Machine learning model for predicting pancreatic fistula after pancreatoduodenectomy

Abstract

Pancreaticoduodenectomy (PD) is a complex surgery used mainly to treat tumors and other pancreas disorders. PD is considered one of the most challenging surgeries because it may have several complications, including bleeding and infections in the surgical area, temporary or permanent diabetes, and pancreatic leakage (PL), which may lead to morbidity and mortality. In this study, we build an accurate and medically oriented machine learning model that predicts PL after PD based on patient markers collected only before the PD operation. The study is made using a real-world dataset for 397 Egyptian patients. The proposed machine learning pipeline starts with a data preprocessing step that handles the missing data values by the median values. In the next step, diverse interpretable classifiers, including logistic regression, random forest, decision tree, support vector machine, XGBoost, and AdaBoost, are utilized to predict the PL. Hyperparameter optimization is done using grid search with k-fold cross-validation. The results indicate that XGBoost achieves the highest marks, outperforming the state-of-the-art techniques in several evaluation metrics (i.e., accuracy= 91%, precision= 90.96%, recall= 91.0%, F1-score= 90.97%, and AUC=89.78%. The resulting model is accurate enough to be medically relevant for PL prediction in real healthcare settings.