A Time-Series Feature-Based Nomogram for the Prediction of Severe Acute Pancreatitis

A Time-Series Feature-Based Nomogram for the Prediction of Severe Acute Pancreatitis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Time-Series Feature-Based Nomogram for the Prediction of Severe Acute Pancreatitis Yang Chen, Zhidong Fu, Yongzhi Liu, Feng Jiang, Tieming Zhu, Lihui Jiang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7029031/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 24 Apr, 2026 Read the published version in BMC Gastroenterology → Version 1 posted 11 You are reading this latest preprint version Abstract Background The annual incidence of acute pancreatitis is approximately 30 per 100,000, with 20% progressing to severe acute pancreatitis and a mortality rate of 20%-40%. Traditional scoring models suffer from data lag or insufficient accuracy, while existing machine learning models mostly overlook the dynamic characteristics of vital signs. Methods Vital signs, laboratory and imaging indices within 24 hours of admission were collected. First, a bidirectional long short-term memory network model was constructed using time-series data. Then,key indices from laboratory and imaging data were screened by LASSO. Eight machine learning models were constructed and compared. Finally, a predictive nomogram was developed based on the Random Forest model and SHAP values. Result After propensity score matching, among 193 patients, there were 124 cases in the MSAP group and 69 cases in the SAP group, with no significant differences in baseline characteristics between the two groups. The BiLSTM model showed an average AUC of 0.9551, accuracy of 0.9222, F1-score of 0.8956, training loss of 0.2992 ± 0.0328, and validation loss of 0.4132 ± 0.0651 in 10-fold cross-validation. Features including Rmax, Pdiff_mean, and Tdiff_std extracted from time-series data, together with those screened by LASSO (PE, Neu, HCT, Ca, TG, AMY, and CRP), were used to construct 8 ML models. The Random Forest model demonstrated the best comprehensive performance, with an accuracy of 0.8793, ROC-AUC of 0.9588. SHAP value analysis identified key features as Rmax, Pdiff_mean, HCT, Tdiff_std, PE, Neu, and serum calcium. The nomogram constructed based on these features achieved AUC values of 0.969 and 0.964 in the training and test sets, respectively. Decision curve analysis showed that the net benefit exceeded 0.2 at high-risk thresholds (0.2–0.8), outperforming both the "treat all" and "treat none" strategies. Conclusion The BiLSTM-RF model constructed in this study improves the accuracy of SAP prediction by extracting time-series features of vital signs. The nomogram built based on key features demonstrates good clinical practicability, providing a visual tool for the early assessment of SAP. Severe acute pancreatitis Bidirectional long short-term memory network Machine learning Random forest SHAP Nomogram Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Introduction Acute pancreatitis (AP) ranks among the most prevalent acute abdominal emergencies requiring hospitalization, with an annual incidence of approximately 30 cases per 100,000 population in high-income countries [ 1 ]. Recent epidemiological evidence highlights a rising trend in regional incidences, such as the average annual percentage change (APCC) of 3.67% in North America and 2.77% in Europe [ 2 ]. The socioeconomic burden of AP is substantial, with annual healthcare costs exceeding $ 2.6 billion in the United States. For patients with AP complicated by severe morbidities, the mean per-patient expenditure escalates to approximately $ 90,000 [ 3 ]. Approximately 20% of AP cases progress to moderate-severe or severe phenotypes, where local and systemic inflammatory cascades drive a mortality rate of 20%-40% [ 4 ]. The overall mortality across all AP severities remains at 2%, underscoring the critical need for early prediction of disease severity to inform timely interventional strategies [ 4 ]. The evolution of traditional AP predictive models began five decades ago with the seminal publication of the Ranson score by Ranson et al.[ 5 ], the first grading system for AP severity. Subsequent decades have seen the emergence of numerous assessment tools, including the Acute Physiology and Chronic Health Evaluation II (APACHE II) [ 6 ], Bedside Index of Severity in Acute Pancreatitis (BISAP) [ 7 ], and Balthazar CT scoring system [ 8 ]. However, these scoring criteria exhibit notable limitations in clinical timeliness and accuracy. For instance, while five parameters of the Ranson score are obtained at admission, six require 48-hour follow-up data, introducing a critical lag that compromises early disease progression assessment. The BISAP score, derived from 24-hour admission data (blood urea nitrogen > 8.9 mmol/L, altered mental status, systemic inflammatory response syndrome (SIRS), age > 60 years, and pleural effusion), prioritizes accessibility but lacks optimal diagnostic performance. A study by Anum et al. demonstrated that BISAP had lower accuracy (76.2% vs. 82.2%) and sensitivity (69.2% vs. 97.4%) than the Ranson score for predicting severe pancreatitis [ 9 ]. The APACHE II score, comprising acute physiology, chronic health, and age components, requires comprehensive 24-hour data collection, hindering its utility in early severe pancreatitis assessment. The Balthazar CT score, dependent on contrast-enhanced CT for pancreatic necrosis evaluation, achieves 90% early detection sensitivity (72–96 hours post-onset), which improves to 100% after four days, thus limiting its application in acute-phase severity stratification [ 10 ]. With the continuous refinement of statistical theories and the rapid advancement of computer technology, machine learning (ML), as a subset of artificial intelligence (AI), has been increasingly adopted in clinical practice [ 11 ]. Most established and reliable AI tools fall within the domain of non-generative AI, encompassing both supervised and unsupervised machine learning techniques [ 12 ]. Compared to generative AI, supervised machine learning models in medicine offer enhanced interpretability, as their outputs (e.g., class labels or continuous values) can be directly linked to clinical outcomes. Additionally, a robust framework of statistical metrics—including accuracy, precision, sensitivity, and specificity—has been developed to evaluate model performance [ 13 ].Therefore, the inclusion of comprehensive and systematic feature variables is pivotal for model construction. Previous studies on AP predictive models have rarely focused on the extraction of time-series features from vital signs (e.g., body temperature, pulse, respiratory rate, etc.). As a branch of Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) addresses the common vanishing/exploding gradient problems in RNNs by introducing gated memory units, making it highly suitable for sequence-related problems [ 14 ]. Graves et al. proposed the BiLSTM architecture, which can fully utilize contextual information and outperforms unidirectional LSTM [ 15 ]. Meanwhile, SHAP (SHapley Additive exPlanations) provides a new approach to solve the "black-box" effect of machine learning models. Therefore, this study aims to develop a multimodal model for predicting SAP based on time-series features and construct a nomogram according to SHAP values to provide evidence for clinical decision-making. Methods Patients A retrospective analysis was performed on clinical data of patients with AP treated at Xiaoshan Hospital Affiliated to Hangzhou Normal University between January 2018 and December 2024. Inclusion criteria were as follows: ① 85 years ≥ age ≥ 18 years; ② AP defined as serum amylase elevation to at least three times the upper limit of normal, accompanied by typical symptoms of acute pancreatitis. SAP was defined according to the revised Atlanta Consensus Document. Exclusion criteria included: ① history of severe cardiovascular, cerebrovascular, respiratory, renal, or hepatic diseases; ② previous or current history of malignant tumors; ③ Child-Pugh class ≥ C;④ Long-term oral administration of hormonal and psychotropic medications. This retrospective medical record study has obtained the approval of the Ethics Committee of Zhejiang Xiaoshan Hospital (Approval Number: Y2024088). All procedures involving human participants in this study were conducted in accordance with the ethical standards of the aforementioned institutional ethics committee and with the 1964 Declaration of Helsinki and its subsequent amendments. In line with the requirements of the Declaration of Helsinki, the ethics committee of Zhejiang Xiaoshan Hospital granted a waiver of informed consent for this study. The waiver was justified based on the following considerations: the study is of low-risk nature, involving only the analysis of de-identified historical medical records, which poses no threat to the rights or interests of patients and will not cause any harm to them. The research team will strictly ensure the confidentiality of all medical records and abide by all relevant ethical and legal requirements throughout the study. Data collection The present study collected general characteristic indices, time - series indices, laboratory indices, and imaging indices within 24 hours of admission. For the binary variable labels, SAP was labeled as 1, while MSAP was labeled as 0.Male was labeled as 1, while female was labeled as 0.Smoking: coded as 1, otherwise 0.Alcohol consumption: coded as 1, otherwise 0.Presence of pleural effusion: coded as 1, otherwise 0. General characteristic indices collected for each patient included demographic characteristics (gender and age), American Society of Anesthesiologists classification (ASA-class), Charlson Comorbidity Index (CCI), body mass index (BMI), and Nutritional Risk Score (NRS2002). Time-series indices comprised body temperature(T), pulse༈P༉, and respiratory rate༈R༉ .Upon admission, a set of data was recorded and labeled as T0, P0, and R0. Thereafter, data were recorded every 2 hours within the subsequent 24 hours, labeled as T1 to T13, P1 to P13, and R1 to R13. Laboratory indices included white blood cell count(WBC), C-reactive protein (CRP), neutrophil count(Neu), red blood cells(RBC), platelets(Plt), hematocrit(HCT), alanine transaminase (ALT), aspartate transaminase (AST), serum amylase(AMY), serum lipase(LIP), total bilirubin(TBil), serum calcium(Ca), blood glucose(Glu), albumin(Alb), and blood urea nitrogen (BUN). Imaging indices focused on chest CT or chest radiograph indicates the presence of pleural effusion(PE). Data processing To achieve a more balanced representation of the sample data characteristics, we incorporated demographic features (gender, age), American Society of Anesthesiologists classification (ASA-class), Charlson Comorbidity Index (CCI), Body Mass Index (BMI), and Nutritional Risk Score 2002 (NRS2002). Subsequently, propensity score matching(PSM) was employed to select the final cases based on the severity of acute pancreatitis (AP).In this study, missing values were managed using multiple imputation techniques[ 16 ]. Specifically, logistic regression imputation was employed for binary variables, whereas predictive mean matching imputation was utilized for continuous variables. Variables with univariate missing values exceeding 30% were directly excluded from the analysis. For variables requiring standardization, standard deviation standardization (i.e., z-score normalization) was applied to ensure consistency across measurements. Model Construction and Evaluation 1)Extraction of features from time-series models Features of the T/P/R time series data in the study sample were extracted by integrating the characteristics of T0-12, P0-12, and R0-12 datasets. The number of time steps was determined to be 13, as each time series comprised observations at 13 discrete time points (indexed from 0 to 12), which comprehensively captured the temporal dynamics of the variables.Feature engineering encompasses time series features (upward/downward streaks, volatility characteristics, and stability characteristics) and statistical features (mean, standard deviation, skewness, kurtosis, median, range, interquartile range(iqr), slope, and delta).A dual-branch hybrid architecture based on a Bidirectional Long Short-Term Memory (BiLSTM) neural network is utilized for processing time-series data features. In the first branch, temporal features are extracted using a BiLSTM layer with 16 hidden neurons. To enhance generalization and prevent overfitting, Dropout and Batch Normalization techniques are incorporated. The second branch consists of a fully connected layer with 16 neurons, which processes manually engineered features. This branch employs Rectified Linear Unit (ReLU) activation and L2 regularization to mitigate overfitting. Subsequently, a concatenation layer integrates the outputs from both branches. The fused data is then passed through another fully connected layer with 16 neurons, utilizing ReLU activation and L2 regularization. Finally, the output layer uses a sigmoid activation function, which is appropriate for binary classification tasks. The Adam optimizer is employed to adaptively adjust the learning rate during training. Binary cross-entropy is selected as the loss function to measure the discrepancy between predicted probabilities and true labels. Model performance is assessed using metrics such as the Area Under the Curve (AUC) and loss curves. Additionally, early stopping and learning rate decay strategies are implemented to improve model stability and convergence. Furthermore, SHAP (SHapley Additive exPlanations) values are calculated to provide interpretable feature importance rankings and visualizations, enabling the identification of the most critical features derived from T/P/R metrics. 2)Optimal Model 2.1 Feature Extraction We implemented LASSO (Least Absolute Shrinkage and Selection Operator) for feature screening of hematological indices and imaging markers. LASSO extends linear regression by incorporating an L1 regularization term, which zeros out the weights of irrelevant features to achieve feature selection. Simultaneously, by constraining the magnitude of weights, it enhances the model’s generalization capability on unseen data and mitigates overfitting. In this study, ten-fold cross-validation was employed on the training set to select the λ value with the minimum error. Meanwhile, a path diagram was drawn to visualize the gradual shrinkage of feature coefficients to zero as λ increases, yielding the feature subset screened by LASSO. Ultimately, features extracted from the time series model were integrated with LASSO-selected features to form the final feature set for model construction. 