Explainable artificial intelligence in pancreatic cancer prediction: from transparency to clinical decision-making

Background/Objectives: Pancreatic cancer (PC) remains among the most lethal malignancies worldwide, with a persistently low 5-year survival rate despite advances in systemic therapies and surgical innovation. Machine learning (ML) has emerged as a transformative tool for early detection, prognostic modelling, and treatment planning in PC, yet widespread clinical use is constrained by the “black box” nature of many models. Explainable artificial intelligence (XAI) offers a pathway to reconcile model accuracy with clinical trust, enabling transparent, reproducible, and clinically meaningful predictions.

Methods: We reviewed literature from 2020–2025, focusing on ML-based studies in PC that incorporated or discussed XAI techniques. Methods were grouped by model architecture, data modality, and interpretability framework. We synthesized findings to evaluate the technical underpinnings, interpretability outcomes, and clinical relevance of XAI applications.

Results: Across 21 studies on ML in PC, only three studies explicitly integrated XAI, primarily using SHAP and SurvSHAP. These methods helped identify key biomarkers, comorbidities, and survival predictors, while enhancing clinician trust. XAI approaches were categorized by staging (ante-hoc vs. post-hoc), compatibility (model-agnostic vs. model-specific), and scope (local vs. global explanations). Barriers to adoption included methodological instability, limited external validation, weak workflow integration, and lack of standardized evaluation.

Conclusions: XAI has the potential to serve as a cornerstone for advancing transparent, trustworthy ML in PC prediction. By clarifying model reasoning, XAI enhances clinical interpretability and regulatory readiness. This review provides a technical and clinical synthesis of current XAI practices, positioning explainability as essential for translating ML innovations into actionable oncology tools.

Introduction

Pancreatic cancer (PC) remains one of the most lethal malignancies worldwide, with a 5-year survival rate persistently below 10% despite advances in surgical techniques and systemic therapies (1–3) Globally, PC ranks as the 12th most common cancer, yet it is the third leading cause of cancer-related mortality in high-income countries such as Australia and the United States (4–6). In Saudi Arabia, PC represents a smaller but notable health burden (7). The incidence of PC has risen steeply in recent years, with reported new cases increasing from just 131 in 2005 to 579 in 2022 (8). Alarmingly, PC also carries the lowest five-year survival rate among all cancers in the Kingdom and is ranked as the eighth leading cause of cancer death (8, 9). Although its incidence remains lower compared to Western countries, the disease in Saudi patients often appears at advanced stages, reflecting global challenges of late detection and limited treatment options (8, 9). This disproportionate mortality burden underscores the urgent need for innovative approaches to improve early detection and patient outcomes.

The dismal prognosis of PC is primarily driven by late-stage diagnosis, aggressive tumour biology, and limited efficacy of available treatments (1, 2). Surgical resection combined with adjuvant chemotherapy remains the gold standard for long-term survival, but eligibility is generally restricted to patients with localised disease and good performance status (3, 6). Although neoadjuvant strategies and more radical surgical approaches have expanded resection possibilities to select stage III and even stage IV patients, the overall survival benefit remains modest, highlighting the necessity for earlier and more precise diagnosis (6).

Efforts to refine prognostic stratification have highlighted the potential of serum carbohydrate antigen 19-9 (CA19-9) as a biomarker for tailoring treatment intensity. Recent studies suggest that markedly elevated CA19–9 levels (>500 U/mL) in anatomically resectable PC may justify the use of intensive neoadjuvant chemotherapy (NAC), helping to identify subgroups more likely to benefit from systemic disease control and improved survival. However, the absence of standardised thresholds and inconsistent predictive performance of CA19–9 underscore a broader challenge: translating biomarkers into reliable, actionable tools for clinical decision-making (10).

No safe and effective population-level screening method currently exists for detecting PC at asymptomatic or early stages (11, 12). Conventional imaging modalities, including endoscopic ultrasonography (EUS), computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET), are hindered by high costs, limited sensitivity, and anatomical constraints associated with the pancreas’s retroperitoneal location (2, 3, 13–15). Consequently, microscopic premalignant lesions often evade detection, delaying diagnosis until advanced stages when therapeutic options are severely limited (2, 3).

The integration of computational intelligence into oncology has opened new avenues for overcoming diagnostic delays in PC (16, 17). Machine learning (ML), a core subset of artificial intelligence (AI), can leverage heterogeneous clinical, imaging, and molecular variables to detect patterns undetectable to human observation (18, 19). By incorporating demographic variables, patient history, laboratory markers, imaging features, and pathology findings, ML-based systems can support earlier diagnosis, improve risk stratification, and assist in treatment planning (20, 21). Unlike conventional rule-based diagnostics, ML can dynamically adapt to evolving datasets, potentially improving sensitivity and specificity in early disease diagnosis. This adaptability positions ML as a promising transformative technology to address one of the greatest challenges in PC care: late detection (2, 22). Despite promising advances, the real-world adoption of ML in PC remains limited by several challenges, including small and heterogeneous datasets, lack of external validation, and limited interpretability of model outputs. Clinical records are often incomplete or inconsistent, while imaging and biomarker datasets show substantial variability across institutions (2, 3, 22–24). Moreover, “black box” ML models, although accurate, frequently fail to provide clinicians with transparent explanations for their predictions, undermining trust and slowing regulatory approval. Given that even minor errors in oncology diagnostics can have serious consequences, ensuring model reliability, reproducibility, and interpretability is essential for safe clinical deployment (2, 23, 24).

