The place of advanced machine learning techniques in building pancreatic adenocarcinoma survival and recurrence prognosis models

Background: Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal malignancy, and traditional prognostic methods, such as TNM staging, often fail to accurately predict outcomes. This review evaluates the use of machine learning (ML) to improve PDAC prognosis.

Methods: A systematic literature search of PubMed and Google Scholar was conducted, identifying 12 studies that applied ML algorithms to predict survival, recurrence, and metastasis in patients with PDAC.

Results: Various algorithms, including Random Forests, XGBoost, and Deep Learning, demonstrated superior predictive performance compared to the TNM staging. Models using multimodal data—combining clinical, radiomic, and genomic features-yielded the highest accuracy for predicting overall survival and early liver metastasis.

Conclusion: ML offers a significant advantage in analyzing complex medical data to refine risk stratification and support personalized PDAC treatment. However, current models are limited by their small datasets and retrospective designs. Future research requires prospective validation to translate these ML tools into clinical practice.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is one of the most resistant and fatal cancers, ranking as the seventh highest cause of cancer-related mortality globally (1). The incidence and mortality of PDAC are still rising, and the disease is expected to have a poor prognosis and a low resection rate despite major advancements in detection and treatment (2). PDAC, because of its retroperitoneal anatomical positioning and subtle early manifestations, leads to approximately 80%–85% of patients presenting at an advanced stage at the time of diagnosis (3).

Various criteria can be used to predict PDAC prognosis. Systemic immunological inflammation in cancer patients is strongly linked to metastasis and poor prognosis, according to a growing body of research (4–7). The most prevalent tumor marker for PDAC in clinical practice is carbohydrate antigen 19 9 (CA19-9), however its low sensitivity and specificity still limit its use in prognostic prediction (7–9). To predict the recurrence and survival of patients with PDAC, TNM stage and pathological grade are also used (10). There have also been reports of a correlation between the overall survival of patients with PDAC and serum indices such as albumin, serum bilirubin, and serum alkaline phosphatase (11). Nonetheless, the correlation between variables and the prognosis of patients is still controversial and requires further additional investigation.

In medicine, the use of AI and machine learning mostly helps with outcome prediction, risk assessment, and disease diagnosis. Researchers have examined diseases that require early diagnosis, including those of the skin, liver, heart, and Alzheimer’s disease, using various AI-based machine learning models (12). Thus, machine learning could improve the prognosis and outcome prediction of pancreatic ductal adenocarcinoma.

In this review, we investigated the potential of machine learning models in pancreatic adenocarcinoma prognosis, with an emphasis on the outcomes and the use of advanced machine learning techniques in the prediction of survival and recurrence.

This review is structured as follows:

● Introduction—contains information about PCAD and its prognosis and the use of ML in survival prediction, the purpose of the paper and its content.

● Material and methods—describing the selection process of included studies.

● Results—a summary of the selected studies organized in subsections according to their aim: survival (after surgery and overall), recurrence, and liver metastasis prediction. Data selection, feature extraction, ML algorithms used, and performance evaluation are included.

● Discussion—it highlights the findings, compares them to other prognostic methods and highlights their advantages.

● Challenges and future directions—assessing the existing issues and prospective development possibilities.

● Conclusions—summarizes the key findings and their implications.

Materials and methods

We conducted a comprehensive literature review to identify studies describing pancreatic adenocarcinoma prognostic models based on machine-learning algorithms. PubMed and Google Scholar were systematically searched using the selected keywords and terms. These included “artificial intelligence,” “pancreatic cancer,” “prognosis,” and “machine learning.” Multiple combinations of these terms were applied to each database using appropriate Boolean operators (OR/AND).

All studies on the use of ML for PDAC prognosis were included in this review. The following were excluded: (1) Other types of studies than original (meta-analyses, systematic reviews, narrative reviews, comments, and case reports); (2) Articles lacking full text in English; (3) Studies that did not use ML algorithms; and 4. Studies that did not address PCAD prognosis.

The literature search retrieved 1,017 articles, from two databases (PubMed and Google Scholar). A total of 620 abstracts were screened, and 46 full-text articles were assessed for eligibility. Studies in which no prognostic models were developed, AI was not used to develop the model, and future perspective articles were excluded. Finally, 12 studies were included (Table 1). The flow diagram is shown in Supplementary Figure 1 (25).

