Model construction method based on preoperative evaluation of postoperative complication risk of hepatic resection

Abstract

The invention relates to a construction method of a model for preoperative evaluation of postoperative complications of hepatic resection. Performing high-order interactive modeling on the plurality of preoperative characteristic variables, and identifying the potential influence of the multi-factor joint effect on the postoperative complications; comprising the interaction between hepatic insufficiency and diabetes and the relationship between preoperative anemia and malnutrition; using a decision tree model to screen out features matched with the complication risk by calculating each variable and interaction item; for the modeling of the nonlinear relationship between the variables, capturing the nonlinear relationship between the variables by adopting a support vector machine (SVM) method; the prediction of the increase of the preoperative liver hardness is in a non-linear relationship with the risk of postoperative complications; constructing a multi-level regression model including hierarchical Bayesian regression, performing risk assessment on different levels including an individual level and a group level, and integrating individual differences and group features of patients; and each layer of model independently evaluates different risk factors, and integrates data of different layers for prediction.

Description

Technical Field

[0001] The present invention relates to a method for constructing a complication risk model, and in particular to a method for constructing a complication risk model after liver resection based on preoperative evaluation. Background Art

[0002] Although the current methods for constructing risk models for preoperative assessment of complications after liver resection have made some progress in some aspects, they still have some obvious deficiencies and drawbacks in practical applications. Compared with the postoperative complication prediction model training methods described in relevant comparative documents (such as Chinese inventions, publication number CN116313053A), the existing preoperative risk assessment models still face many challenges in data processing, model comprehensiveness, personalized adjustment and prediction accuracy.

[0003] First, the existing preoperative evaluation methods mostly rely on traditional clinical feature data, such as the patient's age, gender, underlying diseases, preoperative physiological indicators, etc. These models often treat each feature independently as an input variable, and lack effective modeling of high-order interactions between features. In some cases, the risk of complications of patients is not determined by a single feature or a simple combination of certain features, but by the complex interaction of multiple factors. Therefore, the predictive ability of traditional models performs poorly when faced with complex and nonlinear medical data. In contrast, the prediction method based on multi-task learning (such as the model described in CN116313053A) can optimize the model by processing the multi-task loss function, thereby improving the accuracy of complication prediction. However, this method does not completely avoid the defect of the model being too dependent on a single feature, and the implementation method of multi-task learning may also lead to increased model complexity, longer training time, and a large amount of labeled data is required to achieve the desired effect.

[0004] Secondly, the existing methods are relatively simple and extensive in data preprocessing and feature selection. When preprocessing patient clinical data, existing models usually use conventional missing value filling, normalization and other methods, but these methods have limited effect on the complexity and incompleteness of clinical data. For example, data missing may be due to the fact that some test items of the patient were not done or not entered in time, and traditional missing value filling methods (such as mean filling, linear interpolation, etc.) may introduce bias, especially when dealing with a large amount of patient data, which can easily lead to instability or overfitting of model training. The methods in the technical documents of the literature try to improve the quality and representativeness of training data by expanding the dimension, normalizing and screening features of the data set. Although this improves the selectivity of features and the balance of data, it may also affect the generalization ability of the model due to over-expansion or over-simplification. In addition, the problem of class imbalance of data is very common in complication prediction tasks. The frequency of occurrence of different types of complications varies greatly. Traditional methods may ignore the prediction of minority complications. Multi-task learning methods solve this problem to some extent, but if handled improperly, they may still fail to effectively capture the characteristics of low-frequency complications. Furthermore, existing preoperative risk assessment methods usually rely on traditional algorithms such as single linear regression or decision trees, and the performance of the model is often limited when dealing with medical data with high dimensions and complex features. Especially when faced with nonlinear relationships, traditional algorithms may not be able to fully capture the interaction effects between various features. For example, the interaction between preoperative diabetes and cirrhosis, and the complex relationship between the patient's underlying disease and preoperative physical examination indicators, may not be accurately represented by simple regression analysis or tree models.

[0005] In summary, the existing methods for constructing risk models for complications after liver resection based on preoperative assessment have certain deficiencies and limitations in data processing, model adaptability, personalized prediction ability, interpretability, and computational efficiency. The present invention not only makes up for the shortcomings of traditional methods by introducing technologies such as multi-level modeling, dynamic adjustment, and Bayesian framework, but also better cope with the complex and changeable situations in clinical practice. Summary of the invention

[0006] The purpose of the present invention is to provide a method for constructing a risk model for complications after liver resection based on preoperative evaluation, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0007] The present invention solves the above-mentioned technical problems by adopting the following technical solution, comprising the following steps:

[0008] S1. High-order interaction modeling of preoperative risk factors:

[0009] S1.1. Conduct high-order interaction modeling on multiple preoperative characteristic variables to identify the potential impact of multiple combined factors on postoperative complications, including the interaction between liver dysfunction and diabetes, and the relationship between preoperative anemia and malnutrition;

[0010] S1.2. Use the decision tree model to screen out features that match the risk of complications by calculating each variable and interaction term;

[0011] S1.3. For modeling of nonlinear relationships between variables, the support vector machine (SVM) method was used to capture the nonlinear relationships between variables, including the prediction of the increase in preoperative liver stiffness and the risk of postoperative complications.

[0012] S2. Modeling of preoperative evaluation model:

[0013] S2.1. Construct a multilevel regression model including hierarchical Bayesian regression to conduct risk assessment at different levels including individual level and group level, integrating individual differences of patients with group characteristics;

[0014] S2.2. Each layer of the model independently evaluates different risk factors and integrates data from different levels for prediction; including first conducting group-level modeling based on the evaluation of the underlying diseases of patients with diabetes and cirrhosis, and then adjusting the model through regression at the individual level;

[0015] S3. Dynamic monitoring and real-time update of risks during surgery:

[0016] S3.1. Collect real-time data through intraoperative monitoring equipment, including blood loss, liver resection range, anesthesia depth, and blood sugar level; use real-time data to update the preoperative risk assessment model so that the model can be dynamically adjusted during the operation;

[0017] S3.2. Input intraoperative data into the model to update the complication risk score in real time; use the recurrent neural network (RNN) algorithm to combine historical data with real-time data to adjust the expected risk of postoperative complications for patients;

[0018] S4. Prediction of postoperative complications:

[0019] S4.1. Postoperative risk prediction: Combine preoperative data, intraoperative data, and data from the first day after surgery to build a personalized complication prediction model;

[0020] S4.2. Generate personalized intervention recommendations based on the risk prediction results of postoperative complications.

[0021] Furthermore, the preoperative risk factor high-order interaction modeling method includes:

[0022] Based on capturing the interaction effect, the enhanced decision tree algorithm adopted integrates the direct influence of a single feature, generates interaction terms, and incorporates the interaction terms into the node splitting process of the decision tree; in this process, the generation of interaction terms is carried out according to the following formula:

[0023]

[0024] Among them, X i and X j are preoperative characteristic variables, representing various characteristics in clinical data, including liver function, diabetes, and anemia; α i It is feature X i The weight coefficient reflects the direct impact of the characteristics on postoperative complications; β ij is the feature interaction term X i X j The weight coefficient of feature X i and X j The interaction between X i X j The weight β ij It will be adjusted dynamically based on the distribution of the data and its relationship with postoperative complications.

