Introduction

The global incidence of acute pancreatitis has been steadily increasing, with significant variability across European countries, ranging from 4.6 to 100 per 100,000 population. Multiple studies across Europe have reported an upward trend in incidence, with a median annual increase of 3.4% [1–4]. This trend highlights the growing burden of acute pancreatitis, driven by factors like metabolic diseases, alcohol consumption, and other regional risk factors. Approximately 20% to 30% of patients progress to moderately severe or severe forms of the disease, with mortality rates reaching up to 30% in cases of infected pancreatic necrosis [5–7].

Several prognostic tools have been developed to assess risk in acute pancreatitis. The BISAP score, a five-variable tool assessable within 24 hours, offers high specificity (~90%) but moderate sensitivity (~50–60%) for mortality prediction [8]. Ranson’s criteria, while historically useful, require 48-hour data, which delays risk stratification. Meta-analytic evidence suggests that while Ranson’s criteria offer greater sensitivity, they lack the specificity of BISAP [9]. A major limitation of many existing systems, including APACHE II, is their complexity and reliance on numerous variables, which limits their practical applicability in the modern management of acute pancreatitis [10,11]. Moreover, recent machine learning models, although promising, face challenges in design, validation, and reporting, further underscoring the need for simpler, more clinically feasible stratification tools [12].

In response, we developed a simplified early inpatient risk score based on a limited number of routinely available clinical and laboratory variables assessed within the first 24 hours of hospitalization. By avoiding complex or resource-intensive assessments, the proposed score aims to facilitate practical identification of patients at increased risk of in-hospital mortality during the initial hospitalization period. The objective of this study was to develop and internally validate a points-based risk score for predicting in-hospital mortality in adults with acute pancreatitis, with prospective evaluation required to determine its impact on clinical outcomes or healthcare processes.

Materials and methods

Study population and eligibility criteria

All consecutive adult patients (≥18 years) admitted with a primary diagnosis of acute pancreatitis between January 1, 2020, and December 31, 2024, were screened for eligibility. Acute pancreatitis was diagnosed according to the revised Atlanta criteria, requiring at least two of the following: (1) abdominal pain consistent with acute pancreatitis; (2) serum lipase or amylase levels ≥3 times the upper limit of normal; and (3) imaging findings consistent with acute pancreatitis on contrast-enhanced computed tomography or transabdominal ultrasound [13,14].

Exclusion criteria.

Patients with chronic pancreatitis were excluded during data extraction. Patients were further excluded if they had post–endoscopic retrograde cholangiopancreatography (post-ERCP) pancreatitis or missing data for the primary outcome (in-hospital mortality). Patients with missing data for more than two candidate predictor variables were excluded due to insufficient data completeness. Patients with missing values in two or fewer candidate predictors were retained and included in the analysis using multiple imputation. The patient selection process and exclusions are summarized in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Study flow diagram.

Flow diagram illustrating patient selection and exclusions for the acute pancreatitis cohort admitted between 2020 and 2024. Of 1,059 initially identified patients, 6 with post–endoscopic retrograde cholangiopancreatography pancreatitis and 12 with missing data for more than two candidate predictor variables were excluded. The final analytical cohort comprised 1,041 patients, including 53 in-hospital deaths.

https://doi.org/10.1371/journal.pone.0352980.g001

Outcomes.

The primary outcome was in-hospital mortality, defined as death from any cause occurring during the index hospitalization for acute pancreatitis.

Results

Among the 1,059 patients initially identified, 6 were excluded due to post–endoscopic retrograde cholangiopancreatography pancreatitis, and 12 were excluded because of missing data for more than two predictor variables. The final analytical cohort included 1,041 patients, with 53 in-hospital deaths.

Baseline characteristics of the study population stratified by in-hospital mortality are shown in Table 1. Compared with survivors, patients who died during hospitalization were older and had higher levels of inflammatory and renal biomarkers during the first 24 hours of hospitalization, particularly C-reactive protein, creatinine, and urea. Early multi-organ dysfunction within the first 24 hours was also more frequent among deceased patients. Unadjusted between-group comparisons are reported in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Baseline characteristics of the study population according to in-hospital mortality.

https://doi.org/10.1371/journal.pone.0352980.t001

Continuous variables are presented as median (interquartile range) and were compared between survivors and deceased patients using the Mann–Whitney U test. Categorical variables are presented as number (%) and were compared using Fisher’s exact test. Comparisons were based on observed, non-imputed values; missing values were excluded on a variable-specific basis.