Table 1

Table 1. Overview of the included studies.

Results

The studies included had three major predictive objectives: survival, recurrence, and metastasis. Figure 1 summarizes these studies.

Figure 1

Figure 1. Overall view of the studies, data used, best performing algorithms, and outcome.

Survival prediction

Survival after surgery prediction

To predict long-term survival (2 years) in patients with borderline resectable pancreatic cancer undergoing upfront surgery, Huang et al. used four variables: age, vascular invasion length, vascular morphological malformation, and local lymphadenopathy. Logistic regression along with support vector machines (SVM), random forest, decision tree, and XGBoost were tested. The AUC stratified cross-validation in the logistic regression (0.864) was better than SVM (0.693), random forest (0.789), decision tree (0.790), and XGBoost (0.726) (13). Their findings may contribute to the selection of a personalized treatment option for these patients.

As curative surgery, including pancreaticoduodenectomy, does not always provide the expected results in terms of survival, with a significant percentage of patients not reaching the 1-year survival point, Ahmed et al. used machine learning algorithms to predict the odds of futile surgery. They utilized preoperative input variables (age, tumor size, and poor differentiation as key predictors of fatal outcomes, while neoadjuvant therapy was associated with better survival) to develop a predictive model. The algorithms employed were: gradient boosting model (showed the best performance with an AUC of 0.689), logistic regression (0.679), random forests (0.675), and decision tree (0.664) (14). Although the performance of the model is modest, it provides a good starting point for future research.

More complex algorithms are not necessarily superior. In predicting survival for borderline resectable patients, simple Logistic Regression outperformed advanced models such as SVM, Random Forest, and XGBoost. ML is used to identify patients for whom surgery will not provide a survival benefit (death <1 year). The key predictors of fatal outcomes are age, tumor size, and poor differentiation.

Overall survival prediction

Hu et al. created a complex prognostic model created by using a pathological risk model combined with a clinical risk model was constructed. For the pathological model, pretrained deep neural network models (MobileNet_v3_small, ResNet18, and DenseNet121) were used to extract features from representative hematoxylin and eosin (H&E)-stained digital slides/patients. The DenseNet121 model performed best, with a C-index of 0.73 in the external validation. For the clinical model, nine preoperative laboratory tests, sex, age, tumor site, tumor differentiation, and TNM classification were included. The combined model, using the 0.42 pathological model plus the 0.26 clinical model, showed a performance of 0.77 for the testing set, showing an improvement compared to the performance of the single models as well as the survival prediction made by only utilizing TNM classification (15).

Xiao et al., in a complex study using data from 117 patients with PDAC, extracted three features: Branched-chain Amino Acid Transaminase 1(BCAT1), amylase (AMY), and carbohydrate antigen 12-5 (CA12-5). The data were used to create a COX proportional hazards model and a DeepSurv model (a deep learning survival analysis model based on the Faraggi–Simon network model) for survival prediction. The DL model outperformed the COX model, with a C-index of 0.724 for the validation set. The authors used this model to further elaborate on personalized treatment recommendations to provide survival benefits (16).

Another promising study on PDAC prognostication using DL models was conducted by Schuurmans et al. They created two AI models. The first was a clinical model developed as a two-layer neural network using data retrospectively collected from 401 patients with PDAC. The variables used were age, CA19-9, CEA, BMI, unexpected weight loss, and ECOG performance status. The prediction of long- vs short-term survival reached an AUC of 0.638 in the external validation. An imaging model based on CT images from the same patients, created using a pure convolutional model (EfficientNet-B4), provided an AUC of 0.55 on external validation. The best predictive power was for the combined multimodal model, which consisted of a logistic regression using the risk scores provided by the previous two models. The performance on external validation reached an AUROC of 0.675 (17). This study is one of the few that provides data on the external validation of prognostic values. Although the discrimination power is not ideal, it shows an improvement compared to traditional PDAC prognostication using TNM staging.