[0025] Furthermore, the high-order interaction modeling method of the former risk factors includes:

[0026] Through the interaction term generation technology, combined with the machine learning algorithm including L1 regularization, potential interaction terms are identified, and the model is dynamically optimized according to the feature weights of the interaction terms during the training process; the generation of interaction terms is achieved through the following formula:

[0027]

[0028] Among them, w i is a single feature X i The weight coefficient of feature X i Direct contribution to the risk of postoperative complications; w ij is the interaction term X i X j The weight coefficient of feature X i and X j the contribution of interactions between the two groups to the risk of postoperative complications; is a measure of the weight of the interaction term, representing the total contribution of all features and interaction terms.

[0029] Furthermore, the high-order interaction modeling method of the former risk factors includes:

[0030] An integrated learning algorithm including LightGBM is used to capture the nonlinear relationship between variables through the gradient boosting framework. The sample weights in each round of training are adjusted through multiple rounds of training to optimize the model's predictive ability. The model's predictive ability is optimized through the following loss function:

[0031]

[0032] in, is the result of the i-th sample predicted by the model, indicating the predicted risk value of postoperative complications; i is the true label of the i-th sample, indicating the actual risk of postoperative complications; θ j is a parameter in the model including the split coefficient of each decision tree node, which represents the weight of the feature in the model; λ and γ are regularization parameters; X k is the kth feature of the input, μ k and σ k are the mean and standard deviation of the feature respectively.

[0033] Furthermore, the modeling method of the preoperative evaluation model includes:

[0034] At the population level, risk modeling is performed based on the common characteristics of patients' underlying diseases, including diabetes and liver cirrhosis. The population-level model captures the overall impact of characteristics on postoperative complications by modeling patient population data and generates a unified risk assessment baseline. Hierarchical Bayesian regression is used to represent the baseline risk of various diseases in the population by setting a prior distribution. The description of the population-level model is expressed by the following formula:

[0035]

[0036] Among them, P(y i |θ,φ) is the probability of postoperative complications in patient i, where y i Indicates whether patient i has complications, where 1 means it has occurred and 0 means it has not occurred; θ is the regression parameter of the group-level model, and φ is the prior distribution at the group level, capturing the group risk; represents the risk calculation for all N patients; Bernoulli(y i |θ·x i ) is the Bernoulli distribution model, which represents the probability of postoperative complications in patient i, where x i It is a group characteristic of patients with underlying diseases including diabetes and cirrhosis.

[0037] Furthermore, the modeling method of the preoperative evaluation model includes:

[0038] In the process of individual-level modeling, the baseline risk at the group level is used as prior information and dynamically adjusted according to the individual characteristics of the patient, including age, preoperative physiological indicators, and physical examination results. Through the Bayesian posterior update mechanism, individual characteristics correct the prediction results of the group-level model, and the adjustment at the individual level is performed through the following formula:

[0039]

[0040] Among them, P(θ i |y i ,x i ) is the posterior distribution at the individual level, indicating that when the individual characteristics of the patient x are known i and postoperative outcomes i In the case of i The update probability of P(y i |θ i ,x i ) is the patient i in the given individual characteristic x i and individual regression parameters θ i The conditional probability of postoperative complications under i |φ) is the prior distribution of the group level, which means that in the absence of individual data, the individual regression parameter θ is based on group characteristics. i Prior information; P(y i |x i ) represents the individual feature x i The overall predicted probability of .

[0041] Furthermore, the modeling method of the preoperative evaluation model includes:

[0042] Each level of the model independently assesses risk factors and integrates the results of the group and individual level assessments through a Bayesian framework. The model combines the baseline risk at the group level with the influence of individual characteristics to generate a risk prediction for postoperative complications. When integrating the risk predictions at the group and individual levels, it is expressed by the following formula:

[0043] P(y i |x i )=P(y i |x i )·P(θ i |y i ,x i )

[0044] Among them, P(y i |x i ) is the population level based on patient characteristics x i Predict the probability of postoperative complications, reflecting the prediction results at the group level; P(θ i |yi ,x i ) is the prediction at the individual level after Bayesian posterior update, which represents the assessment of the risk of postoperative complications for patient i by combining individual characteristics and group-level baseline predictions.

[0045] The method for constructing a risk model for complications after liver resection based on preoperative evaluation of the present invention has the following beneficial effects:

[0046] By combining group-level and individual-level risk assessments, the present invention fully considers individual characteristics such as the patient's underlying disease, age, gender, and group-level risks of common diseases such as diabetes and cirrhosis before surgery, thereby more accurately predicting the probability of postoperative complications in patients. The application of the Bayesian framework effectively combines the baseline risk at the group level with the influence of individual characteristics, further improving the personalization and accuracy of risk prediction.

[0047] Reveal the interactive effects between different preoperative characteristics and dynamically adjust the risk of postoperative complications according to the patient's specific situation. This personalized risk assessment method allows doctors to provide more precise and targeted preoperative and postoperative intervention measures based on the specific situation of each patient. For example, for patients with diabetes, blood sugar control can be strengthened before surgery to reduce the risk of postoperative complications.

[0048] The intraoperative dynamic monitoring and real-time risk update function enables the risk prediction model to be adjusted in real time according to new data obtained during the operation (such as bleeding volume, liver resection range, anesthesia depth, etc.). By introducing algorithms such as recurrent neural networks (RNN), the model dynamically adapts to changes in intraoperative data, adjusts complication risk prediction in a timely manner, and enhances decision-making support capabilities during surgery.

[0049] The hierarchical Bayesian regression method, which uses separate modeling at the group level and the individual level, not only ensures the model's effective prediction of the entire patient population, but also makes fine adjustments to different patients at the individual level, improving the model's computational efficiency and interpretability. This hierarchical modeling structure makes the model easier for clinicians to understand and operate while retaining its efficiency and accuracy, and has good prospects for clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The present invention is a flow chart of the method for constructing a risk model for complications after liver resection based on preoperative evaluation.

[0051] Figure 2 This is a flow chart of the high-order interaction modeling method for preoperative risk factors of the present invention.

[0052] Figure 3 The figure is a flow chart of the modeling method of the preoperative evaluation model of the present invention.

[0053] Figure 4 A is an axial view of Example 3 of the present invention, showing an image of the patient's lungs on an axial section.

[0054] Figure 4 B is the three-dimensional reconstructed lung image of the respiratory system three-dimensional model of Example 3 of the present invention.

[0055] Figure 4 C is a coronal view of Example 3 of the present invention, showing a coronal section of the patient's lungs.

[0056] Figure 4 D is a sagittal view of Example 3 of the present invention showing a longitudinal section of the lungs. DETAILED DESCRIPTION

[0057] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0058] The S1 step in the method of constructing a risk model for postoperative complications after liver resection based on preoperative evaluation, namely high-order interaction modeling of preoperative risk factors, aims to identify the potential impact of the combined effects of different factors on the risk of postoperative complications by performing high-order interaction analysis on multiple preoperative characteristic variables.

