We organized the review around successive decisions in pancreatic cancer surgery and examined five prespecified clinical domains: preoperative evaluation of resectability and vascular involvement; three-dimensional reconstruction and surgical planning; stratification for neoadjuvant and adjuvant treatment; management of perioperative complication risk; and models predicting lymph node metastasis, recurrence, and survival.

This approach enabled comparison of evidence maturity and limits to clinical use across tasks without combining findings into a single effect estimate. We adopted a structured narrative design and synthesized evidence according to the surgical decision being addressed. The review was not intended as a systematic review or meta-analysis; accordingly, we did not pool effect estimates or report according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework. Our synthesis examined individual study designs and findings alongside similarities and differences in populations, data modalities, validation levels, and intended clinical applications.

PubMed/MEDLINE served as the main bibliographic database. The core-evidence search covered publications from 1 January 2020 to 28 February 2026. We performed a structured search covering five predefined domains within surgical care for pancreatic cancer: evaluation of resectability and vascular involvement; three-dimensional reconstruction, segmentation, and operative planning; stratification for perioperative therapy; postoperative complications; and prediction of lymph node metastasis, recurrence, survival, and prognosis. We paired terms for pancreatic cancer and pancreatic ductal adenocarcinoma (PDAC) with domain-specific clinical terms and terms for AI, machine learning, deep learning, radiomics, and pathomics. We also screened the reference lists of relevant studies for reports that the search might have missed. A supplementary search of PubMed/MEDLINE, covering publications through 17 September 2026, aimed to identify recent developments. Supplementary Method 1 and Supplementary Figure 1 detail the search concepts, eligibility criteria, and study-selection process.

We considered original English-language studies that used computational approaches such as AI, machine learning, deep learning, radiomics, or pathomics for clinically relevant tasks within the pancreatic cancer surgical pathway. Inclusion as a direct AI study required reporting of model development, validation, or clinical evaluation based on patient-level clinical, imaging, pathological, operative, or omics data. The direct-AI evidence set excluded reviews, editorials, conference abstracts, animal and cell studies, treatment studies without AI, conventional three-dimensional reconstruction lacking an AI component, and studies that did not address a relevant pancreatic surgical task. We retained selected adjacent studies and protocols in a separate category when they clarified technical limitations or emerging routes to clinical translation.

The core evidence set ultimately included 34 studies: 31 direct AI studies and 3 adjacent studies or protocols. Formal assessment of evidence maturity was restricted to the 31 direct AI studies. Supplementary Table 1 summarizes their characteristics. References cited for disease background, clinical context, methodological frameworks, regulation, or wider discussion of translation were not graded for evidence maturity. Li Z (Zhike Li) and Peng G (Gongze Peng) performed literature retrieval and evidence synthesis independently, and resolved differences through discussion. Full-text eligibility assessment was not performed independently in duplicate. The 112 records retained after relevance screening were candidates for detailed eligibility assessment. The historical records do not document a formal process for resolving eligibility disagreements.

For each study in the core set, we extracted information on study design, target population, data sources, sample characteristics and event composition, clinical task, type of model, methods of development and validation, key performance findings, calibration, decision thresholds, analyses of utility, workflow features, clinical outcomes, and major methodological limitations.

Our methodological appraisal drew on the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis-Artificial Intelligence (TRIPOD+AI) statement, the Prediction model Risk Of Bias Assessment Tool (PROBAST), the Checklist for Artificial Intelligence in Medical Imaging (CLAIM), principles of decision-curve analysis, and the Fairness, Universality, Traceability, Usability, Robustness, and Explainability for Artificial Intelligence (FUTURE-AI) guideline^[7-11]. We used these frameworks to guide critical assessment, without deriving a numerical risk-of-bias score.

We focused particularly on cohort design, sample and event adequacy, separation of development and evaluation datasets, independent validation, calibration, decision utility, integration into workflows, and clinical applicability. When source articles provided confidence intervals, we reported them; where they did not, we retained the published point estimates.

