Pancreatic fistula risk assessment system after pancreatic duodenectomy based on multi-factor analysis

Abstract

The invention discloses a pancreatic fistula risk assessment system after pancreaticoduodenectomy based on multi-factor analysis, and relates to the technical field of medical information. Comprising a data acquisition module which is used for acquiring preoperative CT images and clinical data of a patient; the feature extraction module is used for quantitatively extracting radiomics features, pancreatic duct diameters and body component indexes from the CT images; the model construction module is used for integrating the features through multivariable logistic regression to construct a prediction model; the verification module is used for resampling through a self-service method and evaluating the performance of the model through an ROC curve; and an output module. According to the method, through multi-factor integration, the accuracy of CR-POPF preoperative prediction is remarkably improved, and the method is superior to a single-index model; and subjective deviation is reduced based on objective data, and clinical availability is improved in a network calculator form.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a pancreatic fistula risk assessment system based on multifactor analysis after pancreaticoduodenectomy. Background Technology

[0002] Pancreaticoduodenectomy is the preferred surgical procedure for treating diseases of the pancreatic head, but the incidence of postoperative complications is high. Clinically related pancreatic fistula (CR-POPF) is a common and serious complication with an incidence of approximately 10%-28%, which seriously affects patient prognosis and increases medical costs.

[0003] Currently, commonly used clinical risk assessment tools such as the pancreatic fistula risk score (FRS) and the modified pancreatic fistula risk score (A-FRS) are mainly based on intraoperative factors, which are difficult to meet the needs of preoperative assessment, and the assessment of pancreatic texture is subjective.

[0004] With the development of radiomics technology, high-throughput features extracted from CT images can be used as biomarkers to predict disease prognosis. However, existing studies are mostly limited to single indicators and lack comprehensive prediction systems that integrate multiple factors.

[0005] In addition, body composition analysis indicators such as visceral fat area (VFA) and total abdominal muscle area index (TAMAI) are associated with CR-POPF, but there is no systematic approach that combines radiomics, pancreatic duct diameter, and body composition analysis for preoperative risk assessment.

[0006] Therefore, the present invention aims to develop a multi-factor integrated risk assessment system to improve prediction accuracy through objective preoperative data. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing a multifactor analysis-based risk assessment system for pancreatic fistula after pancreaticoduodenectomy.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A multivariate analysis-based system for assessing the risk of pancreatic fistula after pancreaticoduodenectomy includes: The data acquisition module is used to acquire the patient's preoperative CT images and clinical data; A feature extraction module is used to quantitatively extract radiomics features, pancreatic duct diameter, and body composition indicators from CT images; The model building module is used to integrate the features through multivariate logistic regression to build a prediction model. The prediction model uses Rad-score, pancreatic duct diameter and VFA / TAMAI as independent variables and outputs the probability of CR-POPF occurrence. The validation module is used to evaluate model performance through bootstrap resampling and ROC curves. An output module is used to generate risk assessment results and present them in the form of a web calculator.

[0009] Preferably, the body composition index includes the ratio of visceral fat area to total abdominal muscle area index.

[0010] Preferably, the data acquisition module specifically includes a CT scan equipment parameter setting unit and a clinical data entry unit, wherein: The CT scanning equipment parameter setting unit is used to configure the enhanced CT scanning parameters, specifically including voltage, current, slice thickness, contrast agent iohexol injection flow rate, and to ensure that the portal venous phase image acquisition is delayed by 60-70 seconds. The clinical data entry unit is used to collect perioperative data in a structured manner, including patient age, BMI, smoking history, diabetes history, laboratory indicators (such as total bilirubin and albumin levels) and postoperative CR-POPF diagnosis results. The data acquisition module automatically exports DICOM format image data through the PACS system interface and extracts clinical data through the electronic medical record system. All data is anonymized and stored in an encrypted database to ensure data security and integrity, providing standardized input for subsequent feature extraction.

[0011] Preferably, the feature extraction module includes a radiomics feature extraction unit, a pancreatic duct diameter measurement unit, and a body composition analysis unit, wherein: The radiomics feature extraction unit uses 3D Slicer software to segment the ROI of the pancreatic body and tail, extracts shape features, first-order features and texture features, and filters features with ICC values ​​greater than 0.75 through intragroup correlation coefficient (ICC) analysis. Then, Lasso regression is used to select 11 feature combinations as Rad-score. The pancreatic duct diameter measuring unit manually outlines the pancreatic duct in SyngoVia software and measures the maximum diameter with an accuracy of 0.1 mm. The body composition analysis unit measures VFA at the umbilicus level with a HU threshold set from -150 to -50, and measures TAMA at the third lumbar vertebra level with a HU threshold set from -29 to +150, and calculates the VFA / TAMAI index. The feature extraction module uses a semi-automatic algorithm combined with manual correction to ensure feature accuracy and repeatability. All extracted data is output as structured tables for model building.