2.2 Construction and selection of multiple machine learning models For the training dataset, we constructed eight machine learning models, namely Logistic Regression(LR), Support Vector Machine (SVM), XGBoost, Naive Bayes(NB), Random Forest(RF), K-Nearest Neighbors (KNN), Decision Tree(DT), and Neural Network(NN). To optimize the hyperparameters of all models based on the ROC-AUC metric, we employed a 5-fold stratified cross-validation approach using StratifiedKFold and conducted grid search on the training set. This process yielded the optimal configurations for the aforementioned eight models. Subsequently, on the test dataset, we comprehensively evaluated the performance of these eight optimized models using a variety of metrics, including Accuracy, ROC-AUC, Sensitivity, Specificity, Precision, F1 Score, Matthews Correlation Coefficient (MCC), and Cohen's Kappa coefficient. A radar chart was utilized to visually aggregate and compare these metrics, thereby facilitating the identification of the most optimal model. Model Interpretation We utilized SHAP values as a means to interpret machine - learning models, thereby circumventing the "black - box" issue inherent in such models. SHAP values are an interpretability approach grounded in game theory. We quantify the significance of each feature for the model's output by computing the average contribution of that feature across all possible feature combinations to the model's prediction. Consequently, both the local interpretation of individual samples and the global ranking of feature importance are rendered transparent and lucid, bolstering the credibility of our findings. Employing this method, we identified the key features of the optimal model.Subsequently, the screened key features were utilized to develop a nomogram, which is a graphical tool that intuitively illustrates the functional relationships among multiple variables. The predictive accuracy and consistency of the model were assessed using the receiver operating characteristic (ROC) curve and calibration curve, respectively. Additionally, decision curve analysis (DCA) was conducted to evaluate the net benefit of the model for clinical application. Statistical analysis Measurement data that follow a normal distribution are expressed as mean ± standard deviation (SD), and independent t-tests are employed for comparisons between groups. For skewed data, the median and interquartile range (IQR) are used for description, with the Mann-Whitney U test applied to evaluate differences between groups. Categorical data are presented as frequencies and percentages, and chi-square tests or Fisher's exact test are utilized for intergroup comparisons. A p-value < 0.05 is considered indicative of statistical significance. All statistical analyses were conducted using R (version 4.4.3) and Python via PyCharm (Professional Edition, version 2024.3). Result Patient characteristics The flowchart of the research protocol is presented in Fig. 1 . Initially, 292 patients were enrolled in this study following strict inclusion and exclusion criteria. To balance the demographic characteristics, a 1:2 propensity score matching (PSM) strategy was employed, resulting in the final inclusion of 193 patients. Specifically, the MSAP group comprised 87 males and 37 females, while the SAP group included 51 males and 18 females. The mean age of patients was 42 years in both the MSAP and SAP groups. As shown in Table 1 , no significant statistical differences were observed in the demographic characteristics between the two groups after matching. Feature selection Feature selection for time series data In this study, we constructed a time - series model using a bidirectional LSTM neural network with a dual - branch hybrid architecture. To evaluate the model training process, we monitored the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve and the loss functions for both the training set and the validation set, as depicted in Fig. 2 .For the AUC curves (Fig. 2 A), both the training set (blue line) and validation set (orange line) exhibited a rapid ascent in the initial training iterations, plateauing at an AUC value close to 0.9 + interval(Train: 0.64 ~ 0.98, Val: 0.31 ~ 0.95). This indicates the model’s strong capability to distinguish between positive and negative samples, with consistent performance across training and validation data—suggesting no severe overfitting.Regarding the loss curves (Fig. 2 B), the training loss (blue line) and validation loss (orange line) showed a synchronized decline as training epochs increased.In the initial stage (the first 50 rounds), the training loss rapidly decreased from 3.258 to 1.097; the validation loss dropped from 3.220 to 1.032. In the middle stage (rounds 50 to 100), the descent trends of both curves slowed down, with the training loss reducing from 1.0973 to 0.509; the validation loss decreased from 1.032 to 0.580. In the later stage (rounds 100 to 167), both curves approached lower and stable values without significant divergence, with the training curve stabilizing at 0.349 and the validation curve stabilizing at 0.446. Both curves converged to low, stable values without significant divergence, which reflects effective model learning and good generalization ability (i.e., the model did not overfit to the training data, as validation loss did not rise while training loss continued to decrease).To evaluate the model's performance and generalization ability, we employed 10-fold cross-validation. As shown in the Fig. 3 , the results are as follows: AUC mean of 0.9551 ± 0.0158, accuracy mean of 0.9222 ± 0.0300, F1-score mean of 0.8956 ± 0.0364, training loss of 0.2992 ± 0.0328, and validation loss of 0.4132 ± 0.0651.To quantify feature importance in the trained model, we employed SHAP values, visualized in Fig. 4 .This plot ranks features based on the contributions of the statistical characteristics and temporal features of the time-series data to the model's predictions.As illustrated in the figure, Rmax, Pdiff_mean, and Tdiff_std emerge as the most critical features for R, P, and T datasets, respectively. Feature selection for blood test index data and imaging index data The initial dataset contains 16 features. By using the LASSO method with the optimal lambda value of 0.035 obtained through cross - validation, the target variables are reduced to 7 (as illustrated in Fig. 5 ). This process enhances the model’s balance between training and test set performance, yielding superior generalization ability. These 7 variables are: PE, Neu, HCT, Ca, TG, AMY, and CRP. Performance of machine learning models We conducted an analysis of eight supervised learning classification models, comprising single models (Logistic Regression, Support Vector Machine, Naive Bayes, K-Nearest Neighbors, Decision Tree, Neural Network) and ensemble models (Random Forest, XGBoost).The evaluation metrics encompass Accuracy, ROC-AUC, Sensitivity, Specificity, F1-Score, Precision, MCC, and Cohen’s Kappa, as demonstrated in Fig. 6 and quantified in Table 2 . Based on Fig. 6 and Table 2 , the Random Forest (RF) model outperforms other models across multiple key metrics: achieving an Accuracy of 0.8793, an F1 - Score of 0.8293, a Matthews Correlation Coefficient (MCC) of 0.7365, and a Cohen’s Kappa of 0.6951. In terms of ROC - AUC, RF scores 0.9588, trailing only the XGBoost model (0.9614), while the Decision Tree (DT) model records the lowest value at 0.8591. For Specificity, RF aligns with both the XGBoost and DT models, achieving a value of 0.9189—outperforming other models in this metric. As depicted in Fig. 6 , the RF model exhibits superior performance across multiple evaluation dimensions, with its polygonal area in the radar chart being the largest, which directly reflects its strongest comprehensive predictive capability. Conversely, the Neural Network (NN) model shows subpar performance across all evaluated metrics, which is reflected in the smallest polygonal area and weaker predictive power. Thus, the RF model is selected as the optimal model for subsequent research and analysis. Feature importance analysis To interpret the Random Forest(RF) model, we employed SHAP (SHapley Additive exPlanations) to systematically quantify the importance of features and their impacts on model predictions. As illustrated in Fig. 7 A (global feature importance) and Fig. 7 B (local SHAP effects at the feature level), we conducted a comprehensive analysis of both the hierarchical structure of feature contributions and their respective influence patterns. In Fig. 6 A, Rmax exhibits the most significant global influence, as evidenced by the broad distribution of its SHAP values, underscoring its central role as the primary driver in modulating model predictions. Pdiff_mean and HCT follow as key contributors, with their considerable mean SHAP values highlighting their essential roles in the model's decision-making architecture. By contrast, features including TG, CRP, and AMY demonstrate a clustering of SHAP values around zero, indicating their marginal impact on overall predictive outcomes. For Fig. 7 B, an observable trend is noted for Pdiff_mean, HCT, Tdiff_std, and PE, wherein increasing feature values are accompanied by a shift of SHAP values from negative to positive. This pattern suggests that these features systematically influence predictive outcomes: lower values suppress predictions, whereas higher values promote upward adjustments. Inverse trends are observed for Ca. Although no strict monotonic relationship is evident between Rmax and its SHAP values, extreme Rmax values (both upper and lower thresholds) are associated with pronounced deviations of SHAP values from zero, emphasizing the heightened impact of outliers in Rmax on model predictions. Visualization analysis of the optimal model To translate the random forest model into a clinically interpretable visual tool, we developed a predictive nomogram for estimating the occurrence probability of severe acute pancreatitis (SAP) (Fig. 8 ). It incorporates seven key features, namely Rmax, Pdiff_mean, HCT, Tdiff_std, PE, Neu, and Ca. For each feature, a specific value is mapped to a respective score via the upper score axis. Summing these individual scores yields a cumulative total score, which is subsequently mapped to the predicted SAP probability via the lower probability axis.The Bias-corrected predicted probability fitted the actual event rate (Ideal diagonal "IDEAL") well, with a mean absolute error of 0.019 and a mean square error of 0.00061 for both the training and testing sets (Fig. 9 A- 9 B). The ROC curve in Fig. 9 C showed that the AUC values of the training set and the test set were slightly different (0.969vs0.964). Finally, the decision curve of the model indicated consistently higher than the None and All lines at "high risk thresholds between 0.2 and 0.8", net benefit > 0.2, and cost-effectiveness ratios between 1:4 and 4:1.(as in Fig. 9 D) Discussion In this study, we constructed a dual-branch Bidirectional Long Short-Term Memory (BiLSTM) model based on time-series features, combined with LASSO feature selection for the optimal model (Random Forest model). The model was interpreted using SHAP values, and a convenient yet accurate nomogram model was subsequently established.During the research process, we innovatively used a BiLSTM network model to screen the most important features from time-series data (T, R, and P) within 24 hours after admission. These features were combined with those from blood test indicators and imaging indicators to construct eight ML models.We found that the comprehensive performance of the RF model outperformed other models, with an Accuracy of 0.8793, ROC-AUC of 0.9588, Sensitivity of 0.8095, Specificity of 0.9189, F1-Score of 0.8293, Precision of 0.85, MCC of 0.7365, and Cohen’s Kappa of 0.736.To avoid the black-box effect of ML models, we used the SHAP value interpretation tool to rank the importance of model features.To enable bedside triage of severe acute pancreatitis, we developed a visually intuitive nomogram incorporating the top 7 SHAP-identified features (e.g., Rmax, HCT). The nomogram demonstrated excellent discriminative ability (ROC-AUC = 0.964), with decision curve analysis confirming a net benefit > 0.2 across high-risk thresholds (0.2–0.8), outperforming both 'treat all' and 'treat none' strategies. Recently, an increasing number of scholars have constructed ML models to predict the severity and complications of AP. For example, Callum B et al. used a kernel logistic regression model with 8 variables (age, C-reactive protein, respiratory rate, partial pressure of oxygen in inhaled air, arterial blood pH, serum creatinine, white blood cell count, and Glasgow Coma Scale [GCS]) to predict severe complications, outperforming the traditional APACHE II score (AUC: 0.82 vs 0.74) [17] . Hong Wangdong et al. conducted a study on predicting severe pancreatitis based on a classification and regression tree model. They used logistic regression to screen out four variables—systemic inflammatory response syndrome (SIRS), pleural effusion, serum calcium, and blood urea nitrogen—to construct a model, which showed superior performance in early identification of high-risk severe pancreatitis compared with APACHE II (AUC: 0.84 vs 0.68) [18] . In addition, a multicenter, multinational prospective observational study published by Balázs et al. used binary classifiers (decision tree, random forest, logistic regression, support vector machine, CatBoost, and XGBoost) and identified XGBoost as the optimal model for predicting early SAP (AUC: 0.81 ± 0.03). The study also first used SHapley Additive exPlanations (SHAP) to screen the most influential features: respiratory rate, body temperature, abdominal reflex, gender, age, and blood glucose [19] . While state-of-the-art machine learning models outperform traditional scoring systems in AP severity prediction, the scientific rationale underlying feature extraction remains underdeveloped.Notably,longitudinal vital sign trajectories—critical for capturing the temporal dynamics of inflammatory cascades—are typically reduced to static univariate statistics (e.g., mean, interquartile range), neglecting sequential pattern recognition (e.g., autoregressive trends, volatility clustering). Our study addressed this gap by leveraging the power of BiLSTM networks to capture temporal dependencies in vital sign data. The advantages of the BiLSTM architecture in effectively mining temporal and nonlinear features from data have been confirmed by scholars.Shahid F et al. constructed predictive models including Autoregressive Integrated Moving Average (ARIMA), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (Bi-LSTM) to forecast time-series data of confirmed cases, deaths, and recoveries in ten major countries affected by COVID-19. The Bi-LSTM model demonstrated optimal performance in most scenarios, particularly for predicting China's mortality rate, which achieved the lowest MAE (0.0070) and RMSE (0.0077). Additionally, the Bi-LSTM model showed the highest r2_score (0.9997) for predicting recovered cases in China[ 20 ]. In this study, we evaluated the performance of the BiLSTM model using accuracy, AUC value, F1 score, and cross-entropy loss function. Ten-fold cross-validation showed that: the average accuracy was 0.922; the average AUC value was 0.9551; the average F1 score was 0.8956; the average training set loss function was 0.2992; and the average validation set loss function was 0.4132,indicating that the model maintains strong fitting capability alongside robust generalization performance.Finally, Rmax, Pdiff_mean, Tdiff_std, and other indicators were included to construct a nomogram model, which is similar to the characteristic variables screened by the above prediction model [17–19] .This further demonstrated that the systemic inflammatory response syndrome (manifested as T > 38°C or 20 breaths pear minute; P > 90 beats pear minute) plays a crucial role in the progression of acute pancreatitis[ 21 ].Traditionally, SIRS and CARS (compensatory anti-inflammatory response syndrome)are believed to occur sequentially, with their intensity and temporal order serving as key prognostic factors for AP. During the progression of the disease, regulated cell death (RCD) of pancreatic acinar cells plays a pivotal role in influencing AP severity through the following mechanisms: extrinsic apoptosis (activation of caspase-8 via death receptors like Fas/TNF receptors) and intrinsic apoptosis (mitochondrial membrane permeabilization triggering caspase-9 activation through mitochondrial proteins and HtrA serine peptidases), impaired autophagy compromising cellular homeostasis, caspase-independent necroptosis mediated by receptor-interacting protein kinases (RIP1-RIP3) and mixed lineage kinase domain-like protein (MLKL), pyroptosis activated via caspase-1-dependent canonical and caspase-4/5/11-dependent non-canonical pathways, and ferroptosis induced by iron-mediated oxidative damage.Recently,Sendler M et al. had proposed that interleukin-18 (IL-18), a cytokine secreted by macrophages, activates both SIRS and CARS in parallel[ 23 ]. Among the three features mentioned above, as a feature of importance, Rmax shows results consistent with those of Lei H et al., who found that a median respiratory rate of 22 breaths per minute is associated with lung injury in SAP[ 24 ]. Lung injury is the most common extrapancreatic organ dysfunction in SAP[ 25 ],but the specific injury mechanism remains unclear.Hu Q et al. found that emodin alleviates SAP-associated acute lung injury (ALI) by reducing pancreatic exosome secretion, regulating their protein composition, activating PPARγ to inhibit the NF-κB pathway, thereby suppressing M1 polarization of alveolar macrophages and release of pro-inflammatory factors[ 26 ].More importantly, the mortality rate of early SAP complicated with acute lung injury reaches 70% [ 27 ], further confirming the clinical value of the Rmax feature