Interpretable machine learning prediction of 1-year overall survival in pancreatic cancer patients aged 65 years and older

Introduction

Pancreatic cancer (PC) is one of the most aggressive gastrointestinal malignancies globally, characterized by insidious onset, rapid progression, and high mortality (1). The 1-year survival rate of PC of seven high-income countries is 25–30%, making it a major threat to public health (2). Notably, elderly patients aged ≥65 years account for approximately 68% of all PC cases, and this population often presents with age-related physiological decline, multiple comorbidities, and reduced treatment tolerance (3). These characteristics lead to more heterogeneous disease progression and worse short-term prognosis compared with younger patients (4). Therefore, as an important short-term prognostic endpoint reflecting initial therapeutic response and early disease progression risk, accurate prediction of 1-year survival rate is essential for guiding clinical decision-making, optimizing treatment regimens, and improving patient management outcomes in elderly PC patients (5).

In recent years, machine learning (ML) algorithms have been increasingly applied to oncology prognosis prediction, overcoming the limitations of traditional statistical methods (e.g., Cox proportional hazards regression) in handling nonlinear relationships and high-dimensional data. Common ML models such as decision tree (DT), stochastic gradient descent algorithm, random forest (RF) algorithm, support vector machine (SVM) algorithm, logistics algorithm and AdaBoost algorithm have shown promising performance in breast cancer benign or malignant classification (6) or distant metastasis prediction research for thyroid cancer (7). However, most existing studies focus on the general PC population and lack systematic comparisons of multiple ML models, leading to suboptimal model selection (8). Moreover, the “black-box” nature of high-performance ML models limits their clinical translation, as clinicians require transparent interpretations of key prognostic factors to trust and apply these models.

The Surveillance, Epidemiology, and End Results (SEER) database, established by the U.S. National Cancer Institute (NCI), provides population-based, comprehensive data on cancer incidence, treatment, and survival, making it a reliable resource for epidemiological and prognostic studies (9). Least absolute shrinkage and selection operator (LASSO) regression is widely used for feature selection in high-dimensional data, which can reduce model complexity and improve generalization ability (10). SHapley Additive exPlanations (SHAP) combines optimal credit allocation with local explanations to visually represent the importance of each variable in the model, thereby providing a more interpretable output. SHAP analysis, as an interpretable ML tool, can quantify the impact of each feature on the prediction result and enhance the clinical interpretability of the model (11).

To address the existing research gaps, this study aimed to: (I) construct and compare six ML models for predicting 1-year survival in PC patients aged ≥65 years using SEER database data [2011–2015]; (II) screen key prognostic features using LASSO regression; (III) validate the performance of the optimal model; and (IV) interpret the optimal model using SHAP analysis to clarify the influence of key features on 1-year survival. This study is expected to provide an accurate and interpretable prognostic tool for clinical practice in elderly PC patients. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0914/rc).

Methods

Patients source and variables inclusion

The study used the database of the Surveillance, Epidemiology and End Results (SEER) NCI (https://seer.cancer.gov/), a free US cancer registry. Specifically, we used the SEER Research Data, 18 Registries, Nov 2019 Sub [2000–2017] as the data source. The exact SEER variable names used for analysis, along with their corresponding recoding procedures, American Joint Committee on Cancer (AJCC) staging definitions, T/N stage categorization methods, insurance coding definitions, and age grouping rationale, are now fully presented in Table S1. The extraction workflow was provided at https://cdn.amegroups.cn/static/public/tcr-2026-0914-1.pdf. The original data was at https://cdn.amegroups.cn/static/public/tcr-2026-0914-2.xls. The clinicopathological data of all PC patients from 2011 to 2015 were downloaded from the SEER database. It collects information on patient demographics, year of diagnosis, marital status, race, age, survival state, insurance, tumor-node-metastasis (TNM) stage, surgery methods, radiotherapy, chemotherapy, AJCC 7th staging and follow-up for survival. The inclusion criteria were as follows: (I) ICD-O-3/WHO = ‘Pancrea’; (II) diagnostic confirmation = ‘Positive Histology’; (III) survival months >0 day; (IV) year at diagnosis is from 2011–2015. The exclusion criteria were as follows: (I) patients with diagnosed with autopsy or death certificate; (II) cause of death or survival months is unknown; (III) age <65 years; (IV) distant metastasis. Finally, data of 5,601 PC cancer patients were included in this study.

For this study, we have signed authorization and received permission from SEER to access and use the SEER information. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The institutional review board approval and informed consent were not required in the current study because the data source is public database.

