Borderline resectable pancreatic cancer (BRPC) sits in a gray zone — the tumor touches or partially wraps around major blood vessels, making surgery technically possible but high-risk. Unlike clearly resectable tumors, BRPC has a higher chance of cancer cells remaining at the surgical margin, which significantly worsens survival. The central debate is whether to operate immediately (upfront surgery) or try chemotherapy first to shrink the tumor.
Predicting which BRPC patients will survive long-term after upfront surgery would help surgeons make individualized treatment decisions. Current tumor markers like CA19-9 have poor sensitivity, and standard staging systems do not capture enough of the biological variation between patients. A machine learning model tailored to this specific group could fill this gap.
The study reviewed records from 104 BRPC patients who had upfront surgery with portal vein removal and reconstruction between 2011 and 2020. Patients were classified as short-term survivors (died within 2 years) or long-term survivors (alive beyond 2 years). This 2-year cutoff was chosen based on the median survival observed in this population.
Standard statistical methods (univariate and multivariate analysis) first identified independent predictors of survival. These were then used to build a logistic regression model, visualized as a nomogram — a graphical scoring chart. SHAP (Shapley Additive Explanations) values confirmed which variables drove the model's predictions. Five other machine learning algorithms including SVM, random forest, decision tree, and XGBoost were also tested for comparison.
Four variables emerged as independent predictors of survival: patient age, the length of blood vessel invasion by the tumor, whether the vessel's shape was malformed (morphological malformation), and whether nearby lymph nodes were enlarged (local lymphadenopathy). SHAP analysis independently confirmed these same four factors as the most influential.
The logistic regression model achieved AUC scores of 0.881 in the training set and 0.875 in the validation set — strong and consistent performance. In 10-fold cross-validation, logistic regression (AUC 0.864) outperformed SVM (0.693), random forest (0.789), decision tree (0.790), and XGBoost (0.726), suggesting simpler, interpretable models may be preferable when datasets are small.
A surgeon considering whether to operate on a BRPC patient could use this nomogram to calculate a survival probability score based on four readily available clinical and imaging measurements. Patients with low predicted long-term survival might be better served by neoadjuvant chemotherapy first, while high-scoring patients could proceed directly to surgery.
This matters because surgery for BRPC is complex and carries real risks. Unnecessary operations in patients unlikely to benefit impose surgical morbidity without survival gain. Conversely, delaying surgery in patients who would do well upfront may allow disease progression and lost curative opportunity.
One of the key takeaways is that logistic regression — one of the oldest and most interpretable statistical models — outperformed more complex machine learning methods in this small, specialized dataset. This is a reminder that model complexity should be matched to data availability: bigger, more complex models need bigger datasets to realize their advantages.
Future studies should validate this nomogram in multicenter cohorts and explore whether dynamic factors like tumor response to initial chemotherapy can be incorporated to further improve predictions for BRPC patients.