Construction method and system of pancreatic fistula occurrence probability prediction model after far-end pancreas operation

Abstract

The invention provides a construction method and system for a prediction model of the probability of occurrence of pancreatic fistula after a far-end pancreatic operation. The constructed model is used for predicting the risk of pancreatic fistula after a pancreatic tumor operation. According to the method, preoperative nuclear magnetic resonance (MRI) parameters, radiomics features and key clinical indexes are fused, a feature set highly related to pancreatic fistula occurrence is extracted through feature screening and multi-factor regression analysis, a prediction model with optimized performance is constructed, and further, the complex prediction model is converted into a Nomogram which is convenient for doctors to understand and apply. Meanwhile, the invention also provides a system for realizing the method, which comprises a data collection module; a radiomics feature extraction module; a clinical feature analysis module; a model construction and screening module; and a model verification and visualization module.

Description

Technical Field

[0001] This invention relates to the field of medical technology, specifically to a method and system for constructing a predictive model for the probability of pancreatic fistula after distal pancreatic surgery. Background Technology

[0002] Pancreatic tumors are malignant tumors with a high incidence rate in the digestive system. Distal pancreatectomy (DP) is the standard surgical procedure for treating benign and malignant tumors in the pancreatic body and tail, and it has achieved significant clinical efficacy. However, postoperative pancreatic fistula (POPF) is a common and potentially life-threatening complication that can lead to serious consequences such as abdominal infection and massive bleeding, and may even cause organ failure and death in patients.

[0003] Currently, various scoring systems exist for predicting the probability of postoperative pancreatic fistula (POPF), including predictive indicators such as pancreatic texture, main pancreatic duct size, and pancreatic thickness. However, predictive indicators such as pancreatic texture rely heavily on the surgeon's palpation and experience during surgery, making them highly subjective and prone to errors. Furthermore, the fact that predictive parameters can only be obtained intraoperatively limits the preoperative application of these scoring systems, diminishing their value in providing preoperative decision-making support. Therefore, there is a need to develop a model that can predict the probability of POPF before surgery.

[0004] Studies have shown that pancreatic fibrosis and fatty infiltration can affect pancreatic texture, thereby influencing the occurrence of postoperative pancreatic fistula (POPF). Therefore, preoperative assessment of pancreatic fibrosis and fatty infiltration can indirectly reflect the probability of POPF. The MTR (magnetization transfer rate) parameter in preoperative MRI reflects the interaction between macromolecules and free water in pancreatic tissue and is positively correlated with histological fibrosis. Previous studies have also shown a correlation between FF (pancreatic fat fraction) obtained from MRI and histopathology. The ADC (apparent diffusion coefficient) value of DWI (diffusion-weighted imaging) can reflect the diffusion of water molecules within pancreatic tissue, helping to assess the cell density and structural integrity of pancreatic tissue.

[0005] Clinical laboratory indicators such as ALB (albumin) and WBC (white blood cell count) are also important indicators reflecting a patient's preoperative physiological status. ALB (albumin) levels can reflect a patient's nutritional status and liver synthetic function, which is of great significance for assessing a patient's surgical tolerance and postoperative recovery ability; WBC (white blood cell count) can reflect the degree of a patient's inflammatory response and help to determine potential infection risks.

[0006] Meanwhile, individual characteristics such as age, gender, BMI (body mass index), and whether the patient has diabetes may also affect the probability of POPF (postoperative pancreatic fistula) to varying degrees.

[0007] As can be seen from the above analysis, there are many factors related to the probability of POPF (postoperative pancreatic fistula). It is difficult to make accurate predictions based on a single indicator. In order to improve the accuracy of preoperative prediction and provide a reliable reference for doctors and patients' decision-making, it is necessary to organically integrate multiple factors and jointly construct a predictive model for the probability of POPF. Summary of the Invention

[0008] To construct a multi-factor predictive model for the probability of postoperative pancreatic fistula (POPF) used before surgery, this invention provides a method and system for constructing a predictive model for the probability of postoperative pancreatic fistula after distal pancreatic surgery. A method for constructing a predictive model for the probability of postoperative pancreatic fistula after distal pancreatic surgery includes the following steps:

