Use of artificial intelligence for risk prediction of postoperative complication after pancreaticoduodenectomy

Introduction: Despite advances in medical management, the incidence of complications after pancreatic surgery remains still high (between 40% to 50%). Morbidity could lead to a longer length of hospital stay, higher mortality, worse quality of life, higher hospital cost, and finally worsen oncologic outcome also in consequence of delay of adjuvant therapy. Several studies have tried to investigate the preoperative risk of surgical complication and a lot of predictive scores have been proposed to stratify a patient’s risk of postoperative morbidity and mortality; however, this predictive score has been developed with standard statistical methods, considering that the variables interact in a linear and additive fashion linear while the interaction of comorbidities and markers of disease could be different from linear. Conversely, machine learning, thanks to the combination of a vast number of variables in a non-linear way, may be superior to the standard statistical methods currently used by clinicians to predict the postoperative outcome. Our primary end point was to investigate through artificial intelligence the preoperative fields that could affect postoperative outcome after pancreaticoduodenectomy for benign and malignant disease and develop a calculator that can define the real risk of any postoperative complication for each patient candidate to pancreaticoduodenectomy. Material and methods: 496 patients who underwent pancreaticoduodenectomy for benign and malignant tumors between 2011 and 2022 at the Department of Hepato-pancreatobiliary Surgery, Pederzoli Hospital (Peschiera del Garda, Verona) were retrospectively collected from a prospectively maintained database. A random forest model was developed to predict the risk of any postoperative complication after pancreaticoduodenectomy. The study population has been divided in training cohort (80%) and a testing cohort (20%). Results: The primary model evaluation metric was the area under the receiver operating characteristic curve (ROC-AUC) that was good as for training cohort (AUC= 0.87) as for testing cohort (AUC=0.72). The twelve most influential variables were dilated pancreatic duct, weight loss >10% prior to surgery, diagnosis of cystic lesion, diabetes, previous abdominal surgery, preoperative CA 19.9 serum level, high bilirubin level, preoperative chemotherapy, dilated bile duct, ASA Score, Surgical Approach and “Other diagnosis” (GIST, groove pancreatitis, duodenal tumor). A calculator called "PanRisk Calculator" has been developed based on the algorithm, which is available online and it is easy to use in daily clinical practice. Conclusion: Preoperative machine learning prediction of the development of any postoperative complication may improve preoperative planning, postoperative mitigation strategy, and, subsequently, patient outcomes.