Explainable artificial intelligence (XAI) offers a pathway to bridge the gap between high-performing models and clinical trust. XAI methods, such as Shapley Additive Explanations (SHAP), provide granular insights into model predictions by quantifying the relative contribution of individual features and aligning algorithmic reasoning with established clinical knowledge. Beyond supporting physician decision-making, XAI facilitates model refinement, bias detection, and regulatory compliance. In the context of PC, where treatment decisions often include complex trade-offs between surgical risk, chemotherapy tolerance, and disease progression, transparent AI systems hold the potential to advance precision and personalization of care (2, 23, 24). To our knowledge, this work is among the first technically rigorous review dedicated exclusively to the role of XAI in ML-based prediction of PC. By integrating perspectives from oncology, data science, and clinical informatics, this work provides a structured synthesis of XAI methods according to model architecture, data modality, and interpretability scope. The review critically appraises the clinical relevance, emphasising how enhanced transparency can strengthen oncologic decision-making and foster clinician trust. Furthermore, it identifies key translational barriers spanning methodological, operational, and regulatory domains, highlighting areas that require focused attention to enable the future integration of explainability into predictive pipelines. This work forms part of a dual-paper series, complemented by a companion article focused on feature engineering strategies and clinical integration of ML models for PC prediction, together intended to serve as a reference for advancing clinically actionable and trustworthy AI tools in pancreatic oncology.

Scope, objectives, and review methodology

This article is the first in a two-part series examining ML applications for PC prediction. While the companion paper focuses on feature engineering and clinical integration, the present review is dedicated exclusively to XAI, a critical yet underdeveloped component of translational AI in oncology. The objective of this review is to provide a structured synthesis of XAI methods applied to ML-based PC prediction, classify these techniques according to model architecture, data modality, and interpretability domain, and critically evaluate their capacity to support oncological decision-making. By centering on explainability, this article highlights both the promise and current limitations of XAI in PC prediction, while identifying methodological and regulatory challenges that must be addressed to enable safe clinical deployment.

A focused literature search was conducted using PubMed as the primary database and Google Scholar for supplementary retrieval. Due to institutional access constraints, subscription-based databases such as Scopus and Web of Science were not included. Although this may limit exhaustive coverage, the resulting dataset reflects the major peer-reviewed studies published between 2020–2025 relevant to ML and XAI in PC prediction. The search strategy, filtering steps, and eligibility criteria are summarised in Figure 1.

Figure 1

Figure 1. Flow diagram summarizing the literature selection process for ML studies in pancreatic cancer (2020–2025). From 177 initial records, 21 studies met the inclusion criteria, of which 3 incorporated XAI methods.

From an initial 177 records, removal of duplicates, restriction to full-text peer-reviewed articles, and screening for studies specifically addressing ML-based early prediction of PC resulted in 21 eligible ML studies, of which only three directly incorporated XAI methods. To contextualise this observation, we additionally present a temporal trend plot (Figure 2) illustrating the emergence of ML and XAI studies in PC from 2020 to 2025.

Figure 2

Figure 2. Temporal trend in ML and XAI studies for pancreatic cancer (2020–2025). Temporal trend (2020–2025) in machine-learning studies on pancreatic cancer and the smaller subset integrating XAI. Only 3 of 21 studies used XAI, despite a gradual increase in ML publications.

Taken together with its companion review, this work aims to serve as a comprehensive reference for researchers and clinicians developing interpretable, trustworthy, and clinically actionable ML models for PC.

Explainability in practice: XAI applications

Machine learning in oncology

ML has become a transformative force in oncology, providing data-driven solutions for early detection, prognostication, and treatment planning, particularly as the volume and complexity of medical data continue to expand. Broadly, ML methods are classified into supervised and unsupervised learning. Supervised learning algorithms, such as logistic regression (LR), support vector machines (SVM), random forests (RF), and gradient boosting (GB), use labelled datasets to predict known outcomes and are applied to both classification tasks (e.g., distinguishing malignant from benign lesions) and regression tasks (e.g., estimating survival probabilities or treatment responses). Unsupervised learning techniques, including principal component analysis (PCA), k-means clustering (KNN), Gaussian mixture models, density-based spatial clustering of applications with noise (DBSCAN), and balanced iterative reducing and clustering (BIRCH), operate on unlabelled data to uncover latent patterns, identify patient subgroups, and reduce dimensionality (22, 25–29). Deep learning (DL), a specialized branch of ML based on artificial neural networks (ANNs), expands these capabilities by using multi-layered architectures, such as convolutional and recurrent neural networks, to extract hierarchical representations from high-dimensional data. DL models have achieved state-of-the-art performance in oncology tasks such as tumour detection, segmentation, staging, mutation prediction, and radiomics-based analysis. Collectively, ML and DL have supported major advances in cancer susceptibility prediction, recurrence risk assessment, survival modelling, treatment toxicity forecasting, and genomic profiling (22, 25, 27, 30–32). These strengths are especially relevant to PC, where subtle early signals are often missed by conventional diagnostic pathways; by integrating demographics, comorbidities, biomarker trends, imaging features, and pathology findings, ML systems can identify high-risk individuals earlier and more consistently than rule-based approaches (1–34).

However, despite these advantages, several barriers limit the real-world application of ML in oncology, including heterogeneous and incomplete clinical datasets, limited external validation, challenges in multimodal data harmonisation, poor workflow integration, and the inherent opacity of many high-performing models (22, 24, 35, 36). These limitations underscore the urgent need for XAI, which can provide transparent, clinically aligned insights into model decision-making (1, 22, 23). The following sections therefore examine XAI frameworks most relevant to PC prediction and illustrate how explainability can bridge the gap between algorithmic performance and clinical usability.

Overview of XAI methods

Building on this momentum, the integration of XAI into clinical ML systems has shifted from a complementary option to a fundamental requirement. As ML models are increasingly applied in diagnostics, prognostics, and risk stratification (1, 37), explainability becomes essential for ensuring transparency, reliability, and clinical accountability (37–39). To contextualize this evolution, Figure 3 presents an overview of the XAI pipeline for PC prediction, summarizing how data sources, ML architectures, selected features and interpretability layers interact to yield clinically meaningful insights. In PC prediction, where decisions are high-stakes, clinicians must be able to trace and understand the rationale behind each risk score or treatment suggestion, whether derived from structured health records, imaging, or biomarker trends. Without such interpretability, even models with high predictive accuracy may be disregarded by healthcare professionals or encounter resistance from regulatory authorities (37–39).