[0059] First, considering that patients in clinical practice often have multiple potential risk factors at the same time, such as the interaction between liver dysfunction and diabetes, and preoperative anemia and malnutrition, simply analyzing the independent impact of each variable cannot comprehensively and accurately assess the risk of postoperative complications. Therefore, a high-order interaction modeling method is needed to address the joint effects of these multiple factors. Specifically, by constructing a decision tree model, important interaction terms can be automatically calculated and screened between each feature node. The decision tree model takes each preoperative variable as input, recursively splits different combinations of patient characteristics, and finally predicts the risk of postoperative complications at the end of the tree. In this way, the interaction between different preoperative variables can be reflected by the split structure of the tree, so that the contribution of each interaction term can be fully explored. Furthermore, when capturing the nonlinear relationship between variables, the support vector machine (SVM) method is introduced. SVM is a powerful classification and regression method that effectively handles complex relationships between variables by constructing nonlinear decision boundaries in high-dimensional space. In the preoperative assessment model, SVM can help identify those complex nonlinear relationships that exist outside the linearity, especially when the interaction between some variables cannot be captured by a simple linear regression model. For example, the relationship between the increase in preoperative liver stiffness and the risk of postoperative complications is not a simple linear relationship, but has obvious nonlinear characteristics. Through the SVM method, we can introduce a kernel function to map low-dimensional data to a higher-dimensional space, and segment the data in the new space through a linear model, thereby capturing these complex nonlinear patterns.

[0060] Step S2, i.e., the modeling of the preoperative evaluation model, mainly involves constructing a multilevel regression model including hierarchical Bayesian regression to conduct risk assessment at different levels, integrating the individual differences and group characteristics of patients, so as to more accurately predict the risk of postoperative complications.

[0061] This process first models the population level. Through population-level modeling, the overall risk trend of a specific group (such as patients with underlying diseases such as diabetes and cirrhosis) can be identified. At this level, the population characteristics are first analyzed using large sample data to obtain the baseline risk of the group. For example, the risk of postoperative complications in patients with diabetes is different from that in patients with cirrhosis. Therefore, through the population-level model, the overall impact of such underlying diseases on postoperative complications can be quantified and a reasonable population risk baseline can be set for it. On this basis, the model combines the pre-set prior distribution with the existing data through the hierarchical Bayesian regression method, and uses the mechanism of Bayesian reasoning to estimate the risk at the population level. The advantage of hierarchical Bayesian regression is that it integrates prior knowledge with existing data and dynamically updates the parameters of the model, thereby better reflecting the risk characteristics at the population level. Next, the model enters the regression adjustment stage at the individual level. Through individual-level modeling, combined with the patient's personalized information (such as age, preoperative physiological indicators, living habits, etc.), the population-level prediction results are adjusted. For example, although a patient belongs to the high-risk group for diabetes, if his blood sugar is well controlled before surgery, the individual-level model can correct the risk prediction through the posterior update mechanism of Bayesian regression, thereby improving the accuracy of the prediction. Each layer of the model will independently evaluate different risk factors in this process. The group-level model focuses on identifying the relationship between group characteristics and risks, while the individual-level model pays more attention to the personalized differences of patients.

[0062] Step S3, i.e., dynamic intraoperative monitoring and real-time risk update, mainly uses intraoperative monitoring equipment to collect various physiological data of patients in real time, such as bleeding volume, liver resection range, anesthesia depth, blood sugar level, etc., and inputs these real-time data into the existing preoperative risk assessment model, so as to dynamically update and adjust the prediction of postoperative complication risk during the operation. Intraoperative monitoring equipment continuously provides detailed data about the patient's physiological state, which is usually not taken into account in traditional static preoperative assessment models. However, during the actual operation, the patient's physiological state will change dramatically, such as excessive bleeding, increased liver resection range, unstable anesthesia depth, blood sugar fluctuations, etc., all of which affect the risk of postoperative complications. Therefore, incorporating these real-time intraoperative data into the model update makes risk assessment more accurate and timely. To achieve this goal, the use of recurrent neural network (RNN) algorithm is an ideal choice, because RNN is particularly good at processing sequence data and time series data, which enables it to dynamically adjust historical data and real-time data. Through the loop structure of RNN, the model can not only process the data at the current moment, but also retain the information at the past moment, so that the model can capture the temporal dependency of the data. During surgery, by combining the results of preoperative evaluation with real-time collected data, the RNN algorithm can adjust the patient's expected risk of postoperative complications in real time. For example, during surgery, if the patient bleeds too much, the model will process this information in combination with historical data and dynamically update the risk score; if the scope of liver resection is expanded, resulting in an increased risk of postoperative complications, the RNN will adjust this risk increment in the model to provide a more accurate prediction of postoperative complications. In this way, the model not only provides a static risk assessment before surgery, but also updates the complication risk score in real time as various changes occur during the operation, helping doctors make timely decisions and adjust treatment plans.

[0063] Step S4, namely, postoperative complication prediction, mainly builds a personalized complication prediction model by comprehensively considering the data before, during, and on the first day after surgery, so as to accurately predict the complications that may occur after surgery and generate personalized intervention recommendations based on the prediction results.

[0064] In this process, preoperative data, including the patient's basic medical history, underlying diseases, preoperative examination results, and other information, are first used to establish a preliminary risk assessment framework for the patient. Then, during the intraoperative stage, the preoperative assessment results are dynamically updated and adjusted through real-time collected physiological data, such as blood loss, liver resection range, anesthesia depth, blood sugar level, etc. Intraoperative data reflects the changes that patients experience during surgery, and therefore plays an important supplementary role in predicting postoperative complications. Finally, data from the first day after surgery, including the patient's recovery, changes in vital signs, laboratory test results, etc., should also be included in the complication prediction model. The data from the first day after surgery reflects the patient's early postoperative physiological response and helps to further improve the assessment of postoperative risks. After integrating these data, the model generates a personalized complication risk prediction for each patient. This prediction is not only a static calculation based on historical data, but also reflects the patient's current health status and postoperative recovery through dynamic adjustment and real-time updates. Based on the risk prediction results of postoperative complications, the model can further generate personalized intervention recommendations. For example, if the model predicts that a patient has a higher risk of infection after surgery, the system can automatically generate recommendations for intervention measures such as strengthening the use of antibiotics and strengthening nursing monitoring; if the predicted bleeding risk is high, it is recommended to strengthen bleeding monitoring or take appropriate intervention measures in a timely manner.

[0065] Embodiment 1:

[0066] A patient with diabetes, liver dysfunction and anemia was selected, and the enhanced decision tree algorithm was used to capture the interaction effects between different preoperative characteristic variables, and further establish an accurate postoperative complication risk prediction model.

[0067] The patient's preoperative characteristic variables are set as:

[0068] X1: Liver function (e.g., metabolic capacity of the liver, liver enzyme levels, etc.), the value range is: 0≤X1≤100 (e.g., AST and ALT values, AST is 25, ALT is 30, indicating relatively poor liver function).

[0069] X2: History of diabetes, the value range is: 0≤X2≤1, 1 means the patient has diabetes, 0 means the patient does not have diabetes.

[0070] X3: Anemia degree, the value range is: 0≤X3≤1, 1 indicates anemia, and 0 indicates no anemia.

[0071] In order to capture the high-order interaction effects between these features, an enhanced decision tree model is first constructed, and the above formula is used to calculate the impact of each feature and its interaction terms. The formula is:

[0072]

[0073] In this formula, X1, X2, and X3 are the preoperative characteristics of liver function, diabetes, and anemia, respectively; α i is the direct influence coefficient of each feature on postoperative complications, β ij is the interaction coefficient, which indicates the effect of the interaction of two features on postoperative complications. Through the decision tree model, the weight coefficient α of each variable i and the weight coefficient of the interaction term β ij It will be adjusted dynamically based on the distribution of the data and its relationship with postoperative complications.

[0074] First, a certain amount of historical clinical data was collected. The data set was set to contain information on 500 patients who underwent liver resection. The data of each patient contained the following features: liver function (X1), diabetes (X2), and anemia (X3). By preprocessing these data, the mean and standard deviation of each feature variable were calculated for standardization. On this basis, a preliminary prediction model was constructed using the enhanced decision tree algorithm.