[0009] S1. Collect preoperative enhanced magnetic resonance imaging (MRI) reports of patients who have previously undergone distal pancreatic surgery (DP), collect clinical data of patients, integrate them into a dataset, and divide the dataset into training set and test set;

[0010] S2. Obtain regions of interest (ROIs) for lesions from all magnetic resonance imaging (MRI) images;

[0011] S3. Process the dataset and filter out the candidate feature set;

[0012] S4. Using pancreatic fistula or non-pancreatic fistula as the dependent variable, perform univariate and multivariate logistic regression analysis on the training set for the features in the candidate feature set to select the feature set suitable for model construction.

[0013] S5. Obtain the feature set for model construction, and use a combination of several features or all features to construct multiple prediction models on the training set; plot the receiver operating characteristic (ROC) curve on the test set; and select the best joint model by comparing prediction performance.

[0014] S6. Visualize the best joint forecasting model as a Nomograph;

[0015] S7. Validate the best joint prediction model.

[0016] Furthermore, step S1 includes:

[0017] S11. Data Acquisition: This includes clinical and imaging data; add a label for pancreatic fistula or non-pancreatic fistula to each data set.

[0018] S12. Data partitioning: Randomly partition the training set / test set proportionally to ensure class balance and that the proportion of pancreatic fistula patients and non-pancreatic fistula patients is consistent.

[0019] S13. Data cleaning: Handling missing values, filling clinical data with the mean; removing damaged files from imaging data; outlier detection, filtering extreme values ​​in clinical indicators based on the 3σ principle.

[0020] Furthermore, step S3 includes:

[0021] S31. Preprocess the region of interest of the lesion to obtain radiomics data; conduct repeatability assessment and feature screening of radiomics features to obtain a set of radiomics features;

[0022] S32. Determine the preoperative magnetic resonance imaging (MRI) parameters that need to be included in the feature screening, and collect data for the above-mentioned magnetic resonance imaging (MRI) parameters;

[0023] S33. Perform statistical analysis on the clinical data of patients in the training set, obtain variables whose differences meet the set threshold, and generate a set of clinical features.

[0024] S34. Merge the radiomics feature set, magnetic resonance (MRI) parameters, and clinical feature set into a candidate feature set.

[0025] Furthermore, in step S32, the preoperative magnetic resonance (MRI) parameters that need to be included in the feature screening include: magnetization transmissibility (MTR), fat fraction (FF), apparent diffusion coefficient (ADC), and perfusion parameters.

[0026] Furthermore, step S33 includes:

[0027] S331. Extract the original clinical data matrix of patients in the training set and divide it into two categories, continuous variables and categorical variables, according to the variable fields.

[0028] S332. Perform the Kolmogorov-Smirnov test on continuous variables to determine whether they follow a normal distribution. Normally distributed data are expressed as mean ± standard deviation.

[0029] S333. Perform an independent samples t-test on continuous variables that follow a normal distribution, and perform a Mann-Whitney U test on continuous variables that do not follow a normal distribution. Express the results in median (interquartile range) and calculate the significance of the difference between continuous variables in patients with and without pancreatic fistula.

[0030] S334. Express the categorical variable as frequency, construct a contingency table, perform the Pearson chi-square test, and calculate the significance of the difference between the categorical variable and patients with and without pancreatic fistula.

[0031] S335. Combine variables whose differences meet the set threshold into a set of clinical features.

[0032] Furthermore, step S4 includes:

[0033] S41. Perform univariate logistic regression analysis using the combined radiomics feature set, preoperative MRI parameters, and clinical feature set to calculate the regression coefficient, standard error, and p-value for each variable;

[0034] S42. Perform multivariate logistic regression analysis on single-factor variables with P values ​​less than the preset threshold. Use the forward selection method to introduce variables one by one for regression calculation and screen out feature terms that are significantly collinear with other variables.

[0035] S43. Select variables with non-zero coefficients and P-values ​​not exceeding a set threshold from the multi-factor regression to obtain the feature set for model construction.

