Risk assessment model construction method and system based on deep learning and machine learning
By constructing a risk assessment model based on deep learning and machine learning, the existing surgical risk assessment methods have strong subjectivity, low efficiency, insufficient standardization and low data utilization, and a more objective, reliable and efficient surgical risk assessment is achieved.
Patent Information
- Application Number
- CN202411491988.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2025-02-25
AI Technical Summary
The existing surgical risk assessment methods have problems such as strong subjectivity, low efficiency, insufficient standardization and low data utilization, and it is difficult to effectively deal with complex or nonlinear data relationships.
A risk assessment model construction method based on deep learning and machine learning is adopted. By collecting multi-dimensional surgical correlation data, preprocessing and feature extraction, a risk assessment model including machine learning cart decision tree model and deep learning Fe-DCN model is constructed, and the output results of the model are comprehensively scored for weighted scores to generate risk assessment results.
It improves the objectivity and reliability of surgical risk assessment, reduces subjective deviations caused by individual differences in doctors, ensures the consistency and repeatability of evaluation results, and improves medical work efficiency and patient safety.
Smart Images
Figure CN119517391A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of network model construction, and in particular to a method and system for constructing a risk assessment model based on deep learning and machine learning. Background Art
[0002] At present, surgical risk assessment mainly relies on doctors' professional knowledge and clinical experience. Doctors will judge the risk of surgery for patients based on their own experience and the handling of similar cases in the past, taking into account factors such as the patient's age, health status, previous medical history, and drug allergy history. With the accumulation of medical knowledge, some standardized assessment tools and scoring systems have gradually been developed to guide doctors in risk assessment. For example, the American Society of Anesthesiologists (ASA) physical status classification system is a commonly used assessment standard. It divides patients into five levels, each corresponding to a different health status, to help doctors judge surgical risks. Preoperative consultation is an indispensable part of the assessment process. The doctor will discuss the necessity of surgery, possible complications, postoperative recovery, etc. with the patient in detail, and conduct a comprehensive assessment based on the patient's personal situation. In addition, multidisciplinary teamwork is also an important part of preoperative assessment. Doctors from different specialties will participate in the discussion to ensure the comprehensiveness and accuracy of the assessment results. However, the above method of assessing the surgical risk of patients has the following disadvantages: 1. Strong subjectivity: It is highly dependent on the doctor's personal experience and judgment and is easily affected by individual differences. 2. Low efficiency: It takes a lot of time and energy, especially when dealing with a large number of patients. 3. Insufficient standardization: The lack of unified standards and processes may lead to poor consistency in evaluation results. 4. Low data utilization: Failure to fully utilize existing patient data for systematic analysis makes it difficult to discover potential risk factors.
[0003] In order to reduce the time and energy of doctors in risk assessment and improve work efficiency, many medical institutions use rule-based systems to assess patients' surgical risks. Such systems usually consist of a set of predefined rules based on clinical guidelines and expert opinions. Although this approach is relatively simple and easy to understand, it lacks flexibility and adaptability and cannot handle complex or nonlinear data relationships. There are also some traditional statistical models, including logistic regression, Cox proportional hazard model, etc., which are also widely used to predict patient risks. These models require less data preprocessing, but may perform poorly when processing high-dimensional data, and may have difficulty capturing complex interaction effects. Therefore, a model method technology is needed to extract knowledge from a large amount of patient data and build a predictive model based on this knowledge. Summary of the invention
[0004] In order to solve the problems existing in traditional risk assessment models, the purpose of the present invention is to provide a method and system for constructing a risk assessment model based on deep learning and machine learning, extract knowledge from a large amount of patient data, and construct an assessment model based on this knowledge using machine learning and deep learning related technologies to improve the efficiency of medical work and the safety of patients.
[0005] The present invention is achieved through the following technical solutions:
[0006] This solution provides a risk assessment model construction method based on deep learning and machine learning, including:
[0007] Collect multi-dimensional surgery-related data;
[0008] Pre-process the surgery-related data to create feature data;
[0009] Constructing a risk assessment model and training the risk assessment model based on the feature data; the risk assessment model includes a machine learning cart decision tree model and a deep learning Fe-DCN model;
[0010] The output results of the machine learning cart decision tree model and the deep learning Fe-DCN model are integrated, and the risk assessment model is output after the assessment strategy of the risk assessment model is configured.
