Medical record intelligent agent system for carrying out intelligent auxiliary diagnosis and treatment on liver, gall and pancreas diseases by utilizing medical record information

Abstract

The invention discloses a medical record intelligent agent system for performing intelligent auxiliary diagnosis and treatment on liver, gall and pancreas diseases by using medical record information, relates to a medical record intelligent agent system, and belongs to the field of medical information processing. The invention aims to solve the problems that a system in the prior art is insufficient in medical record information mining, various types of data are not utilized well, and the whole reasoning process is invisible, so that the interpretability of the system is poor. The system comprises a data processing module, a numerical value type liver, gall and pancreas disease classification module, a natural language type liver, gall and pancreas disease classification module, an image examination data type liver, gall and pancreas disease classification module, an integration module and an evidence judgment module. The numerical type liver, gall and pancreas disease classification module is used for acquiring a numerical type liver, gall and pancreas disease classification result; the natural language type liver, gall and pancreas disease classification module is used for acquiring a natural language type liver, gall and pancreas disease classification result; the image examination data type liver, gall and pancreas disease classification module is used for obtaining an image examination type liver, gall and pancreas disease classification result.

Description

Technical Field

[0001] The present invention relates to a medical record intelligent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information, and belongs to the field of medical information processing. Background Art

[0002] With the rapid development of artificial intelligence technology, more and more medical fields have begun to explore the application of machine learning and deep learning technologies in disease prediction. At present, three methods, namely disease diagnosis based on symptom description, disease diagnosis based on tabular numerical data, and inference rules based on disease evolution, have become important research directions in the field of medical intelligence. These methods have unique advantages and application scenarios.

[0003] Disease diagnosis based on symptom description:

[0004] This method predicts the type of disease the patient may have by analyzing the symptom information provided by the patient. Its core technology is natural language processing, which uses machine learning algorithms to analyze a large amount of medical text data, learn the relationship model between symptoms and diseases, and thus predict the possible type of disease the patient may have. For example, by establishing a relatively complete symptom and disease knowledge base, and then using a model based on naive Bayesian related symptom matching and deep learning technology, good progress has been made in symptom diagnosis and treatment. Applying genetic algorithms to decision trees and Bayesian classification methods further mines the rich semantics of medical record information, and has achieved a good improvement in the accuracy of disease diagnosis. The advantage of this method is that it can quickly make a preliminary diagnosis through simple question answering, and can reduce the interference of subjective factors in the diagnosis process of doctors. However, this method requires a large amount of medical text data for training, and insufficient data can easily lead to low prediction accuracy.

[0005] Disease diagnosis based on tabular numerical data:

[0006] This method uses machine learning algorithms to analyze and model the tabular data in the medical record to predict the type of disease the patient may have. Given that tabular data contains a large amount of numerical information, which is different from natural language information such as symptom descriptions, a highly scalable unified format processing tool is required. Wei et al. used the CDSS architecture to study a universal case analysis method that supports the rapid analysis of different types of disease cases for case storage (Wei Hongpeng. Research and Implementation of Openness of Clinical Diagnosis and Treatment Decision Support System [D]. Xidian University, 2017.) By annotating the columns of tabular data and fully exploring the relationship between columns in the tabular data, the prediction accuracy in this field is improved. The advantage of this method is that it can make full use of the structured information in the medical record data to make a more comprehensive and accurate prediction of the patient's condition, while avoiding the interference of the doctor's subjective factors. The disadvantage of this method is that it requires the establishment of a complex data model and requires a large amount of medical record data for training. Insufficient data volume can easily lead to low prediction accuracy.

[0007] Inference rules based on disease evolution:

[0008] This method is based on the in-depth study of the disease evolution process in the medical field, and constructs inference rules to better predict the development trend and treatment plan of the disease. Cosnes et al. (Cosnes, Jacques, et al. "Long-term evolution of disease behavior of Crohn's disease." Inflammatory bowel diseases 8.4 (2002): 244-250.) retrospectively studied the information data of 2,002 Crohn's disease patients to find the occurrence and predictive factors of stenosis and / or penetrating complications. Finally, these inference methods are used to complete the automated reasoning process using finite state automata and other methods. The core of this method is expert knowledge and a comprehensive understanding of the patient's disease development process to establish a complete disease model and inference rules. The advantage of this method is that it can better understand the nature and evolution of the disease, thereby improving the treatment effect and survival rate of patients. However, the disadvantage of this method is that it requires a large amount of clinical data for training, and at the same time requires rich medical expert knowledge to guide the construction of rules.

[0009] Although each method has its own advantages and disadvantages, the combined application of these methods can maximize their advantages and reduce the impact of their disadvantages on disease prediction. For example, when predicting a disease, you can first use a disease diagnosis pre-training model based on symptom description to quickly make a preliminary diagnosis, and then use a pre-training model based on tabular data and inference rules based on disease evolution for further analysis and prediction. In this case, you can use the symptom information provided by the patient to quickly make a preliminary diagnosis, and then use more information such as medical record data and clinical experience to further diagnose and reason, thereby improving the accuracy of the prediction.

[0010] There are three main flaws in the current research methods. First, only one or two types of data in medical record information are used. Although there are many methods that can extract sufficient information in a specific type of data, such as chief complaint information, some structured data such as examination indicators are still discarded, resulting in limited application scenarios. Second, medical record information contains a large amount of personal information of patients, which is highly privacy-sensitive, but many current methods do not clearly indicate how to desensitize data privacy. Third, most current methods use deep learning technology, but the disadvantage is that the entire reasoning process is invisible. In sensitive fields such as medicine, doctors and patients often cannot believe in unfounded reasoning, which does not conform to the theory of evidence-based medicine, resulting in difficulties in the actual application and promotion of various intelligent diagnosis and treatment technologies. Summary of the invention

[0011] The purpose of the present invention is to solve the problems that the existing technical systems do not fully mine medical record information, do not make good use of various types of data, and the overall reasoning process is invisible, resulting in poor system interpretability. A medical record intelligent system is proposed that uses medical record information for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases.