Figure 3

Figure 3. Overview of the XAI pipeline for pancreatic cancer prediction. Sequential workflow of an XAI-enhanced ML pipeline for PC prediction, from data sources and modalities, data preprocessing and feature engineering, modeling architecture to predictions, post-hoc XAI interpretation, and resulting clinical insights.

To address these concerns, XAI introduces explanatory logic into the ML pipeline, enabling users to identify which input variables, such as biomarker levels, temporal disease sequences, or lab results, drive a given prediction. By enhancing transparency, XAI deepen model understanding, facilitates error analysis, supports fairness audits, and strengthens regulatory readiness. In complex clinical settings such as cancer care, where multidisciplinary collaborative decision-making is the norm, XAI provides each specialist, from radiologist to oncologist, with a rationale that can be interrogated and trusted (37–39). XAI methods are broadly categorised across three key dimensions that shape their operational behaviour and clinical suitability. The first dimension, staging, differentiates between ante-hoc and post-hoc methods. Ante-hoc (or intrinsic) explainability method is embedded directly within the model architecture itself, such as decision trees or rule-based models, making the model inherently interpretable. In contrast, post-hoc methods are applied after model training, serving to approximate or uncover the reasoning behind predictions from complex, dense architectures such as deep neural networks (DNNs). The second dimension, model compatibility, assesses whether an explanation method is model-agnostic, meaning it can be applied to any ML algorithm regardless of structure (e.g., Local Interpretable Model-agnostic Explanations (LIME) or SHAP), or model-specific, tailored to specific architectures such as CNNs (e.g., Gradient-weighted Class Activation Mapping (GradCAM)). The third dimension relates to the scope of explanation, differentiates between local methods, which provide interpretability for individual predictions and are critical for case-by-case clinical decision-making, and global methods, which explain and characterise the overall behaviour of the model, offering insights into feature importance and decision trends across the dataset (37, 39, 40). Table 1 outlines a comparative landscape of widely used XAI techniques, highlighting their classification across these dimensions, their technical underpinnings, and representative healthcare applications. Understanding this framework is essential for choosing an appropriate XAI strategy tailored to the clinical task at hand, ensuring that the resulting model is not only accurate but also transparent, justifiable, defensible, and aligned with the expectations of clinicians, patients, and regulators (39).

Table 1

Table 1. Overview of widely used XAI techniques: staging, scope, and clinical use cases.

Among the widely adopted XAI methods, SHAP has emerged as a reference standard for clinical model interpretation, offering both local and global explanations through additive feature attributions grounded in cooperative game theory. While SHAP values are inherently local, quantifying the contribution of each feature to an individual prediction, they can be aggregated across patients to yield global insights into model behaviour. Aggregated SHAP summaries, such as mean absolute SHAP values or global importance plots, enable population-level interpretability and highlight consistently influential predictors across the cohort (39–41). For example, a large-scale oncological study utilised aggregated SHAP values to identify global biomarker patterns influencing colorectal-cancer risk. More recently, Almisned et al. (2), applied SHAP in an ensemble ML framework for early detection of PC; by presenting SHAP-based global feature rankings alongside patient-level attributions, the authors demonstrated how SHAP can inform both individual clinical decisions and cohort-wide predictor assessment. These studies illustrate how SHAP transcends strictly local explanation to enable global model interpretability in clinical-AI applications (37, 39).

LIME, although less stable, generates local surrogate models around specific predictions and its broadly model-agnostic applicability. For imaging applications, GradCAM highlights key regions within radiological scans that drive classifications, supporting visual diagnostic interpretation. PDPs provide global insights into the marginal effect of individual features but are limited by their assumption of feature independence and vulnerability to feature correlation. DeepLIFT, optimised for deep architectures and networks, delivers efficient layer-wise attributions by comparing neuron activations, and has proven useful in complex biomedical domains such as genomics (39–41).In practice, post-hoc approaches dominate clinical applications due to their flexibility; they can be applied without retraining the models and generate outputs (e.g., feature rankings, heatmaps, or rule sets) that align with clinical reasoning. However, choosing the right method remains nontrivial; as techniques and approaches vary in assumptions, stability, and interpretive depth. While some methods are inherently interpretable (ante-hoc), others retrospectively approximate explanation (post-hoc); some offer patient-level insights (local), while others describe model-wide behaviour (global) (37, 39).

This section aimed to equip researchers and practitioners with a foundational understanding of XAI tools most applicable to PC prediction. Clarifying these distinctions enhances transparency, supports responsible deployment, and ultimately bridges the gap between ML innovations and actionable clinical insight.

Studies utilising XAI in PC prediction

XAI has become increasingly critical in healthcare, particularly for high-stakes decisions such as cancer diagnosis and prognosis. In our structured review of 21 studies on PC prediction, only three explicitly applied XAI methods (2, 41, 42), with one further study relying on feature importance analysis of ICD codes (33). These XAI applications, principally SHAP and SurvSHAP, illustrate how interpretability can strengthen clinical trust and yield actionable insights (2, 41, 42). A detailed comparison of the three studies that integrated XAI into ML-based PC prediction is provided in Table 2. A full overview of all 21 studies is provided in Supplementary Table S1, with details on model type, data source, and feature design. Here, however, we focus on the three studies that directly integrated XAI, while feature-engineering-centred works, clinical variables, and integration into care pathways are addressed in a forthcoming companion review. Taken together, the two papers offer a comprehensive assessment of how ML can be made both interpretable and clinically actionable in the context of PC.

Table 2

Table 2. Studies integrating XAI methods in pancreatic cancer (PC) prediction.

In a recent study by Almisned et al. (2), researchers proposed an ensemble ML framework for early PC detection using clinical and biomarker features. Six ML algorithms were evaluated alongside an ensemble voting classifier, with SHAP applied for model interpretation. SHAP was used to quantify and visualise the contribution of each feature to the model’s predictions. The top predictive variables identified by SHAP included benign sample diagnosis, TFF1, and LYVE1 as top predictors with strong positive influence on early-stage PC prediction. Notably, SHAP suggested interpretable thresholds and feature relevance that clinicians could interpret and validate. For instance, elevated LYVE1 levels were consistently associated with malignancy, suggesting the need for targeted interventions such as non-invasive imaging or liquid biopsies. Furthermore, SHAP also enabled patient-level risk assessments, reinforcing its value for multidisciplinary decision-making in designing tailored treatment plans. The study emphasised that model interpretability is not merely vital for regulatory transparency but also for clinical adoption, reinforcing XAI’s function in developing trust and informed decision-making in oncology settings (2).

