Interpretable multimodal artificial intelligence model for predicting advanced neoplasia in pancreatic cystic lesions

Abstract

BACKGROUND

International guidelines for pancreatic cystic lesions (PCLs) rely primarily on morphology-based criteria and demonstrate limited discrimination for advanced neoplasia. The integration of clinical, endoscopic ultrasound (EUS), and cyst fluid biomarkers into interpretable risk models may improve preoperative risk stratification.

AIM

To develop and internally validate an interpretable multimodal model for predicting advanced neoplasia in PCLs.

METHODS

This retrospective single-center cohort study included 187 adults who underwent surgical resection for PCLs between 2012 and 2025. Clinical variables, cross-sectional imaging findings, EUS features, and cyst fluid biomarkers (carcinoembryonic antigen and glucose) were analyzed. Using least absolute shrinkage and selection operator-penalized logistic regression, a core clinical-EUS model and an integrated multimodal model were developed (training set, n = 131) and internally validated (n = 56). Discrimination was assessed using the area under the receiver operating characteristic curve (AUC). Calibration and decision curve analysis were also performed. Model performance was compared with American Gastroenterological Association 2015, Fukuoka 2017, European 2018, and Kyoto 2024 criteria.

RESULTS

Advanced neoplasia was identified in 58 of 187 patients (31.0%). In the validation cohort (17 advanced; 39 non-advanced), the integrated multimodal model achieved an AUC of 0.91 (95%CI: 0.82-0.97), with a sensitivity of 82.4% and specificity of 89.7%, significantly outperforming international guideline-based criteria (AUC range: 0.70-0.79; all P < 0.01). Performance remained stable in the intraductal papillary mucinous neoplasm subset (AUC 0.90) and in cysts without mural nodules (AUC 0.88). Decision curve analysis demonstrated a superior net benefit across clinically relevant thresholds.

CONCLUSION

An interpretable multimodal model integrating clinical, EUS, and cyst fluid data improves the discrimination of advanced neoplasia in surgically resected PCLs and may support preoperative risk stratification. Prospective external validation is required before clinical implementation.

Key Words: Pancreatic cystic lesions; Artificial intelligence; Endoscopic ultrasound; Cyst fluid biomarkers; Intraductal papillary mucinous neoplasm; Risk stratification

Core Tip: Current guideline-based algorithms for pancreatic cystic lesions predominantly rely on morphology and show limited discrimination for advanced neoplasia. In this retrospective cohort study of 187 surgically resected lesions, we developed and internally validated an interpretable multimodal model integrating clinical variables, endoscopic ultrasound features, and cyst fluid biomarkers. The model achieved superior discrimination compared with international guideline criteria and demonstrated stable performance across key subgroups. This interpretable framework may support preoperative risk stratification, pending external validation.

INTRODUCTION

The widespread use of cross-sectional imaging has made pancreatic cystic lesions a routine clinical challenge. The central difficulty is no longer detection itself, but discrimination—determining which cyst requires continued surveillance, which warrants endoscopic ultrasound (EUS) with fine-needle aspiration, and which justifies surgical resection. Achieving this balance without exposing patients to unnecessary pancreatic surgery remains one of the most nuanced decisions in pancreatology.

Contemporary management frameworks, including the American Gastroenterological Association (AGA) 2015 guideline, the Fukuoka 2017 international consensus, the European evidence-based guidelines, and the updated Kyoto 2024 consensus for intraductal papillary mucinous neoplasms (IPMNs), provide structured, rule-based algorithms to guide these decisions[1-5]. These systems primarily rely on morphological variables, such as cyst size, main pancreatic duct (MPD) diameter, and the presence of enhancing mural nodules. However, operational definitions are not uniform across guidelines. Thresholds for mural nodule size, anatomical sites of MPD measurement, and the categorization of lesions as “worrisome features” or “high-risk stigmata” differ meaningfully between documents. Such definitional variability affects reported sensitivity and specificity and complicates fair benchmarking of predictive models[6,7].

