The role of AI models in enhancing choledocholithiasis diagnosis: A systematic review and meta‑analysis

Introduction

Choledocholithiasis (CDL), the presence of gallstones in the common bile duct (CBD), affects approximately 10-15% of patients with cholelithiasis over 10 years (1). In the U.S., gallstone disease contributes to over 1.2 million emergency department visits annually (2). The incidence of CDL is rising, likely due to an aging population and increased use of diagnostic imaging (3). Prompt and accurate diagnosis is critical, as undetected CDL can lead to complications such as cholangitis, pancreatitis, or biliary obstruction, while unnecessary endoscopic retrograde cholangiopancreatography (ERCP) may cause iatrogenic harm (4).

While ERCP remains the gold standard for diagnosing and treating CDL, its invasive nature demands judicious patient selection (5). The American Society for Gastrointestinal Endoscopy (ASGE) revised its risk stratification criteria in 2019, classifying patients into low, intermediate, or high-risk groups based on clinical, biochemical, and imaging data (6,7). However, these criteria have shown inconsistent accuracy, particularly in high-volume settings, raising concerns about overuse of ERCP and missed diagnoses (8).

Artificial intelligence (AI) transforms medicine by enabling data-driven, personalized care (9). In gastroenterology, AI has been applied in real-time endoscopic image analysis, polyp detection, colorectal neoplasia and malignant biliary strictures, and disease risk stratification (10-12). Supervised machine learning models (MLMs), such as gradient-boosting machines (GBMs), learn from labeled input-output pairs by building and correcting sequential decision trees (10,13). Deep learning (DL) models, including artificial neural networks (ANNs) and convolutional neural networks (CNNs), can integrate diverse clinical inputs such as biochemical studies and imaging findings, and continuously refine predictions through iterative learning (10,13). Unsupervised models, which learn from unlabeled data, are increasingly used to uncover hidden disease subtypes and patterns (10,13,14).

Data-driven ML and DL algorithms in diagnosing and early predicting CDL may improve their therapeutic implementations and decision-making. Applying AI to CDL diagnosis may reduce unnecessary ERCP, lower complication rates, and improve outcomes by enabling noninvasive, individualized risk assessment. However, evidence of their diagnostic performance remains fragmented. This systematic review and meta-analysis aim to evaluate AI-assisted tools' sensitivity, specificity, and overall diagnostic accuracy in predicting CDL and compare their performance with guideline-based approaches such as those from the ASGE.

Materials and methods

Search strategy

Two independent investigators systematically searched MEDLINE, EMBASE, PubMed, and Web of Science from inception to March 2, 2025. Keywords included ‘Artificial intelligence’, ‘machine learning’, ‘computer-aided diagnosis’, ‘biliary obstruction’, ‘choledocholithiasis’, and ‘endoscopic retrograde cholangiopancreatography or ECRP’ with Boolean operators used to optimize search results. A PRISMA diagram was designed according to the PRISMA 2020 guidelines, as shown in Fig. 1, including our meta-analysis studies (15-25). Our meta-analysis focused on the diagnostic accuracy of AI models, which does not fall within the scope of PROSPERO registration, as it primarily registers reviews of interventional studies; therefore, it was not registered.

Figure 1 PRISMA 2020 flow diagram of study selection for meta-analysis of AI models in choledocholithiasis diagnosis. This diagram illustrates of the structured identification, screening, eligibility assessment, inclusion process in this review. To ensure transparency, reproducibility of the literature search and selection methodology, this flow chart documents the number of records retrieved, excluded at each stage, and the full articles assessed for eligibility and the final group of studies included in the quantitative synthesis. AI, artificial intelligence.

Inclusion and exclusion criteria

Studies were eligible if they: i) Evaluated AI or MLMs for CDL diagnosis or prediction, ii) involved human subjects without age restrictions, and iii) were observational, cohort, retrospective, or prospective studies. Studies predicting recurrence or procedural complexity were excluded since they did not directly evaluate diagnostic performance for initial CDL detection, which was our primary outcome. Specifically, exclusion criteria included: i) Models not aimed at diagnostic prediction, ii) studies not using AI or machine learning (ML) tools, iii) models focused on predicting procedural complexity or recurrence prediction after stone removal rather than initial diagnostic performance for CDL, and iv) studies lacking detailed model diagnostic accuracy metrics such as sensitivity, specificity, positive predictive value, negative predictive value, to limit heterogeneity. All metrics are defined based on the reference standard, ERCP or intraoperative cholangiography (IOC) confirmation.

