Imaging-independent model for predicting choledocholithiasis in acute biliary pancreatitis

Abstract

BACKGROUND

Predicting choledocholithiasis (CDL) in acute biliary pancreatitis remains challenging, and unnecessary endoscopic retrograde cholangiopancreatographies (ERCPs) should be avoided due to their risks. Current stratification tools for indicating ERCP include imaging findings, potentially leading to diagnostic variability across imaging modalities and between observers.

AIM

To stratify patients independent of imaging we developed a machine learning–based model using easily available parameters.

METHODS

In this observational study we retrospectively reviewed the clinical data of 601 patients coded with ICD-10 for biliary pancreatitis. After exclusions, we analyzed 348 patients with acute biliary pancreatitis at LMU University Hospital, between 2005 and 2021. Using 14 features-excluding common bile duct (CBD) width-a random forest model was trained and evaluated for sensitivity, specificity, predictive values, and area under the receiver operating characteristic (AUC-ROC) curve.

RESULTS

Among 348 patients, 167 (48.0%) had confirmed CBD stones. The model achieved a sensitivity of 92.3%, a negative predictive value (NPV) of 78.6%, and an AUC-ROC of 0.739. Compared to guideline-recommended predictors such as CBD width ≥ 8 mm (sensitivity 79.5%, NPV 68.0%) or bilirubin ≥ 4 mg/dL (sensitivity 30.8%, NPV 47.1%), our model showed superior performance and missed significantly fewer intraductal concrements [4.3% (95% confidence interval (CI): 1.5%-11.9%] vs 11.4% (95%CI: 5.9%-21.0%) and 38.6% (95%CI: 28.1%-50.3%).

CONCLUSION

This model for predicting CDL in acute pancreatitis without relying on CBD width may be a valuable tool to reduce ERCPs, especially in centers without access to daily endoscopic ultrasound.

Key Words: Acute pancreatitis; Biliary pancreatitis; Gallstone pancreatitis; Choledocholithiasis; Machine learning

Core Tip: We developed an imaging-independent prediction model for choledocholithiasis in acute biliary pancreatitis using only routinely available admission clinical and laboratory variables. Unlike existing approaches, the model intentionally excluded common bile duct diameter to avoid observer-dependent imaging variability. It achieved high sensitivity and a low false rule-out rate, supporting its potential role as an early triage tool. Although not suitable as a stand-alone indication for endoscopic retrograde cholangiopancreatography, it may help prioritize further diagnostic evaluation, especially where endoscopic ultrasound is not readily available.

INTRODUCTION

Acute pancreatitis is a potentially life-threatening condition, with gallstones identified as the underlying cause in 30%-50% of cases[1]. The prevalence of gallstone-related diseases is increasing. In the United States, it has doubled over the past three decades and the worldwide incidence of pancreatitis is rising-highlighting the growing burden on healthcare systems worldwide[2-4]. Mechanistically, the migration of gallstones from the gallbladder into the common bile duct (CBD) can lead to temporary or persistent obstruction of pancreatic outflow, triggering acute biliary pancreatitis[5,6]. Recent prospective multicentre trials have shown that, in patients with biliary pancreatitis and a predicted severe course (but without cholangitis), early endoscopic sphincterotomy and stone extraction within 24 hours-regardless of whether preceded by endoscopic ultrasound (EUS)-does not improve outcomes in terms of persistent organ failure or mortality[7,8]. Despite this, the detection and removal of CBD stones, microlithiasis, and sludge remains a cornerstone of therapy in biliary pancreatitis management[9].

Due to its broad availability and non-invasiveness ultrasound (US) remains the first-line imaging modality to assess gallbladder and CBD. However, its diagnostic accuracy is highly dependent on the observing clinician and can be limited by factors such as bowel gas or patient discomfort in acute symptomatic pancreatitis. Reported sensitivities for CBD stone detection via US vary widely. They are often based on cohorts with suspected choledocholithiasis (CDL) rather than patients specifically with biliary pancreatitis[1,10,11].

To reduce the risk of unnecessary endoscopic retrograde cholangiopancreatographies (ERCPs)-procedures associated with significant complication rates-while maintaining a high diagnostic yield for CDL, precise patient selection prior to EUS or ERCP is essential[7,8,12].