The Kyoto 2024 update further refines IPMN-specific risk stratification and emphasizes standardized imaging and EUS characterization[4]. By modifying surveillance intervals and surgical triggers, it alters the escalation logic in clinically relevant ways, making it an essential contemporary comparator when evaluating novel risk-stratification tools.

EUS plays a central role in the evaluation of pancreatic cystic lesions. In addition to cyst size and MPD diameter, EUS enables a detailed assessment of mural nodules and their vascularity, cyst wall architecture, duct-cyst communication, and, when available, elastographic stiffness patterns. In routine clinical practice, many key management decisions are refined after EUS reassessment rather than based on cross-sectional imaging alone. Nevertheless, most guideline algorithms remain predominantly morphology-driven, and the systematic integration of detailed EUS phenotyping into quantitative risk prediction models remains limited.

Against this background, we aimed to develop an interpretable multimodal prediction model for advanced neoplasia, defined as high-grade dysplasia or invasive carcinoma, in pancreatic cystic lesions. The model integrates clinical features, cross-sectional imaging findings, detailed EUS characteristics, and cyst fluid biochemistry. Because our cohort included only surgically resected lesions with histopathological confirmation, the model was designed for preoperative risk stratification among surgical candidates rather than for population-level surveillance triage[8-10]. Rather than replacing clinical judgment, this approach seeks to complement multidisciplinary decision-making by providing a transparent, individualized risk estimate.

MATERIALS AND METHODS

Study design and population

This retrospective, single-center cohort study was conducted at Kayseri City Hospital, a tertiary referral center for pancreatic diseases in Türkiye. We reviewed consecutive adults (≥ 18 years) who underwent surgical resection for a histopathologically confirmed pancreatic cystic lesion between January 2012 and August 2025.

The study was reported in accordance with the STROBE[11] and TRIPOD[12] statements.

The primary objective was to perform preoperative risk stratification among patients already considered for surgical management rather than to develop a population-level surveillance tool. Restricting the cohort to surgically resected lesions ensured that histopathology served as the reference standard and minimized the circularity inherent in radiology-defined outcomes.

Inclusion required preoperative radiologic or endosonographic identification of a pancreatic cystic lesion on contrast-enhanced computed tomography (CT), magnetic resonance imaging (MRI), or EUS before surgical resection. The exclusion criteria were as follows: (1) Incomplete preoperative clinical, imaging, or laboratory data precluding application of at least one of the four major international guidelines (AGA 2015[1], Fukuoka 2017[2], European 2018[3], Kyoto 2024[4]); (2) Non-diagnostic cyst fluid aspirate when biochemical analysis was clinically indicated; and (3) Final pathology demonstrating a non-neoplastic cyst (e.g., pseudocyst), solid pseudopapillary neoplasm, or neuroendocrine tumor.

Of the 214 consecutive patients, 187 met the eligibility criteria and comprised the final analytic cohort.

The study protocol was approved by the Ethics Committee of the University of Health Sciences, Kayseri City Hospital (protocol No. 602; October 13, 2025). The requirement for informed consent was waived due to the retrospective study design and use of anonymized data.

Data collection and variable definitions

Preoperative data were extracted from the electronic medical records. Clinical variables included age, sex, presenting symptoms (abdominal pain, jaundice, unintentional weight loss > 5% within six months), history of pancreatitis, and new-onset diabetes mellitus.

Serum carbohydrate antigen 19-9 (CA 19-9) levels were measured using an electrochemiluminescence immunoassay (Roche Diagnostics). Elevation was defined as > 37 U/mL, consistent with international guideline practice[2].

All imaging and EUS examinations were independently re-reviewed by two experienced gastroenterologists and one abdominal radiologist who were blinded to the final histopathology. Discrepancies were resolved by consensus. Interobserver agreement was assessed in a predefined random subset using Cohen’s κ coefficient.

Imaging and EUS protocol

Cross-sectional imaging: For CT and MRI, the following variables were recorded: (1) Maximum cyst diameter (mm); (2) MPD diameter measured at the site of maximal dilation; (3) Presence of mural nodules; (4) Wall thickening; and (5) Calcifications.

MPD measurements were standardized to the maximal visible diameter across the head, body, or tail. Sensitivity analyses were used to evaluate alternative measurement conventions.

