A pancreatic fistula risk assessment model construction method and device

Abstract

The present application relates to the field of bioinformatics analysis, in particular to a pancreatic fistula risk assessment model construction method and device, the method comprises: according to the obtained patient data feature extraction obtains initial feature, patient data includes CT imaging data and clinical data, the initial feature includes the feature of pancreas and spleen CT, and clinical characteristics;According to the correlation of the initial feature and pancreatic fistula risk, the initial feature is screened, and the first screening feature is obtained;Mutual information method is used to determine the mutual information value of the first screening feature, and the first screening feature with the mutual information value greater than the preset threshold is determined as the second screening feature;According to the second screening feature, a logistic regression model is constructed, and a pancreatic fistula risk assessment model is obtained.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics analysis, and in particular to a method and apparatus for constructing a pancreatic fistula risk assessment model. Background Technology

[0002] Pancreaticoduodenectomy is a classic surgical procedure for the radical treatment of pancreatic tumors and is one of the most technically challenging abdominal surgeries. Postoperative complication rates range from 40% to 60%, with pancreatic fistula being the most common complication. Pancreatic fistula activates a large number of bioactive digestive enzymes in pancreatic juice, eroding adjacent tissues and causing symptoms such as abdominal tenderness and fever, leading to other serious complications. It is the leading determinant of postoperative mortality, with an incidence rate of 5% to 50%. Domestic and international studies show that the main risk factors for pancreatic fistula after pancreaticoduodenectomy include age >60 years, body mass index >25 kg / m², serum albumin <35 g / L, history of upper abdominal surgery, preoperative jaundice, soft pancreatic texture, main pancreatic duct diameter <3 mm, no preoperative prophylactic use of somatostatin analogues, and operation duration >285 minutes. These risk factors provide some guidance for predicting the occurrence of pancreatic fistula after pancreaticoduodenectomy. While these factors are relatively crude and rely heavily on human judgment, their predictive efficacy is effective.

[0003] Therefore, a new method and device for constructing a pancreatic fistula risk assessment model are needed. Summary of the Invention

[0004] Therefore, this invention provides a method for constructing a pancreatic fistula risk assessment model in an attempt to solve the problems mentioned above.

[0005] According to a first aspect of the present invention, a method for constructing a pancreatic fistula risk assessment model is provided. The method includes: extracting initial features from acquired patient data, wherein the patient data includes CT imaging data and clinical data, and the initial features include CT features of the pancreas and spleen, as well as clinical features; screening the initial features based on the correlation between the initial features and the risk of pancreatic fistula to obtain first screening features; determining the mutual information value of the first screening features using a mutual information method, and determining the first screening features with a mutual information value greater than a preset threshold as second screening features; and constructing a logistic regression model based on the second screening features to obtain a pancreatic fistula risk assessment model.

[0006] Optionally, in the method according to the present invention, screening the initial features based on their correlation with the risk of pancreatic fistula to obtain a first screening feature includes: determining a correlation determination method based on the data type of the initial features, wherein the data type includes continuous normally distributed data, continuous non-normally distributed data, ordered categorical data, and unordered categorical data, and the correlation determination method includes independent samples t-test, nonparametric verification, and chi-square test; determining the correlation of the initial features based on the determined correlation determination method, wherein the correlation is determined based on a probability value p; and determining the initial features as the first screening feature if the probability value p is less than a preset probability threshold. Optionally, in the method according to the present invention, determining the mutual information value of the first screening feature using the mutual information method, and determining the first screening feature with a mutual information value greater than a preset threshold as the second screening feature includes: determining the marginal distribution of the first screening feature, the outcome variable of pancreatic fistula, and the joint probability distribution of the first screening feature and the outcome variable; determining the mutual information value of the first screening feature based on the marginal distribution and the joint probability distribution; and determining the first screening feature as the second screening feature when the mutual information value of the first screening feature is greater than the preset threshold.

[0007] Optionally, in the method according to the present invention, constructing a logistic regression model based on the second screening feature to obtain a pancreatic fistula risk assessment model includes: using the second screening feature as an input variable to determine the probability that the input variable belongs to the positive or negative class; determining the parameters and regression coefficients in the logistic regression model based on the probability of belonging to the positive or negative class; and determining the logistic regression model based on the parameters and the regression coefficients.

[0008] Optionally, the method according to the present invention further includes: determining the regression coefficient of the second screening feature using the minimum absolute contraction and selection operator method; using the second screening feature as the third screening feature when the regression coefficient of the second screening feature is non-zero; and constructing a logistic regression model based on the third screening feature to obtain a pancreatic fistula risk assessment model.

[0009] Optionally, the method according to the present invention further includes: determining the expansion factor of the first screening feature; and removing the first screening feature if the expansion factor of the first screening feature is greater than a preset expansion factor threshold.