This study aimed to develop an imaging-independent predictive model for CDL using clinical and laboratory parameters, which are typically available during the early phase of hospitalization for acute biliary pancreatitis. Current stratification tools for assessing the probability of CDL and the subsequent indication for ERCP include CBD width as a parameter. To reduce observer-dependent and modality-dependent diagnostic variability our model intentionally excludes the CBD diameter[13-15]. Crucially, the model is designed to support risk stratification and reduce unnecessary invasive procedures, particularly in hospitals where EUS is not routinely available.

MATERIALS AND METHODS

The statistical analysis and methods of this study were performed and reviewed by Hans Stubbe from the LMU University Hospital in Munich, Germany.

Study population

We reviewed the clinical data of 601 patients coded with ICD-10 (K85.10, K85.11) for biliary pancreatitis between 2005 and 2021 at LMU University Hospital in Munich, Germany. Of these, 199 patients were excluded due to duplicate records, incomplete data sets, or unconfirmed diagnoses of acute pancreatitis. An additional 43 patients were excluded because their pancreatitis etiology was non-biliary, including post-ERCP, tumor-related, or other causes. Eleven patients presenting with stent occlusion or other anatomical abnormalities were also excluded. Patients with signs of acute cholangitis and indication for emergency ERCP were not included in the analytic cohort. The diagnosis of acute biliary pancreatitis was established based on the presence of gallstones, microlithiasis, or sludge within the gallbladder or CBD, CBD dilation, or alanine transaminase (ALT) levels exceeding twice the upper limit of normal. CBD dilation was defined as a diameter greater than 8 mm in patients aged 75 years or younger, and greater than 10 mm in patients older than 75, consistent with thresholds used in the APEC trial[7,8]. After applying these criteria, 348 patients with acute biliary pancreatitis were identified, of whom 167 had confirmed CBD stones (Figure 1 and Table 1). The study was conducted in accordance with the updated STARD guideline of 2015[16]. The study was reviewed and approved by the ethics committee of the LMU under reference numbers 22-0084/22-0073. Due to the retrospective nature of the study, the LMU ethics committee waived the requirement for obtaining informed consent.

Figure 1 Flowchart of patient selection. Screening of 601 patients with ICD-10 codes for biliary pancreatitis at LMU University Hospital (2005-2021), application of exclusion criteria, and derivation of the final cohort of 348 patients with acute biliary pancreatitis. ERCP: Endoscopic retrograde cholangiopancreatographies; CDL: Choledocholithiasis.

Table 1 Baseline characteristics, n (%).

	No CBD stone	CBD stone	P value
Patients	181 (100)	167 (100)
Sex			> 0.9
Female	92 (51)	86 (51)
Male	89 (49)	81 (49)
Age, median (IQR: Q1, Q3)	62 (45, 74)	68 (54, 80)	< 0.001
BMI (kg/m2), median (IQR: Q1, Q3)	27.7 (24.5, 32.2)	26.3 (24.0, 30.6)	0.13
Unknown, n	40	28
Concrement localization			< 0.001
CBD	0 (0)	74 (44)
CBD + GB	0 (0)	93 (56)
CBD width (mm), median (IQR: Q1, Q3)	6.0 (5.0, 8.0)	10.0 (8.2, 12.0)	< 0.001
Unknown, n	27	8
Laboratory values at admission
Bilirubin (mg/dL), median (IQR: Q1, Q3)	1.60 (0.90, 3.45)	2.60 (1.40, 4.90)	0.001
Bilirubin unknown	1	1
AP (U/L), median (IQR: Q1, Q3)	134 (91, 203)	190 (117, 290)	< 0.001
AP unknown	3	4
GGT (U/L), median (IQR: Q1, Q3)	283 (119, 500)	373 (196, 646)	< 0.001
GGT unknown	1	4
ICU stay	9 (5.0)	22 (13)	0.007
Severity (RAC)			0.5
Mild	96 (55)	80 (48)
Moderate	62 (35)	65 (39)
Severe	18 (10)	20 (12)
Unknown	5	2
Comorbidities
None	70 (38.7)	41 (24.6)
Cholecystectomy	1 (0.6)	22 (13.2)	< 0.001
Gastritis/reflux	17 (9.4)	15 (9.0)
Liver disease	10 (5.5)	15 (9.0)
Cardiac disease	47 (26.0)	44 (26.3)
Carcinoma in history/current	20 (11.0)	26 (15.6)
Upper GI tract surgery	7 (3.9)	5 (3.0)
Dyslipidemia	31 (17.1)	22 (13.2)
Diabetes mellitus II	26 (14.4)	28 (16.7)
Arterial hypertension	76 (42.0)	91 (54.5)
Medication on admission
None	76 (42.0)	46 (27.5)
PPI	38 (21.9)	49 (29.3)
Antihypertensive drugs	72 (40.0)	91 (54.5)
Lipid lowering drugs	32 (17.7)	37 (22.2)
Diuretics	20 (11.0)	30 (18.0)
Analgesics	5 (2.8)	5 (3.0)
Pancreatic enzymes	0 (0.0)	1 (0.6)
UDCA	1 (0.6)	2 (1.2)
Immunosuppressant drugs	8 (4.4)	11 (6.6)
Unknown	9 (5.0)	8 (4.8)