Endoscopic ultrasound: EUS was performed by one of three experienced endosonographers using standardized linear echoendoscopes.

Recorded variables included: (1) Cyst location (head, body, or tail); (2) Maximum cyst diameter (mm); (3) Morphology (unilocular, multilocular, macrocystic ≥ 10 mm, microcystic < 10 mm); (4) MPD diameter and duct-cyst communication; (5) Mural nodule presence and size (mm); (6) Doppler-confirmed vascularity; (7) Wall thickening (> 3 mm); and (8) Elastography pattern [soft (score 1-2) vs hard (score 3-5)].

An enhancing mural nodule was defined as a solid, vascularized structure projecting into the cyst lumen, irrespective of size. This definition is broader than certain guideline thresholds (e.g., ≥ 5 mm in Kyoto 2024[4]) but reflects standardized clinical practice at the study center. Sensitivity analyses were conducted using alternative cutoffs (≥ 3 mm and ≥ 5 mm).

Cyst fluid analysis

Cyst fluid was obtained via EUS-guided fine-needle aspiration (19- or 22-gauge needle) when clinically indicated. Samples were processed within 30 minutes of collection.

Carcinoembryonic antigen (CEA) was measured using an electrochemiluminescence immunoassay. Glucose was measured using the hexokinase method. Amylase was measured using an enzymatic colorimetric assay.

Established thresholds (CEA > 192 ng/mL; glucose < 50 mg/dL) were reported descriptively and were considered suggestive of mucinous differentiation[8,9]. Because these cutoffs are validated primarily for lineage discrimination rather than grade prediction, CEA and glucose were treated as continuous predictors in the primary model (CEA was log-transformed owing to its skewed distribution).

Cytological findings were categorized as benign, atypical, suspicious, or malignant. Suspicious or malignant cytology was considered a high-risk feature.

Molecular analysis (exploratory subgroup)

Next-generation sequencing (NGS) was performed in 78 patients (41.7%) with adequate samples, at the discretion of the treating clinicians.

A targeted 7-gene pancreatic panel (KRAS, GNAS, TP53, SMAD4, CDKN2A, RNF43, and VHL) was used, with a minimum sequencing depth of > 500 × and a variant allele frequency limit of detection of 1%. The assay failure rate was 12.4%.

Lineage-associated mutations were interpreted according to established molecular diagnostic frameworks[10] and were not used for primary outcome classification to avoid incorporation bias. Mutations were categorized as lineage-associated (KRAS, GNAS) or progression-associated (TP53, SMAD4, CDKN2A).

Molecular data were not imputed and were excluded from primary model derivation.

Outcome definition

The primary endpoint was advanced neoplasia, which was defined histopathologically as high-grade dysplasia or invasive carcinoma (including T1a carcinoma in situ) in the surgical specimen. Low-grade dysplasia and benign lesions were used as comparators.

Pathology reports were reviewed by two authors blinded to preoperative risk estimates.

Model development

The cohort was randomly divided (stratified by outcome) into: (1) Training set (n = 131; 70%); and (2) Internal validation set (n = 56; 30%).

Least absolute shrinkage and selection operator (LASSO)-penalized logistic regression was used for feature selection and overfitting reduction. The optimal regularization parameter (λ) was selected using 10-fold cross-validation in the training set.

Missingness was < 8% for all variables. Missing data were handled using multiple imputation by chained equations and pooled according to Rubin’s rules[13].

Threshold determination and validation

The optimal probability threshold was determined exclusively in the training set using the Youden index. This threshold was fixed and applied unchanged to the validation set without re-optimization.

Optimism correction was performed using 200 bootstrap resamples of the training data.

Discrimination was assessed using the area under the receiver operating characteristic curve (AUC). Comparisons between correlated receiver operating characteristic (ROC) curves were performed using the nonparametric method described by DeLong et al[14].

Calibration was evaluated using the calibration intercept, calibration slope, Hosmer-Lemeshow test, and Brier score.

Decision-analytic evaluation

Clinical utility was assessed using decision curve analysis as described by Vickers and Elkin[15] across threshold probabilities of 5%-50%. Threshold ranges were selected to reflect clinically plausible decision points discussed in multidisciplinary tumor boards.

