Explainable artificial intelligence for personalized prognosis in pancreatic cancer: A nationwide study from Taiwan.

Pancreatic cancer is highly aggressive with poor outcomes; current artificial intelligence (AI) prognostic models often lack interpretability and underutilize large-scale data. This study develops an explainable AI prognostic model for pancreatic cancer survival using Taiwan's national registry data, aiming to identify key prognostic factors, their interactions, non-linear relationships, and patient-specific survival variability. We analyzed 8,864 pancreatic cancer cases diagnosed between 2013 and 2021 from the Taiwan Cancer Registry. We developed three classes of prognostic models using regression-based, machine learning, and deep learning methods. The models were evaluated using nested cross-validation and time-dependent metrics, with Shapley additive explanations enhancing interpretability. XGBoost outperformed random survival forest and deep learning models in predicting pancreatic cancer survival. Key determinants included surgery, histological type, chemotherapy, tumor stage, and their interactions. Adenocarcinoma was associated with the highest mortality risk, whereas acinar cell and neuroendocrine carcinomas had lower risks (hazard ratios 0.768 and 0.660, respectively, vs adenocarcinoma). Chemotherapy showed the greatest mortality risk reduction in adenocarcinoma, while surgery was most strongly associated with reduced mortality in neuroendocrine tumors and adenocarcinoma, particularly in early-stage disease. The mortality reduction associated with chemotherapy increased in advanced stages and with age, plateauing around 65 years. Mortality risk rose faster with age in neuroendocrine carcinoma. Non-linear relationships emerged for age, smoking duration, and BMI: mortality increased gradually (0.69% per year) before age 65, rapidly afterward (2.41% per year); risk increased with longer smoking duration but plateaued between 10 and 30 years; and BMI exhibited a U-shaped risk, lowest at 26. This study demonstrates the potential of explainable AI for predicting pancreatic cancer survival by identifying key prognostic factors, nonlinear relationships, interactions, and patient-level variability, thereby revealing substantial heterogeneity in prognosis.