[0036] Furthermore, step S5 includes:

[0037] S51. Select several or all features from the feature set for model construction and combine them to construct different joint input vectors. Input the joint input vectors into the multifactor logistic regression model.

[0038] S52. Train the prediction model on the training set, calculate the prediction error using the pancreatic fistula label of each patient in the training set, and iteratively adjust the parameter set of the model using the maximum likelihood estimation method to achieve the minimum average cross-entropy loss.

[0039] S53. Plot the receiver operating characteristic (ROC) curve on the test set, calculate the AUC and 95% confidence interval; determine the optimal cutoff value based on the maximum Yoden index, and calculate the accuracy, sensitivity and specificity.

[0040] S54. Based on the area under the receiver operating characteristic curve (ROC), accuracy, sensitivity, and specificity, the predictive efficacy of different prediction models is judged, and the best joint prediction model is selected.

[0041] Furthermore, step S7 includes:

[0042] S71. Randomly divide the original dataset into 3 parts (ratio 1:1:1) to ensure that the ratio of positive and negative samples in each fold is consistent with that in the original dataset; repeat the 3-fold cross-validation 5 times, selecting 2 folds as the training set and 1 fold as the test set in each iteration, plot the receiver operating characteristic curve on the test set, and calculate the AUC and 95% confidence interval; determine the optimal cutoff value based on the maximum Yangen index, and calculate the accuracy, sensitivity and specificity, and calculate the mean ± standard deviation of each performance index in the 5 repeated validations;

[0043] S72. Plot the calibration curve of the best joint prediction model, use the Hosmer-Lemeshow test to calculate the statistic and significance level, and judge the consistency between the predicted probability value of the prediction model and the true label.

[0044] S73. Set several probability threshold intervals, construct decision curves (DCA) corresponding to different models, calculate the net benefit curves respectively, plot the net benefit-probability threshold relationship graph, and evaluate the clinical efficacy of the best model. To implement the above method for constructing predictive models, this invention also provides a system for constructing a predictive model for the probability of pancreatic fistula occurrence after distal pancreatic surgery. The system includes:

[0045] Data collection module: It has system interfaces with the Hospital Information System (HIS) and Picture Archiving System (PACS) to collect patient clinical data and imaging data as required;

[0046] Radiomics Feature Extraction Module: Includes a human-computer interaction interface for lesion region (ROI) annotation and review; extracts radiomics features and magnetic resonance (MRI) parameters; and filters out sets of influencing radiomics features.

[0047] Clinical Feature Analysis Module: Performs statistical analysis on clinical data, filters out clinical feature sets, and merges different feature sets into a complete feature set;

[0048] Model building and selection module: Uses univariate logistic regression and multivariate logistic regression to select the feature set for building the model; selects several feature combinations from the feature set to form different joint variables, builds models for different joint variable groups, and outputs the model formula; plots receiver operating characteristic (ROC) curves for different models; plots decision curves (DCA) for different models; and selects the best predictive model.

[0049] Model validation and visualization module: internally validates the model; converts the model into a nomogram; clinical application interface: web or mobile tools that automatically generate a pancreatic fistula risk report after inputting patient data.

[0050] The beneficial effects of this invention are as follows:

[0051] 1. This invention constructs a predictive model for postoperative pancreatic fistula (POPF) by integrating preoperative magnetic resonance imaging (MRI) parameters and clinical indicators. Subsequent validation results demonstrate that the predictive model constructed using this method improves the accuracy of predicting the probability of postoperative pancreatic fistula before surgery, providing a reliable reference for doctors and patients to make decisions before surgery.

[0052] 2. This invention, through a reasonable feature selection method, for the first time uses magnetization transfer rate (MTR) to construct a predictive model for postoperative pancreatic fistula (POPF). The validation results of this predictive model demonstrate that magnetization transfer rate (MTR) has a strong correlation with postoperative pancreatic fistula (POPF), providing a new approach for image-based assessment of clinical risk.