Explicit counterfactual decision tables were generated for threshold probabilities of 20%, 30%, and 40%.

Guideline benchmarking

Model performance was benchmarked against AGA 2015[1], Fukuoka 2017[2], European 2018[3], and Kyoto 2024[4] criteria.

Each guideline was operationalized as an ordinal risk score based on the number of worrisome features or high-risk stigmata present. Missing guideline components were treated as absent to reflect routine clinical documentation; sensitivity analyses excluding incomplete cases yielded comparable results.

Benchmarking against Kyoto 2024 was restricted to IPMN cases, consistent with its intended scope[4].

Subgroup and sensitivity analyses

Pre-specified subgroup analyses included: (1) IPMN-only; (2) Mucinous cysts only (IPMN + MCN); and (3) Cysts without mural nodules.

Sensitivity analyses evaluated alternative mural nodule thresholds (≥ 3 mm vs ≥ 5 mm) and alternative MPD measurement conventions.

Model interpretability

SHapley Additive exPlanations (SHAP) were applied to quantify feature importance and directionality at both population and individual levels.

Statistical software and review

All statistical tests were two-sided, and P values < 0.05 were considered statistically significant. Analyses were conducted using R version 4.3.1. The statistical methods of this study were reviewed by a biomedical statistician.

RESULTS

Cohort characteristics and outcomes

Between January 2012 and August 2025, 214 consecutive patients underwent surgical resection for a pancreatic cystic lesion. After applying the eligibility criteria, 187 patients comprised the final analytic cohort.

The median age was 66 years (interquartile range: 60-73), and 98 patients (52.4%) were female. Histopathology confirmed advanced neoplasia—defined as high-grade dysplasia or invasive carcinoma—in 58 patients (31.0%), whereas 129 patients (69.0%) had low-grade dysplasia or benign histopathology.

The cohort was predominantly mucinous, comprising 127 branch-duct IPMNs (67.9%), 34 main-duct or mixed-type IPMNs (18.2%), and 21 mucinous cystic neoplasms (11.2%). Five lesions (2.7%) were serous cystadenomas or other benign cysts.

Baseline characteristics stratified by outcome are summarized in Table 1. Patients with advanced neoplasia were significantly older (median 69 years vs 63 years; P = 0.002). They more frequently had unintentional weight loss (31.0% vs 10.1%, P < 0.001), new-onset diabetes (22.4% vs 10.1%, P = 0.04), enhancing mural nodules (56.9% vs 11.6%, P < 0.001), MPD ≥ 5 mm (53.4% vs 16.3%, P < 0.001), cyst size > 30 mm (67.2% vs 32.6%, P < 0.001), and elevated serum CA 19-9 (> 37 U/mL) (46.6% vs 11.6%, P < 0.001).

Table 1 Baseline clinical, imaging, and cyst fluid characteristics stratified by outcome (n = 187).

Characteristic	Total (n = 187)	Advanced neoplasia (n = 58)	Non-advanced (n = 129)	P value
Age, median (IQR), year	66 (60-73)	69 (63-76)	63 (57-70)	0.002
Female sex	98 (52.4)	28 (48.3)	70 (54.3)	0.52
Unintentional weight loss	31 (16.6)	18 (31.0)	13 (10.1)	< 0.001
New-onset diabetes	26 (13.9)	13 (22.4)	13 (10.1)	0.04
Enhancing mural nodule on EUS	48 (25.7)	33 (56.9)	15 (11.6)	< 0.001
MPD ≥ 5 mm	52 (27.8)	31 (53.4)	21 (16.3)	< 0.001
Cyst size > 30 mm	81 (43.3)	39 (67.2)	42 (32.6)	< 0.001
Serum CA 19-9 > 37 U/mL	42 (22.5)	27 (46.6)	15 (11.6)	< 0.001
Cyst fluid CEA > 192 ng/mL	58 (31.0)	35 (60.3)	23 (17.8)	< 0.001
Cyst fluid glucose < 50 mg/dL	51 (27.3)	31 (53.4)	20 (15.5)	< 0.001
Hard elastography pattern (score ≥ 3)	41 (21.9)	24 (41.4)	17 (13.2)	< 0.001
Cyst type				< 0.001
BD-IPMN	127 (67.9)	29 (50.0)	98 (76.0)
MD/mixed IPMN	34 (18.2)	20 (34.5)	14 (10.9)
MCN	21 (11.2)	6 (10.3)	15 (11.6)
Other	5 (2.7)	3 (5.2)	2 (1.5)