in the model. In addition to the lung, which is a common extrapancreatic organ injured in SAP, the heart is also similarly affected. Cardiac injury manifests diversely, including coronary vascular dysfunction, autonomic nervous dysfunction, and autophagic dysregulation. Heart rate variability (HRV), which reflects pathological changes in the sinoatrial node of cardiac cycle variations, serves as an important non-invasive quantitative indicator for evaluating cardiac autonomic neuropathy[ 28 ].Zhang et al. found that high-frequency norm (nHF) is a good indicator for predicting infected pancreatic necrosis (IPN) and multiple organ dysfunction syndrome (MODS). The areas under the ROC curves of nHF for predicting IPN and MODS were 0.927 and 0.821, respectively, which were superior to those of procalcitonin (AUC = 0.709 and 0.722) and APACHE II (AUC = 0.785 and 0.899).As the review published by Liau et al., HRV parameters obtained through time-domain analysis or frequency-domain analysis may be used to predict the severity and prognostic outcome of AP[ 29 ].In the model of this study, the mean value of heart rate difference (Pdiff_mean) was included, which reflects the degree of deviation from the baseline value and avoids the influence of the basal value. HCT, PE, Neu and Ca are another group of important features screened by the RF model.We found that high HCT values were more likely to be associated with SAP, which is consistent with the view of Koutroumpakis et al. They found that hematocrit ≥ 44% on admission and an elevated BUN level after 24 hours were the most accurate indicators for predicting persistent organ failure (AUC 0.67 and 0.71, respectively) and pancreatic necrosis (AUC 0.66 and 0.67, respectively), outperforming other laboratory parameters and the Acute Physiology and Chronic Health Evaluation-II score[ 30 ].In SAP, the elevation of HCT values may be associated with increased permeability of vascular endothelial cells, leading to a state of hemoconcentration. To this end, Komara et al. constructed a fluid model of vascular-interstitial-third space to analyze the permeability changes of capillaries to albumin (Alb) and non-albumin plasma proteins (NAPP = TP-Alb). The HCT in the MOF group increased by 5.00% (3.70%-8.70%) compared with the baseline, while that in the non-MOF group changed by only − 0.10% (-1.55–1.40%). An HCT > 3% had an OR of 17.7 (p = 0.014) for predicting MOF[ 31 ].PE is a prevalent complication of AP and serves as a precise predictor of SAP, as evidenced by Zeng et al.'s meta-analysis on AP with PE:the pooled incidence of PE in AP patients was 34% (95% CI: 28%-39%), with pooled incidences of 49% for unilateral PE and 59% for small PE. Notably, the PE incidence in SAP patients (67%) was significantly higher than that in mild acute pancreatitis (MAP, 14%). The meta-analysis further revealed a 3.99-fold increased risk of mortality in AP patients with PE (95% CI: 1.73–9.2), underscoring PE's prognostic significance in disease severity[ 32 ].Neutrophil recruitment also plays a critical role in the development of AP. For example, depleting neutrophils or inhibiting neutrophil infiltration by targeting specific adhesion molecules such as P-selectin, lymphocyte function antigen-1 (CD11a/CD18), and intercellular adhesion molecule-1 can improve tissue damage in AP[ 33 ].Niu et al. found that inhibiting neutrophil-specific ORAI1 calcium channels reduces pancreatitis-associated acute lung injury, with approximately 40% reduction in alveolar macrophage and neutrophil infiltration[ 34 ].Lei et al. also found that acetate produced by Parabacteroides can reduce neutrophil infiltration, thereby alleviating heparanase (Hpa)-exacerbated acute pancreatitis (AP)[ 35 ].The reduction in serum calcium may be fundamentally attributed to the inhibition of membrane Ca²⁺-ATPases during the onset of AP, which leads to enhanced calcium influx and intracellular calcium overload[ 36 ].Petersen et al. published a review article elaborating on the core roles of Ca²⁺ and ATP in the physiology and pathology of exocrine pancreas. For instance, in AP, excessive Ca²⁺ signals induced by factors like alcohol, fatty acids, bile acids, or physical pressure lead to mitochondrial dysfunction and reduced ATP production, thereby triggering cell necrosis[ 37 ].Li et al. analyzed the trajectories of total serum calcium (TSC) within 24 hours using group-based trajectory modeling and found that the in-hospital mortality rate was significantly higher in the "very low TSC" group (odds ratio [OR], 7.2; 95% CI, 3.7–14.0), the "moderately low TSC" group (OR, 5.0; 95% CI, 3.8–6.7), and the "fluctuating high TSC" group (OR, 5.6; 95% CI, 1.5–20.6) compared with the "stable normal-calcium" group[ 38 ]. Conversely, the last three features (CRP,AMY and TG) selected based on the RF model were not included in the construction of the nomogram. CRP is associated with SAP, which is also reflected in our study. However, as demonstrated by Tarján et al., for predicting necrotizing pancreatitis, the area under the curve (AUC) of CRP within 72 hours after admission is 0.69 (confidence interval [CI]: 0.62–0.76), while the pooled AUC of CRP after 72 hours is 0.88 (CI: 0.75–1.00)[ 39 ]. CRP measured after 72 hours is more clinically valuable, but the CRP data in this study were obtained within 24 hours of admission. Therefore, the CRP feature was excluded.AMY is considered to have no contributory value in the assessment of the severity of AP[ 40 ].Although some scholars have shown that in AP with hypertriglyceridemia, the triglyceride (TG) levels in severe or moderately severe patients within 24 and 48 hours after symptom onset are significantly higher than those in mild patients (P < 0.050), and TG levels are associated with clinical outcomes[ 41 ].However, the AP included in this study are not limited to hyperlipidemic pancreatitis.This is also a shortcoming of this study, as it does not elaborate on the classification of AP types in detail. This study also has several limitations: 1) The sample size is small and the data are from a single center, requiring multicenter large-sample validation of the model's generalization ability and external validation to test the performance of the nomogram; 2) Time-series data collection could record data at a higher frequency to capture more precise temporal features. Conclusion We constructed a BiLSTM-RF model based on time-series features to predict SAP, and screened out 7 key features including Rmax and HCT to develop a nomogram. The model innovatively utilizes dynamic features of vital signs to enhance prediction accuracy, providing a visual tool for early assessment of SAP. Table 1 demographic characteristics before and after propensity score matching MSAP n = 213 SAP n = 79 P vs MSAP SAP P n = 124 n = 69 Gender 167/46 58/21 0.368 87/37 51/18 0.580 (M/F) ASA 88/125 25/54 0.132 50/74 22/47 0.245 (I/II) Smoking 77/136 36/43 0.142 49/75 29/40 0.733 (Y/N) Drinking 59/154 35/44 0.007 42/82 29/40 0.260 (Y/N) Age 44 ± 13.14 40 ± 14.04 0.093 42 ± 12.08 42 ± 14.68 0.961 CCI 1.16 ± 1.24 0.94 ± 1.21 0.170 0.96 ± 1.14 1.01 ± 1.27 0.759 BMI 26.16 ± 3.58 27.81 ± 3.90 0.001 26.88 ± 3.75 27.34 ± 3.75 0.414 NRS2002 0.21 ± 0.41 0.30 ± 0.46 0.120 0.23 ± 0.42 0.28 ± 0.45 0.445 Table 2 Table of Performance Metrics for 8 Classification Models Model Accuracy ROC-AUC Sensitivity Specificity F1 Score Precision MCC Cohen's Kappa XGBoost 0.8621 0.9614 0.7619 0.9189 0.8 0.8421 0.6972 0.6951 RF 0.8793 0.9588 0.8095 0.9189 0.8293 0.85 0.7365 0.736 LR 0.8621 0.9228 0.8571 0.8649 0.8182 0.7826 0.7093 0.7074 SVM 0.7931 0.9228 0.9048 0.7297 0.76 0.6552 0.6099 0.5862 KNN 0.7931 0.9151 0.8571 0.7568 0.75 0.6667 0.5915 0.5782 NB 0.8103 0.8938 0.8571 0.7838 0.766 0.6923 0.6194 0.6095 NN 0.7586 0.8662 0.8095 0.7297 0.7083 0.6296 0.5196 0.5079 DT 0.8621 0.8591 0.7619 0.9189 0.8 0.8421 0.6972 0.6951 Abbreviations AP Acute pancreatitis MSAP Moderate severe acute pancreatitis SAP Severe acute pancreatitis APCC Average annual percentage change APACHE II Acute Physiology and Chronic Health Evaluation II BISAP Bedside Index of Severity in Acute Pancreatitis SIRS Bedside Index of Severity in Acute Pancreatitis CARS Compensatory anti-inflammatory response syndrome ML Machine learning RNNs Recurrent Neural Networks LSTM Long Short-Term Memory BiLSTM Bidirectional long short-term memory network SHAP SHapley Additive exPlanations ASA-class Anesthesiologists classification CCI Charlson comorbidity index BMI Body mass index NRS Nutritional Risk Score T Temperature P Pulse R Respiratory rate WBC White blood cell CRP C-reactive protein Neu Neutrophil RBC Red blood cell Plt Platelets HCT Hematocrit ALT Alanine transaminase AST Aspartate transaminase AMY Amylase LIP Lipase TBil Total bilirubin Ca Calcium Glu Glucose Alb Albumin BUN Blood urea nitrogen PE Pleural effusion PSM Propensity score matching ReLU Rectified linear unit AUC Area under the curve LASSO Least absolute shrinkage and selection operator LR Logistic regression SVM Support vector machine NB Naive bayes RF Random forest KNN K-nearest neighbors DT Decision tree NN Neural network ROC Receiver operating characteristic DCA Decision curve analysis Rmax Maximum respiratory rate Pdiff_mean The average value of the difference in pulse rates Tdiff_std The standard deviation of the difference in body temperatures HRV Heart rate variability IPN Infected pancreatic necrosis Declarations Ethics approval and consent to participate Approval from the Ethics Committee of Zhejiang Xiaoshan Hospital(NO.Y2024088). Consent for publication Not applicable. Availability of data and material Yes. Competing interests The authors declare no competing interests. Funding This work was supported by Zhejiang Health Information Association Research Program,number 2024XHZN-C04. Authors' contributions Y C conceived and designed the study, collected the data, analyzed the data, and drafted the manuscript; ZD F: conceived and designed the study, collected the data, analyzed the data YZ L and F J: collected the data, analyzed the data and draw the figures. TM Z and LH J: conceived and designed the study, and revised the manuscript; All authors have read and approved the submitted manuscript. Acknowledgements Thank you for the contributions of all authors. YC,ZDF, YZL, and LHJ conceived and designed the study, analyzed the data, wrote the manuscript, and contributed to the methodological designs of this study. References Boxhoorn L, Voermans RP, Bouwense SA, et al. Acute pancreatitis [J]. Lancet, 2020, 396(10252): 726-34.doi: 10.1016/S0140-6736(20)31310-6. Iannuzzi JP, King JA, Leong JH, et al. Global Incidence of Acute Pancreatitis Is Increasing Over Time: A Systematic Review and Meta-Analysis [J]. Gastroenterology, 2022, 162(1): 122-34.doi: 10.1053/j.gastro.2021.09.043. Peery AF, Crockett SD, Murphy CC, et al. Burden and Cost of Gastrointestinal, Liver, and Pancreatic Diseases in the United States: Update 2021 [J]. Gastroenterology, 2022, 162(2): 621-44.doi: 10.1053/j.gastro.2021.10.017. Singh VK, Yadav D, Garg PK. Diagnosis and Management of Chronic Pancreatitis: A Review [J]. JAMA, 2019, 322(24): 2422-34.doi: 10.1001/jama.2019.19411. Ranson JH, Rifkind KM, Roses DF, et al. Prognostic signs and the role of operative management in acute pancreatitis [J]. Surg Gynecol Obstet, 1974, 139(1): 69-81. Al-Hadeedi S, Fan ST, Leaper D. APACHE-II score for assessment and monitoring of acute pancreatitis [J]. Lancet, 1989, 2(8665): 738. Singh VK, Wu BU, Bollen TL, et al. A prospective evaluation of the bedside index for severity in acute pancreatitis score in assessing mortality and intermediate markers of severity in acute pancreatitis [J]. Am J Gastroenterol, 2009, 104(4): 966-71. Balthazar EJ, Robinson DL, Megibow AJ, et al. Acute pancreatitis: value of CT in establishing prognosis [J]. Radiology, 1990, 174(2): 331-6. Arif A, Jaleel F, Rashid K. Accuracy of BISAP score in prediction of severe acute pancreatitis [J]. Pak J Med Sci, 2019, 35(4): 1008-12. Balthazar EJ. Acute pancreatitis: assessment of severity with clinical and CT evaluation [J]. Radiology, 2002, 223(3): 603-13. Deo RC. Machine Learning in Medicine [J]. Circulation, 2015, 132(20): 1920-30. Pantanowitz L, Pearce T, Abukhiran I, et al. Non-Generative Artificial Intelligence (AI) in Medicine: Advancements and Applications in Supervised and Unsupervised Machine Learning [J]. Mod Pathol, 2024, 100680. Rashidi HH, Hu B, Pantanowitz J, et al. Statistics of Generative Artificial Intelligence and Nongenerative Predictive Analytics Machine Learning in Medicine [J]. Mod Pathol, 2024, 38(3): 100663. Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput. 1997;9(8):1735-80. Graves A, Schmidhuber J. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Netw. 2005 Jun-Jul;18(5-6):602-10. Deforth M, Heinze G, Held U. The performance of prognostic models depended on the choice of missing value imputation algorithm: a simulation study. J Clin Epidemiol. 2024;176:111539. Pearce CB, Gunn SR, Ahmed A, Johnson CD. Machine learning can improve prediction of severity in acute pancreatitis using admission values of APACHE II score and C-reactive protein. Pancreatology. 2006;6(1-2):123-31. Hong W, Dong L, Huang Q, et al. Prediction of severe acute pancreatitis using classification and regression tree analysis [J]. Dig Dis Sci, 2011, 56(12): 3664-71. Kui B, Pintér J, Molontay R, et al. EASY-APP: An artificial intelligence model and application for early and easy prediction of severity in acute pancreatitis [J]. Clin Transl Med, 2022, 12(6): e842. Shahid F, Zameer A, Muneeb M. Predictions for COVID-19 with deep learning models of LSTM, GRU and Bi-LSTM. Chaos Solitons Fractals. 2020 Nov;140:110212. Bone RC, Balk RA, Cerra FB, et al. Definitions for sepsis and organ failure and guidelines for the use of innovative therapies in sepsis. The ACCP/SCCM Consensus Conference Committee. American College of Chest Physicians/Society of Critical Care Medicine. Chest. 1992;101(6):1644-55. Li H, Wu D, Zhang H, Li P. New insights into regulatory cell death and acute pancreatitis. Heliyon. 2023;9(7):e18036. Sendler M, van den Brandt C, Glaubitz J, et al. NLRP3 Inflammasome Regulates Development of Systemic Inflammatory Response and Compensatory Anti-Inflammatory Response Syndromes in Mice With Acute Pancreatitis. Gastroenterology. 2020;158(1). Lei H, Minghao W, Xiaonan Y, et al. Acute lung injury in patients with severe acute pancreatitis. Turk J Gastroenterol. 2013;24(6):502-7. Garg PK, Singh VP. Organ Failure Due to Systemic Injury in Acute Pancreatitis. Gastroenterology. 2019;156(7):2008-23. Hu Q, Yao J, Wu X, et al. Emodin attenuates severe acute pancreatitis-associated acute lung injury by suppressing pancreatic exosome-mediated alveolar macrophage activation. Acta Pharm Sin B. 2022;12(10):3986-4003. Hu Q, Zhang S, Yang Y, et al. Extracellular Vesicle ITGAM and ITGB2 Mediate Severe Acute Pancreatitis-Related Acute Lung Injury. ACS Nano. 2023;17(8):7562-75. Luo Y, Li Z, Ge P, et al. Comprehensive Mechanism, Novel Markers and Multidisciplinary Treatment of Severe Acute Pancreatitis-Associated Cardiac Injury - A Narrative Review. J Inflamm Res. 2021;14:3145-69. Liau MYQ, Liau JYJ, Selvakumar SV, et al. Heart rate variability in acute pancreatitis: a narrative review. Transl Gastroenterol Hepatol. 2024 Aug 21;9:68. Koutroumpakis E, Wu BU, Bakker OJ, et al. Admission Hematocrit and Rise in Blood Urea Nitrogen at 24 h Outperform other Laboratory Markers in Predicting Persistent Organ Failure and Pancreatic Necrosis in Acute Pancreatitis: A Post Hoc Analysis of Three Large Prospective Databases. Am J Gastroenterol. 2015 Dec;110(12):1707-16. Komara NL, Paragomi P, Greer PJ, et al. Severe acute pancreatitis: capillary permeability model linking systemic inflammation to multiorgan failure. Am J Physiol Gastrointest Liver Physiol. 2020 Nov 1;319(5):G573-G583. Zeng T, An J, Wu Y, et al. Incidence and prognostic role of pleural effusion in patients with acute pancreatitis: a meta-analysis. Ann Med. 2023;55(2):2285909. Renström E, Luo L, Mörgelin M, et al. Neutrophil Extracellular Traps Induce Trypsin Activation, Inflammation, and Tissue Damage in Mice With Severe Acute Pancreatitis. Gastroenterology. 2015 Dec;149(7):1920-1931.e8. Niu M, Zhang X, Wu Z, et al. Neutrophil-specific ORAI1 Calcium Channel Inhibition Reduces Pancreatitis-associated Acute Lung Injury. Function (Oxf). 