A prediction system and method for predicting the risk of ICU transfer in patients with acute pancreatitis.

Abstract

This invention discloses a system and method for predicting the risk of ICU transfer in patients with acute pancreatitis, belonging to the field of medical data processing technology. The system includes a data acquisition and preprocessing module, a feature selection module, a machine learning prediction module, and a risk output module. Addressing the shortcomings of existing prediction tools in terms of early immediacy, accuracy, interpretability, and consideration of patient heterogeneity, this invention uses multi-algorithm recursive feature elimination and consistency analysis to select a core prediction feature set, and employs a prediction model trained with a multi-objective optimization framework to calculate the probability of ICU transfer risk. This invention can achieve early risk warning based on routine clinical indicators at patient admission, improving the accuracy and interpretability of predictions, and supporting stratified analysis of heterogeneous factors such as patient origin.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and more specifically, to a system and method based on machine learning for predicting the risk of transferring patients with acute pancreatitis to the intensive care unit. Background Technology

[0002] Acute pancreatitis (AP) is a common gastrointestinal emergency, and some patients rapidly progress to severe illness requiring transfer to the intensive care unit (ICU). Early and accurate identification of these high-risk patients is crucial for optimizing clinical triage, allocating medical resources appropriately, and improving patient outcomes.

[0003] Currently, the main tools used in clinical practice to assess the severity and prognosis of acute pancreatitis (AP) include the Ranson score, APACHE II score, BISAP score, and biomarkers such as C-reactive protein (CRP). However, these traditional methods have significant limitations: most scoring systems require the collection of multiple dynamic data points over 24 to 72 hours after admission, making it impossible to provide immediate assessment in the early stages of hospitalization; some indicators (such as CRP) exhibit time lag; and the complex parameter calculations also limit their rapid application in emergency settings.

[0004] In recent years, machine learning technology has shown potential in predicting medical risks. However, many existing studies are based on ICU databases rather than real-world cohorts of acute exacerbation (AP) patients, and the predicted outcomes are mostly in-hospital mortality or traditional severity grading, rather than the more clinically relevant "ICU transfer." Furthermore, existing models often lack sufficient interpretability, making it difficult for clinicians to understand the model's decision-making rationale, and they do not adequately consider the role of "patient origin" (such as direct admission or inter-hospital referral), a significant confounding factor that may reflect the stage and initial state of disease progression, in the predictive model.

[0005] Therefore, there is an urgent need to develop a predictive tool for ICU transfer risk of AP patients that is based on routine data at the time of admission, has accurate predictions, is highly interpretable, and can take into account patient heterogeneity. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of existing early prediction tools for the risk of ICU transfer of AP patients in terms of immediacy, accuracy, interpretability and consideration of patient heterogeneity.

[0007] To address the aforementioned technical problems, the first aspect of this invention provides a risk prediction system for ICU transfer of patients with acute pancreatitis, comprising: The data acquisition and preprocessing module is used to acquire the patient's clinical indicator data upon admission and to process the missing and outlier values ​​in the clinical indicator data. The feature selection module is used to perform multi-algorithm recursive feature elimination and consistency analysis on the preprocessed clinical indicator data to select the core predictive feature set for predicting ICU transfer risk. The machine learning prediction module is used to calculate the probability of ICU transfer risk for patients based on the core prediction feature set using a prediction model trained with a multi-objective optimization framework. The risk output module is used to output the probability of risk associated with the ICU transfer.

[0008] Preferably, the core predictive feature set includes activated partial thromboplastin time, blood urea nitrogen, aspartate aminotransferase, fibrinogen, international normalized ratio, neutrophil / lymphocyte ratio, patient origin, thrombin time, and hemoglobin. These nine features have been consistently selected in multiple machine learning algorithms and have clear clinical pathophysiological significance.

[0009] Preferably, the data acquisition and preprocessing module adopts a domain-specific strategy for handling outliers, setting different outlier judgment and processing rules for different types of clinical indicators.

[0010] As a preferred approach, the domain-specific outlier handling strategy sets processing thresholds and exemption conditions based on clinical knowledge for different types of outliers, such as inflammatory markers, coagulation markers, and liver and kidney function markers, thus avoiding the possibility of mistakenly deleting clinically significant extreme values ​​using a uniform statistical method.

[0011] Preferably, the prediction model used by the machine learning prediction module is a neural network model; the optimization objective function of the multi-objective optimization framework integrates the model's sensitivity, area under the curve, calibration degree, and prediction stability.

[0012] Preferably, the scoring function of the multi-objective optimization framework comprehensively considers the model's discriminative ability, calibration accuracy, sensitivity, and stability. The progressive penalty mechanism deducts points from candidate models that perform poorly in calibration, probability distribution, etc., thereby guiding the optimization process to select models that not only have strong predictive power but also high clinical reliability.