[0010] According to a second aspect of the present invention, a method for assessing the risk of pancreatic fistula is provided. The method includes: acquiring CT imaging data and clinical data of a patient to be assessed; determining input features based on the CT imaging data and clinical data, the input features including pancreatic deterioration ratio, main pancreatic duct diameter, amylase content in drainage fluid on the first postoperative day, serum albumin value on the third postoperative day, and body mass index; and inputting the input features into a pancreatic fistula risk assessment model to determine the risk of the patient developing pancreatic fistula.

[0011] According to a third aspect of the present invention, a device for constructing a pancreatic fistula risk assessment model is provided. The device includes: an initial feature extraction module, configured to extract initial features based on acquired patient data, the patient data including CT imaging data and clinical data, the initial features including CT features of the pancreas and spleen, and clinical features; a first screening feature determination module, configured to screen the initial features based on the correlation between the initial features and the risk of pancreatic fistula, to obtain first screening features; a second screening feature determination module, configured to determine the mutual information value of the first screening features using a mutual information method, and determine the first screening features with a mutual information value greater than a preset threshold as second screening features; and a logistic regression model construction module, configured to construct a logistic regression model based on the second screening features to obtain a pancreatic fistula risk assessment model.

[0012] According to a fourth aspect of the present invention, a computing device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, characterized in that the one or more processors execute the one or more computer programs to implement the steps of a method for constructing a pancreatic fistula risk assessment model.

[0013] According to a fifth aspect of the present invention, a computer-readable storage medium is provided having a computer program or instructions stored thereon, characterized in that the computer program or instructions, when executed by a processor, implement the steps of a method for constructing a pancreatic fistula risk assessment model.

[0014] According to a sixth aspect of the present invention, a computer program product is provided, comprising a computer program or instructions, characterized in that, when the computer program or instructions are executed by a processor, the steps of a method for constructing a pancreatic fistula risk assessment model are implemented. Attached Figure Description

[0015] To achieve the foregoing and related objectives, certain illustrative aspects are described herein in conjunction with the following description and accompanying drawings. These aspects indicate various ways in which the principles disclosed herein may be practiced, and all aspects and their equivalents are intended to fall within the scope of the claimed subject matter. The foregoing and other objectives, features, and advantages of this disclosure will become more apparent from the following detailed description, taken in conjunction with the accompanying drawings. Throughout this disclosure, the same reference numerals generally refer to the same parts or elements.

[0016] Figure 1 A schematic diagram of a method 100 for constructing a pancreatic fistula risk assessment model according to an embodiment of the present invention is shown; Figure 2 A nomogram showing the risk of pancreatic fistula after pancreaticoduodenectomy according to one embodiment of this application is presented; Figure 3 The ROC curve of a pancreatic fistula risk assessment model after pancreaticoduodenectomy according to an embodiment of this application is shown; Figure 4 A schematic diagram showing the model performance calibration curve results according to an embodiment of this application is illustrated; Figure 5 A schematic diagram showing the clinical decision curve results of a model performance according to an embodiment of this application is illustrated; Figure 6 A schematic diagram of a pancreatic fistula risk assessment model construction apparatus 600 according to an embodiment of the present invention is shown. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. The same reference numerals generally refer to the same parts or elements.

[0018] Given the harm caused by pancreatic fistula and patients' needs for rehabilitation and quality of life, it is crucial to identify the precipitating factors, preventive techniques, and early diagnosis and management strategies for pancreatic fistula after pancreaticoduodenectomy. Currently, there is a lack of objective and effective methods for constructing pancreatic fistula risk assessment models; therefore, this application is submitted.

[0019] Figure 1 A schematic diagram of a method 100 for constructing a pancreatic fistula risk assessment model according to an embodiment of the present invention is shown. Figure 1 As shown, step 110 is executed first, and initial features are obtained by feature extraction based on the acquired patient data. The patient data includes CT imaging data and clinical data. The initial features include CT features of the pancreas and spleen, as well as clinical features.

[0020] According to one embodiment, clinical data and CT imaging data of patients after pancreaticoduodenectomy are first acquired. When acquiring patient data, certain inclusion and exclusion criteria are set, and patient data is acquired according to these criteria. Exclusion criteria include: patients with pre-existing pancreatic fistula, and patients who underwent pancreaticoduodenectomy due to other complications. The acquired clinical data of pancreaticoduodenectomy patients includes age, sex, body mass index, history of jaundice reduction (PTCD or ERCP), preoperative prophylactic use of somatostatin analogues, preoperative serum bilirubin level, preoperative serum albumin level, operation duration, amylase content in drainage fluid on the first day, serum bilirubin level on the third postoperative day, and serum albumin level on the third postoperative day. CT imaging data includes pancreatic density, spleen density, and diameter of the main pancreatic duct. This application improves the objectivity of the data, reduces manual judgment, and avoids subjective errors by incorporating CT imaging data as a basis for judging pancreatic condition.