P values were determined using Wilcoxon rank sum test for numerical features and Pearson’s χ² test for categorical features. Continuous variables as median (IQR: Q1, Q3), categorical as n (%). P values are descriptive and were not used for predictor selection. Comorbidity and medication categories are not mutually exclusive. CBD: Common bile duct; BMI: Body mass index; PPI: Proton-pump inhibitors; UDCA: Ursodeoxycholic acid; GI: Gastrointestinal; GB: Gall bladder; IQR: Interquartile range; AP: Alkalic phosphatase; GGT: Gamma-glutamyl transferase; ICU: Intensive care unit; RAC: Revised Atlanta Classification.

Study variables

Laboratory parameters were recorded at admission, on days 3 and 7, and at discharge. Clinical data-including sex, age, height, weight, and body mass index-were also incorporated into the analysis. Additionally, patients’ medical histories, current medications, and nicotine and alcohol use were documented. Data on clinical interventions, surgical procedures, stays in intermediate and intensive care units, readmissions, and analgesic requirements were further included in the dataset. Importantly, this study excluded EUS- or US-guided measurements of the CBD to avoid observer-dependent variability and diagnostic imprecision inherent in imaging-based assessments. The presence of intraductal concrements (IDCs) was confirmed through ERCP, magnetic resonance cholangiopancreatography (MRCP), EUS, or US.

Random forest model and statistical analysis

To predict the presence of IDCs, we used a Gradient Boosted Trees Model based on TensorFlow Decision Forests (TFDF-GB; version 1.9.1) with Python (version 3.11.4). To ensure practicality, the model incorporated clinical and laboratory parameters routinely collected to diagnose pancreatitis, selecting only those readily available in standard clinical settings. The dataset was divided into training and testing portions (80% and 20%). Seventy data sets were used as a training data set. To address class imbalance, we employed the Synthetic Minority Over-sampling Technique from imblearn package (version 0.12.2) on the training dataset.

Our TFDF-GB model was configured with 300 trees and evaluated through 10-fold cross-validation. The final performance assessment was conducted on the separate testing dataset using established metrics, including the receiver operating characteristic (ROC) curve and area under the ROC curve, with calculations performed in R (version 4.3.2). Predictors were not selected with a univariate P value screen. Instead, the notebook pre-specified routinely available admission demographics and laboratory variables and excluded variables with substantial missingness before model fitting. The exported model specification table retains CDL as the target column, with the remaining entries corresponding to the final set of admission predictors (Supplementary Table 1). The event-per-variable ratio and the single hold-out validation (Supplementary Table 2) should be interpreted conservatively.