Statistical analysis

All eligible elderly patients with PC were randomized in a 7:3 ratio into the training set and validation set for model development and external validation, respectively. The Chi-squared test was used to compare categorical clinicopathological characteristics (e.g., sex, race, surgery type, AJCC stage) between the two groups. For continuous variables (only age in this study), the Shapiro-Wilk test was first performed to verify normality. If normally distributed, the independent samples t-test was used for comparison; otherwise, the Mann-Whitney U test was applied. The interpretable ML prediction model for 1-year overall survival (OS) in elderly patients with PC was developed based on the training cohort, following these steps: firstly, the LASSO regression method was employed for predictor selection and regularization to avoid overfitting, which helped screen out the most valuable predictive factors from all baseline characteristics (e.g., age, grade, surgery, radiation). Secondly, significant factors selected by LASSO regression were incorporated to construct the interpretable ML prediction model, with hyperparameters optimized via 10-fold cross-validation in the training set. Thirdly, the primary endpoint of the model was the 1-year OS of the study population, defined as the time from diagnosis to all-cause death or last follow-up.

The performance of the ML model was comprehensively evaluated from three aspects: discrimination, calibration, and clinical utility. Discrimination refers to the ability of the model to distinguish between patients who survived for ≥1 year and those who died within 1 year, which was primarily assessed using the receiver operating characteristic (ROC) curve and the area under the curve (AUC). Additionally, the precision-recall (PR) curve was supplemented to evaluate the model’s performance in imbalanced survival outcome datasets, as PR curves are more robust than ROC curves when dealing with uneven distribution of positive and negative cases. For clinical utility assessment, decision curve analysis (DCA) was performed to quantify the net benefit of the ML model at different threshold probabilities, which directly reflects the clinical value of the model for guiding clinical decision-making. Based on the optimal cut-off value determined by the ROC curve (maximizing the Youden index), patients were stratified into low-risk and high-risk groups according to the model’s predicted 1-year OS probabilities. Furthermore, SHAP values were employed for feature importance analysis to enhance the interpretability of the ML model, which quantitatively clarified the direction and magnitude of each predictive factor’s contribution to the 1-year OS prediction.

The baseline characteristic statistics, LASSO regression-based feature selection, multi-ML model training, ROC curve and analysis SHAP-based feature importance analysis were performed using Python software (version 3.8) with pandas, scipy, scikit-learn and XGBoost packages. For all statistical analyses, a two-tailed P value <0.05 was considered statistically significant.

Results

Patients’ baseline characteristics

The total cohort included 5,601 elderly PC patients (aged ≥65 years). The training cohort included 3,920 elderly PC patients, whereas the validation cohort included 1,681 elderly PC patients. The demographic and clinicopathologic characteristics of the patients in both cohorts are listed in Table 1. With the exception of age, no significant differences were observed in clinicopathologic characteristics between the training and validation sets.

LASSO Cox regression analysis

LASSO Cox regression analysis was used to screen the potential factors of PC patients (12). After the predictor selection and regularization of LASSO regression analysis. Six significant parameters (coefficient >0.05) were selected. The parameters including surgery (coefficient 0.4008), grade (coefficient 0.3048), T stage (coefficient 0.1992), chemotherapy (coefficient 0.1823), radiation (coefficient 0.0716), N (coefficient 0.0602). The original feature importance chart analysis is shown in Figure 1.

Figure 1 Original feature importance chart. AJCC, American Joint Committee on Cancer; N, node; T, tumor.

Performance comparison of ML models

Six ML algorithms, specifically DT, logistic regression (LR), RF, SVM, K-nearest neighbor (KNN), light gradient boosting machine (LightGBM). To comprehensively evaluate the predictive capability of our ML models, we employed a suite of established classification metrics. The overall discriminative power was quantified using accuracy, precision, and the F1-Score. To gain deeper insights into the model’s behavior and to diagnose potential issues such as overfitting or class-specific biases, we conducted a detailed examination of the confusion matrices generated on both the training and independent test sets. This dual-dataset analysis provides a holistic view of model performance and generalizability, informing subsequent optimization strategies.

The performance of the six ML models in the training and validation sets is shown in Table 2. The RF model achieved the best predictive performance, with an AUC of 0.875 [95% confidence interval (CI): 0.852–0.898] in the training set and 0.799 (95% CI: 0.774–0.824) in the validation set. The RF model also had the highest accuracy (76.3%), sensitivity (88.1%), specificity (60.2%), and F1-score (0.831) in the validation set. The ROC curves of the six models in the validation set showed that the RF model had the largest AUC (Figure 2).

