Interpreting Decision Support from Multiple Classifiers for Predicting Length of Stay in Patients with Colorectal Carcinoma

A precise estimation of patient length of stay is important for systematically managing both hospital unit resources (medication, equipment, beds) and the distribution of personnel. This is true for hospitalization following any disease, however the particularities of each trigger a different observation/recovery period. The current study investigates this problem in the context of cancer of the colorectal type on a discrete data set. Several classifiers from distinct conceptual families provide an estimation or even further information on the length of stay of patients that had been operated of cancer in certain stages and invasion at various parts of the colon or rectum. Support vector machines and neural networks give a black box prediction of the hospitalization period, while decision trees and evolutionary algorithms additionally offer the underlying rules of decision. Results are also compared to those of ensemble state-of-the-art techniques: bagging, boosting and random forests. A Wilcoxon rank-sum test demonstrates that the support vector machines, the decision trees and the ensembles are significantly better than the neural networks and the evolutionary algorithms. They also show substantial agreement following Cohen’s kappa coefficient to the original outputs. The highest agreement is between the results of support vector machines (SVM)–bagging (0.84) and decision trees (DT)–bagging (0.87). A potential SVM–EA tandem is also investigated, as a more collaborative means towards supporting decision making; its accuracy came similar to that of the plain EA. Faced with the results of each, the professional is given a manner of how to interpret the amalgam of computational opinions and justifications given in support of his/her decision.