2023 Oct 23;5(1):zqad061. Lei Y, Tang L, Liu S, et al. Parabacteroides produces acetate to alleviate heparanase-exacerbated acute pancreatitis through reducing neutrophil infiltration. Microbiome. 2021 May 20;9(1):115. Lee PJ, Papachristou GI. New insights into acute pancreatitis. Nat Rev Gastroenterol Hepatol. 2019 Aug;16(8):479-496. Petersen OH, Gerasimenko JV, Gerasimenko OV, et al. The roles of calcium and ATP in the physiology and pathology of the exocrine pancreas. Physiol Rev. 2021;101(4):1691-744. Li CL, Lin XC, Jiang M. Identifying novel acute pancreatitis sub-phenotypes using total serum calcium trajectories. BMC Gastroenterol. 2024 Apr 23;24(1):141. Tarján D, Szalai E, Lipp M, et al. Persistently High Procalcitonin and C-Reactive Protein Are Good Predictors of Infection in Acute Necrotizing Pancreatitis: A Systematic Review and Meta-Analysis. Int J Mol Sci. 2024 Jan 20;25(2):1273. Yadav D, Agarwal N, Pitchumoni CS. A critical evaluation of laboratory tests in acute pancreatitis. Am J Gastroenterol. 2002 Jun;97(6):1309-18. Dong X, Pan S, Zhang D, et al. Hyperlipemia pancreatitis onset time affects the association between elevated serum triglyceride levels and disease severity. Lipids Health Dis. 2022 May 30;21(1):49. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 24 Apr, 2026 Read the published version in BMC Gastroenterology → Version 1 posted Editorial decision: Revision requested 27 Feb, 2026 Reviews received at journal 22 Feb, 2026 Reviewers agreed at journal 17 Feb, 2026 Reviewers agreed at journal 28 Jan, 2026 Reviews received at journal 24 Aug, 2025 Reviewers agreed at journal 14 Aug, 2025 Reviewers invited by journal 05 Aug, 2025 Editor assigned by journal 30 Jul, 2025 Editor invited by journal 09 Jul, 2025 Submission checks completed at journal 08 Jul, 2025 First submitted to journal 08 Jul, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7029031","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":495836005,"identity":"a32fe2ce-1002-4c75-8bfa-d1da4e6d8b99","order_by":0,"name":"Yang Chen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAy0lEQVRIiWNgGAWjYBACxobDBx///SdR38/eQKQW5sZjyQY8bBaMM3sOEKmFvfmMmgQPWwXjhhsJRGrhbTvDJiHBI8FscPPxxhsMNTbRBLVI9pw9bGEgIcEmeTut2ILhWFpuAyEthjPOJd5IMJDg4budYyYBDAzCWuzvvzGQOJAgIcFw8wyRWhgbzhhJNhyQMBC4wUO0lmPJxowNEgmSPUC/JBDjF3BUMjbUJfCzH95440ONDWEtyMBAIoEU5RAtpOoYBaNgFIyCkQEALYlCvqJ0ykEAAAAASUVORK5CYII=","orcid":"","institution":"Affiliated Hangzhou Normal University","correspondingAuthor":true,"prefix":"","firstName":"Yang","middleName":"","lastName":"Chen","suffix":""},{"id":495836006,"identity":"a579dfee-17b2-4f05-b1a0-9377fa5fa325","order_by":1,"name":"Zhidong Fu","email":"","orcid":"","institution":"Affiliated Hangzhou Normal University","correspondingAuthor":false,"prefix":"","firstName":"Zhidong","middleName":"","lastName":"Fu","suffix":""},{"id":495836007,"identity":"7315141e-8733-473b-96f0-25e5f636d8bc","order_by":2,"name":"Yongzhi Liu","email":"","orcid":"","institution":"Affiliated Hangzhou Normal University","correspondingAuthor":false,"prefix":"","firstName":"Yongzhi","middleName":"","lastName":"Liu","suffix":""},{"id":495836008,"identity":"4d4607fa-fba9-46f4-b57a-464cb3da20ad","order_by":3,"name":"Feng Jiang","email":"","orcid":"","institution":"Affiliated Hangzhou Normal University","correspondingAuthor":false,"prefix":"","firstName":"Feng","middleName":"","lastName":"Jiang","suffix":""},{"id":495836009,"identity":"732a0296-1663-4c19-88af-734c253ff355","order_by":4,"name":"Tieming Zhu","email":"","orcid":"","institution":"Affiliated Hangzhou Normal University","correspondingAuthor":false,"prefix":"","firstName":"Tieming","middleName":"","lastName":"Zhu","suffix":""},{"id":495836010,"identity":"f77f63a8-fb7c-4d36-8231-032c5b3c00cd","order_by":5,"name":"Lihui Jiang","email":"","orcid":"","institution":"Affiliated Hangzhou Normal University","correspondingAuthor":false,"prefix":"","firstName":"Lihui","middleName":"","lastName":"Jiang","suffix":""}],"badges":[],"createdAt":"2025-07-02 11:53:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7029031/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7029031/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12876-026-04856-9","type":"published","date":"2026-04-24T16:00:12+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":88644017,"identity":"cf43bbe6-bcf8-47b8-a95a-124165857be5","added_by":"auto","created_at":"2025-08-08 16:18:57","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":293248,"visible":true,"origin":"","legend":"\u003cp\u003eflowchart of the study\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/30188ae43a0f329e71f78ff4.png"},{"id":88644019,"identity":"47f62f30-d221-404a-b4d4-b6012b827658","added_by":"auto","created_at":"2025-08-08 16:18:57","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":56372,"visible":true,"origin":"","legend":"\u003cp\u003eModel performance during training. (A) AUC curves on training and validation sets. (B) Training and validation loss curves.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/6cb86587a39d84da301b8c8e.png"},{"id":88644018,"identity":"43cf31e8-1d90-4f4d-81e4-66407dd3289d","added_by":"auto","created_at":"2025-08-08 16:18:57","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":95729,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of 10-fold cross-validation results\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/780b81fdbc121d408092012d.png"},{"id":88644045,"identity":"897507cd-1b1f-4022-9c25-eac54ca70a49","added_by":"auto","created_at":"2025-08-08 16:18:58","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":74157,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP Importance ranking of engineered features for interpretability of the dual - branch hybrid bidirectional LSTM Time - Series Model\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/9e09e78769513391dabf31a4.png"},{"id":88644022,"identity":"a1ef5d31-3d7c-4900-a315-2a2e3fda515a","added_by":"auto","created_at":"2025-08-08 16:18:57","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":171289,"visible":true,"origin":"","legend":"\u003cp\u003e(A)Feature selection in Lasso regression analysis;(B)Effects of regularization parameter λ variation on coefficients of 16 predictor variables\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/3bf09d897707df8966866b37.png"},{"id":88645577,"identity":"3eea7a5c-7119-4e0c-baeb-3b4ea521fa62","added_by":"auto","created_at":"2025-08-08 16:26:57","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":228463,"visible":true,"origin":"","legend":"\u003cp\u003eRadar Chart Comparing Performance of 8 Classification Models Across Key Metrics\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/3828e54fcef5297719785d0c.png"},{"id":88644024,"identity":"1c9b3caa-b572-472a-90de-99c92c48afae","added_by":"auto","created_at":"2025-08-08 16:18:57","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":402733,"visible":true,"origin":"","legend":"\u003cp\u003e(A) Random Forest SHAP feature importance, showing average impact of each feature on model output via SHAP values (x - axis) and feature value magnitude (color: red = high, blue = low). (B) Individual SHAP scatter plots. Each shows a feature’s value (x - axis) vs. its SHAP value (y - axis, prediction contribution), with color encoding an additional feature for interaction analysis.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/74c81f286385628883db096b.png"},{"id":88644033,"identity":"ac36d051-251d-42c8-a4c4-fa31be89253e","added_by":"auto","created_at":"2025-08-08 16:18:57","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":19743,"visible":true,"origin":"","legend":"\u003cp\u003eNomogram for predicting the occurrence probability of severe acute pancreatitis (SAP). This nomogram integrates seven key features (Rmax, Ca, PE, HCT, Pdiff_mean, Tdiff_std, and Neu).\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/640a2dfe6da199451cd82699.png"},{"id":88644047,"identity":"ef3de6af-3e3a-46b3-b5cf-645f72609c07","added_by":"auto","created_at":"2025-08-08 16:18:58","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":175888,"visible":true,"origin":"","legend":"\u003cp\u003eCalibration curves of the nomogram in the training set (A) and test set (B); Receiver operating characteristic (ROC) curve comparison between the training and test sets (C); Decision curve analysis of the nomogram in the validation set (D).\u003c/p\u003e","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/dcc8a9f16fe7b9a989dcae11.png"},{"id":107929153,"identity":"14515d17-5a39-4944-a5fa-fddbc7415cdb","added_by":"auto","created_at":"2026-04-27 16:13:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1750019,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7029031/v1/b0497076-b76e-424b-b6c7-61f4c08a9927.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Time-Series Feature-Based Nomogram for the Prediction of Severe Acute Pancreatitis","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAcute pancreatitis (AP) ranks among the most prevalent acute abdominal emergencies requiring hospitalization, with an annual incidence of approximately 30 cases per 100,000 population in high-income countries [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Recent epidemiological evidence highlights a rising trend in regional incidences, such as the average annual percentage change (APCC) of 3.67% in North America and 2.77% in Europe [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The socioeconomic burden of AP is substantial, with annual healthcare costs exceeding \u003cspan\u003e$\u003c/span\u003e2.6\u0026nbsp;billion in the United States. For patients with AP complicated by severe morbidities, the mean per-patient expenditure escalates to approximately \u003cspan\u003e$\u003c/span\u003e90,000 [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eApproximately 20% of AP cases progress to moderate-severe or severe phenotypes, where local and systemic inflammatory cascades drive a mortality rate of 20%-40% [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The overall mortality across all AP severities remains at 2%, underscoring the critical need for early prediction of disease severity to inform timely interventional strategies [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe evolution of traditional AP predictive models began five decades ago with the seminal publication of the Ranson score by Ranson et al.[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], the first grading system for AP severity. Subsequent decades have seen the emergence of numerous assessment tools, including the Acute Physiology and Chronic Health Evaluation II (APACHE II) [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], Bedside Index of Severity in Acute Pancreatitis (BISAP) [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], and Balthazar CT scoring system [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. However, these scoring criteria exhibit notable limitations in clinical timeliness and accuracy. For instance, while five parameters of the Ranson score are obtained at admission, six require 48-hour follow-up data, introducing a critical lag that compromises early disease progression assessment. The BISAP score, derived from 24-hour admission data (blood urea nitrogen \u0026gt; 8.9 mmol/L, altered mental status, systemic inflammatory response syndrome (SIRS), age \u0026gt; 60 years, and pleural effusion), prioritizes accessibility but lacks optimal diagnostic performance. A study by Anum et al. demonstrated that BISAP had lower accuracy (76.2% vs. 82.2%) and sensitivity (69.2% vs. 97.4%) than the Ranson score for predicting severe pancreatitis [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. The APACHE II score, comprising acute physiology, chronic health, and age components, requires comprehensive 24-hour data collection, hindering its utility in early severe pancreatitis assessment. The Balthazar CT score, dependent on contrast-enhanced CT for pancreatic necrosis evaluation, achieves 90% early detection sensitivity (72–96 hours post-onset), which improves to 100% after four days, thus limiting its application in acute-phase severity stratification [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eWith the continuous refinement of statistical theories and the rapid advancement of computer technology, machine learning (ML), as a subset of artificial intelligence (AI), has been increasingly adopted in clinical practice [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Most established and reliable AI tools fall within the domain of non-generative AI, encompassing both supervised and unsupervised machine learning techniques [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Compared to generative AI, supervised machine learning models in medicine offer enhanced interpretability, as their outputs (e.g., class labels or continuous values) can be directly linked to clinical outcomes. Additionally, a robust framework of statistical metrics—including accuracy, precision, sensitivity, and specificity—has been developed to evaluate model performance [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].Therefore, the inclusion of comprehensive and systematic feature variables is pivotal for model construction. Previous studies on AP predictive models have rarely focused on the extraction of time-series features from vital signs (e.g., body temperature, pulse, respiratory rate, etc.). As a branch of Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) addresses the common vanishing/exploding gradient problems in RNNs by introducing gated memory units, making it highly suitable for sequence-related problems [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Graves et al. proposed the BiLSTM architecture, which can fully utilize contextual information and outperforms unidirectional LSTM [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Meanwhile, SHAP (SHapley Additive exPlanations) provides a new approach to solve the \"black-box\" effect of machine learning models. Therefore, this study aims to develop a multimodal model for predicting SAP based on time-series features and construct a nomogram according to SHAP values to provide evidence for clinical decision-making.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cb\u003ePatients\u003c/b\u003e\u003c/p\u003e\u003cp\u003eA retrospective analysis was performed on clinical data of patients with AP treated at Xiaoshan Hospital Affiliated to Hangzhou Normal University between January 2018 and December 2024. Inclusion criteria were as follows: ① 85 years ≥ age ≥ 18 years; ② AP defined as serum amylase elevation to at least three times the upper limit of normal, accompanied by typical symptoms of acute pancreatitis. SAP was defined according to the revised Atlanta Consensus Document. Exclusion criteria included: ① history of severe cardiovascular, cerebrovascular, respiratory, renal, or hepatic diseases; ② previous or current history of malignant tumors; ③ Child-Pugh class ≥ C;④ Long-term oral administration of hormonal and psychotropic medications.\u003c/p\u003e\u003cp\u003eThis retrospective medical record study has obtained the approval of the Ethics Committee of Zhejiang Xiaoshan Hospital (Approval Number: Y2024088). All procedures involving human participants in this study were conducted in accordance with the ethical standards of the aforementioned institutional ethics committee and with the 1964 Declaration of Helsinki and its subsequent amendments.\u003c/p\u003e\u003cp\u003eIn line with the requirements of the Declaration of Helsinki, the ethics committee of Zhejiang Xiaoshan Hospital granted a waiver of informed consent for this study. The waiver was justified based on the following considerations: the study is of low-risk nature, involving only the analysis of de-identified historical medical records, which poses no threat to the rights or interests of patients and will not cause any harm to them.\u003c/p\u003e\u003cp\u003eThe research team will strictly ensure the confidentiality of all medical records and abide by all relevant ethical and legal requirements throughout the study.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData collection\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe present study collected general characteristic indices, time - series indices, laboratory indices, and imaging indices within 24 hours of admission.