Even after successful surgery to remove a pancreatic tumor, cancer returns in a large proportion of patients, often within the first year. This high recurrence rate is one of the main reasons pancreatic cancer remains so deadly despite surgical advances. Current monitoring methods — including CT scans and tumor markers measured after surgery — are often too insensitive to detect recurrence at an early, still-treatable stage.
Knowing which patients have the highest risk of recurrence at the time of surgery could allow doctors to provide more intensive monitoring and adjuvant treatment for high-risk individuals, potentially catching recurrences earlier when interventions are more effective. This study built machine learning models to predict 12-month pancreatic cancer recurrence using data available at the time of initial treatment.
The study used data from 585 pancreatic cancer patients treated at three clinical centers in Tehran between 2019 and 2023. The dataset included pathological features (tumor grade, histological type), treatment information (surgery type, chemotherapy, radiotherapy), and demographic data. Ten different machine learning algorithms were tested and compared, including Random Forest, Support Vector Machine, neural networks, and several ensemble methods.
Feature selection was performed using binary logistic regression to identify the most important predictors before building the ML models. Data was split into 70% training, 15% testing, and 15% validation sets. Model performance was evaluated using AUROC, sensitivity, specificity, positive predictive value, and F-score across all three data splits.
Random Forest and Support Vector Machine emerged as the best-performing algorithms, both achieving an AUROC of approximately 0.90. Random Forest achieved a positive predictive value of 92.05%, sensitivity of 90.67%, specificity of 92.95%, and F-score of 91.35% — performance levels that are clinically meaningful for a recurrence prediction task.
The weaker algorithms, including Naive Bayes and K-nearest neighbors, achieved AUROCs in the range of 0.52-0.59, essentially no better than chance. This wide variation in performance highlights the importance of algorithm selection and confirms that not all machine learning approaches are equally suited to clinical prediction problems.
Feature importance analysis from the Random Forest model revealed that lymph node metastasis (RI = 0.48), tumor grade (RI = 0.46), and whether the patient received radiotherapy (RI = 0.46) and chemotherapy (RI = 0.43) were the four most powerful predictors of 12-month recurrence. In contrast, gender and BMI showed minimal predictive value.
The strong predictive role of radiotherapy and chemotherapy is clinically interesting: patients who received these treatments had lower recurrence rates, which is consistent with their biological effects, and confirms that the model is capturing real clinical relationships rather than artifacts. Resection margin status and resectability classification were also important predictors, consistent with established clinical knowledge.
At the time of surgery, the model's key inputs — lymph node status, tumor grade, resection margins, and treatment plan — are all available. This means a recurrence risk score could be generated immediately after surgery, before the patient even leaves the hospital. High-risk patients could be immediately enrolled in more intensive follow-up programs with more frequent imaging.
The three-center design of this study, incorporating hospitals across Tehran, provides some degree of external validation and suggests the model may generalize across different clinical settings. The relatively large sample size of 585 patients also lends more confidence to the findings compared to single-center studies with smaller cohorts.
This study demonstrates that Random Forest and Support Vector Machine can predict 12-month pancreatic cancer recurrence with approximately 90% accuracy using clinical data routinely collected at the time of treatment. The models achieved consistent performance across training, testing, and validation sets, indicating genuine predictive ability rather than overfitting.
Future work should validate these models in non-Iranian populations and explore whether adding molecular biomarkers or imaging data can push performance even higher. The authors also note that real-time clinical deployment would require integration with hospital information systems, but the core algorithm is ready for prospective testing.