[0012] A medical record intelligent agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information includes:

[0013] Data processing module, numerical type hepatobiliary and pancreatic disease classification module, natural language type hepatobiliary and pancreatic disease classification module, imaging examination data type hepatobiliary and pancreatic disease classification module, integration module, evidence judgment module;

[0014] The data processing module is used to collect hepatobiliary and pancreatic medical record data, unify the format of the collected hepatobiliary and pancreatic medical record data, perform privacy desensitization on the hepatobiliary and pancreatic medical record data after the unified format, and classify the privacy desensitized hepatobiliary and pancreatic medical record data;

[0015] The numerical type hepatobiliary and pancreatic disease classification module is used to process the numerical type data in the medical records processed by the data processing module to obtain the numerical type hepatobiliary and pancreatic disease classification results;

[0016] The natural language type hepatobiliary and pancreatic disease classification module is used to process the natural language type data in the medical records processed by the data processing module to obtain the natural language type hepatobiliary and pancreatic disease classification results;

[0017] The imaging examination data type hepatobiliary and pancreatic disease classification module is used to process the basic imaging examination type data in the medical record processed by the data processing module to obtain the imaging examination type hepatobiliary and pancreatic disease classification result;

[0018] The integration module is used to integrate the numerical type hepatobiliary and pancreatic disease classification results, the natural language type hepatobiliary and pancreatic disease classification results, and the imaging examination type hepatobiliary and pancreatic disease classification results to obtain an integrated classification result;

[0019] The evidence judgment module is used to judge whether the integrated classification result is consistent with the doctor's diagnosis opinion, and output supporting evidence if consistent; if inconsistent, output opposing evidence.

[0020] The beneficial effects of the present invention are:

[0021] While solving the problem of automated and intelligent diagnosis of common hepatobiliary and pancreatic diseases, the method proposed in the present invention will also focus on enhancing the interpretability and reliability of the entire intelligent diagnosis process, thereby making the entire intelligent diagnosis process transparent and the entire system decision-making process credible, truly helping doctors make final decisions.

[0022] The medical record agent is a medical auxiliary tool developed based on artificial intelligence technology to assist doctors and medical teams in disease diagnosis and treatment decision-making. The medical record agent of the present invention analyzes the patient's medical record data, including symptoms, medical history, physical signs and other information, and uses advanced machine learning algorithms and big data analysis technology to predict the patient's possible disease type and provide doctors with relevant diagnostic suggestions and treatment plans. The field of medical record agents is in a stage of rapid development, and current research focuses on improving the accuracy and comprehension of natural language processing technology in medical texts to extract key information more finely. Machine learning algorithms play an important role in assisting doctors in making diagnoses, formulating personalized treatment plans, and predicting patients' conditions.

[0023] It can truly assist doctors and medical teams in disease diagnosis and treatment decisions by analyzing patients' medical records, including symptoms, medical history, physical signs and other information, and using advanced machine learning algorithms and big data analysis technology to predict the patient's possible disease type and provide doctors with relevant diagnostic recommendations and corresponding diagnostic reasons and evidence.

[0024] The medical record agent method of the present invention has high accuracy and precision, can quickly process a large amount of medical record data, and give high-quality diagnosis and treatment suggestions. In addition, the medical record agent method of the present invention can also continuously update and optimize its own algorithm model at a low cost, and expand new functions or business requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the system of the present invention;

[0026] Figure 2 This is a schematic diagram of input and output of this method;

[0027] Figure 3 The figure is a specific application example of this method. DETAILED DESCRIPTION

[0028] Specific implementation method 1: In this implementation method, a medical record intelligent agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information includes:

[0029] Data processing module, numerical type hepatobiliary and pancreatic disease classification module, natural language type hepatobiliary and pancreatic disease classification module, imaging examination data type hepatobiliary and pancreatic disease classification module, integration module, evidence judgment module;

[0030] The data processing module is used to collect hepatobiliary and pancreatic related medical record data, unify the format of the collected hepatobiliary and pancreatic related medical record data, perform privacy desensitization on the hepatobiliary and pancreatic related medical record data after the unified format, and classify the privacy desensitized hepatobiliary and pancreatic related medical record data;

[0031] The numerical type hepatobiliary and pancreatic disease classification module is used to process the numerical type data in the medical records processed by the data processing module to obtain the numerical type hepatobiliary and pancreatic disease classification results;

[0032] The natural language type hepatobiliary and pancreatic disease classification module is used to process the natural language type data in the medical records processed by the data processing module to obtain the natural language type hepatobiliary and pancreatic disease classification results;

[0033] The imaging examination data type hepatobiliary and pancreatic disease classification module is used to process the basic imaging examination type data in the medical record processed by the data processing module to obtain the imaging examination type hepatobiliary and pancreatic disease classification result;

[0034] The integration module is used to integrate the numerical type hepatobiliary and pancreatic disease classification results, the natural language type hepatobiliary and pancreatic disease classification results, and the imaging examination type hepatobiliary and pancreatic disease classification results to obtain an integrated classification result;

[0035] The evidence judgment module is used to judge whether the integrated classification result is consistent with the doctor's diagnosis opinion, and output supporting evidence if consistent; if inconsistent, output opposing evidence.

[0036] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that: the data processing module is used to collect hepatobiliary and pancreatic related medical record data, unify the format of the collected hepatobiliary and pancreatic related medical record data, perform privacy desensitization on the hepatobiliary and pancreatic related medical record data after the unified format, and classify the privacy desensitized hepatobiliary and pancreatic related medical record data;

[0037] The specific process is:

[0038] 1) Collect medical records related to hepatobiliary and pancreatic diseases; the specific process is as follows:

[0039] The number of diseases in the hepatobiliary and pancreatic medical record dataset is K, and the medical record data corresponding to each disease accounts for no less than

[0040] The hepatobiliary and pancreatic medical record data set contains numerical data (examination indicators), natural language data (chief complaint information), and imaging examination data;

[0041] This method makes use of the currently emerging deep learning and machine learning technologies. One of the keys to using these technologies is data collection and processing. According to actual experimental experience, the corresponding medical record data for each disease should be more than 100, and these medical record data must be directly related to the patient's diagnosis of the corresponding hepatobiliary and pancreatic disease, rather than other historical information in the past. In addition, this method combines the chief complaint information, examination indicators and imaging examination results in the patient's medical records. Therefore, the medical record data must contain all the contents of these three items. In addition, since the results of model training are strongly affected by data distribution, there are two precautions that need to be emphasized.

[0042] 2) Unify the format of the collected hepatobiliary and pancreatic medical records; the specific process is:

[0043] The collected hepatobiliary and pancreatic medical records were aligned to the JSON format;

[0044] This is to prevent the final trained model from having obvious bias in the prediction results. Second, the distribution of natural language data in medical records in different hospitals or regions is slightly different, and the units or evaluation standards for laboratory test indicators are also different. Therefore, in actual use, if the user is specific to a specific hospital, it is best to use the hospital's real clinical data. If it is to be promoted to multiple hospitals, it is important to pay attention to combining patient data from multiple different hospitals in the data set to enhance the generalization of the system. The evaluation or measurement standards for inspection indicators must also be unified;

[0045] 3) Desensitize the hepatobiliary and pancreatic medical records after the unified format; the specific process is as follows:

[0046] Use natural language processing or text matching technology to scan and identify the hepatobiliary and pancreatic medical records in a unified format, find text containing personal identification information, and remove or replace it;

[0047] In the medical field, it is very important to protect the privacy and personal information of patients (Kayaalp, Mehmet. "Patientprivacy in the era of big data." Balkan medical journal 35.1 (2018): 8-17.). Therefore, when using medical record data for medical research or developing intelligent medical auxiliary tools, the present invention must take a series of measures to protect the privacy and personal information of patients. The following is the process of processing privacy information in medical record data in the present invention:

[0048] First, the present invention must identify and remove personally identifiable information contained in medical records, such as name, date of birth, phone number, address, etc.