The model was internally validated with a random 80/20 split. Additionally, 10-fold × 10-fold cross-validation, nested cross-validation, out-of-bootstrap resampling, and a temporal development (2005-2016 vs 2017-2021) validation design were used. Imputation, oversampling and model fitting were repeated within each training set (Supplementary Table 3). The threshold table facilitates clinical risk stratification. In the current implementation, the notebook used a probability cut-off of 0.25, which prioritizes sensitivity and negative predictive value (NPV) at the expense of specificity. The decision threshold was defined by a prespecified clinical objective-the most specific probability threshold achieving at least 90% sensitivity. It was derived by cross-validation within the training set only, locked, and evaluated once in the test set. This selected a cut-off of 0.25. Clinically, this pattern is rather consistent with a triage or rule-out tool than with a definitive rule-in test for immediate ERCP (Supplementary Tables 4 and 5). Calibration describes the agreement between predicted probabilities and observed event rates: A well-calibrated model predicts, for example, 30% CDL risk in patients who actually have CDL 30% of the time. Calibration is distinct from discrimination [area under the curve (AUC)/sensitivity/specificity], which only assesses ranking of patients. A model can discriminate well but still be poorly calibrated. Calibration assessment is applicable only to the TFDF random forest (RF) model, because it outputs continuous predicted probabilities (pred_prob). The binary rule-based comparators (CBD ≥ 8 mm; CBD + bilirubin) are deterministic decision rules that classify patients as positive or negative without an underlying probability estimate. Followingly, conventional calibration metrics are not meaningful for them. All calibration metrics are computed on the held-out test set (test_df) to avoid optimistic in-sample estimates. The key calibration statistics-Brier score, calibration intercept, and calibration slope-are already included in the primary model performance table (Supplementary Table 4). The calibration plot (Supplementary Figure 1) provides the complementary visual assessment. At the selected cut-off, the model is intentionally sensitivity-oriented. The trade-off is a comparatively low specificity, meaning that false positives would be expected if the model were used to directly trigger invasive procedures as ERCP. In practice, the more defensible use case is therefore as a triage or rule-out tool to identify patients who may benefit from intensified diagnostic work-up, while avoiding sole reliance on the model for ERCP referral (Supplementary Figure 2). Pairwise correlations among the numeric admission predictors were evaluated using Spearman’s rho to document collinearity within the admission candidate predictor set. This complements the clinical pre-specification and missingness-based feature filtering reported above. The strongest pairwise absolute Spearman correlation among the numeric admission predictors was 0.811. These results document pairwise collinearity among the clinically pre-specified candidate predictors (Supplementary Table 6).

Consistent with ASGE guidelines, we compared our predictive model to the criteria used to define a high-risk subgroup for CDL. According to these guidelines, high risk is characterized by the presence of a CDL on imaging (US or EUS), ascending cholangitis, or the combination of a dilated CBD (≥ 6 mm, or ≥ 8 mm in patients post-cholecystectomy) with an elevated bilirubin level (≥ 4 mg/dL)[17]. We chose to apply the higher CBD diameter cut-off (≥ 8 mm), acknowledging that CBD dilatation is common in post-cholecystectomy patients and tends to increase with age[18]. Patients with signs of acute cholangitis and the need for an emergency ERCP were excluded from the overall cohort and were therefore not part of the guideline-comparison analysis.

RESULTS

Baseline characteristics

Among the 348 patients included in the study, 167 (48.0%) were diagnosed with CDL, whereas 181 (52.0%) had no detectable gallstones, microlithiasis, or biliary sludge within the CBD. In the CDL group, 86 patients (51%) were female, with a median age of 68 years; in the non-CDL group, 92 patients (51%) were female, and the median age was 62 years (P < 0.001). Within the CDL group, stone localization was categorized as CBD only in 74 patients (44%) and both CBD and gallbladder in 93 patients (56%). The median CBD diameter was significantly greater in patients with stones compared to those without (10.0 mm vs 6.0 mm; P < 0.001). Markers of cholestasis were also significantly elevated in the CDL group, with median alkaline phosphatase (AP) levels of 190 U/L vs 134 U/L (P < 0.001), gamma-glutamyl transferase (GGT) levels of 373 U/L vs 283 U/L (P < 0.001), and bilirubin levels of 2.6 mg/dL vs 1.6 mg/dL (P = 0.001). Patients with CDL had to be treated in the intensive care unit significantly more often [22/167 (13%) vs 9/181 (5%), P = 0.007], although there was no significant difference in the severity of pancreatitis (defined by the Revised Atlanta Classification) between the groups (P = 0.5).

