Abstract
Background
Acute pancreatitis (AP) is a global health issue that can lead to acute kidney injury (AKI), especially in critically ill patients. Timely identification of AP-AKI is vital for early intervention. This study aimed to develop and validate a machine learning (ML) model to predict AP-AKI in ICU patients.
Methods
Data were collected from Shandong Provincial Hospital (training/internal validation) and the MIMIC-IV database (external validation). Thirty-three clinical variables within 24 hours of ICU admission were selected using LASSO regression. Six ML algorithms were tested, with model performance evaluated through discrimination, calibration, and clinical utility. SHAP analysis was used for interpretability, and a web-based interface was developed for clinical use.
Results
Among 317 patients in the internal cohort, 219 (69.1%) developed AKI. A 7-variable Random Forest model showed the best performance with AUCs of 0.970 (training), 0.811 (internal validation), and 0.820 (external validation). Key predictors included neutrophil percentage, platelet-to-neutrophil ratio, creatinine, APTT, lactate, glucose, and vasopressor use. A friendly user interface has been developed for clinician use.
Conclusion
The ML-based model demonstrated strong predictive performance for AP-AKI and good generalizability. SHAP analysis enhanced model transparency, supporting clinical decision-making and early intervention.
1. Introduction
Acute pancreatitis (AP) is a rapid-onset gastrointestinal disorder characterized by pancreatic inflammation, with its incidence steadily rising globally. The incidence of AP has been increasing globally, with up to 25% cases progressing to severe conditions.1,2 AP can lead to various complications, among which acute kidney injury (AKI) is one of the most frequent and clinically significant. The occurrence of AKI in AP patients is associated with increased mortality, prolonged hospital stays, and greater healthcare burden, particularly among those requiring intensive care unit (ICU) admission. 3 Therefore, the early identification of patients at high risk for AKI is crucial for timely intervention and improved clinical outcomes in AP.
AKI is currently diagnosed according to the Kidney Disease: Improving Global Outcomes (KDIGO) clinical practice guidelines, which rely on serum creatinine and urine output. 4 However, serum creatinine is a delayed biomarker of kidney injury and may fail to detect early-stage renal dysfunction, especially in critically ill patients. Additionally, urine output is influenced by multiple factors (e.g., fluid resuscitation, diuretics), making it unreliable in critically ill patients.5,6 Several studies have identified risk factors associated with AP-AKI, but many proposed models are either lack external validation, or require costly and complex biomarkers that limit their routine clinical use.7–11
Machine learning (ML) has emerged as a promising tool in medical research, demonstrating superior predictive performance in various clinical applications, including AP-AKI risk assessment. Despite the advantages of ML, many ML models are criticized for their “black-box” nature, where the decision-making process lacks transparency, hindering clinical interpretation. 12 While several ML-based studies have ranked feature importance for AKI prediction in critically ill adult patients, only a limited number have provided a clear interpretation of their models.13,14 To address this issue, The SHAP method have been increasingly applied to enhance transparency and facilitate clinical decision-making. 15
Therefore, this study aimed to develop and externally validate an interpretable machine learning model for early ICU-based prediction of AKI in critically ill patients with acute pancreatitis. We further used SHAP analysis to improve model transparency and deployed the final model as a web-based calculator to facilitate bedside risk estimation.
2. Method
2.1. Patients and study design
We conducted a retrospective study of acute pancreatitis (AP) patients admitted to the ICU at Shandong Provincial Hospital Affiliated to Shandong First Medical University between January 2014 and June 2024. The inclusion criteria were all patients diagnosed with AP during the study period, while the following exclusion criteria applied: (1) End-stage renal disease, current hemodialysis, or AKI diagnosed before ICU admission; (2) Missing information; (3) patients under 18 years of age; (4) hospital stay of less than 24 hours; and (5) pregnancy. An external validation cohort was obtained from the MIMIC-IV database. The author Li Zhao obtained the license of Collaborative Institutional Training Program (CITI) (license number: 58262711) and the right to use the database according to relevant regulations. The research methodology adheres to the principles declared in the Declaration of Helsinki, with approval of the Ethics Committee of Shandong Provincial Hospital Affiliated to Shandong First Medical University (approval number: 2024-706). Given the retrospective nature of the study, informed consent was waived. Additionally, the study followed the TRIPOD+AI 16 guidelines for reporting prognostic or diagnostic prediction models.