In univariable logistic regression analyses, increasing age, early multi-organ dysfunction within 24 hours, and early inpatient levels of inflammatory and metabolic biomarkers, including C-reactive protein, creatinine, white blood cell count, glucose, and urea, were associated with increased odds of in-hospital mortality (Table 2). Sex, alanine aminotransferase, sodium, and hematocrit were not significantly associated with in-hospital mortality in univariable analyses.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Univariable logistic regression analysis for in-hospital mortality.

https://doi.org/10.1371/journal.pone.0352980.t002

In the multivariable logistic regression model, increasing age, early multi-organ dysfunction within the first 24 hours of admission, and higher early inpatient levels of C-reactive protein and urea were independently associated with increased odds of in-hospital mortality (Table 3). Creatinine was not independently associated with mortality after adjustment for other variables. Accordingly, creatinine was excluded from the final four-predictor continuous prediction model and from the derived simplified points-based score; Table 3 is presented to document the multivariable predictor-selection process. Although white blood cell count and glucose were associated with mortality in univariable analyses, they were not retained in the AIC-guided multivariable selection procedure.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Multivariable logistic regression analysis informing predictor selection for the simplified risk score.

https://doi.org/10.1371/journal.pone.0352980.t003

Risk classification performance

For the underlying continuous prediction model, classification performance was additionally evaluated at a predicted-probability threshold of 0.05. At this threshold, the model achieved a sensitivity of 67.9% and a specificity of 79.55%. The 0.05 threshold was selected for descriptive performance reporting based on the observed baseline in-hospital mortality rate and clinical considerations, rather than through formal statistical optimization; it was not intended to define a clinical decision-making cut-off or the points-based risk-group threshold described below.

Following development of the continuous multivariable prediction model, a simplified points-based clinical score was constructed for bedside use, as described in the Methods (Table 4). Creatinine was not incorporated into the simplified score because it was not independently associated with mortality after multivariable adjustment. C-reactive protein and urea remained continuous in the underlying regression model and were categorized only for construction of the simplified bedside score.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Risk score construction for prediction of in-hospital mortality.

https://doi.org/10.1371/journal.pone.0352980.t004

The observed total risk score range in the study population was 1–15 points. The theoretical maximum score is not fixed, as age contributes incrementally without an upper cap within the model structure.

Model discrimination

The discriminative performance of the final prediction model was evaluated using receiver operating characteristic (ROC) curve analysis. The model achieved good discrimination for in-hospital mortality, with an area under the ROC curve (AUC) of 0.81 (95% CI 0.74–0.88) (Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Receiver operating characteristic (ROC) curve of the final prediction model for in-hospital mortality.

https://doi.org/10.1371/journal.pone.0352980.g002

Illustrating good discriminative performance in the full study cohort (AUC = 0.81, 95% CI 0.74–0.88). The dashed diagonal line represents the line of no discrimination.

Model calibration

Model calibration was assessed using a calibration plot comparing predicted and observed in-hospital mortality probabilities (Fig 3). The model demonstrated good agreement between predicted and observed risks across the range of clinically relevant predicted probabilities.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Calibration plot of the final prediction model for in-hospital mortality.

Predicted and observed in-hospital mortality probabilities across deciles of predicted risk are shown. The solid line represents the bootstrap-corrected calibration curve, and the dashed line indicates ideal calibration. Most observations fall within lower predicted risk ranges, reflecting the low overall event rate.

https://doi.org/10.1371/journal.pone.0352980.g003

The bootstrap optimism-corrected calibration intercept was 0.00 (95% CI −0.73 to 0.73), indicating no evidence of systematic over- or underestimation of risk, and the bootstrap optimism-corrected calibration slope was 1.00 (95% CI 0.75 to 1.25), suggesting adequate agreement between predicted and observed probabilities without clear evidence of substantial overfitting. The optimism-corrected Brier score was 0.0410, indicating good overall predictive accuracy.

Decision curve analysis

Decision curve analysis demonstrated that the model provided a higher net benefit than treat-all and treat-none strategies across clinically relevant threshold probabilities ranging from approximately 2% to at least 15%, supporting potential utility for early inpatient risk stratification (Fig 4).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Decision curve analysis of the prediction model for in-hospital mortality.

Net benefit of the prediction model compared with treat-all and treat-none strategies across a range of clinically relevant threshold probabilities, demonstrating potential utility for early risk stratification.

https://doi.org/10.1371/journal.pone.0352980.g004

Applying the descriptive risk-group classification rule, in-hospital mortality was substantially higher in the high-risk group than in the low-to-intermediate-risk group (8.8% vs 1.8%, respectively; Table 5). The ≥ 6-point threshold was selected based on inspection of the score distribution and corresponding observed mortality across score levels; formal sensitivity–specificity optimization was not performed to avoid overfitting given the limited number of outcome events.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Risk stratification according to the derived mortality risk score.

https://doi.org/10.1371/journal.pone.0352980.t005

Low- and intermediate-risk categories were combined due to the low number of outcome events in the lowest-risk group. High risk was defined as a total score ≥6 points. In-hospital mortality is reported as the proportion of deaths within each risk group.