[0013] Preferably, the system also includes an interpretability analysis module configured to calculate and visualize the contribution of each feature in the core predictive feature set to the predicted risk probability of a single patient, based on the SHapley AdditiveexPlanations method.

[0014] Preferably, the system also includes a risk stratification analysis module, configured to divide patients into different subgroups based on their origin, and to analyze the nonlinear association differences between the same clinical indicator and ICU transfer risk in different subgroups based on a restricted cubic spline model.

[0015] A second aspect of this invention provides a method for predicting the risk of ICU transfer in patients with acute pancreatitis, the steps of which correspond to the functional modules of the aforementioned system. Specifically, it includes the following steps: S1. Obtain the patient's clinical indicator data upon admission, and preprocess the missing and outlier values ​​in the clinical indicator data; S2. Perform multi-algorithm recursive feature elimination and consistency analysis on the preprocessed clinical indicator data to select the core predictive feature set; S3. Input the core prediction feature set into the prediction model trained by the multi-objective optimization framework to obtain the patient's ICU transfer risk probability; S4. Output the probability of ICU transfer risk.

[0016] Preferably, after the step of outputting the risk probability, the method further includes: analyzing the contribution of each feature to the prediction result based on the SHapley AdditiveexPlanations method, and performing stratified risk trajectory analysis according to the patient's origin.

[0017] The beneficial effects of this invention are as follows: 1. Early detection and practicality: Only routine laboratory test results and basic information of patients upon admission are required, without waiting for dynamic observation, thus achieving true early risk warning.

[0018] 2. High accuracy and strong generalization ability: Through rigorous internal and external time validation, the model has been demonstrated to have excellent discriminative ability and good calibration, and can be reliably applied to new patient cohorts.

[0019] 3. Good interpretability: The model’s decision-making logic and the contribution of each feature are intuitively demonstrated through SHAP values, risk thresholds (e.g., NLR≈14-15) and stratified risk curves, which improves clinical acceptance.

[0020] 4. Capturing clinical heterogeneity: For the first time, “patient origin” is systematically incorporated into the model as a stratified variable, revealing a significantly different risk trajectory between referred patients and directly admitted patients, making predictions more personalized.

[0021] 5. End-to-end solution: From data processing, feature engineering, model building to final deployment as a lightweight network application, it provides a complete set of clinical decision support tools. Attached Figure Description

[0022] Figure 1Sensitivity evaluation of six missing data imputation methods. Bar charts compare the predictive performance of eight machine learning classifiers on the training set (A) and test set (B) after employing different imputation methods. Abbreviations: LR, Logistic Regression; DT, Decision Tree; RF, Random Forest; KNN, k-Nearest Neighbors; SVM, Support Vector Machine; NB, Naive Bayes; XGB, Extreme Gradient Boosting; GBT, Gradient Boosting; NNET, Neural Network; AUC, Area Under the Curve; NPV, Negative Predicted Value.

[0023] Figure 2 Distribution of clinical variables before and after domain-specific outlier handling. Box plots compare the distribution characteristics of key clinical variables in the original data (Before) and the outlier-handled data (After). Outliers were defined based on clinically significant, variable-specific thresholds (e.g., quantile-based methods or adaptive IQR methods). Analysis was performed on both the training set (A) and the test set (B). All outlier handling thresholds were determined solely based on the training set and uniformly applied to the test set to avoid data leakage. Abbreviations: SII, Systemic Immune Inflammatory Index; WBC, White Blood Cell Count; NLR, Neutrophil / Lymphocyte Ratio; PLR, Platelet / Lymphocyte Ratio; MLR, Monocyte / Lymphocyte Ratio; Coagulation-related indicators: INR, International Normalized Ratio; APTT, Activated Partial Thromboplastin Time; Liver function enzymes: ALT, Alanine Aminotransferase; AST, Aspartate Aminotransferase; Renal function indicators: BUN, Blood Urea Nitrogen.

[0024] Figure 3 : Frequency distribution of feature selection across various machine learning algorithms. This horizontal bar chart illustrates the consistency in feature selection across different clinical variables across algorithms of varying feature importance. Features are ranked by frequency of selection, where frequency represents the number of algorithms that identified that feature as important.

[0025] Figure 4 Feature importance ranking based on multi-algorithm consistency analysis. A corresponding feature importance score is generated for each algorithm to reflect the relative contribution of each feature to the model's predicted output, thereby enabling comparative analysis between different models.

[0026] Figure 5 Results of internal validation for eight machine learning (ML) models. This figure shows the ROC curves, sensitivity, and specificity of eight ML models during internal validation, including DT, NNET, NB, SVM, GBT, LR, RF, and XGB.