2.2. Definitions
The diagnostic criteria for AP were established according to the revised 2012 Atlanta classification, 17 patients needed to satisfy at least two of the following three criteria: (1) experiencing typical abdominal pain; (2) having serum amylase levels surpassing three times the upper limit of normal; and (3) displaying imaging evidence demonstrating characteristic AP finding. According to the definition of kidney disease: Improving Global Outcomes (KDIGO) guidelines, AKI was defined in one of the following situations: (1) an increase in serum creatinine (SCr) of 0.3 mg/dl (≥26.5µmol/L) within 48 h; (2) known or presumed kidney damage occurring within 7 days, with SCr rising to more than 1.5 times the baseline value. 4 Baseline creatinine was defined as the lowest serum creatinine value in the last 6 months before the onset of AKI, or the lowest value in patients who had not measured and were without dialysis during their hospitalization.
2.3. Outcomes
The primary outcome was new-onset AKI occurring between 24 hours and 7 days after ICU admission in patients with AP.
2.4. Data collection
The data for variables were obtained within 24 hours of admission to the ICU from the patient’s hospitalization electronic medical records (EMRs), including demographic data, medical history, laboratory indicators, vital signs, and treatment. The demographic data included age and sex. The medical history included hypertension, diabetes mellitus, and cardiovascular disease. The laboratory indicators included white blood cell (WBC), neutrophil percentage (NEUT), neutrophil (N), lymphocyte (LYM), platelets (PLT), monocyte (MONO), alanine aminotransferase (ALT), aspartate aminotransferase (AST), albumin (ALB), total bilirubin (TBIL), creatinine (Cr), blood urea nitrogen (BUN), prothrombin time (PT), activated partial thromboplastin time (APTT), international normalized ratio (INR), potential of hydrogen (PH), lactate (Lac), glucose (Glu), calcium (Ca2+). The vital signs included temperature (Temp), respiratory rate (RR), mean arterial blood pressure (MAP), and heart rate (HR). The treatment included mechanical ventilation and vasopressor. The ratio of platelets to neutrophils (PNR) was also included in the study. Additionally, SOFA score reflecting disease severity were incorporated into the assessment. When the data was measured several times a day, the worst one was chosen. External validation data is from MIMIC IV 2.2, using database management software and Structured query language (SQL) to extract clinical data of patients.
The time origin for prediction was defined as ICU admission. To minimize outcome leakage and ensure temporal separation between predictor assessment and outcome ascertainment, we used a 24-hour landmark design. The landmark time point was defined as 24 hours after ICU admission. Candidate predictors were extracted only from the first 24 hours after ICU admission. Patients who had AKI before ICU admission, at ICU admission, or developed AKI during the first 24 hours after ICU admission were excluded from model development and validation. The primary outcome was defined as new-onset AKI occurring after the 24-hour landmark and within 7 days after ICU admission. Therefore, all predictors, including serum creatinine and vasopressor use, were measured before the outcome assessment window. Serum creatinine values measured after the 24-hour landmark were used only for outcome ascertainment and were not included as predictors. Vasopressor use was defined as exposure during the first 24 hours after ICU admission.