A different approach to building ML prognosis models for PDAC was used with RNA-sequencing data. Ojha et al. created a 3-year prognosis model using gene expression data. By testing different ML techniques, the best-performing algorithm was determined to be random forests. To provide additional prognostic improvement, the model was individualized for male and female patients, thus providing supplementary personalization and an increase in the AUC value during external validation (18). This study illustrates the utility of ML in handling complex datasets and creating robust predictive models for PDAC.

Luo et al. employed an ML algorithm to construct a survival model based on innate immune cell barrier genes for pancreatic cancer, selected from a set of 1,356 unique genes associated with monocytes/macrophages, dendritic cells, NK cells, and neutrophils extracted from multiple databases. To select the best-performing algorithm, 162 models were constructed based on a combination of the following: RandomSurvivalForests, Enet, StepCox, CoxBoost, PartialLeastSquareRcox, supervisedPrincipalComponents, Gradient Boosting, survival-SVM, Ridge, obliqueRandomSurvivalForest, xgboost, CForest, and CTree. On external validation, the model based on the Random Survival Forest algorithm showed the best performance when constructed using 84 cancer diagnosis risk genes (CDRG) with a c-index-0.615. This model proved to have a good prognostic value for predicting survival at 1-, 3-, and 5-year (AUC—0.77, 0.79, 5–0.76) (19).

By integrating transcriptomic, methylation, and mutational data from 168 pancreatic cancer samples, Ge et al. identified molecular subtypes and built a pancreatic cancer prognostic model. The researchers used ten machine learning algorithms—random survival forest, survival-SVM, ridge regression, LASSO, elastic net, SuperPC, GBM, plsRcox, CoxBoost, and stepwise Cox—forming 101 model combinations. The final prognostic signature was built using ridge regression, which performed best (20). It divides patients into low- and high-risk groups, with high risk indicating worse survival outcomes. This study established a robust multi-omics-based classification of pancreatic cancer and provided a foundation for personalized therapeutic strategies.

Using a random forest algorithm trained on 598 differentially methylated CPG sites extracted from a publicly available dataset, along with clinical features and top mutated genes, Zaccaria et al. developed an ML model that has a good ability to predict KRAS, TP53, and KRAS and TP53 co-mutations on external validation. Furthermore, they used consensus clustering of methylation profiles to create four subgroups with different survival characteristics. An eXtreme Gradient Boosting classifier was used to identify epigenomic prognostic determinants. They successfully used a combination of several ML algorithms to integrate diverse and voluminous data to create a prognostic model for PDAC (21).

Another PDAC prognostic model was developed by Wang et al. using transcriptomic data from 1,034 samples, focusing on programmed cell death-related genes. The authors selected 17 genes, which were further reduced to a six-gene signature, to build a prognostic model. A total of 117 algorithmic combinations were generated using a combination of ML algorithms (StepCox, GBM, RSF, CoxBoost, Enet, plsRcox, survival-SVM, SuperPC, Lasso, and Ridge regression). The best performing model proved to be the one combining StepCox[both] and Ridge, providing an AUC for prediction of survival at 1-year = 0.795, 3-year = 0.776. and 5-year = 0.761 on external retrospective validation (22).

In summary, combining different data types yields better predictions than using a single data source. Deep Learning models extracting features from H&E slides (DenseNet121) combined with clinical data outperformed TNM staging. Combining CT imaging features (Radiomics/Deep Learning) with clinical markers (CA19-9) improves risk stratification.

In studies using specific biomarkers (e.g., BCAT1, Amylase), Deep Learning models (DeepSurv) demonstrated superior performance (C-index 0.724) compared to traditional Cox proportional hazards models. ML algorithms (specifically Random Forests and Ridge Regression) excel at processing complex genomic data (RNA-Seq, methylation, and immune genes) to build prognostic signatures. The creation of separate models for male and female patients improved the prognostic accuracy.

Predicting recurrence

To predict early recurrence (after less than 6 months) and create a tool for personalized treatment, Yan et al. used a combination of radiomics data (from preoperative contrast-enhanced CT, a rad-score was created using LASSO-selected radiomics features), visceral to subcutaneous fat volume ratio, and CA19–9 value to train an ML model based on logistic regression. Although other ML algorithms were not tested, this basic ML proved to have a fair predictive power (C-index = 0.712) on external validation (23).