Data extraction

Two reviewers independently extracted data on study design, model type, comparator (e.g., ASGE guidelines), cohorts, and performance metrics. Extracted values included sensitivity, specificity, PPV, LR+, true/false positives and negatives, and accuracy. LR+ and 95% confidence intervals (CI) were calculated from raw data where needed. Validation cohort data were prioritized over training data. Discrepancies were resolved by consensus. Only clearly reported data was included. Each study was assessed for data originality to avoid duplication.

Study quality assessment

Reviewers independently assessed the risk of bias and applicability concerns using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool, designed for diagnostic accuracy studies. This tool evaluates four bias domains (patient selection, index test, reference standard, and flow/timing) and three applicability domains. All studies were included regardless of bias ratings, though heterogeneity due to bias and applicability concerns was acknowledged and considered in data synthesis and interpretation.

Statistical analysis

All statistical analyses were performed using R Studio (Version 2020). A bivariate diagnostic random-effects meta-analysis was conducted, with variance components estimated using the restricted maximum likelihood (REML) method to minimize bias. Python and the Clopper-Pearson (exact binomial) method were used to calculate 95% CI for sensitivity when raw data were available. Heterogeneity was assessed by calculating the I2 and τ² statistics within the bivariate random-effects framework, following the Zhou and Dendukuri approach for diagnostic test accuracy (DTA) meta-analysis (26). We also visually inspected forest plots and summary receiver operating characteristic (SROC) curves to detect threshold effects or outliers. Because moderate heterogeneity was present in several outcomes, we used a random-effects model for all pooled estimates and interpreted summary estimates with caution, emphasizing the direction rather than the absolute magnitude of effect. SROC curves and corresponding area under the curve (AUC) values were generated to evaluate pooled AI model performance and to compare against 2019 ASGE guidelines. P<0.05 was considered to indicate a statistically significant difference. Using RStudio, pooled sensitivity, specificity plots, likelihood ratio positive (LR+) and Deeks' asymmetry test funnel plots for publication bias were also created.

Results

Study characteristics

Our systematic search across four databases yielded 578 articles. After screening titles, abstracts, and full texts using predefined criteria, 20 studies were assessed for eligibility, and 11 studies were included in the final meta-analysis (Fig. 1). We excluded Huerta-Reyna et al for using AI-assisted tools outside the scope of our review (27) and Shi et al focusing on recurrence prediction post-ERCP (28). Studies by Huang et al and Wang et al which assessed the complexity of stone extraction rather than CDL prediction, were also excluded (29,30). Li et al's study focused on visualization enhancement of biliary pathology, rather than stone prediction, which was similarly omitted (31). Akabane et al's large multi-center study was excluded due to missing sensitivity, specificity, and other key diagnostic metrics (32). Additionally, studies applying AI solely to imaging without biochemical or clinical data were excluded (31,33,34).

The 11 included studies encompassed over 7,000 patients evaluated with MLMs. Most studies benchmarked MLM performance against the ASGE guidelines (2010 or 2019 versions) (7,35), while several also compared their models to logistic regression, support vector machines (SVMs), and k-nearest neighbors (KNN) algorithms, enhancing data robustness. Table I summarizes AI models, comparators, patient numbers, sensitivity, specificity, mean age, sex ratio, PPV, NPV, TP, TN, FN, accuracy, AUC, number of input variables, and CBD diameter cut-off values. Input variables varied, incorporating laboratory markers (total bilirubin (TB), alanine aminotransferase (AST), alanine transaminase (ALT), imaging features (CBD dilation, ductal stones), and clinical characteristics. Seven studies compared their models against ASGE criteria and additional classifiers such as logistic regression, CART, and KNN.

Table I Detailed data characteristics of each study, including sensitivity, specificity, AI model type, comparator group and total number of patients.

Table I

Detailed data characteristics of each study, including sensitivity, specificity, AI model type, comparator group and total number of patients.

