Machine learning method-based acute necrotizing pancreatitis severe prediction model construction method

Abstract

The invention discloses an acute necrotizing pancreatitis severe prediction model construction method based on a machine learning method. The acute necrotizing pancreatitis severe prediction model construction method comprises the following steps: S1, collecting patient data based on an inclusion standard and an exclusion standard; s2, data preprocessing and feature selection; s3, obtaining an optimal differential diagnosis prediction model in combination with radiomics characteristics and a plurality of machine learning algorithms; s4, obtaining an evaluation index through prediction performance comparison; and S5, constructing a pancreas model, a pancreas surrounding model and a combination model, and producing an ROC curve interpretation result. The present invention develops and verifies a machine learning model for distinguishing between severe and moderate ANPs (i.e., ANSP and ANMSP). These models will be based on radiomics features extracted from pancreatic parenchyma portal vein phase CECT images, peripancreatic necrotic lesions, and combinations thereof. By evaluating the diagnostic performance of the models, early recognition of disease severity is realized, and support is provided for clinical decisions related to treatment.

Description

Technical Field

[0001] This invention relates to the field of disease diagnosis technology, specifically to a method for constructing a predictive model for severe acute necrotizing pancreatitis based on machine learning. Background Technology

[0002] Acute pancreatitis (AP) is a common acute abdominal condition characterized by local and systemic inflammatory responses. Its clinical course ranges from self-limiting mild to moderate-to-severe AP. Globally, the annual incidence of AP is approximately 33.74 per 100,000 person-years, with a mortality rate of approximately 1.16 per 100,000 people. Furthermore, its incidence is showing an increasing trend. The revised Atlanta classification in 2012 classifies AP into two types based on morphology and pathology: interstitial edematous pancreatitis and necrotizing pancreatitis (ANP). ANP (morphological type) is the more severe type, often involving multiple organ systems and accompanied by more severe clinical manifestations, leading to higher mortality and poorer prognosis. Approximately 20% of AP patients progress to moderate-to-severe acute pancreatitis (MSAP) or severe acute pancreatitis (SAP). SAP is the most critical type, characterized by high mortality (20%-40%) and poor prognosis. Although the terms ANP (morphological) and SAP (clinical) are often used interchangeably in the literature, their definitions and clinical manifestations are not entirely the same.

[0003] To establish a clearer distinction based on disease severity within the ANP population, this invention introduces the new terms "acute necrotizing moderate-to-severe pancreatitis (ANMSP)" and "acute necrotizing severe pancreatitis (ANSP)." These terms integrate imaging and clinical features, providing a more precise framework for classification and management.

[0004] Contrast-enhanced computed tomography (CECT) is the primary imaging modality for assessing the morphological features of necrotizing pancreatitis (AP). CECT is more widely used than magnetic resonance imaging (MRI) in diagnosing AP and assessing its severity. This may be attributed to the greater availability of CT equipment, faster scan times, and the ease of interpretation of its results by clinicians. CECT clearly displays necrotic areas (peripancreatic necrosis only, pancreatic necrosis only, or both) by highlighting the difference between parenchymal enhancement and peripancreatic vascularization, thus comprehensively assessing the severity of ANP and the extent of surrounding tissue involvement. Furthermore, the concept of radiomics, first proposed by Lambin et al. in 2012, refers to the high-throughput extraction and analysis of a large number of advanced quantitative imaging features from medical images. In the early stages of AP, morphological changes in the pancreas may not be obvious on imaging in some patients (especially those with pancreatic necrosis), leading to an underestimation of disease severity. As a non-invasive method, radiomics can capture subtle heterogeneity in lesions that are undetectable by conventional early imaging examinations. By quantitatively analyzing these characteristics, it establishes a crucial link between imaging findings and clinical practice, thereby aiding in the selection of treatment options.

[0005] In recent years, radiomics has been mainly applied to the diagnosis of pancreatic tumors and the differentiation of different types of pancreatitis. However, few studies have focused on the assessment of ANP severity, and no studies have explored early differential diagnosis by combining the radiomic characteristics of pancreatic parenchyma and peripancreatic necrosis foci. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing a predictive model for severe acute necrotizing pancreatitis (ANP) based on machine learning. This invention develops and validates machine learning models for distinguishing between severe and moderate ANP (i.e., ANSP and AMNSP). These models are based on radiomics features extracted from portal venous phase CECT images of the pancreatic parenchyma, peripancreatic necrotic lesions, and combinations thereof. By evaluating the diagnostic performance of these models, early identification of disease severity can be achieved, supporting treatment-related clinical decisions.