[0021] Next, the CT imaging data and clinical data were preprocessed to obtain preprocessed data. The preprocessing steps included missing value handling and categorical transformation. Missing value handling included: when missing values ​​were <5%, simple imputation was used to fill in the missing values ​​using the mean, median, or mode. When the missing value in a column (one feature dimension) accounts for less than 5% of the total data in that column, it is considered "few missing values" and can be directly filled using the mean, median, or mode of that column. This approach is suitable for scenarios where the missing value is small and the imputation error has little impact on the overall data distribution. When filling missing values, the mean is suitable for continuous data with an approximately normal distribution (such as height and weight), the median is suitable for continuous data with outliers, and the mode is suitable for categorical data (such as gender).

[0022] If the missing value ratio is ≥40%, the data in that dimension will be removed. This is because too many missing values ​​will result in extremely low reliability of imputation, and forced imputation will introduce a lot of noise. Retaining the dimension will not only fail to provide effective information, but may also interfere with model training or data analysis results.

[0023] For features falling between these two extremes, the decision to imput or retain them is based on the feature importance assessment results. If a feature is highly important (e.g., plays a crucial role in prediction results or analysis conclusions), a suitable imputation method (which can be simple imputation or more complex methods such as KNN imputation or model imputation) should be selected to retain the feature. If a feature is of low importance (e.g., has negligible impact on the results), it can be directly removed to reduce data redundancy.

[0024] Categorical data transformation involves converting non-numerical categorical variables into binary variables (dummy variables) containing only "0" and "1", enabling computers or models to recognize and process them. Categorical variables are data divided according to "categories" and have no magnitude or order significance; binarization mapping turns each "category" into an independent variable, taking only 0 or 1 values. The binary dummy variable is the final generated "0 / 1 variable," essentially a "numerical encoding" of the original categories, allowing the model to understand it. This application improves data accuracy through data missing value imputation and binarization mapping preprocessing.

[0025] According to one implementation, feature extraction is performed on preprocessed data or directly on acquired data to obtain CT features of the pancreas and spleen, as well as clinical features.

[0026] Subsequently, step 120 is executed, where the initial features are screened based on their correlation with the risk of pancreatic fistula, resulting in the first screening features. Specifically: the correlation determination method is determined based on the data type of the initial features. Data types include continuous normally distributed data, continuous non-normally distributed data, ordered categorical data, and unordered categorical data. The correlation determination methods include independent samples t-test, nonparametric verification, and chi-square test. The correlation of the initial features is determined based on the determined correlation determination method, and the correlation is determined according to the probability value p. If the probability value p is less than a preset probability threshold, the initial features are determined as the first screening features.

[0027] According to one implementation method, multiple testing methods can be applied to the data simultaneously, or the testing methods can be determined based on the data type. For example, the data types of CT imaging data and clinical data are first determined, including continuous normally distributed data, continuous non-normally distributed data, ordered categorical data, and unordered categorical data. For continuous normally distributed data, an independent samples t-test can be used to compare whether there is a significant difference in the characteristic means between the pancreatic fistula group and the non-pancreatic fistula group; for ordered and unordered categorical data, a non-parametric test (Mann-Whitney U test) can be used to compare whether there is a significant difference in the characteristic distribution positions (median or rank) between the two groups; for unordered categorical data, a chi-square test can be used to test whether there is an association between the characteristic category and the occurrence of pancreatic fistula.

[0028] According to one implementation method, the independent samples t-test is specifically implemented as follows:

[0029] Where t is a statistic. The mean values ​​are for the two groups (pancreatic fistula group and no pancreatic fistula group). This is the standard error of the mean difference. These represent the overall mean values ​​for the pancreatic fistula group and the group without pancreatic fistula, respectively.

[0030] According to one implementation method, the specific implementation of nonparametric validation is as follows:

[0031]

[0032] Where Z is the statistic. The rank sum of the pancreatic atrophy group, The sample sizes for the pancreatic fistula group and the non-pancreatic fistula group are respectively. , Let j be the number of identical ranks. .

[0033] According to one implementation method, the chi-square test is specifically implemented as follows:

[0034] in, For statistical purposes, and represents the number of variables and the number of samples, respectively; i represents the values ​​of the variables; and j represents the sample values ​​of patients who underwent pancreaticoduodenectomy. This represents the actual value of a variable that takes the value i and belongs to the j-th pancreaticoduodenectomy patient sample. Let represent the expected value of the variable i belonging to the j-th pancreaticoduodenectomy patient sample, where The definition is as follows:

[0035] According to one embodiment, this application collected a dataset of 518 patients, including 286 males (55.2%) and 232 females (44.8%), with an average age of 61.25 ± 11.10 years and a BMI of 22.61 ± 3.20 kg / m². A total of 86 patients developed post-pancreatic pulmonary fibrosis (POPF), with an incidence of 16.6% in the modeling group. Of these, 47 were male (54.7%) and 39 were female (45.3%), with an average age of 62.27 ± 11.52 years and a BMI of 24.14 ± 3.60 kg / m². Table 1 shows the results of the first screening of clinical and CT imaging characteristics of patients after pancreaticoduodenectomy. According to one embodiment, the preset probability threshold can be set to 0.001. As shown in Table 1, the risk of postoperative pancreatic fistula was significantly correlated with the diameter of the main pancreatic duct, the pancreatic atrophy ratio (pancreatic density / spleen density), the amylase content of the drainage fluid on the first postoperative day, the serum albumin level on the third postoperative day, and the body mass index, but not with other indicators.