First author, year	AI model	Comparator	Patient number	Sens., %	Spec., %	95% CI Sens.	Accuracy	AUC	Variables	CBD cutoff	F/M	Mean age, years	PPV	NPV	(Refs.)
Blum, 2024	RF, LR, KNN, XGBoost	ASGE	222	86	72	0.794; 0.923	0.79	0.83	9	6	1.13:1	63.6	77	82	(15)
Cohen, 2021	MLR, ΚΝΝ	ASGE	316	40.8	90.3	0.339; 0.477	0.72	0.738	4	NM	1.08:1	13.8	71.4	72.1	(16)
Dalai, 2021	RF	ASGE	52	61.4	100	0.789; 0.896	0.77	0.791	9	9	1.22:1	46	100	32	(17)
Jovanovic, 2014	ANN	LR	291	92.74	68.42	0.753; 0.847	0.92	0.844	10	7	1.49:1	63	92.34	69.64	(18)
Mena-Camilo, 2024	CNN	LR, LDA, ASGE	292	96.77	92.86	0.946; 0.990	0.93	0.927	12	NM	1.94:1	46	92.86	72.35	(19)
Floan Sachs, 2023	Extra-Trees ML	Pediatric DUCT Score	1,597	91.3	93.4	0.881; 0.945	0.72	0.935	9	NM	2.86:1	13.9	75	98	(20)
Steinway, 2023	GBM	ASGE, ESGE	1,378	70.3	72.3	0.673; 0.731	0.72	0.79	8	6	1.58:1	43	78.1	63.4	(21)
Stojadinovic, 2015	CART	LR	157	80.9	94.7	0.581; 0.946	0.93	0.939	3	4	1.44:1	57	70.8	96.3	(22)
Tranter-Entwistle, 2020	GBM	No comparison	1,315	37	96	0.458; 0.597	0.85	0.840	9	NM	1.33:1	70	67	87	(23)
Vukicevic, 2016	ANN	LR, DT, NB, SVM, KNN	303	88.2	95.8	0.844; 0.920	0.92	0.934	8	7	2.12:1	57	78.9	97.8	(24)
Zhang, 2023	Model Arts AI	ASGE, ESGE	1,199	97	97	0.957; 0.983	0.97	~0.97	11	6	1.07:1	66	97.1	50	(25)

[i] ANN, artificial neural network; CART, classification and regression tree; GBM, gradient-boosting machine; CNN, convolutional neural network; MLR, multivariant logistic regression; KNN, k-nearest neighbors; RF, random forest; ML, machine learning; LR, logistic regression; DT, decision tree; NB, naive Bayes; SVM, support vector machine; LDA, linear discriminant analysis; Sens., sensitivity; Spec., specificity; NM, not mentioned; F/M, female/male.

Sensitivity and specificity of AI models

The pooled sensitivity and specificity for MLM predicting CDL were 83.2% [95% CI: 68.9; 91.8] (Fig. 2A) and 91.1% [95% CI: 84.7; 95.0] (Fig. 2B), respectively. These values, calculated using a bivariate random-effects model, reflect consistently high diagnostic performance across diverse study settings. Individual studies, such as Blum et al and Floan Sachs et al reported sensitivities >90% (15,20). In contrast, others, like Cohen et al exhibited trade-offs between sensitivity and specificity depending on the model threshold chosen (16).

Figure 2 Forest plots displaying (A) pooled sensitivity estimates and (B) pooled specificity estimates across included studies, for AI-assisted models in choledocholithiasis diagnosis. Each square represents an individual study estimate with 95% CI bars, scaled by sample size (total n=7,053). Pooled estimates were calculated using a bivariate random-effects model. AI models demonstrated pooled sensitivity 83.2% (95% CI 68.9-91.8) and specificity 91.1% (95% CI 84.7-95.0). These plots highlight a consistently strong diagnostic accuracy among diverse machine learning models. AI, artificial intelligence; CI, confidence interval.

Positive predictive value and likelihood ratio

The pooled positive likelihood ratio (LR+) from the common-effect model was 8.39 [95% CI: 4.4; 9.5], indicating that patients testing positive via an MLM algorithm were approximately seven times more likely to truly have CDL compared to those who tested negative (Fig. 3A). The pooled PPV was 87.7 [95% CI: 77.4; 93.7] based on the calculated PPV from the included studies (Fig. 3B).