[0049] For example, gender, race, religious belief, marital status, etc. Similarly, the present invention uses similar technology to scan and identify medical record data, and removes or replaces text containing sensitive information.

[0050] For example, in this method, the data usage is as follows: the number of medical records is 1,280, and the data information in the medical records that is not related to this method is desensitized and privacy protected. The name, region, and Chinese resident identity information are all hidden as meaningless text;

[0051] 4) Classify the privacy-desensitized hepatobiliary and pancreatic medical records; the specific process is as follows:

[0052] The types of hepatobiliary and pancreatic diseases were set, including gallbladder polyps, liver cysts, cholecystitis, gallstones, liver cancer, pancreatic cancer, bile duct stones, bile duct cysts, pancreatitis, and bile duct cancer;

[0053] Classify the medical records related to the liver, gallbladder and pancreas after privacy desensitization.

[0054] According to the common types of hepatobiliary and pancreatic diseases, there are 10 common hepatobiliary and pancreatic diseases, including gallbladder polyps, liver cysts, cholecystitis, gallbladder stones, liver cancer, pancreatic cancer, bile duct stones, bile duct cysts, pancreatitis and bile duct cancer. The case data of each different type of disease are classified and sorted according to the chief complaint information, numerical type information and imaging examination information in the medical record information.

[0055] The other steps and parameters are the same as those in the first embodiment.

[0056] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that: the numerical type hepatobiliary and pancreatic disease classification module is used to process the numerical type data in the medical records processed by the data processing module to obtain the numerical type hepatobiliary and pancreatic disease classification results; the specific process is:

[0057] 1) Process the numerical data (test report) in the medical records processed by the data processing module; the specific process is:

[0058] If the data in the medical record is of numeric type, it will not be processed;

[0059] If the data in the medical record is non-numeric data, convert the non-numeric data into numerical information, such as positive / negative into 1 / -1;

[0060] Classifying and arranging numerical data is a key task in medical record analysis, which usually involves various examination indicators. In special examinations, the output results of the model are not continuous values, but are presented in categories such as negative / positive. Therefore, in order to process these results in the model, it is necessary to establish rules based on prior conditions to convert them into numerical information, such as 1 / -1.

[0061] 2) The numerical type hepatobiliary and pancreatic disease classification model is the XGBoost model;

[0062] The Gradient Boosting Decision Trees (GBDT) method is an effective ensemble learning technique. In ensemble learning, multiple weak learners (usually decision trees) are combined to build a powerful model. The core idea of ​​gradient boosting is to iteratively train a series of weak learners, each of which is trained on the residual of the previous learner. The final prediction result is the weighted sum of all weak learners. In the selection of base learners, this method uses the XGBoost (eXtreme Gradient Boosting) model widely used in the industry as a reference, and uses decision trees as base learners by default. A decision tree is a recursive binary tree structure in which each node represents a judgment condition for a feature, and each leaf node represents a predicted value. XGBoost uses CART (Classification and Regression Trees) type decision trees.

[0063] 3) Train the numerical type hepatobiliary and pancreatic disease classification model to obtain a trained numerical type hepatobiliary and pancreatic disease classification model; the specific process is:

[0064] The input of the XGBoost model is 1) the numerical data in the processed medical records, and the output of the XGBoost model is the numerical hepatobiliary and pancreatic disease classification results;

[0065] The objective function is L t ; The loss function is the cross entropy loss function;

[0066] Obtain a trained numerical type hepatobiliary and pancreatic disease classification model;

[0067] 4) Obtain the medical record data related to the hepatobiliary and pancreatic function to be tested, and process it through the data processing module, and process the numerical data (test report) in the medical record processed by the data processing module to obtain the numerical data to be tested;

[0068] Based on the trained numerical type hepatobiliary and pancreatic disease classification model, the numerical type data to be tested is classified to obtain the numerical type hepatobiliary and pancreatic disease classification results.

[0069] The other steps and parameters are the same as those in the first or second embodiment.

[0070] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that: the objective function L t for:

[0071]

[0072] Where, L t represents the objective function of the t-th decision tree, n represents the number of data in a batch of updated data; y i Represents sample x i Tags, Represents the first t-1 decision trees for sample x i The prediction result is generally obtained by adding the prediction results of the first t-1 decision trees. t (x i ) represents the current t-th decision tree for sample x i The prediction results of Ω(f t ) represents the complexity of the t-th decision tree;

[0073] loss means for The cross entropy loss function can be written as:

[0074]

[0075] Where p(x i ) represents the true sample distribution, q(x i ) represents the sample distribution predicted by the model, and n refers to the number of data in a batch of updated data.

[0076] The complexity of the t-th decision tree Ω(f t ) represents the acquisition process as follows:

[0077] Ω(f t ) represents the complexity of the entire decision tree. Generally, a linear combination of L1 regularization term and L2 regularization term is used. The L1 regularization term is calculated as the sum of the absolute values ​​of the features. The L2 regularization term is calculated as the sum of the squares of the features.

[0078] For the tth decision tree f t , L1 regularization term Ω1(f t ) is calculated as follows:

[0079]

[0080] Among them, Q is the total number of leaf nodes of the t-th decision tree; w j is the weight of the jth leaf node; γ is a hyperparameter that controls the strength of regularization;

[0081] For the tth decision tree f t , L2 regularization term Ω2(f t ) is calculated as follows:

[0082]

[0083] Among them, Q is the total number of leaf nodes of the t-th decision tree; w j is the weight of the jth leaf node; λ is a hyperparameter that controls the strength of regularization;

[0084] The complexity of the t-th decision tree Ω(f t ) is a linear combination of L1 regularization term and L2 regularization term:

[0085]

[0086] In practical applications, the strength of regularization can be controlled by adjusting the hyperparameters γ and λ. Larger γ and λ will lead to stronger regularization, thereby limiting the growth of the tree and reducing the risk of overfitting.