[0053] 3. This invention provides clinicians with a visualization tool. By constructing a visualization nomograph, the complex model prediction results are presented to doctors in an intuitive way, making it easier for doctors to understand and apply them. Attached Figure Description

[0054] Figure 1 The present invention provides a flowchart of a method for constructing a predictive model for pancreatic fistula after distal pancreatic surgery;

[0055] Figure 2 This invention provides a structural diagram of a system for constructing a predictive model for pancreatic fistula after distal pancreatic surgery;

[0056] Figure 3 Receiver Operating Characteristic (ROC) curves of different prediction models in this embodiment of the invention;

[0057] Figure 4 Decision curve (DCA) analysis of different prediction models in this embodiment of the invention;

[0058] Figure 5 Nomograph of the optimal prediction model in this embodiment of the invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and should not be used to limit the scope of protection of the present invention.

[0060] Example 1:

[0061] To verify the feasibility of this invention in practice, it was applied to a study on predicting pancreatic fistula after distal pancreatic surgery. A multi-factor fusion model for predicting pancreatic fistula after distal pancreatic surgery was constructed, and its discriminative performance, stability, and clinical applicability were systematically verified.

[0062] S1:

[0063] The study included multiple groups of patients with pathologically confirmed pancreatic tumors. Preoperative enhanced MRI images, complete clinical data, and postoperative pathological diagnosis information were collected. Based on whether the distal pancreas developed into a pancreatic fistula after surgery, the data of each patient were labeled and randomly divided into training and test sets in a 7:3 ratio to ensure class balance (consistent proportion of patients with and without pancreatic fistulas).

[0064] S2:

[0065] All images were reviewed by senior radiologists, and another physician manually delineated the regions of interest (ROIs) layer by layer on the sequence images without MT pulses. The ROIs were located within the pancreatic parenchyma at the surgical margin and copied to the same location on the images with MT pulses and the fat fraction map. All ROIs avoided intrapancreatic blood vessels and dilated pancreatic ducts.

[0066] S3:

[0067] S31: No radiomics features were extracted in this embodiment;

[0068] S32: Preoperative magnetic resonance (MRI) parameter extraction: The magnetization transfer rate (MTR) of the lesion region (ROI) was calculated using the formula MTR = (M0 - M1) / M0 × 100%, where M0 is the signal intensity without MT pulse and M1 is the signal intensity with MT pulse; Fat fraction (FF) was obtained and recorded using a fat fraction map;

[0069] S33: In terms of constructing clinical characteristics, normality tests and difference analyses were performed on continuous variables, and chi-square tests were performed on categorical variables, identifying multiple variables with statistical significance.

[0070] S34: Combine the variables selected in the above steps to form a candidate feature set.

[0071] S4:

[0072] S41: A one-way logistic regression analysis was performed on each variable in the candidate feature set. Variables with a p-value less than or equal to 0.05 were selected as candidate variables. The results showed that the occurrence of postoperative pancreatic fistula (POPF) was related to age, white blood cell count, platelet count, postoperative amylase on the first day, and magnetization transfer rate (MTR).

[0073] S42: Multivariate logistic regression analysis was performed on the candidate variables selected in the previous step. The results showed that the magnetization transfer rate (MTR) was an independent risk factor for postoperative pancreatic fistula (POPF) (OR: 6.66 × 10⁻⁶). -10 ; 95% CI; 1.14 × 10 -17 -5.76×10-04 (P = 0.008); The white blood cell count (WBC) in the multivariate analysis had a P value of 0.105, close to 0.05, indicating that after adjusting for other factors, the association between white blood cell count (WBC) and the studied outcome was no longer significant, but a certain degree of association may still exist, although the strength of the association was weaker than in the univariate analysis; The age in the multivariate analysis had a P value of 0.071, close to 0.05, suggesting that after adjusting for other factors, a certain degree of association may still exist between age and the studied outcome, but the strength of the association was weaker than in the univariate analysis.

[0074] S43: Finally, magnetization transfer rate (MTR), white blood cell index (WBC), and age were selected as the feature combination to construct the model.

[0075] Table 1. Univariate and multivariate logistic regression analyses for predicting pancreatic fistula.