Figure 3 (A) Forest plot of LR+ for AI-based diagnostic models across the 11 included studies (>7,000 patients), illustrating their ability to correctly ‘rule in’ CDL when the test result is positive. The pooled LR⁺ from the common-effect model was 8.39 (95% CI: 7.41-9.50), indicating that AI-assisted tools substantially increase the post-test probability of disease compared to pre-test estimates. (B) Forest plot summarizing PPV estimates from studies evaluating AI-based diagnostic tools for CDL. Pooled PPV was 87.7% (95% CI: 77.4-93.7), demonstrating strong predictive performance in clinical practice. Cis for individual studies highlight variability related to patient selection, disease prevalence, and input variable differences, but overall estimates indicate robust diagnostic accuracy across diverse populations and model architectures. CDL, choledocholithiasis; CI, confidence interval; LR+, positive likelihood ratio; PPV, positive predictive value. AI, artificial intelligence.

SROC analysis

SROC curves and AUC values were generated to assess pooled diagnostic performance across MLMs (Fig. 4A). The primary SROC curve indicated excellent discriminative ability, confirming AI models' robust and consistent performance across diverse study populations and clinical settings.

Figure 4 (A) Bivariate random-effects SROC curve of AI-assisted diagnostic tools derived from 11 studies involving over 7,000 individuals. The pooled AUC was 0.94 (95% CI: 0.83-0.98), indicating high discriminative performance across diverse clinical populations. Threshold effect testing showed no significant heterogeneity (P>0.05). Using Deeks' test, funnel plot analysis for publication bias did not demonstrate significant asymmetry (t=1.84, P=0.13). (B) Comparative SROC curves of AI-assisted models vs. the 2019 ASGE high-risk criteria. AI models demonstrated significantly higher pooled diagnostic accuracy with an AUC of 0.91 (97.5% CI: 0.76-0.96) compared to ASGE criteria (AUC 0.76; 97.5% CI: 0.72-0.80). Mixed-effects meta-regression (generalized linear mixed model, likelihood ratio test P=0.019) confirmed the superior performance of AI models. Sensitivity and specificity estimates for AI models were 86.1% (95% CI: 68.5-94.7) and 93.3% (95% CI: 71.4-98.7), respectively, compared to ASGE guideline performance of 78.4% (95% CI: 67.5-86.4) and 57.5% (95% CI: 37.3-75.6). CDL, choledocholithiasis; SROC, summary receiver operating characteristic; AI, artificial intelligence; AUC, area under the curve; CI, confidence interval; ASGE, American Society for Gastrointestinal Endoscopy.

Subgroup analysis: AI-models vs. ASGE guidelines

Seven studies directly compared their MLM to the 2019 ASGE risk stratification criteria. Pooled sensitivity and specificity values were used to generate SROC curves for both models. As shown in Fig. 4B, MLMs consistently achieved higher AUC values and diagnostic accuracy than ASGE high-risk criteria. This suggests that AI may enhance precision and reduce reliance on subjective guideline interpretation in pre-ERCP decision-making.

In bivariate random-effects meta-analysis, MLMs achieved an AUC of 0.915 (95% CI: 0.763-0.961) compared to 0.766 (95% CI: 0.7;-0.7) for ASGE guidelines. The absolute difference in AUC (ΔAUC=0.148) reflects a clinically meaningful improvement favoring AI tools. Although the 95% CI for this difference (-0.002 to 0.205) includes zero, the P-value (0.053) indicates a trend toward statistical significance. The pooled sensitivity for AI models was 82.3%, with a false positive rate of 9.8%. Heterogeneity across studies ranged from 52.4% (Zhou and Dendukuri method) to 97.8% (Holling unadjusted).

Quality assessment for bias

QUADAS-2 tool evaluation identified ‘some concerns’ in several studies, mainly in patient selection (convenience sampling, exclusion of cholangitis, etc.) and flow/timing (delayed or unclear index/reference timing) domains (Fig. 5). Specifically, Tranter-Entwistle et al faced issues with missing clinical data, 36% misclassification, and manual diagnosis reassignment (23). Temporal alignment of index and reference tests was inconsistently reported. Steinway et al lacked clarity regarding blinding details and applied a retrospective cohort design, excluding cholangitis, potentially limiting generalizability (21). Mena-Camilo et al despite ERCP-confirmed diagnoses, demonstrated significant class imbalance (86% positive) and heavy reliance on imputed clinical data, raising data integrity concerns (19). QUADAS-2 revealed moderate risk biases likely inflated accuracy estimates in some models by limiting negative cases or excluding diagnostically challenging scenarios. Therefore, while AI models demonstrated strong pooled performance, conclusions should be interpreted with caution given these study-level limitations.