[0027] Figure 6This study examines the performance of machine learning models in predicting ICU transfers of patients with acute pancreatitis (AP) on both the training and test sets. It includes ROC curve analysis (A, B), decision curve analysis (DCA) (C, D), calibration curves for each model (E, F), and parallel line plots (G, H) of various evaluation metrics for nine machine learning models predicting ICU transfers on both the training and test sets. Abbreviations: LR (Logistic Regression); DT (Decision Tree); RF (Random Forest); SVM (Support Vector Machine); NB (Naive Bayes); XGB (Extreme Gradient Boosting); GBT (Gradient Boosting); NNET (Neural Network); NPV (Negative Predictive Value); PPV (Positive Predictive Value).

[0028] Figure 7 Explanation of the NNET model based on SHAP. (A) Feature contribution summary plot. Each point represents the SHAP value corresponding to a certain feature for a single patient; red and blue indicate higher and lower feature values, respectively. The higher the SHAP value, the greater the predicted probability of ICU transfer. (B–C) Waterfall plots, showing the cumulative contribution of each feature to the final prediction result in high-risk patients (B) and low-risk patients (C), respectively. (D) SHAP dependency plot, showing the influence relationship of a single feature on the model output, with each point representing a patient; the vertical axis is the SHAP value, and the horizontal axis is the actual value of the corresponding feature. Abbreviations: INR, International Normalized Ratio; APTT, Activated Partial Thromboplastin Time; AST, Aspartate Aminotransferase; BUN, Blood Urea Nitrogen; NLR, Neutrophil / Lymphocyte Ratio.

[0029] Figure 8 : Predictive characteristic correlation heatmaps stratified by patient origin. (A) Spearman correlation heatmap for patients with first-episode pancreatitis. (B) Spearman correlation heatmap for patients referred for pancreatitis. Abbreviations: INR, International Normalized Ratio; APTT, Activated Partial Thromboplastin Time; AST, Aspartate Aminotransferase; BUN, Blood Urea Nitrogen; NLR, Neutrophil / Lymphocyte Ratio.

[0030] Figure 9 Distribution of key predictive factors (SHAP) values ​​stratified by patient origin. The (A–H) scatter plot compares the distribution of SHAP values ​​for each key predictive factor in patients with acute pancreatitis (AP) between those admitted for the first time and those referred. Abbreviations: INR (International Normalized Ratio); APTT (Activated Partial Thromboplastin Time); AST (Aspartate Aminotransferase); BUN (Blood Urea Nitrogen); NLR (Neurocyte / Lymphocyte Ratio).

[0031] Figure 10Association between biomarker values ​​stratified by patient origin and ICU transfer probability. (A–H) Biomarkers analyzed included: neutrophil / lymphocyte ratio (NLR), international normalized ratio (INR), activated partial thromboplastin time (APTT), blood urea nitrogen (BUN), thrombin time, fibrinogen, hemoglobin, and aspartate aminotransferase (AST). Restricted cubic spline (RCS) curves were used to show the predicted probabilities after stratification by patient origin; shaded areas represent 95% confidence intervals. P-values ​​were based on the Mann–Whitney U test, comparing differences between different patient origin groups within clinically significant intervals; asterisks indicate statistical significance (*P<0.05, **P<0.01, ***P<0.001).

[0032] Figure 11 This is a network calculator for ICU transport risk assessment of acute pancreatitis (AP) patients based on the constructed model. Users can input the following variables within 24 hours of patient admission to obtain individualized ICU transport risk assessment results: activated partial thromboplastin time (APTT), international normalized ratio (INR), thrombin time, fibrinogen, blood urea nitrogen (BUN), aspartate aminotransferase (AST), neutrophil / lymphocyte ratio (NLR), hemoglobin, and patient origin. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0034] One embodiment of the present invention provides a risk prediction system for ICU transfer of patients with acute pancreatitis, which mainly includes: a data acquisition and preprocessing module, a feature screening module, a machine learning prediction module, an interpretability analysis module, a risk stratification analysis module, and a risk output module.

[0035] I. Data Acquisition and Preprocessing: This was a retrospective cohort study that included patients diagnosed with acute pancreatitis (AP) at the Affiliated Hospital of Zunyi Medical University between January 1, 2014 and December 31, 2022. Inclusion criteria: (1) age ≥18 years; conforming to the 2012 revised Atlanta classification and ICD code (K85 / 577); (2) complete baseline clinical data; (3) completion of routine laboratory tests within 24 hours of admission. Exclusion criteria: (1) not a first-time AP hospitalization; (2) missing laboratory data rate >20%; (3) lack of key hematological parameters.

[0036] Of the 3,996 patients initially screened, 3,629 were ultimately eligible for inclusion. Patients admitted between 2014 and 2020 (n = 2,809) were used for model development, and patients admitted between 2021 and 2022 (n = 820) served as an external validation cohort.

[0037] Predictor variables Baseline demographic, socioeconomic, and clinical characteristics were extracted from the electronic medical record system. Admission laboratory indicators included: complete blood count; coagulation-related indicators (International Normalized Ratio [INR], Activated Partial Thromboplastin Time [APTT], fibrinogen, thrombin time); liver function indicators; renal function markers; and nutritional indicators. Derived inflammatory indicators included the systemic immune inflammation index (SII), platelet / lymphocyte ratio (PLR), neutrophil / lymphocyte ratio (NLR), and monocyte / lymphocyte ratio (MLR).