\u003c/p\u003e\u003cp\u003eFor the binary variable labels, SAP was labeled as 1, while MSAP was labeled as 0.Male was labeled as 1, while female was labeled as 0.Smoking: coded as 1, otherwise 0.Alcohol consumption: coded as 1, otherwise 0.Presence of pleural effusion: coded as 1, otherwise 0.\u003c/p\u003e\u003cp\u003eGeneral characteristic indices collected for each patient included demographic characteristics (gender and age), American Society of Anesthesiologists classification (ASA-class), Charlson Comorbidity Index (CCI), body mass index (BMI), and Nutritional Risk Score (NRS2002).\u003c/p\u003e\u003cp\u003eTime-series indices comprised body temperature(T), pulse༈P༉, and respiratory rate༈R༉ .Upon admission, a set of data was recorded and labeled as T0, P0, and R0. Thereafter, data were recorded every 2 hours within the subsequent 24 hours, labeled as T1 to T13, P1 to P13, and R1 to R13.\u003c/p\u003e\u003cp\u003eLaboratory indices included white blood cell count(WBC), C-reactive protein (CRP), neutrophil count(Neu), red blood cells(RBC), platelets(Plt), hematocrit(HCT), alanine transaminase (ALT), aspartate transaminase (AST), serum amylase(AMY), serum lipase(LIP), total bilirubin(TBil), serum calcium(Ca), blood glucose(Glu), albumin(Alb), and blood urea nitrogen (BUN).\u003c/p\u003e\u003cp\u003eImaging indices focused on chest CT or chest radiograph indicates the presence of pleural effusion(PE).\u003c/p\u003e\u003cp\u003e\u003cb\u003eData processing\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo achieve a more balanced representation of the sample data characteristics, we incorporated demographic features (gender, age), American Society of Anesthesiologists classification (ASA-class), Charlson Comorbidity Index (CCI), Body Mass Index (BMI), and Nutritional Risk Score 2002 (NRS2002). Subsequently, propensity score matching(PSM) was employed to select the final cases based on the severity of acute pancreatitis (AP).In this study, missing values were managed using multiple imputation techniques[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Specifically, logistic regression imputation was employed for binary variables, whereas predictive mean matching imputation was utilized for continuous variables. Variables with univariate missing values exceeding 30% were directly excluded from the analysis. For variables requiring standardization, standard deviation standardization (i.e., z-score normalization) was applied to ensure consistency across measurements.\u003c/p\u003e\u003cp\u003e\u003cb\u003eModel Construction and Evaluation\u003c/b\u003e\u003c/p\u003e\u003cp\u003e1)Extraction of features from time-series models\u003c/p\u003e\u003cp\u003eFeatures of the T/P/R time series data in the study sample were extracted by integrating the characteristics of T0-12, P0-12, and R0-12 datasets. The number of time steps was determined to be 13, as each time series comprised observations at 13 discrete time points (indexed from 0 to 12), which comprehensively captured the temporal dynamics of the variables.Feature engineering encompasses time series features (upward/downward streaks, volatility characteristics, and stability characteristics) and statistical features (mean, standard deviation, skewness, kurtosis, median, range, interquartile range(iqr), slope, and delta).A dual-branch hybrid architecture based on a Bidirectional Long Short-Term Memory (BiLSTM) neural network is utilized for processing time-series data features. In the first branch, temporal features are extracted using a BiLSTM layer with 16 hidden neurons. To enhance generalization and prevent overfitting, Dropout and Batch Normalization techniques are incorporated. The second branch consists of a fully connected layer with 16 neurons, which processes manually engineered features. This branch employs Rectified Linear Unit (ReLU) activation and L2 regularization to mitigate overfitting. Subsequently, a concatenation layer integrates the outputs from both branches. The fused data is then passed through another fully connected layer with 16 neurons, utilizing ReLU activation and L2 regularization. Finally, the output layer uses a sigmoid activation function, which is appropriate for binary classification tasks.\u003c/p\u003e\u003cp\u003eThe Adam optimizer is employed to adaptively adjust the learning rate during training. Binary cross-entropy is selected as the loss function to measure the discrepancy between predicted probabilities and true labels. Model performance is assessed using metrics such as the Area Under the Curve (AUC) and loss curves. Additionally, early stopping and learning rate decay strategies are implemented to improve model stability and convergence. Furthermore, SHAP (SHapley Additive exPlanations) values are calculated to provide interpretable feature importance rankings and visualizations, enabling the identification of the most critical features derived from T/P/R metrics.\u003c/p\u003e\u003cp\u003e2)Optimal Model\u003c/p\u003e\u003cp\u003e2.1 Feature Extraction\u003c/p\u003e\u003cp\u003eWe implemented LASSO (Least Absolute Shrinkage and Selection Operator) for feature screening of hematological indices and imaging markers. LASSO extends linear regression by incorporating an L1 regularization term, which zeros out the weights of irrelevant features to achieve feature selection. Simultaneously, by constraining the magnitude of weights, it enhances the model’s generalization capability on unseen data and mitigates overfitting. In this study, ten-fold cross-validation was employed on the training set to select the λ value with the minimum error. Meanwhile, a path diagram was drawn to visualize the gradual shrinkage of feature coefficients to zero as λ increases, yielding the feature subset screened by LASSO.\u003c/p\u003e\u003cp\u003eUltimately, features extracted from the time series model were integrated with LASSO-selected features to form the final feature set for model construction.\u003c/p\u003e\u003cp\u003e2.2 Construction and selection of multiple machine learning models\u003c/p\u003e\u003cp\u003eFor the training dataset, we constructed eight machine learning models, namely Logistic Regression(LR), Support Vector Machine (SVM), XGBoost, Naive Bayes(NB), Random Forest(RF), K-Nearest Neighbors (KNN), Decision Tree(DT), and Neural Network(NN). To optimize the hyperparameters of all models based on the ROC-AUC metric, we employed a 5-fold stratified cross-validation approach using StratifiedKFold and conducted grid search on the training set. This process yielded the optimal configurations for the aforementioned eight models.\u003c/p\u003e\u003cp\u003eSubsequently, on the test dataset, we comprehensively evaluated the performance of these eight optimized models using a variety of metrics, including Accuracy, ROC-AUC, Sensitivity, Specificity, Precision, F1 Score, Matthews Correlation Coefficient (MCC), and Cohen's Kappa coefficient. A radar chart was utilized to visually aggregate and compare these metrics, thereby facilitating the identification of the most optimal model.\u003c/p\u003e\u003cp\u003e\u003cb\u003eModel Interpretation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe utilized SHAP values as a means to interpret machine - learning models, thereby circumventing the \"black - box\" issue inherent in such models. SHAP values are an interpretability approach grounded in game theory. We quantify the significance of each feature for the model's output by computing the average contribution of that feature across all possible feature combinations to the model's prediction. Consequently, both the local interpretation of individual samples and the global ranking of feature importance are rendered transparent and lucid, bolstering the credibility of our findings. Employing this method, we identified the key features of the optimal model.Subsequently, the screened key features were utilized to develop a nomogram, which is a graphical tool that intuitively illustrates the functional relationships among multiple variables. The predictive accuracy and consistency of the model were assessed using the receiver operating characteristic (ROC) curve and calibration curve, respectively. Additionally, decision curve analysis (DCA) was conducted to evaluate the net benefit of the model for clinical application.\u003c/p\u003e\u003ch2\u003eStatistical analysis\u003c/h2\u003e\u003cp\u003eMeasurement data that follow a normal distribution are expressed as mean ± standard deviation (SD), and independent t-tests are employed for comparisons between groups. For skewed data, the median and interquartile range (IQR) are used for description, with the Mann-Whitney U test applied to evaluate differences between groups. Categorical data are presented as frequencies and percentages, and chi-square tests or Fisher's exact test are utilized for intergroup comparisons. A p-value \u0026lt; 0.05 is considered indicative of statistical significance. All statistical analyses were conducted using R (version 4.4.3) and Python via PyCharm (Professional Edition, version 2024.3).\u003c/p\u003e"},{"header":"Result","content":"\u003cp\u003e\u003cb\u003ePatient characteristics\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe flowchart of the research protocol is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Initially, 292 patients were enrolled in this study following strict inclusion and exclusion criteria. To balance the demographic characteristics, a 1:2 propensity score matching (PSM) strategy was employed, resulting in the final inclusion of 193 patients. Specifically, the MSAP group comprised 87 males and 37 females, while the SAP group included 51 males and 18 females. The mean age of patients was 42 years in both the MSAP and SAP groups. As shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, no significant statistical differences were observed in the demographic characteristics between the two groups after matching.\u003c/p\u003e\u003cp\u003e\u003cb\u003eFeature selection\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eFeature selection for time series data\u003c/b\u003e\u003c/p\u003e\u003cp\u003eIn this study, we constructed a time - series model using a bidirectional LSTM neural network with a dual - branch hybrid architecture. To evaluate the model training process, we monitored the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve and the loss functions for both the training set and the validation set, as depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.For the AUC curves (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA), both the training set (blue line) and validation set (orange line) exhibited a rapid ascent in the initial training iterations, plateauing at an AUC value close to 0.9 \u003csup\u003e+\u003c/sup\u003e interval(Train: 0.64 ~ 0.98, Val: 0.31 ~ 0.95). This indicates the model’s strong capability to distinguish between positive and negative samples, with consistent performance across training and validation data—suggesting no severe overfitting.Regarding the loss curves (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB), the training loss (blue line) and validation loss (orange line) showed a synchronized decline as training epochs increased.In the initial stage (the first 50 rounds), the training loss rapidly decreased from 3.258 to 1.097; the validation loss dropped from 3.220 to 1.032. In the middle stage (rounds 50 to 100), the descent trends of both curves slowed down, with the training loss reducing from 1.0973 to 0.509; the validation loss decreased from 1.032 to 0.580. In the later stage (rounds 100 to 167), both curves approached lower and stable values without significant divergence, with the training curve stabilizing at 0.349 and the validation curve stabilizing at 0.446. Both curves converged to low, stable values without significant divergence, which reflects effective model learning and good generalization ability (i.e., the model did not overfit to the training data, as validation loss did not rise while training loss continued to decrease).To evaluate the model's performance and generalization ability, we employed 10-fold cross-validation. As shown in the Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the results are as follows: AUC mean of 0.9551 ± 0.0158, accuracy mean of 0.9222 ± 0.0300, F1-score mean of 0.8956 ± 0.0364, training loss of 0.2992 ± 0.0328, and validation loss of 0.4132 ± 0.0651.To quantify feature importance in the trained model, we employed SHAP values, visualized in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.This plot ranks features based on the contributions of the statistical characteristics and temporal features of the time-series data to the model's predictions.As illustrated in the figure, Rmax, Pdiff_mean, and Tdiff_std emerge as the most critical features for R, P, and T datasets, respectively.\u003c/p\u003e\u003cp\u003e\u003cb\u003eFeature selection for blood test index data and imaging index data\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe initial dataset contains 16 features. By using the LASSO method with the optimal lambda value of 0.035 obtained through cross - validation, the target variables are reduced to 7 (as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). This process enhances the model’s balance between training and test set performance, yielding superior generalization ability. These 7 variables are: PE, Neu, HCT, Ca, TG, AMY, and CRP.\u003c/p\u003e\u003cp\u003e\u003cb\u003ePerformance of machine learning models\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe conducted an analysis of eight supervised learning classification models, comprising single models (Logistic Regression, Support Vector Machine, Naive Bayes, K-Nearest Neighbors, Decision Tree, Neural Network) and ensemble models (Random Forest, XGBoost).The evaluation metrics encompass Accuracy, ROC-AUC, Sensitivity, Specificity, F1-Score, Precision, MCC, and Cohen’s Kappa, as demonstrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e and quantified in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eBased on Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e and Table \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the Random Forest (RF) model outperforms other models across multiple key metrics: achieving an Accuracy of 0.8793, an F1 - Score of 0.8293, a Matthews Correlation Coefficient (MCC) of 0.7365, and a Cohen’s Kappa of 0.6951. In terms of ROC - AUC, RF scores 0.9588, trailing only the XGBoost model (0.9614), while the Decision Tree (DT) model records the lowest value at 0.8591. For Specificity, RF aligns with both the XGBoost and DT models, achieving a value of 0.9189—outperforming other models in this metric.\u003c/p\u003e\u003cp\u003eAs depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, the RF model exhibits superior performance across multiple evaluation dimensions, with its polygonal area in the radar chart being the largest, which directly reflects its strongest comprehensive predictive capability. Conversely, the Neural Network (NN) model shows subpar performance across all evaluated metrics, which is reflected in the smallest polygonal area and weaker predictive power.\u003c/p\u003e\u003cp\u003eThus, the RF model is selected as the optimal model for subsequent research and analysis.\u003c/p\u003e\u003cp\u003e\u003cb\u003eFeature importance analysis\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo interpret the Random Forest(RF) model, we employed SHAP (SHapley Additive exPlanations) to systematically quantify the importance of features and their impacts on model predictions. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eA (global feature importance) and Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eB (local SHAP effects at the feature level), we conducted a comprehensive analysis of both the hierarchical structure of feature contributions and their respective influence patterns.\u003c/p\u003e\u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA, Rmax exhibits the most significant global influence, as evidenced by the broad distribution of its SHAP values, underscoring its central role as the primary driver in modulating model predictions. Pdiff_mean and HCT follow as key contributors, with their considerable mean SHAP values highlighting their essential roles in the model's decision-making architecture. By contrast, features including TG, CRP, and AMY demonstrate a clustering of SHAP values around zero, indicating their marginal impact on overall predictive outcomes.