[0007] The objective of this invention is achieved as follows:

[0008] A method for constructing a predictive model for severe acute necrotizing pancreatitis based on machine learning includes the following steps:

[0009] Step S1: Collect patient data based on inclusion and exclusion criteria for future use;

[0010] Step S2: Preprocess and feature select the patient data from Step 1 to obtain preprocessed patient data for later use;

[0011] Specifically, the following steps were performed: Feature extraction and analysis of portal venous phase CECT images were conducted using 3D-Slicer software (https: / / www.slicer.org / ); the Region of Interest (ROI) was manually delineated, comprising two independent components: the first part being the entire pancreatic parenchyma, including necrotic areas but excluding bile ducts and vessels; the second part being the peripancreatic necrotic accumulation at the corresponding anatomical level; radiomics features of the ROI were automatically extracted using the PyRadiomics Python package, and discretized into 10 intervals using equal-width and equal-frequency binning methods; based on the LightGBM algorithm, the discretized features were randomly rearranged, and the optimal feature subset was iteratively derived using 10-fold cross-validation; the optimal feature subset is the preprocessed patient data; a classification model was constructed using the optimal radiomics features and the 10-fold cross-validation framework.

[0012] Step S3: Combine the radiomics features of the preprocessed patient data from Step 2 with various machine learning algorithms to obtain the best differential diagnosis prediction model;

[0013] Step S4: Obtain evaluation indicators by comparing the predictive performance of the differential diagnosis prediction model in Step 3;

[0014] Step S5: A pancreatic model, a peripancreatic model, and a combined model were constructed, and ROC curve interpretation results were generated.

[0015] The specific operation of step S1 is as follows: The inclusion criteria include the following conditions: (1) hospitalization within 7 days of the onset of ANP and undergoing CECT examination; (2) complete laboratory data, medical records and imaging examination data; (3) age ≥18 years; The exclusion criteria include the following conditions: (1) hospitalization after 7 days of the onset of ANP; (2) hospitalization within 7 days of the onset of ANP but without undergoing CECT examination; (3) history of chronic pancreatitis, pancreatic malignancy or pancreatic surgery; (4) pregnancy or age <18 years; (5) incomplete imaging examination and lack of portal venous phase axial images, poor image quality leading to inability to assess, missing key clinical data, hospitalization due to other acute abdominal diseases (such as gastrointestinal bleeding, intestinal obstruction); The severe group is characterized by persistent organ dysfunction >48 hours, modified Marshall score ≥2 points, and meeting any two of the following criteria: Ranson score ≥3 points, Glasgow-Imrie score >3 points or BISAP A score of ≥3 points is required; the moderate to severe group is defined as transient organ dysfunction lasting <48 hours and resolving within 48 hours, and does not meet the criteria for severe illness.

[0016] The specific operation of step S3 is as follows: By applying the forest learning algorithm to analyze the portal venous phase CECT data, three prediction models are finally established. The three prediction models include the pancreas model, the peripancreatic model, and the combined model. The model performance evaluation indicators include area under the curve, accuracy, sensitivity, specificity, F1 score, and Brier score. Among them, the Brier score quantifies the model performance by the mean square error between the predicted probability and the actual result. The lower the value, the better the model performance.

[0017] The specific operations of step S4 are as follows: 1520, 1520, and 3040 radiomic features were extracted from the pancreatic parenchyma, peripancreatic necrotic lesions, and combined regions, respectively; after consistency testing, the 1089 features in the pancreatic group, the 1101 features in the peripancreatic group, and the 2190 features in the combined group showed good consistency; 620, 515, and 1135 features in each model showed significant differences among different severity groups; finally, the optimal feature subset determined by LightGBM and cross-validation included 10 pancreatic features, 9 peripancreatic features, and 14 combined features.

[0018] The specific operation of step S5 is as follows: Based on radiomics, a model is developed, and five machine learning algorithms and the optimal feature subset extracted from the portal venous phase CECT data are used to distinguish the severity of the disease.