Data are presented as n (%) or median (interquartile range). Categorical variables were compared using χ² or Fisher’s exact test, as appropriate; continuous variables were compared using the Mann-Whitney U test. MPD: Main pancreatic duct; BD-IPMN: Branch-duct intraductal papillary mucinous neoplasm; MD: Main-duct; MCN: Mucinous cystic neoplasm; IQR: Interquartile range; CEA: Carcinoembryonic antigen; CA 19-9: Carbohydrate antigen 19-9; IPMN: Intraductal papillary mucinous neoplasms.

Cyst fluid biomarkers differed significantly between the groups: CEA > 192 ng/mL was present in 60.3% vs 17.8% (P < 0.001), and glucose < 50 mg/dL in 53.4% vs 15.5% (P < 0.001). A hard elastography pattern (score ≥ 3) was present in 41.4% vs 13.2% (P < 0.001).

Model development and validation

The cohort was randomly divided (stratified by outcome) into: (1) Training set: n = 131; and (2) Validation set: n = 56.

Advanced neoplasia was present in 41 of 131 (31.3%) patients in the training set and in 17 of 56 (30.4%) patients in the validation set.

Guideline benchmarking and overall discrimination

A comparison of the AUC values for the integrated multimodal model, the core clinical-EUS model, and international guideline-based strategies is presented in Figure 1 and summarized in Table 2.

Figure 1 Area under the curve comparison of prediction models and guideline-based criteria. Bar plot illustrating the discriminative performance of the integrated multimodal model and the core clinical-endoscopic ultrasound model in the validation cohort (n = 56; 17 advanced neoplasia, 39 non-advanced), alongside the area under the curve (AUC) values of international guideline-based criteria (American Gastroenterological Association 2015, Fukuoka 2017, European 2018, Kyoto 2024). The integrated multimodal model demonstrated the highest AUC (0.91), followed by the core model (0.87). Guideline-based criteria showed lower discriminative performance (AUC range: 0.70-0.79). The Kyoto 2024 AUC reflects evaluation within the intraductal papillary mucinous neoplasm subset (n = 161), consistent with its intended scope. AGA: American Gastroenterological Association; AUC: Area under the curve.

Table 2 Diagnostic performance of prediction models in the validation cohort and area under the curve values of guideline-based criteria.

Model/criteria	AUC (95%CI)	Sensitivity	Specificity	PPV	NPV
Core clinical-EUS model	0.87 (0.76-0.95)	76.5% (13/17)	79.5% (31/39)	61.9% (13/21)	88.6% (31/35)
Integrated multimodal model	0.91 (0.82-0.97)	82.4% (14/17)	89.7% (35/39)	77.8% (14/18)	92.1% (35/38)
Kyoto 2024 criteria1	0.79	-	-	-	-
Fukuoka 2017 criteria	0.77	-	-	-	-
European 2018 criteria	0.76	-	-	-	-
AGA 2015 criteria	0.70	-	-	-	-

¹Kyoto 2024 was evaluated in the intraductal papillary mucinous neoplasms subset (n = 161), consistent with its intended scope.

Validation distribution (model evaluation): 17 advanced neoplasia, 39 non-advanced (n = 56). DeLong tests comparing the integrated model with each guideline-based strategy were significant (all P < 0.01). AUC: Area under the receiver operating characteristic curve; PPV: Positive predictive value; NPV: Negative predictive value; EUS: Endoscopic ultrasound; AGA: American Gastroenterological Association.