[0087] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0088] Specific implementation method 5: This implementation method is different from any one of the specific implementation methods 1 to 4 in that the information gain entropy H(k) of different features k (output features of different layers in the model) obtained in the numerical type hepatobiliary and pancreatic disease classification model is calculated as follows:

[0089]

[0090] In the formula, Z represents the number of all decision trees;

[0091] C(t) represents the total number of non-leaf nodes in the t-th decision tree;

[0092] B(t) represents the number of all nodes in the t-th decision tree (including the total number of non-leaf nodes and the total number of leaf nodes in the t-th decision tree); O represents the Oth node in the t-th decision tree;

[0093] m represents whether the current node belongs to feature k, if it does, it is 1, if not, it is 0;

[0094] Loss(t,i,L) and Loss(t,i,R) represent the losses on the left and right nodes of the i-th non-leaf node in the t-th decision tree, respectively;

[0095] D(t,i,L) and D(t,i,R) are the second-order derivatives of Loss(t,i,L) and Loss(t,i,R) respectively;

[0096] b and γ are the hyperparameters of the regularization term, respectively;

[0097] The splitting situation I(k) of the nodes to which different features k are obtained in the numerical type hepatobiliary and pancreatic disease classification model belong in the entire decision tree; the calculation formula is shown as follows:

[0098]

[0099] Where B(t) represents the number of all nodes in the t-th decision tree (including the total number of non-leaf nodes and the total number of leaf nodes in the t-th decision tree);

[0100] Based on the information gain entropy H(k) and the splitting situation I(k), the importance of feature k for the numerical type hepatobiliary and pancreatic disease classification model is obtained. The calculation formula is shown as follows:

[0101] Importance(k)=αH(k)+(1-α)I(k)Importance(k)

[0102] In the formula, α represents the adjustment parameter;

[0103] H(k) is convenient for finding the most important features for decision trees, but it often gives very low calculation results for variables with enumeration properties, leading to them being ignored. However, in medical diagnosis, the patient's examination indicators often have enumeration data, for example, a certain indicator is positive or weakly positive, etc.

[0104] Therefore, the present invention introduces the I(k) index, which is more friendly to the enumeration features and balances the influence relationship between the two by adjusting the parameter α.

[0105] Of course, different diseases will get different feature rankings during training. Depending on the patient's input, some features may not be input, and some features may be amplified. Therefore, although the model is static, the important features found in the end will also be different due to different inputs.

[0106] The reasoning of decision trees is an invisible black box process. In order to make the reasoning process of the medical record agent transparent and the conclusion more interpretable, this paper proposes a new method for calculating the contribution gain of decision tree nodes. The changes in the loss function and the changes in node splitting during the training phase are comprehensively considered to calculate the most important feature nodes for model decision making. The features extracted from the patient data of these nodes are used as the decision basis of the model.

[0107] The other steps and parameters are the same as those in Specific Implementation 1 to 4-1.

[0108] Specific implementation method 6: This implementation method is different from the specific implementation methods 1 to 5 in that: the natural language type hepatobiliary and pancreatic disease classification module is used to process the natural language type data in the medical records processed by the data processing module to obtain the natural language type hepatobiliary and pancreatic disease classification results; the specific process is:

[0109] 1) The natural language type hepatobiliary and pancreatic disease classification model is the Bert-chinese model;

[0110] 2) Train the natural language type hepatobiliary and pancreatic disease classification model to obtain a trained natural language type hepatobiliary and pancreatic disease classification model; the specific process is:

[0111] The input of the natural language type hepatobiliary and pancreatic disease classification model is the natural language type data in the medical record processed by the data processing module (the natural language type data (chief complaint information) in the medical record is processed based on the BERT model to obtain the processed natural language type data;), and the output of the natural language type hepatobiliary and pancreatic disease classification model is the hepatobiliary and pancreatic disease type;

[0112] The loss function is the cross entropy loss function;

[0113] Obtain a trained natural language type hepatobiliary and pancreatic disease classification model;

[0114] 3) Obtain the hepatobiliary and pancreatic related medical record data to be tested, process it through the data processing module, and process the natural language type data (chief complaint information) in the medical record processed by the data processing module to obtain the natural language type data to be tested;

[0115] Based on the trained numerical type hepatobiliary and pancreatic disease classification model, the natural language type data to be tested is classified to obtain the natural language type hepatobiliary and pancreatic disease classification results.

[0116] The medical record information data is processed into training data and validation set data of the classification task type. The bidirectional language model based on the Chinese pre-trained language model of Canchian in this method is the most well-known open source model Bert-chinese in the industry. BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained model based on the Transformer architecture, proposed by Google in 2018. One of the main innovations of BERT is the use of bidirectional contextual information to pre-train the model, which has achieved significant performance improvements in multiple natural language processing tasks, including text classification. In the pre-training stage, BERT learns language representations through unsupervised learning on a large-scale text corpus. The model structure is as follows: Figure 2 Bert-chinese replaces the Chinese vocabulary based on Bert and supplements the training of Chinese full-text MASK.

[0117] BERT uses the Transformer architecture with multiple self-attention mechanism layers. These layers allow the model to consider all positions in the input sequence simultaneously, not just individual words. This helps the model capture global contextual information. For text classification tasks, the input is one or more text sequences. A special "[CLS]" marker is added to the beginning of each text sequence to indicate that this is a classification task. Each word in the sentence is embedded as a vector. During the fine-tuning phase, the output layer of BERT is usually modified to adapt to a specific classification task. The output layer usually includes a fully connected layer and a softmax activation function to predict the category of the text sequence. The loss function generally uses the cross entropy loss function.

[0118]

[0119] Where p(x i ) represents the true sample distribution, q(x i ) represents the sample distribution predicted by the model, and n refers to the number of data in a batch of updated data.

[0120] The hyperparameters that need to be set in advance for model training are shown in Table 1. These parameters are usually related to neural network training in machine learning, especially when using deep learning models for classification tasks. The following is an explanation of these parameters:

[0121] ①hidden_dropout_prob

[0122] Significance: Control the dropout probability of the hidden layer in the neural network

[0123] Explanation: Dropout is a regularization technique used to prevent neural networks from overfitting. hidden_dropout_prob specifies the probability of randomly dropping hidden layer neurons at each training step. During training, these neurons will be randomly ignored with the specified probability, which helps improve the generalization ability of the model.

[0124] ②num_labels

[0125] Significance: Defines the number of categories in a classification task.

[0126] Explanation: In classification problems, num_labels indicates how many different categories need to be distinguished by the model. In this example, the model is designed to handle a classification problem with two categories.

[0127] ③learning_rate

[0128] Meaning: Learning rate, used to control the update amplitude of model weights in each training iteration.

[0129] Explanation: The learning rate is an important hyperparameter that determines the step size of the model's weight update in each training iteration. A smaller learning rate may lead to slower convergence but more stable; a larger learning rate may lead to faster convergence but may be unstable. learning_rate=2e-5 means that the learning rate is 2 times 10 to the negative fifth power.

[0130] ④weight_decay

[0131] Meaning: L2 regularization weight.

[0132] Explanation: Weight decay is a method to prevent overfitting by adding a regularization term to the loss function. The weight_decay parameter controls the weight of the regularization term, which penalizes the model's weights and encourages the model to use smaller weight values.

[0133] ⑤epochs=5:

[0134] Meaning: The number of training iterations.