Figure 5 The Quality Assessment of Diagnostic Accuracy Studies 2 tool was used to evaluate risk of bias and applicability concerns for the 11 included studies (>7,000 patients). The round-shaped plot (top) summarizes the judgment of each study across four bias domains, patient selection, index test, reference standard, and flow/timing, and three applicability domains, using grey (low risk), and white (some concerns). No areas of high risk were noted. The summary plot (bottom) depicts the overall proportion of studies rated at low risk or with some concerns within each domain. Overall, most studies demonstrated low risk in the index test and reference standard domains, while patient selection and flow/timing presented the most frequent concerns, primarily due to retrospective designs, incomplete clinical data, or unclear temporal alignment of index and reference tests. Applicability concerns were minimal for most studies, reflecting consistent patient populations and appropriate test applications. These results indicate an overall moderate to high quality of included studies, supporting the robustness of the meta-analysis findings.

Discussion

Accurate stratification of patients with suspected CDL remains clinically challenging. To improve specificity and reduce unnecessary ERCPs, the 2019 ASGE guidelines revised high-risk criteria to include only patients with both TB >4 mg/dl and CBD dilation on imaging (Fig. 6) (6). This change, supported by prior studies showing improved specificity with these combined features (36,37), raised specificity from 55% under the 2010 criteria to approximately 80% (38). The 2019 update also incorporated computed tomography (CT) alongside ultrasonography (USG) for risk stratification (6). Per the current guidelines, in the absence of definitive radiologic evidence of CDL, ERCP is recommended for patients with a suggestive clinical presentation, elevated liver enzymes, CBD dilation >6 mm, TB >4 mg/dl, or signs of cholangitis (Fig. 6). Those not meeting high-risk criteria are advised to undergo magnetic resonance cholangiopancreatography (MRCP) or EUS (Fig. 6) (6). While MRCP is a sensitive and specific tool (39), its overuse may prolong hospital stays and increase costs (40). Despite refinements, current guidelines lack tools for quantitative CDL risk estimation. Our meta-analysis is the first comprehensive synthesis of AI-assisted prediction models for CDL, highlighting their potential for non-invasive risk stratification and optimized ERCP selection (Fig. 6).

Figure 6 Proposed diagnostic pathway incorporating AI-assisted risk stratification for suspected CDL into ASGE guidelines. Conceptual diagram illustrating the integration of AI-based predictive tools into clinical workflows for patients with suspected CDL, particularly among intermediate-risk and high-risk groups lacking definitive imaging or clinical confirmation. AI, artificial intelligence; ASGE, American Society for Gastrointestinal Endoscopy; CDL, choledocholithiasis; TB, total bilirubin.

The reviewed studies employed diverse MLMs, including traditional and DL approaches like ANNs, CNNs, and GBMs, and supervised learning algorithms such as classification and regression trees (CART), logistic regression, KNN, Naïve Bayes (NB), and Scikit-learn frameworks Several models demonstrated superior performance: Stojadinovic et al found that CART outperformed logistic (22); Tranter-Entwistle et al reported that GBM outperformed random forest and KNN (23); Dalai et al tested GLM, SVM, and random forest, with the latter performing best and included in our meta-analysis (17). Cohen et al utilized a KNN-based model (16), while Zhang et al developed the ModelArts ExeML tool (Huawei), outperforming logistic regression, LDA, QDA, NB, KNN, and Ranger (25). Floan Sachs et al applied Extra-Trees to a 9-feature pediatric model, outperforming the 3-feature Pediatric DUCT score (20). Blum et al compared logistic regression, XGBoost, random forest, and KNN, with random forest performing best in adults with suspected CDL (15).

ANNs were among the earliest tools. Jovanovic et al introduced an ANN-based model (18), while Vukicevic et al enhanced it by automating configuration using genetic algorithms, outperforming logistic regression, decision trees, NB, SVMs, and KNN (24). Mena-Camilo et al applied a CNN-based model that exceeded logistic regression and LDA in sensitivity, supporting DL's role in non-invasive CDL diagnosis (19).