[0019] The beneficial effects of this invention are as follows: 1. The radiomics-based acute necrotizing pancreatitis severity prediction model constructed in this invention is used to distinguish between severe and moderate-severe acute necrotizing pancreatitis, thereby achieving more refined severity grading. The combined differential diagnostic prediction model, by integrating the features of pancreatic parenchyma and peripancreatic necrosis foci in portal venous phase enhanced CT, exhibits superior performance. The model's AUC values ​​on the training and test sets are 0.973 and 0.896, respectively, with accuracies of 89.8% and 83.9%, respectively. These technical results demonstrate that radiomics can capture early lesion heterogeneity in acute necrotizing pancreatitis, thereby enabling early identification of severe cases and guiding individualized treatment strategies.

[0020] 2. This invention provides a refined grading system for the severity of acute necrotizing pancreatitis, enabling radiologists and clinicians to more accurately classify patients with radiographically confirmed necrotizing pancreatitis as severe or moderately severe. This predictive model aids clinical decision-making in developing treatment strategies and supports early triage and stratification of such patients, thereby optimizing resource allocation. Attached Figure Description

[0021] Figure 1 This is a flowchart of the patient screening process for the present invention;

[0022] Figure 2This invention presents a schematic diagram of necrosis in different locations of acute necrotizing pancreatitis (mixed necrosis) and delineated representative structures; (a) Schematic diagram of mixed pancreatic necrosis: necrotic foci (N) in the pancreatic body and tail accompanied by peripancreatic fat fragments (FD); (b) Portal vein CECT of a 55-year-old female patient with mixed necrosis showing: parenchymal necrosis (black asterisk), peripancreatic necrosis foci containing fat fragments (white arrow), and residual normal pancreatic tissue (white asterisk); (c) Corresponding region of interest (ROI): The green area manually delineates the entire pancreatic parenchyma (including necrotic tissue, but excluding ducts and blood vessels); the yellow area represents peripancreatic necrosis foci at the same anatomical level as the pancreas;

[0023] Figure 3 This invention presents a schematic diagram of necrosis in acute necrotizing pancreatitis (peripancreatic necrosis only) and delineated representative structures; (a) Schematic diagram of peripancreatic necrosis only: fat fragments (FD) in the peripancreatic adipose tissue, without pancreatic parenchymal necrosis; (b) Portal venous phase CECT of a 43-year-old male patient with peripancreatic necrosis only, showing peripancreatic necrotic effusion (white arrow) containing scattered fat fragments (black arrow). Normal pancreatic parenchyma is marked with a white asterisk; (c) Corresponding manually delineated regions of interest (ROIs): pancreatic parenchyma (green) and peripancreatic necrotic effusion (yellow);

[0024] Figure 4 The ROC curve of the radiomics-based pancreatic parenchyma identification model of this invention is shown in Figure 4a, constructed using five machine learning algorithms (SVM, RF, KNN, GBDT, and XGBoost); Figure 4b shows the training cohort and the test cohort.

[0025] Figure 5 The ROC curves for the radiomics-based identification model for peripancreatic necrotizing effusion of this invention are shown below. The model was constructed using five machine learning algorithms (SVM, RF, KNN, GBDT, and XGBoost). 5a shows the training cohort plot, and 5b shows the testing cohort plot.

[0026] Figure 6 The diagram shows the ROC curve of the radiomics-based differential diagnosis model of this invention, and the diagnostic efficacy of five machine learning algorithms (SVM, RF, KNN, GBDT, and XGBoost) for pancreatic parenchyma with peripancreatic necrotizing effusion; 6a is the training cohort diagram; 6b is the test cohort diagram. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] A method for constructing a predictive model for severe acute necrotizing pancreatitis based on machine learning includes the following steps:

[0029] Step S1: Collect patient data based on inclusion and exclusion criteria for future use;

[0030] Specifically as follows, such as Figure 1 As shown, the trial has been approved by the Ethics Committee of our hospital (Approval No.: cqbykyll-20240918-108). A retrospective analysis was conducted on the medical records of ANP patients treated between May 2016 and June 2024. The diagnosis of acute pancreatitis (AP) is based on the 2012 revised Atlanta classification criteria, which requires at least one of the following three criteria: (1) persistent upper abdominal pain; (2) serum amylase and / or lipase levels at least three times the upper limit of normal; or (3) imaging findings consistent with AP characteristics. ANP is defined as AP with pancreatic parenchymal necrosis and / or peripancreatic necrosis. Pancreatic parenchymal necrosis refers to the presence of non-enhancing or low-enhancing areas (<30 HU) within the pancreas, while peripancreatic necrosis refers to the accumulation of fluid containing fatty necrotic fragments.