2.5. Model development and evaluation
The data were divided into a training set (70%) and an internal validation set (30%). In addition, external validation data was used for testing. To reduce model complexity and the risk of overfitting, LASSO regression with 10-fold cross-validation was first applied in the training cohort for feature selection. The optimal regularization parameter λ was selected according to the one-standard-error criterion, and variables with non-zero coefficients were retained for model development. In this study, six ML algorithms, extreme gradient boosting (XGB), support vector machine (SVM), logistic regression (LR), random forest (RF), K-nearest neighbor (KNN), and decision tree (DT), were used to construct the prediction models. Fivefold cross-validation was used to ensure the stability of the model. The optimal hyperparameters of the model are obtained by using a grid search (Supplementary Table 1). The final tuned models were subsequently evaluated in the independent internal validation cohort and the external validation cohort. This workflow was designed to limit model complexity, reduce optimism in training performance, and assess model generalizability.
Several commonly used evaluation indexes, such as the area under the receiver-operating-characteristic (ROC) curve, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy, were used to evaluate the reliability of these models. Calibration curves were utilized to appraise the accuracy of the model. In addition, decision curve analysis (DCA) were plotted to demonstrate the true clinical utility. In the process of parameter adjustment, the highest area under the curve (AUC) of the receiver operating characteristic (ROC) was selected as the optimal model.
2.6. Model interpretation and deployment
We selected the best-performing algorithm based on its superior performance metrics. The SHAP method is a game theory-based technique that ranks the importance of input features and explains the results of the prediction model, overcoming the “black box” problem. The SHAP method provides local and global explanations by calculating the contribution of each feature to the prediction results, thus increasing the model’s transparency and interpretability.18,19 Namely, the higher SHAP values indicate an increased probability of AP-AKI. The predictors’ contribution to AP-AKI reflects their cumulative impact on overall risk. To facilitate model application in a clinical setting, the final model was deployed as an interactive web-based Shiny app. When the values of the relevant features in the final model are provided, this application returns the probability of AKI in patients with AP.
2.7. Statistical analysis
Continuous variables with normal distribution are presented as mean ± standard deviation and were compared with the t-test. Continuous variables with skewed distributions are presented as medians with interquartile ranges and compared with the Mann–Whitney U test or the Kruskal–Wallis H test. Categorical variables were expressed as percentages (%) and compared with the chi-square test. Missing values were imputed using mean imputation via the glmnet preprocessing function. Mean imputation was selected because the overall proportion of missing data was less than 30% and because this approach allowed preservation of sample size and compatibility with the LASSO-based feature selection workflow. To avoid information leakage, imputation parameters were derived from the training cohort and then applied to the internal and external validation cohorts.
All statistical analyses were performed using R software (version: 4.4.0). A two-tailed P-value< 0.05 was considered statistically significant.
3. Results
3.1. Population characteristics
A total of 317 patients with AKI from Shandong First Medical University affiliated Provincial Hospital served as the training (N = 221, 70%) and internal validation set (N = 96, 30%), and 61 patients from the MIMIC-IV database served as external validation set, as delineated in Figure 1. The comparison of demographic and clinical variables among the training, internal validation, and external validation sets is shown in Table 1. The incidence of AP-AKI in the training, internal validation, and external validation set was 69.2% (153/221), 66.8% (66/96), and 71.7% (43/61), respectively. The flowing chart. Patient demographics and baseline characteristics. MAP: mean arterial blood pressure, HR: heart rate, Temp: temperature, RR: respiratory rate, WBC: white blood cell, NEUT: neutrophil percentage, N: neutrophil, LYM: lymphocyte, Hb: hemoglobin, PLT: platelets, PNR: platelets -to- neutrophil ratio, MONO: monocyte, ALT: alanine aminotransferase, AST: aspartate aminotransferase, ALB: albumin, TBIL: total bilirubin, Cr: creatinine, BUN: blood urea nitrogen, PT: prothrombin time, APTT: active partial thromboplastin time, INR: international normalized ratio, PH: potential of hydrogen, Lac: lactate, Glu: glucose, Ca2+: calcium, SOFA: sequential organ failure assessment.