Performance of the novel imaging-independent model

Our RF model identified 14 key variables predictive of IDCs. Age emerged as the strongest predictor, followed by GGT and AP. Additional important predictors included C-reactive protein, bilirubin, weight, aspartate aminotransferase, ALT, triglycerides, height, hematocrit, creatinine, nicotine use, and sex (Figure 2). Using these features, the RF model predicted the presence of IDC with an overall accuracy of 67.1%, demonstrating a high sensitivity of 92.3% but a lower specificity of 35.5%. The positive predictive value (PPV) was 64.3%, while the NPV reached 78.6%. When compared to the predictive performance of CBD width alone, using a threshold of ≥ 8 mm which achieved an accuracy of 68.6%, our RF model showed comparable accuracy (67.1%) but substantially improved sensitivity (92.3% vs 79.5%) and NPV (78.6% vs 68.0%) (Figure 3A and Table 2).

Figure 2 Variable importance. The importance of each variable included in the random forest model is estimated by the inverted mean minimal depth of each variable in the decision tree. GGT: Gamma-glutamyl transferase; AP: Alkaline phosphatase; CRP: C-reactive protein; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase.

Figure 3 Comparison of decision rules. A: Mosaic diagram in which each field’s size corresponds to the number of correct and false rule-in and rule-out classifications (i.e., the four cells of the respective contingency tables). False rule-outs in red signify missed CDL, and correct rule-outs in yellow correspond to true negatives for CDL; B: Two receiver operating characteristic (ROC) curves were plotted for the random forest model and the common bile duct diameter; C: ROC curve for our predictive model. CDL: Choledocholithiasis; AUC: Area under the curve; CI: Confidence interval; FPR: False positive rate; CBD: Common bile duct; GTE: Greater than or equal to.

Table 2 Performance of different models in predicting intraductal concrements (values in square brackets represent 95% confidence interval).

	CBD width (≥ 8 mm)	Bilirubin (≥ 4 mg/dL)	CBD + bilirubin	RF model
Sensitivity	0.795 (0.645; 0.892)	0.308 (0.186; 0.464)	0.256 (0.146; 0.411)	0.923 (0.797; 0.973)
Specificity	0.548 (0.378; 0.708)	0.774 (0.602; 0.886)	0.903 (0.751; 0.967)	0.355 (0.211; 0.531)
PPV	0.689 (0.543; 0.805)	0.632 (0.410; 0.809)	0.769 (0.497; 0.918)	0.643 (0.512; 0.755)
NPV	0.680 (0.484; 0.828)	0.471 (0.341; 0.605)	0.491 (0.366; 0.617)	0.786 (0.524; 0.924)
Correct rule-out (true negative), n (%)	17 (24.29)	24 (34.29)	28 (40.00)	11 (15.71)
False rule-out (missed IDCs), n (%)	8 (11.43)	27 (38.57)	29 (41.43)	3 (4.29)
Accuracy	0.686 (0.570; 0.782)	0.514 (0.400; 0.628)	0.543 (0.427; 0.654)	0.671 (0.555; 0.770)

IDC: Intraductal concrements; CBD: Common bile duct; PPV: Positive predictive value; NPV: Negative predictive value; RF: Random forest.

Correlation with clinical guideline recommendations

Our risk model demonstrated markedly superior sensitivity (92.3% vs 25.6%) and NPV (78.6% vs 49.1%) compared to the combined predictive criteria of CBD width and bilirubin. Moreover, it significantly reduced the rate of missed IDCs to 4.3% [95% confidence interval (CI): 1.5%-11.9%], compared to 41.4% (95%CI: 30.6%-53.1%) using the combined CBD width and bilirubin thresholds (Table 2 and Figure 3A). Three of 39 patients with CDL were missed (false-negative rate 7.7%, 95%CI: 2.7-20.3). Of 14 model-negative patients, 3 (21.4%, 95%CI: 7.6-47.6) had CDL. Corresponding false-negative rates were 20.5% for CBD ≥ 8 mm and 69.2% for bilirubin ≥ 4 mg/dL. Figure 3B and C show ROC curves plotted for a CBD width of ≥ 8 mm and bilirubin of ≥ 4 mg/dL vs our RF model. AUC was higher for the RF model than both cut-offs recommended by the ASGE (AUC_RF: 0.739 vs AUC_Bili: 0.663 and AUC_CBD: 0.702).