[0011] Working principle of this solution: This solution is based on multi-dimensional surgery-related data to train risk assessment models, which can automatically learn rules from data and reduce reliance on doctors' personal experience and judgment, thereby improving the objectivity and reliability of assessment results; through the automated assessment of machine learning models, subjective biases caused by individual differences among doctors are avoided, ensuring the consistency of assessment results; the training of risk assessment models is based on unified data sets and assessment standards, ensuring the standardization of the assessment process and reducing errors caused by inconsistent standards. Through standardized data processing and model training processes, the repeatability and consistency of assessment results are ensured.
[0012] A further optimization scheme is that the method of collecting multi-dimensional surgery-related data includes:
[0013] Collect recorded data of surgical targets based on the medical system, and collect on-site data of surgical targets on-site;
[0014] Input the recorded data and field data into the constructed large-scale language model to generate simulated data, and then clean the simulated data;
[0015] The recorded data, on-site data and cleaned simulation data are used as surgery-related data.
[0016] A further optimization scheme is that the surgery-related data is preprocessed to create feature data; including the method:
[0017] De-duplicate, fill in, and correct errors in surgery-related data;
[0018] Inputting each surgery-related data into the constructed target surgery feature extraction model to extract feature data; the feature data is composed of a plurality of different category variable sets;
[0019] The categorical variables of each categorical variable set are obtained, and encoding processing is performed on the categorical variables.
[0020] A further optimization scheme is that the method for constructing the machine learning cart decision tree model includes:
[0021] The cart decision tree model is constructed using the feature data as the sample set: the categorical variable set is used as the attribute set, the GINI coefficient of each attribute set is calculated, and the attribute set with the smallest GINI coefficient is used as the split attribute of the root node to recursively establish the decision tree child nodes until the samples of all child nodes belong to the same attribute or there is no attribute that can be selected as the split attribute, and finally the cart decision tree model is obtained;
[0022] Dynamically improve the constructed cart decision tree model based on the boundary point judgment principle;
[0023] Calculate the relevant scores of the improved cart decision tree model to determine the category of each leaf node.
[0024] A further optimization scheme is to construct a cart decision tree model using feature data as a sample set, including the following method:
[0025] In the classification task, for a given attribute set D, which contains N samples, the attribute set D is divided into attribute set D1 and attribute set D2 according to the i-th value of feature A. The GINI coefficient of attribute set D is:
[0027] .
[0028] A further optimization scheme is to dynamically improve the constructed cart decision tree model based on the boundary point determination principle, including the following method:
[0029] Sort all data in a continuous attribute set;
[0030] The midpoints of two sorted adjacent data are selected as candidate segmentation points;
[0031] For each candidate segmentation point, calculate the Gini index of the segmented subset;
[0032] The candidate segmentation point with the smallest Gini index is selected as the final segmentation point.
[0033] A further optimization scheme is that the deep learning Fe-DCN model includes a cross network, a feature enhancement module and a deep network.
[0034] The feature enhancement module is used to first perform global average pooling on the feature data, and then convolve to capture context information, while reducing the amount of calculation based on deep separable convolution; based on the SCA attention mechanism, the weight distribution between channel dimensions is obtained to obtain enhanced new feature data;
[0035] The deep network is used to capture the nonlinear interaction between the combined feature data; the combined feature data is the data obtained by combining the new feature data with the feature data;
[0036] The cross network is used to capture the cross influence between feature data.
[0037] A further optimization scheme is to output the risk assessment model after configuring the assessment strategy of the risk assessment model based on the output results of the integrated machine learning cart decision tree model and the deep learning Fe-DCN model, including the method:
[0038] Get the output results of the machine learning cart decision tree model and the deep learning Fe-DCN model;
[0039] The risk assessment result is obtained by performing weighted scoring calculation on the two output results.
[0040] The present solution also provides a risk assessment model construction system based on deep learning and machine learning, which is used to implement the above-mentioned risk assessment model construction method based on deep learning and machine learning. The system includes:
[0041] A collection module, used to collect multi-dimensional surgery-related data;
[0042] A preprocessing module is used to preprocess the surgery-related data to create feature data;
[0043] A construction module is used to construct a risk assessment model and train the risk assessment model based on feature data; the risk assessment model includes a machine learning cart decision tree model and a deep learning Fe-DCN model;
[0044] The output module is used to collect the output results of the comprehensive machine learning cart decision tree model and the deep learning Fe-DCN model, configure the evaluation strategy of the risk assessment model, and then output the risk assessment model.