AI research in pancreatic cancer surgery addresses diverse tasks, from segmentation and evaluation of vascular invasion to stratification of treatment sensitivity, prediction of complications, and estimation of survival. These tasks differ in their intended populations, consequences of error, and resulting clinical actions. Consequently, the same area under the receiver operating characteristic curve (AUC) or accuracy can have different clinical implications. Failure to detect vascular involvement may underestimate operative complexity and margin risk, while falsely predicting lymph node metastasis may classify additional patients as high risk unnecessarily. The acceptable trade-off between sensitivity and specificity therefore varies by task. Discrimination measures how well a model distinguishes higher-risk from lower-risk cases; calibration measures how closely predicted probabilities match observed event rates. Decision thresholds translate continuous estimates of risk into actions, including further testing, closer monitoring, or discussion of treatment options. For uncommon outcomes, assessment of an operating point must also account for positive predictive value (PPV), sensitivity, baseline risk, and the clinical consequences of errors.

Performance marks the start of the evidence chain. Internal validation mainly examines stability within the environment of the development data, while independent external or temporal validation tests generalizability across institutions, devices, populations, and periods. Prospective performance validation tests a fixed model on prospectively collected data for the same task. Collecting data prospectively does not, by itself, establish an independent evaluation or show that predictions alter clinical decisions. Actual use in prospective workflows and its influence on decisions are assessed separately at Level 4. In multistage systems that combine segmentation, feature extraction, and classification, reporting should distinguish errors at each step and show how they affect the final output. Identifying the main points of failure is equally necessary, since strong performance early in a pipeline can conceal weaknesses in its ultimate clinical task. Progress from predictive validity to clinical utility requires interpretable and actionable thresholds, compatibility with real-world workflows, and prospective comparative evidence of net decision benefit or better patient outcomes^[11,12].

We assigned each of the 31 direct AI studies the highest cumulative evidence level it met by satisfying all required criteria. Level 1 covered development and internal validation of a model for the target clinical task. To reach Level 2, a fixed model required independent external, temporal, or prospective performance validation for that same task. Evaluation data must not have informed feature selection, parameter tuning, or the choice of operating threshold. When evaluation data were used to update a model, an additional independent evaluation was necessary.

Level 3 required all Level 2 criteria to be met, together with clinically interpretable evidence of decision utility appropriate to the task and output. Probabilistic risk models had to demonstrate calibration suitable for their intended use and connect predictions with a specific clinical action using a prespecified or clinically interpretable threshold. They also needed evidence of net benefit or additional decision value relative to an appropriate clinical comparator. Neither calibration nor decision-curve analysis was sufficient on its own. For locked treatment-predictive classifiers, investigators had to fix the classification rule and threshold before independent evaluation. Independent data also had to demonstrate differential treatment value using a formal test of biomarker-treatment interaction or an equivalent validated analysis incorporating a treatment comparator. Prognostic stratification, a significant association within one treatment group, or significance in one group but not another did not establish treatment-predictive utility.

Level 4 required prospective use within a clinical workflow and measurement of its effects on decisions, management, or workflow. Validating performance prospectively was insufficient to meet this level. Level 5 required evidence that an AI-assisted strategy affected patient outcomes, resource use, efficiency, or safety.

Li Z and Peng G jointly assessed study-level evidence maturity and verified the supporting evidence. Ambiguous or borderline classifications were reviewed with Liu L (Liping Liu) and He Y (Yongzhu He) until a final classification was agreed upon. No inter-rater agreement statistic was calculated.

Every direct AI study met the Level 1 criteria; of the 31 studies, 13 attained at least Level 2 and 3 attained Level 3. Classification by the highest level achieved placed 18 studies at Level 1, 10 at Level 2, and 3 at Level 3, with none at Levels 4 or 5. Most evidence thus concerned model development and internal validation or independent validation of performance. Only a few studies supplied sufficient clinically interpretable evidence of decision utility to exceed Level 2. No study prospectively measured the effect of integrating AI on clinical decisions, as required for Level 4, or demonstrated benefits for patient outcomes, resource use, or safety, as required for Level 5. Table 1 gives an overview of evidence across decision points, while Figure 2 displays its maturity distribution. Table 2 presents representative studies, and Figure 3 outlines the route to clinical translation.