In the validation cohort, the integrated multimodal model achieved the highest discriminative performance (AUC 0.91), followed by the core clinical-EUS model (AUC 0.87). Guideline-based criteria demonstrated lower discrimination, with AUC values of 0.70 for AGA 2015, 0.77 for Fukuoka 2017, 0.76 for European 2018, and 0.79 for Kyoto 2024.

The comparisons between the integrated model and each guideline-based strategy were statistically significant (all P < 0.01, DeLong test).

Because Kyoto 2024 is intended specifically for IPMNs, its performance was evaluated within the IPMN subset (n = 161). In this subgroup, the integrated model achieved an AUC of 0.90 (95%CI: 0.82-0.96), compared with 0.79 for Kyoto 2024.

Core clinical-EUS model

LASSO regression retained four predictors: Age, enhancing mural nodule, MPD diameter (continuous), and serum CA 19-9 (dichotomous).

In the validation set, the core model achieved: (1) AUC = 0.87 (95%CI: 0.76-0.95); (2) Sensitivity = 76.5% (13/17); (3) Specificity = 79.5% (31/39); (4) Positive predictive value (PPV) = 61.9% (13/21); and (5) Negative predictive value (NPV) = 88.6% (31/35).

Model performance metrics for the models and guideline-based criteria are summarized in Table 2.

Integrated multimodal model and calibration

After adding cyst fluid CEA and glucose as continuous predictors, discrimination improved.

In the validation set, the integrated model achieved: (1) AUC = 0.91 (95%CI: 0.82-0.97); (2) Sensitivity = 82.4% (14/17); (3) Specificity = 89.7% (35/39); (4) PPV = 77.8% (14/18); and (5) NPV = 92.1% (35/38).

Improvement over the core model was statistically significant (DeLong P = 0.02).

Calibration demonstrated good agreement between the predicted and observed risks (Figure 2): Calibration intercept 0.06, calibration slope 0.95, Hosmer-Lemeshow P = 0.61, and Brier score 0.11.

Figure 2 Calibration plot of the integrated multimodal model in the validation cohort (n = 56). Calibration plot comparing predicted vs observed probabilities of advanced neoplasia across deciles of risk. The diagonal line represents perfect calibration. The model demonstrated good agreement between predicted and observed risk (calibration intercept 0.06; slope 0.95; Brier score 0.11), indicating stable probability estimation in internal validation.

Model interpretability

SHAP analysis (Figure 3) demonstrated that enhancing mural nodules had the highest relative mean absolute SHAP value, indicating the greatest contribution to model prediction. This was followed by MPD diameter, serum CA 19-9, cyst fluid glucose, cyst fluid CEA, and age. The relative ranking of predictors was clinically plausible and consistent across subgroup analyses.

Figure 3 Relative SHapley Additive exPlanations feature importance for the integrated multimodal model. The bar plot illustrates the relative mean absolute SHapley Additive exPlanations (SHAP) values of predictors in the integrated multimodal model. Enhancing mural nodules demonstrated the greatest contribution to predicted risk of advanced neoplasia, followed by main pancreatic duct diameter, serum carbohydrate antigen 19-9, cyst fluid glucose, cyst fluid carcinoembryonic antigen, and age. Feature contributions are presented as relative mean absolute SHAP values, reflecting overall predictor importance in the model. MPD: Main pancreatic duct; CEA: Carcinoembryonic antigen; CA 19-9: Carbohydrate antigen 19-9; SHAP: SHapley Additive exPlanations.

Subgroup and sensitivity analyses

Performance remained stable across pre-specified subgroups: (1) IPMN-only (n = 161): AUC 0.90; (2) Mucinous-only (IPMN + MCN, n = 160): AUC 0.89; and (3) No mural nodules (n = 131): AUC 0.88.

Sensitivity analyses using alternative mural nodule thresholds (≥ 3 mm vs ≥ 5 mm) and alternative MPD measurement sites altered the AUC by ≤ 0.02.

Exploratory molecular analyses

NGS data were available for 78 patients (41.7%), with a failure rate of 12.4%. Progression-associated mutations (TP53, SMAD4, CDKN2A) were detected in 22/78 patients (28.2%).

In this exploratory subgroup, adding the progression-associated mutation status resulted in a modest increase in AUC from 0.91 to 0.93. Given the limited sample size and non-systematic testing, this finding is considered hypothesis-generating and was not incorporated into the primary model.