[0036]

[0037] Table 1 According to one implementation method, the Box-Tidwell method is used to test the linear relationship between the first screening feature and LogitP. For those exhibiting a non-linear relationship, the continuous variable is converted into a categorical variable, and the optimal classification cutoff value is obtained using Receiver Operating Characteristic (ROC) curve analysis based on the principle of maximizing the Youden index. Specifically, this includes: First, the Box-Tidwell method is used to test the linear relationship between the first screening feature and Logit P. The Box-Tidwell method is a statistical method specifically designed to verify whether a continuous independent variable is linearly related to Logit P, and it is a prerequisite test tool for Logistic regression. Logit P is the logarithmic odds ratio of pancreatic fistula occurrence, i.e., Log(probability of pancreatic fistula occurrence / probability of not occurring), and it is the core dependent variable in Logistic regression. A linear relationship means that the change in Logit P is "uniform" as the feature value changes (e.g., for every 1 mm decrease in the diameter of the main pancreatic duct, Logit P increases uniformly). The core logic of the test is that if the test result p > 0.05, it indicates that "a linear relationship exists between the feature and Logit P," and the feature can be directly included in the Logistic regression model (no transformation required); if the test result p ≤ 0.05, it indicates that "a non-linear relationship exists," and the feature is not suitable for direct modeling and needs to be transformed into a categorical variable (e.g., "≤3mm" is high risk, ">3mm" is low risk).

[0038] Subsequently, given the existence of a non-linear relationship, the continuous variable is converted into a categorical variable. This involves dividing the continuous variable into two groups based on the risk inflection point (e.g., main pancreatic duct diameter ≤3mm and >3mm), transforming it into a categorical variable with clear risk differentiation. The key to this conversion is finding the cutoff point that best distinguishes between high and low risk. To determine the appropriate risk level, ROC curve analysis can be used, and the optimal classification cutoff value can be obtained by maximizing the Youden index.

[0039] The ROC curve is a curve plotted with sensitivity (true positive rate) on the vertical axis and 1-specificity (false positive rate) on the horizontal axis. It is used to assess the discriminative power of categorical variables. For example, using the diameter of the main pancreatic duct as the basis, ROC curves are plotted to distinguish between the occurrence and absence of pancreatic fistula when different diameters are used as cutoff values. Sensitivity (true positive rate) is the proportion of patients who are correctly judged as "high risk" among those who actually have pancreatic fistula; specificity (true negative rate) is the proportion of patients who are correctly judged as low risk among those who do not actually have pancreatic fistula; the Youden index measures the overall discriminative power of the cutoff value; the optimal cutoff value is the characteristic value corresponding to the maximum Youden index, which is the best dividing point for distinguishing between high and low risk.

[0040] According to one implementation method, when determining the optimal cutoff value for the main pancreatic duct diameter, the main pancreatic duct diameter (continuous values: 2mm, 2.5mm, 3mm, 3.5mm) and pancreatic fistula outcome (occurrence or non-occurrence) of all patients are collected. Each diameter value is tried as a cutoff value one by one, and the corresponding sensitivity and specificity are calculated (e.g., using 2mm as the boundary: ≤2mm → high risk, >2mm → low risk, calculate sensitivity / specificity; then use 2.5mm as the boundary, and repeat the calculation). Each cutoff value corresponds to one Youden index, and an ROC curve is plotted (each point corresponds to the sensitivity / specificity of one cutoff value). The point with the largest Youden index on the ROC curve is found, and the corresponding diameter value is the optimal cutoff value, such as 3mm.

[0041] According to one implementation, the final output is a clinically quantifiable cutoff value, wherein the cutoff value for the main pancreatic duct diameter is 3 mm, the cutoff value for the pancreatic decline ratio using variance is 0.785, and the cutoff value for the amylase content of the drainage fluid on the first day is 966.5 U / L.

[0042] The optimal cutoff value plays a crucial role in subsequent model building and risk assessment. It transforms continuous variables into categorical variables. When a feature exhibits a non-linear relationship with the Logit P (log-odds ratio) of pancreatic fistula occurrence (e.g., Box-Tidwell test showing p≤0.05), continuous variables cannot be directly included in the Logistic regression model. In this case, the optimal cutoff value serves as a boundary point, converting continuous variables (such as main pancreatic duct diameter and postoperative drainage amylase content) into categorical variables with clear risk distinctions (e.g., "main pancreatic duct diameter ≤3mm" and ">3mm"), thus resolving the non-linearity issue and making the feature usable for subsequent modeling.