Subgroup analysis

We conducted a subgroup analysis by categorizing IDCs into two groups: Microlithiasis and/or sludge, and gallstones. We then evaluated the predictive performance of our RF model compared to CBD width for each subgroup. The RF model outperformed CBD width in predicting gallstones, achieving an accuracy of 68.9% vs 64.4%, with higher sensitivity (96.0% vs 76.0%) and NPV (87.5% vs 62.5%). Conversely, for microlithiasis and sludge, CBD width showed superior accuracy (72.0% vs 64.0%), specificity (63.6% vs 36.4%), PPV (75.0% vs 63.2%), and NPV (77.8% vs 66.7%) compared to the RF model. Both approaches demonstrated equal sensitivity (85.7%) in detecting microlithiasis and sludge. According to the previous comparison, our model performed better in predicting stones than both a dilated CBD (≥ 8 mm) and elevated bilirubin (≥ 4 mg/dL) (AUC_RF: 0.8 vs AUC_Bili: 0.746 and AUC_CBD: 0.633, Figure 4A). Regarding microlithiasis or sludge, the CBD width was more accurate than our model (AUC_RF: 0.623 vs AUC_CBD: 0.808), but the RF model performed better than bilirubin only (AUC_RF: 0.623 vs AUC_Bili: 0.487, Figure 4B). The overall rate of missed IDCs was significantly lower with our RF model (4.3%; 95%CI: 1.5%-11.9%) compared to CBD width (11.4%; 95%CI: 5.9%-21.0%). This difference was even more pronounced in the gallstone subgroup, with missed rates of 2.2% (95%CI: 0.4%-11.6%) vs 13.3% (95%CI: 3.3%-26.2%) for the RF model and CBD width, respectively (Figure 4A and Supplementary Figure 3).

Figure 4 Comparison of decision rules for microlithiasis and sludge. Model performances were computed for two subgroups. A: Patients with gallstones (n = 45); B: Patients with sludge and/or microlithiasis (n = 25). Two receiver operating characteristic curves were plotted for each subgroup for the random forest model, the common bile duct diameter and bilirubin. AUC: Area under the curve; CDL: Choledocholithiasis; CBD: Common bile duct; CI: Confidence interval; FPR: False positive rate.

DISCUSSION

Our study introduces the first predictive model for CDL in acute pancreatitis that functions independently of CBD width and is based on routinely available admission variables. In particular, the model showed a lower false rule-out rate than the evaluated bilirubin- and CBD-based comparator thresholds, suggesting potential value for early risk stratification when the primary goal is to avoid missed CDL.

Given the rising incidence of gallstones and the limited availability of EUS in many centers, along with the suboptimal accuracy of standard cholestasis markers and conventional US, there is a pressing need for improved clinical decision-making tools[4,17]. This novel model may operate as a stratification tool for ERCP, potentially reducing unnecessary endoscopic procedures and associated complications[13]. The strengths of our model are grounded in three key aspects: (1) All model parameters are routinely collected during the standard diagnostic workup for acute pancreatitis and can be rapidly assessed without additional imaging or invasive tests; (2) By excluding CBD width and imaging-based assessment of IDCs, the model avoids observer-dependent variability, especially in US; and (3) It eliminates the risks and costs linked to further diagnostic procedures such as MRCP or EUS. Despite this independence from CBD width, our model achieves remarkable sensitivity and a high NPV. Most existing literature and predictive models focus on cohorts with suspected CDL rather than patients specifically with pancreatitis, and the majority of the studies incorporate CBD width as a central predictor[14,19]. While we deliberately excluded CBD measurements, our data demonstrated a significant correlation between bilirubin levels and CBD diameter, showing that increased CBD width coupled with elevated bilirubin significantly raises the likelihood of CDL. This aligns with the pathophysiology of bile duct obstruction, where stone impaction leads to biliary dilation and cholestasis. Importantly, this relationship is time-dependent, as bilirubin often normalizes more quickly than CBD dimensions following stone passage, which underscores the limitations of relying on either parameter alone and highlights the advantage of our multifactorial risk stratification approach[20,21].