Clinical utility: Decision-analytic evaluation

Decision curve analysis demonstrated a superior net benefit of the integrated model compared with guideline strategies across threshold probabilities of approximately 10%-45% (Figure 4).

Figure 4 Decision curve analysis comparing net clinical benefit of the integrated model and guideline-based strategies. Decision curve analysis illustrating net benefit across threshold probabilities between 5% and 50% for recommending surgical resection. The integrated multimodal model demonstrates superior net benefit across clinically relevant thresholds compared with guideline-based criteria. Horizontal lines represent “treat all” and “treat none” strategies. Because all patients in the cohort underwent resection, net benefit estimates reflect counterfactual model-based scenarios.

Counterfactual decision tables for the validation cohort (n = 56) at representative risk thresholds are shown in Table 3.

Table 3 Counterfactual decision table at representative risk thresholds (validation cohort, n = 56).

Risk threshold	Surgical candidates (n)	Advanced neoplasia correctly classified	Patients without advanced neoplasia potentially spared surgery	Advanced neoplasia missed (n)
20%	20	14/17	33/39	3
30%	16	13/17	36/39	4
40%	13	12/17	37/39	5

Validation distribution: 17 advanced neoplasia, 39 non-advanced. All patients in the cohort underwent resection; therefore, the number of surgeries potentially avoided represents counterfactual estimates based on model classification rather than observed clinical avoidance of surgery.

At a 20% threshold: (1) Surgical candidates: 20; (2) Correctly classified advanced neoplasia: 14/17; (3) Patients without advanced neoplasia potentially spared surgery: 33/39; and (4) Advanced neoplasia missed: 3.

At a 30% threshold: (1) Surgical candidates: 16; (2) Correctly classified advanced neoplasia: 13/17; (3) Patients without advanced neoplasia potentially spared surgery: 36/39; and (4) Advanced neoplasia missed: 4.

At a 40% threshold: (1) Surgical candidates: 13; (2) Correctly classified advanced neoplasia: 12/17; (3) Patients without advanced neoplasia potentially spared surgery: 37/39; and (4) Advanced neoplasia missed: 5.

Because all patients underwent resection, these represent counterfactual model-based scenarios rather than observed clinical avoidance of surgery.

DISCUSSION

In this retrospective surgical cohort, we developed and internally validated an interpretable multimodal model for preoperative prediction of advanced neoplasia in pancreatic cystic lesions. By anchoring the outcome to definitive histopathology, we minimized the misclassification inherent to radiologic surrogates. However, resection cohorts are intrinsically enriched for higher-risk lesions, and the prediction-model literature cautions that discrimination derived from enriched samples may overestimate performance relative to surveillance populations[16]. Accordingly, the model is intended for preoperative risk stratification among surgical candidates rather than for population-level surveillance triage.

The integrated model demonstrated improved discrimination compared with contemporary guideline-based strategies (AUC 0.91 in internal validation). Its incremental value should be interpreted in the context of definitional variability across guideline frameworks. Differences in mural nodule size thresholds, MPD measurement sites, and feature categorization materially influence apparent diagnostic performance[6,7]. By explicitly harmonizing operational definitions and conducting sensitivity analyses for key thresholds, we sought to reduce structural benchmarking bias. Because Kyoto 2024 is specific to IPMN, its evaluation was appropriately restricted to IPMN cases, consistent with its intended scope[4].

Multimodal integration beyond morphology

Existing guidelines are largely morphology-driven. Although mural nodules and MPD dilation are established high-risk features[4-6], rule-based systems cannot readily incorporate continuous or biologically graded variables. Our model integrates clinical data, cross-sectional imaging, detailed EUS characteristics, and cyst fluid biomarkers within a probabilistic framework.