[0045] The present solution also provides a computer-readable medium having a computer program stored thereon, and the computer program is executed by a processor to implement the risk assessment model construction method based on deep learning and machine learning as described above.
[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0047] The present invention provides a method and system for constructing a risk assessment model based on deep learning and machine learning; the risk assessment model is trained based on multi-dimensional surgery-related data, which can automatically learn rules from the data and reduce dependence on the doctor's personal experience and judgment, thereby improving the objectivity and reliability of the assessment results; through the automated evaluation of the machine learning model, the subjective bias caused by individual differences among doctors is avoided, and the consistency of the assessment results is ensured; the training of the risk assessment model is based on a unified data set and evaluation criteria, which ensures the standardization of the evaluation process and reduces errors caused by inconsistent standards. Through standardized data processing and model training processes, the repeatability and consistency of the assessment results are ensured. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:
[0049] Figure 1 Construct a process diagram for risk assessment models based on deep learning and machine learning;
[0050] Figure 2 Schematic diagram for building a risk assessment model based on deep learning and machine learning;
[0051] Figure 3 Schematic diagram of the deep learning Fe-DCN model structure;
[0052] Figure 4 Schematic diagram of SCA attention mechanism;
[0053] Figure 5 Schematic diagram of the feature enhancement module structure. DETAILED DESCRIPTION
[0054] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.
[0055] In order to reduce the time and energy of doctors in risk assessment and improve work efficiency, many medical institutions use rule-based systems to assess patients' surgical risks. Such systems usually consist of a set of predefined rules based on clinical guidelines and expert opinions. Although this method is relatively simple and easy to understand, it lacks flexibility and adaptability and cannot handle complex or nonlinear data relationships. There are also some traditional statistical models, including traditional statistical methods such as logistic regression and Cox proportional hazard model, which are also widely used to predict patient risks. These models require less data preprocessing, but may perform poorly when processing high-dimensional data, and may have difficulty capturing complex interaction effects. In view of this, this solution provides the following embodiments to solve the above technical problems:
[0056] Example 1
[0057] This embodiment provides a method for constructing a risk assessment model based on deep learning and machine learning, such as Figure 1 and Figure 2 As shown, the method includes:
[0058] Step 1: Collect multi-dimensional surgery-related data; this step specifically includes the following methods:
[0059] S11, collecting recorded data of the surgical target based on the medical system, and collecting on-site data of the surgical target on-site;
[0060] S12, inputting the recorded data and the on-site data into the constructed large-scale language model to generate simulated data, and performing cleaning processing on the simulated data;
[0061] S13, using the recorded data, the on-site data and the cleaned simulation data as surgery-related data.
[0062] The recorded data includes the personal information of the surgical target (age, gender), medical history, preoperative examination results, test results of various departments, type of surgery, conditions during the operation, anesthesia method used, postoperative recovery, etc. Data sources include hospital information system (HIS): including basic patient information, medical history, laboratory test reports, etc., electronic health records (EHR): containing the patient's complete medical history, surgical records: recording various indicators during the operation, such as blood pressure, heart rate, etc. Patient follow-up data: including the occurrence of postoperative complications, recovery status, etc.
[0063] Field data mainly refers to new data entered by medical staff on site. An automated data synchronization mechanism is implemented for the collected data to ensure that the data can be updated to the central database in real time. Use API interface to integrate with hospital information system and electronic health record system. Design background services to automatically synchronize data regularly. Implement a two-way synchronization mechanism to ensure data consistency. At the same time, based on security considerations, data transmission needs to be encrypted to protect patient privacy. Restrict access rights so that only authorized users can enter or view data.
[0064] Use a pre-trained large language model to generate simulated data based on existing disease tree knowledge. Generation process: Analyze existing data sets and extract disease tree structure and relevance. Use a large model to generate simulated patient data that conforms to the disease tree logic, and evaluate it with professionals to ensure that the generated data is statistically similar to actual cases. Clean and remove duplicates of the generated simulated data to ensure the quality of the simulated data.