[0075] Through the training process of the decision tree, the model automatically generates each preoperative feature variable and its interaction term, and dynamically calculates the weight of each interaction term based on the training data. For example, during the training process, the model will find that the interaction between liver dysfunction (X1) and diabetes (X2) (i.e., the interaction terms X1 and X2 have a greater impact on postoperative complications, so the coefficient β of this interaction term is 12 will be adjusted to a higher value.

[0076] Through calculation, the specific values ​​of the following interaction terms are obtained:

[0077] α1=0.45: indicates the direct influence of liver function (X1) on postoperative complications. The larger α1 is, the more significant the influence of liver function on postoperative complications is.

[0078] α2=1.2: represents the influence coefficient of diabetes (X2) on postoperative complications.

[0079] α3=0.8: represents the effect of anemia (X3) on postoperative complications.

[0080] β 12 =0.7: represents the influence coefficient of the interaction effect between liver dysfunction and diabetes, the value is positive, indicating that the combined effect of these two characteristics will further increase the risk of postoperative complications.

[0081] β 13 =0.5: represents the influence coefficient of the interaction effect between liver dysfunction and anemia.

[0082] β 23 =0.4: represents the influence coefficient of the interaction effect between diabetes and anemia.

[0083] After training and generating the above coefficients through the enhanced decision tree model, the model can be used for risk prediction. For example, the preoperative data of a patient is set as follows:

[0084] X1=40 (poor liver function, AST is 40, ALT is 45)

[0085] X2=1(history of diabetes)

[0086] X3=1(anemia)

[0087] Substitute these data into the trained model for calculation:

[0088] f(X1,X2,X3)=α1X1+α2X2+α3X3+β 12 X1X2+β 13 X1X3+β 23 X2X3

[0089] Substitute the coefficient values ​​and variables:

[0090] f(40,1,1)=0.45×40+1.2×1+0.8×1+0.7×40×1+0.5×40×1+0.4×1×1

[0091] Calculation results:

[0092] f(40,1,1)=18+1.2+0.8+28+20+0.4=68.4

[0093] Based on this risk score, the model predicts that the patient has a high risk of postoperative complications, and the specific value can be used to determine the subsequent intervention plan. For example, if the risk score exceeds 60, enhanced intraoperative monitoring and early postoperative intervention will be recommended.

[0094] As intraoperative data is collected, such as changes in bleeding volume (E) and anesthesia depth (A), the model will be dynamically adjusted based on real-time data. If intraoperative bleeding volume increases and anesthesia depth is unstable, the model can update risk predictions through real-time data and further adjust the patient's complication risk score. This dynamic adjustment method makes postoperative risk assessment more timely and personalized, thereby providing clinicians with more accurate decision support.

[0095] This example continues to use specific patient cases and explains in detail how to generate interaction terms and how to use L1 regularization to optimize the model to achieve more accurate prediction of postoperative complication risks.

[0096] In this example, patients who are still considered for liver resection have the following preoperative characteristics:

[0097] X1: Liver function (such as liver metabolic capacity, AST and ALT values). The patient's liver function is assumed to be poor, with AST of 40 and ALT of 45.

[0098] X2: history of diabetes, assuming the patient has diabetes, with a value of 1;

[0099] X3: anemia, assume that the patient has anemia, the value is 1;

[0100] X4: Preoperative nutritional status (such as body mass index (BMI)), set BMI to 23.

[0101] The generation of interaction terms is to capture the combined effects between different features. For example, liver function (X1) and diabetes (X_2) will jointly affect the risk of postoperative complications. It is necessary to use interaction term generation techniques combined with L1 regularization to identify these potential interaction terms.

[0102] The formula is as follows:

[0103]

[0104] Among them, w i is a single feature X i The weight coefficient of feature X i Direct contribution to the risk of postoperative complications; w ij is the interaction term X i X j The weight coefficient of feature X i and X j the contribution of interactions between the two groups to the risk of postoperative complications; is the total contribution measure of all features and interaction terms.

[0105] The weights of the following features are obtained from the data (through the training process of the machine learning algorithm):

[0106] w1 = 0.6 (weight coefficient of liver function, indicating that poor liver function will increase the risk of postoperative complications)

[0107] w2 = 0.5 (weight coefficient for diabetes, indicating that diabetic patients have a higher risk of postoperative complications)

[0108] w3 = 0.4 (weight coefficient of anemia, indicating that anemia increases the risk of postoperative complications)

[0109] w4 = 0.3 (weight coefficient of nutritional status, indicating that poor nutritional status also increases the risk of postoperative complications)

[0110] The weight coefficient of the interaction term can be calculated by the algorithm. For example, for the interaction term between liver function and diabetes, w 12=0.8, indicating that the combined effect of liver function and diabetes has a greater impact on postoperative complications.

[0111] In order to identify and optimize the most important interaction terms and remove unimportant features, an L1 regularized machine learning algorithm (such as Lasso regression) is used for feature selection. The basic idea of ​​L1 regularization is to introduce a regularization term to penalize excessive coefficients in the model, thereby prompting the model to select only features and interaction terms that have a greater impact on the results. The calculation formula of the L1 regularization term is as follows:

[0112]

[0113] Among them, λ is the regularization coefficient, which controls the strength of the penalty and is usually determined by cross-validation during the training process. A larger λ value will increase the penalty strength, causing more coefficients to approach zero, thereby removing unimportant features and interactions.

[0114] The model is trained through machine learning algorithms, and the value of λ is adjusted. After training, the optimal λ is set to 0.5. Then the model will remove some unimportant interaction terms through regularization, and finally retain the following important interaction terms:

[0115] X1X2 (interaction between liver function and diabetes, w 12 =0.8

[0116] X2X3 (interaction between diabetes and anemia, w 23 =0.6

[0117] In this way, when considering preoperative characteristics, the model will pay special attention to the joint effect between liver dysfunction and diabetes, as well as the interaction between diabetes and anemia. The existence of these interaction terms has a high explanatory power for predicting the risk of postoperative complications. By substituting the trained interaction terms and feature weights into the model, the patient's risk score can be obtained. The preoperative data of the patient is set as follows:

[0118] X1=40 (liver dysfunction, AST is 40, ALT is 45)

[0119] X2=1(history of diabetes)

[0120] X3=1(anemia)

[0121] X4=23(normal nutritional status)

[0122] Substitute into the model formula for calculation:

[0123] f(X1,X2,X3,X4)=w1X1+w2X2+w3X3+w4X4+w 12 X1X2+w 23 X2X3

[0124] Substitute the corresponding values ​​into the calculation:

[0125] f(40,1,1,23)=0.6×40+0.5×1+0.4×1+0.3×23+0.8×40×1+0.6×1×1

[0126] Calculation results:

[0127] f(40,1,1,23)=24+0.5+0.4+6.9+32+0.6=64.4

[0128] Therefore, the model predicts that the risk of postoperative complications for this patient is 64.4, indicating a high risk, and clinical monitoring and intervention should be strengthened. Through this specific patient case, it is explained in detail how to use the interaction term generation technology and L1 regularization algorithm to optimize the preoperative complication risk assessment model. By generating important interaction terms, training and optimizing the model, the complex relationship between different preoperative features can be more accurately captured, thereby improving the accuracy of postoperative complication risk prediction.