Cyst fluid CEA and glucose levels were modeled as continuous predictors rather than dichotomized thresholds. Although CEA levels > 192 ng/mL and glucose levels < 50 mg/dL are widely used to distinguish mucinous from nonmucinous cysts[8,9], these cutoffs are lineage markers rather than validated grade predictors. In our analysis, the biomarkers appeared primarily to improve rule-in specificity in mucinous lesions. Subtype-restricted analyses (IPMN-only and mucinous-only) demonstrated that the integrated model retained incremental discrimination beyond lineage classification, suggesting that multimodal integration provides added value over morphology alone.

Within the molecular subgroup, progression-associated alterations (TP53, SMAD4, CDKN2A) showed an incremental predictive contribution, consistent with established pancreatic tumorigenesis models[17,18]. However, molecular testing was not systematic and the subgroup size was limited. Therefore, these findings should be interpreted as hypothesis-generating rather than confirmatory. We deliberately separated lineage-associated mutations (KRAS, GNAS) from progression-associated alterations to avoid conflating cyst type with grade prediction.

Artificial intelligence and explainability

Previous artificial intelligence (AI)-driven approaches for pancreatic cyst assessment have primarily focused on radiomics or imaging-based deep learning models[19,20]. Although such models may achieve strong discrimination in selected datasets, concerns persist regarding their interpretability, reproducibility, and clinical deployment[21].

To address these limitations, we prioritized transparency. By applying SHAP within a penalized logistic regression framework[22], we generated an explainable model that quantifies feature-level contributions at both the population and individual levels. This approach aligns with emerging reporting standards for clinical AI systems, which emphasize transparency and reproducibility[23]. Rather than functioning as a black-box classifier, the model provides individualized risk estimates that remain clinically interpretable.

Clinical utility and decision framing

Because all patients in this cohort underwent surgical resection, the reduction of unnecessary surgery could not be directly observed. Therefore, we implemented explicit counterfactual decision tables across clinically relevant risk thresholds. This decision-analytic framework is consistent with recommended approaches for evaluating prediction models in clinical decision contexts[24]. Across threshold probabilities between approximately 10% and 45%, the integrated model demonstrated a superior net benefit compared with guideline-based strategies.

Importantly, we do not propose replacing guideline algorithms. Instead, the model may complement multidisciplinary tumor board discussions by providing individualized risk quantification in cases in which guideline criteria yield equivocal or discordant recommendations.

Strengths and limitations

The principal strength of this study is its histopathologically confirmed outcome, which reduces the misclassification bias inherent to imaging-defined endpoints. Additional strengths include blinded imaging review, explicit calibration assessment (intercept, slope, Brier score), bootstrap optimism correction, sensitivity analyses addressing definitional variability, and transparent decision-analytic modeling.

The limitations of this study include its retrospective single-center design and spectrum enrichment inherent to surgical cohorts[16]. External validation in prospective, multicenter surveillance populations is essential before clinical implementation. Furthermore, molecular profiling was available only in a subset, which may have introduced selection bias and limited inferences regarding the incremental value of genomic markers.

Future directions

Prospective external validation, impact analysis on surgical referral patterns, and evaluation of oncologic outcomes are required to determine the real-world clinical utility[25]. The model architecture is compatible with the integration of emerging biomarkers, including cell-free DNA fragmentomics and proteomic classifiers[17,25], and could be implemented as a web-based calculator or integrated into structured radiology and endoscopy reporting systems.

CONCLUSION

In this histopathologically confirmed surgical cohort, we developed and internally validated an interpretable multimodal prediction model for advanced neoplasia in pancreatic cystic lesions. By integrating clinical variables, cross-sectional imaging, detailed EUS features, and cyst fluid biomarkers within a transparent modeling framework, the model demonstrated improved discrimination and calibration compared with guideline-based criteria.

Given the surgical nature of the cohort, the model is intended for preoperative risk stratification among patients already considered for surgery rather than for general surveillance triage. Counterfactual decision modeling suggests that individualized risk estimation may support more informed surgical decision-making; however, prospective external validation is required before clinical adoption.

This interpretable AI-assisted framework provides a scalable and clinically transparent approach for multimodal risk integration in the management of pancreatic cystic lesions.

ACKNOWLEDGEMENTS

The authors thank the Department of Radiology and the Department of Pathology at the University of Health Sciences, Kayseri City Hospital for their collaboration and support in data collection and image analysis.