\u003c/p\u003e\u003cp\u003eFor Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eB, an observable trend is noted for Pdiff_mean, HCT, Tdiff_std, and PE, wherein increasing feature values are accompanied by a shift of SHAP values from negative to positive. This pattern suggests that these features systematically influence predictive outcomes: lower values suppress predictions, whereas higher values promote upward adjustments. Inverse trends are observed for Ca. Although no strict monotonic relationship is evident between Rmax and its SHAP values, extreme Rmax values (both upper and lower thresholds) are associated with pronounced deviations of SHAP values from zero, emphasizing the heightened impact of outliers in Rmax on model predictions.\u003c/p\u003e\u003cp\u003e\u003cb\u003eVisualization analysis of the optimal model\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo translate the random forest model into a clinically interpretable visual tool, we developed a predictive nomogram for estimating the occurrence probability of severe acute pancreatitis (SAP) (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e). It incorporates seven key features, namely Rmax, Pdiff_mean, HCT, Tdiff_std, PE, Neu, and Ca. For each feature, a specific value is mapped to a respective score via the upper score axis. Summing these individual scores yields a cumulative total score, which is subsequently mapped to the predicted SAP probability via the lower probability axis.The Bias-corrected predicted probability fitted the actual event rate (Ideal diagonal \"IDEAL\") well, with a mean absolute error of 0.019 and a mean square error of 0.00061 for both the training and testing sets (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eA-\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eB). The ROC curve in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eC showed that the AUC values of the training set and the test set were slightly different (0.969vs0.964). Finally, the decision curve of the model indicated consistently higher than the None and All lines at \"high risk thresholds between 0.2 and 0.8\", net benefit \u0026gt; 0.2, and cost-effectiveness ratios between 1:4 and 4:1.(as in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eD)\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we constructed a dual-branch Bidirectional Long Short-Term Memory (BiLSTM) model based on time-series features, combined with LASSO feature selection for the optimal model (Random Forest model). The model was interpreted using SHAP values, and a convenient yet accurate nomogram model was subsequently established.During the research process, we innovatively used a BiLSTM network model to screen the most important features from time-series data (T, R, and P) within 24 hours after admission. These features were combined with those from blood test indicators and imaging indicators to construct eight ML models.We found that the comprehensive performance of the RF model outperformed other models, with an Accuracy of 0.8793, ROC-AUC of 0.9588, Sensitivity of 0.8095, Specificity of 0.9189, F1-Score of 0.8293, Precision of 0.85, MCC of 0.7365, and Cohen\u0026rsquo;s Kappa of 0.736.To avoid the black-box effect of ML models, we used the SHAP value interpretation tool to rank the importance of model features.To enable bedside triage of severe acute pancreatitis, we developed a visually intuitive nomogram incorporating the top 7 SHAP-identified features (e.g., Rmax, HCT). The nomogram demonstrated excellent discriminative ability (ROC-AUC\u0026thinsp;=\u0026thinsp;0.964), with decision curve analysis confirming a net benefit\u0026thinsp;\u0026gt;\u0026thinsp;0.2 across high-risk thresholds (0.2\u0026ndash;0.8), outperforming both 'treat all' and 'treat none' strategies.\u003c/p\u003e\u003cp\u003eRecently, an increasing number of scholars have constructed ML models to predict the severity and complications of AP. For example, Callum B et al. used a kernel logistic regression model with 8 variables (age, C-reactive protein, respiratory rate, partial pressure of oxygen in inhaled air, arterial blood pH, serum creatinine, white blood cell count, and Glasgow Coma Scale [GCS]) to predict severe complications, outperforming the traditional APACHE II score (AUC: 0.82 vs 0.74) \u003csup\u003e[17]\u003c/sup\u003e. Hong Wangdong et al. conducted a study on predicting severe pancreatitis based on a classification and regression tree model. They used logistic regression to screen out four variables\u0026mdash;systemic inflammatory response syndrome (SIRS), pleural effusion, serum calcium, and blood urea nitrogen\u0026mdash;to construct a model, which showed superior performance in early identification of high-risk severe pancreatitis compared with APACHE II (AUC: 0.84 vs 0.68) \u003csup\u003e[18]\u003c/sup\u003e. In addition, a multicenter, multinational prospective observational study published by Bal\u0026aacute;zs et al. used binary classifiers (decision tree, random forest, logistic regression, support vector machine, CatBoost, and XGBoost) and identified XGBoost as the optimal model for predicting early SAP (AUC: 0.81\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03). The study also first used SHapley Additive exPlanations (SHAP) to screen the most influential features: respiratory rate, body temperature, abdominal reflex, gender, age, and blood glucose\u003csup\u003e[19]\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eWhile state-of-the-art machine learning models outperform traditional scoring systems in AP severity prediction, the scientific rationale underlying feature extraction remains underdeveloped.Notably,longitudinal vital sign trajectories\u0026mdash;critical for capturing the temporal dynamics of inflammatory cascades\u0026mdash;are typically reduced to static univariate statistics (e.g., mean, interquartile range), neglecting sequential pattern recognition (e.g., autoregressive trends, volatility clustering).\u003c/p\u003e\u003cp\u003eOur study addressed this gap by leveraging the power of BiLSTM networks to capture temporal dependencies in vital sign data. The advantages of the BiLSTM architecture in effectively mining temporal and nonlinear features from data have been confirmed by scholars.Shahid F et al. constructed predictive models including Autoregressive Integrated Moving Average (ARIMA), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (Bi-LSTM) to forecast time-series data of confirmed cases, deaths, and recoveries in ten major countries affected by COVID-19. The Bi-LSTM model demonstrated optimal performance in most scenarios, particularly for predicting China's mortality rate, which achieved the lowest MAE (0.0070) and RMSE (0.0077). Additionally, the Bi-LSTM model showed the highest r2_score (0.9997) for predicting recovered cases in China[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn this study, we evaluated the performance of the BiLSTM model using accuracy, AUC value, F1 score, and cross-entropy loss function. Ten-fold cross-validation showed that: the average accuracy was 0.922; the average AUC value was 0.9551; the average F1 score was 0.8956; the average training set loss function was 0.2992; and the average validation set loss function was 0.4132,indicating that the model maintains strong fitting capability alongside robust generalization performance.Finally, Rmax, Pdiff_mean, Tdiff_std, and other indicators were included to construct a nomogram model, which is similar to the characteristic variables screened by the above prediction model \u003csup\u003e[17\u0026ndash;19]\u003c/sup\u003e.This further demonstrated that the systemic inflammatory response syndrome (manifested as T\u0026thinsp;\u0026gt;\u0026thinsp;38\u0026deg;C or \u0026lt;\u0026thinsp;36\u0026deg;C; R\u0026thinsp;\u0026gt;\u0026thinsp;20 breaths pear minute; P\u0026thinsp;\u0026gt;\u0026thinsp;90 beats pear minute) plays a crucial role in the progression of acute pancreatitis[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].Traditionally, SIRS and CARS (compensatory anti-inflammatory response syndrome)are believed to occur sequentially, with their intensity and temporal order serving as key prognostic factors for AP. During the progression of the disease, regulated cell death (RCD) of pancreatic acinar cells plays a pivotal role in influencing AP severity through the following mechanisms: extrinsic apoptosis (activation of caspase-8 via death receptors like Fas/TNF receptors) and intrinsic apoptosis (mitochondrial membrane permeabilization triggering caspase-9 activation through mitochondrial proteins and HtrA serine peptidases), impaired autophagy compromising cellular homeostasis, caspase-independent necroptosis mediated by receptor-interacting protein kinases (RIP1-RIP3) and mixed lineage kinase domain-like protein (MLKL), pyroptosis activated via caspase-1-dependent canonical and caspase-4/5/11-dependent non-canonical pathways, and ferroptosis induced by iron-mediated oxidative damage.Recently,Sendler M et al. had proposed that interleukin-18 (IL-18), a cytokine secreted by macrophages, activates both SIRS and CARS in parallel[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Among the three features mentioned above, as a feature of importance, Rmax shows results consistent with those of Lei H et al., who found that a median respiratory rate of 22 breaths per minute is associated with lung injury in SAP[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Lung injury is the most common extrapancreatic organ dysfunction in SAP[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e],but the specific injury mechanism remains unclear.Hu Q et al. found that emodin alleviates SAP-associated acute lung injury (ALI) by reducing pancreatic exosome secretion, regulating their protein composition, activating PPARγ to inhibit the NF-κB pathway, thereby suppressing M1 polarization of alveolar macrophages and release of pro-inflammatory factors[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].More importantly, the mortality rate of early SAP complicated with acute lung injury reaches 70% [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], further confirming the clinical value of the Rmax feature in the model.\u003c/p\u003e\u003cp\u003eIn addition to the lung, which is a common extrapancreatic organ injured in SAP, the heart is also similarly affected. Cardiac injury manifests diversely, including coronary vascular dysfunction, autonomic nervous dysfunction, and autophagic dysregulation. Heart rate variability (HRV), which reflects pathological changes in the sinoatrial node of cardiac cycle variations, serves as an important non-invasive quantitative indicator for evaluating cardiac autonomic neuropathy[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].Zhang et al. found that high-frequency norm (nHF) is a good indicator for predicting infected pancreatic necrosis (IPN) and multiple organ dysfunction syndrome (MODS). The areas under the ROC curves of nHF for predicting IPN and MODS were 0.927 and 0.821, respectively, which were superior to those of procalcitonin (AUC\u0026thinsp;=\u0026thinsp;0.709 and 0.722) and APACHE II (AUC\u0026thinsp;=\u0026thinsp;0.785 and 0.899).As the review published by Liau et al., HRV parameters obtained through time-domain analysis or frequency-domain analysis may be used to predict the severity and prognostic outcome of AP[\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].In the model of this study, the mean value of heart rate difference (Pdiff_mean) was included, which reflects the degree of deviation from the baseline value and avoids the influence of the basal value.\u003c/p\u003e\u003cp\u003eHCT, PE, Neu and Ca are another group of important features screened by the RF model.We found that high HCT values were more likely to be associated with SAP, which is consistent with the view of Koutroumpakis et al. They found that hematocrit\u0026thinsp;\u0026ge;\u0026thinsp;44% on admission and an elevated BUN level after 24 hours were the most accurate indicators for predicting persistent organ failure (AUC 0.67 and 0.71, respectively) and pancreatic necrosis (AUC 0.66 and 0.67, respectively), outperforming other laboratory parameters and the Acute Physiology and Chronic Health Evaluation-II score[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e].In SAP, the elevation of HCT values may be associated with increased permeability of vascular endothelial cells, leading to a state of hemoconcentration. To this end, Komara et al. constructed a fluid model of vascular-interstitial-third space to analyze the permeability changes of capillaries to albumin (Alb) and non-albumin plasma proteins (NAPP\u0026thinsp;=\u0026thinsp;TP-Alb). The HCT in the MOF group increased by 5.00% (3.70%-8.70%) compared with the baseline, while that in the non-MOF group changed by only \u0026minus;\u0026thinsp;0.10% (-1.55\u0026ndash;1.40%). An HCT\u0026thinsp;\u0026gt;\u0026thinsp;3% had an OR of 17.7 (p\u0026thinsp;=\u0026thinsp;0.014) for predicting MOF[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].PE is a prevalent complication of AP and serves as a precise predictor of SAP, as evidenced by Zeng et al.'s meta-analysis on AP with PE:the pooled incidence of PE in AP patients was 34% (95% CI: 28%-39%), with pooled incidences of 49% for unilateral PE and 59% for small PE. Notably, the PE incidence in SAP patients (67%) was significantly higher than that in mild acute pancreatitis (MAP, 14%). The meta-analysis further revealed a 3.99-fold increased risk of mortality in AP patients with PE (95% CI: 1.73\u0026ndash;9.2), underscoring PE's prognostic significance in disease severity[\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].Neutrophil recruitment also plays a critical role in the development of AP. For example, depleting neutrophils or inhibiting neutrophil infiltration by targeting specific adhesion molecules such as P-selectin, lymphocyte function antigen-1 (CD11a/CD18), and intercellular adhesion molecule-1 can improve tissue damage in AP[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e].Niu et al. found that inhibiting neutrophil-specific ORAI1 calcium channels reduces pancreatitis-associated acute lung injury, with approximately 40% reduction in alveolar macrophage and neutrophil infiltration[\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e].Lei et al. also found that acetate produced by Parabacteroides can reduce neutrophil infiltration, thereby alleviating heparanase (Hpa)-exacerbated acute pancreatitis (AP)[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].The reduction in serum calcium may be fundamentally attributed to the inhibition of membrane Ca\u0026sup2;⁺-ATPases during the onset of AP, which leads to enhanced calcium influx and intracellular calcium overload[\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e].Petersen et al. published a review article elaborating on the core roles of Ca\u0026sup2;⁺ and ATP in the physiology and pathology of exocrine pancreas. For instance, in AP, excessive Ca\u0026sup2;⁺ signals induced by factors like alcohol, fatty acids, bile acids, or physical pressure lead to mitochondrial dysfunction and reduced ATP production, thereby triggering cell necrosis[\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e].Li et al. analyzed the trajectories of total serum calcium (TSC) within 24 hours using group-based trajectory modeling and found that the in-hospital mortality rate was significantly higher in the \"very low TSC\" group (odds ratio [OR], 7.2; 95% CI, 3.7\u0026ndash;14.0), the \"moderately low TSC\" group (OR, 5.0; 95% CI, 3.8\u0026ndash;6.7), and the \"fluctuating high TSC\" group (OR, 5.6; 95% CI, 1.5\u0026ndash;20.6) compared with the \"stable normal-calcium\" group[\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eConversely, the last three features (CRP,AMY and TG) selected based on the RF model were not included in the construction of the nomogram. CRP is associated with SAP, which is also reflected in our study. However, as demonstrated by Tarj\u0026aacute;n et al., for predicting necrotizing pancreatitis, the area under the curve (AUC) of CRP within 72 hours after admission is 0.69 (confidence interval [CI]: 0.62\u0026ndash;0.76), while the pooled AUC of CRP after 72 hours is 0.88 (CI: 0.75\u0026ndash;1.00)[\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. CRP measured after 72 hours is more clinically valuable, but the CRP data in this study were obtained within 24 hours of admission. Therefore, the CRP feature was excluded.AMY is considered to have no contributory value in the assessment of the severity of AP[\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e].Although some scholars have shown that in AP with hypertriglyceridemia, the triglyceride (TG) levels in severe or moderately severe patients within 24 and 48 hours after symptom onset are significantly higher than those in mild patients (P\u0026thinsp;\u0026lt;\u0026thinsp;0.050), and TG levels are associated with clinical outcomes[\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e].However, the AP included in this study are not limited to hyperlipidemic pancreatitis.This is also a shortcoming of this study, as it does not elaborate on the classification of AP types in detail.