[0129] This example further refines the high-order interaction modeling process of preoperative risk factors by introducing the LightGBM ensemble learning algorithm. In this example, the LightGBM algorithm is used to capture the nonlinear relationship between variables, optimize the model prediction ability, and train and tune through a specific loss function. The feasibility and application of this method will be gradually demonstrated through actual data and calculation results.

[0130] First, a patient undergoing liver resection is selected as an example. The patient's preoperative characteristics include:

[0131] X1: Liver function (AST, ALT, etc.) - Assuming that the patient's AST is 40 and ALT is 45, it indicates poor liver function;

[0132] X2: history of diabetes mellitus—the patient is diabetic, with a value of 1;

[0133] X3: Anemia - the patient has anemia, the value is 1;

[0134] X4: Preoperative nutritional status (BMI) - the patient's BMI is 28, indicating that the patient is slightly overweight;

[0135] X5: Age—the patient is 65 years old;

[0136] X6: Gender - male, value is 1.

[0137] It is assumed that 1,000 similar patient data samples are collected, and these data will be used as a training set to train the model.

[0138] LightGBM (LightGradientBoostingMachine) is an integrated learning algorithm based on the gradient boosting tree (GBDT) framework. It effectively captures the nonlinear relationship between variables and adjusts the sample weights in each round of training through multiple rounds of training to optimize the model's prediction ability. In this model, there are complex nonlinear relationships between feature variables, and traditional linear regression cannot effectively capture these relationships, so LightGBM is chosen to model.

[0139] LightGBM gradually optimizes the loss function through the gradient boosting algorithm, thereby continuously reducing the prediction error. Its loss function is as follows:

[0140]

[0141] in, is the risk value of postoperative complications of the i-th sample predicted by the model; y i is the actual risk of postoperative complications (i.e., the true label); θ j are model parameters, including the split coefficient of each decision tree node, which indicates the importance of the feature in the model; λ and γ are regularization parameters used to control the complexity of the model; X k is the kth feature of the input, μ k and σ k are the mean and standard deviation of the feature respectively.

[0142] During the model training process, LightGBM will continuously optimize the sample weights in each round of training and adjust the prediction results. It is assumed that after multiple rounds of training, the model has learned how to effectively assess the risk of postoperative complications.

[0143] In the training process of the LightGBM model, the optimization of the loss function is very critical. The loss function guides the update of the model by calculating the difference between the predicted value and the actual value. Specifically, the first part of the loss function:

[0144]

[0145] is a typical mean square error (MSE), which represents the difference between the model's predicted value and the true value. This term is used to measure the accuracy of the model in predicting the risk of postoperative complications. If the predicted value is far from the actual label, the model needs more adjustments.

[0146] The second part, regularization term:

[0147]

[0148] Used to control the complexity of the model and avoid overfitting. By adjusting the regularization parameter λ, the model can be made more concise when fitting the data and avoid capturing unnecessary noise.

[0149] Part III:

[0150]

[0151] It is a regularization of feature variables, taking into account the mean and standard deviation of each feature. By standardizing the features, the model better understands the contribution of different features. The regularization parameters are set to λ=0.1 and γ=0.5, and the best model hyperparameters are selected through cross-validation during training.

[0152] After several rounds of training, the prediction results of the model are as follows:

[0153] For this patient, the model predicted a postoperative complication risk of 0.75 (ie, a 75% probability of complications occurring).

[0154] These predicted values ​​can be compared with actual postoperative complications. If the patient does develop complications after surgery, the prediction is accurate; if no complications occur, the model's prediction will be slightly biased, but overall, the model achieves high accuracy in most patients. Based on the prediction results, doctors can combine the patient's postoperative complication risk value (such as 0.75) with other relevant information to make more personalized clinical decisions. For example, if the patient's risk is higher, additional postoperative intervention measures are needed, such as strengthening postoperative monitoring, adjusting medication regimens, etc.

[0155] Embodiment 2:

[0156] This embodiment adopts a hierarchical Bayesian regression framework, and sets a prior distribution to represent the baseline risk of various diseases in the group, thereby calculating the overall probability of postoperative complications in the patient group.

[0157] First, a group of patients undergoing liver resection were selected as the research subjects. These patients have some common underlying diseases, such as diabetes and cirrhosis, and the common characteristics of these diseases have an important impact on the occurrence of postoperative complications. Preoperative data of 1,000 patients were collected, including clinical characteristics of diseases such as diabetes and cirrhosis. These characteristics were modeled at the population level in order to capture the overall impact of these underlying diseases in the population on postoperative complications and generate a unified risk assessment baseline.

[0158] The following features are selected to model population-level risk:

[0159] X1 represents the patient marker of diabetes, with a value of 1 indicating that the patient has diabetes and 0 indicating that the patient does not have diabetes;

[0160] X2 represents the patient marker for liver cirrhosis, with a value of 1 indicating that the patient has liver cirrhosis and 0 indicating that the patient does not have liver cirrhosis;

[0161] X3 represents age, which is set between 40 and 80 years old.

[0162] Through the population-level model, it is hoped that the impact of these characteristics on postoperative complications can be captured from the overall data of the patient population and ultimately generate a unified risk assessment baseline.

[0163] Hierarchical Bayesian regression is used to model risk assessment at the population level. The hierarchical Bayesian regression model models each patient's data and uses the prior distribution and posterior update mechanism to obtain the baseline risk of various diseases (such as diabetes and cirrhosis) in the population. The mathematical expression of the population hierarchical model is as follows:

[0164]

[0165] in:

[0166] P(y i |θ,φ) represents the probability of postoperative complications for patient i, y i is whether patient i has complications, 1 means yes, 0 means no;

[0167] θ is the regression parameter of the group-level model, representing the weight of the effect of the characteristics on the risk of postoperative complications;

[0168] φ is the population-level prior distribution, representing the baseline risk of different characteristics in the population, for example, the postoperative risk in patients with diabetes; x i is the characteristic data of patient i, including the common characteristics of diseases such as diabetes and liver cirrhosis;

[0169] Bernoulli i |θ·x i ) is a Bernoulli distribution, which is used to describe the probability of whether patient i will have postoperative complications, depending on the patient's characteristics. The key to this formula is to capture the overall impact of population characteristics on postoperative complications through the prior distribution φ. For example, the risk of postoperative complications in patients with diabetes and cirrhosis will be affected by these underlying diseases, so it is necessary to model them in the model through Bayesian regression.

[0170] Suppose there is a set of patient data, which includes whether the patient has diabetes, whether they have cirrhosis, and their age.

[0171] In order to simplify the problem, the following actual data is set:

[0172] Patient 1: diabetes (X1=1), liver cirrhosis (X2=0), age 45 years;

[0173] Patient 2: diabetes (X1=0), liver cirrhosis (X2=1), age 60 years;

[0174] Patient 3: Diabetes (X1=1), liver cirrhosis (X_2=1), age 55 years old.

[0175] The goal of the population-level model is to predict the probability of postoperative complications for each patient based on these characteristics. In order to calculate the risk of postoperative complications for each patient, it is necessary to estimate the regression parameters θ and the prior distribution φ, as well as calculate the contribution of each patient's characteristics.

[0176] The following regression parameters and prior distributions are estimated from historical data (for simplicity, there are two features and a constant term):

[0177] θ0 = -2.0 (constant term, representing the group baseline risk);

[0178] θ1 = 1.5 (regression parameter for diabetes);

[0179] θ2 = 0.5 (regression parameter for cirrhosis);

[0180] φ = 0.1 (population-level prior distribution).