A previous predictive model, such as the nine-feature model developed by Floan Sachs et al[22] for pediatric patients-which included CBD stones on US and CBD diameter-achieved a sensitivity of 91.3%, comparable to our 92.3%; however, their cohort consisted of children undergoing cholecystectomy for biliary disease rather than acute pancreatitis. Steinway et al[14] published a model for patients admitted with suspected CDL that incorporated CBD width and achieved an accuracy of 71.5%, slightly higher than our 67.1%; nevertheless, our model demonstrated superior sensitivity (92.3% vs 70.3%) and NPV (78.6% vs 63.4%) despite focusing on the more specific biliary pancreatitis population and excluding CBD dilation or stones. A recent notable contribution by Mínguez et al[13] presented a predictive model for persistent CDL specifically in acute biliary pancreatitis patients, identifying independent predictors including CBD diameter, GGT levels, and medical history. Their model, which centrally relied on CBD measurements, achieved sensitivity between 84% and 89% and specificity from 77% to 80%. Although effective, their approach depends on the potentially variable CBD measurement, whereas our model avoids this parameter while maintaining comparable sensitivity, offering a more universally applicable tool less reliant on operator skill or imaging quality. Silva-Santisteban et al[23] evaluated the accuracy of these guidelines in a prospective trial involving 359 patients and found that 10% and 13.1% of patients underwent unnecessary ERCPs following the ASGE guidelines from 2019 and 2010, respectively, while ESGE guidelines resulted in 8.1% unnecessary ERCPs. Retrospective data from Lourido Gamboa et al[24] suggest even higher rates of avoidable ERCPs with current guidelines. De Jong et al[25] demonstrated that EUS and MRCP can reduce unnecessary ERCPs; however, guideline adherence remains suboptimal. Our model offers an initial stratification method to select patients prior to EUS or ERCP, potentially streamlining the diagnostic workflow. Although pancreatitis severity did not differ between patients with and without IDCs, those with IDCs were more frequently admitted to intensive care units, possibly due to a higher rate of endoscopic interventions and related complications such as papilla bleeding, highlighting the importance of accurate risk stratification prior to endoscopic procedures.

The comparison with ASGE-based predictors should be interpreted as a pragmatic benchmark rather than a full comparison with the complete guideline algorithm. Because patients with acute cholangitis were excluded and the criterion “CBD stone on imaging” already constitutes near-direct evidence of the outcome, only bilirubin and CBD dilatation could be meaningfully evaluated as comparators. These variables were therefore assessed individually and in combination, using a CBD threshold of ≥ 8 mm as a conservative and clinically pragmatic cut-off.

The performance profile of our model reflects an intentional prioritization of sensitivity over specificity. In acute biliary pancreatitis, missing a persistent CDL may delay appropriate biliary intervention, whereas a false-positive classification will in most cases lead to additional diagnostic clarification rather than immediate harm. In this context, the high sensitivity of 92.3% and the low false rule-out rate support the model’s use as an early triage tool. At the same time, the relatively low specificity of 35.5% indicates that the model should not be used as a stand-alone rule-in test for ERCP, since this would lead to an unacceptable number of unnecessary invasive procedures. Instead, the model is best interpreted as a rule-out-oriented risk stratification tool that may help identify patients in whom urgent invasive management is less likely to be necessary, while prioritizing model-positive patients for further diagnostic evaluation, such as EUS, MRCP, or closer clinical reassessment depending on local resources and guideline-based management. This distinction is particularly important in centers without immediate access to EUS, where a highly sensitive first-line model based on routinely available clinical and laboratory parameters may help guide early decision-making. However, because the NPV remained moderate in our cohort, the model should not be used in isolation to definitively exclude CDL, especially in patients with cholangitis, progressive cholestasis, or other signs of persistent biliary obstruction. The high sensitivity and NPV of our model make it particularly valuable for initial risk assessment, which in clinical practice may translate to fewer unnecessary ERCPs, reduced procedure-related complications and costs, improved patient selection for advanced imaging or therapeutic intervention, and a standardized approach adaptable to centers with limited access to advanced diagnostics.