Three-dimensional reconstruction, automated segmentation, and three-dimensional printing are connected technologies, but each requires different evidence. Segmentation algorithms delineate structures such as the pancreas, tumor, pancreaticobiliary ducts, and vessels. Reconstruction then represents these structures in an interactive spatial model, while printing creates a physical version. Segmentation and reconstruction can be components of an AI-driven imaging workflow; printing is not itself AI. Evaluation standards also differ. Segmentation is usually assessed against manual contours. Three-dimensional planning should be compared with conventional two-dimensional interpretation or manually constructed three-dimensional models, and physical printing must add information beyond a digital model viewed on screen. Keeping these distinctions clear helps identify whether a study validates segmentation, anatomical visualization, or clinical planning.

A multicenter study developed a model to segment pancreatic cancer and adjacent anatomy automatically. Overlap was favorable for the pancreas and several vessels, but tumor segmentation was less successful: DSCs were approximately 0.57 internally and 0.55 externally^[21]. Pancreatic tumors can be small, have poorly defined margins, and overlap with inflammatory or atrophic changes, making segmentation harder than for larger solid organs with clear contrast. These weaknesses matter for surgical planning. Missing tumor regions or displaced boundaries can distort measurements of tumor-vessel contact and remnant pancreatic volume, as well as three-dimensional displays. In addition to overall DSC, planning systems should therefore report failures involving small lesions, unclear boundaries, and differing acquisition protocols. Reporting should also cover how often manual correction is needed, how long it takes, and how consistently observers perform it. Extensive expert correction may erode the workflow gains from automation.

Three-dimensional vascular models connect anatomical displays with individualized indicators of risk. In a retrospective study at two institutions, deep learning enabled virtual models and measurements of superior mesenteric artery zones. Selected branching features were statistically associated with postoperative pancreatic fistula (POPF), other severe complications, and blood loss during surgery^[22]. The cohort included resections for several pancreatic tumor types. Accordingly, these anatomical features should be interpreted as indicators of operative complexity across diseases, rather than as oncological predictors specific to PDAC. They could draw attention to vascular variants, branching patterns, and technical challenges in preoperative discussions, helping with operative planning and communication of risk. The associations do not show that adapting the operative strategy to the model reduces complications. Prospective workflow evaluations should document whether the model alters the surgeon's plan and whether intraoperative findings support that alteration.

Physical printing requires more resources than digital three-dimensional visualization. A small retrospective study of 16 patients with borderline-resectable disease found that segmentation and three-dimensional reconstruction appeared to improve visualization of venous involvement, whereas a printed model supplied no additional information beyond the on-screen version^[23]. Producing the digital model took approximately 3 h, and printing took approximately 36 h, at a cost of about €700-1,000 per model. These requirements can limit access when treatment sequencing or operative scheduling demands prompt decisions. Although this is adjacent evidence without an AI component, it highlights a wider challenge in assessing surgical-planning tools. An intuitive visual display does not necessarily add clinical value. New approaches need direct comparisons against two-dimensional interpretation, on-screen three-dimensional models, and standard MDT assessment.

The most immediate benefits of AI-assisted operative planning are likely to be faster segmentation, clearer depiction of complex spatial relationships, and quantitative description of individual vascular anatomy^[21-23]. Assessment can proceed at three levels. Technical assessment addresses segmentation errors and consistency between centers; process assessment covers manual correction, interpretation time, and agreement within teams; clinical assessment examines changes to operative plans and outcomes. Future studies should select endpoints appropriate to each level, including correction burden, time spent interpreting and discussing findings, interobserver agreement, changes in surgical approach, blood loss, complications, and margin status. They should also determine which complex vascular anatomies or borderline-resectable cases gain additional value, rather than applying three-dimensional processing routinely to all patients. For automated segmentation and three-dimensional modeling to function as evaluable decision support, they must deliver reliable, reproducible information that adds value within everyday workflows.