[0181] For patient 1, diabetes has a greater positive impact on the risk of postoperative complications, while cirrhosis has a smaller impact. According to the above formula, the probability of postoperative complications for patient 1 is:

[0182] P(y1∣θ,φ)=Bernoulli(y1∣∣(-2.0+1.5×1+0.5×0))=Bernoulli(y1∣-0.5)

[0183] Therefore, the probability of postoperative complications in patient 1 is:

[0184]

[0185] For patient 2, cirrhosis has a greater impact on postoperative complications, while diabetes has a smaller impact. Calculate the risk probability for patient 2:

[0186] P(y2∣θ,φ)=Bernoulli(y2∣∣(-2.0+1.5×0+0.5×1))=Bernoulli(y2∣-1.5)

[0187] Therefore, the probability of postoperative complications in patient 2 is:

[0188]

[0189] For patient 3, both diabetes and cirrhosis have an impact on the risk of postoperative complications. Calculate the risk probability for patient 3:

[0190] P(y3∣θ,φ)=Bernoulli(y3∣∣(-2.0+1.5×1+0.5×1))=Bernoulli(y3∣0.0)

[0191] Therefore, the probability of postoperative complications in patient 3 is:

[0192]

[0193] The group hierarchical model can be used to derive the probability of postoperative complications for different patients. For example, the risk for patient 1 is 37.7%, the risk for patient 2 is 18.2%, and the risk for patient 3 is 50%. These predictions can provide a reference for doctors to help them develop personalized postoperative management and intervention plans. For patients at higher risk, doctors can strengthen postoperative monitoring and take preventive measures, such as closer monitoring of vital signs and adjusting drug therapy. The group hierarchy is modeled using a hierarchical Bayesian regression method, and the common characteristics of patients, such as diabetes and cirrhosis, are combined to predict the risk of postoperative complications.

[0194] Individual-level modeling is a crucial step in this case. This process not only relies on the baseline risk provided by the group-level model, but also dynamically adjusts the risk of postoperative complications by combining individual patient characteristics such as age, preoperative physiological indicators, and physical examination results. Through the Bayesian posterior update mechanism, individual characteristics correct the prediction results of the group-level model to generate a more accurate risk assessment. This modeling method is illustrated and calculated through specific examples.

[0195] When modeling the risk at the patient population level, an estimate of the population baseline risk has been obtained, and the model captures the impact of common characteristics such as diabetes and cirrhosis on postoperative complications. The results of the population-level model provide a baseline risk level, which will then be adjusted at the individual level based on the patient's specific characteristics.

[0196] Suppose there is a specific patient, patient A. Her preoperative data are as follows:

[0197] Age: 60 years old

[0198] Physical examination results: liver function indexes were slightly abnormal;

[0199] Preoperative physiological indicators: high blood sugar level (not diagnosed as diabetes).

[0200] The population-level model has derived a base risk of 40%, ie, in the absence of further individual data, patient A has a 40% chance of developing a postoperative complication.

[0201] The adjustment at the individual level is achieved through the Bayesian posterior updating mechanism, and the formula is as follows:

[0202]

[0203] in:

[0204] P(θ i |y i ,x i ) is the posterior distribution at the individual level, indicating that when the individual characteristics of the patient x are known i and postoperative outcomes i In the case of i The update probability of

[0205] P(y i |θ i ,x i ) is the patient i in the given individual characteristic x i and individual regression parameters θ i The conditional probability of postoperative complications under i |φ) is the prior distribution of the group level, which means that in the absence of individual data, the individual regression parameter θ is based on group characteristics. i Prior information; P(y i |x i ) is based on individual characteristics x i The overall predicted probability reflects the impact of individual characteristics on the probability of postoperative complications.

[0206] Taking the actual data of patient A as an example, individual-level adjustments are made. The baseline risk given by the population-level model is set to 40%, that is, the probability of postoperative complications for patient A is 40% without any individual characteristic data.

[0207] The population-level prior distribution φ reflects the overall impact of diseases such as diabetes and cirrhosis on the risk of postoperative complications at the population level. The following parameters have been estimated from historical data:

[0208] θ0 = -0.8 (population-level constant term, representing baseline risk);

[0209] θ1 = 0.6 (regression parameter of diabetes, reflecting the effect of diabetes on postoperative risk);

[0210] θ2=0.4 (regression parameter of cirrhosis, reflecting the effect of cirrhosis on postoperative risk).

[0211] Therefore, the population-level prior distribution P(θ i |φ) means that in the absence of individual data, the individual regression parameter θ is calculated based on group characteristics. i Estimates.

[0212] The individual characteristics of patient A are as follows:

[0213] Age: 60 years old

[0214] Preoperative physiological indicators: high blood sugar level, in the pre-diabetes stage;

[0215] Physical examination results: Slight abnormality in liver function.

[0216] Based on these characteristics, the risk of postoperative complications for patient A can be calculated by the following steps.

[0217] First, set the regression parameter θ of patient A A =θ0+θ1×X1+θ2×X2, where X1 represents the characteristics of diabetes (setting X1=1 represents diabetes, X1=0 represents no diabetes), and X2 represents the characteristics of cirrhosis (setting X2=0 represents no cirrhosis).

[0218] For patient A, it is assumed that she is in the pre-diabetes stage, so X1 = 1 and X2 = 0. Therefore, the regression parameters of patient A are:

[0219] θ A =-0.8+0.6×1+0.4×0=-0.2

[0220] Then, according to the Bayesian posterior update mechanism, the probability of postoperative complications of patient A is calculated:

[0221]

[0222] According to the above formula, in the case of patient A, the baseline risk provided by the population-level model is assumed to be 40%, so the risk of patient A can be adjusted through the Bayesian updating mechanism.

[0223] Set the group baseline risk P(y A ∣x A )=0.4, then:

[0224]

[0225] Through a specific calculation method, it can be concluded that the adjusted risk of postoperative complications for patient A is 40%*1.1=44% (the adjustment factor after Bayesian updating is set to 1.1).

[0226] Through the Bayesian posterior updating mechanism, the risk of postoperative complications for patient A was adjusted from 40% to 44%. This adjustment reflects that the risk of postoperative complications for patient A in the case of prediabetes is slightly higher than the baseline level.

[0227] Ultimately, by adjusting the individual-level model, doctors can provide a more accurate risk assessment of postoperative complications for patient A based on his specific characteristics and combined with the baseline risk at the group level, and thus provide a personalized intervention plan for postoperative management. By using the baseline risk at the group level as prior information and dynamically adjusting it in combination with the individual characteristics of the patient, the individual-level modeling method provides a more refined and personalized risk prediction for preoperative evaluation.

[0228] The population-level risk assessment is based on the common characteristics of patients, such as underlying diseases such as diabetes and cirrhosis. A patient population is set up, including 1,000 patients. The population-level model analyzes the population data and derives a baseline risk. For example, the baseline risk of the population-level assessment is set to 60%, which means that in the absence of any individual data, the probability of a patient developing a postoperative complication is 60%.

[0229] In this example, the group-level evaluation formula is:

[0230] P(y i |x i )=0.6

[0231] The characteristics of patient A are set as follows: age 60, relatively normal physiological indicators before surgery, and mild diabetes (high blood sugar level but not reaching the diabetes standard). Under the population hierarchical model, based on the characteristics of patient A, the predicted probability of postoperative complications is still 60%.