Another study, Keyl et al. (42), examined the application of SHAP within a random survival forest (RSF) model trained on clinical data from 203 patients with advanced pancreatic ductal adenocarcinoma (PDAC). Baseline predictors included CA19-9, C-reactive protein (CRP), neutrophil-to-lymphocyte ratio (NLR), age, and metastatic status. SHAP analysis revealed CRP and NLR as the dominant and most influential predictors of poor survival, followed by age and CA19-9. In contrast, higher serum protein levels and absence of metastasis disease (M0 status) were associated with improved survival outcomes. SHAP enabled transparent visualisation of each variable’s directional influence on survival predictions, enhancing clinician understanding and guiding more nuanced prognostic evaluations. This study validates the potential of XAI not just in classification tasks but also in survival modelling critical to advanced-stage cancer management (42).

SurvSHAP visualisations uncovered heterogeneous predictive logic across models, increasing transparency and confidence in outputs, with implications for targeted screening strategies.

In an effort to improve early detection of sporadic PC, another study by Chen et al. (41) compared the predictive performance of RSF, eXtreme gradient boosting (XGB), and Cox proportional hazards regression models using electronic health record (EHR) data from two main US healthcare systems. The study incorporated SurvSHAP, an XAI method specifically tailored for survival models, to explain feature contribution across algorithms. Age emerged as the most influential predictor across all three models, while abdominal pain contributed minimally in the RSF and XGB models but was more prominent in Cox regression model. SurvSHAP visualisations uncovered heterogeneous predictive logic across models, increasing transparency and confidence in model outputs, with implications for targeted screening strategies (41).

Despite the proliferation of ML-based approaches in PC prediction, the adoption and integration of XAI remains limited. The few studies that integrated XAI, primarily through SHAP and its survival adaptation SurvSHAP, demonstrated how interpretability can enhance model transparency, clinical trust, and actionability. These studies reflect the growing recognition that interpretability is not optional but a core necessity for translating AI solutions into meaningful clinical practice (2, 41, 42).

Interpretability and explainability in clinical AI: conceptual boundaries, cross-domain perspectives, and relevance to pancreatic cancer

The distinction between interpretability and explainability is fundamental to evaluating the transparency, safety, and trustworthiness of AI systems in high-stakes healthcare applications. Although these terms are often used interchangeably across the XAI literature, leading scholars have emphasized that they embody distinct, and sometimes competing, objectives (38, 43–46). We explicitly align our definitions with established frameworks while acknowledging that different fields conceptualize these terms differently. Our synthesis therefore represents both a consolidation of prior literature and a clinical contextualization tailored to PC prediction.

Interpretability refers to models whose internal structure is directly understandable by humans without requiring external explanatory tools. Interpretability is achieved through domain-aligned constraints, such as sparsity, rule-based logic, monotonicity, or additive structure, that make predictions transparent and traceable. Examples include LR with clinically meaningful coefficients, decision trees, Certifiably Optimal Rule Lists (CORELS), and generalized additive models (GAMs) (43–45, 47). In oncology, such models enable clinicians to confirm that predictors align with biological and epidemiological knowledge, e.g., elevated CA19–9 or rapid weight loss increasing PC risk, thereby supporting real-time auditing, error detection, and the integration of contextual cues absent from EHR data (10). This advantage parallels findings in other high-stakes fields, including criminal justice, materials engineering, and air-quality forecasting, where interpretable models have repeatedly prevented bias, confounding, and model misuse (43).

Explainability, in contrast, refers to post hoc analytic techniques that attempt to summarize or approximate the behaviour of complex, non-transparent (“black-box”) models after training. Methods such as SHAP, SurvSHAP, LIME, Grad-CAM, Anchors, and Partial Dependence Plots (PDP) generate feature-level or instance-level explanations without exposing the model’s internal computations (45, 47). Some of these tools are widely used in PC prediction because high-performing models, such as XGBoost, random forests, or DNNs, often sacrifice inherent transparency to maximize predictive accuracy (2, 41, 48). However, as Rudin (43) argues, post hoc explanations cannot fully replicate a black-box model’s reasoning; if an explanation perfectly captured the model, the model itself would be unnecessary (43, 44). This fidelity gap is particularly consequential in medicine, where misaligned or incomplete explanations can mislead clinicians, obscure confounding, or mask algorithmic biases.

To reconcile these perspectives, we present interpretability, explainability, and completeness as overlapping but non-identical constructs (Figure 4). Interpretability focuses on human comprehension of the model’s inherent logic. Explainability provides approximations of the reasoning process for otherwise opaque models. Completeness describes the degree to which either approach faithfully captures the true computational or causal pathways (49, 50).

Figure 4

Figure 4. Interpretability, explainability, and completeness in clinical AI.

This conceptual model aligns with the XAI distinctions articulated (43), recent engineering literature (45), and foundational ML theory (44), while acknowledging that our synthesis is tailored to clinical decision-making.

Cross-domain comparisons also illustrate that the stakes differ by discipline. In materials science and nanoporous materials modelling, for example, SHAP is routinely used as a global and local post hoc tool to analyse physics-related feature importance despite inherent limitations in capturing mechanistic pathways (37, 38, 45). In contrast, medical domains, where decisions directly affect patient outcomes, emphasize accountability, causal faithfulness, fairness, and regulatory acceptance (37, 38). This divergence reinforces the argument that explainability alone may be insufficient for clinical deployment, yet practically necessary when high-performing black-box models are unavoidable.