AI in pancreatic surgery is expanding beyond preoperative planning to recognize and visualize anatomy in operative video. One study trained a deep-learning model to identify two vascular structures: the superior mesenteric vein–portal vein axis and the splenic vein at pixel level. In videos of laparoscopic distal pancreatectomy and pancreatoduodenectomy, recall reached 85.0%, precision 91.1%, and mean DSC 0.754, with processing at roughly 11 frames per second^[24]. A separate study used 78 robotic pancreatoduodenectomy videos to develop a continuous AI score for superior mesenteric artery exposure and the extent of dissection. This score correlated strongly with expert surgeons’ ratings (Spearman’s rho = 0.767, P < 0.001), with high agreement among surgeons [intraclass correlation coefficient (ICC), 0.81]. Displaying AI overlays also improved surgeons’ confidence in judging the dissection extent (P < 0.001)^[25]. These studies illustrate emerging intraoperative recognition of anatomy, quantitative assessment of dissection, and assisted visualization. They are cited as contextual reports and were not among the 31 direct AI studies graded for evidence maturity.

The clinical purpose of a perioperative risk model depends on when its inputs become available. Preoperative imaging and variables can guide risk discussions, surgical preparation, and monitoring plans. In a single-center study, a decision tree based on preoperative characteristics predicted major bleeding with similar accuracy in training and temporal testing. A neural-network comparator performed less accurately on the test set, underscoring the importance of assessing overfitting^[33]. Anatomical information obtained during surgery may inform immediate management, while postoperative pathology and early laboratory findings better explain subsequent changes in risk. POPF, hemorrhage, infection, and venous thromboembolism (VTE) can prolong hospitalization, increase reintervention, and delay adjuvant therapy. Prediction endpoints should therefore capture clinically meaningful events affecting recovery. Under the International Study Group of Pancreatic Surgery classification, clinically relevant grade B and C POPF are distinguished from biochemical leakage^[34]. Consistent outcome definitions and explicit prediction times are essential for model comparison and for specifying intended use. They also help prevent bias from grouping minor biochemical changes with complications that alter management.

Studies of POPF demonstrate how different sources contribute complementary information. An analysis of a national surgical quality database included 8,666 patients undergoing pancreatoduodenectomy and predicted clinically relevant POPF using variables available before surgery alone. XGBoost attained an AUC of 0.72^[35]. Although the large cohort reduces random error, the findings also indicate that routine preoperative variables offer only moderate initial stratification and incompletely characterize gland texture, duct features, and local histology. A multicenter digital-pathology study applied AI to resection-margin slides to quantify tissue components. The area of fibrosis showed an independent association with clinically relevant POPF; reported AUCs were approximately 0.58-0.73^[36]. The preoperative model may aid counseling before surgery; the slide-based model instead offers a postoperative histological explanation because its inputs are obtained after surgery. Intended use must respect this timing. A model requiring postoperative slides cannot serve as a preoperative decision tool.

Models incorporating body composition seek to capture systemic condition beyond what routine clinical variables describe. A two-institution study evaluated 230 distal pancreatectomy patients, including 42 in an external test cohort. Combining clinical variables with two- and three-dimensional body-composition features yielded an external AUC of 0.82, sensitivity of about 81%, and specificity of about 76%, exceeding the performance of a clinical-only model^[37]. The results suggest additional predictive information from body composition and demonstrate a role for automated segmentation in perioperative risk research. However, the cohort was limited to distal pancreatectomy, and external testing involved few patients. Differences between procedures in duct management, pancreatic remnants, and fistula mechanisms leave it uncertain whether the model transfers to pancreatoduodenectomy or other workflows for assessing body composition.

VTE prediction demonstrates the importance of interpreting discrimination at a stated operating threshold. One study analyzed 51,916 pancreatectomy records, including 743 postdischarge VTE events (1.4%), using a stratified 80/20 split between training and holdout data^[38]. The leading gradient-boosting model produced two performance estimates: 0.83 (95%CI: 0.80-0.87) for the receiver operating characteristic AUC (ROC-AUC) and 0.34 for the area under the precision-recall curve (AUPRC). Supplementary results for a threshold of 0.10 showed 39.6% sensitivity, 99.5% specificity, 51.8% PPV, and 99.1% negative predictive value. These results neither identify a clinically optimal cutoff for extended prophylaxis nor demonstrate prospective benefit from model-guided prophylaxis.