[0012] Preferably, the model building module employs a multivariate logistic regression algorithm, specifically including a feature weight calculation unit and a model training unit. The feature weight calculation unit determines the coefficients of each feature in the Rad-score using Lasso regression and 10-fold cross-validation, as shown in Table 2, where the intercept is -1.2034 and the coefficients of each feature range from -0.8381 to 0.6832. The model training unit uses Rad-score, pancreatic duct diameter, and VFA / TAMAI as independent variables and the CR-POPF binary classification result as the dependent variable to fit the logistic regression equation: Logit(P) = β0 + β1Rad-score + β2pancreatic duct diameter + β3*VFA / TAMAI, where β0 is the intercept, and β1, β2, and β3 are regression coefficients calculated using the maximum likelihood estimation method. The model building module is implemented in R software, using the glm function for parameter estimation, and optimizing model stability through 1000 bootstrap resampling operations to ensure that the predicted probabilities are consistent with the true values.

[0013] Preferably, the validation module includes a performance evaluation unit and a clinical utility analysis unit; wherein: The performance evaluation unit calculates the area under the curve (AUC), sensitivity, and specificity using the receiver operating characteristic (ROC) curve. The AUC of the joint model is 0.839 (95% CI: 0.757-0.907), and the DeLong test is used to compare the differences between different models. The clinical utility analysis unit plots calibration curves to assess the accuracy of predicted probability calibration, and performs decision curve analysis (DCA) to calculate the net benefit from risk thresholds ranging from 10% to 28%. The validation module uses the pROC and rmda packages in R language to implement automated scripts, and outputs validation reports including confidence intervals and statistical significance to ensure model reliability and clinical applicability.

[0014] Preferably, the output module includes a visualization unit and a network interface unit; The visualization unit generates a nodal plot that linearly superimposes Rad-score, pancreatic duct diameter, and VFA / TAMAI score to convert them into CR-POPF risk probability and provides a probability scale. The network interface unit is based on the Shiny framework to develop an online calculator, which is deployed at https: / / whyyjyljz.shinyapps.io / DynNomapp / . After the user inputs feature values ​​through the webpage, the backend R script calculates and returns the risk percentage and risk level in real time. The output module supports exporting in multiple formats, such as PDF reports or CSV files, and features user authentication to ensure data privacy, thereby improving the system's usability and accessibility in clinical settings.

[0015] A method for assessing the risk of pancreatic fistula after pancreaticoduodenectomy, comprising the following steps: S1: Data Acquisition, collecting patient data through CT scans and clinical records; S2: Feature extraction, quantitatively extracting Rad-score, pancreatic duct diameter and VFA / TAMAI from CT images; S3: Model calculation, which inputs the extracted features into the logistic regression model to obtain the risk probability; S4: Validation, verifying model performance through bootstrap resampling; S5: Output, presenting the results as a nomogram or network calculator; In this method, the Rad-score is a linear combination of 11 radiomics features, and VFA / TAMAI is calculated by HU threshold measurement. The model AUC is not less than 0.839, and the entire process is completed preoperatively to avoid interference from subjective factors during the operation, thus achieving accurate and objective risk assessment.

[0016] Step S2 includes the following steps: S21: The ROI of the pancreatic body and tail was delineated using 3D Slicer software. After extracting 1130 radiomics features, 196 features were retained through ICC analysis and correlation screening. Then, 11 features were selected through Lasso regression. S22: Pancreatic duct diameter measurement was performed manually on CT images during the portal venous phase, and the maximum value was taken; S23: Body composition analysis measures VFA and TAMA in a specific anatomical plane and normalizes them to an index.

[0017] The feature extraction step is performed independently by two doctors, and any discrepancies are resolved through consensus. All feature data is stored in a database for the model to use, ensuring process traceability and result consistency.

[0018] The beneficial effects of this invention are as follows: 1. This invention significantly improves the accuracy of preoperative prediction of CR-POPF by integrating multiple factors, which is superior to single-index models; it also reduces subjective bias based on objective data and improves clinical usability by using an online calculator. Attached Figure Description

[0019] Figure 1 This is a flowchart of the assessment method for the pancreatic fistula risk assessment system based on multifactor analysis proposed in this invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0021] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection or setting, a detachable connection or setting, or an integral connection or setting. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0022] Example 1: Data Preparation and Acquisition During implementation, data from 139 patients who underwent PD surgery were retrospectively collected from the hospital information system. Inclusion criteria included: complete enhanced CT images within one month preoperatively, complete perioperative clinical data (such as BMI and laboratory indicators), and a clear diagnosis of CR-POPF postoperatively (according to the ISGPS criteria). CT scans were performed using a Toshiba Aquilion One device with parameters set to 120kV, 350mAs, slice thickness 5mm, contrast agent iohexol (300mg / mL), injection flow rate 3-3.5mL / s, and portal venous phase images acquired (delay 60-70 seconds). Clinical data, including age, gender, smoking history, and diabetes history, were extracted from electronic medical records. All data were anonymized to comply with ethical requirements (informed consent was waived with the approval of the ethics committee).