In PC, these distinctions have concrete implications. Interpretable models allow clinicians to validate risk factors, temporal disease signatures, and biomarker behaviour, strengthening trust and facilitating uptake into clinical pathways. Explainability tools, such as SHAP and SurvSHAP, allow high-accuracy black-box models to be interrogated for biological plausibility, temporal trends, and subgroup bias, critical for early detection where false reassurance or misinterpretation carries significant risk. Together, these perspectives highlight that a hybrid strategy, prioritizing interpretable models when feasible and applying rigorous, high-fidelity post hoc XAI when necessary, provides the most robust pathway toward clinically actionable and ethically sound AI for PC prediction.

Gaps in interpretability and physician trust

While ML holds considerable promise for healthcare, its integration into clinical practice remains constrained by persistent challenges in interpretability and physician trust. Interpretability is not optional; it is foundational for responsible adoption. Yet many high-performing models often operate as black boxes, composed of deeply layered, nonlinear structures with millions of parameters, making their decision processes inherently opaque. Even with the aid of XAI tools, explanations often remain partial: attribution heatmaps may highlight influential features without explaining their clinical significance. For frontline clinicians lacking advanced ML or statistical training, these explanations may remain inaccessible or insufficiently actionable, widening the gap between developers and end-users (37–40).

Trust is further undermined by the inconsistency of some XAI methods, whose explanations can be unstable or misleading. Several studies question the reliability of post-hoc explanations, citing their susceptibility to instability or misrepresentation. The lack of standardised metrics for assessing explanation quality worsens this issue, limiting comparison across clinical contexts (37, 38).

At the system level, barriers such as limited infrastructure, poor EHR integration, non-intuitive interfaces, and shifting regulatory standards further erode clinician confidence in AI-driven recommendations, particularly in time-sensitive scenarios (37–39).

Nonetheless, XAI remains central to fostering trust. Techniques such as SHAP and LIME can help translate complex predictions into clinically interpretable elements, enabling clinicians to question, validate, or contextualise predictions based on patient-specific data. This transparency reframes the clinician–AI relationship from one of scepticism to collaboration. Bridging the gap between interpretability and physician trust will, however, demand more than technical advances. It essentially requires aligning XAI outcomes with clinical reasoning, simplifying user interaction, and developing rigorous evaluation standards. Only through such multidimensional progress can AI move beyond theoretical promise and become a trusted tool in routine care (37–39).

Discussion and future directions

The integration of XAI into ML-based prediction of PC represents a pivotal opportunity to bridge the gap between computational innovation and clinical translation. While conventional models achieve strong predictive performance, their opacity undermines physician trust, regulatory approval, and ultimately, real-world clinical deployment (2, 23, 24). Our synthesis highlights that XAI is increasingly regarded as a foundational requirement for precision oncology. However, the review also uncovers limitations that must be addressed to achieve clinical maturity.

First, the adoption of XAI in PC prediction is limited, with only a handful of studies explicitly incorporating interpretability frameworks such as SHAP or SurvSHAP (2, 41, 42). Even within these applications, explanations are typically restricted to surface-level feature attributions, rarely aligned with pathophysiological knowledge. Bridging this gap requires moving beyond feature rankings to clinically grounded narratives that link model predictions to biological plausibility and treatment relevance.

Second, methodological fragility is pressing concern. Many post-hoc techniques are unstable, with outputs sensitive to minor data perturbations, undermining reproducibility. Developing standardised benchmarks for explanation quality, analogous to accuracy or area under the curve (AUC) metrics for predictive performance, is essential for robust clinical evaluation (23, 37).

Future work should also prioritise scalability and external validation. Current studies are largely confined to single-institution cohorts or limited datasets, which restricts generalisability. Cross-institutional federated learning frameworks, combined with XAI, may strengthen robustness while preserving data privacy (22, 24, 35). Furthermore, integration into EHR systems and clinical workflows is similarly underdeveloped. Effective deployment will require intuitive, clinician-facing interfaces where model explanations are seamlessly presented alongside conventional diagnostic tools. Such integration must align with evolving regulatory standards that increasingly demand transparency in algorithmic decision-making (37, 38, 41, 51).

Multimodal explainability also demands attention. PC prediction increasingly involves EHR variables, biomarkers, imaging, and genomics. Future work should aim to unify explanations across modalities, enabling a holistic understanding of patient-level insights and supporting collaborative decision-making in multidisciplinary oncology boards (33, 52, 53).

Finally, physician trust must remain central to future directions. Technical innovation in explainability will have limited impact unless it resonates with clinician reasoning and patient communication. Human-centred design, co-development with end-users, and iterative feedback between data scientists and healthcare providers will be indispensable for aligning XAI with clinical decision-making (2, 37, 38, 41, 51). Such multidimensional approach, blending methodological rigour, clinical grounding, and human-centric design, will be essential if explainable ML is to move from proof-of-concept to a transformative tool in PC care.

Conclusion

This review establishes explainability as a cornerstone for translating ML innovations into clinically actionable tools for PC prediction. By appraising XAI methods, we demonstrate their potential to enhance transparency, refine prognostic reasoning, and build clinician trust, while also outlining the barriers that currently hinder their adoption. The limited number of studies explicitly applying XAI underscores both the novelty of this domain and the urgency of advancing it. Future research must prioritise methodological robustness, multimodal integration, external validation, and user-centred deployment to unlock the full clinical potential of explainable ML.

As the first article in a two-part series, this review defines the technical and clinical landscape of XAI in PC prediction, while its companion paper addresses feature engineering and integration strategies. Together, they provide a comprehensive roadmap for developing interpretable, trustworthy, and impactful AI solutions in oncology.

In parallel, we are conducting an IRB-approved study (NRR25/67/3, KAIMRC) that aims to develop an ML-based model for predicting the risk of PC in patients with chronic metabolic disorders, using data from local EHRs. The ultimate aim of this project is to support early diagnosis of PC and targeted surveillance, improving patient outcomes. Importantly, the integration of XAI will ensure that this predictive framework remains interpretable and clinically trustworthy, directly addressing the translational gaps highlighted in this review.