A priority for complication models is to connect predicted risk with a defined action and the consequences of incorrect classification. False-negative results can deprive high-risk patients of closer monitoring or additional assessment. False-positive results can lead to more testing, inpatient observation, or prophylactic treatment. The balance of benefit and harm differs across these actions; consequently, drain management, nutritional support, and thromboembolic-risk assessment may need different thresholds even when based on one model^[37,38]. Protocols should also document whether and when predictions alter the team's initial judgment and why clinicians decline recommendations^[12]. The present findings support further assessment of risk stratification at the appropriate clinical time point. In particular, a discharge-based VTE model is not a preoperative tool. Prospective comparisons remain necessary to determine whether model use reduces complications or hospital stay, prevents unnecessary interventions, or improves completion of later treatment.

Moving AI models for pancreatic cancer surgery into practice requires interpretable, reproducible findings applicable to the intended population. Gaps in reporting of populations, data sources, prediction timing, outcome definitions, missing-data methods, training and validation splits, feature selection, or optimization obscure whether bias, information leakage, or overfitting has inflated performance. Estimates can be overly optimistic when feature selection uses the entire dataset, images from one patient appear in both training and test sets, or investigators repeatedly consult the test set while tuning the model. Alongside point estimates, reports should provide confidence intervals, examples of failure, and additional value relative to a basic clinical model. This information helps separate random variation and technical improvements from clinically meaningful gains. TRIPOD+AI, PROBAST, and CLAIM offer complementary guidance on prediction-model reporting, assessment of bias, and reporting of medical-imaging AI^[7-9].

Independent external validation is necessary to assess generalizability between centers, but the strength of that evidence depends on where the external cohort originates and how heterogeneous it is. Testing a small cohort from the same region, using similar equipment or identical preprocessing, addresses a different question from testing across regions, devices, and healthcare systems. The 2026 multicenter vascular-invasion study reported discrimination in five external-center cohorts, but model-locking and data isolation could not be fully verified for Level 2^[20]. Conversely, substantial deterioration during temporal validation of a body-composition prognostic model^[47] highlights the need to monitor changes in data distributions over time. Reports of external validation should specify whether patient selection matched the development cohort, whether outcome definitions were consistent, and how missing inputs were managed. Otherwise, performance differences cannot reliably be attributed to the population, workflow, or algorithm. A practical route to implementation begins with silent prospective validation at the intended center, observing performance, calibration, and failure patterns without affecting care. Once local performance is understood, studies of clinician-AI collaboration can follow. Even consistently strong external performance does not establish that clinical decisions improve.

Translating predictions into action requires calibration and a clear understanding of the consequences of the chosen threshold. Decision-curve analysis adds to discrimination by evaluating the net benefit of model-assisted decisions with intervening in all patients or none across a range of threshold probabilities^[10]. The VTE model demonstrates how sensitivity and PPV vary with the threshold; the lymph node metastasis model demonstrates the burden associated with high sensitivity and low specificity^[38,39]. Both patterns must be judged in relation to the proposed action, rather than by AUC alone. Differences in event rates between institutions may alter predictive values and calibration, so a cutoff optimized in development cannot simply become a universal clinical threshold. Every threshold should correspond to a defined action. More false positives may be tolerable when the consequence is additional imaging, whereas changes to surgery or treatment demand stronger evidence. Before implementation, investigators should establish actionable thresholds in the target population and report calibration, predictive values, and the consequences of misclassification. Sensitivity analyses should assess whether net benefit remains stable across thresholds.

Interpretable outputs should enable clinicians to scrutinize the basis for a prediction and identify uncertainty. Models based on tumor-vessel features and the multicenter model of vascular invasion convert selected imaging signs into quantitative features that can be traced and reviewed clinically^[15,20]. Work on structured reporting assisted by large language models suggests that text tools can improve completeness, although image review, external validation, and human oversight remain essential^[50]. Further assessment should test whether explanations remain stable when inputs are perturbed, accord with clinical knowledge, and help identify errors. Studies of clinician-AI collaboration can document acceptance, modification, or rejection of recommendations, compare benefits across levels of clinician experience, and establish procedures for resolving disagreement between experts and models. Linking data quality, predictions, clinical review, and escalation of abnormal findings within an accountable process would make interpretability an auditable part of practice rather than merely a display feature.