3.2. Model development and comparison
The predictive variables identified by lasso were as the common elements for developing the prediction model. seven predictive variables were included: NEUT, PNR, Cr, APTT, Lac, Glu, and vasopressor (Figure 2). In the training set, the algorithm with the greatest performance was the RF model with AUROC=0.970 (95% CI 0.952-0.988), accuracy=0.919, sensitivity=0.922, specificity=0.912, PPV=0.959, and NPV=0.838. In the internal validation set, the RF model with AUC=0.811 (95% CI: 0.708–0.914), accuracy=0.781, sensitivity=0.848, and specificity=0.633. In the external validation set, the best-predicting model was the RF which had AUROC=0.820 95%CI (0.701-0.939), sensitivity=0.907, specificity=0.588, PPV=0.847, and NPV=0.714. The ROC curves in internal and external validations are shown in Figure 3, and confusion matrices results of the model in the training, internal validation, and external validation set are shown in Table 2. The calibration curves for the six algorithms are presented in Figure 4, displaying the relative consistency between predictions and actual values. We also conduct DCA for six algorithms in Figure 5, showing good clinical applicability of RF model. Finally, comprehensively comparing the prediction performance of the training set, internal validation set, and the generalization of the external validation set, the RF algorithm showed good overall performance. Penalty chart of predictive factors for acute kidney injury based on LASSO regression analysis. The receiver operating characteristics curves. (a). Training set; (b). Internal validation set; (c). External validation set. LR: logistic regression; DT: decision tree; RF: random forest; SVM: support vector machine; KNN: K-nearest neighbors; XGB: extreme gradient boosting. Performance parameters of the six machine learning prediction models. Abbreviations: LR: logistic regression; DT: decision tree; RF: random forest; SVM: support vector machine; KNN: K-nearest neighbors; XGB: extreme gradient boosting. PPV positive predicted value, NPV negative predicted value, AUC, area under the curve. The calibration curves. (a). The internal validation set; (b). The external validation set. LR: logistic regression; DT: decision tree; RF: random forest; SVM: support vector machine; KNN: K-nearest neighbors; XGB: extreme gradient boosting. The decision curve analyses. (a). Internal validation set; (b). External validation set. LR: logistic regression; DT: decision tree; RF: random forest; SVM: support vector machine; KNN: K-nearest neighbors; XGB: extreme gradient boosting.
We compared the predictive value of RF and SOFA in patients. The RF model showed numerically higher discrimination and clinical application than SOFA in the validation cohorts, as shown in Supplementary Figure 1, Supplementary Figure 2, and Supplementary Figure 3.
3.3. Model interpretability
To better explain the clinical significance of certain features, this study quantified the features importance as SHAP values. As shown in Figure 6, variables were given a ranking based on their contribution to the risk prediction. The SHAP summary plot for the RF model are shown in Figure 7, highlighting the relationship between high and low values of each feature and their corresponding SHAP values. The plots also identify the features that influenced the model predictions the most. The SHAP dependence plots of the RF model are shown in Figure 8, which shows the nonlinear association between the predictors and the risk of AP-AKI. The plots show how changes in a single feature can affect model output. The feature importance rankings in the RF model of the training set. Cr: creatinine; Lac: lactate; PNR: platelets-to-neutrophil ratio; Glu: glucose; APTT: active partial thromboplastin time; NEUT: neutrophil percentage. SHAP summary plot of seven features of the RF in the training set. Cr: creatinine; Lac: lactate; PNR: platelets-to-neutrophil ratio; Glu: glucose; APTT: active partial thromboplastin time; NEUT: neutrophil percentage. SHAP dependence plot of seven features the RF in the training set. Cr: creatinine; Lac: lactate; PNR: platelets-to-neutrophil ratio; Glu: glucose; APTT: active partial thromboplastin time; NEUT: neutrophil percentage.
3.4. Model application
The final prediction model was implemented into the web application via Shiny, upon inputting the actual values of the seven predicted variables required by the model, this application will automatically predict the risk of AP-AKI patients, as depicted in Figure 9. The web application is accessible online at https://zzll.shinyapps.io/AP-AKI/. AP-AKI web app. Cr: creatinine; Lac: lactate; PNR: platelets-to-neutrophil ratio; Glu: glucose; APTT: active partial thromboplastin time; NEUT: neutrophil percentage.