Figure 2 The ROC curves of training set (A) and validation set (B). AUC, area under the curve; KNN, K-nearest neighbor; LightGBM, light gradient boosting machine; ROC, receiver operating characteristic; SVM, support vector machine.

Figure 2 shows the ROC curve. The AUC, represents the area beneath the ROC curve and can be employed to predict precision in binary classification models. All six models outperformed the “Random Guess” (AUC =0.500), confirming their capacity to distinguish 1-year survivors from non-survivors. Among training set, LightGBM achieved the highest AUC (0.802), followed closely by RF (AUC =0.799) and DT (AUC =0.788). Other models (SVM: AUC =0.749; LR: AUC =0.729; KNN: AUC =0.721) also exhibited acceptable discriminative performance on the training set. Consistent with training set results, all models maintained AUC values above 0.500, indicating robust discriminative capacity on unseen samples. The top-performing models included RF (AUC =0.799), LightGBM (AUC =0.798), and DT (AUC =0.797). Their AUCs were nearly identical to those on the training set, demonstrating strong generalization. Notably, KNN’s AUC decreased from 0.721 (training set) to 0.687 (validation set), suggesting slight overfitting to the training data. Other models (e.g., LR, SVM) showed minor AUC fluctuations between sets, which is typical for model generalization. Overall, the ROC curves and AUC values confirm that most models (particularly LightGBM, RF, and DT) possess strong, generalizable discriminative performance for predicting 1-year survival in PC patients. Given that RF achieved the highest AUC on the validation set (additionally, it exhibited excellent generalization stability, with identical AUC values between the training and validation phases), we selected it as the optimal model.

The predicted performance of RF

RF exhibited highest AUC among six models. The area under the precision-recall curves (AUPRCs) are 0.875 and 0.867 in training set and validation set, respectively, indicating excellent OS predictive performance (Figure 3). The DCA curve demonstrates that the RF model has excellent clinical utility (Figure 4).

Figure 3 The PR curves of training set (A) and validation set (B). AUPRC, area under the precision-recall curve; PR, precision-recall.

Figure 4 The DCA curves of training set (A) and validation set (B). DCA, decision curve analysis.

A total of 1,681 patients were enrolled in the test set, among which the predictive outcome was concordant with the actual status in 1,283 patients, while discordant results were observed in 398 patients. In terms of the 1-year survival stratification, 1,112 patients had an OS time of more than 1 year, and 980 of these cases were predicted as survivors by the model. In contrast, 529 patients did not survive beyond 1 year, and 303 of them were accurately identified as non-survivors by the model (Figure 5).

Figure 5 The confusion matrix of validation set.

The interpretability of RF

SHAP analysis was performed to interpret the RF model. The SHAP violin plot showed the importance of the 6 key features and their impact on 1-year survival (Figure 6). It presented a comprehensive swarm plot illustrating the variables in the RF model. The horizontal axis represented SHAP values, while the vertical axis displayed features sorted by their cumulative SHAP value impact. Each data pointed corresponds to a specific instance, with its position along the x-axis indicating the SHAP value for that particular instance and feature. Surgery, chemotherapy and grade emerged as the three most important factors in predicting 1-year OS.

Figure 6 Violin plot illustrating the distribution of SHAP values for each feature. N, node; RF, random forest; SHAP, SHapley Additive exPlanations; T, tumor.

Figure 7A displays a SHAP force plot for an individual sample, elaborating the predictive reasoning process of the model (noted as “RF =0” herein) for this specific case. Within this visualization, red segments signify positive contributions that elevate the predictive outcome (corresponding to the “higher” direction), while the blue segment represents a negative contribution that dampens the predicted value (aligning with the “lower” direction).

Figure 7 Local model explanation by the SHAP method. (A) SHAP force plot. (B) SHAP waterfall plot. N, node; RF, random forest; SHAP, SHapley Additive exPlanations; T, tumor.

Figure 7B depicts a SHAP waterfall plot that delineates the model’s predictive reasoning for the 1-year survival probability of a specific sample. Here, the base value (0.661) represents the model’s average predicted 1-year survival probability across all samples, while the final f(x) value (0.816) denotes the predicted 1-year survival probability for this individual case—indicating that the model estimates a higher survival likelihood for this sample relative to the overall average.