Author contributions

WA: Funding acquisition, Methodology, Formal Analysis, Data curation, Conceptualization, Resources, Visualization, Writing – original draft. AA: Validation, Project administration, Conceptualization, Writing – review & editing, Supervision.

Funding

The author(s) declared financial support was received for this work and/or its publication. This research received no external funding. The project was conducted under the approval of the King Abdullah International Medical Research Center (KAIMRC; IRB #00000134225). The Article Processing Charges (APCs) will be fully covered by KAIMRC upon acceptance.

Acknowledgments

The authors thank the King Abdullah International Medical Research Center (KAIMRC) for institutional support.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fonc.2025.1720039/full#supplementary-material

Supplementary Table 1 | Structured summary of included studies.

Abbreviations

AI, Artificial Intelligence; AUROC, Area Under the Receiver Operating Characteristic Curve; CA19-9, Carbohydrate Antigen 19-9; CT, Computed Tomography; DL, Deep Learning; EHRs, Electronic Health Records; EUS, Endoscopic Ultrasonography; FOLFIRINOX, Fluorouracil, Irinotecan, and Oxaliplatin (chemotherapy regimen); LYVE1, Lymphatic vessel endothelial hyaluronan receptor 1; ML, Machine Learning; MRI, Magnetic Resonance Imaging; NAC, Neoadjuvant Chemotherapy; OS, Overall Survival; PC, Pancreatic Cancer; PDAC, Pancreatic ductal adenocarcinoma; PET, Positron Emission Tomography; PFS, Progression-Free Survival; RSF, Random survival forests; SHAP, Shapley Additive Explanations; TFF1, Trefoil Factor 1; XAI, eXplainable Artificial Intelligence

References

1. Láinez Ramos-Bossini AJ, Gámez Martínez A, Luengo Gómez D, Valverde-López F, Morillo Gil AJ, González Flores E, et al. Computed tomography-based sarcopenia and pancreatic cancer survival—A comprehensive meta-analysis exploring the influence of definition criteria, prevalence, and treatment intention. Cancers (Basel). (2025) 17:607. doi: 10.3390/cancers17040607

PubMed Abstract | Crossref Full Text | Google Scholar

2. Almisned FA, Usanase N, Ozsahin DU, and Ozsahin I. Incorporation of explainable artificial intelligence in ensemble machine learning-driven pancreatic cancer diagnosis. Sci Rep. (2025) 15:14038. doi: 10.1038/s41598-025-98298-0

PubMed Abstract | Crossref Full Text | Google Scholar

3. Vaiyapuri T, Dutta AK, Punithavathi ISH, Duraipandy P, Alotaibi SS, Alsolai H, et al. Intelligent deep-learning-enabled decision-making medical system for pancreatic tumor classification on CT images. Healthcare. (2022) 10:677. doi: 10.3390/healthcare10040677

PubMed Abstract | Crossref Full Text | Google Scholar

4. Australian Institute of Health and Welfare. Cancer data in Australia. Canberra: Australian Institute of Health and Welfare (2024). Available at: https://www.aihw.gov.au/reports/cancer/cancer-data-in-australia.

Google Scholar

5. Siegel RL, Kratzer TB, Giaquinto AN, Sung H, and Jemal A. Cancer statistics, 2025. CA Cancer J Clin. (2025) 75:10–45. doi: 10.3322/caac.21871

PubMed Abstract | Crossref Full Text | Google Scholar

6. Lockie EB, Sylivris A, Pandanaboyana S, Zalcberg J, Skandarajah A, and Loveday BP. Relationship between pancreatic cancer resection rate and survival at population level: systematic review. BJS Open. (2025) 9. doi: 10.1093/bjsopen/zraf007

PubMed Abstract | Crossref Full Text | Google Scholar

7. AlGhamdi HJ, Alfaifi SA, Alolayan AA, Musaad SMA, and Jazieh AM. Pancreatic cancer in Saudi patients treated at tertiary institution. Ten years retrospective study. Saudi Med J. (2013) 34:604–8. doi: 10.1200/jco.2011.29.15_suppl.e14520

Crossref Full Text | Google Scholar

8. Badheeb AM, Awad MA, Al Masad AG, Alyami MS, Fagihi MA, Al Walani M, et al. Pancreatic cancer: A retrospective study from the Najran region of Saudi Arabia. Cureus. (2024) 16:e65685. doi: 10.7759/cureus.65685

PubMed Abstract | Crossref Full Text | Google Scholar

9. Elwali NE, AlShareef SM, Khamis AH, and Elhassan MMA. Pancreatic cancer in Saudi Arabia (2005–2020): increasing trend. BMC Cancer. (2024) 24:653. doi: 10.1186/s12885-024-12401-8

PubMed Abstract | Crossref Full Text | Google Scholar

10. Omiya K, Oba A, Sugiura K, Maekawa A, Mie T, Kobayashi K, et al. Resectable pancreatic cancer with CA19-9 > 500 U/mL: A biological indicator for survival benefit with intensive neoadjuvant chemotherapy. Ann Surg Oncol. (2025) 32:5411–20. doi: 10.1245/s10434-025-17407-5

PubMed Abstract | Crossref Full Text | Google Scholar

11. Cao K, Xia Y, Yao J, Han X, Lambert L, Zhang T, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med. (2023) 29:3033–43. doi: 10.1038/s41591-023-02640-w

PubMed Abstract | Crossref Full Text | Google Scholar

12. Yang J, Xu R, Wang C, Qiu J, Ren B, and You L. Early screening and diagnosis strategies of pancreatic cancer: a comprehensive review. Cancer Commun. (2021) 41:1257–74. doi: 10.1002/cac2.12204

PubMed Abstract | Crossref Full Text | Google Scholar

13. Debnath P, Ata NA, Anton CG, Kocaoglu M, Zhou Y, Zhang B, et al. Associations between MRI and endoscopic ultrasound findings in the pancreas in children and young adults with suspected chronic pancreatitis. Pancreatology. (2025) 25:641–7. doi: 10.1016/j.pan.2025.07.002