[0135] Explanation: epochs specifies the number of times the entire training dataset is used by the model for learning. In each epoch, the model performs a forward propagation and a backward propagation through the entire training dataset. Increasing the number of epochs helps the model learn the patterns of the data better, but may increase training time.

[0136] ⑥batch_size

[0137] Meaning: The number of samples used in each training iteration.

[0138] Explanation: The training dataset is divided into batches, each batch contains a certain number of samples. batch_size specifies the number of samples in each batch. Larger batches can increase training speed, but may require more memory.

[0139] Table 1 Bert model hyperparameters list

[0140]

[0141]

[0142] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5-1.

[0143] Specific implementation method 7: This implementation method is different from any one of the specific implementation methods 1 to 6 in that the calculation formula of the dot product attention in the Bert-chinese model is:

[0144]

[0145] Among them, Dot(x) represents dot product attention;

[0146] Q represents the output of the input text X after a linear layer with a shape of [Embed,Embed]; Embed represents the embedding length;

[0147] K represents the output of the input text X after a linear layer with a shape of [Embed,Embed];

[0148] Q and K represent the output of the input X after two different linear layers with shapes [Embed, Embed];

[0149] Q×K T The result is a matrix of [Sequence_length, Sequence_length], which can be regarded as a weight matrix of the influence of each token in the input on the output. The i-th row represents the influence factor of the i-th token in the sentence on the result; Sequence_length is the length of the input text X; d represents the embedding dimension; the superscript T represents the transposition;

[0150] Set a threshold T′, extract the text to which all tokens greater than the threshold T′ belong in the input, and use it as the basis for the Bert-chinese model decision (the decision result of the Bert-chinese model is determined by the extracted sentences);

[0151] Token represents a word token.

[0152] Obviously, the selection of threshold T is very important for the extraction of diagnostic evidence.

[0153] During model inference, the input is converted into a Tensor vector of shape [Batch_size, Sequence_length, Embed]. Batch_size represents the number of samples in a batch, Sequence_length is the length of the input text, and Embed is an embedding length defined during model training.

[0154] The reasoning of the Bert model is also an invisible black box process. In order to make the reasoning process of the medical record agent transparent and the conclusion more interpretable, this paper proposes a new interpretability method based on dot product attention and secondary verification analysis.

[0155] Therefore, the present invention uses a two-stage verification process, taking the extracted sentences as new inputs and letting the model re-judge. If the results are consistent with the previous inputs, then it can be considered that these sentences are sufficient basis for model diagnosis. If the model output is inconsistent with the previous one, it means that there is missing information. The present invention appropriately enlarges the threshold T and loops to obtain the final result.

[0156] The other steps and parameters are the same as those in Specific Implementations 1 to 6-1.

[0157] Specific embodiment eight: This embodiment differs from specific embodiments one to seven in that: the imaging examination data type hepatobiliary and pancreatic disease classification module is used to process the imaging examination type data in the medical record processed by the data processing module to obtain the imaging examination type hepatobiliary and pancreatic disease classification result; the specific process is:

[0158] Process the imaging examination data in the medical records into regular expressions to obtain the classification results of hepatobiliary and pancreatic diseases in the form of regular expressions;

[0159] Extracting specific disease diagnosis results from medical imaging results usually involves the extraction and analysis of text information. Regular matching and finite state automaton (FSA) are one possible approach, but they need to be customized according to the specific text structure and pattern. Analyze the text structure of medical imaging results to understand the general format and pattern of different disease diagnosis results. This may involve regular matching patterns, such as keywords, phrases, or formats for specific diagnoses. Use regular expressions to match key information of disease diagnosis results. This may include information such as disease name, degree, and location. Regular expressions can be used to capture patterns of this information. If the structure of medical imaging results is more complex, finite state automaton can be used to capture specific text patterns more flexibly. Finite state automaton is a formal tool that can describe various patterns of strings. Below is a demonstration of regular expression design.

[0160] 1) Extract disease name:

[0161] Regular expression: (liver cancer|liver cyst|gallbladder polyp|cholecystitis|tumor)

[0162] 2) Extract the lesion location:

[0163] Regular expression: (left side | right side | both sides)

[0164] 3) Extract the extent of lesions:

[0165] Regular expression: (mild|moderate|severe)

[0166] 4) Extract keywords from image descriptions:

[0167] Regular expression: (image displayed|discovered|observed|visible)

[0168] 5) Extract digital information (such as size, density, etc.):

[0169] Regular expression: (\d+(\.\d+)?)\s*(cm|mm|cm 2 |mm 2 )

[0170] (\d+(\.\d+)?) This part is used to match the numeric part, which may be an integer or a decimal. \d+: matches one or more digits, that is, the integer part. For example, 123 or 45 can both be matched. (\.\d+)?: matches the optional decimal part. Specifically, \.: matches the decimal point. \d+: matches one or more digits, that is, the digits in the decimal part. ()?: indicates that the decimal point and the decimal part are optional, that is, the decimal part may or may not exist. If there is no decimal point, the integer part can also be matched alone. Therefore, this subexpression (\d+(\.\d+)?) can match the following situations: integers: such as 123; numbers with decimals: such as 45.67; separate decimals: such as 0.5; \s* This part is used to match the blank characters between the number and the unit. \s: matches any blank character, including space, tab, newline, etc. *: represents zero or more blank characters. So it can be matched even if there is no space between the number and the unit; cm|mm|cm 2 |mm 2 In Chinese, | means or.

[0171] The other steps and parameters are the same as those in Specific Implementation 1 to 7-1.

[0172] Specific embodiment 9: This embodiment differs from any one of specific embodiments 1 to 8 in that: the integration module is used to integrate the numerical type hepatobiliary and pancreatic disease classification results, the natural language type hepatobiliary and pancreatic disease classification results, and the imaging examination type hepatobiliary and pancreatic disease classification results to obtain an integrated classification result;

[0173] The specific process is:

[0174] 1) The integrated module classification model is a multi-layer perceptron MLP;

[0175] The working principle of the multi-layer perceptron is as follows:

[0176] MLP(X)=WX+B

[0177] Among them, X is the input of the multilayer perceptron, and W and B are the parameters to be trained.

[0178] 2) Train the integrated module classification model to obtain a trained integrated module classification model; the specific process is:

[0179] The input of the integrated module classification model is the predicted label and predicted probability. The predicted label is the numerical type hepatobiliary and pancreatic disease classification result, the natural language type hepatobiliary and pancreatic disease classification result, and the imaging type hepatobiliary and pancreatic disease classification result. The predicted probability refers to the probability that the current sample belongs to a certain type of hepatobiliary and pancreatic disease.