[0031] Inclusion criteria were as follows: (1) hospitalization within 7 days of ANP onset and CECT examination; (2) complete laboratory data, medical records and imaging examination data; (3) age ≥ 18 years. Exclusion criteria were as follows: (1) hospitalization after 7 days of ANP onset; (2) hospitalization within 7 days of ANP onset but without CECT examination; (3) history of chronic pancreatitis, pancreatic malignancy or pancreatic surgery; (4) pregnancy or age < 18 years; (5) incomplete imaging examination (no portal venous phase axial images), poor image quality leading to inability to assess, missing key clinical data, or hospitalization mainly due to other acute abdominal diseases (such as gastrointestinal bleeding, intestinal obstruction). In addition, the severe group was characterized by persistent organ dysfunction (>48 hours), modified Marshall score ≥ 2, and meeting any two of the following criteria: Ranson score ≥ 3, Glasgow-Imrie score > 3, or BISAP score ≥ 3. The moderate-to-severe group was defined as transient organ dysfunction (<48 hours) that resolved within 48 hours and did not meet the criteria for severe illness. A total of 184 ANP patients were included (see patient screening flowchart). Figure 1 Of these, 72 were in the severe case group (ANSP) and 112 were in the moderate-severe case group (ANMSP). Subsequently, these patients were randomly assigned in a 7:3 ratio to the training cohort (n=128; 52 severe cases and 76 moderate cases) and the testing cohort (n=56; 20 severe cases and 36 moderate cases).

[0032] After screening for contraindications to CECT, all patients underwent examination within 7 days of symptom onset. The revised Atlanta classification describes the dynamic course of acute pancreatitis (AP) as having two mortality peaks (early and late stages). This study focused on analyzing portal venous phase CT images acquired within 7 days of ANP onset. Patient positioning: head-first, supine, with arms elevated overhead. Scan range: from the diaphragmatic fornix to below the lower poles of both kidneys. CT parameters are detailed in Table 1. A non-contrast scan was performed first, followed by intravenous bolus injection of iohexol (1-2 mL / kg) via the antecubital vein at a flow rate of 3.5-4 mL / s using a high-pressure injector. Arterial and portal venous phase scans were acquired 25-30 seconds and 65-75 seconds after injection, respectively. All images were reconstructed and uploaded to a dedicated workstation for analysis, and then stored in the Picture Archiving and Communication System (PACS). Two abdominal radiologists with more than 4 years of experience independently reviewed all studies for diagnostic interpretation.

[0033]

[0034] Step S2: Preprocess and feature select the patient data from Step 1 to obtain preprocessed patient data for later use;

[0035] Specifically, the following steps were performed: Feature extraction and analysis of portal venous phase CECT images were conducted using 3D-Slicer software (https: / / www.slicer.org / ); the region of interest (ROI) was manually delineated, comprising two independent components, such as... Figure 2 and Figure 3 As shown: The first part is the entire pancreatic parenchyma, including necrotic areas but excluding bile ducts and blood vessels; the second part is the peripancreatic necrotic accumulation at the corresponding anatomical level. These accumulations are distributed in spaces such as the prerenal space, perirenal space, postrenal space, small cysts, perihepatic space, gastrosplenic space, and pancreatosplenic space. All images were first resampled to an isotropic 5 mm³ voxel size using bilinear interpolation.

[0036] The PyRadiomics Python package was used to automatically extract radiomics features from volumes of interest. The extracted feature set included first-order gray-level histogram features, second-order and higher-order texture features, and other features. Inter-group and intra-group correlation coefficients (ICCs) were calculated to assess radiologist concordance; an ICC value > 0.75 indicated good concordance. Independent samples t-tests were then performed on features with good concordance (ICC > 0.75) to identify features with statistically significant differences (P < 0.05). Radiomics features were discretized into 10 intervals using equal-width and equal-frequency binning methods. This binning process transforms features to an appropriate scale without normalization or standardization, while simplifying the logistic regression model and reducing the risk of overfitting. Based on the LightGBM algorithm, the discretized features were randomly rearranged, and the optimal feature subset was iteratively derived using 10-fold cross-validation. The optimal feature subset is the preprocessed patient data.

