Preoperative Features-based Machine Learning Models with XAI for Improved Pancreatic Cancer Survival Prediction

Pancreatic cancer is the sixth principal cause of cancer-related deaths with a dismal survival rate largely due to aggressive tumor behavior. It is often asymptomatic in the early phase of the disease but already emerging with nodal and distant spread at the time of the initial diagnosis. An effective risk stratification for survival needs to be made to determine options for individualized and precise treatment options. Accurate prediction of pancreatic cancer survival remains a daunting challenge due to diverse tumor characteristics and varying clinical presentation at the time of diagnosis. In this study, several machine learning models were examined to predict survival outcomes based on pre-operative clinical and laboratory attributes. Random forest obtained the highest area under the precision recall curve (77.92), excellent recall (96.70) and F2-score (90.16) clearly pointing to its capability to predict reliably the positive class or patients who shall survive for more than one year where surgical resection would be very valuable due to its potential cure. The excellent recall also minimizes the false negatives thereby lessening the likelihood of withholding surgical treatment as potential cure for patients surviving more than one year. The most important attributes to determine survival for more than one year based on SHAP were abdominal pain, CA19-9 (carbohydrate antigen 19-9) and CEA (carcinoembryonic antigen). The strong predictive performance demonstrates the feasibility of integrating ML models into clinical practice, providing valuable diagnostic and prognostic support. This would pave the way for an enhanced decision-making process in identifying patients who shall benefit from curative surgical tumor resection and improving the planning of the social and care needs of pancreatic cancer patients leading to a positively enhanced outlook for this malignancy given its current abysmal prognosis.