[0180] The output of the integrated module classification model is the integrated hepatobiliary and pancreatic disease type;

[0181] The loss function is the cross entropy loss function;

[0182] Obtain a trained integrated module classification model;

[0183] 3) Based on the trained integrated module classification model, the numerical type hepatobiliary and pancreatic disease classification results, natural language type hepatobiliary and pancreatic disease classification results and imaging examination type hepatobiliary and pancreatic disease classification results are processed to obtain an integrated classification result.

[0184] The integration of the results of the above three parts needs to be carried out according to professional medical standard guidelines. The following Table 2 lists the commonly used diagnostic methods for related hepatobiliary and pancreatic diseases used in the experiment of this method. Specifically, for example, gallbladder polyps can be diagnosed by basically only using imaging examination results during the examination. Therefore, the diagnosis of this disease needs to focus on the results of step four, and the results of step three are used as an auxiliary, while the results of step two should not occupy too important a position.

[0185] From Table 2, the present invention also notes that some diseases require multi-dimensional diagnosis as a basis. For example, the numerical data processed in step 2 of the present invention (in reality, the test indicators) and the analysis of the imaging examination results in step 4 may be very important for the diagnosis of a certain disease. If the conclusions of these two steps are consistent, then the overall result can be directly obtained. If the conclusions are inconsistent, then the conclusions of the two need to be screened and merged.

[0186] For this purpose, the present invention trains a perceptron MLP, and when the conclusions of the above steps conflict, the perceptron is used to screen and merge the conclusions. The working principle of the perceptron is as follows.

[0187] MLP(X)=WX+B

[0188] To a certain extent, it is equivalent to learning a more fine-grained distinction between diagnostic results of different dimensions in specific diseases in the training samples.

[0189] The input X consists of two parts, the predicted label and the predicted probability. The predicted label is the disease prediction result obtained in step 2, step 3 and step 4. The predicted probability refers to the probability that the current sample belongs to this disease. The specific method of obtaining it is as follows:

[0190] The XGBoost model in step 2 uses the SoftMax formula to calculate the probability corresponding to each classification label in the multi-classification task, and outputs the label with the highest probability as the result. The method of obtaining does not require explicit secondary calculations. For example, in the Python library, it can be directly obtained through the XGBoost.predict_proba() method;

[0191] The bert-chinese model in step 3 also outputs probabilities according to the SoftMax formula. First, get the corresponding vector of the position of the [CLS] tag in the BERT input in the last layer of the model, and call the formula for calculation. torch.softmax(logits, dim=-1), and take the largest probability as the predicted probability of this step.

[0192] Table 2 Prediction probability of different hepatobiliary and pancreatic diseases

[0193]

[0194] Step 4 does not involve the participation of deep models. In this method, the human prediction probabilities of different diseases are given based on medical clinical diagnosis, as shown in Table 3 and actual experience, as shown in Table 2 above.

[0195] To a certain extent, it is equivalent to learning a more fine-grained distinction between diagnostic results of different dimensions in specific diseases in the training samples.

[0196] Table 3 Diagnostic basis for common hepatobiliary and pancreatic diseases

[0197]

[0198]

[0199] The other steps and parameters are the same as those in Specific Implementation 1 to 8-1.

[0200] Specific implementation method 10: This implementation method is different from the specific implementation methods 1 to 9 in that: the evidence judgment module is used to judge whether the integrated classification result is consistent with the doctor's diagnosis opinion, and outputs supporting evidence if they are consistent; if they are inconsistent, outputs opposing evidence; the specific process is:

[0201] If the integrated classification result obtained by the integrated module is consistent with that of the doctor, then the integrated classification result obtained by the integrated module is supporting evidence;

[0202] If the integrated classification result obtained by the integration module is inconsistent with that of the doctor, the integrated classification result obtained by the integration module is dissenting evidence.

[0203] If the final diagnosis of the medical record agent is consistent with that of the doctor, then the medical record information that the medical record agent relies on for decision-making will become supporting evidence. These evidences are divided into three parts. The first is based on some key examination indicators obtained in step two. These indicators are often key evidence for disease diagnosis in medicine. These data can be obtained based on the nodes in the decision tree in step two. In the decision tree, each node represents the corresponding examination result. The more critical the node, the higher the information gain and the smaller the depth in the decision tree, which means that the examination indicator is more important. Based on this, a series of important examination indicators can be output as reasoning evidence. The second is the relevant description of the chief complaint obtained in step three, such as disease symptoms, patient physical information, etc. The third is the relevant information obtained from the imaging examination report obtained in step four. Different diseases may rely on different imaging examinations such as enhanced CT, MRI or color ultrasound. The results of imaging examinations are often an important basis for distinguishing some similar diseases in medicine. The relevant report information located in the imaging results can be used as the result to become the basis for the diagnosis of the medical record agent. The input and output examples of this method are as follows. Figure 2 shown.

[0204] In addition, the medical record agent should also provide missing evidence based on the diagnostic guidance for different diseases in step 5. For example, if the patient is diagnosed with pancreatitis after only a CT scan, the medical record agent will provide feedback to the doctor based on the guidance manual in step 5 that the relevant test results of serum amylase and lipase are missing as the basis for diagnosis.

[0205] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 9.

[0206] The evaluation index of the present invention uses precision, recall and F1 value. Precision, recall and F1 value are commonly used indicators for evaluating model performance in classification problems. Precision refers to the proportion of samples that are actually positive in all samples predicted by the model as positive. Precision measures the accuracy of the model when predicting as positive, that is, how many positive examples predicted by the model are truly positive. Recall refers to the proportion of samples that are correctly predicted as positive by the model in all actual positive samples. Recall measures the coverage of the model for positive samples, that is, how many positive examples are successfully predicted. The F1 value is the harmonic mean of precision and recall, which comprehensively considers the accuracy and coverage of the model. The range of F1 value is between 0 and 1, and the closer to 1, the better the overall performance of the model. The F1 value is suitable for balancing precision and recall in different problems.

[0207] Other commonly used machine learning methods for disease diagnosis include the Naive Bayes method and the classifier method based on the RoBERTa (Robustly optimized BERT approach) model.

[0208] The Naive Bayes method is a probabilistic statistical classification algorithm based on Bayes' theorem, which is often used in applications such as text classification. The basic idea is to classify by calculating the probability of input features under the condition of a given category. Based on the training data, the prior probability (c) of each category and the conditional probability P(x|c) of each feature under each category are calculated. For a given input feature vector x, the posterior probability of each category is calculated using Bayes' theorem.

[0209]

[0210] P(x) is a normalization factor, summed over all categories so that the sum of the posterior probabilities is 1. The category with the largest posterior probability is selected as the prediction result.