PubMed Abstract | Crossref Full Text | Google Scholar

14. Li J, Fu C, Zhao S, Pu Y, Yang F, Zeng S, et al. The progress of PET/MRI in clinical management of patients with pancreatic Malignant lesions. Front Oncol. (2023) 13. doi: 10.3389/fonc.2023.920896

PubMed Abstract | Crossref Full Text | Google Scholar

15. Pawelec N, Durko Ł, and Małecka-Wojciesko E. Changes connected to early chronic pancreatitis and early pancreatic cancer in endoscopic ultrasonography (EUS): clinical implications. Cancers (Basel). (2025) 17:1891. doi: 10.3390/cancers17111891

PubMed Abstract | Crossref Full Text | Google Scholar

16. Jin D, Khan NU, Gu W, Lei H, Goel A, and Chen T. Informatics strategies for early detection and risk mitigation in pancreatic cancer patients. Neoplasia. (2025) 60:101129. doi: 10.1016/j.neo.2025.101129

PubMed Abstract | Crossref Full Text | Google Scholar

17. Nagaraju GP, Sandhya T, Srilatha M, Ganji SP, Saddala MS, and El-Rayes BF. Artificial intelligence in gastrointestinal cancers: Diagnostic, prognostic, and surgical strategies. Cancer Lett. (2025) 612:217461. doi: 10.1016/j.canlet.2025.217461

PubMed Abstract | Crossref Full Text | Google Scholar

18. Prasad VK, Verma A, Bhattacharya P, Shah S, Chowdhury S, Bhavsar M, et al. Revolutionizing healthcare: a comparative insight into deep learning’s role in medical imaging. Sci Rep. (2024) 14:30273. doi: 10.1038/s41598-024-71358-7

PubMed Abstract | Crossref Full Text | Google Scholar

19. Maleki Varnosfaderani S and Forouzanfar M. The role of AI in hospitals and clinics: transforming healthcare in the 21st century. Bioengineering. (2024) 11:337. doi: 10.3390/bioengineering11040337

PubMed Abstract | Crossref Full Text | Google Scholar

20. Dixon D, Sattar H, Moros N, Kesireddy SR, Ahsan H, Lakkimsetti M, et al. Unveiling the influence of AI predictive analytics on patient outcomes: A comprehensive narrative review. Cureus. (2024) 16:e59954. doi: 10.7759/cureus.59954

PubMed Abstract | Crossref Full Text | Google Scholar

21. Rani S, Kumar R, Panda BS, Kumar R, Muften NF, Abass MA, et al. Machine learning-powered smart healthcare systems in the era of big data: applications, diagnostic insights, challenges, and ethical implications. Diagnostics. (2025) 15:1914. doi: 10.3390/diagnostics15151914

PubMed Abstract | Crossref Full Text | Google Scholar

22. Santos CS and Amorim-Lopes M. Externally validated and clinically useful machine learning algorithms to support patient-related decision-making in oncology: a scoping review. BMC Med Res Methodol. (2025) 25:45. doi: 10.1186/s12874-025-02463-y

PubMed Abstract | Crossref Full Text | Google Scholar

23. Guo Z, Zhang Z, Liu L, Zhao Y, Liu Z, Zhang C, et al. Explainable machine learning for predicting lung metastasis of colorectal cancer. Sci Rep. (2025) 15:13611. doi: 10.1038/s41598-025-98188-5

PubMed Abstract | Crossref Full Text | Google Scholar

24. Fountzilas E, Pearce T, Baysal MA, Chakraborty A, and Tsimberidou AM. Convergence of evolving artificial intelligence and machine learning techniques in precision oncology. NPJ Digit Med. (2025) 8:75. doi: 10.1038/s41746-025-01471-y

PubMed Abstract | Crossref Full Text | Google Scholar

25. Bibault JE, Giraud P, and Burgun A. Big Data and machine learning in radiation oncology: State of the art and future prospects. Cancer Lett. (2016) 382:110–7. doi: 10.1016/j.canlet.2016.05.033

PubMed Abstract | Crossref Full Text | Google Scholar

26. Jensen PB, Jensen LJ, and Brunak S. Mining electronic health records: towards better research applications and clinical care. Nat Rev Genet. (2012) 13:395–405. doi: 10.1038/nrg3208

PubMed Abstract | Crossref Full Text | Google Scholar

27. Shen L, Kann BH, Taylor RA, and Shung DL. The clinician’s guide to the machine learning galaxy. Front Physiol. (2021) 12. doi: 10.3389/fphys.2021.658583

PubMed Abstract | Crossref Full Text | Google Scholar

28. Jordan MI and Mitchell TM. Machine learning: Trends, perspectives, and prospects. Science. (2015) 349:255–60. doi: 10.1126/science.aaa8415

PubMed Abstract | Crossref Full Text | Google Scholar

29. Kourou K, Exarchos TP, Exarchos KP, Karamouzis MV, and Fotiadis DI. Machine learning applications in cancer prognosis and prediction. Comput Struct Biotechnol J. (2015) 13:8–17. doi: 10.1016/j.csbj.2014.11.005

PubMed Abstract | Crossref Full Text | Google Scholar

30. LeCun Y, Bengio Y, and Hinton G. Deep learning. Nature. (2015) 521:436–44. doi: 10.1038/nature14539

PubMed Abstract | Crossref Full Text | Google Scholar

31. Krogh A. What are artificial neural networks? Nat Biotechnol. (2008) 26:195–7. doi: 10.1038/nbt1386

PubMed Abstract | Crossref Full Text | Google Scholar

32. Nagendran M, Chen Y, Lovejoy CA, Gordon AC, Komorowski M, Harvey H, et al. Artificial intelligence versus clinicians: systematic review of design, reporting standards, and claims of deep learning studies. BMJ. (2020) 368:m689. doi: 10.1136/bmj.m689

PubMed Abstract | Crossref Full Text | Google Scholar

33. Placido D, Yuan B, Hjaltelin JX, Zheng C, Haue AD, Chmura PJ, et al. A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories. Nat Med. (2023) 29:1113–22. doi: 10.1038/s41591-023-02332-5