[0023] Feature extraction module Feature extraction was performed in 3D Slicer software (version 4.11.20210226), and the specific steps included: ROI segmentation: Two experienced physicians independently delineate the pancreatic body and tail plane, excluding tumor areas, ensuring ICC > 0.75. The outline is then manually drawn.

[0024] Radiomics feature calculation: 1130 features were extracted, standardized using Z-scores, and redundant features were removed using Spearman correlation analysis (correlation coefficient > 0.9). Finally, 11 features were selected (as shown in the table below), for example, the coefficient for "Wavelet-LLH_glszm_Small Region Emphasis" was 0.1364. Pancreatic duct diameter measurement: The maximum value was manually measured at the same level using SyngoVia software (version vb20), with the unit accurate to 0.1 mm.

[0025] Body composition analysis: SFA and VFA are measured at the umbilical plane, and TAMA is measured at the L3 plane. The software automatically calculates the HU value, and the boundaries are manually corrected. The VFA / TAMAI calculation formula is VFA / (height^2), where the height unit is converted to meters.

[0026] Model building module Model training uses R software (version 3.6.3), and the code is open source and available. The steps are as follows: Data preprocessing: The median of continuous variables is standardized, and categorical variables are coded as numerical values.

[0027] Lasso regression screening: 10-fold cross-validation was performed using the glmnet package, and the minimum error point was selected for the λ value, resulting in the Rad-score formula: Rad-score = -1.2034 + (-0.5046)feature1 + 0.1242feature2 + ... + 0.1628*feature11.

[0028] Logistic regression modeling: The glm function was used, with the dependent variable being CR-POPF (binary classification), and the independent variables including Rad-score, pancreatic duct diameter, and VFA / TAMAI. In the multivariate analysis, only variables with P < 0.05 were retained (pancreatic duct diameter OR = 0.8, VFA / TAMAI OR = 1.55).

[0029] Model validation: Write a Bootstrap resampling script to loop 1000 times and calculate the average AUC, sensitivity (0.802), specificity (0.719) and other indicators.

[0030] Verification module Performance evaluation is achieved using a custom R function: ROC analysis: Curves were plotted using the pROC package, and the DeLong test was used to compare the differences between the models (joint model vs. clinical model, P=0.24).

[0031] Calibration curve: Plot a scatter plot of the predicted probability and the actual probability, and use the fitted curve to evaluate the deviation.

[0032] DCA: Calculates the net return for threshold probabilities from 0.1 to 0.3. The joint model yields the highest return in the 10%-28% threshold range.

[0033] Output module The online calculator was developed using the Shiny framework and deployed on a cloud server. Front-end interface: Design input forms, including sliders for pancreatic duct diameter (range 0-10mm), numeric input for VFA / TAMAI (range 0-5), and text boxes for Rad-score.

[0034] Backend logic: Load the pre-trained model using an R script, calculate the risk probability in real time, and output the percentage and risk level (e.g., <10% low risk, >30% high risk).

[0035] Nodal chart generation: Based on the RMS package, the scores of each factor are summed and converted into probabilities.

[0036] 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 multivariate analysis-based system for assessing the risk of pancreatic fistula after pancreaticoduodenectomy, characterized in that, include: The data acquisition module is used to acquire the patient's preoperative CT images and clinical data; A feature extraction module is used to quantitatively extract radiomics features, pancreatic duct diameter, and body composition indicators from CT images; The model building module is used to integrate the features through multivariate logistic regression to build a prediction model. The prediction model uses Rad-score, pancreatic duct diameter and VFA / TAMAI as independent variables and outputs the probability of CR-POPF occurrence. The validation module is used to evaluate model performance through bootstrap resampling and ROC curves. An output module is used to generate risk assessment results and present them in the form of a web calculator.

2. The pancreatic fistula risk assessment system based on multifactor analysis after pancreaticoduodenectomy according to claim 1, characterized in that, The body composition index includes the ratio of visceral fat area to total abdominal muscle area index.

3. The pancreatic fistula risk assessment system based on multifactor analysis after pancreaticoduodenectomy according to claim 1, characterized in that, The data acquisition module specifically includes a CT scan equipment parameter setting unit and a clinical data entry unit, wherein: The CT scanning equipment parameter setting unit is used to configure the enhanced CT scanning parameters, specifically including voltage, current, slice thickness, contrast agent iohexol injection flow rate, and to ensure that the portal venous phase image acquisition is delayed by 60-70 seconds. The clinical data entry unit is used to collect perioperative data in a structured manner, specifically including patient age, BMI, smoking history, diabetes history, laboratory indicators, and postoperative CR-POPF diagnosis results.