Discussion

In this retrospective cohort analysis, we developed and internally validated an admission-phase clinical risk score for predicting in-hospital mortality among adults hospitalized with acute pancreatitis. Using routinely available variables obtained within the first 24 hours of hospitalization, the model demonstrated good discrimination, good calibration, and stable performance across multiple internal validation strategies, including bootstrap resampling, complete-case analysis, and temporal validation in a more recent patient cohort. Collectively, these findings suggest that the proposed score may support early inpatient risk stratification within the first 24 hours of hospitalization. Potential clinical applications include supporting early inpatient risk stratification and identifying patients who may warrant closer clinical observation during the initial hospitalization period.

For the underlying continuous prediction model, a predicted-probability threshold of 0.05 was used for descriptive classification-performance reporting, yielding a sensitivity of 67.9% and a specificity of 79.55%; this threshold was separate from the ≥ 6-point threshold used for the simplified bedside score. This threshold reflects a pragmatic balance between sensitivity and specificity for early inpatient risk stratification and was informed by the observed baseline in-hospital mortality rate rather than formal statistical optimization. While alternative thresholds could increase sensitivity at the expense of specificity, the 0.05 threshold was chosen to illustrate model performance and risk stratification characteristics without implying a prescriptive clinical action or decision-making cut-off.

The final model incorporated age, early multi‑organ dysfunction within the first 24 hours of admission, C-reactive protein, and urea levels.

These predictors are biologically plausible and have been associated with worse clinical outcomes and mortality in acute pancreatitis in previous studies; for instance, age and multi‑organ dysfunction have been linked to increased mortality risk [18,19], while CRP has shown good prognostic accuracy in predicting mortality when measured at 24 hours [20], and urea levels have been consistently associated with poor outcomes [18]. Although both creatinine and urea are markers of renal function, urea remained independently associated with in-hospital mortality after multivariable adjustment, which may reflect its ability to capture not only renal impairment but also systemic dehydration, catabolic stress, and neurohormonal activation commonly observed in severe disease.

Early multi-organ dysfunction emerged as the strongest predictor of mortality, underscoring the critical role of early systemic involvement in adverse outcomes [21,22]. Elevated inflammatory markers and renal dysfunction, reflected by CRP and urea levels, likely capture both disease severity and physiological stress, while increasing age has consistently been associated with poorer outcomes in acute pancreatitis [23,24].

The derived risk score achieved good discrimination, with an area under the receiver operating characteristic curve of approximately 0.81 in the primary analysis. Internal validation using bootstrap resampling revealed minimal optimism, indicating limited overfitting. Temporal validation restricted to patients admitted during the later study period further supported the stability of model performance in a contemporary cohort. When applied to the study population, the risk score stratified patients into low-to-intermediate-risk and high-risk categories, with substantially higher observed in-hospital mortality in the high-risk group.

Recent studies have proposed alternative mortality prediction models in acute pancreatitis, often focusing on selected high-risk populations. For example, Chen et al. developed and internally validated a multivariable model for 30-day mortality among patients with severe acute pancreatitis admitted to the intensive care unit, reporting high discriminatory performance [25]. In contrast to that ICU-focused approach, the present study was conducted in a broader, unselected hospital cohort and targeted in-hospital mortality across the full spectrum of disease severity. While Chen et al. identified biochemical markers such as AST, ALP, triglycerides, and creatinine as key predictors, our model emphasizes early multi-organ dysfunction and routinely available inflammatory and renal markers obtained within the first 24 hours of hospitalization, allowing practical early inpatient risk stratification during this period.

Several prognostic tools for acute pancreatitis, including Ranson’s criteria, APACHE II, BISAP, and the revised Atlanta classification, are widely used in clinical practice [10,26]. However, many of these systems either require repeated assessments, extensive physiological and laboratory measurements, imaging-based variables, or reassessment at 24–48 hours, which may limit their applicability for early inpatient severity stratification within the first 24 hours of hospitalization [27].