[0232] The individual-level assessment updates the group-level assessment through a Bayesian framework, combining the specific characteristics of patient A to produce a more accurate risk prediction. In the Bayesian posterior update, individual characteristics such as age, weight, and preoperative physiological indicators are used to adjust the results of the group-level prediction. The baseline risk given by the group level is set to 60%, and then the risk is adjusted based on the individual characteristics of patient A.

[0233] The individual level adjustment formula is:

[0234]

[0235] Among them, P(y i |θ i ,x i ) indicates that when the individual characteristics of the patient x are known i and postoperative outcomes i In the case of i The update probability of P(θi |φ) represents the prior distribution at the group level, that is, the known baseline risk information in the group; and P(y i |x i ) represents the predicted probability of postoperative complications at the individual level.

[0236] The prior distribution φ given at the population level was set to include the following information: the regression coefficient of diabetes on the risk of postoperative complications was 0.5, and patients aged over 60 years had a higher risk of postoperative complications, with a regression coefficient of 0.3.

[0237] Through Bayesian updating, adjustments are made based on the specific characteristics of patient A (such as diabetes and age). For example, patient A's diabetes characteristic X1 = 1 (mild diabetes) and age characteristic X2 = 60, so the individual-level adjustment coefficient θ can be obtained A =0.5×1+0.3×1=0.8.

[0238] Through Bayesian posterior updating, the individual-level risk assessment of patient A was obtained. Next, the group-level and individual-level assessment results were integrated through the Bayesian framework to obtain the final postoperative complication risk prediction for patient A.

[0239] The integration formula is:

[0240] P(y i |x i )=P(y i |x i )·P(θ i |y i ,x i )

[0241] In this formula:

[0242] P(y i |x i ) is the baseline risk at the population level, set at 60%;

[0243] P(θ i |y i ,x i ) is the posterior update probability at the individual level, which is set to 0.8.

[0244] Therefore, the final postoperative complication risk prediction is:

[0245] P(y A |x A )=0.6×0.8=0.48

[0246] This means that after combining the population-level baseline risk with the individual characteristics of patient A, the final risk of postoperative complications in patient A is 48%. Compared with relying solely on population-level risk assessment (60%), individualized adjustment makes risk prediction more accurate.

[0247] Based on the final risk prediction, the doctor can better formulate a personalized postoperative management plan for patient A. For example, before the operation, intervention measures for diabetes and advanced age, such as blood sugar control and preoperative evaluation, can be arranged to reduce the occurrence of postoperative complications. At the same time, the doctor can also provide more precise postoperative care recommendations based on the adjusted 48% risk assessment to help patient A achieve better postoperative recovery.

[0248] The Bayesian framework is used to integrate the evaluation results at the group level and the individual level, significantly improving the accuracy of postoperative complication risk prediction. In practical applications, the integration process not only considers the baseline risk at the group level, but also combines the individual characteristics of the patient, making the risk prediction more personalized and accurate. This method is of great significance in the preoperative evaluation of liver resection, helping doctors make more reasonable preoperative and postoperative decisions and improving patients' treatment outcomes and quality of life.

[0249] Embodiment 3:

[0250] This embodiment combines three-dimensional computed tomography (CT) technology with the method of constructing a preoperative assessment model for the risk of postoperative complications after liver resection of the present invention, and is applied to the preoperative risk assessment of patients undergoing liver resection. Through accurate lung imaging data, combined with the individual characteristics of the patient (such as underlying diseases, age, gender, etc.), this method provides a more personalized and accurate prediction of postoperative complications, thereby providing strong support for clinical decision-making. The following describes in detail the data collection, processing, and calculation process, showing how to assess the risk of postoperative complications through three-dimensional imaging data and clinical characteristics.

[0251] First, the patient's three-dimensional lung image data is obtained through CT scanning and preprocessed (such as denoising, standardization, etc.) for subsequent analysis.

[0252] Figure 4A (axial view) shows an image of the patient's lungs in an axial section. Using this view, the cross-sectional area of ​​the lungs is measured and the lung volume is estimated. In the axial view, the cross-sectional area of ​​the right lung is assumed to be A. right =180cm 2 , the cross-sectional area of ​​the left lung is A left =160cm 2 , the lung volume can be calculated by the formula:

[0253] V lung =A right ×L right +Aleft ×L left

[0254] Among them, L right and L left are the lengths of the right lung and the left lung respectively (the length is set to 20 cm). Therefore, the lung volume is:

[0255] V lung =180×20+160×20=3600+3200=6800cm 3

[0256] This data will be used as one of the input features in the risk assessment model to help analyze the impact of the patient's lung health status on postoperative complications.

[0257] Figure 4B (3D model of the respiratory system) is a 3D reconstructed lung image. This figure shows the patient's entire respiratory system, including the right lung, left lung, and trachea. The 3D model can better identify the airway structure of the lungs, and then evaluate the patency of the airways and lung volume. The airway patency index C = 0.92 (ranging from 0 to 1, the closer the value is to 1, the better the patency) is obtained from the 3D model. Combined with this data, the model can further predict the risk of postoperative respiratory failure.

[0258] Figure 4C (coronal view) shows a coronal section of the patient's lungs. Through this view, doctors can identify problems such as structural asymmetry and airway obstruction in the lungs. In this view, there is mild airway stenosis in the right lung, and the lung volume in the stenotic area is V. right,narrow =900cm 3 , the left lung is V left =1600cm 3 The risk of postoperative complications is weighted by calculating the degree of airway obstruction in the patient's lungs and incorporating that data.

[0259] The 4D (sagittal view) in the figure shows the longitudinal section of the lungs. In this view, the morphology of the alveoli and the health of the airways can be observed. The set image analysis showed that the alveolar morphology of the right upper lobe was irregular, indicating impaired gas exchange function, which is expected to affect lung function by about 15%. This data will also be included in the risk assessment model to adjust the prediction of postoperative complications.

[0260] The preoperative assessment model of the present invention combines the three-dimensional image data of the lungs with clinical characteristics (such as the patient's age, underlying diseases, physical examination indicators, etc.). The risk assessment process uses a Bayesian regression model to dynamically adjust risk prediction by combining group-level and individual-level data.

[0261] The clinical data of the patient are set as follows:

[0262] Age: 60

[0263] Do you have diabetes: Yes (x1=1)

[0264] Whether there is underlying disease of liver cirrhosis: Yes (x2=1)

[0265] Lung volume: 6800cm 3 (from Figure 4 A calculated)

[0266] Airway patency: 0.92 (from Figure 4 B gets)

[0267] Combined with the above characteristics, the Bayesian model will be initialized using the population-level baseline risk (such as diabetes, cirrhosis, etc.). The prior distribution of the population baseline risk is set as:

[0268]

[0269] Where μ = 0.5 and σ 2 =0.2, indicating the average level and standard deviation of group risk. Based on the individual characteristics of the patient (such as diabetes, lung volume, airway patency, etc.), the model will dynamically adjust the posterior probability at the individual level to predict the risk of postoperative complications. In the calculation, the patient's lung volume (6800cm 3 The contribution weights of airway patency (0.92) and postoperative respiratory failure are w1=0.4 and w2=0.3 respectively, so the predicted probability of postoperative respiratory failure is:

[0270] P(y i |x i )=σ(w1·V lung +w2·C)

[0271] Among them, σ is the Sigmoid function, which is used to map the result to between 0 and 1. Substitute the specific data for calculation:

[0272] P(y i |x i )=σ(0.4·6800+0.3·0.92)=σ(2720+0.276)=σ(2720.276)

[0273] The calculation results are:

[0274] P(y i |x i )≈1

[0275] This indicates that the patient is at a very high risk of developing respiratory failure after surgery, and clinicians need to pay special attention to this and prepare for postoperative respiratory support in advance. Through the preoperative risk assessment method based on three-dimensional lung images of the present invention, doctors can have a clearer understanding of the patient's lung condition and its impact on postoperative recovery before surgery. For example, in the above calculations, the patient's lung volume is small and the airway patency is low, which greatly increases the risk of postoperative respiratory failure. Combined with the patient's underlying diseases (diabetes, cirrhosis, etc.), the model gives a higher risk of postoperative complications. Based on this result, doctors can develop personalized postoperative care plans, such as strengthening respiratory support, using ventilators in advance, etc., to reduce the probability of postoperative complications.