4. The pancreatic fistula risk assessment system based on multifactor analysis after pancreaticoduodenectomy according to claim 1, characterized in that, The feature extraction module includes a radiomics feature extraction unit, a pancreatic duct diameter measurement unit, and a body composition analysis unit, wherein: The radiomics feature extraction unit uses 3D Slicer software to segment the ROI of the pancreatic body and tail, extracts shape features, first-order features and texture features, and filters features with ICC values ​​greater than 0.75 through intragroup correlation coefficient (ICC) analysis. Then, Lasso regression is used to select 11 feature combinations as Rad-score. The pancreatic duct diameter measurement unit manually outlines the pancreatic duct in SyngoVia software and measures the maximum diameter with an accuracy of 0.1 mm. The body composition analysis unit measures VFA at the umbilicus level with a HU threshold set from -150 to -50, and measures TAMA at the third lumbar vertebra level with a HU threshold set from -29 to +150, and calculates the VFA / TAMAI index.

5. The pancreatic fistula risk assessment system based on multifactor analysis after pancreaticoduodenectomy according to claim 1, characterized in that, The model building module employs a multivariate logistic regression algorithm, specifically including a feature weight calculation unit and a model training unit. The feature weight calculation unit determines the coefficients of each feature in the Rad-score using Lasso regression and 10-fold cross-validation. The model training unit uses Rad-score, pancreatic duct diameter, and VFA / TAMAI as independent variables and CR-POPF binary classification results as dependent variables to fit the logistic regression equation: Logit(P) = β0 + β1Rad-score + β2pancreatic duct diameter + β3*VFA / TAMAI, where β0 is the intercept, and β1, β2, and β3 are regression coefficients calculated using the maximum likelihood estimation method.

6. The pancreatic fistula risk assessment system based on multifactor analysis after pancreaticoduodenectomy according to claim 1, characterized in that, The validation module includes a performance evaluation unit and a clinical utility analysis unit; wherein: The performance evaluation unit calculates the area under the curve (AUC), sensitivity, and specificity using the receiver operating characteristic (ROC) curve. The AUC of the joint model is 0.839, and the DeLong test is used to compare the differences between different models. The clinical utility analysis unit plots calibration curves to assess the degree of prediction probability calibration, and the decision curve analysis (DCA) calculates the net benefit from a risk threshold of 10% to 28%.

7. The pancreatic fistula risk assessment system based on multifactor analysis after pancreaticoduodenectomy according to claim 1, characterized in that, The output module includes a visualization unit and a network interface unit; The visualization unit generates a column chart; The network interface unit is based on the Shiny framework and is used to develop an online calculator.

8. A method for assessing the risk of pancreatic fistula after pancreaticoduodenectomy, which is the assessment method of the pancreatic fistula risk assessment system after pancreaticoduodenectomy as described in any one of claims 1-7, comprising the following steps: S1: Data Acquisition, collecting patient data through CT scans and clinical records; S2: Feature extraction, quantitatively extracting Rad-score, pancreatic duct diameter and VFA / TAMAI from CT images; S3: Model calculation, which inputs the extracted features into the logistic regression model to obtain the risk probability; S4: Validation, verifying model performance through bootstrap resampling; S5: Output, presenting the results as a nomogram or network calculator.

9. A method for assessing the risk of pancreatic fistula after pancreaticoduodenectomy according to claim 1, characterized in that, Step S2 includes the following steps: S21: The ROI of the pancreatic body and tail was delineated using 3D Slicer software. Radiomic features were extracted, and then ICC analysis and correlation screening were performed to screen features. Finally, Lasso regression was used to screen features again. S22: Pancreatic duct diameter measurement was performed manually on CT images during the portal venous phase, and the maximum value was taken; S23: Body composition analysis measures VFA and TAMA in a specific anatomical plane and normalizes them to an index.

10. A clinical application device for a pancreatic fistula risk assessment system, characterized in that, The pancreatic fistula risk assessment system after pancreaticoduodenectomy as described in claim 1 includes hardware and software components; wherein: The hardware components include a CT scanner, a server, and a user terminal. The CT scanner is used to acquire images, the server is used to run models and store data, and the user terminal is used to input and display results. The software component includes a data preprocessing algorithm, a model inference engine, and a user interface. The model inference engine loads a pre-trained logistic regression model and processes input features in real time. The clinical application device is used in preoperative outpatient clinics or surgical planning. After the doctor inputs patient data, the device automatically outputs risk suggestions, such as recommending nutritional intervention or surgery by an experienced physician for high-risk patients. The device connects to the hospital information system via a network to achieve data synchronization and remote access, thereby improving clinical decision support capabilities.

Images