[0076]

[0077]

[0078] S51: Combine the three variables of age (Age), white blood cell count (WBC), and magnetization transfer rate (MTR) in pairs and in all groups;

[0079] S52: Train the prediction model on the training set, input different joint variables into the multifactor logistic regression model, calculate the prediction error using the pancreatic fistula label of each patient in the training set, and iteratively adjust the parameter set of the model using the maximum likelihood estimation method to minimize the average cross-entropy loss, so as to obtain prediction models with different joint variables. In this embodiment, four different prediction models are obtained: WBC+Age model, WBC+MTR model, MTR+Age model, and Age+WBC+MTR model.

[0080] S53: Plot the receiver operating characteristic curves of the four prediction models on the test set, calculate the AUC and 95% confidence interval; determine the optimal cutoff value based on the maximum Yoden index, and calculate the accuracy, sensitivity and specificity.

[0081] S74: The results showed that the AUC for predicting postoperative pancreatic fistula (POPF) was 0.706 (95% CI, 0.556-0.857) for the WBC+Age model, 0.797 (95% CI, 0.682-0.912) for the WBC+MTR model, 0.779 (95% CI, 0.653-0.905) for the MTR+Age model, and 0.802 (95% CI, 0.686-0.918) for the combined WBC+Age+MTR model. The combined WBC+ALB+MTR model showed better predictive performance than the other three models.

[0082] Table 2. Predictive performance of different models for postoperative pancreatic fistula (POPF)

[0083]

[0084] After the above screening and adjustments, the optimal joint prediction model was determined to be:

[0085] Y=3.628+(-0.0411)*Age(years)+0.301*WBC(*10^9 / L)+(-21.129)*MTR

[0086] The regression weights and intercepts in the above prediction model are obtained by iterative adjustment using the maximum likelihood estimation method in step S72.

[0087] S6:

[0088] The best joint prediction model selected will be used to generate a visual nomograph, such as... Figure 5 As shown.

[0089] S7:

[0090] S71: The performance of the selected best joint prediction model was verified by using 3-fold cross-validation repeated 5 times. The results of repeated cross-validation showed that the average AUC of the joint model was 0.724, indicating that the model has a certain ability to distinguish between positive and negative examples and can accurately classify and predict samples to a certain extent.

[0091] S72: The Hosmer-Lemeshow goodness-of-fit test showed that the chi-square value of the selected best joint prediction model was 12.902, P = 0.115 (P > 0.05), indicating a good fit;

[0092] S73: Decision curve analysis (DCA) was used to evaluate the clinical efficacy of different models. The WBC+ALB+MTR combined model performed best in the low-to-medium and high-to-medium threshold ranges; the Age+WBC model performed best at very low thresholds (close to 0.0), but its performance declined rapidly as the threshold increased; the Age+MTR model performed well in the low-to-medium range (0.2-0.4), but overall it was not as stable and efficient as the combined model. The WBC+ALB+MTR combined model provided optimal clinical efficacy in most clinical decision-making scenarios.

[0093] Overall analysis shows that the WBC+Age+MTR combined model constructed using the method provided in this invention for predicting the probability of postoperative pancreatic fistula performs well, with multiple key performance indicators maintaining a high level, and can provide a quantitative and reliable auxiliary tool for risk stratification of pancreatic fistula after pancreatic surgery.

[0094] While radiomics features were not incorporated in the above embodiments, the method for constructing a predictive model for pancreatic fistula after distal pancreatic surgery provided by this invention includes the fusion of radiomics features. Specifically, step S31 includes:

[0095] S311: Region of Interest (ROI) Marking and Review: Doctors manually delineate the ROI on MRI images, supporting multi-planar (axial, coronal) marking. The review mechanism involves senior physicians conducting secondary confirmation of the ROI to ensure the accuracy of the marking (error rate must be <5%).

[0096] S312: Image Preprocessing

[0097] Normalization: linearly scales pixel values ​​to the range of [0,1] to eliminate device differences;

[0098] Resampling: Unify image resolution to 1×1×1mm 3 To ensure comparability of features;

[0099] S313: Feature Extraction

[0100] Statistical features: histogram parameters (mean, standard deviation), contrast and correlation of gray-level co-occurrence matrix (GLCM); shape features: lesion volume, surface area, sphericity; texture features: energy and entropy after wavelet transform (implemented through PyRadiomics library);

[0101] S14: Feature selection: Calculate the intragroup correlation coefficient (ICC) and retain features with ICC ≥ 0.75 (to ensure repeatability); use LASSO regression to further reduce dimensionality and select several features with the strongest correlation to pancreatic fistula.