Recent evidence further supports the growing role of clinically interpretable prediction tools in pancreatic disease. Song et al. developed a multicentre model for early prediction of infected pancreatic necrosis in patients with acute pancreatitis using routinely available variables collected within the first 24 hours of admission, including temperature, respiratory rate, calcium, blood urea nitrogen, and glucose [28]. Although their outcome differed from the present study’s endpoint of in-hospital mortality, both studies emphasize the feasibility of early inpatient risk stratification using readily available clinical information. More broadly, Sherpa et al. highlighted that prediction tools in pancreatic disorders require clear clinical applicability, transparent reporting, and external validation before routine implementation [29]. These considerations support the intended role of the present score as a practical early inpatient stratification tool that requires validation in independent cohorts before clinical use.

The present score was intentionally designed to rely exclusively on variables routinely available within the first 24 hours of hospitalization. Direct head-to-head comparison with established prognostic systems such as BISAP, Ranson’s criteria, or APACHE II was not performed, as these tools require variables collected beyond the early admission phase or repeated assessments that were not uniformly or prospectively available in the present dataset. Retrospective reconstruction of these scores would risk incomplete data capture and biased performance estimates. Rather than performing potentially misleading retrospective comparisons, the present model prioritizes feasibility, parsimony, and early applicability within real-world admission workflows. Accordingly, the primary advantage of the proposed score lies not in superior predictive accuracy, but in its simplicity and suitability for early inpatient risk stratification during the first 24 hours of hospitalization.

Several methodological strengths merit emphasis. This study includes a relatively large cohort of consecutive patients spanning multiple years, enhancing the representativeness of the study population. Missing data were handled using multiple imputation, reducing potential bias and preserving statistical power. In addition, model performance was evaluated using complementary validation approaches, including calibration assessment, bootstrap internal validation, and temporal validation, in accordance with TRIPOD recommendations for prediction model development.

Nevertheless, several limitations should be considered. Importantly, the proposed model is designed for early inpatient risk stratification using variables available within the first 24 hours of hospitalization and should not be interpreted as a point-of-triage prediction tool at emergency department presentation. Early multi-organ dysfunction within the first 24 hours was deliberately included as a predictor to enhance clinical utility during the initial hospitalization period rather than to serve as a strict baseline prognostic marker. This reflects an intentional balance between early physiological assessment and practical risk stratification, acknowledging that MODS may partially capture evolving disease severity. However, sensitivity analysis excluding early MODS demonstrated preserved discriminative performance (AUC 0.79), supporting that the model’s predictive ability is not solely driven by inclusion of this variable.

This was a single-center study, which may limit generalizability to other healthcare settings with different patient populations, pancreatitis etiologies, or management practices. Although internal and temporal validation demonstrated stable model performance, the absence of external validation in independent cohorts represents a key limitation. The original predictor-selection strategy may be susceptible to overfitting given the limited number of mortality events. Specifically, the 53 observed deaths among 11 evaluated candidate predictor parameters corresponded to 4.82 events per candidate predictor parameter, and the precision and stability of individual coefficient estimates should therefore be interpreted cautiously.

However, sensitivity analysis using ridge-penalized regression incorporating all candidate predictors demonstrated only modestly greater optimism-corrected discrimination and did not improve optimism-corrected overall predictive accuracy compared with the existing four-predictor model underlying the simplified score. In addition, categorization of CRP and urea in the simplified points-based score may result in some loss of predictive information compared with the underlying continuous regression model, reflecting a deliberate trade-off in favor of bedside usability. Calibration at higher predicted risk levels should also be interpreted carefully due to the limited number of high-risk patients and outcome events. Consequently, the proposed score should be considered a preliminary, clinically interpretable risk-stratification tool requiring external validation in independent, preferably multicenter cohorts before clinical implementation.

From an implementation perspective, the proposed score is intended to support structured risk reassessment during the first 24 hours of hospitalization, when its component variables are routinely available. Patients classified as higher risk may warrant closer clinical monitoring, earlier senior review, and increased vigilance for deterioration or need for escalation of care. However, the score should not be used as a stand-alone decision rule or guide management changes before external validation and prospective evaluation of its clinical impact.

Finally, the retrospective design precludes assessment of whether implementation of the score improves clinical decision-making or patient outcomes, which will require prospective evaluation. Despite these limitations, the model demonstrated consistent performance across complementary internal validation strategies, supporting its potential utility for early inpatient risk stratification.

Conclusions

In conclusion, this study developed and internally validated a simple, admission-phase clinical risk score for predicting in-hospital mortality in patients with acute pancreatitis. The model demonstrated good discrimination, calibration, and clinical interpretability in internal analyses. External validation in independent, preferably multicenter cohorts is required before broader clinical implementation.

Supporting information

Acknowledgments

The authors utilized ChatGPT (OpenAI) as a language editing assistant to aid in manuscript structuring and enhance clarity. All scientific content, data analysis, and interpretation remain the sole responsibility of the authors.