[0065] Step 2: Preprocess the surgery-related data to create feature data; this step specifically includes the following methods:
[0066] S21, deduplicate, fill in and correct the surgery-related data; first, it is necessary to identify and delete duplicate records in the surgery-related data set. Duplicates can be identified by comparing key fields (such as patient ID, surgery date, etc.), and the latest or most complete records can be retained. There may be some missing fields in the surgery-related data set. For numerical variables, statistical methods such as mean, median or mode are used to fill in missing values; for categorical variables, appropriate filling strategies are selected based on contextual information, such as using the "unknown" category to mark missing values or predicting missing values based on other variables. Outliers may be caused by measurement errors or data entry problems. Statistical analysis and domain knowledge are needed to identify outliers and decide whether to correct or delete them. For example, if a patient's weight is recorded as 10 kg, this is obviously unreasonable and needs further verification and correction.
[0067] S22, inputting each surgery-related data into the constructed target surgery feature extraction model to extract feature data; the feature data is composed of a plurality of different category variable sets;
[0068] S23, obtaining the category variables of each category variable set, and encoding the category variables.
[0069] The routine data collected can be used directly for analysis and modeling without additional preprocessing steps. However, for certain types of data, it is necessary to set certain rules for further processing and conversion so that they can be better used for subsequent analysis and model training. For example, according to the patient's usual blood pressure level and medication status, we set rules through professional medical evaluation standards, and convert them into whether the blood pressure regularity and medication control are good, which are used as our features for model training. In this way, the original complex data can be converted into more intuitive, easy to understand and use features, thereby improving the accuracy and practicality of the model.
[0070] The constructed feature extraction model includes the existing medical disease tree, etc., which can determine which features are crucial for predicting surgical outcomes based on the characteristics of the surgery and relevant medical knowledge. For example, features related to the respiratory and digestive systems, because they may affect the risk of anesthesia and postoperative recovery. These features may include but are not limited to the patient's age, BMI, smoking history, asthma, sputum volume, hemoptysis or other history of respiratory diseases. In order to convert the categorical variables in the feature data into a form that can be understood by the machine learning algorithm, the categorical variables also need to be encoded: surgery-related data often contains many categorical variables, such as surgery type, gender, presence or absence of chronic diseases, etc.; One-Hot Encoding is used to process them.
[0071] Step three, construct a risk assessment model, and train the risk assessment model based on the feature data; the risk assessment model includes a machine learning cart decision tree model and a deep learning Fe-DCN model;
[0072] The risk assessment model of this scheme is composed of two types of models: a machine learning cart decision tree model and a deep learning Fe-DCN model; the deep learning Fe-DCN model in this scheme is an improved model obtained by adding a feature enhancement module to the DCN model. The advantage of the machine learning cart decision tree model is that it can meet the high requirements of medicine for interpretability. The deep learning Fe-DCN model combines the advantages of deep neural networks and cross networks, can effectively capture the complex relationships between data, and highlight important feature combinations through the attention mechanism. The risk assessment model of this scheme is particularly suitable for processing data containing a large number of features and potential interactions. The two models have their own advantages and can improve the accuracy of the overall predictive assessment in a complementary way.
[0073] In this step, the method for constructing a machine learning cart decision tree model includes:
[0074] S31, constructing a cart decision tree model using feature data as a sample set: using a categorical variable set as an attribute set, calculating the GINI coefficient of each attribute set, and recursively establishing a decision tree child node using the attribute set with the smallest GINI coefficient as the splitting attribute of the root node, until the samples of all child nodes belong to the same attribute or there is no attribute that can be selected as the splitting attribute, and finally obtaining a cart decision tree model;
[0075] Using feature data as a sample set to build a cart decision tree model, including methods:
[0076] For continuous attribute sets, such as blood pressure and blood sugar, set a segmentation threshold, discretize the sample set according to the segmentation threshold, and then calculate the GINI coefficient of the continuous attribute set;
[0077] For discrete attribute sets, such as whether there is hyperthyroidism or AIDS, dyspnea questionnaire mMRC and other features, the sample set needs to be divided according to the possible subsets of the discrete attribute values (except the full set and the empty set); if the discrete attribute has n values, then its valid subsets are 2 n -2. For example, there are only 2 values for whether there is AIDS. Then its valid subset is 2, and the mMRC feature of the dyspnea questionnaire contains 5 values of 0-4, so its valid subset is 30; then the subset with the smallest GINI coefficient is selected as the division method of the discrete attribute, and the minimum GINI coefficient is used as the GINI coefficient of the discrete attribute.