4. Discussion
In this study, we developed and validated a ML-based predictive model for AP-AKI in critically ill patients. By leveraging electronic medical record (EMR) data, the model demonstrated strong predictive performance, validated both internally and externally. Notably, the incorporation of SHAP enhanced model interpretability, offering a transparent view of how individual variables contribute to risk prediction, a critical feature for clinical applicability. The high incidence of AKI in this study likely reflects the severity of the ICU population rather than the general AP population. As our cohort consisted exclusively of critically ill ICU patients, the model is primarily applicable to ICU-based AKI risk stratification and should be generalized to non-ICU AP populations with caution.
Several studies have explored the risk factors associated with AP-AKI, primarily using traditional statistical models.20,21 Wu et al. 21 developed a nomogram for AP-AKI prediction using variables from a public database, achieving an AUC of 0.795 in the training cohort and 0.772 in the validation cohort. Similarly, Jiang et al. 22 constructed an AP-AKI nomogram based on variables selected through LASSO regression, with an AUC of 0.82 in the training set and 0.76 in the validation set. While these models demonstrate good predictive performance, they fail to capture complex, nonlinear interactions among variables and did not undergo external validation, limiting their generalizability. More recent ML-based studies have suggested that data-driven methods may improve prediction by modeling nonlinear relationships. However, existing studies vary in patient population, validation strategy, and interpretability. Some were based on a single database or lacked independent external validation, and the clinical meaning of model outputs was not always clearly explained.21–24 Our study adds to this literature by developing an ICU-based ML model using routinely collected variables, validating it in an independent external cohort, and applying SHAP to enhance interpretability. The web-based interface further improves accessibility for bedside risk estimation. Nevertheless, the incremental value of this model should be confirmed by future prospective studies directly comparing it with established clinical scores and other AP-AKI prediction tools.
Given these limitations, ML algorithms have emerged as promising alternatives, demonstrating superior predictive performance in critical care settings.25,26 Yuan et al. 27 employed a ML-based approach to predict AP progression, achieving an AUC of 0.895 in the internal validation and 0.873 in the external validation. Similarly, a retrospective study demonstrated that the random forest classifier (RFC) model improved AKI prediction in AP patients. 14 However, these studies lack of interpretability, which hinders their integration into clinical workflows. To address this gap, we employed an RF model, which handling high-dimensional data and capturing intricate interactions among variables. The model achieved an AUC of 0.970 in the training set, 0.811 in the internal validation set, and 0.820 in the external validation set, highlighting its robust predictive capability across diverse populations. More importantly, we utilized Shapley Additive Explanations (SHAP) analysis to elucidate feature contributions, enhancing model interpretability and facilitating informed clinical decision-making.
It should be noted that the variables highlighted by SHAP analysis may reflect both AKI risk and the overall severity of acute pancreatitis-related critical illness. In ICU patients with AP, AKI is commonly driven by systemic inflammation, hemodynamic instability, coagulation abnormalities, metabolic stress, and tissue hypoperfusion. Therefore, the final model is unlikely to capture kidney-specific injury signals alone. Instead, it may identify a broader high-risk phenotype characterized by severe pancreatitis and critical illness burden, in which AKI is more likely to develop. This interpretation is clinically relevant because early recognition of this phenotype may prompt closer renal monitoring, optimization of hemodynamics, avoidance of nephrotoxic exposures, and timely preventive strategies.