Within this visualization, blue segments correspond to negative contributions (which reduce the predicted survival probability), whereas red segments signify positive contributions (which elevate the predicted survival probability). Concretely, multiple features exert initial minor negative impacts: chemotherapy (SHAP =−0.024), surgery (SHAP =−0.03), grade (SHAP =−0.013), radiation (SHAP =−0.02), T (SHAP =−0.019), and N (SHAP =−0.001). These downward pressures are subsequently offset by sequential positive contributions (including SHAP values of +0.05, +0.07, and a prominent +0.1), which collectively drive the predicted value upward from the baseline (0.661) to the final f(x) of 0.816.

Discussion

Principal findings

Precise survival prediction can help improve the quality of future clinical trials as it could reduce potential biases during the assignment of different treatments. However, an increasing number of studies have pointed out that the currently used TNM clinical staging system for PC cancer could not provide very accurate prognostic information (13-15). Despite the high prognostic accuracy achieved by ML models, the influence of individual variables on these models often remains unknown (16). This lack of transparency limits the application of ML in clinical practice (17). Our study focused on predicting 1-year OS in PC patients, systematically constructing and validating an optimal predictive model through a complete workflow of “LASSO Cox feature selection - ML algorithm validation - SHAP interpretability analysis”.

Firstly, LASSO Cox regression screened 6 prognostic factors with robust prognostic value from candidate features, among which surgery emerged as the core protective factor with the highest coefficient. Secondly, among the 6 ML algorithms, the RF model exhibited the best performance, achieving an AUC of 0.799 (95% CI: 0.774–0.824) in the validation set, along with excellent accuracy (76.3%), sensitivity (88.1%), and F1-score (0.831). Thirdly, the AUPRC of the RF model reached 0.875 and 0.867 in the training and validation sets, respectively, and DCA confirmed its superior clinical utility; validation in a large-scale test set (1,681 patients) showed that the concordance rate between predicted outcomes and actual status was 76.3% (1,283/1,681), among which 88.1% (980/1,112) of 1-year survivors (1,112 cases) were accurately predicted, and 53.25% (303/569) of non-survivors (569 cases) were correctly identified. Finally, SHAP interpretable analysis identified surgery, chemotherapy, and histological grade as the three core factors predicting 1-year OS, quantifying the marginal contribution of each factor to individual predictions and addressing the “black box” dilemma of ML models. Collectively, these results confirm that the RF model possesses excellent discriminative performance, clinical utility, and interpretability in predicting 1-year OS in elderly PC patients, demonstrating significant methodological and clinical application advantages over traditional nomograms (18-20).

The multi-dimensional performance advantages of the RF model essentially stem from its accurate capture of complex relationships in PC prognosis. Unlike the linear regression relied on by nomograms, the RF model, as a tree-based ensemble algorithm, can automatically mine non-linear interactions between predictive factors (e.g., the synergistic protective effect of surgery combined with chemotherapy), which is the core reason why its AUPRC (0.867 in the validation set) and AUC (0.799) are superior to those of traditional nomograms (usually with a C-index of 0.711) (4). Especially in predicting 1-year OS of PC, there is a certain imbalance in outcomes (survivors accounted for 66.1% in the test set). Compared with AUC, AUPRC better reflects the balance between precision and recall of the model. However, nomogram studies rarely report AUPRC, often ignoring prediction biases in imbalanced data. DCA further confirmed that the RF model had a positive net benefit across the entire range of clinically relevant thresholds, meaning its application could effectively optimize clinical decisions (e.g., avoiding overtreatment in low-risk patients and intensifying interventions in high-risk patients), which is unattainable by nomograms. Most nomograms only validate discriminative performance through C-index, lacking direct quantification of clinical utility. The high concordance rate (76.3%) and excellent predictive accuracy for survivors (88.1%) in the large-scale test set prove that the model has good generalization ability, overcoming the poor generalization of nomograms often caused by single-center small-sample validation (21).