[0078] In the classification task, for a given attribute set D, which contains N samples, the attribute set D is divided into attribute set D1 and attribute set D2 according to the i-th value of feature A. The GINI coefficient of attribute set D is:
[0080] .
[0081] S32, dynamically improving the constructed cart decision tree model based on the boundary point determination principle; this step specifically includes the following method:
[0082] Sort all data in a continuous attribute set;
[0083] The midpoints of two adjacent sorted data are selected as candidate segmentation points; specifically, assuming that the sorted sample values are x1, x2, ..., x n , then the candidate segmentation points are
[0084] For each candidate segmentation point, calculate the Gini index of the segmented subset;
[0085] The candidate segmentation point with the smallest Gini index is selected as the final segmentation point.
[0086] In surgical risk prediction, it is necessary to evaluate various medical examination indicators of patients, such as blood pressure, blood sugar level, concentration of specific chemicals in the blood, etc. These indicators are usually continuous values, and their changes can directly reflect the patient's physical condition and are crucial to predicting surgical risks. When building a decision tree, the processing of continuous attributes usually requires finding an optimal split point to divide the data set into two subsets. Ideally, this split point should be able to maximize information gain or minimize evaluation indicators such as the Gini index. However, directly enumerating all possible split points is very time-consuming, especially when the data volume is large. Directly enumerating all possible split points is a very time-consuming process, especially when the data set is large. For example, the measurement value of blood pressure can range from 80mmHg to more than 200mmHg. If every possible value is to be tried as a split point, the amount of calculation will be huge. Therefore, when processing continuous attributes, the selection of the segmentation threshold can be dynamically adjusted. The boundary point judgment principle is used to select the segmentation threshold of continuous attributes. This method does not need to check every possible split point, but only needs to check the specific points after the sample is sorted. This can improve efficiency while ensuring the effectiveness of segmentation. When processing continuous attributes such as blood pressure, all blood pressure values can be sorted first. Then, we only look for split points between adjacent blood pressure values, that is, we choose the average of two adjacent samples as possible split points. The advantage of this is that we only need to check a small number of key points instead of all possible points.
[0087] S33, calculating the relevant scores of the improved cart decision tree model to determine the category of each leaf node.
[0088] When the decision tree selects the category label of the leaf node, the sample set of the leaf node adopts the "majority voting" method, that is, the majority class is selected as the category label of the leaf node. In order to overcome the numerical advantage of the majority class in the biased sample set and provide the small category with the opportunity to show its own data characteristics, the present invention proposes a measurement method of the correlation score, which selects the category label with the largest correlation score instead of selecting the category label of the majority class. In the traditional decision tree algorithm, when most of the samples in a leaf node belong to a certain class, the leaf node will be marked as this class. However, in the prediction of surgical risks, there may be some risk factors that are less common, but once they occur, they are very serious. For example, a rare underlying disease may significantly increase the risk of surgery, but because there are fewer such cases, the traditional majority voting method may ignore this risk factor. In order to ensure that the model can take into account these less common but important risk factors, a measurement method of the correlation score can be used. This method no longer simply relies on the majority class to determine the category of the leaf node, but selects the category label of the leaf node based on a certain correlation score. Relevance scores can be calculated based on the degree of association between each category and surgical risk, ensuring that even small categories that are at a disadvantage in terms of sample size can influence the model's decision. When assessing surgical risk, if an underlying disease (such as a heart disease) occurs less frequently in the training set but is highly correlated with postoperative complications, we can increase the surgical risk level of such patients by calculating the relevance score. In this way, even if such patients are a minority in the data set, the model will be more inclined to mark them as high risk, thereby helping doctors better prepare and prevent potential complications.