\u003c/p\u003e\u003cp\u003eThis study also has several limitations: 1) The sample size is small and the data are from a single center, requiring multicenter large-sample validation of the model's generalization ability and external validation to test the performance of the nomogram; 2) Time-series data collection could record data at a higher frequency to capture more precise temporal features.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe constructed a BiLSTM-RF model based on time-series features to predict SAP, and screened out 7 key features including Rmax and HCT to develop a nomogram. The model innovatively utilizes dynamic features of vital signs to enhance prediction accuracy, providing a visual tool for early assessment of SAP.\u003c/p\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 1\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003edemographic characteristics before and after propensity score matching\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eMSAP\u003c/p\u003e\n \u003cp\u003en\u0026thinsp;=\u0026thinsp;213\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eSAP n\u0026thinsp;=\u0026thinsp;79\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eP\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003evs\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMSAP\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSAP\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eP\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003en\u0026thinsp;=\u0026thinsp;124\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003en\u0026thinsp;=\u0026thinsp;69\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGender\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e167/46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e58/21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" rowspan=\"2\"\u003e\n \u003cp\u003e0.368\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e87/37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e51/18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" rowspan=\"2\"\u003e\n \u003cp\u003e0.580\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(M/F)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eASA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e88/125\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e25/54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" rowspan=\"2\"\u003e\n \u003cp\u003e0.132\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e50/74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e22/47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" rowspan=\"2\"\u003e\n \u003cp\u003e0.245\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(I/II)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSmoking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e77/136\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e36/43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" rowspan=\"2\"\u003e\n \u003cp\u003e0.142\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e49/75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e29/40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" rowspan=\"2\"\u003e\n \u003cp\u003e0.733\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(Y/N)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDrinking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e59/154\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e35/44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" rowspan=\"2\"\u003e\n \u003cp\u003e0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e42/82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003e29/40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" rowspan=\"2\"\u003e\n \u003cp\u003e0.260\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(Y/N)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e44\u0026thinsp;\u0026plusmn;\u0026thinsp;13.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e40\u0026thinsp;\u0026plusmn;\u0026thinsp;14.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.093\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42\u0026thinsp;\u0026plusmn;\u0026thinsp;12.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42\u0026thinsp;\u0026plusmn;\u0026thinsp;14.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.961\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCCI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.16\u0026thinsp;\u0026plusmn;\u0026thinsp;1.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.94\u0026thinsp;\u0026plusmn;\u0026thinsp;1.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.170\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;1.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.01\u0026thinsp;\u0026plusmn;\u0026thinsp;1.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.759\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e26.16\u0026thinsp;\u0026plusmn;\u0026thinsp;3.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e27.81\u0026thinsp;\u0026plusmn;\u0026thinsp;3.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e26.88\u0026thinsp;\u0026plusmn;\u0026thinsp;3.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e27.34\u0026thinsp;\u0026plusmn;\u0026thinsp;3.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.414\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNRS2002\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.21\u0026thinsp;\u0026plusmn;\u0026thinsp;0.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.30\u0026thinsp;\u0026plusmn;\u0026thinsp;0.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.120\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.23\u0026thinsp;\u0026plusmn;\u0026thinsp;0.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.28\u0026thinsp;\u0026plusmn;\u0026thinsp;0.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.445\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ctable id=\"Tab2\" border=\"1\" class=\"fr-table-selection-hover\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 2\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eTable of Performance Metrics for 8 Classification Models\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eROC-AUC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF1 Score\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMCC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCohen\u0026apos;s Kappa\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8621\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9614\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7619\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9189\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8421\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6972\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6951\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8793\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9588\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8095\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9189\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8293\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7365\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.736\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8621\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9228\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8571\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8649\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8182\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7826\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7093\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7074\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7931\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9228\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9048\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7297\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6552\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6099\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.5862\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7931\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9151\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8571\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7568\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6667\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.5915\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.5782\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8103\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8938\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8571\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7838\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.766\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6923\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6194\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6095\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7586\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8662\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8095\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7297\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7083\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6296\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.5196\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.5079\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8621\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8591\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7619\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9189\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8421\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6972\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.6951\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n"},{"header":"Abbreviations","content":"\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eAcute pancreatitis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eMSAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eModerate severe acute pancreatitis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eSAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eSevere acute pancreatitis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eAPCC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eAverage annual percentage change\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eAPACHE II\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eAcute Physiology and Chronic Health Evaluation II\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eBISAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eBedside Index of Severity in Acute Pancreatitis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eSIRS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eBedside Index of Severity in Acute Pancreatitis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eCARS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eCompensatory anti-inflammatory response syndrome\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eMachine learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eRNNs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eRecurrent Neural Networks\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eLSTM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eLong Short-Term Memory\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eBiLSTM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eBidirectional long short-term memory network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eSHAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eSHapley Additive exPlanations\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eASA-class\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eAnesthesiologists classification\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eCCI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eCharlson comorbidity index\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eBody mass index\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eNRS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eNutritional Risk Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eTemperature\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003ePulse\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eRespiratory rate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eWBC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eWhite blood cell\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eCRP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eC-reactive protein\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eNeu\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eNeutrophil\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eRBC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eRed blood cell\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003ePlt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003ePlatelets\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eHCT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eHematocrit\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eALT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eAlanine transaminase\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eAST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eAspartate transaminase\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eAMY\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eAmylase\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eLIP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eLipase\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eTBil\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eTotal bilirubin\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eCa\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eCalcium\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eGlu\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eGlucose\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eAlb\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eAlbumin\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eBUN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eBlood urea nitrogen\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003ePE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003ePleural effusion\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003ePSM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003ePropensity score matching\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eReLU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eRectified linear unit\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eArea under the curve\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eLASSO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eLeast absolute shrinkage and selection operator\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eLogistic regression\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eSupport vector machine\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eNB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eNaive bayes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eRandom forest\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eK-nearest neighbors\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eDT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eDecision tree\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eROC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eReceiver operating characteristic\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eDCA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eDecision curve analysis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eRmax\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eMaximum respiratory rate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003ePdiff_mean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eThe average value of the difference in pulse rates\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eTdiff_std\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eThe standard deviation of the difference in body temperatures\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eHRV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eHeart rate variability\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 148px;\"\u003e\n \u003cp\u003eIPN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 392px;\"\u003e\n \u003cp\u003eInfected pancreatic necrosis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eApproval from the Ethics Committee of Zhejiang Xiaoshan Hospital(NO.Y2024088).