It is important to objectively highlight that the RF model exhibited high sensitivity (88.1%) in identifying non-survivors among elderly PC patients, a notable strength for clinical risk stratification. Nevertheless, this performance should be contextualized alongside the high heterogeneity of PC. Our current model does not incorporate critical prognostic factors such as molecular subtypes and tumor microenvironment features. These uncaptured variables may still contribute to occasional missed diagnoses of individual high-risk patients, underscoring room for further model refinement. Previous studies on PC prognosis prediction have mostly focused on the general population (18-20), and few have targeted elderly patients aged ≥65 years (4). For example, Teng et al. (22) constructed a CatBoost model to predict 3-year survival in PC patients, but their study included patients of all ages and did not focus on short-term survival. This study fills the gap by constructing a targeted prediction model for elderly PC patients, considering their unique clinical characteristics. Moreover, this study systematically compared six ML models and confirmed the superiority of the RF model, which is consistent with previous studies showing that ensemble learning models have better performance in handling complex clinical data (23). Another advantage of this study is the combination of LASSO regression and SHAP analysis. LASSO regression reduced model complexity by screening key features, while SHAP analysis solved the “black-box” problem of ML models, making the prediction results more interpretable (24). Previous studies on PC prognosis rarely used SHAP analysis to interpret ML models, which limited their clinical translation (25). This study’s SHAP analysis clarified the quantitative impact of each key feature on 1-year survival, providing clinicians with clear prognostic information. Our model identified surgery as the strongest prognostic factor. However, this finding should be interpreted with caution, as it may partially reflect treatment selection bias. Patients eligible for surgery are generally healthier and have earlier-stage disease. Our model did not adjust for frailty, comorbidity burden, or functional status due to the unavailability of these variables in the SEER database. Therefore, the observed survival benefit associated with surgery may overestimate the true treatment effect. We emphasize that our model is intended as a population-based prognostic tool for risk stratification, not as a causal inference of treatment effects.

Limitations

In addition, the model’s analysis of chemotherapy subgroup benefits is still based on clinical features. Several clinically important prognostic variables, including performance status, frailty, comorbidity burden, CA19-9 levels, resection margin status, tumor location, chemotherapy regimen details, and molecular characteristics, were not available in the SEER database and therefore could not be incorporated into our model. These factors, particularly frailty and comorbidity burden, are known to be highly prognostic in elderly cancer patients and may influence outcomes beyond chronological age alone. Future studies integrating geriatric assessment and biomarker data are needed to validate and potentially improve our model. Furthermore, compared with nomograms, the limitations of the RF model can be remedied through subsequent optimization, while the inherent linear constraints of nomograms and the problem of redundant features caused by lack of regularization are difficult to break through technical improvements.

Conclusions

In summary, this study confirmed through LASSO Cox feature selection and multi-ML algorithm validation that the RF model is an optimal predictive tool for 1-year OS in PC patients. Its excellent AUC, AUPRC, and DCA performance, combined with the interpretability of SHAP analysis and the generalization ability of large-scale test sets, are significantly superior to traditional nomograms. Surgery, chemotherapy, and histological grade, as the three core predictive factors, provide clear targets for clinical risk stratification and personalized treatment (1,6). Although the model has limitations such as generalization and deployment thresholds, it is expected to become an important auxiliary tool for PC clinical management through subsequent multi-center validation, molecular indicator integration, and hybrid nomogram construction, promoting the transformation of PC prognosis prediction from “linear estimation” to an ML-driven model of “precision, interpretability, and clinical applicability”, and ultimately improving patients’ survival outcomes.