[0089] The definition of the correlation score is as follows:
[0090] Entropy reduction: In information theory, entropy is a measure of uncertainty. In a decision tree, we can determine which split point can best reduce uncertainty by calculating the entropy of each leaf node. For leaf node j, its entropy E (i,j) The calculation formula is as follows:
[0091]
[0092] Among them, p ij Represents the proportion of category i in leaf node j;
[0093] Support and confidence: Support refers to the frequency of category i appearing in leaf node j, which can be calculated using the following formula:
[0094]
[0095] Confidence refers to the ratio of the frequency of category i appearing in leaf node j to the total number of samples:
[0096]
[0097] in, Represents the sum of the support of all categories in leaf node j.
[0098] When selecting the category label of the leaf node, you can give priority to categories with higher support and higher confidence. This measurement method can better reflect the relative importance of category i in leaf node j, rather than just relying on absolute number.
[0099] Relevance score: Combining the entropy reduction with support and confidence, the present invention designs a new metric relevance score, which is defined as follows:
[0100] score(i,j)=w1E (i,j) +w2support(i,j)×confidence(i,j)
[0101] Wherein, w1 and w2 are weight factors, which can be adjusted according to actual conditions. In this scheme, the values of w1 and w2 are 0.6 and 0.4 respectively.
[0102] like Figure 3 As shown, the deep learning Fe-DCN model includes a cross network, a feature enhancement module and a deep network.
[0103] It is used to first perform global average pooling on the feature data, and then convolve to capture context information, while reducing the amount of calculation based on deep separable convolution; based on the SCA attention mechanism, the weight distribution between channel dimensions is obtained to obtain enhanced new feature data; the SCA attention mechanism structure of the feature enhancement module is as follows Figure 4 As shown, X is the input feature map, pool is the global average pooling, W is the fully connected layer, Represents element-wise product. The input feature map is firstly subjected to global average pooling and then to the fully connected layer to obtain the feature map as attention and multiply it with the input feature map. The structure of the feature enhancement module is as follows: Figure 5 As shown in the figure, the input features first pass through the global average pooling layer to average the pixel values of the entire feature map to generate a single vector representation. This operation can help capture the overall information of the feature, and then pass through a 7x1 depthwise separable convolution. The convolution with a large convolution kernel helps capture a wider range of contextual information. At the same time, the depthwise separable convolution reduces the amount of calculation. Then, the SCA attention mechanism is used to obtain the weight distribution between channel dimensions, thereby enhancing the model's attention to important features. After the sigmoid activation function, it is added to the input feature and output as the first output feature.
[0104] The second output feature is normalized by LayerNorm based on the first output feature, and then further extracted by a 7x1 depth-separable convolution and sigmoid activation function, and then added to the first output feature as the second output feature for output.
[0105] After the input features pass through the feature enhancement module, the receptive field is improved while capturing a wider range of contextual information. At the same time, the attention module is added to enhance the model's attention to important features, and new features containing richer information are obtained. In subsequent operations, the new features are merged with the original features to obtain the final features for the next step.
[0106] The deep network is used to capture the nonlinear interaction between the combined feature data; the combined feature data is the data obtained by combining the new feature data with the feature data;
[0107] The cross network is used to capture the cross-influence between feature data.
[0108] The DCN-v2 network consists of a cross network and a deep network. After the features pass through the two networks respectively, the output results are finally combined in the merged output layer. Based on DCN-v2, this solution discretizes the input data and further extracts the processed features into new features through the feature enhancement module. After feature enhancement, a series of new features are generated. These features can supplement the information of the original features and enrich the feature space. The new features after feature enhancement are combined with the original features. The purpose of this is to expand the feature space so that the model can learn more information. After the enhanced features are combined with the original features, they are input into the deep network part of the DCN-v2 network. Through the multi-layer perceptron (MLP) structure, the features are combined at multiple levels, so that the nonlinear interactions between features can be captured, which is very important for understanding complex medical data. The cross network part is responsible for capturing the cross-effects between features. Through linear combination, the cross network can effectively model high-order interactions between features, which helps to understand the complex relationship between different medical indicators.
[0109] Step 4: Combine the output results of the machine learning cart decision tree model and the deep learning Fe-DCN model, configure the risk assessment model's assessment strategy, and then output the risk assessment model. Specifically, obtain the output results of the machine learning cart decision tree model and the deep learning Fe-DCN model; perform weighted scoring calculations on the two output results to obtain the risk assessment results.