[0037] Step S3: Combine the radiomics features of the preprocessed patient data from Step 2 with the forest algorithm to obtain the optimal differential diagnosis prediction model;

[0038] The specific operations of step S3 are as follows: A classification model is constructed using optimal radiomics features and a 10-fold cross-validation framework. The portal venous phase CECT data is analyzed using the forest algorithm, ultimately establishing three prediction models: a pancreas model, a peripancreatic model, and a combined model. Simultaneously, other machine learning algorithms such as SVM, KNN, GBDT, and XGBoost are used for comparative analysis. Model performance evaluation metrics include area under the curve, accuracy, sensitivity, specificity, F1 score, and Brier score. The Brier score quantifies model performance by the mean squared error between the predicted probability and the actual result; a lower value indicates better model performance.

[0039] The principles and model building processes of various machine learning algorithms are as follows:

[0040] 1. Support Vector Machine (SVM)

[0041] Mathematical principle: Finding the optimal hyperplane to maximize the classification margin.

[0042]

[0043]

[0044]

[0045]

[0046] Modeling steps:

[0047] ① Transform the feature space using the RBF kernel function:

[0048]

[0049]

[0050] ② Set the penalty parameters C=10 and γ=0.01

[0051] ③ Solve the dual problem to obtain support vectors

[0052] ④ Decision function:

[0053]

[0054]

[0055] 2. Random Forest (RF)

[0056] Mathematical principles:

[0057] Integrating multiple decision trees:

[0058]

[0059]

[0060] Modeling steps:

[0061] ① Bootstrap sampling creates 200 decision trees

[0062] ② Selecting the optimal feature and split point when splitting a node:

[0063] a. Minimize Gini impurity:

[0064]

[0065]

[0066] b. Maximizing information gain:

[0067]

[0068]

[0069] ③ Set the maximum depth to 8 and the minimum number of samples per leaf node to 5.

[0070] 3. K-Nearest Neighbors (KNN)

[0071] Mathematical principles:

[0072] Similarity classification based on distance metrics:

[0073]

[0074] symbol type meaning y^ scalar Class prediction value of the sample to be predicted <![CDATA[argmax c ]]> Operator Category that takes the maximum value c scalar Categorical variables <![CDATA[N k (x)]]> gather The set of k nearest neighbors of the sample to be predicted <![CDATA[I(y i=c )]]> Indicator Function Category judgment function <![CDATA[y i ]]> scalar The true label of training sample i

[0075] Modeling steps:

[0076] ① Calculate similarity using Manhattan distance:

[0077]

[0078]

[0079] ② Set k=5 nearest neighbors

[0080] ③Distance-weighted voting:

[0081]

[0082]

[0083] ④ Probability estimation:

[0084]

[0085]

[0086] 4. Gradient Boosting Decision Tree (GBDT)

[0087] Mathematical principles:

[0088] Iterative optimization and improvement model:

[0089]

[0090]

[0091] Modeling steps:

[0092] ① Initialize the model:

[0093]

[0094]

[0095] ②For m=1 to 200:

[0096] f. Calculate pseudo residuals:

[0097]

[0098]

[0099] g. Fit the new tree hm(x) to the residual.

[0100] h. Calculate the output value of the leaf node:

[0101] i.

[0102]

[0103] j. Update the model:

[0104]

[0105]

[0106] ③ Set the learning rate ν=0.1 and the tree depth=3.

[0107] 5. Extreme Gradient Boosting (XGBoost)

[0108] Mathematical principles:

[0109] Regularized gradient boosting framework:

[0110]

[0111]

[0112]

[0113] Modeling steps:

[0114] ① Construct a second-order approximation of the objective function:

[0115]

[0116]

[0117] ② Tree structure optimization:

[0118] a. Greedy algorithm for finding the optimal split point

[0119] b. Split gain calculation:

[0120]

[0121]

[0122] General Symbol Explanation

[0123] ③ Set parameters:

[0124] learning_rate=0.05

[0125] gamma=0.1

[0126] max_depth=6

[0127] reg_lambda=1.

