Pancreatic neuroendocrine neoplasms (PNENs) are a distinct subgroup of pancreatic cancers arising from hormone-producing cells rather than the ductal lining. While rarer than pancreatic ductal adenocarcinoma, their incidence has been rising, now reaching nearly 0.48 cases per 100,000 people.
PNENs account for about 7% of all neuroendocrine tumors. Their behavior ranges widely — from slow-growing tumors cured with surgery to aggressive cancers that spread quickly. Accurate survival prediction is important for guiding treatment and counseling patients.
This study compared four survival prediction models — three modern machine learning approaches and a traditional statistical method — using a large national cancer registry, to find the best tool for personalized prognosis in PNENs.
The researchers collected data on 3,239 PNEN patients from the SEER database, covering 18 cancer registries from 2000 to 2018. SEER is a publicly available US national cancer database, making results transparent and reproducible.
Ten clinicopathological variables were selected for modeling after Cox regression and collinearity analysis: age, sex, marital status, race, primary tumor site, AJCC stage, tumor grade, surgery, radiotherapy, and chemotherapy. Missing data were handled using MissForest imputation.
Patients were split 70/30 into training and testing sets. The four models compared were: traditional Cox proportional hazard (CoxPH), DeepSurv (a neural network extension of Cox), neural multitask logistic regression (NMTLR), and random survival forest (RSF).
Multivariate analysis identified age, sex, stage, and surgery as independent prognostic factors. Stage had the most dramatic effect: stage IV patients had nearly 5 times the mortality risk of stage I patients. Patients who underwent surgery had substantially better survival regardless of stage.
All four models showed good performance with integrated Brier scores well below the acceptable threshold of 0.25. Deep learning models (DeepSurv, NMTLR) performed comparably to RSF and outperformed traditional CoxPH on calibration of survival probabilities.
RSF achieved the best feature importance interpretability, clearly ranking stage, surgery, and age as the top predictors. All four models yielded consistent risk stratification when grouping patients into survival probability categories.
The researchers built an online application using Python's Streamlit package that allows clinicians to enter a patient's clinical characteristics and receive a predicted survival probability curve over time.
The tool is based on the best-performing model from the study, allowing personalized prognostic assessment at the point of care — helping oncologists have more informed conversations with PNEN patients about prognosis and treatment options.
Validating and deploying such a tool in routine clinical practice would require further prospective testing to confirm performance outside the SEER training environment, but the study demonstrates both feasibility and accuracy on a large population.
Advanced machine learning models — particularly DeepSurv and NMTLR — can match or exceed traditional survival statistics in predicting outcomes for PNEN patients, while also handling complex variable interactions that Cox models may miss.
The large, nationally representative SEER dataset ensures findings are relevant to real-world PNEN patients across the US, and the full analytical approach supports reproducibility.
Future work should incorporate molecular markers, imaging features, and treatment response data into these models to further improve predictive accuracy and clinical utility for this heterogeneous cancer type.