\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003eAvailability of data and material\u003c/p\u003e\n\u003cp\u003eYes.\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis work was supported by Zhejiang Health Information Association Research Program,number 2024XHZN-C04.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026apos; contributions\u003c/p\u003e\n\u003cp\u003eY C conceived and designed the study, collected the data, analyzed \u0026nbsp;the data, and drafted the manuscript;\u003c/p\u003e\n\u003cp\u003eZD F: conceived and designed \u0026nbsp; the study, collected the data, analyzed the data\u003c/p\u003e\n\u003cp\u003eYZ L and F J: collected the data, analyzed the data and draw the figures. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTM Z and LH J: conceived and designed the study, and revised the manuscript; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll authors have read and approved the submitted manuscript.\u003c/p\u003e\n\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThank you for the contributions of all authors. YC,ZDF, YZL, and LHJ conceived and designed the study, analyzed the data, wrote the manuscript, and \u0026nbsp;contributed to the methodological designs of this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBoxhoorn L, Voermans RP, Bouwense SA, et al. Acute pancreatitis [J]. Lancet, 2020, 396(10252): 726-34.doi: 10.1016/S0140-6736(20)31310-6.\u003c/li\u003e\n\u003cli\u003eIannuzzi JP, King JA, Leong JH, et al. Global Incidence of Acute Pancreatitis Is Increasing Over Time: A Systematic Review and Meta-Analysis [J]. Gastroenterology, 2022, 162(1): 122-34.doi: 10.1053/j.gastro.2021.09.043.\u003c/li\u003e\n\u003cli\u003ePeery AF, Crockett SD, Murphy CC, et al. Burden and Cost of Gastrointestinal, Liver, and Pancreatic Diseases in the United States: Update 2021 [J]. Gastroenterology, 2022, 162(2): 621-44.doi: 10.1053/j.gastro.2021.10.017.\u003c/li\u003e\n\u003cli\u003eSingh VK, Yadav D, Garg PK. Diagnosis and Management of Chronic Pancreatitis: A Review [J]. JAMA, 2019, 322(24): 2422-34.doi: 10.1001/jama.2019.19411.\u003c/li\u003e\n\u003cli\u003eRanson JH, Rifkind KM, Roses DF, et al. Prognostic signs and the role of operative management in acute pancreatitis [J]. Surg Gynecol Obstet, 1974, 139(1): 69-81.\u003c/li\u003e\n\u003cli\u003eAl-Hadeedi S, Fan ST, Leaper D. APACHE-II score for assessment and monitoring of acute pancreatitis [J]. Lancet, 1989, 2(8665): 738.\u003c/li\u003e\n\u003cli\u003eSingh VK, Wu BU, Bollen TL, et al. A prospective evaluation of the bedside index for severity in acute pancreatitis score in assessing mortality and intermediate markers of severity in acute pancreatitis [J]. Am J Gastroenterol, 2009, 104(4): 966-71.\u003c/li\u003e\n\u003cli\u003eBalthazar EJ, Robinson DL, Megibow AJ, et al. Acute pancreatitis: value of CT in establishing prognosis [J]. Radiology, 1990, 174(2): 331-6.\u003c/li\u003e\n\u003cli\u003eArif A, Jaleel F, Rashid K. Accuracy of BISAP score in prediction of severe acute pancreatitis [J]. Pak J Med Sci, 2019, 35(4): 1008-12.\u003c/li\u003e\n\u003cli\u003eBalthazar EJ. Acute pancreatitis: assessment of severity with clinical and CT evaluation [J]. Radiology, 2002, 223(3): 603-13.\u003c/li\u003e\n\u003cli\u003eDeo RC. Machine Learning in Medicine [J]. Circulation, 2015, 132(20): 1920-30.\u003c/li\u003e\n\u003cli\u003ePantanowitz L, Pearce T, Abukhiran I, et al. Non-Generative Artificial Intelligence (AI) in Medicine: Advancements and Applications in Supervised and Unsupervised Machine Learning [J]. Mod Pathol, 2024, 100680.\u003c/li\u003e\n\u003cli\u003eRashidi HH, Hu B, Pantanowitz J, et al. Statistics of Generative Artificial Intelligence and Nongenerative Predictive Analytics Machine Learning in Medicine [J]. Mod Pathol, 2024, 38(3): 100663.\u003c/li\u003e\n\u003cli\u003eHochreiter S, Schmidhuber J. Long short-term memory. Neural Comput. 1997;9(8):1735-80.\u003c/li\u003e\n\u003cli\u003eGraves A, Schmidhuber J. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Netw. 2005 Jun-Jul;18(5-6):602-10.\u003c/li\u003e\n\u003cli\u003eDeforth M, Heinze G, Held U. The performance of prognostic models depended on the choice of missing value imputation algorithm: a simulation study. J Clin Epidemiol. 2024;176:111539.\u003c/li\u003e\n\u003cli\u003ePearce CB, Gunn SR, Ahmed A, Johnson CD. Machine learning can improve prediction of severity in acute pancreatitis using admission values of APACHE II score and C-reactive protein. Pancreatology. 2006;6(1-2):123-31.\u003c/li\u003e\n\u003cli\u003eHong W, Dong L, Huang Q, et al. Prediction of severe acute pancreatitis using classification and regression tree analysis [J]. Dig Dis Sci, 2011, 56(12): 3664-71.\u003c/li\u003e\n\u003cli\u003eKui B, Pint\u0026eacute;r J, Molontay R, et al. EASY-APP: An artificial intelligence model and application for early and easy prediction of severity in acute pancreatitis [J]. Clin Transl Med, 2022, 12(6): e842.\u003c/li\u003e\n\u003cli\u003eShahid F, Zameer A, Muneeb M. Predictions for COVID-19 with deep learning models of LSTM, GRU and Bi-LSTM. Chaos Solitons Fractals. 2020 Nov;140:110212.\u003c/li\u003e\n\u003cli\u003eBone RC, Balk RA, Cerra FB, et al. Definitions for sepsis and organ failure and guidelines for the use of innovative therapies in sepsis. The ACCP/SCCM Consensus Conference Committee. American College of Chest Physicians/Society of Critical Care Medicine. Chest. 1992;101(6):1644-55.\u003c/li\u003e\n\u003cli\u003eLi H, Wu D, Zhang H, Li P. New insights into regulatory cell death and acute pancreatitis. Heliyon. 2023;9(7):e18036.\u003c/li\u003e\n\u003cli\u003eSendler M, van den Brandt C, Glaubitz J, et al. NLRP3 Inflammasome Regulates Development of Systemic Inflammatory Response and Compensatory Anti-Inflammatory Response Syndromes in Mice With Acute Pancreatitis. Gastroenterology. 2020;158(1).\u003c/li\u003e\n\u003cli\u003eLei H, Minghao W, Xiaonan Y, et al. Acute lung injury in patients with severe acute pancreatitis. Turk J Gastroenterol. 2013;24(6):502-7.\u003c/li\u003e\n\u003cli\u003eGarg PK, Singh VP. Organ Failure Due to Systemic Injury in Acute Pancreatitis. Gastroenterology. 2019;156(7):2008-23.\u003c/li\u003e\n\u003cli\u003eHu Q, Yao J, Wu X, et al. Emodin attenuates severe acute pancreatitis-associated acute lung injury by suppressing pancreatic exosome-mediated alveolar macrophage activation. Acta Pharm Sin B. 2022;12(10):3986-4003.\u003c/li\u003e\n\u003cli\u003eHu Q, Zhang S, Yang Y, et al. Extracellular Vesicle ITGAM and ITGB2 Mediate Severe Acute Pancreatitis-Related Acute Lung Injury. ACS Nano. 2023;17(8):7562-75.\u003c/li\u003e\n\u003cli\u003eLuo Y, Li Z, Ge P, et al. Comprehensive Mechanism, Novel Markers and Multidisciplinary Treatment of Severe Acute Pancreatitis-Associated Cardiac Injury - A Narrative Review. J Inflamm Res. 2021;14:3145-69.\u003c/li\u003e\n\u003cli\u003eLiau MYQ, Liau JYJ, Selvakumar SV, et al. Heart rate variability in acute pancreatitis: a narrative review. Transl Gastroenterol Hepatol. 2024 Aug 21;9:68.\u003c/li\u003e\n\u003cli\u003eKoutroumpakis E, Wu BU, Bakker OJ, et al. Admission Hematocrit and Rise in Blood Urea Nitrogen at 24\u0026thinsp;h Outperform other Laboratory Markers in Predicting Persistent Organ Failure and Pancreatic Necrosis in Acute Pancreatitis: A Post Hoc Analysis of Three Large Prospective Databases. Am J Gastroenterol. 2015 Dec;110(12):1707-16.\u003c/li\u003e\n\u003cli\u003eKomara NL, Paragomi P, Greer PJ, et al. Severe acute pancreatitis: capillary permeability model linking systemic inflammation to multiorgan failure. Am J Physiol Gastrointest Liver Physiol. 2020 Nov 1;319(5):G573-G583.\u003c/li\u003e\n\u003cli\u003eZeng T, An J, Wu Y, et al. Incidence and prognostic role of pleural effusion in patients with acute pancreatitis: a meta-analysis. Ann Med. 2023;55(2):2285909.\u003c/li\u003e\n\u003cli\u003eRenstr\u0026ouml;m E, Luo L, M\u0026ouml;rgelin M, et al. Neutrophil Extracellular Traps Induce Trypsin Activation, Inflammation, and Tissue Damage in Mice With Severe Acute Pancreatitis. Gastroenterology. 2015 Dec;149(7):1920-1931.e8.\u003c/li\u003e\n\u003cli\u003eNiu M, Zhang X, Wu Z, et al. Neutrophil-specific ORAI1 Calcium Channel Inhibition Reduces Pancreatitis-associated Acute Lung Injury. Function (Oxf). 2023 Oct 23;5(1):zqad061.\u003c/li\u003e\n\u003cli\u003eLei Y, Tang L, Liu S, et al. Parabacteroides produces acetate to alleviate heparanase-exacerbated acute pancreatitis through reducing neutrophil infiltration. Microbiome. 2021 May 20;9(1):115.\u003c/li\u003e\n\u003cli\u003eLee PJ, Papachristou GI. New insights into acute pancreatitis. Nat Rev Gastroenterol Hepatol. 2019 Aug;16(8):479-496.\u003c/li\u003e\n\u003cli\u003ePetersen OH, Gerasimenko JV, Gerasimenko OV, et al. The roles of calcium and ATP in the physiology and pathology of the exocrine pancreas. Physiol Rev. 2021;101(4):1691-744.\u003c/li\u003e\n\u003cli\u003eLi CL, Lin XC, Jiang M. Identifying novel acute pancreatitis sub-phenotypes using total serum calcium trajectories. BMC Gastroenterol. 2024 Apr 23;24(1):141.\u003c/li\u003e\n\u003cli\u003eTarj\u0026aacute;n D, Szalai E, Lipp M, et al. Persistently High Procalcitonin and C-Reactive Protein Are Good Predictors of Infection in Acute Necrotizing Pancreatitis: A Systematic Review and Meta-Analysis. Int J Mol Sci. 2024 Jan 20;25(2):1273.\u003c/li\u003e\n\u003cli\u003eYadav D, Agarwal N, Pitchumoni CS. A critical evaluation of laboratory tests in acute pancreatitis. Am J Gastroenterol. 2002 Jun;97(6):1309-18.\u003c/li\u003e\n\u003cli\u003eDong X, Pan S, Zhang D, et al. Hyperlipemia pancreatitis onset time affects the association between elevated serum triglyceride levels and disease severity. Lipids Health Dis. 2022 May 30;21(1):49.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-gastroenterology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmge","sideBox":"Learn more about [BMC Gastroenterology](http://bmcgastroenterol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmge/default.aspx","title":"BMC Gastroenterology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Severe acute pancreatitis, Bidirectional long short-term memory network, Machine learning, Random forest, SHAP, Nomogram","lastPublishedDoi":"10.21203/rs.3.rs-7029031/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7029031/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground\u003c/b\u003e The annual incidence of acute pancreatitis is approximately 30 per 100,000, with 20% progressing to severe acute pancreatitis and a mortality rate of 20%-40%. Traditional scoring models suffer from data lag or insufficient accuracy, while existing machine learning models mostly overlook the dynamic characteristics of vital signs.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMethods\u003c/b\u003e Vital signs, laboratory and imaging indices within 24 hours of admission were collected. First, a bidirectional long short-term memory network model was constructed using time-series data. Then,key indices from laboratory and imaging data were screened by LASSO. Eight machine learning models were constructed and compared. Finally, a predictive nomogram was developed based on the Random Forest model and SHAP values.\u003c/p\u003e\u003cp\u003e\u003cb\u003eResult\u003c/b\u003e After propensity score matching, among 193 patients, there were 124 cases in the MSAP group and 69 cases in the SAP group, with no significant differences in baseline characteristics between the two groups. The BiLSTM model showed an average AUC of 0.9551, accuracy of 0.9222, F1-score of 0.8956, training loss of 0.2992\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0328, and validation loss of 0.4132\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0651 in 10-fold cross-validation. Features including Rmax, Pdiff_mean, and Tdiff_std extracted from time-series data, together with those screened by LASSO (PE, Neu, HCT, Ca, TG, AMY, and CRP), were used to construct 8 ML models. The Random Forest model demonstrated the best comprehensive performance, with an accuracy of 0.8793, ROC-AUC of 0.9588. SHAP value analysis identified key features as Rmax, Pdiff_mean, HCT, Tdiff_std, PE, Neu, and serum calcium. The nomogram constructed based on these features achieved AUC values of 0.969 and 0.964 in the training and test sets, respectively. Decision curve analysis showed that the net benefit exceeded 0.2 at high-risk thresholds (0.2\u0026ndash;0.8), outperforming both the \"treat all\" and \"treat none\" strategies.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusion\u003c/b\u003e The BiLSTM-RF model constructed in this study improves the accuracy of SAP prediction by extracting time-series features of vital signs. The nomogram built based on key features demonstrates good clinical practicability, providing a visual tool for the early assessment of SAP.\u003c/p\u003e","manuscriptTitle":"A Time-Series Feature-Based Nomogram for the Prediction of Severe Acute Pancreatitis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-08 16:18:52","doi":"10.21203/rs.3.rs-7029031/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-27T10:29:00+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-22T21:48:18+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"180587170332839588641944006782516036976","date":"2026-02-17T12:35:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"22766893434956159696177639943449881988","date":"2026-01-28T06:25:43+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-24T09:43:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"209185072969707166365093889133378825139","date":"2025-08-14T11:45:44+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-08-05T05:26:03+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-07-30T12:08:03+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-07-09T10:03:21+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-08T14:30:51+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Gastroenterology","date":"2025-07-08T13:01:10+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-gastroenterology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmge","sideBox":"Learn more about [BMC Gastroenterology](http://bmcgastroenterol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmge/default.aspx","title":"BMC Gastroenterology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"612f1c1b-3d82-41b0-9a24-310e4cc32080","owner":[],"postedDate":"August 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-04-27T16:11:06+00:00","versionOfRecord":{"articleIdentity":"rs-7029031","link":"https://doi.org/10.1186/s12876-026-04856-9","journal":{"identity":"bmc-gastroenterology","isVorOnly":false,"title":"BMC Gastroenterology"},"publishedOn":"2026-04-24 16:00:12","publishedOnDateReadable":"April 24th, 2026"},"versionCreatedAt":"2025-08-08 16:18:52","video":"","vorDoi":"10.1186/s12876-026-04856-9","vorDoiUrl":"https://doi.org/10.1186/s12876-026-04856-9","workflowStages":[]},"version":"v1","identity":"rs-7029031","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7029031","identity":"rs-7029031","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}