[0043] The transformed categorical variables (grouped based on the optimal cutoff value) are key inputs for constructing a pancreatic fistula risk assessment model. For example, grouping characteristics such as main pancreatic duct diameter (≤3mm / >3mm) and postoperative day 1 drainage amylase (≥966.5U / L / <966.5U / L) are incorporated into LASSO regression screening, ultimately becoming the third screening feature, directly participating in model training (e.g., multivariate logistic regression model), improving the model's accuracy and stability. The optimal cutoff value is also a key parameter in ROC curve analysis, directly affecting the model's sensitivity (true positive rate) and specificity (true negative rate). The cutoff value determined in the document using the principle of maximizing Youden's index ensures that the model achieves an optimal balance in distinguishing between pancreatic fistula and non-pancreatic fistula patients, providing a reliable basis for subsequent model performance evaluation (e.g., AUC value, calibration curve).

[0044] According to one implementation, a collinearity diagnosis is performed on first screening features using the expansion factor (VIF). This includes: determining the expansion factor of the first screening feature; and removing the first screening feature if its expansion factor exceeds a preset expansion factor threshold. Collinearity diagnosis uses the VIF statistical index to detect whether there is information overlap (collinearity) among the selected first screening features (such as main pancreatic duct diameter, drainage fluid amylase, etc.), avoiding result distortion during subsequent modeling. This allows for secondary verification of feature quality, ensuring that each screening feature independently provides unique information for pancreatic fistula risk prediction. Collinearity refers to a high correlation between multiple screening features, causing their information to overlap and be superimposed. If both are included in the model simultaneously, the model cannot distinguish which feature is affecting pancreatic fistula risk, ultimately leading to inaccurate risk assessment results.

[0045] The core function of the Variance Inflation Factor (VIF) is to measure the information independence of a feature. The higher the VIF value, the more the feature is explained by other features, and the more severe the collinearity. If VIF=1: the feature is completely independent of other features (no collinearity), and the information is 100% unique. If VIF=5: the feature has high "information redundancy," with 50% of its information replaceable by other features. If VIF=10: collinearity is severe, and most of the feature's information is covered by other features; continuing to retain it would seriously interfere with the model.

[0046] Table 2 shows the collinearity diagnosis results for the first screening feature. As can be seen from Table 2, the tolerance of all first screening features is greater than 0.1, and the variance inflation factor is less than 10.

[0047]

[0048] Table 2 Subsequently, step 130 is executed, which uses mutual information to determine the mutual information value of the first screening feature and determines the first screening feature with a mutual information value greater than a preset threshold as the second screening feature; including: determining the marginal distribution of the first screening feature and the outcome variable of pancreatic fistula, and the joint probability distribution of the first screening feature and the outcome variable; determining the mutual information value of the first screening feature based on the marginal distribution and the joint probability distribution; and determining the first screening feature as the second screening feature if the mutual information value of the first screening feature is greater than the preset threshold.

[0049] According to one implementation method, mutual information (MI) is used to statistically analyze the nonlinear dependency between two variables. The mutual information value between features and pancreatic fistula is calculated using Scikit-learn's `mutual_info_regression` function. The core advantage of mutual information is that it is independent of the variable distribution type, capturing both linear relationships and accurately quantifying nonlinear dependencies. This allows for the selection of key features missed by traditional tests, forming a second set of screening features and making the features input to the subsequent model more comprehensive.

[0050] Mutual information I(X;Y) is an indicator that measures the degree of information sharing between two variables. That is, given a feature X (such as the pancreatic decline ratio), to what extent can we determine the outcome Y (whether pancreatic fistula occurs)? The more information shared (the larger I(X;Y) is), the stronger the association between the feature and pancreatic fistula (whether linear or nonlinear). The order of feature importance is determined by ranking mutual information values ​​from largest to smallest, and the following is retained: The top K features with the highest mutual information values ​​were selected using the Top-K method to complete the data feature screening and obtain the second-selection clinical and CT imaging features. The data feature selection calculation method based on mutual information is as follows:

[0051] in, Let Y be the feature to be selected, and let Y be the outcome variable. It is a joint probability distribution; and These are the marginal distributions of feature X and outcome variable Y, respectively.

[0052] When using the mutual information method to screen features, continuous features are first preprocessed using equal-frequency binning (bin number K=10) to eliminate dimensional differences and enhance the stability of mutual information estimation. Different features have large differences in units or ranges (e.g., BMI is 20-30, postoperative drainage fluid amylase is hundreds to thousands U / L), and direct calculation of mutual information would be affected by dimensions. Binning unifies them into "categories," eliminating dimensional interference. Secondly, it enhances the stability of mutual information estimation. The probability distribution of continuous features is difficult to estimate accurately; discretization (dividing into a finite number of categories) reduces estimation errors and makes mutual information values ​​more reliable. Equal-frequency binning refers to dividing all data of a continuous feature into 10 intervals (bins) with roughly the same number of samples in each bin.