[0038] Ending Definition The primary outcome was ICU transfer within 72 hours of admission. ICU transfer criteria were based on the 2021 Chinese Guidelines for the Diagnosis and Treatment of Acute Pancreatitis.

[0039] Data quality was assessed prior to analysis. Missing values ​​were imputed using multiple imputation. Outliers were identified using interquartile range and clinical plausibility assessment. Continuous variables were treated as numerical variables, and categorical variables were encoded using one-hot encoding. To prevent data leakage, all preprocessing steps for the training and validation sets were performed independently.

[0040] Missing data imputation methods Multiple interpolation is employed, and after comparing methods such as Predictive Mean Matching (PMM), the interpolation strategy that best preserves the data distribution characteristics is selected.

[0041] This study implemented six commonly used missing data imputation methods: Predictive Mean Matching (PMM), Classification and Regression Trees (CART), k-Nearest Neighbors (kNN), Mean / Mode imputation, Bayesian Ridge Regression (Bayes), and Random Forest (RF). Each imputation method was combined with eight machine learning algorithms (LR, DT, RF, KNN, NB, SVM, XGB, SGBT, and NNET), resulting in 54 imputation-model combinations. These combinations were used to systematically evaluate their impact on model performance (Tables 1, 2, and 3). Figure 1 ).

[0042] Table 1: Comparison of features before and after multiple imputation in the training set Characteristic Before multiple imputation After multiple imputation P value1 Age 47.00 (38.00-57.00) 47.00 (38.00-57.00) >0.99 WBC_count 11.66 (8.58-15.13) 11.66 (8.58-15.13) >0.99 NeutrophilCount 9.70 (6.79-13.13) 9.70 (6.79-13.13) >0.99 Lymphocyte Count 1.08 (0.78-1.43) 1.08 (0.78-1.43) 0.988 MonocytesCount 0.64 (0.46-0.86) 0.64 (0.46-0.86) 0.975 Hemoglobin 135.00 (119.00-150.00) 135.00 (119.00-150.00) >0.99 PlateletCount 176.00 (135.00-218.00) 176.00 (135.00-218.00) 0.99 INR 0.92 (0.84-1.02) 0.93 (0.84-1.02) 0.653 APTT 33.30 (28.45-38.55) 33.30 (28.50-38.50) 0.879 Fibrinogen 4.06 (3.11-5.49) 4.13 (3.13-5.54) 0.335 ThrombinTime 18.50 (16.60-20.60) 18.60 (16.60-20.60) 0.649 ALT 30.00 (17.00-70.00) 30.00 (17.00-70.00) 0.999 AST 34.00 (22.00-68.00) 34.00 (22.00-68.00) >0.99 TotalBilirubin 20.10 (14.40-30.60) 20.10 (14.40-30.60) >0.99 Prealbumin 155.00 (105.00-208.00) 155.00 (105.00-208.00) 0.997 BUN 4.38 (3.18-5.96) 4.39 (3.20-5.95) 0.963 Creatinine 69.00 (58.00-84.00) 69.00 (58.00-84.00) 0.99 NLR 9.26 (5.47-14.52) 9.26 (5.47-14.52) >0.99 PLR 160.14 (113.45-227.03) 160.14 (113.45-227.03) >0.99 SII 1502.00 (869.27-2560.73) 1502.00 (869.27-2560.73) >0.99 MLR 0.60 (0.41-0.87) 0.60 (0.41-0.87) >0.99 Sex >0.99 Female 1198 (42.65%) 1198 (42.65%) Male 1611 (57.35%) 1611 (57.35%) OnsetToAdmission_hours >0.99 <=48 1985 (70.67%) 1985 (70.67%) >48 824 (29.33%) 824 (29.33%) PatientSource >0.99 First 1355 (48.24%) 1355 (48.24%) Transferred 1454 (51.76%) 1454 (51.76%) Education Level >0.99 College or above 958 (34.10%) 958 (34.10%) Junior / Senior High 1431 (50.94%) 1431 (50.94%) Primary or below 420 (14.95%) 420 (14.95%) Residence >0.99 Rural 1140 (40.58%) 1140 (40.58%) Urban 1669 (59.42%) 1669 (59.42%) Etiology >0.99 Alcoholic 228 (8.12%) 228 (8.12%) Biliary 1345 (47.88%) 1345 (47.88%) Hyperlipidemic 745 (26.52%) 745 (26.52%) Other 491 (17.48%) 491 (17.48%) RecurrenceStatus >0.99 Non_Recurrence 1958 (69.70%) 1958 (69.70%) Recurrence 851 (30.30%) 851 (30.30%) Table 2: Comparison of features before and after multiple interpolation in the test set

[0043] Handling outliers in clinical data Adopting a domain-specific strategy, differentiated rules are set for different indicator types: Inflammatory markers (such as NLR and SII) were measured using the quantile method (0.5%~99.5%), retaining extreme values ​​associated with severe illness.