[0128] Step S4: Obtain evaluation indicators by comparing the predictive performance of the differential diagnosis prediction model in Step 3;

[0129] The specific operations of step S4 are as follows: 1520, 1520, and 3040 radiomic features were extracted from the pancreatic parenchyma, peripancreatic necrotic lesions, and combined regions, respectively; after consistency testing, the 1089 features in the pancreatic group, the 1101 features in the peripancreatic group, and the 2190 features in the combined group showed good consistency; 620, 515, and 1135 features in each model showed significant differences among different severity groups; finally, the optimal feature subset determined by LightGBM and cross-validation included 10 pancreatic features, 9 peripancreatic features, and 14 combined features.

[0130] Step S5: A pancreatic model, a peripancreatic model, and a combined model were constructed, and ROC curve interpretation results were generated.

[0131] The specific operation of step S5 is as follows: Based on radiomics, a model is developed, and five machine learning algorithms and the optimal feature subset extracted from the portal venous phase CECT data are used to distinguish the severity of the disease.

[0132] Statistical analyses were performed using SPSS 25.0 and Python 3.11.9. Normally distributed continuous variables were expressed as mean ± standard deviation (x ± s), while non-normally distributed continuous variables were expressed as median and interquartile range [M(Q...]. l Q u Presented as n (%). Categorical variables are expressed as n (%). Comparisons between groups for continuous variables are performed using the independent samples t-test or the Mann-Whitney U test, while categorical variables are compared using the χ² test or Fisher's exact test. A p-value less than 0.05 is considered statistically significant.

[0133] Test results

[0134] In the training cohort, there were no significant differences between the severe and moderate-severe ANP groups in terms of sex, alcohol consumption history, white blood cell count, serum lipase level, or type of pancreatic necrosis (P>0.05). In contrast, patients in the severe group were significantly older and had higher serum amylase levels and MCTSI scores (P<0.05) (Table 2).

[0135]

[0136] In the test cohort, there were no significant differences between the severe and moderate-severe ANP groups in terms of age, sex, alcohol consumption history, serum lipase level, or pancreatic necrosis type (P>0.05). However, the white blood cell count, serum amylase level, and MCTSI score were significantly higher in the severe group than in the moderate group (P<0.05) (Table 3).

[0137]

[0138] Feature processing and selection:

[0139] 1520, 1520, and 3040 radiomic features were extracted from the pancreatic parenchyma, peripancreatic necrotic lesions, and combined regions, respectively. After consistency testing, 1089 features in the pancreatic group, 1101 features in the peripancreatic group, and 2190 features in the combined group showed good consistency (ICC > 0.75). Independent samples t-tests revealed significant differences among different severity groups for 620, 515, and 1135 features in each model (P < 0.05). The optimal feature subset determined by LightGBM and cross-validation included 10 pancreatic features, 9 peripancreatic features, and 14 combined features (Tables 4-6).

[0140]

[0141]

[0142]

[0143] Model building and validation

[0144] A model based on radiomics was developed, utilizing five machine learning algorithms (SVM, RF, KNN, GBDT, and XGBoost) and the optimal feature subset extracted from portal venous phase CECT data to distinguish disease severity. Three different models were constructed: a pancreatic model, a peripancreatic model, and a combined model. Their ROC curves are shown below. Figure 4-6 As shown ( Figure 4 The ROC curve of the radiomics-based pancreatic parenchyma identification model was constructed using five machine learning algorithms (SVM, RF, KNN, GBDT, and XGBoost). 4a: training cohort; 4b: test cohort. Figure 5 The ROC curves of the radiomics-based model for identifying peripancreatic necrotizing effusion were constructed using five machine learning algorithms (SVM, RF, KNN, GBDT, and XGBoost). 5a: training cohort; 5b: test cohort. Figure 6Based on the radiomics-based differential diagnostic model ROC curve, five machine learning algorithms (SVM, RF, KNN, GBDT, and XGBoost) were used to assess the diagnostic efficacy of pancreatic parenchyma combined with peripancreatic necrotizing effusion. (6a: training cohort; 6b: test cohort)

[0145] Performance metrics for each model (including area under the curve (AUC), accuracy, sensitivity, specificity, F1 score, and Brier score). When applying the random forest algorithm to the combined feature set, it outperforms all other models (SVM, Random Forest, KNN, GBDT, XGBoost) in distinguishing between severe and moderate ANPs. This combined random forest model performed excellently in the test cohort, achieving an AUC of 0.896 (95% CI: 0.778–0.977), accuracy of 0.839, sensitivity of 0.65, specificity of 0.944, F1 score of 0.743, and Brier score of 0.134.