Bias and limitations

This study has several limitations that should be considered when interpreting the results. First, its retrospective single-center design and the long inclusion period from 2005 to 2021 introduce relevant temporal heterogeneity. Diagnostic pathways for suspected CDL changed substantially over time, particularly with the increasing availability and routine use of EUS and a parallel decline in diagnostic ERCPs. As a result, the reference standard used to confirm CDL was not fully uniform across the study period, and the threshold to proceed to confirmatory testing may also have changed. Although CDL was not confirmed by a single mandatory reference standard in all patients, ERCP was the predominant confirming modality in the CDL-positive cohort. Specifically, among the 167 CDL-positive patients, 160 (95.8%) underwent ERCP, 37 (22.2%) MRCP, 74 (44.3%) EUS, and 138 (82.6%) abdominal US; because multiple modalities could be used in the same patient, these proportions do not sum to 100%. The high proportion of ERCP-confirmed cases strengthens outcome ascertainment and, importantly, enabled a more precise distinction between stones, microlithiasis, and sludge. Nevertheless, because verification was not fully uniform across all patients, partial and differential verification bias cannot be excluded, particularly as patients with higher clinical suspicion were more likely to proceed to ERCP or EUS, whereas a minority were classified based on non-invasive imaging alone. Supplementary Figure 4 summarizes the main diagnostic verification pathways. Second, imaging technology and operator expertise likely improved during the 16-year study period. Advances in US, MRCP, and especially EUS may have increased the detection of smaller CDL in later years, whereas earlier cases may have been missed or classified differently. This issue is particularly relevant for microlithiasis and sludge, which are more difficult to detect than larger stones and may therefore be more susceptible to time-dependent differences in diagnostic sensitivity.

In addition, laboratory assays and reference ranges may have changed over time, which could have introduced further measurement variability in predictors such as bilirubin, AP, and GGT. Although all data were derived from a single center, which reduces variability, assay-related temporal variation cannot be excluded.

Another point that should be considered when interpreting the model performance is the prevalence dependency of predictive values. In contrast to sensitivity and specificity, both NPV and PPV depend on the underlying prevalence of CDL in the evaluated population. Accordingly, the NPV and PPV observed in our study are specific to our cohort and should not be directly extrapolated to centers with a different case mix. In lower-prevalence settings, the NPV would be expected to increase, whereas in higher-prevalence settings it would decrease; conversely, the PPV would increase with higher prevalence. Based on the observed sensitivity and specificity of our model, illustrative calculations show this effect clearly. For example, at a hypothetical CDL prevalence of 10%, the NPV would be expected to be 97.6%, whereas at a prevalence of 50%, it would decrease to approximately 82.2%. In contrast, the PPV would increase from approximately 13.7% at 10% prevalence to 58.9% at 50% prevalence (Supplementary Table 7 and Supplementary Figure 5). These estimates are theoretical and assume stable sensitivity and specificity across populations, but they underline that the predictive values of the model are setting-specific and should be interpreted in light of local disease prevalence. Resampling-based internal validation yielded an AUC of approximately 0.65, while sensitivity remained 89%-93% across all designs (Supplementary Table 3). Without external validation, the model should be regarded as internally validated and exploratory. The model can not exclude CDL on its own and is best used to prioritize patients for further work-up. True-negative classifications are not equivalent to avoided ERCPs.

Finally, the originally used cut-off of 0.25 was not documented as prespecified. A training-only derivation reproduced the same value, though the threshold varied between 0.25 and 0.27 across cross-validation configurations.

To our knowledge, this is the first study to predict CDL without using CBD width in an acute pancreatitis cohort. The promising results warrant follow-up studies in larger, multicenter cohorts to validate the model’s performance. Future research should focus on prospective validation, integration into clinical decision pathways, and potential adaptation for specific patient subgroups. Despite encouraging findings, limitations include the retrospective single-center design, and the need for validation in larger, more diverse populations. While our model performs well without CBD measurements, combining it with selective use of advanced imaging in intermediate-risk patients could further optimize diagnostic algorithms.

CONCLUSION

In conclusion, our model provides a simple, imaging-independent approach for early risk stratification in patients with acute biliary pancreatitis. By prioritizing sensitivity, it may help reduce missed CDL and support the identification of patients who require further diagnostic evaluation. Given its limited specificity, however, the model should not be used as a stand-alone rule-in test for ERCP, but rather as a triage tool to guide subsequent work-up, such as EUS, MRCP, or closer clinical reassessment depending on local resources and clinical presentation. After prospective multicenter validation, this approach may become a useful adjunct for clinical decision-making, particularly in settings without immediate access to EUS.