Despite using a basic Logistic Regression model, the study achieved fair predictive power in external validation. Feature engineering (specifically, extracting radiomic features) and combining different clinical variables are often more important than the complexity of the algorithm itself.

Predicting early liver metastasis

For the prognosis of early liver metastasis after pancreatic cancer surgery, Zhu et al. gathered different variables (BMI, fatty liver, Child–Pugh score, tumor location, N stage, differentiation, tumor emboli, nerve invasion, and adjuvant chemotherapy) from 538 patients and created and tested prognostic models based on different ML algorithms. An advanced ensemble algorithm, XG boost, provided the best performance, outperforming other similar or simpler ML methods (24). Their model was intended to improve the prediction of metastasis, facilitating personalized treatment planning.

Unlike some other prognostic areas where simple models perform best, this study found that an advanced ensemble machine learning algorithm (XGBoost) provided the best performance. It outperformed both simpler and similar ML algorithms. The model succeeded by integrating a wide range of patient and tumor characteristics rather than using only a single biomarker.

Discussion

Machine learning provides significant benefits for the integration and assessment of vast quantities of intricate healthcare data. Compared to conventional biostatistical techniques, machine learning offers the advantages of flexibility and scalability, making it applicable to a wide range of activities, including risk stratification, diagnosis and classification, and survival prediction. The capacity of machine learning algorithms to analyze a variety of data types, such as imaging data, laboratory results, gene combinations, and pathological slides, and integrate them into forecasts for disease risk, diagnosis, prognosis, and suitable treatments is another benefit. ML has evolved from simple, basic supervised algorithms to complex, advanced algorithms, to black-box ML represented by DL.

PDAC has a dismal five-year survival rate of approximately 5%, making it one of the most lethal cancers globally (26). Currently, treating PDAC remains a clinical challenge, with the majority of patients failing to obtain survival improvements while receiving guideline-recommended therapy, which is typically due to late diagnosis or ineffective therapy. Furthermore, different treatment regimens have diverse therapeutic effects on patients, making it challenging for clinicians to choose personalized therapy programs. Using artificial intelligence technology to anticipate the prognosis of various treatment options and then choosing appropriate treatment measures based on these predictions is a promising technique.

The TNM system is the current standard for cancer staging. PDAC treatment strategies are now guided by this staging approach, which aims to identify various survival subgroups within the patient population (stages IA, IB, IIA, IIB, III, and IV). After curative resection, more than 90% of patients have metastases or cancer recurrence (27). Moreover, the overall survival rates of patients with the same TNM stage vary significantly. TNM staging’s prognostic performance reflects this diversity. For TNM staging areas, the reported c-index values for resectable and unresectable patients were 0.57 and 0.611, respectively (28–30). Therefore, it is clear that TNM staging is inaccurate in predicting outcomes, does not substantially correlate with overall survival, and is not a reliable tool for directing treatment choices (27, 31). Therefore, there is a need for improved PDAC stratification systems because of the shortcomings of traditional TNM staging. Artificial intelligence (AI) has shown encouraging results in predicting patient outcomes for different kinds of carcinomas. Its use might help to develop models with superior prediction power.

Compared to the data available for other types of cancer, such as colorectal or breast cancer, PDAC data are limited. As screening is not provided, data concerning early stage forms are scarce; survival is reduced, and successful treatment strategies are limited. Radical treatment rarely provides long-time survival; therefore, new treatment approaches and timing are necessary. Survival and recurrence predictions are vital for this purpose. Unfortunately, as stated above, the current prognosis using the TNM classification does not provide robust results, and other approaches are needed. Machine learning provides alternative variants by building different prognosis models. ML can integrate large amounts of different data (clinical, pathological, radiomics, genetic, and molecular) to create better prognosis models and guide personalized treatments. Simple ML algorithms, such as logistic regression, play a valuable role in building these models, while ensemble ML algorithms, especially Random Forests, have good predictive power. Promising predictive models have been created using a combination of DL and supervised ML or by combining two ML algorithms to improve predictive power.

To our knowledge, no other review addressing the use of ML and Advanced ML in PDAC prognosis has been published, the main reason being the novelty of the studies (less than 1 year at the time when the review was written).

Challenges and future directions

Although the use of ML algorithms is promising, we noted that the current studies lack prospective validation, and external validation in some cases. Most studies rely on a small number of patients; therefore, their generalizability might be affected.