Acknowledgments

We extend our profound gratitude to the SEER program for granting approval for registration and providing access to the SEER database.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0914/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0914/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0914/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Siegel RL, Miller KD, Fuchs HE, et al. Cancer statistics, 2022. CA Cancer J Clin 2022;72:7-33. [Crossref] [PubMed]
2. Arnold M, Rutherford MJ, Bardot A, et al. Progress in cancer survival, mortality, and incidence in seven high-income countries 1995-2014 (ICBP SURVMARK-2): a population-based study. Lancet Oncol 2019;20:1493-505. [Crossref] [PubMed]
3. Xie D, Qian B, Yang J, et al. Can Elderly Patients With Pancreatic Cancer Gain Survival Advantages Through More Radical Surgeries? A SEER-Based Analysis. Front Oncol 2020;10:598048. [Crossref] [PubMed]
4. Zhong J, Liao X, Peng S, et al. A Visualized Dynamic Prediction Model for Overall Survival in Elderly Patients With Pancreatic Cancer for Smart Medical Services. Front Public Health 2022;10:885624. [Crossref] [PubMed]
5. Schneider F, Chen H, Pelzer U, et al. The basis for future personalized therapy approaches - Machine learning-generated 1-year survival rate, metastatic status and therapy-dependent survival in pancreatic cancer patients. Eur J Cancer 2026;234:116189. [Crossref] [PubMed]
6. Zhou S, Hu C, Wei S, et al. Breast Cancer Prediction Based on Multiple Machine Learning Algorithms. Technol Cancer Res Treat 2024;23:15330338241234791. [Crossref] [PubMed]
7. Qiao L, Li H, Wang Z, et al. Machine learning based on SEER database to predict distant metastasis of thyroid cancer. Endocrine 2024;84:1040-50. [Crossref] [PubMed]
8. Zhang L, He J. Development and validation of a machine learning model for predicting early death in metastatic pancreatic ductal adenocarcinoma: a study based on the SEER database. Transl Cancer Res 2026;15:53. [Crossref] [PubMed]
9. Chen H, Dai W, Yang J, et al. Interpretable machine learning-based prediction of liver metastasis risk in elderly patients with small cell lung Cancer: A study based on the SEER database and external validation in a Chinese cohort. Int J Med Inform 2026;209:106274. [Crossref] [PubMed]
10. Wu Z, Deng Q, Pang Y, et al. Development of a novel web-based calculator for predicting overall survival in early-onset cervical cancer patients with positive lymph node metastasis. European Journal of Gynaecological Oncology. 2025;46:20-9. [Crossref]
11. Chen B, Chen J, Cao Z, et al. Explainable multi-omics modeling for risk stratification in pancreatic ductal adenocarcinoma. Gland Surg 2026;15:91. [Crossref] [PubMed]
12. He W, Cao C, Li Q, et al. Machine-learning CT radiomics for prognostication in unresectable pancreatic cancer. Front Pharmacol 2026;17:1750762. [Crossref] [PubMed]
13. Schuurmans M, Saha A, Alves N, et al. End-to-end prognostication in pancreatic cancer by multimodal deep learning: a retrospective, multicenter study. Eur Radiol 2025;35:7537-48. [Crossref] [PubMed]
14. Ren C, Ma Y, Jin J, et al. Development and external validation of a dynamic nomogram to predict the survival for adenosquamous carcinoma of the pancreas. Front Oncol 2022;12:927107. [Crossref] [PubMed]
15. Srivastav AK, Singh R. Artificial intelligence and immune topography transforming prognostication in pancreatic cancer. Gland Surg 2026;15:55. [Crossref] [PubMed]
16. Xin W, Rixin S, Linrui L, et al. Machine learning-based radiomics for predicting outcomes in cervical cancer patients undergoing concurrent chemoradiotherapy. Comput Biol Med 2024;177:108593. [Crossref] [PubMed]
17. Cabitza F, Rasoini R, Gensini GF. Unintended Consequences of Machine Learning in Medicine. JAMA 2017;318:517-8. [Crossref] [PubMed]
18. Peng W, Yu X, Yang R, et al. Construction and validation of a nomogram for cancer specific survival of postoperative pancreatic cancer based on the SEER and China database. BMC Gastroenterol 2024;24:104. [Crossref] [PubMed]
19. Ren H, Fei H, Niu P, et al. An interpretable machine learning model for predicting survival in pancreatic cancer via SHAP: a multicenter study. J Gastroenterol 2026;61:808-20. [Crossref] [PubMed]
20. Yunpeng P, Lingdi Y, Xiaole Z, et al. Establishment and validation of a nomogram based on coagulation parameters to predict the prognosis of pancreatic cancer. BMC Cancer 2023;23:548. [Crossref] [PubMed]
21. Wang N. Interpretable multi-modal machine learning model predicts early recurrence after curative resection of pancreatic ductal adenocarcinoma. Int J Surg 2026;112:10950-1. [Crossref] [PubMed]
22. Teng B, Zhang X, Ge M, et al. Personalized three-year survival prediction and prognosis forecast by interpretable machine learning for pancreatic cancer patients: a population-based study and an external validation. Front Oncol 2024;14:1488118. [Crossref] [PubMed]
23. Oh SS, Kang B, Hong D, et al. A Multivariable Prediction Model for Mild Cognitive Impairment and Dementia: Algorithm Development and Validation. JMIR Med Inform 2024;12:e59396. [Crossref] [PubMed]
24. Wei Z, Bai X, Xv Y, et al. A radiomics-based interpretable machine learning model to predict the HER2 status in bladder cancer: a multicenter study. Insights Imaging 2024;15:262. [Crossref] [PubMed]
25. Zhu W, Chen L, Aphinyanaphongs Y, et al. Identification of patients at risk for pancreatic cancer in a 3-year timeframe based on machine learning algorithms. Sci Rep 2025;15:11697. [Crossref] [PubMed]