Can machine learning predict resecability of a peritoneal carcinomatosis?

Background: Approximately 20% of initially eligible patients in a HIPEC procedure eventually underwent a simple surgical exploration. These procedures are called 'open & close' (O & C) representing up to 48% of surgery. The objective of this study was to predict the resecability of peritoneal carcinomatosis using a machine-learning model for decision-making support, for eligible patients of HIPEC. Methods: The study was conducted as an intention to treat based on three databases including a prospective, between January 2000 and December 2015. A propensity score allowed us to obtain two groups of comparable and matched patients. Subsequently, several algorithm models of classification were studied (simple classification, conditional tree, support vector machine, random forest) to determine the model having the best performance and accuracy. Results: Two groups of 155 patients were obtained: one group without resection and one group with resection. Nine criteria of non-resecability reflecting the organ involvement have been retained. They were coded according to their importance. Five classification algorithms were tested. The training data included 218 patients and 92 test data. The random forest model exhibited the best performance with an accuracy of close to 98%. Only two errors of predictions were observed. Discussion: The largest number of patients will allow us to improve the precision prediction. Gathering more data such as biologic, radiologic, and even laparoscopic features, should improve the knowledge of the disease and decrease the number of 'O & C' procedures.