Our model included seven common clinical risk factors including NEUT, PNR, Cr, APTT, Lac, Glu, and vasopressor. Consistent with current diagnostic criteria, creatinine was also included in the model. It is worth noting that in order to further improve the accuracy of diagnosis, other indicators were integrated into our model. The neutrophil percentage, a well-established inflammatory marker, has been linked to disease severity. A prior study reported that a model incorporating neutrophil percentage achieved an AUC of 0.903 for predicting mortality in severe COVID-19 patients. 28 In this study, we also found that neutrophil percentage is positive in the prediction of AP-AKI. Coagulation abnormalities are frequently observed in AP and contribute to its complications. APTT serves as a critical marker of coagulation status, with prior studies indicating its role in adverse AP outcomes. 29 Notably, our analysis revealed a non-linear relationship between APTT and AKI risk, suggesting that extreme values—rather than a simple linear increase—may drive adverse renal outcomes. PNR, a composite marker integrating inflammation and coagulation, has recently emerged as a prognostic indicator in various critical illnesses.30–32 Moreover, recent study have found that the interaction between neutrophils and platelets plays an important role in the exacerbation mechanism of acute pancreatitis.33,34 A lower PNR may reflect excessive neutrophil activation coupled with platelet consumption or dysfunction, correlating with a heightened inflammatory burden and impaired microcirculation. Prior studies have shown that interactions between neutrophils and platelets, especially through the formation of neutrophil extracellular traps (NETs), contribute to organ damage in acute pancreatitis, including renal injury.35,36 We used the ratio of platelets to neutrophils (PNR) and found that PNR < 60 had a negative effect on AKI. Metabolic disturbances, including hyperlactatemia and stress hyperglycemia, are hallmarks of critical illness. Zhao et al. 37 identified lactate as a key predictor of AP mortality in ICU patients, aligning with our findings that elevated lactate levels were associated with increased AKI risk. Similarly, hyperglycemia has been implicated in oxidative stress, endothelial dysfunction, and coagulopathy, all of which contribute to AKI.38,39 Our study identified an “N”-shaped relationship between blood glucose levels and AKI risk, suggesting that both hypo- and hyperglycemia may be detrimental. The need for vasopressors reflects hemodynamic instability, a major risk factor for AKI. Our previous research has demonstrated that vasopressor use is linked to septic AKI. 40 Mechanistically, vasopressors may exacerbate renal medullary hypoxia, thereby predisposing patients to ischemic injury. 41 Collectively, the inclusion of these clinically accessible variables is supported by biological plausibility and aligns with known mechanisms of inflammation, coagulation, and metabolic disturbance in AP-AKI. Their interpretability enhances the clinical credibility and potential utility of the proposed model.
A direct comparison with established clinical scores is important for evaluating the added value of ML-based prediction models. In this study, we compared the final random forest model with SOFA score. Although SOFA is widely used to assess organ dysfunction and overall illness severity in ICU patients, it was not specifically designed for AP-AKI prediction and may not capture nonlinear relationships among inflammation, coagulation, metabolic stress, and hemodynamic instability. Beyond numerical improvement in discrimination compared with SOFA, the potential clinical value of the model lies in its ability to support earlier and more individualized risk stratification. A high predicted risk should not be interpreted as a diagnosis of AKI, but may prompt closer monitoring of urine output and serum creatinine, reassessment of volume status and hemodynamics, avoidance of nephrotoxic medications when feasible, optimization of fluid and vasopressor strategies, and earlier nephrology consultation for patients with persistent high-risk features. In this sense, the model may complement SOFA by providing a more focused estimate of AP-AKI risk. However, whether model-guided monitoring or intervention improves clinical outcomes requires prospective impact evaluation.
This study has several notable strengths. First, it addresses a key gap in AP-AKI prediction by leveraging machine learning (ML). Unlike traditional statistical models, our ML-based approach captures nonlinear relationships, enabling earlier and more accurate risk stratification for timely intervention. Second, the use of SHAP enhances model transparency, addressing the common “black-box” criticism of ML algorithms. Third, the variables incorporated into the model are routinely collected in ICU settings, ensuring the model’s feasibility. Compared to conventional AP severity scores (e.g., SOFA, BISAP), it integrates a broader range of clinical and laboratory data for a more individualized risk assessment. Finally, the dual-validation approach enhances the model’s generalizability, supporting its potential application across different healthcare institutions.