[0102] After completing the above steps, step S6 can be executed to summarize and process the selected impact omics feature set with other feature sets.

[0103] To achieve the above method, the present invention also provides a system for constructing a predictive model for pancreatic fistula after distal pancreatic surgery, such as... Figure 4 As shown, the system includes the following modules:

[0104] I. Data Collection Module:

[0105] (1) Data acquisition: Patient data is collected synchronously from the Hospital Information System (HIS) and the Picture Archiving System (PACS), including clinical data: WBC (white blood cells), age, gender, etc. (obtained through a structured query interface); and imaging data: enhanced MRI scan sequences (DICOM format, supporting multi-center data access).

[0106] (2) Data partitioning: The training set and test set were randomly partitioned in a 7:3 ratio to ensure class balance (the proportion of patients with pancreatic fistula and those without pancreatic fistula was consistent);

[0107] (3) Data cleaning: missing values ​​are handled, and clinical data are filled with the mean (e.g., when WBC is missing, the mean of the same disease is used as a substitute); damaged files in the imaging data are removed; outlier detection is performed, and extreme values ​​in clinical indicators are filtered based on the 3σ principle.

[0108] II. Radiomics Feature Extraction Module:

[0109] (1) Region of Interest (ROI) Marking and Review: Doctors manually delineate the region of interest (ROI) on the magnetic resonance imaging (MRI) image. Multi-planar (axial and coronal) marking is supported. The review mechanism is to have senior physicians confirm the ROI a second time to ensure the accuracy of the marking (error rate must be <5%).

[0110] (2) Image preprocessing:

[0111] Normalization: linearly scales pixel values ​​to the range of [0,1] to eliminate device differences;

[0112] Resampling: Unify image resolution to 1×1×1mm 3 To ensure comparability of features;

[0113] (3) Feature extraction:

[0114] Statistical features: histogram parameters (mean, standard deviation), contrast and correlation of gray-level co-occurrence matrix (GLCM); shape features: lesion volume, surface area, sphericity; texture features: energy and entropy after wavelet transform (implemented through PyRadiomics library);

[0115] Magnetic resonance imaging (MRI) parameter extraction: quantitative parameters such as MTR (magnetic transmissibility) and ADC (apparent diffusion coefficient) are parsed from DICOM tags; fat fraction (FF) is obtained and recorded through fat fraction maps;

[0116] (4) Feature selection: Calculate the intragroup correlation coefficient (ICC) and retain features with ICC ≥ 0.75 (to ensure repeatability); use LASSO regression to further reduce dimensionality and select 10-15 features with the strongest correlation to pancreatic fistula.

[0117] III. Clinical Feature Analysis Module:

[0118] (1) Statistical analysis: Perform univariate analysis (t test / chi-square test) on the clinical data (such as WBC, age) in the training set, screen variables with p<0.05, and generate a set of clinical input features;

[0119] (2) Data integration: Clinical features, radiomics features, and MRI parameters are combined into a complete feature set (approximately 20-30 dimensions).

[0120] IV. Model Building and Screening Module:

[0121] (1) Univariate Logistic Regression: The association between each feature and pancreatic fistula was evaluated separately, and features with significant OR values ​​(p<0.05) were selected;

[0122] (2) Multivariate Logistic Regression: The features selected by single-factor screening are included in the model, and the stepwise regression method (AIC criterion) is used to determine the final feature set;

[0123] (3) Model combination and screening: Select some variables from the feature set selected in the previous step for combination or perform full variable combination to form different joint variable groups; build a model for each joint variable group on the training set and output the model formula: logit(P)=β0+β1*X1+β2*X2+...+βn*Xn (P is the probability of pancreatic fistula); plot the receiver operating characteristic curve for each model of the joint variable group on the test set, evaluate the model performance through ROC curve (AUC value) and DCA (decision curve analysis), and select the model with the highest AUC and the largest clinical net benefit as the best model.