[0044] Coagulation function indicators (INR, APTT, etc.) were measured using the dynamic IQR method to protect against combined abnormal conditions (e.g., INR > 2.0 and APTT > 50 s).

[0045] Liver function indicators (AST, ALT) were measured using the quantile method, retaining extreme values ​​that suggest acute liver injury (e.g., AST > 1000 U / L).

[0046] Hematological parameters (hemoglobin, platelets) were measured using the IQR method, with exemptions for values ​​with serious clinical risks (such as platelets <30 × 10⁻⁶). 9 / L).

[0047] Renal function indicators (BUN, creatinine) were measured using the intermediate IQR method.

[0048] Nutritional indicators (prealbumin) were measured using the intermediate IQR method.

[0049] All processing thresholds are determined based on the training set to avoid data leakage.

[0050] II. Feature Filtering First, the variance inflation factor (VIF) of all candidate features is calculated, and features with VIF > 10 are removed to alleviate multicollinearity. Then, recursive feature elimination is performed using RF, LR, XGB, LightGBM, GBT, and SVM algorithms respectively. Finally, the features that are stably selected across all algorithms are chosen to form the core prediction feature set. In this embodiment, this feature set is defined as: APTT, BUN, AST, fibrinogen, INR, NLR, patient origin (direct admission / referral), thrombin time, and hemoglobin (…). Figures 3-4 These nine features have clear clinical pathophysiological significance and constitute the final feature set for model development.

[0051] III. Machine Learning Prediction Module: Eight prediction models (LR, DT, RF, SVM, NB, XGBoost, GBT, and NNET) were constructed using the aforementioned core feature set. Hyperparameter tuning was performed using Bayesian optimization within a multi-objective framework, comprehensively considering sensitivity, AUC, calibration accuracy (1-Brier score), and predictive stability. A progressive penalty mechanism was introduced: when a model exhibits abnormalities in the Hosmer-Lemeshow test, a Brier score > 0.12, excessive deviation in calibration slope, or abnormal predicted probability distribution, its optimization score is progressively reduced, with a cumulative penalty cap of 50%. This mechanism ensures a balance between discriminative power, calibration accuracy, and clinical rationality in the selected models.

[0052] The optimization process is conducted within a multi-objective framework, with the basic scoring function being: Score base = 0.30×sensitivity + 0.30×AUC + 0.30×(1 - Brier score) + 0.10×stability.

[0053] The neural network (NNET) was ultimately selected as the prediction model because it demonstrated balanced performance in both internal and external validation.

[0054] IV. Model Validation and Results Internal validation (development cohort): NNET's AUC was 0.882 (95% CI 0.849–0.909), indicating good calibration (HL test, P = 0.344). External validation over time (2021–2022 cohort): AUC was 0.812, sensitivity 0.716, specificity 0.764, HL test, P = 0.126, indicating good generalization ability of the model. Decision curve analysis showed that NNET achieved a stable net benefit within the clinically relevant threshold range of 0.2–0.4.

[0055] V. Interpretability Analysis Module The SHAP method was used to interpret the NNET model. Global feature importance ranking showed that NLR, INR, patient origin, APTT, and BUN contributed the most. The SHAP dependency plot revealed approximate risk thresholds for key indicators: NLR approximately 14–15, INR approximately 1.02, APTT approximately 38–40 s, BUN approximately 5.8 mmol / L, and fibrinogen approximately 5.6 g / L. When multiple indicators simultaneously exceeded the threshold, the risk exhibited an additive effect.

[0056] VI. Risk Stratification Analysis Module Subgroup analyses were performed based on patient origin (direct admission vs. referral). Restricted cubic spline (RCS) models showed that, as Figure 2 As shown, the risk curves of referred patients were steeper in the ranges of NLR>14.5, INR>1.02, and APTT>38.5 s, suggesting that the two groups of patients have different pathophysiological evolution patterns. The comparison of SHAP values ​​also confirmed that the key characteristics of referred patients contributed significantly more (P<0.001).

[0057] Risk output: The trained final model (such as a neural network) and its accompanying interpretation and hierarchical logic are deployed as a web-based application (e.g., using the Streamlit framework). Clinicians input nine core feature values ​​of the patient into the interface, and the system returns the ICU transfer risk probability in real time (e.g., 87% high risk), while also providing a SHAP interpretation graph and risk stratification background information to assist clinical decision-making.

[0058] Identification of independent risk factors for ICU transfer Univariate analysis showed that multiple variables were significantly associated with ICU transport. In multivariate logistic regression analysis, systemic inflammation-related indicators (lymphocyte count, NLR), coagulation dysfunction indicators (INR, APTT, fibrinogen), and organ damage-related indicators (ALT, AST, BUN, creatinine) were still independently associated with ICU transport (Table 3).