Our meta-analysis of 7,000+ subjects across 11 studies showed strong predictive performance of AI-assisted models, with a pooled sensitivity of 83.2%, specificity of 91.1% and a PPV of 87.7% [77.4; 93.7]. Despite heterogeneity in models, patient populations, and predictive variables, consistently high accuracy supports AI as a non-invasive diagnostic tool and underscores its adaptability, especially where guidelines fall short. A key source of heterogeneity is in training/validation cohorts, impacting generalizability. Jovanovic et al excluded patients with prior CCY, cholangitis, and primary sclerosing cholangitis, achieving high AUC (0.884) and sensitivity (92.7%) but lower specificity (68.7%) due to few ERCP-negative cases and low NPV (18).

Other studies also assessed high-suspicion cohorts undergoing ERCP or IOC, excluding cholangitis cases (15,21). Dalai et al focused on Hispanic/Latino patients with prior CCY (17). Again, low ERCP-negative rates skewed the sample distribution. Blum et al applying similar criteria, confirmed stones in only 50% of cases (15). Mena-Camilo et al trained a CNN regardless of CCY status (19); Zhang et al used MRCP or IOC for diagnosis, excluding cholangitis (25). Tranter-Entwistle et al found CDL in 19% of acute biliary cases, with their GBM model showing low sensitivity (37%) and high specificity (96%) due to missing data imputation (23). In elective surgical populations, Stojadinovic et al found 37% CDL prevalence; their CART model exceeded logistic regression in sensitivity (90.9% vs. 57.2%) but not specificity (92.3% vs. 94.7%) (22). Similarly, Vukicevic et al's ANN achieved 88.2% sensitivity and 95.8% specificity among CCY patients with IOC (24).

Although CDL is typically adult-onset, pediatric cases are rising with obesity (41-43). Cohen et al evaluated suspected pediatric CDL using IOC/ERCP, prioritizing PPV by setting 90% specificity, which yielded only 40.8% sensitivity and 71.5% accuracy, marking it an outlier (16). Floan Sachs et al used pediatric elective CCY cases, with CDL confirmed via MRCP, ERCP, and/or IOC. Their 9-feature model outperformed the 3-feature Pediatric DUCT score (AUC 0.957 vs. 0.935; sensitivity 91.3% vs. 78.3%), while maintaining specificity (20), estimating only 8.7% of cases missed, compared to 21.7% with the 3-feature model. These results support ML's role in optimizing ERCP in children, with performance thresholds that are unable to meet institutional needs.

Input variables ranged from 3 to 53 (Table I), typically combining biochemical laboratory data and imaging. USG and CT were common, while MRCP was underused. Dalai et al showed models incorporating CT/MRCP had higher AUC than USG alone (17). Cohen et al excluded CBD stone presence on USG due to its high predictive value; univariate analysis identified (AST), CBD diameter, and ductal stone presence on USG as key predictors (16). Common biochemical values in all studies included TB, AST, ALT, and alkaline phosphatase (ALP). Few used follow-up biochemical values. Steinway et al found no benefit from serial labs (21). Stojadinovic et al emphasized cystic duct diameter (>4 mm), elevated ALP, and multiple small gallstones (‘dangerous’ stones) as key predictors, though their elective CCY cohort may not reflect acute cases (22). Jovanovic et al (18) found lipase and CBD diameter to be influential, with lipase and AST negatively, and CBD diameter positively associated with CDL, supporting the theory that pancreatitis may follow stone passage rather than concurrent CDL (6,18,22).

Seven studies directly compared MLMs to the 2019 ASGE guidelines; others used published ASGE performance data for benchmarking. MLMs consistently outperformed current guidelines, particularly in high-suspicion cases (Table I) (18,19,21). Most models achieved high sensitivity, except Tranter-Entwistle et al (23) and Cohen et al (16), which prioritized specificity above 90%, reducing sensitivity. Most models reported >90% accuracy, though accuracy can be prevalence-dependent. Large sample sizes strengthen these conclusions. Integrating MLMs may enhance individualized CDL risk stratification.