Considering that most studies were published in the past year, it is clear that this is just the beginning of ML algorithms in PDAC prediction. Additionally, the studies presented use different types of data, but the best type of information to use has yet to be discovered. There is no comparison between the algorithms, as the studies compare the results with the preexisting “classical” methods of prediction, not to similar research using ML methods.

Future studies are needed to confirm the reliability of these models are needed, as well as to determine the most suitable data to be analyzed and the best algorithm to be employed.

Conclusions

ML models, from simple Logistic Regression to complex Deep Learning and Ensemble methods (such as Random Forests and XGBoost), have demonstrated a superior ability to predict survival, recurrence, and metastasis in patients with PDAC compared to conventional methods.

ML excels at integrating diverse and voluminous data (clinical, pathological, radiomics, and multi-omics/genetic) to create robust, predictive models. The development of these ML models is important for moving towards personalized therapeutic strategies, allowing clinicians to anticipate patient outcomes for different treatment options (e.g., predicting futile surgery or recurrence). Current models lack extensive external and prospective validation, which is crucial for their reliable implementation in clinical practice. Future research should focus on determining the most suitable data types and algorithms.

Author contributions

M-FA: Conceptualization, Data curation, Investigation, Methodology, Writing – original draft, Writing – review & editing. D-CL: Data curation, Supervision, Visualization, Writing – original draft, Writing – review & editing. M-IM: Data curation, Methodology, Writing – original draft, Writing – review & editing. AC-C: Data curation, Writing – original draft, Writing – review & editing. M-SM: Methodology, Supervision, Writing – original draft, Writing – review & editing.

Funding

The author(s) declared financial support was received for this work and/or its publication. We would like to acknowledge VICTOR BABES UNIVERSITY OF MEDICINE AND PHARMACY TIMISOARA for their support in covering the costs of publication for this research paper.

Conflict of interest

The authors 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.2025.1727806/full#supplementary-material

Supplementary Figure 1 | PRISMA flow diagram of literature search [25].

References

1. Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, and Jemal A. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. (2018) 68:394–424. doi: 10.3322/caac.21492

PubMed Abstract | Crossref Full Text | Google Scholar

2. Rawla P, Sunkara T, and Gaduputi V. Epidemiology of pancreatic cancer: global trends, etiology and risk factors. World J Oncol. (2019) 10:10–27. doi: 10.14740/wjon1166

PubMed Abstract | Crossref Full Text | Google Scholar

3. Zhou X, Guo Z, Liu S, Chen Z, Wang Y, Yang R, et al. Transcriptomics and molecular Docking reveal the potential mechanism of Lycorine against pancreatic cancer. Phytomedicine. (2024) 122:155128. doi: 10.1016/j.phymed.2023.155128

PubMed Abstract | Crossref Full Text | Google Scholar

4. Diakos CI, Charles KA, McMillan DC, and Clarke SJ. Cancer-related inflammation and treatment effectiveness. Lancet Oncol. (2014) 15:e493–503. doi: 10.1016/S1470-2045(14)70263-3

PubMed Abstract | Crossref Full Text | Google Scholar

5. Topkan E, Kucuk A, Ozturk D, Ozkan EE, Kılıç Durankuş N, Şenyürek Ş, et al. High systemic immune-inflammation index values before treatment predict poor pancreatic cancer outcomes after definitive chemoradiotherapy. Clin Med Insights: Oncol. (2024) 18:1–8. doi: 10.1177/11795549241298552

PubMed Abstract | Crossref Full Text | Google Scholar

6. Han R, Tian Z, Jiang Y, Guan G, Sun X, Yu Y, et al. Prognostic significance of systemic immune-inflammation index and platelet-albumin-bilirubin grade in patients with pancreatic cancer undergoing radical surgery. Gland Surg. (2022) 11:576–87. doi: 10.21037/gs-22-117

PubMed Abstract | Crossref Full Text | Google Scholar

7. Madadjim R, An T, and Cui J. MicroRNAs in pancreatic cancer: advances in biomarker discovery and therapeutic implications. Int J Mol Sci. (2024) 25:3914. doi: 10.3390/ijms25073914