A functioning model also needs to fit imaging, pathology, MDT, and perioperative processes. Three-dimensional modeling, digital pathology, and omics assays introduce additional image processing, specimen preparation, or sequencing. Their value in practice depends on whether the resulting information justifies the time, resources, and maintenance involved^[23,26,27]. Implementation studies should identify who obtains and checks inputs, when results become available, how missing or abnormal inputs are addressed, and whether recommendations are included in the medical record and MDT documentation. Evaluation should account for turnaround time, manual correction, repeat tests, and information-system maintenance costs. Following model updates, calibration and scope of use require reconfirmation. Monitoring measures and suspension criteria should address declining performance, safety incidents, and disparities between populations. The Developmental and Exploratory Clinical Investigations of Decision-support systems driven by Artificial Intelligence (DECIDE-AI) guideline highlights safety, human factors, workflows, and possible harms in early clinical evaluation^[12]. FUTURE-AI covers fairness, universality, traceability, usability, robustness, and explainability throughout the AI lifecycle^[11]. These frameworks distinguish successive, connected stages of evidence: predictive performance, deployment in practice, and patient benefit. Prospective assessment of effects on decisions and real-world outcomes remains the major gap for AI in pancreatic cancer surgery.

Implementation should reflect each institution’s capacity. High-volume tertiary centers may be equipped to integrate multimodal platforms locally, monitor performance prospectively, and update models systematically. Regional centers and institutions with fewer resources may gain more from externally validated tools for specific tasks that accept standardized inputs, need little local infrastructure, and permit centralized support. Decisions about deployment should therefore account for case volume, data systems, specialist access, integration with information systems, technical assistance, and the ability to monitor performance over time.

Regulatory translation must also be assessed within each jurisdiction. In the United States, the U.S. Food and Drug Administration (FDA) issued draft guidance in January 2025 on lifecycle management and marketing submissions for AI-enabled device software functions and final guidance in August 2025 on predetermined change control plans^[51,52]. Within the European Union, devices enabled by AI may fall under complementary provisions of the Medical Device Regulation or In Vitro Diagnostic Medical Device Regulation and the Artificial Intelligence Act when relevant high-risk criteria apply^[53]. Comparisons between jurisdictions should examine intended use, requirements for clinical evidence, human oversight, procedures for software changes, cybersecurity, transparency, conformity assessment or marketing authorization, and post-market surveillance. Regulatory status in one jurisdiction should not be assumed to transfer directly to another.

Economic assessments should account for software purchase, annotation of data, system integration, upkeep, manual correction, and subsequent testing. They should also evaluate possible reductions in ineffective interventions, repeated examinations, and time spent making decisions. Conducting budget-impact and cost-effectiveness analyses alongside studies of decision impact can establish whether performance improvements justify routine adoption.

Declining performance on temporal validation, threshold-sensitive trade-offs for rare outcomes, and the tension between sensitivity and specificity pose different obstacles to clinical use^[38,39,47]. Within the evidence chain in Figure 3, these obstacles relate to external validation, calibration and thresholds, and effects on decisions.

Studies of vascular involvement have advanced beyond single-center development. A three-dimensional deep-learning framework demonstrated external and prospective performance validation; a separate multicenter vascular-invasion study reported external-cohort discrimination, but its published methods did not verify all model-locking and data-isolation criteria required for Level 2^[16,20]. The next question is whether these models improve observers' agreement on vascular invasion and resectability during actual MDT practice. Studies should capture effects on further testing, consideration of neoadjuvant therapy, preparation for vascular reconstruction, and surgical selection. Initial silent implementation can characterize local calibration and failure patterns. Before-and-after comparisons, crossover interpretation studies, or cluster designs can then evaluate conventional versus model-assisted assessment. Endpoints should include the proportion of judgments altered, futile operations, readiness for vascular reconstruction, R0 resection rates, and discussion time, extending evaluation beyond accuracy. Work on three-dimensional planning should improve segmentation of small tumors and poorly defined margins while measuring manual correction, planning time, and operative outcomes. It should separately assess the additional value of digital models and physical printing. When a model prompts a change in the operative plan, comparison with intraoperative or pathological findings is needed to establish whether that change was helpful.