[0110] After completing the construction and training of the machine learning cart decision tree model and the deep learning Fe-DCN model, these models need to be integrated into a unified evaluation framework so that the advantages of the decision tree model and the deep learning model (Fe-DCN) can be used to perform multi-model joint evaluation. The decomposed features are input into the decision tree model of each disease system in turn to obtain the risk scores and important influencing feature factors for different systems. At the same time, the same data is also input into the Fe-DCN model to obtain the prediction results.
[0111] Cart decision tree model reasoning: Each cart decision tree model will give a corresponding risk score and key factors affecting the score based on the input features. These factors may be the patient's age, health status, previous medical history, etc. Deep learning model reasoning: The Fe-DCN model will also generate risk assessment results based on the input data, and may provide more detailed feature weight information to help understand which factors play an important role in the model decision.
[0112] Finally, the results obtained from each disease system model are summarized for comprehensive evaluation and decision-making.
[0113] Comprehensive evaluation and decision rule formulation: In order to integrate the inference results of multiple models, a weighted comprehensive score is used to quantify the risk more finely. The specific weight allocation needs to be determined according to the importance of each department system. For example, the health status of the respiratory system may have a greater impact on the risk of surgery, so it is given a higher weight in the comprehensive score.
[0114] The data of surgical objects includes the examination results of various departments, so a decision tree model and Fe-DCN model are trained separately for the data of each department. The results of multiple models are comprehensively evaluated, and the risk is quantified more finely through weighted comprehensive scoring. The weight allocation is determined according to the importance of each department system and the professional opinions of doctors, and the final result of risk prediction is obtained.
[0115] Example 2
[0116] This embodiment provides a risk assessment model construction system based on deep learning and machine learning, which is used to implement the risk assessment model construction method based on deep learning and machine learning in Example 1. The system includes:
[0117] A collection module, used to collect multi-dimensional surgery-related data;
[0118] A preprocessing module is used to preprocess the surgery-related data to create feature data;
[0119] A construction module is used to construct a risk assessment model and train the risk assessment model based on feature data; the risk assessment model includes a machine learning cart decision tree model and a deep learning Fe-DCN model;
[0120] The output module is used to collect the output results of the comprehensive machine learning cart decision tree model and the deep learning Fe-DCN model, configure the evaluation strategy of the risk assessment model, and then output the risk assessment model.
[0121] Embodiment 3, this embodiment provides a computer-readable medium having a computer program stored thereon, and the computer program is executed by a processor to implement the method for constructing a risk assessment model based on deep learning and machine learning as described in Embodiment 1.
[0122] This solution predicts patient data features and gives relevant treatment recommendations based on the improved decision tree model and Fe-DCN model. It can automatically learn rules from the data and reduce reliance on doctors' personal experience and judgment, thereby improving the objectivity and reliability of the evaluation results. Through the automated evaluation of the machine learning model, subjective bias caused by individual differences among doctors is avoided, ensuring the consistency of the evaluation results.
[0123] By automatically processing large amounts of patient data, risk assessment can be completed quickly, saving doctors’ time and energy and improving work efficiency. Compared with traditional manual assessment, the model can provide assessment results in a short time, especially in emergency situations, and can quickly provide decision support.
[0124] The training of the risk assessment model in this solution is based on a unified data set and evaluation criteria, which ensures the standardization of the evaluation process and reduces errors caused by inconsistent standards. The standardized data processing and model training process ensures the repeatability and consistency of the evaluation results; it can make full use of existing patient data to mine potential risk factors and increase the value of the data. It is capable of multi-dimensional analysis, and the model can comprehensively consider multiple factors such as the patient's age, health status, previous medical history, and history of drug allergies to conduct a comprehensive risk assessment; it can be dynamically adjusted according to new data and clinical practice to maintain its advancement and effectiveness; it can capture nonlinear characteristics and interaction effects to improve prediction accuracy.
[0125] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for constructing a risk assessment model based on deep learning and machine learning, characterized in that: The method comprises: Collect multi-dimensional surgery-related data; Pre-process the surgery-related data to create feature data; Constructing a risk assessment model and training the risk assessment model based on the feature data; the risk assessment model includes a machine learning cart decision tree model and a deep learning Fe-DCN model; The output results of the machine learning cart decision tree model and the deep learning Fe-DCN model are integrated, and the risk assessment model is output after the assessment strategy of the risk assessment model is configured.