PubMed Abstract | Crossref Full Text | Google Scholar

8. Ni XG, Bai XF, Mao YL, Shao YF, Wu JX, Shan Y, et al. The clinical value of serum CEA, CA19-9, and CA242 in the diagnosis and prognosis of pancreatic cancer. Eur J Surg Oncol. (2005) 31:164–9. doi: 10.1016/j.ejso.2004.09.007

PubMed Abstract | Crossref Full Text | Google Scholar

9. Vlavcheski F, O’Neill EJ, Gagacev F, and Tsiani E. Effects of Berberine against pancreatitis and pancreatic cancer. Molecules. (2022) 27:8630. doi: 10.3390/molecules27238630

PubMed Abstract | Crossref Full Text | Google Scholar

10. Shen CN, Goh KS, Huang CR, Chiang TC, Lee CY, Jeng YM, et al. Lymphatic vessel remodeling and invasion in pancreatic cancer progression. EBioMedicine. (2019) 47:98–113. doi: 10.1016/j.ebiom.2019.08.044

PubMed Abstract | Crossref Full Text | Google Scholar

11. Wang D, Pan B, Huang JC, Chen Q, Cui SP, Lang R, et al. Development and validation of machine learning models for predicting prognosis and guiding individualized postoperative chemotherapy: a real-world study of distal cholangiocarcinoma. Front Oncol. (2023) 13:1106029. doi: 10.3389/fonc.2023.1106029

PubMed Abstract | Crossref Full Text | Google Scholar

12. Yi F, Yang H, Chen D, Qin Y, Han H, Cui J, et al. XGBoost-SHAP-based interpretable diagnostic framework for Alzheimer’s disease. BMC Med Inf Decis Mak. (2023) 23:137. doi: 10.1186/s12911-023-02238–9

PubMed Abstract | Crossref Full Text | Google Scholar

13. Huang JC, Lyu SC, Pan B, Wang HX, Ma YW, Jiang T, et al. A logistic regression model to predict long-term survival for borderline resectable pancreatic cancer patients with upfront surgery. Cancer Imaging. (2025) 25:10. doi: 10.1186/s40644-025-00830-y

PubMed Abstract | Crossref Full Text | Google Scholar

14. Ahmed KS, Marcinak CT, Issaka SM, Ali MM, and Zafar SN. Machine learning to predict early death despite pancreaticoduodenectomy. J Surg Res. (2025) 310:186–93. doi: 10.1016/j.jss.2025.03.047

PubMed Abstract | Crossref Full Text | Google Scholar

15. Hu K, Bian C, Yu J, Jiang D, Chen Z, Zhao F, et al. Construction of a combined prognostic model for pancreatic ductal adenocarcinoma based on deep learning and digital pathology images. BMC Gastroenterol. (2024) 24:387. doi: 10.1186/s12876-024-03469-4

PubMed Abstract | Crossref Full Text | Google Scholar

16. Xiao Y, Sun S, Zheng N, Zhao J, Li X, Xu J, et al. Development of PDAC diagnosis and prognosis evaluation models based on machine learning. BMC Cancer. (2025) 25:512. doi: 10.1186/s12885-025-13929-z

PubMed Abstract | Crossref Full Text | Google Scholar

17. Schuurmans M, Saha A, Alves N, Vendittelli P, Yakar D, Sabroso-Lasa S, et al. End-to-end prognostication in pancreatic cancer by multimodal deep learning: a retrospective, multicenter study. Eur Radiol. (2025) 35:7537–48. doi: 10.1007/s00330-025-11694-y

PubMed Abstract | Crossref Full Text | Google Scholar

18. Ojha A, Zhao SJ, Akpunonu B, Zhang JT, Simo KA, and Liu JY. Gap-App: A sex-distinct AI-based predictor for pancreatic ductal adenocarcinoma survival as a web application open to patients and physicians. Cancer Lett. (2025) 622:217689. doi: 10.1016/j.canlet.2025.217689

PubMed Abstract | Crossref Full Text | Google Scholar

19. Luo Q, Jiang T, Xie D, Li X, and Xie K. Innate immune cell barrier-related genes inform precision prognosis in pancreatic cancer. Front Immunol. (2025) 16:1559373. doi: 10.3389/fimmu.2025.1559373