[0124] V. Model Validation and Visualization Module

[0125] (1) Internal validation: Three-fold cross-validation (repeated 5 times) was used for the best model to calculate the mean AUC and 95% confidence interval; the Hosmer-Lemeshow test was used to evaluate the calibration (p>0.05 indicates that the model fits well);

[0126] (2) Nomogram visualization: The best model is transformed into a Nomogram, which shows the score of each feature and the total risk probability;

[0127] (3) Clinical application interface: Develop web or mobile tools so that doctors can automatically generate pancreatic fistula risk reports (including nomographs and interpretation suggestions) after inputting patient data.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a predictive model for the probability of pancreatic fistula after distal pancreatic surgery, characterized in that... The method includes the following steps: S1. Collect preoperative enhanced magnetic resonance imaging (MRI) reports of patients who have previously undergone distal pancreatic surgery (DP), collect clinical data of patients, integrate them into a dataset, and divide the dataset into training set and test set; S2. Obtain regions of interest (ROIs) for lesions from all magnetic resonance imaging (MRI) images; S3. Process the dataset and filter out the candidate feature set; S4. Using pancreatic fistula or non-pancreatic fistula as the dependent variable, perform univariate and multivariate logistic regression analysis on the training set for the features in the candidate feature set to select the feature set suitable for model construction. S5. Obtain the feature set for model construction, and use a combination of several features or all features to construct multiple prediction models on the training set; plot the receiver operating characteristic (ROC) curve on the test set; and select the best joint model by comparing prediction performance. S6. Visualize the best joint forecasting model as a Nomograph; S7. Validate the best joint prediction model.

2. The method for constructing a prediction model for the probability of pancreatic fistula after distal pancreatic surgery according to claim 1, characterized in that... Step S1 includes: S11. Data Acquisition: This includes clinical and imaging data; add a label for pancreatic fistula or non-pancreatic fistula to each data set. S12. Data partitioning: Randomly partition the training set / test set proportionally to ensure class balance and that the proportion of pancreatic fistula patients and non-pancreatic fistula patients is consistent. S13. Data cleaning: Handling missing values, filling clinical data with the mean; removing damaged files from imaging data; outlier detection, filtering extreme values ​​in clinical indicators based on the 3σ principle.

3. The method for constructing a prediction model for the probability of pancreatic fistula after distal pancreatic surgery according to claim 1, characterized in that... Step S3 includes: S31. Preprocess the region of interest of the lesion to obtain radiomics data; conduct repeatability assessment and feature screening of radiomics features to obtain a set of radiomics features; S32. Determine the preoperative magnetic resonance imaging (MRI) parameters that need to be included in the feature screening, and collect data for the above-mentioned magnetic resonance imaging (MRI) parameters; S33. Perform statistical analysis on the clinical data of patients in the training set, obtain variables whose differences meet the set threshold, and generate a set of clinical features. S34. Merge the radiomics feature set, magnetic resonance (MRI) parameters, and clinical feature set into a candidate feature set.

4. The method for constructing a prediction model for the probability of pancreatic fistula after distal pancreatic surgery according to claim 3, characterized in that... In step S32, the preoperative magnetic resonance (MRI) parameters that need to be included in the feature screening include: magnetization transfer rate (MTR), fat fraction (FF), apparent diffusion coefficient (ADC), and perfusion parameters.

5. The method for constructing a prediction model for the probability of pancreatic fistula after distal pancreatic surgery according to claim 3, characterized in that... Step S33 includes: S331. Extract the original clinical data matrix of patients in the training set and divide it into two categories, continuous variables and categorical variables, according to the variable fields. S332. Perform the Kolmogorov-Smirnov test on continuous variables to determine whether they follow a normal distribution. Normally distributed data are expressed as mean ± standard deviation. S333. Perform an independent samples t-test on continuous variables that follow a normal distribution, and perform a Mann-Whitney U test on continuous variables that do not follow a normal distribution. Express the results in median (interquartile range) and calculate the significance of the difference between continuous variables in patients with and without pancreatic fistula. S334. Express the categorical variable as frequency, construct a contingency table, perform the Pearson chi-square test, and calculate the significance of the difference between the categorical variable and patients with and without pancreatic fistula. S335. Combine variables whose differences meet the set threshold into a set of clinical features.