PubMed Abstract | Crossref Full Text | Google Scholar

34. Amri F, Belkhayat C, Yeznasni A, Koulali H, Jabi R, Zazour A, et al. Association between pancreatic cancer and diabetes: insights from a retrospective cohort study. BMC Cancer. (2023) 23:856. doi: 10.1186/s12885-023-11344-w

PubMed Abstract | Crossref Full Text | Google Scholar

35. Jeong C, Goh Y, and Kwak J. Challenges and opportunities to integrate artificial intelligence in radiation oncology: a narrative review. Ewha Med J. (2024) 47. doi: 10.12771/emj.2024.e49

PubMed Abstract | Crossref Full Text | Google Scholar

36. Lipkova J, Chen RJ, Chen B, Lu MY, Barbieri M, Shao D, et al. Artificial intelligence for multimodal data integration in oncology. Cancer Cell. (2022) 40:1095–110. doi: 10.1016/j.ccell.2022.09.012

PubMed Abstract | Crossref Full Text | Google Scholar

37. Adeniran AA, Onebunne AP, and William P. Explainable AI (XAI) in healthcare: Enhancing trust and transparency in critical decision-making. World J Advanced Res Rev. (2024) 23:2447–658. doi: 10.30574/wjarr.2024.23.3.2936

Crossref Full Text | Google Scholar

38. Sadeghi Z, Alizadehsani R, CIFCI MA, Kausar S, Rehman R, Mahanta P, et al. A review of Explainable Artificial Intelligence in healthcare. Comput Electrical Engineering. (2024) 118:109370. doi: 10.1016/j.compeleceng.2024.109370

Crossref Full Text | Google Scholar

39. Gupta J and Seeja KR. A comparative study and systematic analysis of XAI models and their applications in healthcare. Arch Comput Methods Eng. (2024) 31:3977–4002. doi: 10.1007/s11831-024-10103-9

Crossref Full Text | Google Scholar

40. Chen V, Yang M, Cui W, Kim JS, Talwalkar A, and Ma J. Applying interpretable machine learning in computational biology—pitfalls, recommendations and opportunities for new developments. Nat Methods. (2024) 21:1454–61. doi: 10.1038/s41592-024-02359-7

PubMed Abstract | Crossref Full Text | Google Scholar

41. Chen W, Zhou B, Jeon CY, Xie F, Lin YC, Butler RK, et al. Machine learning versus regression for prediction of sporadic pancreatic cancer. Pancreatology. (2023) 23:396–402. doi: 10.1016/j.pan.2023.04.009

PubMed Abstract | Crossref Full Text | Google Scholar

42. Keyl J, Kasper S, Wiesweg M, Götze J, Schönrock M, Sinn M, et al. Multimodal survival prediction in advanced pancreatic cancer using machine learning. ESMO Open. (2022) 7:100555. doi: 10.1016/j.esmoop.2022.100555

PubMed Abstract | Crossref Full Text | Google Scholar

43. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. (2019) 1:206–15. doi: 10.1038/s42256-019-0048-x

PubMed Abstract | Crossref Full Text | Google Scholar

44. Rudin C, Chen C, Chen Z, Huang H, Semenova L, and Zhong C. Interpretable machine learning: Fundamental principles and 10 grand challenges. Stat Surv. (2022) 16. doi: 10.1214/21-SS133

Crossref Full Text | Google Scholar

45. Ghasemi A and Barisik M. Machine learning for thermal transport prediction in nanoporous materials: progress, challenges, and opportunities. Nanomaterials. (2025) 15:1660. doi: 10.3390/nano15211660

PubMed Abstract | Crossref Full Text | Google Scholar

46. Molnar C. Interpretable machine learning: a guide for making black box models explainable, 3rd ed. (2025). 379 p. Available at: https://christophm.github.io/interpretable-ml-book.

Google Scholar

47. Barredo Arrieta A, Díaz-Rodríguez N, Del Ser J, Bennetot A, Tabik S, Barbado A, et al. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Inf Fusion. (2020) 58:82–115. doi: 10.1016/j.inffus.2019.12.012

Crossref Full Text | Google Scholar

48. Alkhanbouli R, Matar Abdulla Almadhaani H, Alhosani F, and Simsekler MCE. The role of explainable artificial intelligence in disease prediction: a systematic literature review and future research directions. BMC Med Inform Decis Mak. (2025) 25:110. doi: 10.1186/s12911-025-02944-6

PubMed Abstract | Crossref Full Text | Google Scholar

49. Murdoch WJ, Singh C, Kumbier K, Abbasi-Asl R, and Yu B. Definitions, methods, and applications in interpretable machine learning. Proc Natl Acad Sci. (2019) 116:22071–80. doi: 10.1073/pnas.1900654116

PubMed Abstract | Crossref Full Text | Google Scholar

50. Doshi-Velez F and Kim B. Towards A rigorous science of interpretable machine learning. arXiv: Machine Learning. (2017). doi: 10.48550/arXiv.1702.08608

Crossref Full Text | Google Scholar

51. Chi H, Chen H, Wang R, Zhang J, Jiang L, Zhang S, et al. Proposing new early detection indicators for pancreatic cancer: Combining machine learning and neural networks for serum miRNA-based diagnostic model. Front Oncol. (2023) 13. doi: 10.3389/fonc.2023.1244578

PubMed Abstract | Crossref Full Text | Google Scholar

52. Huang W, Xu Y, Li Z, Li J, Chen Q, Huang Q, et al. Enhancing noninvasive pancreatic cystic neoplasm diagnosis with multimodal machine learning. Sci Rep. (2025) 15:16398. doi: 10.1038/s41598-025-01502-4

PubMed Abstract | Crossref Full Text | Google Scholar

53. Lee D, Lee C, Han K, Goo T, Kim B, Han Y, et al. Machine learning models for pancreatic cancer diagnosis based on microbiome markers from serum extracellular vesicles. Sci Rep. (2025) 15:10995. doi: 10.1038/s41598-025-94183-y

PubMed Abstract | Crossref Full Text | Google Scholar