PubMed Abstract | Crossref Full Text | Google Scholar

20. Ge J, Cai J, Zhang G, Li D, and Tao L. Multi-omics integration and machine learning uncover molecular basal-like subtype of pancreatic cancer and implicate A2ML1 in promoting tumor epithelial-mesenchymal transition. J Transl Med. (2025) 23:741. doi: 10.1186/s12967-025-06711-z

PubMed Abstract | Crossref Full Text | Google Scholar

21. Zaccaria GM, Altini N, Mongelli V, Marino F, and Bevilacqua V. Development and validation of a machine learning prognostic model based on an epigenomic signature in patients with pancreatic ductal adenocarcinoma. Int J Med Inform. (2025) 199:105883. doi: 10.1016/j.ijmedinf.2025.105883

PubMed Abstract | Crossref Full Text | Google Scholar

22. Wang B, Long Z, Zou X, Sun Z, and Xiao Y. Machine learning-based construction of a programmed cell death-related model reveals prognosis and immune infiltration in pancreatic adenocarcinoma patients. Sci Rep. (2025) 15:25156. doi: 10.1038/s41598-025-10847-9

PubMed Abstract | Crossref Full Text | Google Scholar

23. Yan SY, Chen FM, Guo BJ, Hu S, Lin L, Yang YW, et al. Preoperative prediction for early recurrence in patients with pancreatic ductal adenocarcinoma: combining radiomics and abdominal fat analysis. BMC Med Imaging. (2025) 25:251. doi: 10.1186/s12880-025-01773-3

PubMed Abstract | Crossref Full Text | Google Scholar

24. Zhu H, Zhou Y, Shen D, Wu K, Gan X, Xue X, et al. An interpretable machine learning model for predicting early liver metastasis after pancreatic cancer surgery. BMC Cancer. (2025) 25:1117. doi: 10.1186/s12885-025-14503-3

PubMed Abstract | Crossref Full Text | Google Scholar

25. Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. JClinEpidemiol. (2021) 134:178189. doi: 10.1016/j.jclinepi.2021.03.001

PubMed Abstract | Crossref Full Text | Google Scholar

26. Bengtsson A, Andersson R, and Ansari D. The actual 5-year survivors of pancreatic ductal adenocarcinoma based on real-world data. Sci Rep. (2020) 10:16425. doi: 10.1038/s41598-020-73525-y

PubMed Abstract | Crossref Full Text | Google Scholar

27. Shi XY, Wang Y, Zhou X, Xie ML, Ma Q, Wang GX, et al. A population-based nomogram to individualize treatment modality for pancreatic cancer patients underlying surgery. Sci Rep. (2023) 13:4856. doi: 10.1038/s41598-023-31292-6

PubMed Abstract | Crossref Full Text | Google Scholar

28. Hamada T, Nakai Y, Yasunaga H, Isayama H, Matsui H, Takahara N, et al. Prognostic nomogram for nonresectable pancreatic cancer treated with gemcitabine-based chemotherapy. Br J Cancer. (2014) 110:1943–9. doi: 10.1038/bjc.2014.131

PubMed Abstract | Crossref Full Text | Google Scholar

29. Healy GM, Salinas-Miranda E, Jain R, Dong X, Deniffel D, Borgida A, et al. Pre-operative radiomics model for prognostication in resectable pancreatic adenocarcinoma with external validation. Eur Radiol. (2022) 32:2492–505. doi: 10.1007/s00330-021-08314-w

PubMed Abstract | Crossref Full Text | Google Scholar

30. He C, Sun S, Zhang Y, Lin X, and Li S. Score for the overall survival probability of patients with pancreatic adenocarcinoma of the body and tail after surgery: A novel nomogram-based risk assessment. Front Oncol. (2020) 10:590. doi: 10.3389/fonc.2020.00590

PubMed Abstract | Crossref Full Text | Google Scholar

31. Kang H, Kim SS, Sung MJ, Jo JH, Lee HS, Chung MJ, et al. Evaluation of the 8th edition AJCC staging system for the clinical staging of pancreatic cancer. Cancers (Basel). (2022) 14:4672. doi: 10.3390/cancers14194672

PubMed Abstract | Crossref Full Text | Google Scholar