[0059] Table 3: Results of univariate and multivariate analyses of ICU transfer.

[0060] Building and validating machine learning models To evaluate the applicability of different machine learning models in predicting ICU transfer risks for patients with acute pancreatitis, this embodiment systematically constructed and compared the performance of several commonly used classification algorithms. Internal validation results show that models employing nonlinear modeling strategies generally outperform linear models in overall discriminative performance. Specifically, the XGB model exhibits high discriminative ability, with an AUC of 0.897 (95% CI: 0.869–0.923), followed by the GBT model (AUC = 0.891, 95% CI: 0.859–0.918), while NNET's discriminative ability is similar to the aforementioned models (AUC = 0.882, 95% CI: 0.849–0.909). Furthermore, NNET maintains a high level of specificity (0.930–0.944) across different models; relevant results can be found in [link to relevant documentation]. Figure 5 .

[0061] During model training and screening, after comprehensively considering discriminative ability, calibration performance, and potential clinical application value, NNET exhibited a relatively balanced characteristic across various performance indicators. Although its AUC was slightly lower than some tree-based ensemble models, the consistency between its predicted probability and the actual outcome was good, and the Hosmer–Lemeshow calibration test showed no significant deviation. Relevant calibration curves are shown below. Figure 6 A and Figure 6 E. Under the default probability threshold of 0.5, NNET still maintains high specificity and overall accuracy, while maintaining an acceptable level of sensitivity. Figure 6 G).

[0062] Furthermore, threshold optimization analysis based on the Youden index showed that most nonlinear classification models could achieve a relative balance between sensitivity and specificity at a probability threshold of approximately 0.3. Within this threshold range, decision curve analysis showed that NNET could obtain stable net benefit within the clinically relevant threshold range of 0.2–0.4. Figure 6 (C) indicates that it is suitable for auxiliary decision-making scenarios involving ICU transfer risks.

[0063] Temporal external validation based on a patient cohort from 2021–2022 was used to evaluate the model's generalization performance. With an optimized probability threshold of 0.3, the NNET model maintained a relatively balanced performance between discriminative ability and calibration performance, with an AUC of 0.812, sensitivity of 0.716, specificity of 0.764, NPV of 0.893, and PPV of 0.497. Furthermore, the Hosmer–Lemeshow goodness-of-fit test did not show a significant deviation between the predicted results and the actual outcomes (P = 0.126) (see [link to relevant documentation]). Figure 6 B Figure 6 F and Figure 6 H).

[0064] Furthermore, pairwise comparisons based on the DeLong method showed no statistically significant difference in discriminative ability between the NNET model and the GBT and XGB models (all P>0.05); however, compared with various traditional classification algorithms, including DT, NB, LR, RF, and SVM, NNET showed a statistically significant advantage in discriminative performance. In addition, reclassification analysis results (NRI and IDI) indicated that NNET showed a significant improvement in risk stratification compared to relatively weaker models such as NB and DT, while no additional performance gain was observed compared to GBT and XGB.

[0065] DCA in the external validation cohort further showed that NNET consistently achieved stable net benefits within the clinically relevant threshold range of 0.2–0.4, outperforming other competing models and the two reference strategies, “Treat All” and “Treat None” (see Figure 6D).

[0066] In summary, although some tree-based ensemble models have certain advantages in discriminant metrics, NNET demonstrates relatively consistent discriminant ability, calibration performance, and decision-level utility in both internal and external validation. Therefore, in this embodiment, NNET is selected as the risk prediction model for subsequent interpretive analysis and clinical application deployment.

[0067] To interpret and analyze the ICU transfer risk prediction results, the SHAP method was used to quantify the contribution of each input feature in the NNET risk prediction model to the model output. For example... Figure 7 As shown in Figure A, the relative importance of different predictor variables in ICU transfer risk assessment can be obtained through the global analysis results of SHAP.

[0068] Among the nine selected predictor variables, NLR, INR, source of admission, APTT, BUN, fibrinogen, TT, hemoglobin, and AST had high mean absolute SHAP values, indicating that these indicators have significant contribution weights to the model output. A positive SHAP value indicates that the corresponding feature has a positive driving effect on the prediction results of ICU transfer risk, while a negative SHAP value indicates that its changing trend is related to a reduction in predicted risk.

[0069] At the individual level, analyzing the SHAP results of a single patient sample can illustrate the impact of different combinations of biomarkers on the prediction of ICU transfer risk. For example... Figure 7As shown in B, in the representative patient sample with high predictive outcomes, the elevation of APTT and INR had a major weight in the SHAP analysis, playing a major role in predicting increased risk; while in the patient sample with low predictive outcomes ( Figure 7 C) The SHAP values ​​corresponding to each prediction indicator are generally small. The slight increase in some indicators (such as NLR and TT) is offset by the negative contribution of other features in the model, keeping the prediction risk at a low level.