[0211] RoBERTa (Robustly optimized BERT approach) is a natural language processing (NLP) model proposed by Facebook AI Research in 2019. It is an improvement and optimization based on BERT (Bidirectional Encoder Representations from Transformers). The design goal of RoBERTa is to improve the performance of pre-trained models and solve some problems in BERT. The improvements of RoBERTa have enabled it to achieve excellent performance on multiple natural language processing tasks and become one of the most advanced pre-trained language models at the time. Since RoBERTa is an improvement on BERT, it still uses the Transformer architecture and generates context-sensitive word vectors, which makes it perform well in various classification tasks.

[0212] Table 4 Comparison of experimental results with common disease prediction and classification models

[0213]

[0214] From Table 3 above, it can be seen that the method of the present invention is better than some commonly used machine learning and deep learning methods in various classification indicators, which can reflect the accuracy and reliability of the method of the present invention in intelligent disease auxiliary diagnosis. In addition, Table 4 below shows the classification results and indicators of 10 common diseases in hepatobiliary and pancreatic subjects by the method of the present invention.

[0215] Table 5 Classification results and indicators of 10 common hepatobiliary and pancreatic diseases using this method

[0216]

[0217] In addition to its excellent performance in disease prediction and classification tasks, an important feature that distinguishes the method of the present invention from other intelligent auxiliary diagnosis and treatment methods is that the method of the present invention can provide a chain of evidence for the entire disease inference process. Other existing disease prediction models based on machine learning or deep learning are basically black box reasoning processes. Without good interpretability, the present invention cannot explore the reasoning process and basis for diagnosing diseases within the model. When there is a disagreement between the model prediction and the diagnosis results of human experts, other methods often cannot provide explanations or evidence, making it difficult for patients and doctors to believe the model's prediction results. This also causes other methods to encounter difficulties in actual application scenarios, and it is difficult to really help the entire medical consultation process.

[0218] In addition to making full use of multi-source information such as the chief complaint, examination indicators and imaging results in the medical record to improve the accuracy of disease prediction, the method of the present invention also provides sufficient argumentation basis and diagnostic evidence. In this way, doctors can choose whether to trust the prediction results of the model by checking and scrutinizing the diagnostic evidence output by the model, and it can also provide doctors with new ideas and discoveries in the diagnosis process, truly allowing artificial intelligence to serve medical scenarios. Demonstration as follows Figure 3 shown.

[0219] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information, characterized by: The system comprises: Data processing module, numerical type hepatobiliary and pancreatic disease classification module, natural language type hepatobiliary and pancreatic disease classification module, imaging examination data type hepatobiliary and pancreatic disease classification module, integration module, evidence judgment module; The data processing module is used to collect hepatobiliary and pancreatic medical record data, unify the format of the collected hepatobiliary and pancreatic medical record data, perform privacy desensitization on the hepatobiliary and pancreatic medical record data after the unified format, and classify the privacy desensitized hepatobiliary and pancreatic medical record data; The numerical type hepatobiliary and pancreatic disease classification module is used to process the numerical type data in the medical records processed by the data processing module to obtain the numerical type hepatobiliary and pancreatic disease classification results; The natural language type hepatobiliary and pancreatic disease classification module is used to process the natural language type data in the medical records processed by the data processing module to obtain the natural language type hepatobiliary and pancreatic disease classification results; The imaging examination data type hepatobiliary and pancreatic disease classification module is used to process the basic imaging examination type data in the medical record processed by the data processing module to obtain the imaging examination type hepatobiliary and pancreatic disease classification result; The integration module is used to integrate the numerical type hepatobiliary and pancreatic disease classification results, the natural language type hepatobiliary and pancreatic disease classification results, and the imaging examination type hepatobiliary and pancreatic disease classification results to obtain an integrated classification result; The evidence judgment module is used to judge whether the integrated classification result is consistent with the doctor's diagnosis opinion, and output supporting evidence if consistent; if inconsistent, output opposing evidence.

2. According to claim 1, a medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information is characterized by: The data processing module is used to collect hepatobiliary and pancreatic medical record data, unify the format of the collected hepatobiliary and pancreatic medical record data, perform privacy desensitization on the hepatobiliary and pancreatic medical record data after the unified format, and classify the privacy desensitized hepatobiliary and pancreatic medical record data; The specific process is: 1) Collect hepatobiliary and pancreatic medical records; the specific process is: The number of diseases in the hepatobiliary and pancreatic medical record dataset is K, and the medical record data corresponding to each disease accounts for no less than The hepatobiliary and pancreatic medical record dataset contains numerical data, natural language data, and imaging examination data; 2) Unify the format of the collected hepatobiliary and pancreatic medical records data; the specific process is: The collected hepatobiliary and pancreatic medical records data were aligned to the JSON format; 3) Desensitize the hepatobiliary and pancreatic medical records after the unified format; 4) Classify the privacy-desensitized hepatobiliary and pancreatic medical records data; the specific process is as follows: The types of hepatobiliary and pancreatic diseases were set, including gallbladder polyps, liver cysts, cholecystitis, gallstones, liver cancer, pancreatic cancer, bile duct stones, bile duct cysts, pancreatitis, and bile duct cancer; Classify the hepatobiliary and pancreatic medical record data after privacy anesthesia.

3. According to claim 2, a medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information is characterized by: The numerical type hepatobiliary and pancreatic disease classification module is used to process the numerical type data in the medical records processed by the data processing module to obtain the numerical type hepatobiliary and pancreatic disease classification results; the specific process is: 1) Process the numerical data in the medical records processed by the data processing module; the specific process is: If the data in the medical record is of numeric type, it will not be processed; If the data in the medical record is non-numeric data, the non-numeric data is converted into numerical information; 2) The numerical type hepatobiliary and pancreatic disease classification model is the XGBoost model; 3) Train the numerical type hepatobiliary and pancreatic disease classification model to obtain a trained numerical type hepatobiliary and pancreatic disease classification model; the specific process is: The input of the XGBoost model is 1) the numerical data in the processed medical records, and the output of the XGBoost model is the numerical hepatobiliary and pancreatic disease classification results; The objective function is L t ; The loss function is the cross entropy loss function; Obtain a trained numerical type hepatobiliary and pancreatic disease classification model; 4) Obtain the hepatobiliary and pancreatic medical record data to be tested, process it through the data processing module, and process the numerical type data in the medical record processed by the data processing module to obtain the numerical type data to be tested; Based on the trained numerical type hepatobiliary and pancreatic disease classification model, the numerical type data to be tested is classified to obtain the numerical type hepatobiliary and pancreatic disease classification results.