The main challenge for treatment-decision research is to test retrospective stratification hypotheses prospectively using feature-guided designs. Trial-derived transcriptomic signatures, predictors of ypN0 after therapy, and imaging biomarkers of fibrosis capture molecular sensitivity, pathological response, and the tumor microenvironment, respectively^[26-29]. They differ, however, in the populations studied, timing of sampling, and actions they might inform. Future studies should prespecify testing procedures, decision thresholds, and allocation rules. Prospective comparative designs are needed to determine whether model-assisted selection alters treatment completion, opportunities for resection, or long-term outcomes. Evaluation of transcriptomic and pathological features should also cover tissue adequacy, agreement between laboratory platforms, turnaround times, and assay failure rates, establishing whether results are available within the treatment window. Studies comparing benefit between regimens should prespecify a biomarker-treatment interaction in their statistical design; showing a prognostic association is insufficient. Establishing value for treatment selection requires direct comparison of feature-guided allocation with a conventional strategy.

Complication and prognostic models should move beyond stratifying risk to defining action pathways that can be tested. Models predicting POPF or VTE, lymph node metastasis or occult metastasis, and early recurrence or survival should report calibration, decision thresholds, predictive values, and decision curves for their intended population. Investigators should specify whether predictions inform risk communication, further testing, closer monitoring, or trial stratification^{[37,38,41,49]}. Every action needs clear criteria for initiation, duration, and cessation. For instance, a high-risk classification should not be given multiple unvalidated roles in postoperative monitoring, additional imaging, and research stratification. Tools intended to alter anticoagulation, drain management, adjuvant treatment, or surveillance could be tested in cluster-randomized, stepped-wedge, or embedded prospective studies against current practice. Assessment should capture harms caused by false-positive and false-negative results, resource consumption, workload for clinicians, and patient outcomes, rather than focusing only on small AUC differences.

Multimodal approaches can combine local tumor imaging, histology, molecular characteristics, and the patient's systemic condition. More modalities, however, do not automatically add clinical value. Published studies have combined body composition and clinical variables, CT findings and guideline-based information, or intratumoral and peritumoral radiomics and host features^[19,37,42]. These pairings suggest complementary information but add missing data, workflow dependence, and model complexity. Comparisons with unimodal models should use identical cohorts, endpoints, data partitions, and validation conditions. Reports should quantify gains in discrimination, shifts in calibration, net decision benefit, resilience when modalities are missing, and extra costs of acquisition and maintenance. Ablation analyses can establish which modalities account for gains, and comparing complete cases with those missing modalities can expose selection bias. Models also require a safe fallback when inputs from some modalities are unavailable. Multimodal deployment is warranted only if clinically meaningful net benefit over a simpler unimodal or baseline clinical model exceeds the extra cost and workflow burden. A simpler, more reproducible model may otherwise be the better practical choice.

Emerging AI approaches may help overcome several current limitations. Self-supervised learning extracts general representations from unlabeled images or pathology data, potentially reducing reliance on scarce high-quality labels and laborious annotation. Generative AI may assist in controlled augmentation of data, annotation of images or text, and checks for consistency. Safeguards are nevertheless needed to keep synthetic bias and incorrect information out of training datasets. Large language models may be more suitable for combining complex findings from imaging, pathology, laboratory tests, and longitudinal records, as well as supporting structured reporting. A supervised approach to structured reporting of pancreatic cancer resectability has already been investigated^[50]. Broader reviews of surgical AI similarly identify the move from prediction to image guidance, clinical workflows, and ongoing evaluation as a common challenge^[54,55]. Before these methods inform surgical decisions, safeguards should address factual accuracy, privacy, updates, traceability of prompts and versions, and responsibility for human review. Evaluation datasets should cover incomplete structured information, contradictory test findings, unusual anatomy, and post-treatment changes. Omissions, fabricated content, erroneous attribution, and unsuitable recommendations should be counted separately, rather than compressed into a text-similarity metric. After changes to a model or prompt, records of versions, input sources, and human revisions should be retained to enable error investigation and monitoring for performance drift. FUTURE-AI offers a framework for monitoring data drift, safety incidents, and fairness after deployment^[11]. The priority is to develop auditable decision-support systems that work within clinical processes, remain under continuous surveillance, and demonstrate value prospectively, beyond achieving a high score in a single evaluation.