[0146] This invention utilizes a radiomics-based machine learning model to differentiate between severe and moderate-severe acute necrotizing pancreatitis (ANP), enabling more refined severity grading. The combined radiofrequency (RF) model, by integrating features of pancreatic parenchyma and peripancreatic necrosis lesions in portal venous phase CECT, demonstrates superior performance. The model achieved AUC values ​​of 0.973 and 0.896 on the training and test sets, respectively, with accuracies of 89.8% and 83.9%. These results demonstrate that radiomics can capture early ANP lesion heterogeneity—information that is not identifiable by traditional imaging—thus enabling early identification of severe cases and guiding individualized treatment strategies. This refined grading of ANP severity, combined with the proposed new terminology system for acute necrotizing pancreatitis moderate-severe (ANMSP) and acute necrotizing severe (ANSP), allows radiologists and clinicians to more accurately classify radiographically confirmed necrotizing pancreatitis patients as severe or moderate-severe. This identification helps clinical decision-makers formulate treatment strategies and supports early triage and stratification of such patients, thereby optimizing resource allocation.

[0147] Furthermore, radiomics-based models showed a progressively increasing AUC value for the pancreas, peripancreatic region, and combined models. In the training cohort, the AUC values ​​were 0.962, 0.969, and 0.973, respectively; the corresponding values ​​in the test cohort were 0.840, 0.868, and 0.896. This performance pattern may be related to the prevalence of the three ANP subtypes. The 2012 revised Atlanta classification indicates that ANP most commonly presents as a mixed type, less frequently as a peripancreatic necrosis-only type, and very rarely as a pancreatic parenchymal necrosis-only type. These findings collectively suggest a positive correlation between the prevalence of ANP subtypes and the discriminative performance (AUC) of the corresponding radiomics models. Therefore, integrating radiomics features of the pancreas and peripancreatic regions, combined with machine learning, can provide a more comprehensive and accurate assessment of the disease extent of ANP.

[0148] Radiomics noninvasively extracts and analyzes a wealth of quantitative features from medical images to meet clinical diagnostic needs. This approach is strongly complementary to machine learning (a branch of artificial intelligence that uses algorithms to identify risk factors, discover patterns, and build predictive models from complex datasets). The application of machine learning is rapidly evolving. Radiomics has been successfully applied to differentiate pancreatic malignancies from inflammatory diseases and to predict the severity of acute pancreatitis (AP). However, its potential in ANP severity stratification remains largely unexplored. This invention proposes, for the first time, a machine learning-based radiomics model for early stratification of ANP patients into moderate and severe categories. This invention integrates a comprehensive set of peripancreatic necrosis features with pancreatic characteristics to construct a holistic model that exhibits superior discriminative performance.

[0149] In summary, by integrating CECT radiomics features of pancreatic parenchyma and peripancreatic necrotic tissue and employing multiple machine learning algorithms, severe and moderate ANP can be effectively distinguished. The combined random forest model demonstrates optimal diagnostic efficacy. This invention can improve the early diagnosis rate of severe ANP, providing a basis for clinical decision-making. Furthermore, subdividing ANP into severe and moderate categories helps optimize patient triage processes, improve resource allocation efficiency, and is expected to improve patient prognosis.