[0070] To analyze the relationship between predictor variables and their contributions, a SHAP dependency graph is plotted for each predictor feature, such as... Figure 7 As shown in D. The results showed that the SHAP values ​​of some biomarkers exhibited a non-linear trend across different value ranges. When the relevant indicators exceeded a certain value range, the corresponding SHAP values ​​increased more rapidly. The observed approximate ranges included: BUN approximately 5.8 mmol / L, NLR approximately 14–15, INR approximately 1.02, fibrinogen approximately 5.6 g / L, APTT approximately 38–40 s, and cases where the admission source was inter-hospital referral (see Table 4). When multiple indicators were simultaneously within the above value ranges, their corresponding SHAP values ​​could have a cumulative effect in the model, thus having a more significant impact on the prediction results of ICU transfer risk.

[0071] Table 4: Risk Stratification Characteristics of Patients with Acute Pancreatitis in the ICU Feature Risk Direction High Risk Threshold High-Risk Cases Count High Risk ICU Rate Low Risk Threshold Low Risk Cases Count Low Risk ICU Rate NLR Positive 14.52 704 0.4659 5.47 704 0.0497 BUN Positive 5.95 707 0.4484 3.2 714 0.1513 PatientSource Positive 1 1454 0.3343 0 1355 0.1387 Fibrinogen Positive 5.54 711 0.3671 3.13 706 0.1586 APTT Positive 38.5 707 0.4752 28.5 704 0.1449 INR Positive 1.02 745 0.5275 0.84 720 0.0611 AST Positive 68 703 0.3343 22 720 0.1194 ThrombinTime Negative 16.6 721 0.3065 20.6 721 0.2538 Hemoglobin Positive 150 740 0.3054 119 739 0.2842 Key features and their structural differences across different patient origins Spearman correlation analysis revealed significant differences in biomarker interaction patterns between first-admission patients (n = 1,355) and patients referred from other hospitals (n = 1,454). Among first-admission patients, patient origin was significantly different from blood urea nitrogen (BUN) (r = 0.45), International Normalized Ratio (INR) (r = 0.60) and neutrophil / lymphocyte ratio (NLR) (r = The correlation between these biomarkers and the value of 0.44 was negative, suggesting that the association between these biomarkers was relatively weak. Figure 8 A). In contrast, among referred patients, patient origin showed a strong positive correlation with fibrinogen (r = 0.50), INR (r = 0.53), and NLR (r = 0.44), reflecting significantly different correlation structures. Figure 8B). The above results indicate significant heterogeneity in biomarker interactions among patients from different admission sources, a finding highly consistent with the differences in feature contributions observed in subsequent SHAP analyses.

[0072] Analysis of the differences in contribution from different patient origins based on SHAP Comparing the SHAP values ​​of newly admitted patients and referred patients revealed a significant difference in characteristic contribution between the two groups. Figure 9 Compared with first-time admission patients, referred patients had significantly higher SHAP values ​​on several key biomarkers, including neutrophil / lymphocyte ratio (NLR), international normalized ratio (INR), and blood urea nitrogen (BUN) (all P < 0.001). These results indicate that these characteristics contribute more significantly to the prediction of ICU transfer risk in referred patients, reflecting a source-specific attribution pattern in the model across different patient populations.

[0073] Heterogeneity of risk inflection point patterns from different patient origins revealed by RCS analysis Restricted cubic spline (RCS) models showed a clear source-specific nonlinear association between biomarker levels and ICU transfer prediction risk in patients from different admission sources. Figure 10 Of all eight biomarkers, compared with patients admitted for the first time, referred patients showed a steeper and higher risk trajectory within the key value range.

[0074] Regarding inflammation and coagulation-related indicators, patients referred to ICUs with higher NLR levels (approximately 14.5–30.2, P<0.001), INR in the range of 1.0–1.2 (P<0.001), and APTT in the range of 38.5–50.9 s (P<0.001) had a significantly increased predicted risk of ICU transfer. Figure 10 A– Figure 10 C). BUN also showed a similar trend; within the elevated range of 6.0–10.4 mmol / L, the predicted risk for referred patients was significantly higher than that for first-time admission patients (P<0.001). Figure 10 D). Furthermore, patients referred to the hospital with thrombin time (14.7–16.6 s) and fibrinogen (5.5–8.1 g / L) in their high-range values ​​also showed significantly higher risk estimates (both P<0.05). Figure 10 E– Figure 10 F).

[0075] Hemoglobin and AST also showed source-dependent nonlinear differences. Lower hemoglobin levels were associated with higher predictive risk in referred patients (P<0.01), while AST, although showing smaller differences within the intermediate range, was still statistically significant (P<0.05) (Figure 10G– Figure 10 H).

[0076] Overall, RCS analysis revealed a clear nonlinear risk inflection point pattern across all biomarkers, indicating a consistent and significant divergence between first-admission and referred patients on the predicted ICU transfer risk trajectory.