2. The method for constructing a risk assessment model based on deep learning and machine learning according to claim 1, characterized in that: The method of collecting multi-dimensional surgery-related data includes: Collect recorded data of surgical targets based on the medical system, and collect on-site data of surgical targets on-site; Input the recorded data and field data into the constructed large-scale language model to generate simulated data, and then clean the simulated data; The recorded data, on-site data and cleaned simulation data are used as surgery-related data.
3. The method for constructing a risk assessment model based on deep learning and machine learning according to claim 2, characterized in that: Preprocessing the surgery-related data to create feature data; Included methods: De-duplicate, fill in, and correct errors in surgery-related data; Inputting each surgery-related data into the constructed target surgery feature extraction model to extract feature data; the feature data is composed of a plurality of different category variable sets; The categorical variables of each categorical variable set are obtained, and encoding processing is performed on the categorical variables.
4. The method for constructing a risk assessment model based on deep learning and machine learning according to claim 1, characterized in that: The method for constructing the machine learning cart decision tree model includes: The cart decision tree model is constructed using the feature data as the sample set: the categorical variable set is used as the attribute set, the GINI coefficient of each attribute set is calculated, and the attribute set with the smallest GINI coefficient is used as the split attribute of the root node to recursively establish the decision tree child nodes until the samples of all child nodes belong to the same attribute or there is no attribute that can be selected as the split attribute, and finally the cart decision tree model is obtained; Dynamically improve the constructed cart decision tree model based on the boundary point judgment principle; Calculate the relevant scores of the improved cart decision tree model to determine the category of each leaf node.
5. The method for constructing a risk assessment model based on deep learning and machine learning according to claim 4, characterized in that: The method of constructing a cart decision tree model using feature data as a sample set includes: In the classification task, for a given attribute set D, which contains N samples, the attribute set D is divided into attribute set D1 and attribute set D2 according to the i-th data of feature A. The GINI coefficient of attribute set D is: Where Gini() represents the GINI coefficient.
6. The method for constructing a risk assessment model based on deep learning and machine learning according to claim 5, characterized in that: The method of dynamically improving the constructed cart decision tree model based on the boundary point determination principle includes: Sort all data in a continuous attribute set; The midpoints of two adjacent sorted data are selected as candidate segmentation points; For each candidate segmentation point, calculate the Gini index of the segmented subset; The candidate segmentation point with the smallest Gini index is selected as the final segmentation point.
7. The method for constructing a risk assessment model based on deep learning and machine learning according to claim 1, characterized in that: The deep learning Fe-DCN model includes a cross network, a feature enhancement module and a deep network. The feature enhancement module is used to first perform global average pooling on the feature data, and then convolve to capture context information, while reducing the amount of calculation based on deep separable convolution; based on the SCA attention mechanism, the weight distribution between channel dimensions is obtained to obtain enhanced new feature data; The deep network is used to capture the nonlinear interaction between the combined feature data; the combined feature data is the data obtained by combining the new feature data with the feature data; The cross network is used to capture the cross influence between feature data.
8. The method for constructing a risk assessment model based on deep learning and machine learning according to claim 1, characterized in that: The output results of the integrated machine learning cart decision tree model and the deep learning Fe-DCN model are configured to output the risk assessment model after the assessment strategy of the risk assessment model, including the method: Get the output results of the machine learning cart decision tree model and the deep learning Fe-DCN model; The risk assessment result is obtained by performing weighted scoring calculation on the two output results.
9. A risk assessment model construction system based on deep learning and machine learning, characterized in that: The system is used to implement the method for constructing a risk assessment model based on deep learning and machine learning as described in any one of claims 1 to 8, comprising: A collection module, used to collect multi-dimensional surgery-related data; A preprocessing module is used to preprocess the surgery-related data to create feature data; A construction module is used to construct a risk assessment model and train the risk assessment model based on feature data; the risk assessment model includes a machine learning cart decision tree model and a deep learning Fe-DCN model; The output module is used to integrate the output results of the machine learning cart decision tree model and the deep learning Fe-DCN model, configure the evaluation strategy of the risk assessment model, and then output the risk assessment model.