Claims

1. A method for constructing a predictive model for severe acute necrotizing pancreatitis based on machine learning, characterized in that, Includes the following steps: Step S1: Collect patient data based on inclusion and exclusion criteria for future use; Step S2: Preprocess and feature select the patient data from Step 1 to obtain preprocessed patient data for later use; Specifically, the following steps were performed: 3D-Slicer software was used to extract and analyze features from CECT images of the portal venous phase; the Region of Interest (ROI) was manually delineated, comprising two independent components: the first part being the entire pancreatic parenchyma, including necrotic areas but excluding bile ducts and vessels; the second part being the peripancreatic necrotic accumulation at the corresponding anatomical level; radiomics features were automatically extracted from the ROI using the PyRadiomics Python package, and discretized into 10 intervals using equal-width and equal-frequency binning methods; based on the LightGBM algorithm, the discretized features were randomly rearranged, and the optimal feature subset was iteratively derived using 10-fold cross-validation; the optimal feature subset is the preprocessed patient data; a classification model was constructed using the optimal radiomics features and the 10-fold cross-validation framework. Step S3: Combine the radiomics features of the preprocessed patient data in Step 2 with the forest learning algorithm to obtain a differential diagnosis prediction model; use the forest learning algorithm and 10-fold cross-validation to construct pancreatic, peripancreatic, and combined radiomics prediction models based on the best radiomics features selected from pancreatic parenchyma, peripancreatic necrosis, and the two regions. Step S4: Obtain evaluation indicators by comparing the predictive performance of the differential diagnosis prediction model in Step 3; Step S5: A pancreatic model, a peripancreatic model, and a combined model were constructed, and ROC curve interpretation results were generated.

2. The method for constructing a predictive model for severe acute necrotizing pancreatitis based on machine learning as described in claim 1, characterized in that, The specific operation of step S1 is as follows: The inclusion criteria include the following conditions: (1) hospitalization within 7 days of the onset of ANP and undergoing CECT examination; (2) complete laboratory data, medical records and imaging examination data; (3) age ≥18 years; The exclusion criteria include the following conditions: (1) hospitalization after 7 days of the onset of ANP; (2) hospitalization within 7 days of the onset of ANP but without undergoing CECT examination; (3) history of chronic pancreatitis, pancreatic malignancy or pancreatic surgery; (4) pregnancy or age <18 years; (5) incomplete imaging examination and lack of portal venous phase axial images, poor image quality leading to inability to assess, missing key clinical data, hospitalization due to other acute abdominal diseases; The severe group is characterized by persistent organ dysfunction >48 hours, modified Marshall score ≥2 points, and meeting any two of the following criteria: Ranson score ≥3 points, Glasgow-Imrie score >3 points or BISAP A score of ≥3 points is required; the moderate to severe group is defined as transient organ dysfunction lasting <48 hours and resolving within 48 hours, and does not meet the criteria for severe illness.

3. The method for constructing a predictive model for severe acute necrotizing pancreatitis based on machine learning as described in claim 1, characterized in that, The specific operation of step S3 is as follows: By applying the forest learning algorithm to analyze the portal venous phase CECT data, three prediction models are finally established, including the pancreas model, the peripancreatic model, and the combined model. Random Forest (RF): Mathematical Principles: Ensemble Multiple Decision Trees The modeling steps are as follows: ① Bootstrap sampling creates 200 decision trees; ② Selecting the optimal feature and split point when splitting a node: a. Minimize Gini impurity: In the formula, t represents the current node in the decision book, and p(k / t) represents a scalar; b. Maximizing information gain: In the formula, ΔI represents information gain, I(parent) represents the impurity index of the parent node, and ∑ child This represents summing over all child nodes, N. child N represents the number of samples in the child nodes. parent N represents the number of samples in the parent node. child / N parent I(child) represents the proportion of child node samples, and I(child) represents the impurity index of child nodes. ③ Set the maximum depth to 8 and the minimum number of samples per leaf node to 5.

4. The method for constructing a severe acute necrotizing pancreatitis prediction model based on machine learning as described in claim 1, characterized in that, The specific operations of step S4 are as follows: 1520, 1520, and 3040 radiomic features were extracted from the pancreatic parenchyma, peripancreatic necrotic lesions, and combined regions, respectively; after consistency testing, the 1089 features in the pancreatic group, the 1101 features in the peripancreatic group, and the 2190 features in the combined group showed good consistency; 620, 515, and 1135 features in each model showed significant differences among different severity groups; finally, the optimal feature subset determined by LightGBM and cross-validation included 10 pancreatic features, 9 peripancreatic features, and 14 combined features.

5. The method for constructing a predictive model for severe acute necrotizing pancreatitis based on machine learning as described in claim 1, characterized in that, The specific operation of step S5 is as follows: Based on radiomics, a model is developed, and five machine learning algorithms and the optimal feature subset extracted from the portal venous phase CECT data are used to distinguish the severity of the disease.

Images