[0077] VII. Development of Network Prediction Tools The final neural network model (NNET) was deployed using the Streamlit Python framework as the networked prediction tool RICU-APP (Rapid ICU Prediction for Pancreatitis). Figure 11 The application takes nine admission biomarkers used by the model as input and automatically generates an individualized prediction probability of a patient being transferred to the ICU.

[0078] RICU-APP can simultaneously output ICU transfer risk prediction results and SHAP-based feature contribution interpretations, and provide risk distribution maps to help clinicians interpret individual predicted probabilities within the context of a reference population. See the example output below. Figure 11 This tool can be accessed at the following URL: https: / / jsehvkzse7nbwqnjdurcdv.streamlit.app / .

[0079] This study constructed and validated a machine learning model for predicting early ICU transfer in patients with acute pancreatitis (AP) based on routine biomarkers obtained at admission. The model demonstrated stable discriminative ability, good calibration, and clear clinical utility in a large-scale real-world population, and its generalization ability was confirmed through out-of-time validation. Importantly, this study is the first to systematically evaluate the modifying role of admission origin in the risk evolution process. Through SHAP-based interpretability analysis, we identified several reproducible risk thresholds, including NLR ≥14.52, INR ≥1.02, APTT ≥38.5 s, and BUN ≥5.95 mmol / L, all of which are significantly associated with increased ICU transfer risk. These thresholds provide actionable references for early clinical assessment, highlighting the core roles of inflammatory response, coagulation imbalance, and metabolic disorders in early disease progression, and supporting the feasibility of risk stratification based on immediate admission indicators from both biological and clinical perspectives.

[0080] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A risk prediction system for ICU transfer of patients with acute pancreatitis, characterized in that, include: The data acquisition and preprocessing module is used to acquire the patient's clinical indicator data upon admission and to process the missing and outlier values ​​in the clinical indicator data. The feature selection module is used to perform multi-algorithm recursive feature elimination and consistency analysis on the preprocessed clinical indicator data to select the core predictive feature set for predicting ICU transfer risk. The machine learning prediction module is used to calculate the probability of ICU transfer risk for patients based on the core prediction feature set using a prediction model trained with a multi-objective optimization framework. The risk output module is used to output the probability of risk associated with the ICU transfer.

2. The system according to claim 1, characterized in that, The core predictive feature set includes: activated partial thromboplastin time, blood urea nitrogen, aspartate aminotransferase, fibrinogen, international normalized ratio, neutrophil / lymphocyte ratio, patient origin, thrombin time, and hemoglobin.

3. The system according to claim 1, characterized in that, The data acquisition and preprocessing module employs a domain-specific strategy for handling outliers, setting different outlier judgment and processing rules for different types of clinical indicators.

4. The system according to claim 3, characterized in that, The domain-specific strategies include: processing inflammation-related indicators using quantile methods and retaining extreme values ​​related to severe illness; processing coagulation function indicators using dynamic IQR methods and retaining combined abnormal states; and processing liver function indicators using quantile methods and retaining extreme values ​​indicating acute liver injury or severe jaundice.

5. The system according to claim 1, characterized in that, The machine learning prediction module uses a neural network model for prediction; the objective function of the multi-objective optimization framework integrates the model's sensitivity, area under the curve, calibration degree, and prediction stability.

6. The system according to claim 5, characterized in that, The multi-objective optimization framework includes a progressive penalty mechanism, which penalizes the optimization score of a candidate model when it fails to perform calibration, exhibits an abnormal probability distribution, or makes excessively extreme predictions during cross-validation.

7. The system according to claim 1, characterized in that, It also includes an interpretability analysis module, configured to calculate and visualize the contribution of each feature in the core predictive feature set to the predicted risk probability of a single patient based on the SHapley Additive exPlanations method.

8. The system according to claim 1, characterized in that, It also includes a risk stratification analysis module, configured to divide patients into different subgroups based on their origin, and analyze the nonlinear association differences between the same clinical indicator and ICU transfer risk in different subgroups based on a restricted cubic spline model.

9. A method for predicting the risk of ICU transfer in patients with acute pancreatitis, characterized in that, The system based on any one of claims 1 to 8 includes the following steps: S1. Obtain the patient's clinical indicator data upon admission, and preprocess the missing and outlier values ​​in the clinical indicator data; S2. Perform multi-algorithm recursive feature elimination and consistency analysis on the preprocessed clinical indicator data to select the core predictive feature set; S3. Input the core prediction feature set into the prediction model trained by the multi-objective optimization framework to obtain the patient's ICU transfer risk probability; S4. Output the probability of ICU transfer risk.

10. The method according to claim 9, characterized in that, Following the step of outputting risk probabilities, the method also includes: analyzing the contribution of each feature to the prediction results based on the SHapley Additive exPlanations method, and performing stratified risk trajectory analysis by patient origin.

Images