Perform input layer training design on standard medical imaging data, standard gene experiment time series data and standard medical record basic data to generate a decision model feature extraction input layer; perform feature vectorization processing on various medical features corresponding to the decision model feature extraction input layer through an embedded network layer to generate various gastroenterology medical feature vectors; perform feature fusion on various gastroenterology medical feature vectors through the introduction of an attention mechanism fusion layer, and use a fully connected layer to perform joint training on an intelligent decision model to generate a gastroenterology intelligent medical decision model; Step S3: performing feature dependency screening on various gastroenterology medical feature vectors to obtain gastroenterology medical feature dependency descriptors; Step S4: Based on the gastroenterology medical feature dependency descriptor, the gastroenterology intelligent medical decision-making model is updated by deep reinforcement learning to generate a gastroenterology intelligent medical decision-making optimization model.

2. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: obtaining gastroenterology medical imaging data and gastroenterology gene experiment time series data from the gastroenterology medical information system, wherein the gastroenterology medical imaging data include gastrointestinal CT images, MRI images, ultrasound examination images and digestive endoscopy images, and the gastroenterology gene experiment time series data include blood test time series data, liver function experiment time series data, kidney function experiment time series data, intestinal microbiome experiment time series data and digestive tract gene experiment time series data; Step S12: Obtaining basic medical record data of gastroenterology from the electronic medical record system, wherein the basic medical record data of gastroenterology includes basic information of the patient, diagnosis and treatment records, gastroenterology medical history, and gastroenterology medication history; Step S13: performing image data preprocessing on the gastroenterology medical image data, including cropping, rotating, enlarging, and contrast enhancement, to generate standard medical image data; Step S14: performing time series denoising and standardization processing on the gastroenterology gene experiment time series data to generate standard gene experiment time series data; Step S15: extract information data and perform standardization processing on the basic data of gastroenterology medical records to generate standard basic data of medical records.

3. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: using a convolutional neural network to extract image lesion features from standard medical image data to extract gastroenterology medical image lesion features, and constructing a gastroenterology medical image feature extraction layer; Step S22: extracting time series variation features of standard gene experiment time series data using a long short-term memory network to extract time series variation features of gastroenterology medical gene experiments, and constructing a gastroenterology medical gene time series feature extraction layer; Step S23: using a multi-layer perceptron corresponding to a deep neural network to perform multi-level nonlinear feature extraction on the basic data of the standard medical records, so as to extract nonlinear high-level features of the gastroenterology medical records, and construct a gastroenterology medical records feature extraction layer; Step S24: Performing input layer training design on the gastroenterology medical image feature extraction layer, the gastroenterology medical gene time series feature extraction layer, and the gastroenterology medical medical record feature extraction layer to generate a decision model feature extraction input layer; performing feature vectorization processing on the gastroenterology medical image lesion features, gastroenterology medical gene experiment time series change features, and gastroenterology medical medical record nonlinear high-level features corresponding to the decision model feature extraction input layer through an embedded network layer to generate various types of gastroenterology medical feature vectors; Step S25: By introducing the attention mechanism fusion layer, various gastroenterology medical feature vectors are feature-fused, and the fully connected layer is used to jointly train the intelligent decision-making model to generate a gastroenterology intelligent medical decision-making model.

4. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 3 is characterized in that: The decision model feature extraction input layer described in step S24 is reconstructed through a cross-domain hierarchical connection design to eliminate the heterogeneity and feature information asymmetry between the gastroenterology medical image lesion characteristics, the gastroenterology medical gene experiment time series change characteristics and the gastroenterology medical medical record nonlinear high-level features, wherein the gastroenterology medical image feature extraction layer is used as the feature input first layer, the gastroenterology medical gene time series feature extraction layer is used as the feature input second layer, and the gastroenterology medical medical record feature extraction layer is used as the feature input third layer.

5. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 3 is characterized in that: The step S24 includes the following steps of performing feature vectorization processing on the gastroenterology medical image lesion features, gastroenterology medical gene experiment temporal change features, and gastroenterology medical record nonlinear high-level features corresponding to the input layer of the decision model feature extraction through the embedded network layer: Through the embedded network layer, multi-level feature extraction and vectorization are performed on the corresponding gastroenterology medical image lesion features in the decision model feature extraction input layer, from low-level edge features to high-level structural features, to generate gastroenterology medical image feature vectors, wherein the gastroenterology medical image feature vectors are specifically feature vectors formed by the aggregation of lesion area texture features, lesion area morphological features, and lesion area pixel distribution features; Through the embedded network layer, the time series change characteristics of the gastroenterology medical gene experiment corresponding to the decision model feature extraction input layer are dynamically extracted and vectorized to generate the gastroenterology medical gene experiment time series feature vector, wherein the gastroenterology medical gene experiment time series feature vector is specifically a feature vector formed by the aggregation of the activity of the gastroenterology medical gene at different time nodes and the experimental change trend characteristics; Through the embedding network layer, the nonlinear high-level features of the gastroenterology medical records corresponding to the input layer of the decision model feature extraction are subjected to word embedding semantic feature extraction and vectorization, so as to extract the corresponding nonlinear and high-dimensional semantic features in the gastroenterology medical records text and generate the gastroenterology medical records text semantic feature vector, wherein the gastroenterology medical records text semantic feature vector is specifically a feature vector formed by the aggregation of the semantic features of the medical record disease description, pathological changes, and diagnosis and treatment process; The gastroenterology medical image feature quanta, the gastroenterology medical gene experiment time series feature quanta and the gastroenterology medical record text semantic feature quanta are used as various types of gastroenterology medical feature quanta.

6. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 3 is characterized in that: Step S25 includes the following steps: Step S251: Performing medical feature contribution analysis on various gastroenterology medical features to quantum by introducing an attention mechanism fusion layer to obtain feature decision contribution corresponding to various gastroenterology medical features; Step S252: performing feature contribution weighted fusion on various gastroenterology medical feature vectors based on the feature decision contribution corresponding to various gastroenterology medical features, so as to generate a gastroenterology medical global fusion feature vector; Step S253: Obtain the clinical decision-making target of gastroenterology, and based on the clinical decision-making target of gastroenterology, use the fully connected layer to jointly train the intelligent decision-making model of gastroenterology global fusion feature vector to generate an intelligent medical decision-making model for gastroenterology.

7. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: performing feature correlation analysis on each pair of features in each type of gastroenterology medical feature vector to generate a gastroenterology medical feature correlation measurement matrix; Step S32: performing high correlation feature clustering optimization on each gastroenterology medical sub-feature in each gastroenterology medical feature vector based on the gastroenterology medical feature correlation measurement matrix to obtain gastroenterology medical high correlation feature clustering clusters; Step S33: using the inter-feature dependency strength calculation formula to perform feature dependency measurement calculation between each gastroenterology medical sub-feature within the gastroenterology medical high correlation feature cluster, so as to obtain the feature decision dependency strength corresponding to each gastroenterology medical sub-feature; Step S34: performing dependency screening on each gastroenterology medical sub-feature in the gastroenterology medical high correlation feature cluster based on the feature decision dependency strength corresponding to each gastroenterology medical sub-feature, so as to obtain a gastroenterology medical optimized feature dependency subset; Step S35: Perform structured description transformation on the gastroenterology medical optimized feature dependency subset to obtain a gastroenterology medical feature dependency descriptor.

8. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 7, characterized in that: Step S31 includes the following steps: Combining each gastroenterology medical feature sub-feature in each gastroenterology medical feature vector to generate a gastroenterology medical feature combination pair; The Pearson correlation coefficient is calculated between each pair of features of each gastroenterology medical feature combination pair to obtain the Pearson correlation coefficient between each gastroenterology medical feature pair; Based on the Pearson correlation coefficients between each pair of gastroenterology medical features, a correlation measurement matrix is ​​constructed for the corresponding gastroenterology medical feature combination pairs to generate a gastroenterology medical feature correlation measurement matrix.

9. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 7, characterized in that: The calculation formula of the inter-feature dependency strength in step S33 is specifically: Where D ij is the feature decision dependency strength between the i-th gastroenterology sub-feature and the j-th gastroenterology sub-feature, X is the feature space corresponding to the gastroenterology high correlation feature cluster, x i is the i-th gastroenterology sub-feature, x j is the jth gastroenterology sub-feature, Cov(x i ,x j ) is the covariance between the i-th gastroenterology sub-feature and the j-th gastroenterology sub-feature, is the standard deviation corresponding to the i-th gastroenterology sub-feature, is the standard deviation corresponding to the jth gastroenterology sub-feature, α is the dependency contribution coefficient of the gastroenterology sub-feature, exp is the exponential function, β is the distance attenuation rate of the gastroenterology sub-feature, dx is the integral differential, and η is the correction coefficient of the feature decision dependency strength.

10. The training method of the intelligent medical decision-making model based on gastroenterology according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: reconstructing the temporal relationship of various medical features corresponding to the gastroenterology medical feature dependency descriptor to generate a gastroenterology medical feature temporal dependency association network; Step S42: simulating the medical decision path according to the gastroenterology medical feature temporal dependency association network to obtain a gastroenterology medical decision simulation evolution path diagram; Step S43: performing decision optimization selection on each medical decision path in the gastroenterology medical decision simulation evolution path diagram to generate a gastroenterology medical optimization decision path; Step S44: performing deep cross-feedback update on the corresponding clinical medical decision in the gastroenterology intelligent medical decision model based on the gastroenterology medical optimization decision path to generate a gastroenterology intelligent decision feedback update model; Step S45: Based on the gastroenterology intelligent decision-making feedback update model, the corresponding clinical medical decision in the gastroenterology intelligent medical decision-making model is updated by feedback loop reinforcement learning to generate a gastroenterology intelligent medical decision-making optimization model.

Images