Pancreatic neuroendocrine tumors (PNETs) are a distinct subtype of pancreatic cancer arising from the hormone-producing cells of the pancreas. Although they are generally considered less aggressive than pancreatic ductal adenocarcinoma, PNETs still carry significant risk of spreading to other organs, with the liver being the most frequent site of distant metastasis.
At the time of diagnosis, a large proportion of PNET patients already have detectable liver metastases, which dramatically worsen outcomes. Identifying which patients are most likely to have or develop liver metastasis is therefore critical for treatment planning, surveillance intensity, and prognosis counseling.
Traditional clinical assessments rely on imaging and pathological staging, but these methods do not always capture the full complexity of individual patient risk. There is a clear need for evidence-based, quantitative tools that can integrate multiple clinical factors to yield personalized risk estimates.
This study addressed that gap by building and validating machine learning-based nomograms specifically designed to predict both the presence of synchronous liver metastasis at diagnosis and patient survival outcomes following metastasis. These nomograms translate statistical models into visual, clinician-friendly tools.
The study drew on the Surveillance, Epidemiology, and End Results (SEER) database, a large national cancer registry maintained by the National Cancer Institute. A total of 1,998 PNET patients diagnosed between 2010 and 2019 were identified, providing a statistically robust foundation for model building.
To identify the most predictive clinical variables, researchers employed two complementary machine learning approaches: LASSO (Least Absolute Shrinkage and Selection Operator) regression and Random Forest feature importance ranking. LASSO penalizes model complexity and shrinks less informative coefficients toward zero, effectively performing automatic variable selection.
Random Forest builds an ensemble of decision trees and ranks each variable by how much it contributes to accurate predictions across many tree splits. Using both methods together improves confidence that the selected variables are genuinely predictive rather than coincidentally associated.
The dataset was split into training and validation cohorts. A separate external validation cohort consisting of Chinese PNET patients was also used to test whether the models generalize across different populations and healthcare systems, an important step for establishing real-world applicability.
The final selected variables were incorporated into two separate nomograms: one for diagnosing liver metastasis at presentation and another for predicting survival prognosis among patients who already have liver involvement.
The analysis identified several independent risk factors for synchronous liver metastasis in PNET patients. These included histological grade, N stage (lymph node involvement), receipt of surgery, exposure to chemotherapy, tumor size, and the presence of bone metastasis. Higher tumor grade and larger size were consistently associated with increased liver metastasis risk.
The diagnostic nomogram achieved an area under the receiver operating characteristic curve (AUC) of 0.877 in the training cohort, indicating strong discriminative ability. When applied to external Chinese validation data, the AUC improved to 0.893, suggesting the model generalizes well beyond the original dataset.
The prognostic nomogram for predicting survival after liver metastasis demonstrated a C-index of 0.752, which reflects its ability to correctly rank patients by their relative survival risk. Calibration curves showed good agreement between predicted and observed outcomes at multiple time points.
Decision curve analysis confirmed that both nomograms provide net clinical benefit across a wide range of threshold probabilities, meaning they would be useful in real clinical decision-making without causing net harm through over- or under-treatment.
Nomograms are particularly valuable in oncology because they translate complex statistical models into visual scoring systems that clinicians can use at the bedside. By assigning point values to each predictor, a clinician can quickly calculate a patient's total score and read off the corresponding probability of liver metastasis or adverse survival outcome.
For PNET patients presenting with suspicious imaging findings, the diagnostic nomogram can help oncologists decide whether to pursue immediate biopsy or intervention versus watchful waiting. This is especially useful when imaging results are ambiguous or when patient comorbidities make aggressive workup risky.
The prognostic nomogram can guide decisions about treatment intensity after metastasis is confirmed. Patients identified as high risk for early mortality may benefit from more aggressive systemic therapy, enrollment in clinical trials, or earlier palliative care discussions.
The validation in a Chinese patient population is significant because it demonstrates cross-ethnic and cross-institutional generalizability, a common weakness of single-center or single-population oncology tools. This external validation strengthens the case for broader clinical adoption.
Earlier identification of patients at high risk for synchronous liver metastasis allows for more targeted surveillance strategies. Patients scoring high on the diagnostic nomogram might benefit from more frequent imaging or earlier referral to hepatobiliary surgical teams for evaluation of resectability.
The finding that chemotherapy receipt and surgical status are among the key predictors highlights how treatment history itself modifies metastatic risk. This reinforces the importance of integrating treatment variables into risk models rather than relying solely on tumor biology.
The inclusion of bone metastasis as a predictor for liver involvement reflects the broader systemic spread patterns of high-grade PNETs and suggests that a full metastatic workup may be warranted even when liver involvement is not yet confirmed on initial imaging.
Future directions could include integrating molecular markers such as Ki-67 index and mutational profiles into the nomogram framework, potentially further improving predictive accuracy and enabling truly personalized risk stratification for PNET patients.
This study successfully constructed and externally validated two machine learning-based nomograms for PNET patients: one predicting synchronous liver metastasis at diagnosis and one predicting survival prognosis after metastasis. Both models demonstrated strong discriminative performance and good calibration.
The use of both LASSO regression and Random Forest for variable selection ensured that only robust, stable predictors were included, reducing the risk of overfitting and improving interpretability. The nomograms are straightforward to use in clinical settings without requiring specialized software.
External validation in a Chinese cohort confirmed that these tools can be applied across diverse patient populations, which is a prerequisite for wide clinical adoption. The authors encourage prospective validation studies to further confirm real-world utility.
This work adds to the growing evidence that machine learning tools can meaningfully improve risk stratification in rare cancers like PNETs, where large randomized trial evidence is limited and individualized prediction is particularly valuable.