[0053] Secondly, based on the mutual_info_regression algorithm of the Scikit-learn machine learning framework, the system quantifies the nonlinear association strength between each feature and the postoperative pancreatic fistula risk. mutual_info_regression is a tool in the Scikit-learn framework specifically designed to quantify the nonlinear association strength between continuous or discrete features and continuous or binary outcomes, making it suitable for this scenario. Furthermore, compared to manual calculation, this function can automatically process binned features and quickly output the mutual information value between each feature and the pancreatic fistula, resulting in higher efficiency. The function inputs are preprocessed feature data (such as pancreatic decline ratio, main pancreatic duct diameter, etc.) and pancreatic fistula outcome data (0 / 1); the output is the mutual information value corresponding to each feature (such as pancreatic decline ratio MI=0.92, main pancreatic duct diameter MI=0.88, BMI=0.65); its function is to quantify the nonlinear association strength between each feature and the pancreatic fistula, providing a basis for subsequent ranking.

[0054] Finally, a dynamic threshold mechanism was set to retain statistically significant mutual information features while eliminating redundant variables to control model complexity. First, features with mutual information values ​​≥ θ were retained (θ is a preset dynamic threshold to ensure that the retained features have a statistically significant association with pancreatic fistula, while redundant features with excessively low mutual information values, such as gender (MI = 0.05), were removed). Then, the retained features were sorted from largest to smallest mutual information value, and the top K (Top-K) features with the highest mutual information values ​​were selected (K is a preset value to control the number of features and avoid excessive features leading to model complexity). Finally, through these two steps, the second set of selected features was obtained.

[0055] The second screening feature, constructed after multi-stage screening, includes six core variables: pancreatic atrophy ratio, main pancreatic duct diameter, preoperative serum albumin level, amylase content in drainage fluid on the first postoperative day, serum albumin level on the third postoperative day, and body mass index. This application employs the mutual information method to avoid the difficulty of capturing complex nonlinear correlations by traditional linear methods. The order of mutual information values ​​from largest to smallest determines the order of feature importance, further clarifying the target data features.

[0056] According to one implementation, the second screening feature can be screened using the Least Absolute Contraction and Selection Operator (LASSO) method to obtain the third screening feature; including: determining the regression coefficient of the second screening feature using the Least Absolute Contraction and Selection Operator method; if the regression coefficient of the second screening feature is non-zero, using the second screening feature as the third screening feature; and constructing a logistic regression model based on the third screening feature to obtain a pancreatic fistula risk assessment model.

[0057] The unique advantage of the LASSO method is its ability to simultaneously shrink regression coefficients and select features, eliminating redundant or weakly contributing features to make the final feature set more concise and better suited to subsequent prediction models. LASSO is a regression method with a "penalty term" that can shrink and select. Shrinkage includes shrinking the regression coefficients of unimportant features towards 0 to reduce their impact on the model. Selection includes directly compressing the regression coefficients of "very weakly contributing" features to 0 when the penalty is strong enough; these features with coefficients of 0 are automatically eliminated, retaining only features with non-zero coefficients.

[0058] LASSO scores each second screening feature (the size of the regression coefficient represents the contribution), eliminates features with scores close to 0, and keeps only the core features with significantly non-zero scores (i.e., the third screening features).

[0059] The method for selecting the third screening feature using the minimum absolute value contraction and selection operator method is as follows:

[0060] Where Y is an n×1 vector, Y represents the actual value corresponding to sample X, X is an n×p matrix, X represents the input sample for LASSO regression, and β = (β1, β2, β3) / (β4, β5). p ) T It is a p×1 regression coefficient vector. It is a disciplinary measure. >0 is an adjustment parameter used to balance the penalty period and experience risk.

[0061] LASSO analysis can eliminate weakly contributing features and simplify the feature set while retaining the core predictive information of features, and at the same time optimize the risk weight of features, thereby improving the stability and generalization ability of subsequent models.

[0062] The steps to obtain the third screening features based on the second screening features are as follows: First, prepare the input data, including: Feature matrix X: an n×6 matrix (n = number of patients, 6 columns = 6 features of the second screening, already standardized / preprocessed); Outcome vector Y: an n×1 binary vector (pancreatic fistula occurrence = , no occurrence = 0). Then, determine the optimal penalty coefficient λ (a key step) and use k-fold cross-validation (commonly 5-fold / 10-fold): divide the data into k parts, alternately use k-1 parts to train the model and 1 part to validate the model, and finally select the λ value corresponding to the smallest prediction error on the validation set (optimal λ); the purpose is to avoid λ being too large, leading to the elimination of important features, or λ being too small, leading to the retention of redundant features, and to balance the fitting effect and feature simplification.

[0063] Next, a LASSO regression model is fitted, and the optimal λ value is substituted into the LASSO formula to fit the model and calculate the regression coefficient β for each feature.j The core logic is that LASSO's L1 penalty term compresses coefficients, causing the coefficients of features with weak contributions to approach zero. Next, non-zero coefficient features are selected, resulting in a third set of features. The retention rule is to only retain the regression coefficient β. j Features ≠0, remove β j Characteristics of =0; According to one implementation, the final third screening characteristics are: pancreatic degeneration ratio, main pancreatic duct diameter, amylase content in drainage fluid on the first postoperative day, serum albumin level on the third postoperative day, and body mass index (BMI).