[0276] This embodiment combines 3D CT image data with the patient's clinical characteristics to accurately calculate the patient's lung volume, airway patency and other indicators, and uses the Bayesian regression model to perform a personalized assessment of the risk of postoperative complications. This method provides clinicians with more accurate preoperative decision support, reduces the incidence of postoperative complications, and significantly improves the patient's postoperative recovery quality.

[0277] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for constructing a risk model for complications after liver resection based on preoperative evaluation, characterized by The following steps are involved: S1. High-order interaction modeling of preoperative risk factors: S1.

1. Conduct high-order interaction modeling on multiple preoperative characteristic variables to identify the potential impact of multiple combined factors on postoperative complications, including the interaction between liver dysfunction and diabetes, and the relationship between preoperative anemia and malnutrition. S1.

2. Use the decision tree model to screen out features that match the risk of complications by calculating each variable and interaction term; S1.3, for modeling nonlinear relationships between variables, the support vector machine (SVM) method is used to capture the nonlinear relationships between variables; These included increases in preoperative liver stiffness predicting a nonlinear relationship with the risk of postoperative complications; S2. Modeling of preoperative evaluation model: S2.

1. Construct a multilevel regression model including hierarchical Bayesian regression to conduct risk assessment at different levels including individual level and group level, integrating individual differences of patients with group characteristics; S2.

2. Each layer of the model independently evaluates different risk factors and integrates data from different levels for prediction; This includes first conducting group-level modeling based on the assessment of patients’ underlying diseases such as diabetes and liver cirrhosis, and then adjusting the model through regression at the individual level; S3. Dynamic monitoring and real-time update of risks during surgery: S3.

1. Collect real-time data through intraoperative monitoring equipment, including blood loss, liver resection range, anesthesia depth, and blood sugar level; use real-time data to update the preoperative risk assessment model so that the model can be dynamically adjusted during the operation; S3.

2. Input intraoperative data into the model to update the complication risk score in real time; use the recurrent neural network (RNN) algorithm to combine historical data with real-time data to adjust the expected risk of postoperative complications for patients; S4. Prediction of postoperative complications: S4.

1. Postoperative risk prediction: Combine preoperative data, intraoperative data, and data from the first day after surgery to build a personalized complication prediction model; S4.

2. Generate personalized intervention recommendations based on the risk prediction results of postoperative complications.

2. The method for constructing a risk model for complications after liver resection based on preoperative evaluation according to claim 1, characterized in that The preoperative risk factor high-order interaction modeling method comprises: Based on capturing the interaction effect, the enhanced decision tree algorithm adopted integrates the direct influence of a single feature, generates interaction terms, and incorporates the interaction terms into the node splitting process of the decision tree; in this process, the generation of interaction terms is carried out according to the following formula: Among them, X i and X j are preoperative characteristic variables, representing various characteristics in clinical data, including liver function, diabetes, and anemia; α i It is feature X i The weight coefficient reflects the direct impact of the feature on postoperative complications; β ij is the feature interaction term X i X j The weight coefficient of feature X i and X j The interaction between X i X j The weight β ij It will be adjusted dynamically based on the distribution of the data and its relationship with postoperative complications.

3. The method for constructing a risk model for complications after liver resection based on preoperative evaluation according to claim 2, characterized in that The method for modeling high-order interactions of pre-risk factors includes: identifying potential interactions through interaction term generation technology combined with a machine learning algorithm including L1 regularization, and dynamically optimizing the model according to the feature weights of the interaction terms during the training process.

4. The method for constructing a risk model for complications after liver resection based on preoperative evaluation according to claim 3, characterized in that The method for modeling the high-order interaction of pre-risk factors includes: using an integrated learning algorithm including LightGBM to capture the nonlinear relationship between variables through a gradient boosting framework; adjusting the sample weights in each round of training through multiple rounds of training to optimize the predictive ability of the model.

5. The method for constructing a risk model for complications after liver resection based on preoperative evaluation according to claim 1, characterized in that The modeling method of the preoperative evaluation model includes: At the population level, risk modeling is performed based on the common characteristics of patients' underlying diseases, including diabetes and liver cirrhosis. The population-level model captures the overall impact of characteristics on postoperative complications by modeling patient population data and generates a unified risk assessment baseline. Hierarchical Bayesian regression is used to represent the baseline risk of various diseases in the population by setting a prior distribution. The description of the population-level model is expressed by the following formula: Among them, P(y i |θ,φ) is the probability of postoperative complications in patient i, where y i Indicates whether patient i has complications, where 1 means it has occurred and 0 means it has not occurred; θ is the regression parameter of the group-level model, and φ is the prior distribution at the group level, capturing the group risk; To represent the risk calculation for all N patients; is the Bernoulli distribution model, which represents the probability of postoperative complications in patient i, where x i It is a group characteristic of patients with underlying diseases including diabetes and cirrhosis.

6. The method for constructing a risk model for complications after liver resection based on preoperative evaluation according to claim 5, characterized in that The modeling method of the preoperative evaluation model includes: In the process of individual-level modeling, the baseline risk at the group level is used as prior information and dynamically adjusted according to the individual characteristics of the patient, including age, preoperative physiological indicators, and physical examination results. Through the Bayesian posterior update mechanism, individual characteristics correct the prediction results of the group-level model, and the adjustment at the individual level is performed through the following formula: Among them, P(θ i |y i ,x i ) is the posterior distribution at the individual level, indicating that when the individual characteristics of the patient x are known i and postoperative outcomes i In the case of i The update probability of P(y i |θ i ,x i ) is the patient i in the given individual characteristic x i and individual regression parameters θ i The conditional probability of postoperative complications under i |φ) is the prior distribution of the group level, which means that in the absence of individual data, the individual regression parameter θ is based on group characteristics. i Prior information; P(y i |x i ) represents the individual feature x i The overall predicted probability of .

7. The method for constructing a risk model for complications after liver resection based on preoperative evaluation according to claim 6, characterized in that The modeling method of the preoperative evaluation model includes: Each level of the model independently assesses risk factors and integrates the results of the group and individual level assessments through a Bayesian framework. The model combines the baseline risk at the group level with the influence of individual characteristics to generate a risk prediction for postoperative complications. When integrating the risk predictions at the group and individual levels, it is expressed by the following formula: P(y i |x i )=P(y i |x i )·P(θ i |y i , x i ) Among them, P(y i |x i ) is the population level based on patient characteristics x i Predict the probability of postoperative complications, reflecting the prediction results at the group level; P(θ i |y i ,x i ) is the prediction at the individual level after Bayesian posterior update, which represents the assessment of the risk of postoperative complications for patient i by combining individual characteristics and group-level baseline predictions.

Images