6. The method for constructing a prediction model for the probability of pancreatic fistula after distal pancreatic surgery according to claim 1, characterized in that... Step S4 includes: S41. Perform univariate logistic regression analysis using the combined radiomics feature set, preoperative MRI parameters, and clinical feature set to calculate the regression coefficient, standard error, and p-value for each variable; S42. Perform multivariate logistic regression analysis on single-factor variables with P values ​​less than the preset threshold. Use the forward selection method to introduce variables one by one for regression calculation and screen out feature terms that are significantly collinear with other variables. S43. Select variables with non-zero coefficients and P-values ​​not exceeding a set threshold from the multi-factor regression to obtain the feature set for model construction.

7. The method for constructing a prediction model for the probability of pancreatic fistula after distal pancreatic surgery according to claim 1, characterized in that... Step S5 includes: S51. Select several or all features from the feature set for model construction and combine them to construct different joint input vectors. Input the joint input vectors into the multifactor logistic regression model. S52. Train the prediction model on the training set, calculate the prediction error using the pancreatic fistula label of each patient in the training set, and iteratively adjust the parameter set of the model using the maximum likelihood estimation method to achieve the minimum average cross-entropy loss. S53. Plot the receiver operating characteristic (ROC) curve on the test set, calculate the AUC and 95% confidence interval; determine the optimal cutoff value based on the maximum Yoden index, and calculate the accuracy, sensitivity and specificity. S54. Based on the area under the receiver operating characteristic curve (ROC), accuracy, sensitivity, and specificity, the predictive efficacy of different prediction models is judged, and the best joint prediction model is selected.

8. The method for constructing a prediction model for the probability of pancreatic fistula after distal pancreatic surgery according to claim 1, characterized in that... Step S7 includes: S71. Randomly divide the original dataset into 3 parts (ratio 1:1:1) to ensure that the ratio of positive and negative samples in each fold is consistent with that in the original dataset; repeat the 3-fold cross-validation 5 times, selecting 2 folds as the training set and 1 fold as the test set in each iteration, plot the receiver operating characteristic curve on the test set, and calculate the AUC and 95% confidence interval; determine the optimal cutoff value based on the maximum Yangen index, and calculate the accuracy, sensitivity and specificity, and calculate the mean ± standard deviation of each performance index in the 5 repeated validations; S72. Plot the calibration curve of the best joint prediction model, use the Hosmer-Lemeshow test to calculate the statistic and significance level, and judge the consistency between the predicted probability value of the prediction model and the true label. S73. Set several probability threshold intervals, construct decision curves (DCA) for different models, calculate the net benefit curves respectively, draw the net benefit-probability threshold relationship graph, and evaluate the clinical efficacy of the best model.

9. A system for constructing a predictive model for the probability of pancreatic fistula after distal pancreatic surgery, characterized in that... The system includes: Data collection module: It has system interfaces with the Hospital Information System (HIS) and Picture Archiving System (PACS) to collect patient clinical data and imaging data as required; Radiomics Feature Extraction Module: Includes a human-computer interaction interface for lesion region (ROI) annotation and review; extracts radiomics features; filters out the set of influencing omics features; and calculates magnetic resonance (MRI) parameters; Clinical Feature Analysis Module: Performs statistical analysis on clinical data, filters out clinical feature sets, and merges different feature sets into a complete feature set; Model building and selection module: Uses univariate logistic regression and multivariate logistic regression to select the feature set for building the model; selects several feature combinations from the feature set to form different joint variables, builds models for different joint variable groups, and outputs the model formula; plots receiver operating characteristic (ROC) curves for different models; plots decision curves (DCA) for different models; and selects the best predictive model. Model validation and visualization module: internally validates the model; converts the model into a nomogram; clinical application interface: web or mobile tools that automatically generate a pancreatic fistula risk report after inputting patient data.

Images