[0064] Finally, step 140 is executed to construct a logistic regression model based on the second screening feature to obtain a pancreatic fistula risk assessment model; this includes: using the second screening feature as an input variable to determine the probability that the input variable belongs to the positive or negative class; determining the parameters and regression coefficients in the logistic regression model based on the probability of belonging to the positive or negative class; and determining the logistic regression model based on the parameters and regression coefficients.

[0065] According to one implementation, a pancreatic fistula risk assessment model can be constructed based on a second screening feature or a third screening feature. Specifically, the pancreatic fistula risk assessment model can be implemented as a logistic regression model determined using multivariate logistic regression, including:

[0066]

[0067] in, Input variables The probability of belonging to the positive class. Input variables The probability of belonging to the negative class. It is a constant. Input variables The regression coefficients. Input variables These are the second or third screening features that are substituted into the calculation.

[0068] Parameters in a logistic regression model and regression coefficients The calculation method is as follows: ; in, Indicates estimation The likelihood function, Indicates the number of samples. Indicates input variables The predicted value.

[0069] The established logistic regression model is as follows: Logit(p)=-0.5849-2.513MPD-1.172PAI+1.651AMY+0.104BMI According to one implementation, when constructing the logistic regression model, features are also screened based on p-values; among them, the p-value of serum albumin on the third postoperative day is >0.05, which is not statistically significant in logistic regression analysis, and therefore is removed from the final logistic regression model; the final logistic regression model includes pancreatic regression ratio, main pancreatic duct diameter, amylase content in drainage fluid on the first postoperative day, and body mass index.

[0070] According to one implementation, after constructing a pancreatic fistula risk assessment model, a nomogram for predicting the risk of pancreatic fistula after pancreaticoduodenectomy is determined based on the pancreatic fistula risk assessment model to predict the postoperative pancreatic fistula risk; the performance of the nomogram is also evaluated using the area under the receiver operating characteristic curve, calibration curve, and decision curve.

[0071] Figure 2 A nomogram of the risk of pancreatic fistula after pancreaticoduodenectomy according to an embodiment of this application is shown. The method for constructing the nomogram of the pancreatic fistula risk assessment model is as follows: each value level of the influencing factor is scored according to the magnitude of the regression coefficient in the pancreatic fistula risk assessment model, and then the scores are added together to obtain the total score; the probability of postoperative pancreatic fistula is obtained by positioning the total score on the total score scale line.

[0072] Figure 3 The ROC curve of a pancreatic fistula risk assessment model after pancreaticoduodenectomy according to an embodiment of this application is shown. Figure 3 As shown, the AUC value is 0.808.

[0073] Figure 4 A schematic diagram showing the model performance calibration curve results according to one embodiment of this application is illustrated. Figure 4 As shown, the test results for the goodness of fit of HL are χ²=3.5831, degrees of freedom=8, P=0.8926>0.05, indicating that it has a high degree of calibration.

[0074] Figure 5 A schematic diagram showing the clinical decision curve results of a model performance according to an embodiment of this application is illustrated, as follows: Figure 5 As shown, the decision curve lies above the two extreme lines, indicating that it has some clinical usefulness.

[0075] According to one embodiment, this application also provides a method for assessing the risk of pancreatic fistula, the method comprising: acquiring CT imaging data and clinical data of the patient to be assessed; determining input features based on the CT imaging data and clinical data; the input features including pancreatic deterioration ratio, main pancreatic duct diameter, amylase content in drainage fluid on the first postoperative day, serum albumin value on the third postoperative day, and body mass index; inputting the input features into a pancreatic fistula risk assessment model to determine the patient's risk of developing pancreatic fistula.

[0076] Figure 6 A schematic diagram of a pancreatic fistula risk assessment model construction device 200 according to an embodiment of the present invention is shown. The following will be combined with... Figure 6 The device is described in detail. Figure 6 As shown, the pancreatic fistula risk assessment model construction device 600 includes an initial feature extraction module 610, a first screening feature determination module 620, a second screening feature determination module, a model training module 630, and a logistic regression model construction module 640.

[0077] The initial feature extraction module 610 is used to extract initial features based on the acquired patient data, including CT imaging data and clinical data. The initial features include CT features of the pancreas and spleen, as well as clinical features. In one embodiment, the initial feature extraction module 610 can be used to perform step 110 described above, which will not be repeated here.

[0078] The first screening feature determination module 620 is used to screen the initial features based on the correlation between the initial features and the risk of pancreatic fistula, thereby obtaining the first screening features. In one embodiment, the first screening feature determination module 620 can be used to perform step 120 described above, which will not be repeated here.