Steinway et al's GBM achieved sensitivity and specificity above 70%, outperforming ASGE (61.9/62.8%) and European Society of Gastrointestinal Endoscopy (ESGE) (46.9/86.3%), avoiding 22% of unnecessary ERCPs and identifying 48% of true positives missed by ESGE guidelines (21). Dalai et al's best-performing model (77.3% sensitivity, 75% specificity) outperformed the ASGE's high-likelihood group, which had higher sensitivity (90.3%) but lower specificity (25%) (17). Threshold tuning raised sensitivity to 97.7%, illustrating MLM flexibility. Blum et al found Random Forest best among four classifiers (AUC 0.83, sensitivity 94%), exceeding ASGE (15). Zhang et al reported ASGE's high sensitivity (98.5%) but poor specificity, while their ModelArts ExML tool achieved AUCs of 0.77-0.81(25). Jovanovic et al's ANN model (AUC 0.884) surpassed logistic regression (AUC 0.752), despite excluding several ASGE criteria (18). Cohen et al's model prioritized specificity in pediatrics, achieving 40.8% sensitivity but outperforming ASGE in specificity (16).

This meta-analysis showed higher accuracy for MLMs (AUC 0.91) vs. ASGE 2019 guidelines (AUC 0.72). Unlike ASGE's fixed risk categories, MLMs may provide individualized risk stratification. While ASGE guidelines remain a valuable risk stratification tool, their broad criteria often underperform in real-world populations (6). AI models offer personalization, especially in low-resource settings or special populations like pediatrics (16,20). High-specificity models reduce unnecessary ERCPs; high-sensitivity models expedite referrals. From a therapeutic standpoint, these tools may refine patient selection for ERCP, particularly in intermediate/high-risk patients lacking definitive findings. In such cases, MLMs could guide decisions between MRCP and ERCP. The findings also open avenues for developing real-time, bedside decision-support tools trained on multicenter data and embedded within electronic medical records. Though AI tool thresholds may vary, high sensitivity supports timely referral in community hospitals, while high specificity aids triage in tertiary centers. ASGE guidelines recommend direct CCY for low-risk patients, yet some harbor occult stones (6). While the benefit of pre-CCY ERCP is debated, AI may identify low-risk patients needing further evaluation (6,44-46). These findings are consistent with broader AI applications in gastroenterology, including colorectal polyp and neoplasia detection or classification (11,12,47,48), highlighting the potential of AI-based tools to improve diagnostic accuracy and efficiency. However, extrapolation across clinical settings remains challenging, underscoring the need for external validation and standardized performance reporting. Integrating AI into clinical workflows may reduce unnecessary testing, optimize resources, and improve outcomes (49).

This study supports the diagnostic potential of AI tools for CDL, but several limitations exist. We identified 11 distinct AI models using varied input data: labs, imaging, clinical features, introducing heterogeneity that hinders direct comparisons and pooled estimates. A key limitation is the lack of prospective validation; most studies were retrospective, single-centered, limiting generalizability. Homogeneous populations and limited external validation further constrain applicability. Comparisons were inconsistent, using differing ASGE guidelines, impairing uniform benchmarking. Future studies should prioritize standardized comparisons using current ASGE criteria and diverse cohorts. Some studies lacked key statistics (e.g., CI, PPV, NPV), and unclear training-validation cohort separation may have skewed performance. In several cases, unclear differentiation between training and validation cohorts may have distorted performance outcomes. These issues underscore the need for prospective validation and standardized reporting to enable clinical adoption.

In conclusion, this study highlights that ML tools demonstrate promising diagnostic potential for CDL detection and may enhance current or future guideline-based risk stratification. While pooled sensitivity and specificity were high, heterogeneity, reliance on retrospective data, and a somewhat limited prospective validation constrain definitive conclusions. AI-based tools may complement existing guidelines via a non-invasive accurate alternative in refining patient selection for ERCP, enhancing both the safety and efficiency of care, especially in acute and resource-limited settings. Future work should prioritize validation in prospective multicenter studies, ensure model generalizability, and support integration into clinical workflows before clinical implementation.

Acknowledgements

Not applicable.

Funding

Funding: No funding was received.

Availability of data and materials

The data generated in the present study may be requested from the corresponding author.

Authors' contributions

PGD, SGD and AB conceived and designed the study. PGD and SGD performed literature search, data analysis, and interpretation of data and results. PGD and SGD confirm the authenticity of all the raw data. PGD, SGD and AB prepared the draft manuscript and critically revised the manuscript. AB supervised the study. All authors reviewed the results, and read and approved the final manuscript.

Ethics approval and consent to participate

Not applicable.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Authors' information

Panagiotis G. Doukas: https://orcid.org/0000-0001-8050-963X. Sotirios G. Doukas: https://orcid.org/0000-0002-1840-659X.

References