4. According to claim 3, a medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information is characterized by: The objective function L t for: Where, L t represents the objective function of the t-th decision tree, n represents the number of data in a batch of updated data; y i Represents sample x i Tags, Represents the first t-1 decision trees for sample x i The prediction result, f t (x i ) represents the current t-th decision tree for sample x i The prediction results of Ω(f t ) represents the complexity of the t-th decision tree; loss means for The cross entropy loss function of The complexity of the t-th decision tree Ω(f t ) is obtained as follows: For the tth decision tree f t , L1 regularization term Ω1(f t ) is calculated as follows: Among them, Q is the total number of leaf nodes of the t-th decision tree; w j is the weight of the jth leaf node; γ is a hyperparameter that controls the strength of regularization; For the tth decision tree f t , L2 regularization term Ω2(f t ) is calculated as follows: Among them, Q is the total number of leaf nodes of the t-th decision tree; w j is the weight of the jth leaf node; λ is a hyperparameter that controls the strength of regularization; The complexity of the t-th decision tree Ω(f t ) is a linear combination of L1 regularization term and L2 regularization term:

5. According to claim 4, a medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information is characterized by: The calculation formula of the information gain entropy H(k) of different features k obtained in the numerical type hepatobiliary and pancreatic disease classification model is as follows: In the formula, Z represents the number of all decision trees; C(t) represents the total number of non-leaf nodes in the t-th decision tree; B(t) represents the number of all nodes in the t-th decision tree; O represents the O-th node in the t-th decision tree; m represents whether the current node belongs to feature k, if it does, it is 1, if not, it is 0; Loss(t,i,L) and Loss(t,i,R) represent the losses on the left and right nodes of the i-th non-leaf node in the t-th decision tree, respectively; D(t,i,L) and D(t,i,R) are the second-order derivatives of Loss(t,i,L) and Loss(t,i,R) respectively; b and γ are the hyperparameters of the regularization term, respectively; The splitting situation I(k) of the nodes to which different features k are obtained in the numerical type hepatobiliary and pancreatic disease classification model belong in the entire decision tree; the calculation formula is shown as follows: Where B(t) represents the number of all nodes in the t-th decision tree; Based on the information gain entropy H(k) and the splitting situation I(k), the importance of feature k for the numerical type hepatobiliary and pancreatic disease classification model is obtained. The calculation formula is shown as follows: Importance(k)=αH(k)+(1-α)I(k)Importance(k) In the formula, α represents the adjustment parameter.

6. According to claim 5, a medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information is characterized by: The natural language type hepatobiliary and pancreatic disease classification module is used to process the natural language type data in the medical records processed by the data processing module to obtain the natural language type hepatobiliary and pancreatic disease classification results; the specific process is: 1) The natural language type hepatobiliary and pancreatic disease classification model is the Bert-chinese model; 2) Train the natural language type hepatobiliary and pancreatic disease classification model to obtain a trained natural language type hepatobiliary and pancreatic disease classification model; the specific process is: The input of the natural language type hepatobiliary and pancreatic disease classification model is the natural language type data in the medical records processed by the data processing module, and the output of the natural language type hepatobiliary and pancreatic disease classification model is the hepatobiliary and pancreatic disease type; The loss function is the cross entropy loss function; Obtain a trained natural language type hepatobiliary and pancreatic disease classification model; 3) Obtain the hepatobiliary and pancreatic medical record data to be tested, process it through the data processing module, and process the natural language type data in the medical record processed by the data processing module to obtain the natural language type data to be tested; Based on the trained numerical type hepatobiliary and pancreatic disease classification model, the natural language type data to be tested is classified to obtain the natural language type hepatobiliary and pancreatic disease classification results.

7. According to claim 6, a medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information is characterized by: The calculation formula of the dot product attention in the Bert-chinese model is: Among them, Dot(x) represents dot product attention; Q represents the output of the input text X after a linear layer with a shape of [Embed,Embed]; Embed represents the embedding length; K represents the output of the input text X after a linear layer with a shape of [Embed,Embed]; Q×K T The result is a matrix of [Sequence_length, Sequence_length], where Sequence_length is the length of the input text X; d represents the embedding dimension; the superscript T represents the transposition; Set a threshold T′, extract the texts to which all tokens whose dot product attention is greater than the threshold T′ belong in the input, and use this as the basis for the decision of the Bert-chinese model; Token represents a word token.

8. According to claim 7, a medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information is characterized by: The imaging examination data type hepatobiliary and pancreatic disease classification module is used to process the imaging examination type data in the medical record processed by the data processing module to obtain the imaging examination type hepatobiliary and pancreatic disease classification result; the specific process is: Process the imaging examination data in the medical records into regular expressions to obtain the classification results of hepatobiliary and pancreatic diseases in the form of regular expressions; 1) Extract disease name: Regular expression: (liver cancer|liver cyst|gallbladder polyp|cholecystitis|tumor) 2) Extract the lesion location: Regular expression: (left side | right side | both sides) 3) Extract the extent of lesions: Regular expression: (mild|moderate|severe) 4) Extract keywords from image descriptions: Regular expression: (image displayed|discovered|observed|visible) 5) Extract digital information (such as size, density, etc.): Regular expression: (\d+(\.\d+)?)\s*(cm|mm|cm 2 |mm 2 ).

9. The medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information according to claim 8, characterized in that: The integration module is used to integrate the numerical type hepatobiliary and pancreatic disease classification results, the natural language type hepatobiliary and pancreatic disease classification results, and the imaging examination type hepatobiliary and pancreatic disease classification results to obtain an integrated classification result; The specific process is: 1) The integrated module classification model is a multi-layer perceptron MLP; The working principle of the multi-layer perceptron is as follows: MLP(X)=WX+B Among them, X is the input of the multilayer perceptron, W and B are the parameters to be trained; 2) Train the integrated module classification model to obtain a trained integrated module classification model; the specific process is: The input of the integrated module classification model is the predicted label and predicted probability. The predicted label is the numerical type hepatobiliary and pancreatic disease classification result, the natural language type hepatobiliary and pancreatic disease classification result, and the imaging type hepatobiliary and pancreatic disease classification result. The predicted probability refers to the probability that the current sample belongs to a certain type of hepatobiliary and pancreatic disease. The output of the integrated module classification model is the integrated hepatobiliary and pancreatic disease type; The loss function is the cross entropy loss function; Obtain a trained integrated module classification model; 3) Based on the trained integrated module classification model, the numerical type hepatobiliary and pancreatic disease classification results, natural language type hepatobiliary and pancreatic disease classification results and imaging examination type hepatobiliary and pancreatic disease classification results are processed to obtain an integrated classification result.

10. A medical record agent system for intelligent auxiliary diagnosis and treatment of hepatobiliary and pancreatic diseases using medical record information according to claim 9, characterized in that: The evidence judgment module is used to judge whether the integrated classification result is consistent with the doctor's diagnosis opinion, and output supporting evidence if consistent; if inconsistent, output opposing evidence; The specific process is: If the integrated classification result obtained by the integrated module is consistent with that of the doctor, then the integrated classification result obtained by the integrated module is supporting evidence; If the integrated classification result obtained by the integration module is inconsistent with that of the doctor, the integrated classification result obtained by the integration module is dissenting evidence.

Images