[0079] The second screening feature determination module 630 is used to determine the mutual information value of the first screening feature using the mutual information method, and to determine the first screening feature whose mutual information value is greater than a preset threshold as the second screening feature. In one embodiment, the second screening feature determination module 630 can be used to perform step 130 described above, which will not be repeated here.

[0080] The logistic regression model construction module 640 is used to construct a logistic regression model based on the second screening features to obtain a pancreatic fistula risk assessment model. In one embodiment, the logistic regression model construction module 640 can be used to perform step 140 described above, which will not be repeated here.

[0081] It should be noted that the storage medium (computer-readable medium) described above in this invention can be a computer-readable signal medium, a non-transitory computer-readable storage medium, or any combination thereof. A non-transitory computer-readable storage medium can be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a non-transitory computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, and portable compact disk read-only memory (CD-ROM). ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0082] In this invention, a non-transitory computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a non-transitory computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0083] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0084] The above description is merely a partial embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0085] Furthermore, although the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in sequential order. Multitasking and parallel processing may be advantageous in certain environments. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0086] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for constructing a pancreatic fistula risk assessment model, the method comprising: Initial features are obtained by feature extraction based on the acquired patient data, which includes CT imaging data and clinical data. The initial features include CT features of the pancreas and spleen, as well as clinical features. The initial features are screened based on their correlation with the risk of pancreatic fistula to obtain the first screening features; The mutual information method is used to determine the mutual information value of the first screening feature, and the first screening feature with a mutual information value greater than a preset threshold is determined as the second screening feature; Based on the second screening feature, a logistic regression model is constructed to obtain the pancreatic fistula risk assessment model.

2. The method as described in claim 1, wherein, The process of filtering the initial features based on their correlation with the risk of pancreatic fistula to obtain the first screening features includes: The correlation determination method is determined based on the data type of the initial features. The data types include continuous normally distributed data, continuous non-normally distributed data, ordered categorical data, and unordered categorical data. The correlation determination method includes independent samples t-test, nonparametric verification, and chi-square test. The correlation of the initial features is determined according to the determined correlation determination method, wherein the correlation is determined based on the probability value p; If the probability value p is less than a preset probability threshold, the initial feature is determined to be the first screening feature.

3. The method as described in claim 2, wherein, The step of determining the mutual information value of the first screening feature using the mutual information method, and determining the first screening feature with a mutual information value greater than a preset threshold as the second screening feature includes: Determine the marginal distributions of the first screening feature, the outcome variable of pancreatic fistula, and the joint probability distribution of the first screening feature and the outcome variable; The mutual information value of the first screening feature is determined based on the edge distribution and the joint probability distribution of the edge distribution; If the mutual information value of the first screening feature is greater than the preset threshold, the first screening feature is determined to be the second screening feature.

4. The method according to claim 1, wherein, The step of constructing a logistic regression model based on the second screening features to obtain the pancreatic fistula risk assessment model includes: Using the second screening feature as an input variable, determine the probability that the input variable belongs to the positive or negative class; The parameters and regression coefficients in the logistic regression model are determined based on the probability of belonging to the positive or negative class. The logistic regression model is determined based on the parameters and the regression coefficients.

5. The method according to any one of claims 1-4, wherein, The method further includes: The regression coefficients were determined using the minimum absolute shrinkage and selection operator method for the second screening feature; If the regression coefficient of the second screening feature is non-zero, the second screening feature is used as the third screening feature. Based on the third screening feature, a logistic regression model is constructed to obtain the pancreatic fistula risk assessment model.

6. The method according to any one of claims 1-4, wherein, The method further includes: Determine the inflation factor of the first screening feature; If the inflation factor of the first screening feature is greater than a preset inflation factor threshold, the first screening feature is removed.

7. A method for assessing the risk of pancreatic fistula, the method comprising: Obtain CT imaging data and clinical data of the patient to be evaluated, and determine input features based on the CT imaging data and clinical data. The input features include pancreatic deterioration ratio, main pancreatic duct diameter, amylase content of drainage fluid on the first postoperative day, serum albumin value on the third postoperative day, and body mass index. The input features are input into the pancreatic fistula risk assessment model to determine the risk of the patient developing pancreatic fistula. The pancreatic fistula risk assessment model is trained according to the pancreatic fistula risk assessment model construction method according to any one of claims 1-6.

8. A device for constructing a pancreatic fistula risk assessment model, the device comprising: An initial feature extraction module is used to extract initial features from the acquired patient data, which includes CT imaging data and clinical data. The initial features include CT features of the pancreas and spleen, as well as clinical features. The first screening feature determination module is used to screen the initial features based on the correlation between the initial features and the risk of pancreatic fistula, so as to obtain the first screening features; The second screening feature determination module is used to determine the mutual information value of the first screening feature using the mutual information method, and to determine the first screening feature whose mutual information value is greater than a preset threshold as the second screening feature. The logistic regression model construction module is used to construct a logistic regression model based on the second screening feature to obtain a pancreatic fistula risk assessment model.

9. A computing device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1-7.

Images