From a clinical perspective, this predictive model offers a valuable tool for early AKI risk stratification in AP patients. Patients classified as high risk by the model should not be diagnosed with AKI solely based on the model output. Instead, the prediction may prompt closer monitoring of urine output and serum creatinine, reassessment of hemodynamic status, avoidance of nephrotoxic drugs when possible, optimization of fluid and vasopressor therapy, and early nephrology consultation when clinically indicated. Early identification of high-risk individuals could facilitate timely interventions, such as fluid resuscitation and nephroprotective strategies, ultimately improving patient outcomes. Furthermore, ML-driven risk assessment could complement traditional clinical decision-making, fostering a more personalized approach to critical care.
We acknowledge several limitations of this study. First, it was conducted in a single-center cohort, which may limit its generalizability to other healthcare settings. Second, although the model incorporates routinely measured clinical variables, differences in data collection protocols across institutions could impact its performance. Third, Missing data were handled using mean imputation, which is simple and reproducible but may underestimate variability and does not fully account for uncertainty related to missingness. Although the proportion of missing data was less than 30%, future prospective studies should use predefined missing-data strategies and consider multiple imputation to improve robustness. Forth, although we have developed a web-based calculator based on RF algorithm (https://zzll.shinyapps.io/AP-AKI/), we used LASSO-based feature selection, cross validation, and hyperparameter tuning to reduce model complexity, and potential overfitting remains an issue, further prospective validation is necessary. Incorporating more clinical data and leveraging real-world feedback will be crucial for continuously refining and optimizing the model for practical application.
5. Conclusion
In conclusion, we developed and validated an ML-based predictive model for AP-AKI, demonstrating strong performance and clinical interpretability. By incorporating routinely collected clinical parameters and leveraging SHAP analysis, the model offers a practical and transparent approach to the risk of AKI occurrence. Future research should focus on multi-center validation and prospective implementation to optimize its clinical utility.
Supplemental material
Supplemental material - Interpretable machine learning for early prediction of acute kidney injury in critically Ill patients with acute pancreatitis
Supplemental material for Interpretable machine learning for early prediction of acute kidney injury in critically Ill patients with acute pancreatitis by Li Zhao, Lei Tian, Shenglin Zhou, Tuo Zhang, Zeyu Yang, Qiuxia Liu, Wei Fang, Jicheng Zhang, and Man Chen in Digital Health.
Supplemental material
Supplemental material - Interpretable machine learning for early prediction of acute kidney injury in critically Ill patients with acute pancreatitis
Supplemental material for Interpretable machine learning for early prediction of acute kidney injury in critically Ill patients with acute pancreatitis by Li Zhao, Lei Tian, Shenglin Zhou, Tuo Zhang, Zeyu Yang, Qiuxia Liu, Wei Fang, Jicheng Zhang, and Man Chen in Digital Health.
Footnotes
Ethical considerations
This study received approval from the Human Research Ethics Committee of the Provincial Hospital affiliated with Shandong First Medical University (No. 2024-706), and informed consent was waived because of the study design.
Author contributions
LZ conceived and designed the study, collected data, performed data analysis, and drafted the manuscript. LT, SZ and ZY contributed to data collection. TZ and QL provided technical support for data analysis. WF, JZ and MC participated in study design. MC also contributed to manuscript revision. All authors read and approved the final manuscript.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Natural Science Foundation of Shandong Province (Grant No. ZR2024MH008 and ZR2023MH379); Wu Jieping Medical Foundation Special Fund for Clinical Research (Grant No. 320.6750.2024-2-24); and Shandong Provincial Natural Science Foundation Youth Project (Grant No.ZR2021QH344).
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
The data analyzed and the codes used during the current study are available from the corresponding author upon reasonable request.
AI disclosure statement
No artificial intelligence (AI) tools were used in the preparation, writing, or analysis of this manuscript. All work was performed independently by the authors.
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
